On Fri, May 08, 2020 at 02:26:32PM +0200, Thomas Gleixner wrote: > Peter Zijlstra <pet...@infradead.org> writes: > > > On Thu, May 07, 2020 at 11:24:49PM +0200, Thomas Gleixner wrote: > >> But over our IRC conversation I came up with a 3rd variant: > >> > >> For most of the vectors the indirect call overhead is just noise, so > >> we can run them through the ASM switcher, but for the resched IPI > >> we can just use a separate direct call stub in ASM. > > > > Are we sure the rat-poison crap is noise for all the other system > > vectors? I suppose it is for most since they'll do indirect calls > > themselves anyway, right? > > We have different categories: > > 1) Uninteresting > > SPURIOUS_APIC_VECTOR, ERROR_APIC_VECTOR, THERMAL_APIC_VECTOR, > THRESHOLD_APIC_VECTOR, REBOOT_VECTOR, DEFERRED_ERROR_VECTOR > > 2) Indirect call poisoned > > LOCAL_TIMER_VECTOR > X86_PLATFORM_IPI_VECTOR > IRQ_WORK_VECTOR > HYPERV_STIMER0_VECTOR > HYPERVISOR_CALLBACK_VECTOR > POSTED_INTERRUPT_WAKEUP_VECTOR. > CALL_FUNCTION_VECTOR > CALL_FUNCTION_SINGLE_VECTOR > > 3) Quick > > RESCHEDULE_VECTOR > > POSTED_INTR_VECTOR > POSTED_INTR_NESTED_VECTOR > > These two postit ones are weird because they are both empty and > just increment different irq counts. > > HYPERV_REENLIGHTENMENT_VECTOR > > schedules delayed work, i,e. arms a timer which should be > straight forward, but does it matter? > > 4) Others > > UV_BAU_MESSAGE - The TLB flushes are probably more expensive than > ratpoutine > > Hmm?
As we just agreed on IRC, 3) can run without changing stack, and then the rest can use the indirect thing.