Having flush_tlb_mm->smp_flush_tlb_mm() send an IPI to every cpu on the system is occasionally triggering spin_lock contention in generic_smp_call_function_interrupt(). The x86 arch only sends IPIs to the cpus in mm->cpu_vm_mask. Does anybody see any reason why ia64 shouldn't do this as well? I have noticed significant improvement in this contention issue with the following patch. Signed-off-by: Dimitri Sivanich <sivanich@xxxxxxx> --- arch/ia64/kernel/smp.c | 13 +++++-------- 1 files changed, 5 insertions(+), 8 deletions(-) diff --git a/arch/ia64/kernel/smp.c b/arch/ia64/kernel/smp.c index 2ea4199..5c1787b 100644 --- a/arch/ia64/kernel/smp.c +++ b/arch/ia64/kernel/smp.c @@ -300,15 +300,12 @@ smp_flush_tlb_mm (struct mm_struct *mm) return; } + smp_call_function_mask(mm->cpu_vm_mask, + (void (*)(void *))local_finish_flush_tlb_mm, mm, 1); + local_irq_disable(); + local_finish_flush_tlb_mm(mm); + local_irq_enable(); preempt_enable(); - /* - * We could optimize this further by using mm->cpu_vm_mask to track which CPUs - * have been running in the address space. It's not clear that this is worth the - * trouble though: to avoid races, we have to raise the IPI on the target CPU - * anyhow, and once a CPU is interrupted, the cost of local_flush_tlb_all() is - * rather trivial. - */ - on_each_cpu((void (*)(void *))local_finish_flush_tlb_mm, mm, 1); } void arch_send_call_function_single_ipi(int cpu) -- 1.5.4.rc3 -- To unsubscribe from this list: send the line "unsubscribe linux-ia64" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
|