Background:

Several race conditions in the scheduler have cropped up
recently, which Steven and I have tracked down using ftrace.
The most recent one turns out to be a race in how the scheduler
determines a suitable migration target for RT tasks, introduced
recently with commit:

    commit 68e74568
    Date:   Tue Nov 25 02:35:13 2008 +1030

        sched: convert struct cpupri_vec cpumask_var_t.

The original design of cpupri allowed lockless readers to
quickly determine a best-estimate target.  Races between the
pri_active bitmap and the vec->mask were handled in the
original code because we would detect and return "0" when this
occured.  The design was predicated on the *effective*
atomicity (*) of caching the result of cpus_and() between the
cpus_allowed and the vec->mask.

Commit 68e74568 changed the behavior such that vec->mask is
accessed multiple times.  This introduces a subtle race, the
result of which means we can have a result that returns "1",
but with an empty bitmap.

*) yes, we know cpus_and() is not a locked operator across the
   entire composite array, but it is implicitly atomic on a
   per-word basis which is all the design required to work.

Implementation:

Rather than forgoing the lockless design, or reverting to a
stack-based cpumask_t, we simply check for when the race has
been encountered and continue processing in the event that the
race is hit.  This renders the removal race as if the priority
bit had been atomically cleared as well, and allows the
algorithm to execute correctly.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
CC: Rusty Russell <rusty@rustcorp.com.au>
CC: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090730145728.25226.92769.stgit@dev.haskins.net>
Signed-off-by: Ingo Molnar <mingo@elte.hu>