|
This describes the mechanics for the call but there are
several things I don't understand,
especially after seeing the patch:
1) Why would these modules be in Xen at all?
They appear to be (at least very similar to) the
code from Domain0. Why is the
functionality duplicated from Domain0? Are you
trying to eliminate
Domain0?
2) How do these modules get dynamically "loaded"?
Xen/ia64 currently has no disk
access or any driver support; all bits
are currently put in memory at boot time by elilo.
(And if they are not dynamically loaded,
why are they "kernel modules"?)
3) Is there any similarity between this and
Xen/x86(VTx)? I looked for similar code/mechanism
on Xen/x86 to help me understand your intent
but didn't find anything.
Unless I'm completely misunderstanding, this seems to be
heading in the direction
of major architectural departures from the direction of
Xen. I think we should have some
more discussion before applying these "km"
patches.
Could you provide an architectural overview of what you are
trying to do here rather
than just the mechanics of how the VTLB and hypercall
work?
Thanks,
Dan
Hi,
Dan,
We
implemented VTLB infrastructure per domain on XEN/IPF, which tracked guest tlb
information. VTLB has fixed size, then when using out of it, HV will desert
all VTLB and recycle VTLB. We add a flag “lock” in VTLB entry, when HV recycle
VTLB, the VTLB entry with “lock” flag will not be deserted, but if Guest uses
instructions like “ptc” to purge tlb, the VTLB entry with “lock” flag will be
deserted.
Before
invoking hypercall,
Kernel
Module pseudo read parameter once per page size to make sure the translation
for this parameter has been inserted in VTLB infrastructure. Then kernel
Module will call another new hypercall which don’t need pointer parameter to
lock above translation in VTLB infrastructure.
Then
invoking this hypercall,
In HV, HV
will use copy_from_user() or copy_to_user() to get parameter or return result.
In these functions, HV will lookup VTLB infrastructure to get corresponding
guest physical address of the parameter, because this translation has been
locked in VTLB, HV definitely can find it, then HV can get corresponding
machine address from physical to machine address table, as we know HV use
region 7 for identity mapping, HV can get identity virtual address for that
machine address, at last, HV do normal copy operation using this identity
virtual address.
After this
hypercall,
Guest
application definitely will unmap the memory allocated for passing hypercall
parameter, and this operation definitely will purge tlb for this address, so
the “lock” VTLB entry in VTLB infrastructure can be
recycled.
We had
tested this parameter passing mechanism for several hypercalls, such as
GETMEMLIST, and it works well.
Could we
check in this patch and discuss further?
From: Magenheimer, Dan (HP Labs Fort
Collins) [mailto:dan.magenheimer@xxxxxx]
Sent: 2005年6月21日 22:04
To: Xu, Anthony
Cc:
xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
Subject: RE: how to put kernel module in
xen/ipf
Yes, the hypercall
parameter mechanism is still evolving for Xen/ia64. Very
few
hypercalls are
necessary to run domain0, so experimentation with different
hypercall
mechanisms has waited
until multi-domain work.
Can you explain more
about kernel modules? I know (roughly) how they
work for Linux, but
not how they are used on Xen/x86. Others on this
list might like to
learn too, so perhaps you could explain the design
in
detail?
Thanks,
Dan
From: Xu, Anthony
[mailto:anthony.xu@xxxxxxxxx]
Sent: Monday, June 20, 2005 11:10
PM
To: Magenheimer, Dan (HP
Labs Fort Collins)
Cc:
xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
Subject: how to put kernel module in
xen/ipf
Hi,
Dan,
XEN/IPF
kernel module has a lot
of difference from XEN/ia32, especially the mechanism of passing
hypercall parameter. Currently we create directory “km” under xen/arch/ia64,
and put kernel module code in that
directory.
Any
comment?
-Anthony
|
Home