Re: [PATCH] mm/damon: simplify stop mechanism

[Changbin Du]

On Tue, Oct 26, 2021 at 06:42:03PM +0000, SeongJae Park wrote:
> Hello Changbin,
> 
> On Tue, 26 Oct 2021 23:30:33 +0800 Changbin Du <changbin.du@xxxxxxxxx> wrote:
> 
> > An kernel thread can exit gracefully with kthread_stop(). So we don't need a
> > new flag 'kdamond_stop'.  And to make sure the task struct is not freed when
> > accessing it, get task struct on start and put it on stop.
> 
> We previously considered using kthread_stop() here.  However, we resulted in
> current code because kdamond can be self-terminated when all target processes
> are invalid[1].
> 
> Seems this patch is also not fully prepared for the self-termination case.  I
> left some comments below.
> 
> [1] https://lore.kernel.org/linux-mm/20210624102623.24563-1-sjpark@xxxxxxxxx/
> 
> > 
> > And since the return value of 'before_terminate' callback is never used,
> > we make it have no return value.
> 
> This looks nice to me.  Could you please send this again as a separate patch?
>
Sure, I'll do it later.

> > 
> > Signed-off-by: Changbin Du <changbin.du@xxxxxxxxx>
> > ---
> >  include/linux/damon.h |  3 +--
> >  mm/damon/core.c       | 59 +++++++++++++------------------------------
> >  mm/damon/dbgfs.c      |  5 ++--
> >  3 files changed, 20 insertions(+), 47 deletions(-)
> > 
> > diff --git a/include/linux/damon.h b/include/linux/damon.h
> > index a14b3cc54cab..041966786270 100644
> > --- a/include/linux/damon.h
> > +++ b/include/linux/damon.h
> [...]
> > --- a/mm/damon/core.c
> > +++ b/mm/damon/core.c
> [...]  
> > @@ -1069,7 +1048,7 @@ static int kdamond_fn(void *data)
> >  					sz_limit);
> >  			if (ctx->callback.after_aggregation &&
> >  					ctx->callback.after_aggregation(ctx))
> > -				set_kdamond_stop(ctx);
> > +				done = true;
> >  			kdamond_apply_schemes(ctx);
> >  			kdamond_reset_aggregated(ctx);
> >  			kdamond_split_regions(ctx);
> > @@ -1088,16 +1067,12 @@ static int kdamond_fn(void *data)
> >  			damon_destroy_region(r, t);
> >  	}
> >  
> > -	if (ctx->callback.before_terminate &&
> > -			ctx->callback.before_terminate(ctx))
> > -		set_kdamond_stop(ctx);
> > +	if (ctx->callback.before_terminate)
> > +		ctx->callback.before_terminate(ctx);
> >  	if (ctx->primitive.cleanup)
> >  		ctx->primitive.cleanup(ctx);
> >  
> >  	pr_debug("kdamond (%d) finishes\n", current->pid);
> > -	mutex_lock(&ctx->kdamond_lock);
> > -	ctx->kdamond = NULL;
> > -	mutex_unlock(&ctx->kdamond_lock);
> 
> When kdamond is self-terminating, ctx->kdamond will not be nullfified.  As a
> result, this patch can introduce some errors like below:
> 
>     # cd /sys/kernel/debug/damon
>     # sleep 60 &
>     [1] 1926
>     # echo $(pidof sleep) > target_ids
>     # echo on > monitor_on
>     # cat monitor_on
>     on
>     # # after 60 seconds, sleep finishes and kdamond is self-terminated
>     # cat monitor_on
>     off
>     # echo 42 > target_ids
>     bash: echo: write error: Device or resource busy
> 
> If we simply restore the nullification here with the mutex locking, we would
> result in a deadlock because __damon_stop() calls kthread_stop() while holding
> ctx->kdamond_lock.
> 
> Also, the reference count of ctx->kdamond, which increased by __damon_start(),
> would not be decreased in the case.
>

If so, I suppose below change should work correctly (still set ctx->kdamond to
NULL at the end of kdamond).

 static int __damon_stop(struct damon_ctx *ctx)
 {
+       struct task_struct *tsk;
+
        mutex_lock(&ctx->kdamond_lock);
-       if (ctx->kdamond) {
-               ctx->kdamond_stop = true;
+       tsk = ctx->kdamond;
+       if (tsk) {
+               get_task_struct(tsk);
                mutex_unlock(&ctx->kdamond_lock);
-               while (damon_kdamond_running(ctx))
-                       usleep_range(ctx->sample_interval,
-                                       ctx->sample_interval * 2);
+               kthread_stop(tsk);
+               put_task_struct(tsk);
                return 0;
        }
        mutex_unlock(&ctx->kdamond_lock);


> If I'm missing something, please let me know.
> 
> 
> Thanks,
> SJ
> 
> [...]

-- 
Cheers,
Changbin Du