Hi Julia, Thanks for the quick response! On Fri, Mar 23, 2018 at 11:59:21AM -0500, Julia Cartwright wrote: > Hey Joe- > > Thanks for the writeup. > > On Fri, Mar 23, 2018 at 11:09:59AM -0400, joe.korty@xxxxxxxxxxxxxxxxx wrote: > > I see the below kernel splat in 4.9-rt when I run a test program that > > continually changes the affinity of some set of running pids: > > > > do not call blocking ops when !TASK_RUNNING; state=2 set at ... > > ... > > stop_one_cpu+0x60/0x80 > > migrate_enable+0x21f/0x3e0 > > rt_spin_unlock+0x2f/0x40 > > prepare_to_wait+0x5c/0x80 > > ... > > This is clearly a problem. > > > The reason is that spin_unlock, write_unlock, and read_unlock call > > migrate_enable, and since 4.4-rt, migrate_enable will sleep if it discovers > > that a migration is in order. But sleeping in the unlock services is not > > expected by most kernel developers, > > I don't buy this, see below: > > > and where that counts most is in code sequences like the following: > > > > set_current_state(TASK_UNINTERRUPIBLE); > > spin_unlock(&s); > > schedule(); > > The analog in mainline is CONFIG_PREEMPT and the implicit > preempt_enable() in spin_unlock(). In this configuration, a kernel > developer should _absolutely_ expect their task to be suspended (and > potentially migrated), _regardless of the task state_ if there is a > preemption event on the CPU on which this task is executing. > > Similarly, on RT, there is nothing _conceptually_ wrong on RT with > migrating on migrate_enable(), regardless of task state, if there is a > pending migration event. My understanding is, in standard Linux and in rt, setting task state to anything other than TASK_RUNNING in of itself blocks preemption. A preemption is not really needed here as it is expected that there is a schedule() written in that will shortly be executed. And if a 'involuntary schedule' (ie, preemption) were allowed to occur between the task state set and the schedule(), that would change the task state back to TASK_RUNNING, which would cause the schedule to NOP. Thus we risk not having paused long enough here for the condition we were waiting for to become true. > > It's clear, however, that the mechanism used here is broken ... > > Julia Thanks, Joe -- To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
