Hi, I've been reading through the perfmon2 patch for 2.6.23 and it dawned on me that in the powerpc architecture we have the notion of lazy-disabling of interrupts where we don't actually disable interrupts unless we see one come in and then turn them off. I believe a notable exception is the performance monitor interrupt which we allow to break into our critical sections. This is desirable because we want to profile those sections and right now, no in-kernel user of the performance monitor interrupt in powerpc does any synchronization across cpus inside the interrupt handler (oprofile seems to use per-cpu buffers/state for everything). We use hard_irq_disable() when we _really_ have to disable interrupts because we're switching out the kernel stack, etc.
Perfmon2 is more demanding in its synchronization requirements and it uses spin_lock_irqsave() in the generic arch-neutral code in several places, I think this will break on powerpc the way we're handling things now. I'm copying a few people to solicit opinions on how best to deal with this. I think it's desirable to keep the property of letting us profile critical sections in the kernel wherever possible. So we would need to add hard_irq_disable/enable calls to the generic code, but want to check with you guys first. Sonny _______________________________________________ Linuxppc-dev mailing list Linuxppc-dev@ozlabs.org https://ozlabs.org/mailman/listinfo/linuxppc-dev
