On Thu, Apr 09, 2020 at 12:26:37PM +0900, Daniel T. Lee wrote:
>
> So, I think this 'unable to attach bpf program on BPF map function (read,write)'
> is a bug. Or is it desired action?
desired action.
> If it is a bug, bpf_{enable|disable}_instrumentation() should only
> cover stackmap
> as the upper commit intended. Not sure but adding another flag for
> lock might work?
>
> Or if this is an desired action, this should be covered at
> documentation with a limitation
> and tracex6 sample has to be removed.
Right. That test has to be fixed.
These two old commits:
commit 020a32d9581a ("bpf: add a test case for helper bpf_perf_event_read_value")
commit 41e9a8046c92 ("samples/bpf: add tests for more perf event types")
attached kprobe bpf to inner map accessors get_next and lookup.
That's safe only for a subset of map types and not allowed in general.
It works for hash and array because they don't take locks in these operations.
