Re: [PATCH v2 1/4] x86/hyper-v: move synic/stimer control structures definitions to hyperv-tlfs.h

From: Vitaly Kuznetsov
Date: Thu Nov 29 2018 - 06:36:25 EST

Nadav Amit <nadav.amit@xxxxxxxxx> writes:

>> On Nov 28, 2018, at 5:07 AM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
>> On Wed, 28 Nov 2018, Vitaly Kuznetsov wrote:
>>> Nadav Amit <nadav.amit@xxxxxxxxx> writes:
>>>> On a different note: how come all of the hyper-v structs are not marked
>>>> with the âpacked" attribute?
>>> "packed" should not be needed with proper padding; I vaguely remember
>>> someone (from x86@?) arguing _against_ "packed".
>> Packed needs to be used, when describing fixed format data structures in
>> hardware or other ABIs, so the compiler cannot put alignment holes into
>> them.
>> Using packed for generic data structures might result in suboptimal layouts
>> and prevents layout randomization.
> Right, I forgot about the structs randomization. So at least for it, the
> attribute should be needed.

Not sure when randomization.s used but Hyper-V drivers will of course be
utterly broken with it.

> To prevent conflicts, I think that this series should also add the
> attribute in a first patch, which would be tagged for stable.

As the patchset doesn't add new definitions and as Paolo already queued
it I'd go with a follow-up patch adding "packed" to all hyperv-tlfs.h
structures. The question is how to avoid conflicts when Linus will be
merging this. We can do:
- Topic branch in kvm
- Send the patch to x86, make topic branch and reabse kvm
- Send the patch to kvm
- ... ?

Paolo/Thomas, what would be your preference?