On Wed, Feb 5, 2025 at 4:58 AM Ritesh Oedayrajsingh Varma <ritesh@xxxxxxxxxxxxxxx> wrote: > > Given this, while it's not possible to remove the wait entirely > without breaking user space, I was wondering if it would be > possible/acceptable to add a way to opt-out of this behavior for > programs like ours that don't care about this. One way to do so could > be to add an additional flag to the BPF_MAP_CREATE flags, perhaps > something like BPF_F_INNER_MAP_NO_SYNC. Sounds reasonable. The flag name is a bit cryptic, BPF_F_UPDATE_INNER_MAP_NO_WAIT is a bit more explicit, but I'm not sure whether it's better. Also have you considered a batched update? There will be only one sync_rcu() for the whole batch.
