Only Interlagos, and I solved it. It was task_packer doing an
inappropriate copying of the thread's xstate.
John
On Mon, Apr 2, 2012 at 6:05 AM, Adam Young <ayoung(a)redhat.com> wrote:
On 03/30/2012 03:55 PM, John Hawkes wrote:
>
> The most significant problem I'm dealing with now has to do with a
> failing bproc_move() to an Interlagos server. When the bproc kernel
> module on the Interlagos target node returns to user space, the
> process suffers a rather immediate segfault. This works on
> non-Interlagos. This also works when doing a bproc_move() *from* an
> Interlagos node to another non-Interlagos node. The failure just
> occurs when the target node is Interlagos.
>
> A bpsh of a binary image to an Interlagos node works. The difference
> between bpsh and bproc_move() is that the bproc_move'd process has
> already begun to execute. Its libraries are loaded. Versus the bpsh
> case, where the program's memory image is still rather pristine.
>
Sounds like a task_packer problem. To verify, you could run a process on a
head node that dumps to a file, and then read that file back into the
kernel.
Does this happen on other relatively recent AMD machines, or only
Interlagos?