On 11/2/2018 2:45 PM, Chris Friesen wrote:
On 11/2/2018 11:51 AM, Dr. David Alan Gilbert wrote:
so the fix is Fam's 'aio: Do aio_notify_accept only during blocking
aio_poll'. I see you're running the qemu-kvm-ev from centos, if I read
the version tea-leaves right, then I think that patch is in the
2.10.0-21.el7_5.7.1 package I can see.
It looks like we do not have this fix in our code, will definitely be
giving it a try.
Well, it appears that this was not the fix. With the patches applied we
still saw the problem.
One of my coworkers has been investigating further, and he sent me the
following information:
"One of the things the BIOS does before initializing the VGA is to start
the CPUs. More specifically, from CPU-0, it broadcasts a Setup
InterProcess Interrupt (SIPI) to all the remaining processors so that
they wake up and are initialized. CPU-0 busy-waits for all the
Instance’s CPUs to be initialized, and once that happens, it moves on to
continue booting (which includes initializing the VGA). In our case,
one CPU fails to respond to the SIPI causing CPU-0 to keep busy waiting
for it. I am in the process of adding more instrumentation to qemu so
that we can know whether the missing CPU did not receive the SIPI or
what happened."
Chris
