Song noticed switch_mm_irqs_off taking a lot of CPU time in recent
kernels,using 1.8% of a 48 CPU system during a netperf to localhost run.
Digging into the profile, we noticed that cpumask_clear_cpu and
cpumask_set_cpu together take about half of the CPU time taken by
switch_mm_irqs_off.
However, the CPUs running netperf end up switching back and forth
between netperf and the idle task, which does not require changes
to the mm_cpumask. Furthermore, the init_mm cpumask ends up being
the most heavily contended one in the system.
Simply skipping changes to mm_cpumask(&init_mm) reduces overhead.
Signed-off-by: Rik van Riel <r...@surriel.com>
Acked-by: Dave Hansen <dave.han...@intel.com>
Reported-and-tested-by: Song Liu <songliubrav...@fb.com>
---
arch/x86/mm/tlb.c | 17 ++++++++++++-----
1 file changed, 12 insertions(+), 5 deletions(-)
diff --git a/arch/x86/mm/tlb.c b/arch/x86/mm/tlb.c
index e7c6de7eb903..c8644aa12abd 100644
--- a/arch/x86/mm/tlb.c
+++ b/arch/x86/mm/tlb.c
@@ -310,15 +310,22 @@ void switch_mm_irqs_off(struct mm_struct *prev, struct
mm_struct *next,
sync_current_stack_to_mm(next);
}
- /* Stop remote flushes for the previous mm */
- VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(real_prev)) &&
- real_prev != &init_mm);
- cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
+ /*
+ * Stop remote flushes for the previous mm.
+ * Skip the idle task; we never send init_mm TLB flushing IPIs,
+ * but the bitmap manipulation can cause cache line contention.
+ */
+ if (real_prev != &init_mm) {
+ VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu,
+ mm_cpumask(real_prev)));
+ cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
+ }
/*
* Start remote flushes and then read tlb_gen.
*/
- cpumask_set_cpu(cpu, mm_cpumask(next));
+ if (next != &init_mm)
+ cpumask_set_cpu(cpu, mm_cpumask(next));
next_tlb_gen = atomic64_read(&next->context.tlb_gen);
choose_new_asid(next, next_tlb_gen, &new_asid, &need_flush);
--
2.14.4
