Re: [PATCH v4 0/2] refactor sqthread cpu binding logic

On 9/2/21 10:48 AM, Michal Koutný wrote:

Hello Hao.

On Wed, Sep 01, 2021 at 08:43:20PM +0800, Hao Xu [off-list ref] wrote:

quoted

This patchset is to enhance sqthread cpu binding logic, we didn't
consider cgroup setting before. In container environment, theoretically
sqthread is in its container's task group, it shouldn't occupy cpu out
of its container.

I see in the discussions that there's struggle to make
set_cpus_allowed_ptr() do what's intended under the given constraints.

IIUC, set_cpus_allowed_ptr() is conventionally used for kernel threads
[1]. But does the sqthread fall into this category? You want to have it
_directly_ associated with a container and its cgroups. It looks to me
more like a userspace thread (from this perspective, not literally). Or
is there a different intention?

It's an io thread, which is kind of a hybrid - it's a kernel thread in
the sense that it never exits to userspace (ever), but it's a regular
thread in the sense that it's setup like one.

It seems to me that reusing the sched_setaffinity() (with all its
checks and race pains/solutions) would be a more universal approach.
(I don't mean calling sched_setaffinity() directly, some parts would
need to be factored separately to this end.) WDYT?

We already have this API to set the affinity based on when these were
regular kernel threads, so it needs to work with that too. As such they
are marked PF_NO_SETAFFINITY.

[1] Not only spending their life in kernel but providing some
delocalized kernel service.

That's what they do...

-- 
Jens Axboe