Re: [Qemu-devel] [Qemu-ppc] [PATCH v3 0/4] target-ppc: Add FWNMI support in qemu for powerKVM guests

[Alexander Graf]

On 07.08.15 05:37, Sam Bobroff wrote:
> Hello Aravinda and all,
> 
> On Wed, Jul 08, 2015 at 01:58:13PM +0530, Aravinda Prasad wrote:
>> On Friday 03 July 2015 11:31 AM, David Gibson wrote:
>>> On Thu, Jul 02, 2015 at 07:11:52PM +1000, Alexey Kardashevskiy wrote:
>>>> On 04/02/2015 03:46 PM, David Gibson wrote:
>>>>> On Thu, Apr 02, 2015 at 03:28:11PM +1100, Alexey Kardashevskiy wrote:
>>>>>> On 11/19/2014 04:48 PM, Aravinda Prasad wrote:
>>>>>>>
>>>>>>>
>>>>>>> On Tuesday 11 November 2014 08:54 AM, David Gibson wrote:
>>>>>>>
>>>>>>> [..]
>>>>>>>
>>>>>>>>
>>>>>>>> So, this may not still be possible depending on whether the KVM side
>>>>>>>> of this is already merged, but it occurs to me that there's a simpler
>>>>>>>> way.
>>>>>>>>
>>>>>>>> Rather than mucking about with having to update the hypervisor on the
>>>>>>>> RTAS location, they have qemu copy the code out of RTAS, patch it and
>>>>>>>> copy it back into the vector, you could instead do this:
>>>>>>>>
>>>>>>>>   1. Make KVM instead of immediately delivering a 0x200 for a guest
>>>>>>>> machine check, cause a special exit to qemu.
>>>>>>>>
>>>>>>>>   2. Have the register-nmi RTAS call store the guest side MC handler
>>>>>>>> address in the spapr structure, but perform no actual guest code
>>>>>>>> patching.
>>>>>>>>
>>>>>>>>   3. Allocate the error log buffer independently from the RTAS blob,
>>>>>>>> so qemu always knows where it is.
>>>>>>>>
>>>>>>>>   4. When qemu gets the MC exit condition, instead of going via a
>>>>>>>> patched 0x200 vector, just directly set the guest register state and
>>>>>>>> jump straight into the guest side MC handler.
>>>>>>>>
>>>>>>>
>>>>>>> Before I proceed further I would like to know what others think about
>>>>>>> the approach proposed above (except for step 3 - as per PAPR the error
>>>>>>> log buffer should be part of RTAS blob and hence we cannot have error
>>>>>>> log buffer independent of RTAS blob).
>>>>>>>
>>>>>>> Alex, Alexey, Ben: Any thoughts?
>>>>>>
>>>>>>
>>>>>> Any updates about FWNMI? Thanks
>>>>>
>>>>> Huh.. I'd completely forgotten about this.  Aravinda, can you repost
>>>>> your latest work on this?
>>>>
>>>>
>>>> Aravinda disappeared...
>>>
>>> Ok, well someone who cares about FWNMI is going to have to start
>>> sending something, or it won't happen.
>>
>> I am yet to work on the new approach proposed above. I will start
>> looking into that this week.
> 
> The RTAS call being discussed in this thread actually has two vectors to patch
> (System Reset and Machine Check), and the patches so far only address the
> Machine Check part. I've been looking at filling in the System Reset part and
> that will mean basing my code on top of this set.  I would like to keep the
> same style of solution for both vectors, so I'd like to get the discussion
> started again :-)
> 
> So (1) do we use a trampoline in guest memory, and if so (2) how is the
> trampoline code handled?
> 
> (1) It does seem simpler to me to deliver directly to the handler, but I'm
> worried about a few things:
> 
> If a guest were to call ibm,nmi-register and then kexec to a new kernel that
> does not call ibm,nmi-register, would the exception cause a jump to a stale
> address?

Probably - how does that get handled today with pHyp? Does pHyp just
override the actual exception vector code and thus the kexec'ed code
path gets overwritten?

I don't remember the original patch set fully, but if all we need is to
override 0x200, why can't we replace the code with

  mtsprg scratch, r0
  li r0, HCALL_KVM_MC
  sc 1

then there is no complexity in that code at all with dynamically patched
bits. Or am I missing the obvious?

> 
> Because we're adding a new exit condition, presumably an upgraded KVM would
> require an upgraded QEMU: is this much of a problem?

Well, you would keep default behavior identical. On nmi-register QEMU
would send an ioctl to KVM, telling it to route 0x200 to QEMU instead
(just like with breakpoints). So old QEMU would still work the same way
and new QEMU with old KVM would simply get non-working MC intercepts.

> 
> From some investigation it looks like the current upstream KVM already
> forwards (some) host machine checks to the guest by sending it directly to
> 0x200 and that Linux guests expect this, regardless of support in the host for
> ibm,nmi-register (although they do call ibm,nmi-register if it's present).
> 
> (2) If we are using trampolines:
> 
> About the trampoline code in the v3 patches: I like producing the code using
> the assembler, but I'm not sure that the spapr-rtas blob is the right place to
> store it. The spapr-rtas blob is loaded into guest memory but it's only QEMU
> that needs it. It seems messy to me and means that the guest could corrupt it.

If you like, rename the blob. My original proposal was to just use
well-known offsets inside the blob that get indicated through a function
pointer table at the beginning/end/known location.

> 
> Some other other options might be:
> 
> (a) Create a new blob (spapr-rtas-trampoline?) just like the spapr-rtas one 
> but
> only load it when ibm,nmi-register is called, and only into QEMU not the guest
> memory. There would be another "BIOS" blob to install, and it wouldn't really
> actually be BIOS but it seems like it would work easily.  Since we need a
> second, different, trampoline for System Reset, I would then need to add yet
> another blob for that... Still, this doesn't seem so bad. I suppose we could
> add some structure to the blob (e.g. a table of contents at the start) and fit
> both trampolines in, but that's inventing yet another file format... ugh.

Yes, I think inventing our own file format is the best way forward. It
shouldn't be too bad. Just reserve say 10 64bit values somewhere are use
then as function table.

> (b) As above but assemble the trampoline code into an ELF dynamic library
> rather than stripping it down to a raw binary: we could use known symbols to
> find the trampolines, even the patch locations, so at least we wouldn't be
> inventing our own format (using dlopen()/dlsym()... I wonder if this would be
> OK for all platforms...).

We have our own ELF loader in QEMU so it's workable, but I think it's
actually more complicated and harder at the end of the day than (a).

> 
> (c) Assemble it (as above) but include it directly in the QEMU binary by
> objcopying it in or hexdumping into a C string or something similar. This 
> seems
> fairly neat but I'm not sure how people would feel about including "binaries"
> into QEMU this way.  Although it would take some work in the build system, it
> seems like a fairly neat solution to me.

We tried to move away from code as hex arrays in QEMU to make it easier
for people to patch things when they want to. But then again if we're
talking 3 instructions it might not be the worst option.


Alex