netvsc: optimize receive completions

Optimize how receive completion ring are managed.
   * Allocate only as many slots as needed for all buffers from host
   * Allocate before setting up sub channel for better error detection
   * Don't need to keep copy of initial receive section message
   * Precompute the watermark for when receive flushing is needed
   * Replace division with conditional test
   * Replace atomic per-device variable with per-channel check.
   * Handle corner case where receive completion send
     fails if ring buffer to host is full.
Signed-off-by: Stephen Hemminger <sthemmin@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

drivers/net/hyperv/hyperv_net.h

@@ -186,6 +186,7 @@ struct net_device_context;
 
 struct netvsc_device *netvsc_device_add(struct hv_device *device,
 					const struct netvsc_device_info *info);
+int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx);
 void netvsc_device_remove(struct hv_device *device);
 int netvsc_send(struct net_device_context *ndc,
 		struct hv_netvsc_packet *packet,
@@ -657,11 +658,8 @@ struct recv_comp_data {
 	u32 status;
 };
 
-/* Netvsc Receive Slots Max */
-#define NETVSC_RECVSLOT_MAX (NETVSC_RECEIVE_BUFFER_SIZE / ETH_DATA_LEN + 1)
-
 struct multi_recv_comp {
-	void *buf; /* queued receive completions */
+	struct recv_comp_data *slots;
 	u32 first;	/* first data entry */
 	u32 next;	/* next entry for writing */
 };
@@ -750,7 +748,7 @@ struct netvsc_device {
 	u32 recv_buf_size;
 	u32 recv_buf_gpadl_handle;
 	u32 recv_section_cnt;
-	struct nvsp_1_receive_buffer_section *recv_section;
+	u32 recv_completion_cnt;
 
 	/* Send buffer allocated by us */
 	void *send_buf;
@@ -778,8 +776,6 @@ struct netvsc_device {
 	u32 max_pkt; /* max number of pkt in one send, e.g. 8 */
 	u32 pkt_align; /* alignment bytes, e.g. 8 */
 
-	atomic_t num_outstanding_recvs;
-
 	atomic_t open_cnt;
 
 	struct netvsc_channel chan_table[VRSS_CHANNEL_MAX];

drivers/net/hyperv/rndis_filter.c

@@ -928,12 +928,12 @@ static bool netvsc_device_idle(const struct netvsc_device *nvdev)
 {
 	int i;
 
-	if (atomic_read(&nvdev->num_outstanding_recvs) > 0)
-		return false;
-
 	for (i = 0; i < nvdev->num_chn; i++) {
 		const struct netvsc_channel *nvchan = &nvdev->chan_table[i];
 
+		if (nvchan->mrc.first != nvchan->mrc.next)
+			return false;
+
 		if (atomic_read(&nvchan->queue_sends) > 0)
 			return false;
 	}
@@ -1031,11 +1031,6 @@ static void netvsc_sc_open(struct vmbus_channel *new_sc)
 		return;
 
 	nvchan = nvscdev->chan_table + chn_index;
-	nvchan->mrc.buf
-		= vzalloc(NETVSC_RECVSLOT_MAX * sizeof(struct recv_comp_data));
-
-	if (!nvchan->mrc.buf)
-		return;
 
 	/* Because the device uses NAPI, all the interrupt batching and
 	 * control is done via Net softirq, not the channel handling
@@ -1225,6 +1220,15 @@ struct netvsc_device *rndis_filter_device_add(struct hv_device *dev,
 	if (num_rss_qs == 0)
 		return net_device;
 
+	for (i = 1; i < net_device->num_chn; i++) {
+		ret = netvsc_alloc_recv_comp_ring(net_device, i);
+		if (ret) {
+			while (--i != 0)
+				vfree(net_device->chan_table[i].mrc.slots);
+			goto out;
+		}
+	}
+
 	refcount_set(&net_device->sc_offered, num_rss_qs);
 	vmbus_set_sc_create_callback(dev->channel, netvsc_sc_open);