Re: [MirageOS-devel] Crash on x86 with doubles

[Thomas Leonard <talex5@xxxxxxxxx>]

On 2 July 2014 12:45, Anil Madhavapeddy <anil@xxxxxxxxxx> wrote:
> On 2 Jul 2014, at 13:04, Thomas Leonard <talex5@xxxxxxxxx> wrote:
>
>> On 2 July 2014 09:02, Anil Madhavapeddy <anil@xxxxxxxxxx> wrote:
>>> On 1 Jul 2014, at 22:53, Thomas Leonard <talex5@xxxxxxxxx> wrote:
>>>
>>>> I've got printing of floating point values working now on ARM, but
>>>> while testing I discovered it crashes on x86.
>>>>
>>>> This turned out to be unrelated to OCaml or Mirage. It seems that
>>>> passing a double argument to any varargs function causes a crash on
>>>> x86_64 if done from a Mini-OS thread. Can any x86 gurus here shed some
>>>> light on what might cause that? x86 is a mystery to me.
>>>
>>> First thing to check is that %rsp stack alignment is valid when you
>>> enter the function, as it's required to be 16-byte aligned by the
>>> x86_64 ABI if you're using the XMM registers.
>>>
>>> An ancient diff in our local MiniOS that fixed one instance of this
>>> is below. (I can't reproduce this instantly to give you a more precise
>>> diagnosis as I'm travelling at the moment, apologies!)
>>
>> Ah, that was indeed the problem. I changed the x86 arch_create_thread
>> to skip 8 bytes and now it works!
>>
>> Though, it seemed that it *was* 16-byte aligned before, and is now off
>> by 8 bytes, which I don't understand. I'll send the patch to xen-devel
>> and they can decide if it's the right fix.
>
> Stack misalignment is the bug that keeps on giving.  It's really easy
> for an unrelated change to misalign the stack under some conditions, and
> to not notice until a future 16-byte alignment operation happens.
>
> Found several of these in OCaml as well...
> http://caml.inria.fr/mantis/view.php?id=5700
> http://caml.inria.fr/mantis/view.php?id=6038
>
> Incidentally, I notice that the thread machinery in MiniOS is now active
> again.  I removed it from our MiniOS on the basis that it wasn't used
> (but there's no harm in leaving it there now).  But that might explain
> why the issue cropped up again, as I never tested that code path in
> upstream MiniOS.

I've removed it again now (mirage-platform now overrides Mini-OS's
start_kernel).


-- 
Dr Thomas Leonard        http://0install.net/
GPG: 9242 9807 C985 3C07 44A6  8B9A AE07 8280 59A5 3CC1
GPG: DA98 25AE CAD0 8975 7CDA  BD8E 0713 3F96 CA74 D8BA

_______________________________________________
MirageOS-devel mailing list
MirageOS-devel@xxxxxxxxxxxxxxxxxxxx
http://lists.xenproject.org/cgi-bin/mailman/listinfo/mirageos-devel