Re: [PATCH v3 6/7] perf: Rename perf_event_exit_task(.child)

[Peter Zijlstra]

On Mon, Mar 10, 2025 at 08:50:55PM +0530, Ravi Bangoria wrote:
> On 10-Mar-25 8:17 PM, Peter Zijlstra wrote:
> > On Mon, Mar 10, 2025 at 04:38:36PM +0530, Ravi Bangoria wrote:
> >> Hi Peter,
> >>
> >> On 08-Mar-25 1:03 AM, Peter Zijlstra wrote:
> >>> The task passed to perf_event_exit_task() is not a child, it is
> >>> current. Fix this confusing naming, since much of the rest of the code
> >>> also relies on it being current.
> >>>
> >>> Specifically, both exec() and exit() callers use it with current as
> >>> the argument.
> >>
> >> ...
> >>
> >>> -static void perf_event_exit_task_context(struct task_struct *child, bool exit)
> >>> +static void perf_event_exit_task_context(struct task_struct *task, bool exit)
> >>>  {
> >>> -	struct perf_event_context *child_ctx, *clone_ctx = NULL;
> >>> +	struct perf_event_context *ctx, *clone_ctx = NULL;
> >>>  	struct perf_event *child_event, *next;
> >>>  
> >>> -	WARN_ON_ONCE(child != current);
> >>> +	WARN_ON_ONCE(task != current);
> >>
> >> exec() codepath (i.e. copy_process()) passes child pointer, not 'current'.
> > 
> > I am confused, this not a new warning.
> 
> Right, However the WARN was present only in perf_event_exit_task_context()
> before merging it with perf_event_free_task() (patch #5). And
> perf_event_free_task() is getting called for child task.

Argh, yes.

> > Also, copy_process() is clone(), exec() is another code path.
> 
> My bad. I meant clone() code path:
> 
>   copy_process()
>     p = dup_task_struct(current);
>     perf_event_init_task(p);
>       perf_event_free_task(p);
>         perf_event_exit_task_context(p);
>           WARN_ON_ONCE(task != current);
> 
> Another one:
> 
>   copy_process()
>     p = dup_task_struct(current);
>     ...
>     bad_fork_cleanup_perf:
>       perf_event_free_task(p);
>         perf_event_exit_task_context(p);
>           WARN_ON_ONCE(task != current);
> 
> Or am I missing something?

No, the perf_event_free_task() callchain has a problem.

I'll remove that WARN_ON_ONCE() since perf_event_exit_task() has the
same check. I'll do that in the merge patch, not this rename patch.