Re: [RFC][PATCH] taskstats: Fix CGROUPSTATS_TYPE_CGROUP_STATS having same value as TASKSTATS_TYPE_PID

[Nikanth Karthikesan <knikanth@suse.de>]

On Monday 13 July 2009 21:23:10 Balbir Singh wrote:
> * Nikanth Karthikesan <knikanth@suse.de> [2009-07-13 21:07:37]:
> > On Monday 13 July 2009 19:11:58 Balbir Singh wrote:
> > > * Nikanth Karthikesan <knikanth@suse.de> [2009-07-13 18:31:12]:
> > > > Hi
> > > >
> > > > Currently we never get message from kernel to userspace of type
> > > > TASKSTATS_TYPE_PID. Otherwise this could have been spotted earlier.
> > > >
> > > > I was trying to add a new taskstat command that would return response
> > > > of type TASKSTATS_TYPE_PID.
> > > >
> > > > Having the same values would restrict one not to use the same netlink
> > > > socket for a command that would return response of type
> > > > TASKSTATS_TYPE_PID and the CGROUPSTATS_CMD_GET command.
> > > >
> > > > Should we fix it by using values after __TASKSTATS_TYPE_MAX.
> > > >
> > > > Changing this now might break pre-built binaries. Or is this
> > > > intended, and the application is not supposed to use CGROUPSTATS and
> > > > TASKSTATS on the same socket?
> > >
> > > Ideally they are supposed to be on different sockets, but nothing
> > > really is hard and fast in terms of rules.
> >
> > IMHO, If we are adding CGROUPSTATS_CMD_GET to the same netlink family of
> > TASKSTATS, we should allow both the commands in the same socket.
> >
> > > > Thanks
> > > > Nikanth
> > > >
> > > > Currently we never get message from kernel to userspace of type
> > > > TASKSTATS_TYPE_PID. Otherwise this could have been spotted earlier.
> > > > Having the values in the same range would restrict one not to use the
> > > > same netlink socket for a command that would return response of type
> > > > TASKSTATS_TYPE_PID and the CGROUPSTATS_CMD_GET command. Fix it by
> > > > using values after
> > > > __TASKSTATS_TYPE_MAX.
> > > >
> > > > Signed-off-by: Nikanth Karthikesan <knikanth@suse.de>
> > > >
> > > > ---
> > > >
> > > >
> > > > diff --git a/include/linux/cgroupstats.h
> > > > b/include/linux/cgroupstats.h index 3753c33..87b31f0 100644
> > > > --- a/include/linux/cgroupstats.h
> > > > +++ b/include/linux/cgroupstats.h
> > > > @@ -53,7 +53,7 @@ enum {
> > > >  #define CGROUPSTATS_CMD_MAX (__CGROUPSTATS_CMD_MAX - 1)
> > > >
> > > >  enum {
> > > > -	CGROUPSTATS_TYPE_UNSPEC = 0,	/* Reserved */
> > > > +	CGROUPSTATS_TYPE_UNSPEC = __TASKSTATS_TYPE_MAX,	/* Reserved */
> > > >  	CGROUPSTATS_TYPE_CGROUP_STATS,	/* contains name + stats */
> > > >  	__CGROUPSTATS_TYPE_MAX,
> > > >  };
> > >
> > > This seems like the correct fix
> > >
> > > Reviewed-by: Balbir Singh <balbir@linux.vnet.ibm.com>
> >
> > But this would break applications every time a new item is added to the
> > enums in taskstat.h
>
> Yes, correct and the expectation, we will bump the version number (see
> version field) or ignore the fix and send a strong recommendation that
> we should use different sockets for cgroupstats and taskstats.
>
> > The correct fix would be to add all the CGROUPSTATS_* enums, as part of
> > the same enum in taskstat.h. So that when new members are added to enum's
> > in taskstat.h, the cgroup stat enum's values won't change and
> > applications need not be re-built. If you agree, I would send a patch to
> > do that.
>
> I wanted to keep the separate to avoid #ifdef's around the code. One
> other generic fix would be to use different starting bases. We could
> for example use a bit (say bit 63) to indicate the type - taskstats or
> cgroupstats, but we might be too late for that.
>
> Please remember to bump the version field, lets try your suggested
> approach of integration and if that can be done cleanly, I have no
> objection.

I have posted[0] it with subject, "[PATCH] taskstats: Unify cgroupstats.h with 
taskstats.h and use a single nla_policy". Please do review it.

Thanks
Nikanth

[0] http://lkml.org/lkml/2009/7/13/189