On Thu, 17 Oct 2019 18:04:22 +0100, Robert Beckett wrote: > On Thu, 2019-10-17 at 08:44 -0700, Jakub Kicinski wrote: > > On Thu, 17 Oct 2019 12:24:03 +0100, Robert Beckett wrote: > > > On Wed, 2019-10-16 at 16:55 -0700, Jakub Kicinski wrote: > > > > On Wed, 16 Oct 2019 16:47:06 -0700, Jeff Kirsher wrote: > > > > > From: Robert Beckett <bob.beck...@collabora.com> > > > > > > > > > > To allow userland to enable or disable dropping packets when > > > > > descriptor > > > > > ring is exhausted, add RX_DROP_EN private flag. > > > > > > > > > > This can be used in conjunction with flow control to mitigate > > > > > packet storms > > > > > (e.g. due to network loop or DoS) by forcing the network > > > > > adapter to > > > > > send > > > > > pause frames whenever the ring is close to exhaustion. > > > > > > > > > > By default this will maintain previous behaviour of enabling > > > > > dropping of > > > > > packets during ring buffer exhaustion. > > > > > Some use cases prefer to not drop packets upon exhaustion, but > > > > > instead > > > > > use flow control to limit ingress rates and ensure no dropped > > > > > packets. > > > > > This is useful when the host CPU cannot keep up with packet > > > > > delivery, > > > > > but data delivery is more important than throughput via > > > > > multiple > > > > > queues. > > > > > > > > > > Userland can set this flag to 0 via ethtool to disable packet > > > > > dropping. > > > > > > > > > > Signed-off-by: Robert Beckett <bob.beck...@collabora.com> > > > > > Tested-by: Aaron Brown <aaron.f.br...@intel.com> > > > > > Signed-off-by: Jeff Kirsher <jeffrey.t.kirs...@intel.com> > > > > > > > > How is this different than enabling/disabling flow control.. > > > > > > > > ethtool -a/-A > > > > > > Enabling flow control enables the advertisement of flow control > > > capabilites and allows negotiation with link partner. > > > > More or less. If autoneg is on it controls advertised bits, > > if autoneg is off it controls the enabled/disable directly. > > > > > It does not dictate under which circumstances those pause frames > > > will > > > be emitted. > > > > So you're saying even with pause frames on igb by default will not > > backpressure all the way to the wire if host RX ring is full/fill > > ring > > is empty? > > Correct. > Honestly I personally considered it a bug when I first saw it. > > see e6bdb6fefc590 > > Specifically it enables dropping of frames if multiple queues are in > use, ostensibly to prevent head of line blocking between the different > rx queues. > > This patch says that that should be a user choice, defaulting to the > old behaviour.
I'd say just always enable it then. Honestly if it's statically enabled with multi queue (which I presume is the default?) then there's no need for us to ponder the uAPI questions. I can't think of anyone flipping that flag to disabled... Flow control is supposed to prevent _all_ drops it can, AFAIU. If we get an actual user request to change it we can revisit.
