Re: [Qemu-devel] [RFC PATCH v1 16/22] i386: pc: load OS images at fixed location in SEV-enabled guest

[Brijesh Singh]

Hi Michael and Paolo,

On 09/13/2016 01:37 PM, Michael S. Tsirkin wrote:
> On Tue, Sep 13, 2016 at 10:49:29AM -0400, Brijesh Singh wrote:
> > Typically linux kernel, initrd and cmdline are build and loaded
> > into guest memory through linux optionrom. The linux optionrom is
> > probed and executed by SeaBIOS. This method will not work for
> > SEV-enabled guest.
> >
> > In SEV-enabled guest all the code and data must be copied using SEV
> > launch command prior to starting the guest (i.e before first vmrun).
> > The data copied using SEV launch command will be encrypted using guest
> > owner's key. This patch loads kernel, initrd and cmdline blobs at fixed
> > location into guest memory and builds etc/sev_cfg config file. The cfg
> > file provide the below structure
> >
> > struct sev_cfg {
> >         u32 kernel_addr, initrd_addr, cmdline_addr;
> >         u32 kernel_size, initrd_size, cmdline_size;
> > }
> >
> > The config file can be used by SeaBIOS to locate OS images into guest
> > RAM and build linux boot entry code.
> >
> > Signed-off-by: Brijesh Singh <address@hidden>
>
> I don't think we want to give users this kind of
> control over how we manage memory internally
> for what is essentially a debugging feature at this point.
>
> Isn't there a way to first launch guest, and then have it
> encrypt itself once it's running?
>
> If not, I guess it's not too bad if -kernel does not
> work with sev debug feature - just load kernel from disk.

I'm working on v2 and getting ready for another review but not sure how 
to address this feedback. For now, I can drop this patch from the series 
and get other patches reviewed. But I would like to get some direction 
on how do I go about adding -kernel support for SEV guest.

Looking through -kernel option code, I see that qemu uses optionrom 
(linuxboot.bin or linuxboot_dma.bin) to load kernel specified through 
-kernel option.

- linuxboot optionrom uses "rep ins" instruction to fetch the kernel and 
initrd images into guest memory. SEV does not support string I/O, so 
this method will not work for SEV. But we can consider unrolling the 
string I/O operation into a loop, something similar to this KVM patch [1].

- linuxboot_dma - as name suggests it uses DMA. On SEV guest DMA should 
to be done on shared pages and since SeaBIOS runs into non-PAE mode 
hence it will not able to create a shared pages for this to work.

please let me know your thoughts on the following approaches:

- update linuxboot.S to unroll the string I/O instruction into a loop 
for SEV guest. This may work but the boot will be very slow, fetching 
kernel and initrd via 'inb' will increase the boot time significantly.

or

- Create a SEV specific linuxboot rom. The ROM will use a special port 
to communicate kernel and initrd load address to qemu. In qemu when we 
see the request on special port we map the pre-encrypted kernel and 
initrd at the requested location. This maybe similar to linuxboot_dma 
with exception that instead of copying the data into guest memory we 
somehow map the pre-encrypted image. I have not looked into much detail 
and not sure if we can do something like this. If possible then this 
seems like a bit better approach.


[1] http://marc.info/?l=kvm&m=147191048324519&w=2