Re: [PATCH 1/4] x86: move syscall trampolines off the stack

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 19/05/15 07:41, Jan Beulich wrote:
>>>> On 18.05.15 at 20:39, <andrew.cooper3@citrix.com> wrote:
>> On 18/05/15 13:46, Jan Beulich wrote:
>>> This is needed as stacks are going to become non-executable. Use
>>> separate stub pages (shared among suitable CPUs on the same node)
>>> instead.
>>>
>>> Signed-off-by: Jan Beulich <jbeulich@suse.com>
>> Can you please include a description of how you intend the stubs to
>> function, and how they are layed out?  Parts of the code look like a
>> single page per stub, while other bits look like several stubs per page.
> I'm adding this to the already present description:
>
> Stub areas (currently 128 bytes each) are being split into two parts -
> a fixed usage one (the syscall ones) and dynamically usable space,
> which will be used by subsequent changes to hold dynamically generated
> code during instruction eumlation.
>
> While sharing physical pages among certain CPUs on the same node, for
> now the virtual mappings get established in distinct pages for each
> CPU. This isn't a strict requirement, but simplifies VA space
> management for this initial implementation: Sharing VA space would
> require additional tracking of which areas are currently in use. If
> the VA and/or TLB overhead turned out to be a problem, such extra code
> could easily be added.

Thanks - this clarifies things a lot.

>
>> (Personally, I would split the stub allocation/mapping/freeing into a
>> patch separately to moving the syscall trampolines, as each are
>> moderately complicated changes.)
> I'm afraid this wouldn't work: The freeing of the stub page depends
> on finding the first byte of each stub area being other than 0xCC in
> order for the page to not get freed. Yet only the setting up of the
> syscall stubs guarantees this (and I'm not really looking forward to
> add - however little - code to store a placeholder instead).

Ah ok.

~Andrew