Ah. we certainly misunderstand each other.

On 08/29, Jiri Olsa wrote:

On Thu, Aug 29, 2024 at 05:20:33PM +0200, Oleg Nesterov wrote:

SNIP

SNIP

right.. if the event is not added by perf_trace_add on this cpu
it won't go pass this point, so no problem for perf

Yes, and this is what I tried to verify. In your previous email you said

	and I think the same will happen for perf record in this case where instead of
	running the program we will execute perf_tp_event

and I tried verify this can't happen. So no problem for perf ;)

but the issue is with bpf program triggered earlier by return uprobe

Well, the issue with bpf program (with the bpf_prog_array_valid(call) code
in __uprobe_perf_func) was clear from the very beginning, no questions.

and [1] patch seems to fix that

I'd say this patch fixes the symptoms, and it doesn't fix all the problems.
But I can't suggest anything better for bpf code, so I won't really argue.
However the changelog and even the subject is wrong.

I sent out the bpf selftest that triggers the issue [2]

Thanks, I'll try take a look tomorrow.

Oleg.