sched/cpupool: properly update affinity when removing a cpu from a cpupool

sched/cpupool: properly update affinity when removing a cpu from a cpupool

And this time, do it right. In fact, a similar change was
attempted in 93be8285a79c6 ("cpupools: update domU's node-affinity
on the cpupool_unassign_cpu() path"). But that was buggy, and got
reverted with 8395b67ab0b8a86.

However, even though reverting was the right thing to do, it
remains true that:
 - calling the function is better done in the cpupool cpu removal
   code, even if just for simmetry with the cpupool cpu adding path;
 - it is not necessary to call it during cpu teardown (for suspend
   or shutdown) code as we either are going down and will never
   come up (shutdown) or, when coming up, we want everything to be
   as before the tearing down process started, and so we would just
   undo any update made during the process.
 - calling it from the teardown path is not only unnecessary, but
   it can trigger an ASSERT(), in case we get, during the process,
   to remove the last online pcpu of a domain's node affinity:

  (XEN) Assertion '!cpumask_empty(dom_cpumask)' failed at domain.c:466
  (XEN) ----[ Xen-4.6-unstable  x86_64  debug=y  Tainted:    C ]----
  ... ... ...
  (XEN) Xen call trace:
  (XEN)    [<ffff82d0801055b9>] domain_update_node_affinity+0x113/0x240
  (XEN)    [<ffff82d08012e676>] cpu_disable_scheduler+0x334/0x3f2
  (XEN)    [<ffff82d08018bb8d>] __cpu_disable+0x313/0x36e
  (XEN)    [<ffff82d080101424>] take_cpu_down+0x34/0x3b
  (XEN)    [<ffff82d080130ad9>] stopmachine_action+0x70/0x99
  (XEN)    [<ffff82d08013274f>] do_tasklet_work+0x78/0xab
  (XEN)    [<ffff82d080132a85>] do_tasklet+0x5e/0x8a
  (XEN)    [<ffff82d08016478c>] idle_loop+0x56/0x6b
  (XEN)
  (XEN)
  (XEN) ****************************************
  (XEN) Panic on CPU 12:
  (XEN) Assertion '!cpumask_empty(dom_cpumask)' failed at domain.c:466
  (XEN) ****************************************

Therefore, for all these reasons, move the call from
cpu_disable_schedule() to cpupool_unassign_cpu_helper().

While there, add some sanity checking (in the latter function), and
make sure that scanning the domain list is done with domlist_read_lock
held, at least when the system is 'live'.

I re-tested the scenario described in here:
 http://permalink.gmane.org/gmane.comp.emulators.xen.devel/235310

which is what led to the revert of 93be8285a79c6, and that is
working ok after this commit.

Signed-off-by: Dario Faggioli <dario.faggioli@citrix.com>
Acked-by: George Dunlap <george.dunlap@eu.citrix.com>
Acked-by: Juergen Gross <jgross@suse.com>
Release-acked-by: Wei Liu <wei.liu2@citrix.com>