Re: [PATCH 05/10] perf/uprobe: SRCU-ify uprobe->consumer list

[Peter Zijlstra]

On Tue, Jul 09, 2024 at 05:05:04PM +0200, Oleg Nesterov wrote:


> To simplify, suppose we have a single consumer which is not interested
> in this task/mm, it returns UPROBE_HANDLER_REMOVE.
> 
> For example, event->hw.target != NULL and the current task is the forked
> child which hits the breakpoint copied by dup_mmap().
> 
> Now. We need to ensure that another (say system-wide) consumer can't come
> and call register_for_each_vma() before unapply_uprobe().
> 
> But perhaps I missed your point...

Ooh, I see. I failed to consider that particular case. This is
interleaving uprobe_register() and handler_chain(). Silly me only looked
at uprobe_unregister() and handler_chain().

Hmm, easiest would be to add a seqcount to register_mutex and simply
skip the remove case when odd.

Then the handler might get a few extra (unwanted) calls, but it should
be able to handle them, they're fundamentally not different from the
first one where it says REMOVE. Right?