> On Nov 14, 2016, at 3:09 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: > > On Mon, 2016-11-14 at 14:46 -0800, Eric Dumazet wrote: >> On Mon, 2016-11-14 at 16:12 -0600, Eric W. Biederman wrote: >> >>> synchronize_rcu_expidited is not enough if you have multiple network >>> devices in play. >>> >>> Looking at the code it comes down to this commit, and it appears there >>> is a promise add rcu grace period combining by Eric Dumazet. >>> >>> Eric since people are hitting noticable stalls because of the rcu grace >>> period taking a long time do you think you could look at this code path >>> a bit more? >>> >>> commit 93d05d4a320cb16712bb3d57a9658f395d8cecb9 >>> Author: Eric Dumazet <eduma...@google.com> >>> Date: Wed Nov 18 06:31:03 2015 -0800 >> >> Absolutely, I will take a loop asap. > > The worst offender should be fixed by the following patch. > > busy poll needs to poll the physical device, not a virtual one... > > diff --git a/include/net/gro_cells.h b/include/net/gro_cells.h > index > d15214d673b2e8e08fd6437b572278fb1359f10d..2a1abbf8da74368cd01adc40cef6c0644e059ef2 > 100644 > --- a/include/net/gro_cells.h > +++ b/include/net/gro_cells.h > @@ -68,6 +68,9 @@ static inline int gro_cells_init(struct gro_cells *gcells, > struct net_device *de > struct gro_cell *cell = per_cpu_ptr(gcells->cells, i); > > __skb_queue_head_init(&cell->napi_skbs); > + > + set_bit(NAPI_STATE_NO_BUSY_POLL, &cell->napi.state); > + > netif_napi_add(dev, &cell->napi, gro_cell_poll, 64); > napi_enable(&cell->napi); > } >
This solved a ~20 second slowdown between OVS datapath unit tests for me. Jarno
