Re: [PATCH 2/2] io_uring: consider cgroup setting when binding sqpoll cpu

在 2021/9/8 上午12:54, Tejun Heo 写道:

Hello,

On Fri, Sep 03, 2021 at 11:04:07PM +0800, Hao Xu wrote:

quoted

quoted

Would it make sense to just test whether set_cpus_allowed_ptr() succeeded
afterwards?

Do you mean: if (sqd->sq_cpu != -1 && !set_cpus_allowed_ptr(current,
cpumask_of(sqd->sq_cpu)))

I'm not familiar with set_cpus_allowed_ptr(), you mean it contains the
similar logic of test_cpu_in_current_cpuset?

It's kinda muddy unfortunately. I think it rejects if the target cpu is
offline but accept and ignores if the cpu is excluded by cpuset.

quoted

This is a bit beyond of my knowledge, so you mean if the cpu back
online, the task will automatically schedule to this cpu? if it's true,
I think the code logic here is fine.

quoted

offline and online. If the operation takes place while the cpu happens to be
offline, the operation fails.

It's ok that it fails, we leave the option of retry to users themselves.

I think the first thing to do is defining the desired behavior, hopefully in
a consistent manner, rather than letting it be defined by implementation.
e.g. If the desired behavior is the per-cpu helper failing, then it should
probably exit when the target cpu isn't available for whatever reason. If
the desired behavior is best effort when cpu goes away (ie. ignore

Hmm, I see. First I think we should move the set_cpus_allowed_ptr() to
sqthread creation place not when it is running(not sure why it is
currently at the beginning of sqthred itself), then we can have
consistent behaviour.(if we do the check at sqthread's running time,
then no matter we kill it or still allow it to run when cpu_online
check fails, it's hard to let users know the result of their cpu binding
since users don't know the exact time when sqthread waken up and begin
to run, so that they can check their cpu binding result).
Second, I think users' cpu binding is a kind of 'advice', not 'command'.
So no matter cpu_online check succeeds or fails, we still let sqthread
run, meanwhile return the cpu binding result to the userspace.
Anyway, I'd like to know Jens' thoughts on this.

affinity), the creation likely shouldn't fail when the target cpu is
unavailable but can become available in the future.

Thanks.