Re: [RFC v4] mm/damon: add synchronous commit for commit_inputs

[Liew Rui Yan <aethernet65535@gmail.com>]

> > Hi SeongJae,
> > 
> > Follow-up on my previous email regarding the locking concerns.
> > 
> > > I admit I don't have an elegant solution yet. Here are my current
> > > thought:
> > > 
> > > Add a timeout to wait_for_completion() (e.g.,
> > > wait_for_completion_timeout()), using 'damos_watermarks.interval' as the
> > > upper bound. This prevents indefinite blocking of 'param_lock', though
> > > it still holds the global lock for up to several seconds.
> > 
> > After further analysis, I realized that since both 'commit_inputs' and
> > 'enabled' module parameters are protected by the global 'param_lock',
> > stopping kdamond gracefully (via enabled=N) is serialized with
> > 'commit_inputs' writes. Therefore, a classic deadlock scenario should
> > not occur in normal operation.
> > 
> > However, I'm considering edge cases where kdamond might terminate
> > unexpectedly. In such cases, commit_inputs_store() could hold
> > 'param_lock' indefinitely, blocking other module parameter updates
> > system-wide.
> > 
> > To mitigate this risk defensively, would you accept adding a timeout to
> > wait_for_completion()? This ensures 'param_lock' is eventually released
> > even if kdamond fails unexpectedly.
> > 
> > Please let me know your preference. :>
> 
> As I mentioned on the reply to the original mail of this mail, I think such
> infinite wait will not happen because damon_call() is aware of the kdamond
> stop, and therefore the timeout is not needed.  Please double check and let me
> know if I'm missing something, though, as a reply to my reply.

Yes, you are absolutely right. My concern stemmed from a misconception
that kdamond could be forcibly terminated by userspace signals like
'kill -9'.

I just verified this experimentally in virtme-ng, even 'sudo kill -9
<kdamond_pid>' fails to terminate the kdamond thread. This confirms that
kdamond, asa kthread, must exit gracefully via kthread_stop(), ensuring
the completion will always be signaled.

Thank you for guiding me to verify this. I agree that no additional
timeout is needed.

Best regards,
Rui Yan