Re: [PATCH v3] eventpoll: Fix priority inversion problem

[Nam Cao <namcao@linutronix.de>]

On Fri, Jul 11, 2025 at 08:09:12PM +0800, Xi Ruoyao wrote:
> On Fri, 2025-07-11 at 11:58 +0200, Nam Cao wrote:
> > On Fri, Jul 11, 2025 at 05:48:56PM +0800, Xi Ruoyao wrote:
> > > Sadly, still no luck.
> > 
> > That's unfortunate.
> > 
> > I'm still unable to reproduce the issue, so all I can do is staring at the
> > code and guessing. But I'm out of idea for now.
> 
> Same as I.  I tried to reproduce in a VM running Fedora Rawhide but
> failed.
> 
> > This one is going to be hard to figure out..
> 
> And I'm afraid this may be a bug in my userspace...

Yeah I expect the same thing. In the log you sent me, there are some
strange-looking event masks that look like a userspace bug.

I tried those strange looking event masks, but epoll_ctl() correctly
rejects them with EINVAL. So those masks probably aren't the reason, but
they still suggest that userspace may also be broken somewhere else and
triggers hard-to-reproduce problems in kernel's epoll.

> Then I'd feel guilty if this is reverted because of an invalid bug report
> from I :(.

But even when userspace is broken, if the kernel patch makes userspace
non-functional, I would still call it a kernel bug. So it's a valid bug
report alright, don't worry about it.

Until we know exactly the root cause, I think revert is the right thing to
do.

Best regards,
Nam