Re: [dpdk-dev] [PATCH] eal: fix ctrl thread affinity with --lcores

[David Marchand <david.marchand@redhat.com>]

On Tue, Jul 30, 2019 at 6:32 PM Johan Källström
<johan.kallstrom@ericsson.com> wrote:
>
> Hi, for the online check I referred to the check of "default_set" via the initial thread affinity.
>
> I see that pthread_getaffinity_np returns an already and:ed mask, was under the impression that pthread_getaffinity_np would return the same mask as was set using pthread_setaffinity_np.
> Looking on the implementation I see that it has been implemented on this line (https://github.com/torvalds/linux/blob/master/kernel/sched/core.c#L5242) for the last decade. Don’t know how this is implemented on FreeBSD or Windows.

Afaics on FreeBSD, trying to set an unknown core is rejected with the
cpuset tool.
Not sure at which level the refusal is (cpuset, libc, kernel).


> Below is some example runs without the online cpu check running inside the exclusive cpuset 1-3,19,79 with cpu 79 offline.
> Added a print statements after each consecutive calculation just to verify what the different steps.
>
> Nice that you were able to reproduce the bug, the fix looks good otherwise :) .

Ok, I understand your concern now, but afaiu you confirm that there is
no issue with this patch.

We are really close to -rc3.
Can you send a review or test tag?


Thanks.

--
David Marchand