Date: Wed, 22 Nov 2017 11:28:12 -0800 From: Gleb Smirnoff <glebius@FreeBSD.org> To: =?utf-8?Q?Micha=C5=82?= Krawczyk <mk@semihalf.com> Cc: Marcin Wojtas <mw@freebsd.org>, src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r325239 - head/sys/dev/ena Message-ID: <20171122192812.GI1063@FreeBSD.org> In-Reply-To: <CAJMMOfP6tpAKv2EBNiSkwNNKCPNr2s19Ek8QkVZsANYBQiTfCg@mail.gmail.com> References: <201710311631.v9VGVNOs058255@repo.freebsd.org> <20171122000816.GH1063@FreeBSD.org> <CAJMMOfP6tpAKv2EBNiSkwNNKCPNr2s19Ek8QkVZsANYBQiTfCg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
Michał, thanks for explanation. Looks like in this case software counters are indeed superior to hardware ones. On Wed, Nov 22, 2017 at 05:19:54PM +0100, Michał Krawczyk wrote: M> Hi Gleb, M> We are counting the statistics in the software because of the alignment M> with the newest ena-com API (HAL) which is delivered to us from NIC's M> vendor. M> M> Regarding fetching them periodically - there are 2 issues with that. First M> of all, there may be problem with the counter overflow. Secondly, reading M> statistics on demand requires sending admin message to the NIC and then the M> cv_timedwait function is called to wait for the response. However there are M> a lot of witness warnings when the cv_timedwait is called from the M> if_get_counter() context, because of the non sleepable lock which is hold M> there (I don't remember exactly which lock was causing this issue). M> M> Best regards, M> Michal M> M> 2017-11-22 1:08 GMT+01:00 Gleb Smirnoff <glebius@freebsd.org>: M> M> > Hi Martin and Michal, M> > M> > On Tue, Oct 31, 2017 at 04:31:23PM +0000, Marcin Wojtas wrote: M> > M> Author: mw M> > M> Date: Tue Oct 31 16:31:23 2017 M> > M> New Revision: 325239 M> > M> URL: https://svnweb.freebsd.org/changeset/base/325239 M> > M> M> > M> Log: M> > M> Rework counting of hardware statistics in ENA driver M> > M> M> > M> Do not read all statistics from the device, instead count them in the M> > M> driver except from RX drops - they are received directly from the NIC M> > M> in the AENQ descriptor. M> > M> M> > M> Submitted by: Michal Krawczyk <mk@semihalf.com> M> > M> Reviewed by: imp M> > M> Obtained from: Semihalf M> > M> Sponsored by: Amazon.com, Inc. M> > M> Differential Revision: https://reviews.freebsd.org/D12852 M> > M> > Is it possible not to count them in software, as well as not fetch M> > them from hardware periodically, but instead just fetch them on M> > demand, when either if_get_counter() or a sysctl is called? M> > M> > That would be more efficient. M> > M> > -- M> > Gleb Smirnoff M> > -- Gleb Smirnoff
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20171122192812.GI1063>