Date: Tue, 16 Aug 2016 19:22:53 -0700 From: Gleb Smirnoff <glebius@FreeBSD.org> To: Mark Johnston <markj@FreeBSD.org> Cc: src-committers@freebsd.org, svn-src-all@freebsd.org, svn-src-head@freebsd.org Subject: Re: svn commit: r303646 - head/sys/ofed/drivers/infiniband/ulp/ipoib Message-ID: <20160817022253.GD1069@FreeBSD.org> In-Reply-To: <20160817004128.GB94062@wkstn-mjohnston.west.isilon.com> References: <201608012222.u71MMB4E018482@repo.freebsd.org> <20160817000953.GB1069@FreeBSD.org> <20160817004128.GB94062@wkstn-mjohnston.west.isilon.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Tue, Aug 16, 2016 at 05:41:28PM -0700, Mark Johnston wrote:
M> > M> Log:
M> > M> ipoib: Bound the number of egress mbufs buffered during pathrec lookups.
M> > M>
M> > M> In pathological situations where the master subnet manager becomes
M> > M> unresponsive for an extended period, we may otherwise end up queuing all
M> > M> of the system's mbufs while waiting for a response to a path record lookup.
M> > M>
M> > M> This addresses the same issue as commit 1e85b806f9 in Linux.
M> > M>
M> > M> Reviewed by: cem, ngie
M> > M> MFC after: 2 weeks
M> > M> Sponsored by: EMC / Isilon Storage Division
M> > M>
M> > M> Modified:
M> > M> head/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c
M> > M>
M> > M> Modified: head/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c
M> > M> ==============================================================================
M> > M> --- head/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c Mon Aug 1 22:19:23 2016 (r303645)
M> > M> +++ head/sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_main.c Mon Aug 1 22:22:11 2016 (r303646)
M> > M> @@ -660,7 +660,13 @@ ipoib_unicast_send(struct mbuf *mb, stru
M> > M> new_path = 1;
M> > M> }
M> > M> if (path) {
M> > M> - _IF_ENQUEUE(&path->queue, mb);
M> > M> + if (_IF_QLEN(&path->queue) < IPOIB_MAX_PATH_REC_QUEUE)
M> > M> + _IF_ENQUEUE(&path->queue, mb);
M> > M> + else {
M> > M> + if_inc_counter(priv->dev, IFCOUNTER_OERRORS, 1);
M> > M> + m_freem(mb);
M> > M> + }
M> >
M> > Shouldn't that be IFCOUNTER_ODROPS?
M>
M> I'm not sure. I used IFCOUNTER_OERRORS to be consistent with other error
M> cases in this function. This error case doesn't represent the normal
M> source of outbound packet drops but rather indicates that a key routing
M> agent on the network is not responding. OQDROPS seems like it's
M> specifically for the case that we can't buffer packets because the
M> transmitter isn't keeping up.
I see. The fact that this is an if_queue has confused me. I expected that
it is an ifqueue belonging to an ifnet.
btw, may be this ifqueue can be reduced to mbufq.
--
Totus tuus, Glebius.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20160817022253.GD1069>