Merge pull request #174 from virtualopensystems/non-zero-copy

Handle non zero copy case in virtio device

src/apps/vhost/vhost_user.c

@@ -136,13 +136,13 @@ int vhost_user_receive(int sock, struct vhost_user_msg *msg, int *fds,
     return ret;
 }
 
-void* vhost_user_map_guest_memory(int fd, int size)
+void* vhost_user_map_guest_memory(int fd, uint64_t size)
 {
     void *ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
     return ptr == MAP_FAILED ? 0 : ptr;
 }
 
-int vhost_user_unmap_guest_memory(void *ptr, int size)
+int vhost_user_unmap_guest_memory(void *ptr, uint64_t size)
 {
     return munmap(ptr, size);
 }

src/apps/vhost/vhost_user.h

@@ -58,6 +58,6 @@ int vhost_user_connect(const char *path);
 int vhost_user_send(int sock, struct vhost_user_msg *msg);
 int vhost_user_receive(int sock, struct vhost_user_msg *msg, int *fds,
         int *nfds);
-void* vhost_user_map_guest_memory(int fd, int size);
-int vhost_user_unmap_guest_memory(void *ptr, int size);
+void* vhost_user_map_guest_memory(int fd, uint64_t size);
+int vhost_user_unmap_guest_memory(void *ptr, uint64_t size);
 int vhost_user_sync_shm(void *ptr, size_t size);

src/lib/virtio/net_device.lua

@@ -176,18 +176,53 @@ end
 -- Prepared argument for writing a 1 to an eventfd.
 local eventfd_one = ffi.new("uint64_t[1]", {1})
 
+function VirtioNetDevice:more_vm_buffers ()
+   return freelist.nfree(self.vring_transmit_buffers) > 2
+end
+
+-- return the buffer from a iovec, ensuring it originates from the vm
+function VirtioNetDevice:vm_buffer (iovec)
+   local should_continue = true
+   local b = iovec.buffer
+   -- check if this is a zero-copy packet
+   if b.origin.type ~= C.BUFFER_ORIGIN_VIRTIO then
+      -- get buffer from the once supplied by the VM
+      local old_b = b
+      b = freelist.remove(self.vring_transmit_buffers)
+      --assert(iovec.offset + iovec.length <= b.size)
+
+      -- copy the whole buffer data, including offset
+      ffi.copy(b.pointer, old_b.pointer, iovec.offset + iovec.length)
+      buffer.free(old_b)
+      iovec.buffer = b
+
+      if not self:more_vm_buffers() then
+         -- no more buffers, stop the loop
+         should_continue = false
+      end
+   end
+   return should_continue, b
+end
+
 -- Transmit packets from the app input queue to the VM.
 function VirtioNetDevice:transmit_packets_to_vm ()
    local l = self.owner.input.rx
    if not l then return end
-   while not link.empty(l) do
+   local should_continue = not self.not_enough_vm_bufers
+
+   while (not link.empty(l)) and should_continue do
       local p = link.receive(l)
-      local iovec = p.iovecs[0]
-      local b = iovec.buffer
-      local virtio_hdr = b.origin.info.virtio.header_pointer
 
-      --assert(b.origin.type == C.BUFFER_ORIGIN_VIRTIO)
+      -- ensure all data is in a single buffer
+      if p.niovecs > 1 then
+         packet.coalesce(p)
+      end
+
+      local iovec = p.iovecs[0]
+      should_continue, b = self:vm_buffer(iovec)
 
+      -- fill in the virtio header
+      local virtio_hdr = b.origin.info.virtio.header_pointer
       ffi.copy(virtio_hdr, p.info, packet_info_size)
 
       local used = self.txring.used.ring[self.txused%self.tx_vring_num]
@@ -200,6 +235,11 @@ function VirtioNetDevice:transmit_packets_to_vm ()
       packet.deref(p)
    end
 
+   if not should_continue then
+      -- not enough buffers detected, verify once again
+      self.not_enough_vm_bufers = not self:more_vm_buffers()
+   end
+
    if self.txring.used.idx ~= self.txused then
       self.txring.used.idx = self.txused
       C.write(self.callfd[0], eventfd_one, 8)