On Thu, Feb 16, 2017 at 5:49 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: > On Thu, 2017-02-16 at 14:44 +0200, Saeed Mahameed wrote: >> On Wed, Feb 15, 2017 at 4:04 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: >> > On Wed, 2017-02-15 at 05:29 -0800, Eric Dumazet wrote: >> > >> >> >> >> mlx4_eq_int() is a hard irq handler. >> >> >> >> How a tasklet could run in the middle of it ? >> >> >> >> A tasklet is a softirq handler. >> > >> > Speaking of mlx4_eq_int() , 50% of cycles are spent on mb() (mfence) >> > in eq_set_ci() >> > >> >> I wonder why you have so many interrupts ? don't you have some kind of >> interrupt moderation ? >> what test are you running that got your CPU so busy. > > > Simply 8 RX queues. About 140,000 irq per second per RX queue, > for a moderate network load on 40Gbit NIC. >
so i guess you are not busy polling .. and adaptive moderation decided to lower down rx-usecs for you, and you are looking to improve latency. > Interrupt moderation is a latency killer, we want our usec back. > well, for RX we have adaptive moderation. and for TX we have xmit more. so interrupt moderation can be good if it is done right. do i understand from this that you are against interrupt moderation ? >> >> > I wonder why this very expensive mb() is required, right before exiting >> > the interrupt handler. >> >> to make sure the HW knows we handled Completions up to (ci) consumer >> index. so it will generate next irq. > > > So why a mb() is needed exactly ? > > wmb() seems enough. > Yes seems right, not sure why it is mb() though, i will have to check and get back to you on this.
