hyperv: Enable sendbuf mechanism on the send path
authorKY Srinivasan <kys@microsoft.com>
Wed, 30 Apr 2014 17:14:31 +0000 (10:14 -0700)
committerDavid S. Miller <davem@davemloft.net>
Wed, 30 Apr 2014 17:48:46 +0000 (13:48 -0400)
We send packets using a copy-free mechanism (this is the Guest to Host transport
via VMBUS). While this is obviously optimal for large packets,
it may not be optimal for small packets. Hyper-V host supports
a second mechanism for sending packets that is "copy based". We implement that
mechanism in this patch.

In this version of the patch I have addressed a comment from David Miller.

With this patch (and all of the other offload and VRSS patches), we are now able
to almost saturate a 10G interface between Linux VMs on Hyper-V
on different hosts - close to  9 Gbps as measured via iperf.

Signed-off-by: K. Y. Srinivasan <kys@microsoft.com>
Reviewed-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/hyperv/hyperv_net.h
drivers/net/hyperv/netvsc.c
drivers/net/hyperv/netvsc_drv.c

index d1f7826..4b7df5a 100644 (file)
@@ -140,6 +140,8 @@ struct hv_netvsc_packet {
        void *send_completion_ctx;
        void (*send_completion)(void *context);
 
+       u32 send_buf_index;
+
        /* This points to the memory after page_buf */
        struct rndis_message *rndis_msg;
 
@@ -582,6 +584,9 @@ struct nvsp_message {
 
 #define NETVSC_RECEIVE_BUFFER_SIZE             (1024*1024*16)  /* 16MB */
 #define NETVSC_RECEIVE_BUFFER_SIZE_LEGACY      (1024*1024*15)  /* 15MB */
+#define NETVSC_SEND_BUFFER_SIZE                        (1024 * 1024)   /* 1MB */
+#define NETVSC_INVALID_INDEX                   -1
+
 
 #define NETVSC_RECEIVE_BUFFER_ID               0xcafe
 
@@ -607,6 +612,15 @@ struct netvsc_device {
        u32 recv_section_cnt;
        struct nvsp_1_receive_buffer_section *recv_section;
 
+       /* Send buffer allocated by us */
+       void *send_buf;
+       u32 send_buf_size;
+       u32 send_buf_gpadl_handle;
+       u32 send_section_cnt;
+       u32 send_section_size;
+       unsigned long *send_section_map;
+       int map_words;
+
        /* Used for NetVSP initialization protocol */
        struct completion channel_init_wait;
        struct nvsp_message channel_init_pkt;
index bbee446..c041f63 100644 (file)
@@ -28,6 +28,7 @@
 #include <linux/slab.h>
 #include <linux/netdevice.h>
 #include <linux/if_ether.h>
+#include <asm/sync_bitops.h>
 
 #include "hyperv_net.h"
 
@@ -80,7 +81,7 @@ get_in_err:
 }
 
 
-static int netvsc_destroy_recv_buf(struct netvsc_device *net_device)
+static int netvsc_destroy_buf(struct netvsc_device *net_device)
 {
        struct nvsp_message *revoke_packet;
        int ret = 0;
@@ -146,10 +147,62 @@ static int netvsc_destroy_recv_buf(struct netvsc_device *net_device)
                net_device->recv_section = NULL;
        }
 
+       /* Deal with the send buffer we may have setup.
+        * If we got a  send section size, it means we received a
+        * SendsendBufferComplete msg (ie sent
+        * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
+        * to send a revoke msg here
+        */
+       if (net_device->send_section_size) {
+               /* Send the revoke receive buffer */
+               revoke_packet = &net_device->revoke_packet;
+               memset(revoke_packet, 0, sizeof(struct nvsp_message));
+
+               revoke_packet->hdr.msg_type =
+                       NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
+               revoke_packet->msg.v1_msg.revoke_recv_buf.id = 0;
+
+               ret = vmbus_sendpacket(net_device->dev->channel,
+                                      revoke_packet,
+                                      sizeof(struct nvsp_message),
+                                      (unsigned long)revoke_packet,
+                                      VM_PKT_DATA_INBAND, 0);
+               /* If we failed here, we might as well return and
+                * have a leak rather than continue and a bugchk
+                */
+               if (ret != 0) {
+                       netdev_err(ndev, "unable to send "
+                                  "revoke send buffer to netvsp\n");
+                       return ret;
+               }
+       }
+       /* Teardown the gpadl on the vsp end */
+       if (net_device->send_buf_gpadl_handle) {
+               ret = vmbus_teardown_gpadl(net_device->dev->channel,
+                                          net_device->send_buf_gpadl_handle);
+
+               /* If we failed here, we might as well return and have a leak
+                * rather than continue and a bugchk
+                */
+               if (ret != 0) {
+                       netdev_err(ndev,
+                                  "unable to teardown send buffer's gpadl\n");
+                       return ret;
+               }
+               net_device->recv_buf_gpadl_handle = 0;
+       }
+       if (net_device->send_buf) {
+               /* Free up the receive buffer */
+               free_pages((unsigned long)net_device->send_buf,
+                          get_order(net_device->send_buf_size));
+               net_device->send_buf = NULL;
+       }
+       kfree(net_device->send_section_map);
+
        return ret;
 }
 
-static int netvsc_init_recv_buf(struct hv_device *device)
+static int netvsc_init_buf(struct hv_device *device)
 {
        int ret = 0;
        int t;
@@ -248,10 +301,90 @@ static int netvsc_init_recv_buf(struct hv_device *device)
                goto cleanup;
        }
 
+       /* Now setup the send buffer.
+        */
+       net_device->send_buf =
+               (void *)__get_free_pages(GFP_KERNEL|__GFP_ZERO,
+                                        get_order(net_device->send_buf_size));
+       if (!net_device->send_buf) {
+               netdev_err(ndev, "unable to allocate send "
+                          "buffer of size %d\n", net_device->send_buf_size);
+               ret = -ENOMEM;
+               goto cleanup;
+       }
+
+       /* Establish the gpadl handle for this buffer on this
+        * channel.  Note: This call uses the vmbus connection rather
+        * than the channel to establish the gpadl handle.
+        */
+       ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
+                                   net_device->send_buf_size,
+                                   &net_device->send_buf_gpadl_handle);
+       if (ret != 0) {
+               netdev_err(ndev,
+                          "unable to establish send buffer's gpadl\n");
+               goto cleanup;
+       }
+
+       /* Notify the NetVsp of the gpadl handle */
+       init_packet = &net_device->channel_init_pkt;
+       memset(init_packet, 0, sizeof(struct nvsp_message));
+       init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
+       init_packet->msg.v1_msg.send_recv_buf.gpadl_handle =
+               net_device->send_buf_gpadl_handle;
+       init_packet->msg.v1_msg.send_recv_buf.id = 0;
+
+       /* Send the gpadl notification request */
+       ret = vmbus_sendpacket(device->channel, init_packet,
+                              sizeof(struct nvsp_message),
+                              (unsigned long)init_packet,
+                              VM_PKT_DATA_INBAND,
+                              VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
+       if (ret != 0) {
+               netdev_err(ndev,
+                          "unable to send send buffer's gpadl to netvsp\n");
+               goto cleanup;
+       }
+
+       t = wait_for_completion_timeout(&net_device->channel_init_wait, 5*HZ);
+       BUG_ON(t == 0);
+
+       /* Check the response */
+       if (init_packet->msg.v1_msg.
+           send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
+               netdev_err(ndev, "Unable to complete send buffer "
+                          "initialization with NetVsp - status %d\n",
+                          init_packet->msg.v1_msg.
+                          send_recv_buf_complete.status);
+               ret = -EINVAL;
+               goto cleanup;
+       }
+
+       /* Parse the response */
+       net_device->send_section_size = init_packet->msg.
+                               v1_msg.send_send_buf_complete.section_size;
+
+       /* Section count is simply the size divided by the section size.
+        */
+       net_device->send_section_cnt =
+               net_device->send_buf_size/net_device->send_section_size;
+
+       dev_info(&device->device, "Send section size: %d, Section count:%d\n",
+                net_device->send_section_size, net_device->send_section_cnt);
+
+       /* Setup state for managing the send buffer. */
+       net_device->map_words = DIV_ROUND_UP(net_device->send_section_cnt,
+                                            BITS_PER_LONG);
+
+       net_device->send_section_map =
+               kzalloc(net_device->map_words * sizeof(ulong), GFP_KERNEL);
+       if (net_device->send_section_map == NULL)
+               goto cleanup;
+
        goto exit;
 
 cleanup:
-       netvsc_destroy_recv_buf(net_device);
+       netvsc_destroy_buf(net_device);
 
 exit:
        return ret;
@@ -369,8 +502,9 @@ static int netvsc_connect_vsp(struct hv_device *device)
                net_device->recv_buf_size = NETVSC_RECEIVE_BUFFER_SIZE_LEGACY;
        else
                net_device->recv_buf_size = NETVSC_RECEIVE_BUFFER_SIZE;
+       net_device->send_buf_size = NETVSC_SEND_BUFFER_SIZE;
 
-       ret = netvsc_init_recv_buf(device);
+       ret = netvsc_init_buf(device);
 
 cleanup:
        return ret;
@@ -378,7 +512,7 @@ cleanup:
 
 static void netvsc_disconnect_vsp(struct netvsc_device *net_device)
 {
-       netvsc_destroy_recv_buf(net_device);
+       netvsc_destroy_buf(net_device);
 }
 
 /*
@@ -440,6 +574,12 @@ static inline u32 hv_ringbuf_avail_percent(
        return avail_write * 100 / ring_info->ring_datasize;
 }
 
+static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
+                                        u32 index)
+{
+       sync_change_bit(index, net_device->send_section_map);
+}
+
 static void netvsc_send_completion(struct netvsc_device *net_device,
                                   struct hv_device *device,
                                   struct vmpacket_descriptor *packet)
@@ -447,6 +587,7 @@ static void netvsc_send_completion(struct netvsc_device *net_device,
        struct nvsp_message *nvsp_packet;
        struct hv_netvsc_packet *nvsc_packet;
        struct net_device *ndev;
+       u32 send_index;
 
        ndev = net_device->ndev;
 
@@ -477,6 +618,9 @@ static void netvsc_send_completion(struct netvsc_device *net_device,
 
                /* Notify the layer above us */
                if (nvsc_packet) {
+                       send_index = nvsc_packet->send_buf_index;
+                       if (send_index != NETVSC_INVALID_INDEX)
+                               netvsc_free_send_slot(net_device, send_index);
                        q_idx = nvsc_packet->q_idx;
                        channel = nvsc_packet->channel;
                        nvsc_packet->send_completion(nvsc_packet->
@@ -504,6 +648,52 @@ static void netvsc_send_completion(struct netvsc_device *net_device,
 
 }
 
+static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
+{
+       unsigned long index;
+       u32 max_words = net_device->map_words;
+       unsigned long *map_addr = (unsigned long *)net_device->send_section_map;
+       u32 section_cnt = net_device->send_section_cnt;
+       int ret_val = NETVSC_INVALID_INDEX;
+       int i;
+       int prev_val;
+
+       for (i = 0; i < max_words; i++) {
+               if (!~(map_addr[i]))
+                       continue;
+               index = ffz(map_addr[i]);
+               prev_val = sync_test_and_set_bit(index, &map_addr[i]);
+               if (prev_val)
+                       continue;
+               if ((index + (i * BITS_PER_LONG)) >= section_cnt)
+                       break;
+               ret_val = (index + (i * BITS_PER_LONG));
+               break;
+       }
+       return ret_val;
+}
+
+u32 netvsc_copy_to_send_buf(struct netvsc_device *net_device,
+                           unsigned int section_index,
+                           struct hv_netvsc_packet *packet)
+{
+       char *start = net_device->send_buf;
+       char *dest = (start + (section_index * net_device->send_section_size));
+       int i;
+       u32 msg_size = 0;
+
+       for (i = 0; i < packet->page_buf_cnt; i++) {
+               char *src = phys_to_virt(packet->page_buf[i].pfn << PAGE_SHIFT);
+               u32 offset = packet->page_buf[i].offset;
+               u32 len = packet->page_buf[i].len;
+
+               memcpy(dest, (src + offset), len);
+               msg_size += len;
+               dest += len;
+       }
+       return msg_size;
+}
+
 int netvsc_send(struct hv_device *device,
                        struct hv_netvsc_packet *packet)
 {
@@ -513,6 +703,10 @@ int netvsc_send(struct hv_device *device,
        struct net_device *ndev;
        struct vmbus_channel *out_channel = NULL;
        u64 req_id;
+       unsigned int section_index = NETVSC_INVALID_INDEX;
+       u32 msg_size = 0;
+       struct sk_buff *skb;
+
 
        net_device = get_outbound_net_device(device);
        if (!net_device)
@@ -528,10 +722,26 @@ int netvsc_send(struct hv_device *device,
                sendMessage.msg.v1_msg.send_rndis_pkt.channel_type = 1;
        }
 
-       /* Not using send buffer section */
+       /* Attempt to send via sendbuf */
+       if (packet->total_data_buflen < net_device->send_section_size) {
+               section_index = netvsc_get_next_send_section(net_device);
+               if (section_index != NETVSC_INVALID_INDEX) {
+                       msg_size = netvsc_copy_to_send_buf(net_device,
+                                                          section_index,
+                                                          packet);
+                       skb = (struct sk_buff *)
+                             (unsigned long)packet->send_completion_tid;
+                       if (skb)
+                               dev_kfree_skb_any(skb);
+                       packet->page_buf_cnt = 0;
+               }
+       }
+       packet->send_buf_index = section_index;
+
+
        sendMessage.msg.v1_msg.send_rndis_pkt.send_buf_section_index =
-               0xFFFFFFFF;
-       sendMessage.msg.v1_msg.send_rndis_pkt.send_buf_section_size = 0;
+               section_index;
+       sendMessage.msg.v1_msg.send_rndis_pkt.send_buf_section_size = msg_size;
 
        if (packet->send_completion)
                req_id = (ulong)packet;
index c76b665..939e3af 100644 (file)
@@ -236,10 +236,11 @@ static void netvsc_xmit_completion(void *context)
        struct hv_netvsc_packet *packet = (struct hv_netvsc_packet *)context;
        struct sk_buff *skb = (struct sk_buff *)
                (unsigned long)packet->send_completion_tid;
+       u32 index = packet->send_buf_index;
 
        kfree(packet);
 
-       if (skb)
+       if (skb && (index == NETVSC_INVALID_INDEX))
                dev_kfree_skb_any(skb);
 }