On Mon, Sep 24, 2018 at 1:00 PM Song Liu <songliubrav...@fb.com> wrote: > > > > > On Sep 24, 2018, at 12:41 PM, Eric Dumazet <eduma...@google.com> wrote: > > > > On Mon, Sep 24, 2018 at 12:31 PM Song Liu <songliubrav...@fb.com> wrote: > > > >> This triggers dev_watchdog() on a simple netperf TCP_RR on bnxt (I haven't > >> tested other drivers yet). > >> > >> I guess this is because NAPI_STATE_SCHED is set when poll_one_napi() calls > >> napi->poll(). And then cleared by napi->poll(). > > > > Which part of napi->poll() could possibly clear NAPI_STATE_SCHED when > > called by netpoll ? > > > > AFAIK, napi_complete_done() should exit early (before having a chance > > to clear NAPI_STATE_SCHED) > > because of : > > > > if (unlikely(n->state & (NAPIF_STATE_NPSVC | NAPIF_STATE_IN_BUSY_POLL))) > > return false; > > > > Thanks ! > > You are right on this condition. But this does trigger dev_watchdog() for > some reason.
Interesting, maybe a bnxt specific issue. It seems their model is to process TX/RX notification in the same queue, they throw away RX events if budget == 0 It means commit e7b9569102995ebc26821789628eef45bd9840d8 is wrong and must be reverted. Otherwise, we have a possibility of blocking a queue under netpoll pressure.
