| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <20170726175412.GB73972@C02RW35GFVH8.greyhouse.net>
Date: Wed, 26 Jul 2017 13:54:12 -0400
From: Andy Gospodarek <andy@...yhouse.net>
To: Jakub Kicinski <kubakici@...pl>
Cc: netdev@...r.kernel.org, Jiri Pirko <jiri@...lanox.com>,
Or Gerlitz <ogerlitz@...lanox.com>,
Michael Chan <michael.chan@...adcom.com>,
Sathya Perla <sathya.perla@...adcom.com>,
simon.horman@...ronome.com, davem@...emloft.net
Subject: Re: [RFC] switchdev: clarify ndo_get_phys_port_name() formats
On Tue, Jul 25, 2017 at 07:34:47PM -0700, Jakub Kicinski wrote:
> On Tue, 25 Jul 2017 21:48:15 -0400, Andy Gospodarek wrote:
> > On Tue, Jul 25, 2017 at 03:26:47PM -0700, Jakub Kicinski wrote:
> > > On Tue, 25 Jul 2017 11:22:41 -0400, Andy Gospodarek wrote:
> > > > On Mon, Jul 24, 2017 at 10:13:44PM -0700, Jakub Kicinski wrote:
> > > > > We are still in position where we can suggest uniform naming
> > > > > convention for ndo_get_phys_port_name(). switchdev.txt file
> > > > > already contained a suggestion of how to name external ports.
> > > > > Since the use of switchdev for SR-IOV NIC's eswitches is growing,
> > > > > establish a format for ports of those devices as well.
> > > > >
> > > > > Signed-off-by: Jakub Kicinski <jakub.kicinski@...ronome.com>
> > > >
> > > > This is a nice addition and I suspect there could be even more done to
> > > > update this file to cover the VF rep usage.
> > > >
> > > > > ---
> > > > > Documentation/networking/switchdev.txt | 14 +++++++++++---
> > > > > 1 file changed, 11 insertions(+), 3 deletions(-)
> > > > >
> > > > > diff --git a/Documentation/networking/switchdev.txt b/Documentation/networking/switchdev.txt
> > > > > index 3e7b946dea27..7c4b6025fb4b 100644
> > > > > --- a/Documentation/networking/switchdev.txt
> > > > > +++ b/Documentation/networking/switchdev.txt
> > > > > @@ -119,9 +119,17 @@ into 4 10G ports, resulting in 4 port netdevs, the device can give a unique
> > > > > SUBSYSTEM=="net", ACTION=="add", ATTR{phys_switch_id}=="<phys_switch_id>", \
> > > > > ATTR{phys_port_name}!="", NAME="swX$attr{phys_port_name}"
> > > > >
> > > > > -Suggested naming convention is "swXpYsZ", where X is the switch name or ID, Y
> > > > > -is the port name or ID, and Z is the sub-port name or ID. For example, sw1p1s0
> > > > > -would be sub-port 0 on port 1 on switch 1.
> > > > > +Suggested formats of the port name returned by ndo_get_phys_port_name are:
> > > > > + - pA for external ports;
> > > > > + - pAsB for split external ports;
> > > > > + - pfC for PF ports (so called PF representors);
> > > > > + - pfCvfD for VF ports (so called VF representors).
> > > >
> > > > I hate to clutter this up, but might be also need to add:
> > > >
> > > > - pfCsB for split PF ports (so called PF representors);
> > > > - pfCsBvfD for split VF ports (so called VF representors).
> > > >
> > > > or are we comfortable that these additions to the name for split ports
> > > > are implied?
> > >
> > > Hm.. What is a split PF port? Splits happen on the physical port - see
> > > my rant on the thread this is a reply to ;) PFs are PCIe functions,
> > > on the opposite side of the eswitch from the wires.
> >
> > I'm with you that I think there is value in separate netdevs to
> > represent "PFs, VFs and external ports/MACs" -- particularly for the
> > use-case you to create rules to control PF<->VF traffic.
> >
> > So while I'm not saying it is a _great_ idea to support such a thing as
> > port-splitting of PFs, I suggested this addition as I'm not willing to restrict
> > such a design/implementation if a vendor or customer desired. It seemed
> > useful to provde some guidance on how to name them -- even if we do not
> > like them. :-)
>
> If I understand you correctly split PF would be a situation where
> device has multiple port instances on the PCIe PF side? IOW switch sees
> multiple endpoints on the PF side? Let me attempt an ASCII diagram :)
>
>
> HOST A || HOST B
> ||
> PF A | V | V | V | V || PF B | V | V | V
> | F | F | F | F ... || | F | F | F ...
> port A0 | port A1 | 0 | 1 | 2 | 3 || port B0 | port B1 | 0 | 1 | 2
> ||
> PCI Express link || PCI Express link
> \ \ \ | | | | | / / /
> \ \ \ | | | | | / / /
> \______\______\' | | | '____/___/___/
> /---------------------------\
> |<<========== |
> | ==========>> |
> | SR-IOV e-switch |
> |<<========== |
> | ==========>> |
> \---------------------------/
> | | |
> | | |
> || ||
> MAC 0 || MAC 1 || MAC 2
> || ||
>
>
> Seems to be a valid configuration, perhaps this would actually be of
> some use in container workloads, especially if the ports could be
> instantiated at runtime in high numbers. I would be cautious though
> with calling the instances splits. The more different PFs look from
> MACs the better IMHO. Do you actually have that problem today?
>
> Is there any HW supported upstream which would benefit from this? Could we
> decide on naming when we have an example implementation? In theory
> nothing stops us from splitting VFs the same way.
Not that I know about right now.
> Another note on PF netdevs, perhaps the most awkward thing about them,
> is that they result in two netdevs being visible to the host. This is
> not incorrect, since VFs if unassigned to VMs will end up creating an
> "actual" netdev and the switchdev port representor too, but it rubs
> some people the wrong way. Which in turn makes those people try to not
> spawn separate netdevs, which is incorrect IMHO, and breaks down e.g.
> when the real netdev gets assigned to a namespace.
For me, the most awkward part of having a separate netdev for the PF and
the MAC is that is really not how things were thought about in the
nominal switching case (the non e-switch case).
Since idea behind switchdev when it was created was to make sure that
each front-panel port on a switch was represented by a netdev in the
kernel (and the 'CPU interface' was abstracted away by the driver) I was
always a bit uneasy about having a separate netdev allocated for the CPU
port when in the switching case it really wasn't necessary.
> I'm not sure if this clarifies my thinking, I have, however, seem to
> have drawn a moose :)
Which looks great, BTW. The moose may turn out to be one of the major
benefits from this thread!
Powered by blists - more mailing lists