Performance enhancements

Use static arrays to cache pointers to cdata objects in the nd_light
app and the datagram class.

src/apps/ipv6/nd_light.lua

@@ -183,13 +183,21 @@ function nd_light:new (arg)
    sna.eth, sna.ipv6, sna.icmp = unpack(dgram:stack())
    sna.dgram = dgram
    o._sna = sna
+
+   -- Caches for for various cdata pointer objects to avoid boxing in
+   -- the push() loop
+   o._cache = {
+      p = ffi.new("struct packet *[1]"),
+      mem = ffi.new("uint8_t *[1]")
+   }
    return o
 end
 
 -- Process neighbor solicitation
 local function ns (self, dgram, eth, ipv6, icmp)
-   local payload, length = dgram:payload()
-   if not icmp:checksum_check(payload, length, ipv6) then
+   local mem, length = self._cache.mem
+   mem[0], length = dgram:payload()
+   if not icmp:checksum_check(mem[0], length, ipv6) then
       print(self:name()..": bad icmp checksum")
       return nil
    end
@@ -209,8 +217,8 @@ local function ns (self, dgram, eth, ipv6, icmp)
    sna.ipv6:dst(ipv6:src())
    -- The payload of the pre-fabricated packet consists of the NA and
    -- target ll-option
-   local mem, length = sna.dgram:payload()
-   sna.icmp:checksum(mem, length, sna.ipv6)
+   mem[0], length = sna.dgram:payload()
+   sna.icmp:checksum(mem[0], length, sna.ipv6)
    return true
 end
 
@@ -237,10 +245,10 @@ local function na (self, dgram, eth, ipv6, icmp)
 end
 
 local function from_south (self, p)
-   if not self._filter:match(packet.data(p), packet.length(p)) then
+   if not self._filter:match(packet.data(p[0]), packet.length(p[0])) then
       return false
    end
-   local dgram = datagram:new(p, ethernet)
+   local dgram = datagram:new(p[0], ethernet)
    -- Parse the ethernet, ipv6 amd icmp headers
    dgram:parse_n(3)
    local eth, ipv6, icmp = unpack(dgram:stack())
@@ -264,22 +272,24 @@ function nd_light:push ()
       self._next_hop.timer_cb()
    end
 
+   local cache = self._cache
    local l_in = self.input.south
    local l_out = self.output.north
    local l_reply = self.output.south
    while not link.empty(l_in) and not link.full(l_out) do
-      local p = link.receive(l_in)
+      local p = cache.p
+      p[0] = link.receive(l_in)
       local status = from_south(self, p)
       if status == nil then
 	 -- Discard
-	 packet.free(p)
+	 packet.free(p[0])
       elseif status == true then
 	 -- Send NA back south
-	 packet.free(p)
+	 packet.free(p[0])
 	 link.transmit(l_reply, packet.clone(self._sna.packet))
       else
 	 -- Send transit traffic up north
-	 link.transmit(l_out, p)
+	 link.transmit(l_out, p[0])
       end
    end
 
@@ -291,11 +301,12 @@ function nd_light:push ()
 	 -- has completed.
 	 packet.free(link.receive(l_in))
       else
-	 local p = link.receive(l_in)
-         if packet.length(p) >= self._eth_header:sizeof() then
-            self._eth_header:copy(packet.data(p))
-            link.transmit(l_out, p)
-         else packet.free(p) end
+	 local p = cache.p
+	 p[0] = link.receive(l_in)
+         if packet.length(p[0]) >= self._eth_header:sizeof() then
+            self._eth_header:copy(packet.data(p[0]))
+            link.transmit(l_out, p[0])
+         else packet.free(p[0]) end
       end
    end
 end

src/lib/protocol/datagram.lua

@@ -53,6 +53,7 @@ local datagram = subClass(nil)
 local function init (o, p, class)
    if not o._recycled then
       o._parse = { stack = {}, index = 0 }
+      o._packet = ffi.new("struct packet *[1]")
    elseif o._parse.stack[1] then
       for i, _ in ipairs(o._parse.stack) do
 	 o._parse.stack[i]:free()
@@ -62,7 +63,7 @@ local function init (o, p, class)
    end
    o._parse.offset = 0
    o._parse.ulp = class
-   o._packet = p or packet.allocate()
+   o._packet[0] = p or packet.allocate()
    return o
 end
 
@@ -81,7 +82,7 @@ end
 
 -- Push a new protocol header to the front of the packet.
 function datagram:push (proto)
-   packet.prepend(self._packet, proto:header(), proto:sizeof())
+   packet.prepend(self._packet[0], proto:header(), proto:sizeof())
    self._parse.offset = self._parse.offset + proto:sizeof()
 end
 
@@ -113,8 +114,8 @@ function datagram:parse_match (class, check)
    local parse = self._parse
    local class = class or parse.ulp
    if not class then return nil end
-   local proto = class:new_from_mem(packet.data(self._packet) + parse.offset,
-                                    packet.length(self._packet) - parse.offset)
+   local proto = class:new_from_mem(packet.data(self._packet[0]) + parse.offset,
+                                    packet.length(self._packet[0]) - parse.offset)
    if proto == nil or (check and not check(proto)) then
       if proto then proto:free() end
       return nil
@@ -201,7 +202,7 @@ end
 -- Remove <length> bytes from the start of the packet and set upper-layer
 -- protocol to <ulp> if <ulp> is supplied.
 function datagram:pop_raw (length, ulp)
-   packet.shiftleft(self._packet, length)
+   packet.shiftleft(self._packet[0], length)
    if ulp then self._parse.ulp = ulp end
 end
 
@@ -210,16 +211,16 @@ function datagram:stack ()
 end
 
 function datagram:packet ()
-   return(self._packet)
+   return(self._packet[0])
 end
 
 -- Return the location and size of the packet's payload.  If mem is
 -- non-nil, the memory region at the given address and size is
 -- appended to the packet's payload first.
 function datagram:payload (mem, size)
-   if mem then packet.append(self._packet, mem, size) end
-   return packet.data(self._packet) + self._parse.offset,
-          packet.length(self._packet) - self._parse.offset
+   if mem then packet.append(self._packet[0], mem, size) end
+   return packet.data(self._packet[0]) + self._parse.offset,
+          packet.length(self._packet[0]) - self._parse.offset
 end
 
 return datagram