Re: [PATCH] bpf, sockmap: fix deadlock in rcu_report_exp_cpu_mult

[John Fastabend]

Jakub Sitnicki wrote:
> On Mon, Mar 25, 2024 at 01:23 PM +01, Jakub Sitnicki wrote:
> 
> [...]
> 
> > But we also need to cover sock_map_unref->sock_sock_map_del_link called
> > from sock_hash_delete_elem. It also grabs a spin lock.
> 
> On second look, no need to disable interrupts in
> sock_map_unref->sock_sock_map_del_link. Call is enclosed in the critical
> section in sock_hash_delete_elem that has been updated.
> 
> I have a question, though, why are we patching sock_hash_free? It
> doesn't get called unless there are no more existing users of the BPF
> map. So nothing can mutate it from interrupt context.
> 
> [...]

Agree sock_hash_free should be only after all refs are dropped.

Edward, did you want to send a v2 for this? Also if you want fixing the
sockmap case as well would be useful. Also happy to finish up the patches
if you would rather not.

Thanks,
John