[BUG] libxenstore and xentop confused by renamed VIF interfaces

[BUG] libxenstore and xentop confused by renamed VIF interfaces

[Jacek Konieczny]

Hi,

I have noticed a strange behaviour on 'xl top' – it would show no
network statistics for some of my domU. It turned out that 'xl top'
shows no network ('NETS', 'NETTX(k)', 'NETRX(k)' all '0') when the
domU VIF has been renamed by using the 'vifname' option in its xl.cfg
file.

The 'vifname' option is very useful to have predictable interface names
for firewall configuration or network monitoring. The problem is
libxenstore under Linux parses /proc/net/dev to extract 'vif.${domid}.*'
interfaces, which won't be there, when the interfaces have been renamed.

I think the solution would be to look-up the actual interface name or
index in sysfs, e.g. under /sys/devices/vif-${domind}-*.

Greets,
	Jacek

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Jacek Konieczny]

On 05/21/14 09:52, Jacek Konieczny wrote:
> The 'vifname' option is very useful to have predictable interface names
> for firewall configuration or network monitoring. The problem is
> libxenstore under Linux parses /proc/net/dev to extract 'vif.${domid}.*'

I mean _libxenstat_, of course.

Xen 4.4.0.

Greets,
	Jacek

Re: [BUG] libxenstore and xentop confused by renamed VIF interfaces

[Ian Campbell]

On Wed, 2014-05-21 at 09:52 +0200, Jacek Konieczny wrote:
> Hi,
> 
> I have noticed a strange behaviour on 'xl top' – it would show no
> network statistics for some of my domU. It turned out that 'xl top'
> shows no network ('NETS', 'NETTX(k)', 'NETRX(k)' all '0') when the
> domU VIF has been renamed by using the 'vifname' option in its xl.cfg
> file.
> 
> The 'vifname' option is very useful to have predictable interface names
> for firewall configuration or network monitoring. The problem is
> libxenstore under Linux parses /proc/net/dev to extract 'vif.${domid}.*'
> interfaces, which won't be there, when the interfaces have been renamed.

I suppose it is trying to exclude non-vif network devices (xenbr0, eth0
etc).

> I think the solution would be to look-up the actual interface name or
> index in sysfs, e.g. under /sys/devices/vif-${domind}-*.

Sounds plausible.

Or perhaps check that /sys/class/net/${dev}/device/devtype contains
"vif"? Might want to handle tap devices too. Perhaps parse the content
of /sys/class/net/vif1.0/device/nodename to get the domid and devid?

Ian.


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Jacek Konieczny]

On 05/21/14 12:29, Ian Campbell wrote:
> On Wed, 2014-05-21 at 09:52 +0200, Jacek Konieczny wrote:
> libxenstore under Linux parses /proc/net/dev to extract 'vif.${domid}.*'
> interfaces, which won't be there, when the interfaces have been renamed.
> I suppose it is trying to exclude non-vif network devices (xenbr0, eth0
> etc).

That is one reason. The more important reason is to match interface with
a domid. Also, the /proc/net/dev file is the source of the interface
usage statistics collected by libxenstat.

>> I think the solution would be to look-up the actual interface name or
>> index in sysfs, e.g. under /sys/devices/vif-${domind}-*.
> 
> Sounds plausible.
> 
> Or perhaps check that /sys/class/net/${dev}/device/devtype contains
> "vif"? Might want to handle tap devices too. Perhaps parse the content
> of /sys/class/net/vif1.0/device/nodename to get the domid and devid?

Yes, that could work and would probably require less changes in the
code – this can be done while iterating /proc/net/dev.

Another approach would be to use the data from the xenstore, but I guess
that would be more expensive.

Greets,
	Jacek

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Ian Campbell]

On Wed, 2014-05-21 at 12:48 +0200, Jacek Konieczny wrote:
> On 05/21/14 12:29, Ian Campbell wrote:
> > On Wed, 2014-05-21 at 09:52 +0200, Jacek Konieczny wrote:
> > libxenstore under Linux parses /proc/net/dev to extract 'vif.${domid}.*'
> > interfaces, which won't be there, when the interfaces have been renamed.
> > I suppose it is trying to exclude non-vif network devices (xenbr0, eth0
> > etc).
> 
> That is one reason. The more important reason is to match interface with
> a domid.

Yes, I figured that out later ;-)

>  Also, the /proc/net/dev file is the source of the interface
> usage statistics collected by libxenstat.
> 
> >> I think the solution would be to look-up the actual interface name or
> >> index in sysfs, e.g. under /sys/devices/vif-${domind}-*.
> > 
> > Sounds plausible.
> > 
> > Or perhaps check that /sys/class/net/${dev}/device/devtype contains
> > "vif"? Might want to handle tap devices too. Perhaps parse the content
> > of /sys/class/net/vif1.0/device/nodename to get the domid and devid?
> 
> Yes, that could work and would probably require less changes in the
> code – this can be done while iterating /proc/net/dev.
> 
> Another approach would be to use the data from the xenstore, but I guess
> that would be more expensive.

Yes, and I don't think all the stats are in xenstore anyway.

Ian.



_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Jacek Konieczny]

On 05/21/14 12:29, Ian Campbell wrote:
> On Wed, 2014-05-21 at 09:52 +0200, Jacek Konieczny wrote:
>> I think the solution would be to look-up the actual interface name or
>> index in sysfs, e.g. under /sys/devices/vif-${domind}-*.
> 
> Sounds plausible.
> 
> Or perhaps check that /sys/class/net/${dev}/device/devtype contains
> "vif"? Might want to handle tap devices too.

I am not sure we could easily map a renamed tap (can they be renamed?)
device to a domid.

> Perhaps parse the content of /sys/class/net/vif1.0/device/nodename to
> get the domid and devid?

I think I could prepare the patch.

One question: can I assume /sys is always mounted and the
'/sys/.../device/nodename', contains 'backend/vif/$domid/*' or should
there be a fall-back to the old behaviour?

BTW, when 'device/nodename' exists and contains 'backend/vif/*' we don't
need to check 'device/devtype' – one file access less.

Jacek

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Ian Campbell]

On Wed, 2014-05-21 at 13:35 +0200, Jacek Konieczny wrote:
> On 05/21/14 12:29, Ian Campbell wrote:
> > On Wed, 2014-05-21 at 09:52 +0200, Jacek Konieczny wrote:
> >> I think the solution would be to look-up the actual interface name or
> >> index in sysfs, e.g. under /sys/devices/vif-${domind}-*.
> > 
> > Sounds plausible.
> > 
> > Or perhaps check that /sys/class/net/${dev}/device/devtype contains
> > "vif"? Might want to handle tap devices too.
> 
> I am not sure we could easily map a renamed tap (can they be renamed?)

I think so, to ${vifname}-emu IIRC.

> device to a domid.

True.

> > Perhaps parse the content of /sys/class/net/vif1.0/device/nodename to
> > get the domid and devid?
> 
> I think I could prepare the patch.

Thank you!

> One question: can I assume /sys is always mounted and the
> '/sys/.../device/nodename', contains 'backend/vif/$domid/*' or should
> there be a fall-back to the old behaviour?

I think on Linux we can assume these days that /sys must exist. I'm less
sure about the content of nodename, in particular WRT the behaviour of
the classic-Xen forward ports. (Jan, CCd, might know).

I'd be inclined to keep the existing code as a fallback for now, I
think.

> BTW, when 'device/nodename' exists and contains 'backend/vif/*' we don't
> need to check 'device/devtype' – one file access less.

Ack.

Ian.


_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel

Re: [BUG] libxenstat and xentop confused by renamed VIF interfaces

[Jan Beulich]

>>> On 21.05.14 at 13:52, <Ian.Campbell@citrix.com> wrote:
> On Wed, 2014-05-21 at 13:35 +0200, Jacek Konieczny wrote:
>> One question: can I assume /sys is always mounted and the
>> '/sys/.../device/nodename', contains 'backend/vif/$domid/*' or should
>> there be a fall-back to the old behaviour?
> 
> I think on Linux we can assume these days that /sys must exist. I'm less
> sure about the content of nodename, in particular WRT the behaviour of
> the classic-Xen forward ports. (Jan, CCd, might know).
> 
> I'd be inclined to keep the existing code as a fallback for now, I
> think.

I agree - an eventual change in sysfs layout would need hunting
down between 2.6.18 and 3.15; not something I would do without
a very good reason. And I can't tell without doing so...

Jan