Re: [RFC PATCH] mm/damon: reset thread status parameters upon kdamond termination

[SeongJae Park <sj@kernel.org>]

On Thu,  2 Apr 2026 13:34:58 +0800 Liew Rui Yan <aethernet65535@gmail.com> wrote:

> Hi SeongJae,
> 
> On Wed,  1 Apr 2026 08:41:18 -0700 SeongJae Park <sj@kernel.org> wrote:
> 
> > [...]
> > What about checking if damon_call() and damon_commit_ctx() failures via therir
> > return value?  It seems damon_call() fails only when kdamond goes to the
> > termination path.  damon_commit_ctx() failure always causes the kdamond be
> > terminated.  So if I didn't miss something that could be a path forward.  What
> > do you think?
> 
> Thank you for the suggestion. I looked into the code and I think
> focusing on damon_commit_ctx() failures makes sense.
> 
> Specifically, damon_call() is only used in the damon_{lru_sort, reclaim}
> _enabled_store() path, which doesn't trigger unexpected termination. The
> main risk comes from 'commit_inputs', which calls damon_{lru_sort,
> reclaim}_apply_parameters() . When damon_commit_ctx() fails there, it
> reliably indicates that kdamond is terminating.
> 
> My plan is to add the error check directly in apply_parameters(), right
> after damon_commit_ctx():
> 
>     err = damon_commit_ctx(ctx, param_ctx);
>     if (err) {
>         enabled = false;
>         kdamond_pid = -1;
>     }
> 
> This covers the reproducible case (addr_unit=3 + commit_inputs=Y).
> 
> For truly unexpected termination (e.g., memory allocation failure inside
> kdamond), I'm considering a fallback mechanism in enabled_store():
> 
> - When the user writes 'N' to 'enabled', if damon_stop() fails but
>   kdamond is actually terminated, we reset enabled and kdamond_pid.

Sounds good.

> 
> - When the user writes 'Y' to 'enabled', if enabled is already 'Y' but
>   kdamond is terminated, we treat this as a restart request.

This sounds bit odd to me.  User shows enabled=Y and kdamond_pid!=-1.  Why they
would write 'Y' again?  Yes, they would notice kdamond is terminated in real,
using 'ps' like tool.  But even in the case, I'd imagine users would write N to
enabled first, and then write Y to enabled.  In other words, this could improve
the user experience, but seems not really necessary to me, at least as a
hotfix.  What do you think?

> 
> This way, even if kdamond terminates unexpectedly, the next user
> interaction will recover the state automatically.
> 
> Does this approach sound reasonable?

Yes, the plan for the next version sounds good to me.


Thanks,
SJ

[...]