Skip to contentSkip to conversation

non-atomic test victim #1

DISCUSSION Robert Love 1 participant 1 message Sep 17, 2002

lore
Share & tools
lore
#0 lore raw

Current kernel with the non-atomic test I sent, SMP with preemption.

I only get two triggers, both during boot. They are on migration_thread and ksoftirqd startup. The problem lies in set_cpus_allowed():

    Trace; c0116ac4 <schedule+a4/4b0>
    Trace; c011731a <wait_for_completion+11a/1d0>
    Trace; c0116f30 <default_wake_function+0/40>
    Trace; c0115d02 <try_to_wake_up+312/320>
    Trace; c0116f30 <default_wake_function+0/40>
    Trace; c0118f0f <set_cpus_allowed+22f/250>
    Trace; c0118f7d <migration_thread+4d/590>
    Trace; c0118f30 <migration_thread+0/590>
    Trace; c0118f30 <migration_thread+0/590>
    Trace; c010586d <kernel_thread_helper+5/18>

It is obviously the preempt_disable() which we hold past the wake_up().

The issue is that, without this preempt_disable() there have been observed crashes, especially on large n-way machines. Both Andrew Morton and Anton Blanchard have reported the problem and that this fixes it.

Question is, why does set_cpus_allowed() need it? I do not see it... it must be an issue with an early preemption and the resulting migration_thread?

Ingo, ideas?

Robert Love