Question on QEMU and P2M

Question on QEMU and P2M

[Wei Huang]

I am facing a qemu-dm problem when setting up 1GB page table to P2M.

Setting 1GB page table is very straight forward: just skip the last two 
levels of p2m table. While doing gfn_to_mfn translation, I have to check 
PSE bits in level 3 page table entry first. If this bit is 1, I 
calculate the target mfn and return immediately. Two gfn_to_mfn 
translation functions, p2m_gfn_to_mfn_current() and p2m_gfn_to_mfn(), 
are changed in this way.

However, QEMU dm starts to function weirdly with this change. For 
instance, linux guest won't be able to locate hard disk because port IO 
access returns incorrect values. I couldn't control guest keyboard either.

I am thinking that there are some Xen codes (related to qemu) which 
doesn't goes through gfn_to_mfn translation inside p2m.c. It might 
assume p2m table always has L2 P2M table and does fast translation. But 
I haven't found such areas yet.

Any suggestion of suspicious code areas to look at?

-Wei

Re: Question on QEMU and P2M

[Ian Jackson]

Wei Huang writes ("[Xen-devel] Question on QEMU and P2M"):
> I am facing a qemu-dm problem when setting up 1GB page table to P2M.
...
> Any suggestion of suspicious code areas to look at?

This is unlikely to be a problem in qemu-dm; qemu-dm gets told the
guest-physical of IO requests addresses by Xen.  I would suggest that
the starting point would be xen/arch/x86/hvm/io.c and its subroutines
and callers.  That's the place where the IO requests are collected and
sent on to the device model (ie, qemu-dm).

Ian.