On 2021-11-16 18:59:56 [-0800], Davidlohr Bueso wrote: > Updating the per-CPU lport->stats relies on disabling preemption > with get/put_cpu() semantics. However, this is a bit harsh for > PREEMPT_RT and by using a local_lock, it can allow the region > to be preemptible, guaranteeing CPU locality, without making any > difference to the non-RT common case as it will continue to > disable preemption. What about adding a struct u64_stats_sync where the stats are updated. You have to use u64_stats_update_begin() at the beginning and the matching end function instead the get_cpu()/ local_lock(). You need just ensure that there is only one writer at a time. Network wise it is easy as there are often per-queue stats so one writer at a time :) Looking closer, the current approach appears broken in that regard: stats members, such as TxFrames/ TxWords which are incremented in bnx2fc_xmit(), are 64bit. Reading/ writing them requires two operations on 32bit architectures. But we probably don't care because stats and it happens hardly and performance of course. This is the cute part about u64_stats_sync, it magically vanishes on 64bit architectures. Sebastian
