Re: [PATCH] cpuset code prevents binding tasks to new cpus

[Paul Jackson <pj@sgi.com>]

> Will it actually break anything?

It will break any user code that thought it could actually run on the
CPUs listed in its cpuset, if it happens to be in a task in the top
cpuset.  I'm pretty sure I've got some cpuset users who would be
harmed by this change.

One degree of freedom for change we do have is that, as best I know,
no one is doing anything serious with both cpu hotplug/unplug and with
cpusets at the same time on the same system.  For instance, I am
willing to wager that no one is counting on the CPUs in the top cpuset
remaining constant if a CPU comes on or off line.  I say this because
so far as I know, serious cpuset users aren't taking CPUs on and
off line.  I've rather been expecting cpusets and cpu hotplug to
butt heads for a couple of years now.  Looks like the time has come.

So if I'm right, we could change the API to have the top_cpuset
cpus_allowed track the cpus_online_map dynamically, rather than being a
static copy of the cpus_online_map value at system boot, with little
or no negative impact on users.

For adding CPUs, that is easy enough, using a register_cpu_notifier
callback to add new CPUs to top_cpuset.cpus_allowed.  This may seem
like overkill to Nathan, but I think it is necessary, to keep the
top cpusets CPUs tracking what's online, rather than changing it to
what's possible (which can be a -much- bigger set of CPUs on some
configurations, and likely surprising to existing cpuset-aware code.)
Automatically adding newly onlined CPUs to just the top cpuset (but not
to other cpusets) does treat the top cpuset as a special case, but I
doubt this will surprise existing cpuset aware code, and corresponds
nicely to what hotplug aware, cpuset clueless code will expect.

For removing CPUs, this is a bit harder, as one cannot remove a CPU
from a cpuset without first removing it from any child cpusets of
that cpuset.  So I will need to scan the cpuset hierarchy, from the
bottom up, removing the CPU about to be offline'd, and in the case that
this was the last CPU in a cpuset, finding some fallback setting for
that cpuset.  I suspect that means moving any poor tasks left in that
soon to be useless (no CPUs) cpuset into the parent cpuset, then
empty'ing the cpus_allowed for that cpuset.  If the user doesn't like
this default action, they should have done what they wanted first, by
adapting their cpuset configuration in anticipation of taking the CPU
offline.  Taking CPUs offline will likely surprise (as in break)
existing cpuset aware code, but I don't know any way around that.

In any event, as a workaround with existing kernels, I suspect you could
make use of the existing /sbin/hotplug mechanism to run the (bash syntax)
following commands to add a newly online CPU $cpu to the top cpuset's cpus:

	test -d /dev/cpuset || mkdir /dev/cpuset
	test -f /dev/cpuset/cpus || mount -t cpuset cpuset /dev/cpuset
	/bin/echo $(</dev/cpuset/cpus),$cpu > /dev/cpuset/cpus

This workaround presumes that the number of the CPU being added, $cpu,
is visible to the user level hotplug script - offhand I don't know if
that is so or not.

-- 
                  I won't rest till it's the best ...
                  Programmer, Linux Scalability
                  Paul Jackson <pj@sgi.com> 1.925.600.0401
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/