On Mon, Jul 15, 2024 at 3:47 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Mon, Jul 15, 2024 at 12:23:27PM +0200, Radoslaw Zielonek wrote: > > The rcu_dereference() is using rcu_read_lock_held() as a checker, but > > BPF in bpf_prog_test_run_syscall() is using rcu_read_lock_trace() locker. > > To fix this issue the proper checker has been used > > (rcu_read_lock_trace_held() || rcu_read_lock_held()) > > How does that fix it? release_callchain_buffers() does call_rcu(), not > call_rcu_tracing(). > > Does a normal RCU grace period fully imply an RCU-tracing grace period? I don't think so, they are completely independent. So this change doesn't seem correct. I think we should just ensure rcu_read_lock()/rcu_read_unlock() before calling into perf_callchain functionality. Which is what I'm doing in [0]. Radoslaw, can you please help validating if those changes are enough to fix this issue or we need to do some more? [0] https://patchwork.kernel.org/project/netdevbpf/patch/20240709204245.3847811-10-andrii@xxxxxxxxxx/ > > > --- > > kernel/events/callchain.c | 11 +++++++++-- > > 1 file changed, 9 insertions(+), 2 deletions(-) > > [...]
