On Tue, Apr 05, 2022 at 02:04:34AM +0200, KP Singh wrote:
> > >>> Either way, how frequently is call_rcu_tasks_trace() being invoked in
> > >>> your setup? If it is being invoked frequently, increasing delays would
> > >>> allow multiple call_rcu_tasks_trace() instances to be served by a single
> > >>> tasklist scan.
> > >>>
> > >>>> Given that, I do not think bpf_sk_storage_free() can/should use
> > >>>> call_rcu_tasks_trace(),
> > >>>> we probably will have to fix this soon (or revert from our kernels)
> > >>>
> > >>> Well, you are in luck!!! This commit added call_rcu_tasks_trace() to
> > >>> bpf_selem_unlink_storage_nolock(), which is invoked in a loop by
> > >>> bpf_sk_storage_free():
> > >>>
> > >>> 0fe4b381a59e ("bpf: Allow bpf_local_storage to be used by sleepable programs")
> > >>>
> > >>> This commit was authored by KP Singh, who I am adding on CC. Or I would
> > >>> have, except that you beat me to it. Good show!!! ;-)
>
> Hello :)
>
> Martin, if this ends up being an issue we might have to go with the
> initial proposed approach
> of marking local storage maps explicitly as sleepable so that not all
> maps are forced to be
> synchronized via trace RCU.
>
> We can make the verifier reject loading programs that try to use
> non-sleepable local storage
> maps in sleepable programs.
>
> Do you think this is a feasible approach we can take or do you have
> other suggestions?
bpf_sk_storage_free() does not need to use call_rcu_tasks_trace().
The same should go for the bpf_{task,inode}_storage_free().
The sk at this point is being destroyed. No bpf prog (sleepable or not)
can have a hold on this sk. The only storage reader left is from
bpf_local_storage_map_free() which is under rcu_read_lock(),
so a 'kfree_rcu(selem, rcu)' is enough.
A few lines below in bpf_sk_storage_free(), 'kfree_rcu(sk_storage, rcu)'
is currently used instead of call_rcu_tasks_trace() for the same reason.
KP, if the above makes sense, can you make a patch for it?
The bpf_local_storage_map_free() code path also does not need
call_rcu_tasks_trace(), so may as well change it together.
The bpf_*_storage_delete() helper and the map_{delete,update}_elem()
syscall still require the call_rcu_tasks_trace().
