From 503f1bf4d049cfb5b5a8b9cae8a1200f0d5b5cc7 Mon Sep 17 00:00:00 2001
From: Avi Kivity <avi@cloudius-systems.com>
Date: Thu, 11 Dec 2014 19:19:09 +0200
Subject: [PATCH] virtio: batch transmitted packets

Instead of placing packets directly into the virtio ring, add them to
a temporary queue, and flush it when we are polled.  This reduces
cross-cpu writes and kicks.
---
 net/virtio.cc | 28 +++++++++++++++++++++++++---
 1 file changed, 25 insertions(+), 3 deletions(-)

diff --git a/net/virtio.cc b/net/virtio.cc
index e8d7049348..a57a667e41 100644
--- a/net/virtio.cc
+++ b/net/virtio.cc
@@ -270,6 +270,7 @@ private:
     semaphore _available_descriptors = { 0 };
     int _free_head = -1;
     int _free_last = -1;
+    std::vector<uint16_t> _batch;
     std::experimental::optional<reactor::poller> _poller;
     bool _poll_mode = false;
 public:
@@ -300,6 +301,8 @@ public:
     template <typename Iterator>
     void post(Iterator begin, Iterator end);
 
+    void flush_batch();
+
     semaphore& available_descriptors() { return _available_descriptors; }
 private:
     // Let host know about interrupt delivery
@@ -412,12 +415,25 @@ void vring::run() {
         complete();
     } else {
         _poller = reactor::poller([this] {
+            flush_batch();
             do_complete();
             return true;
         });
     }
 }
 
+void vring::flush_batch() {
+    if (_batch.empty()) {
+        return;
+    }
+    for (auto desc_head : _batch) {
+        _avail._shared->_ring[masked(_avail._head++)] = desc_head;
+    }
+    _batch.clear();
+    _avail._shared->_idx.store(_avail._head, std::memory_order_release);
+    kick();
+}
+
 template <typename Iterator>
 void vring::post(Iterator begin, Iterator end) {
     // Note: buffer_chain here is any container of buffer, not
@@ -443,13 +459,19 @@ void vring::post(Iterator begin, Iterator end) {
         }
         auto desc_head = pseudo_head._next;
         _completions[desc_head] = std::move(bc.completed);
-        _avail._shared->_ring[masked(_avail._head++)] = desc_head;
+        if (!_poll_mode) {
+            _avail._shared->_ring[masked(_avail._head++)] = desc_head;
+        } else {
+            _batch.push_back(desc_head);
+        }
         _avail._avail_added_since_kick++;
     });
-    _avail._shared->_idx.store(_avail._head, std::memory_order_release);
-    kick();
     if (!_poll_mode) {
+        _avail._shared->_idx.store(_avail._head, std::memory_order_release);
+        kick();
         do_complete();
+    } else if (_batch.size() >= 16) {
+        flush_batch();
     }
 }