[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <504C9EFCA2D0054393414C9CB605C37F20CBC47B@SJEXCHMB06.corp.ad.broadcom.com>
Date: Tue, 18 Jun 2013 12:11:46 +0000
From: "Dmitry Kravkov" <dmitry@...adcom.com>
To: "Eric Dumazet" <eric.dumazet@...il.com>
cc: "davem@...emloft.net" <davem@...emloft.net>,
"netdev@...r.kernel.org" <netdev@...r.kernel.org>,
"Eilon Greenstein" <eilong@...adcom.com>
Subject: RE: [PATCH net-next 1/2] bnx2x: add support for ndo_ll_poll
> -----Original Message-----
> From: Eric Dumazet [mailto:eric.dumazet@...il.com]
> Sent: Tuesday, June 18, 2013 2:16 PM
> To: Dmitry Kravkov
> Cc: davem@...emloft.net; netdev@...r.kernel.org; Eilon Greenstein
> Subject: RE: [PATCH net-next 1/2] bnx2x: add support for ndo_ll_poll
>
> On Tue, 2013-06-18 at 10:45 +0000, Dmitry Kravkov wrote:
>
> > > Ideally, we could keep a counter of enabled LLS sockets, to
> > > automatically switch off/on GRO, but in net/core, not in every driver.
> >
> > Since we have GRO in FW it will be hard to disable/enable it on the fly, probably better to return LL_FLUSH_FAILED when GRO is
> enabled?
>
> Yes, assuming GRO in FW means latencies ?
>
> (If we receive a frame, FW has to wait a bit for the following frames to
> eventually aggregate them)
That's true !
But problem is that in disabling the feature on the fly will be very SlowPath operation, which can't be done from poll callback...
>
> I thought FW did LRO part on bnx2x, and GRO was the software thing (in
> networking core, not in the bnx2x drivre)
LRO has priority over GRO, when LRO is disabled bnx2x FW provides fully GRO-complaint aggregations,
which are ready for forwarding without host cpu intervention...
>
>
>
Powered by blists - more mailing lists