diff --git a/core/reactor.cc b/core/reactor.cc
index ec0e603ea61..46cce7794df 100644
--- a/core/reactor.cc
+++ b/core/reactor.cc
@@ -1667,14 +1667,31 @@ class reactor::lowres_timer_pollfn final : public reactor::pollfn {
 
 class reactor::smp_pollfn final : public reactor::pollfn {
     reactor& _r;
+    struct aligned_flag {
+        std::atomic<bool> flag;
+        char pad[63];
+        bool try_lock() {
+            return !flag.exchange(true, std::memory_order_relaxed);
+        }
+        void unlock() {
+            flag.store(false, std::memory_order_relaxed);
+        }
+    };
+    static aligned_flag _membarrier_lock;
 public:
     smp_pollfn(reactor& r) : _r(r) {}
     virtual bool poll() final override {
         return smp::poll_queues();
     }
     virtual bool try_enter_interrupt_mode() override {
+        // systemwide_memory_barrier() is very slow if run concurrently,
+        // so don't go to sleep if it is running now.
+        if (!_membarrier_lock.try_lock()) {
+            return false;
+        }
         _r._sleeping.store(true, std::memory_order_relaxed);
         systemwide_memory_barrier();
+        _membarrier_lock.unlock();
         if (poll()) {
             // raced
             _r._sleeping.store(false, std::memory_order_relaxed);
@@ -1687,6 +1704,8 @@ class reactor::smp_pollfn final : public reactor::pollfn {
     }
 };
 
+alignas(64) reactor::smp_pollfn::aligned_flag reactor::smp_pollfn::_membarrier_lock;
+
 class reactor::epoll_pollfn final : public reactor::pollfn {
     reactor& _r;
 public: