Use dynasm to optimize ndpi binding functions

Dynasm is used to construct a wrapper function that avoids a
struct return (that LuaJIT does not support for traced FFI
calls) by using a struct pointer.

We use dynasm to avoid having to compile a C wrapper which
complicates the build process. The wrapper isn't expressible
in Lua.

src/apps/wall/scanner/ndpi.lua

@@ -1,5 +1,6 @@
 local scanner = require("apps.wall.scanner")
 local const   = require("apps.wall.constants")
+local opt     = require("apps.wall.scanner.ndpi_opt")
 local util    = require("apps.wall.util")
 local ndpi    = require("ndpi")
 
@@ -103,12 +104,13 @@ function NdpiScanner:scan_packet(p, time)
    end
 
    flow.proto_master, flow.protocol =
-         self._ndpi:process_packet(flow._ndpi_flow,
-                                   p.data + ip_offset,
-                                   p.length - ip_offset,
-                                   time,
-                                   src_id,
-                                   dst_id)
+         opt.process_packet(self._ndpi,
+                            flow._ndpi_flow,
+                            p.data + ip_offset,
+                            p.length - ip_offset,
+                            time,
+                            src_id,
+                            dst_id)
 
    if flow.protocol ~= ndpi.protocol.PROTOCOL_UNKNOWN then
       return true, flow
@@ -121,9 +123,10 @@ function NdpiScanner:scan_packet(p, time)
       (flow.key.ip_proto == IPv4_PROTO_TCP and flow.packets > 10)
    then
       flow.proto_master, flow.protocol =
-            self._ndpi:guess_undetected_protocol(flow.key.ip_proto,
-                                                 rd32(src_addr), src_port,
-                                                 rd32(dst_addr), dst_port)
+            opt.guess_undetected_protocol(self._ndpi,
+                                          flow.key.ip_proto,
+                                          rd32(src_addr), src_port,
+                                          rd32(dst_addr), dst_port)
       -- TODO: Check whether we should check again for PROTOCOL_UNKNOWN
       return true, flow
    end

src/apps/wall/scanner/ndpi_opt.dasl

@@ -0,0 +1,103 @@
+-- This module creates a wrapper around nDPI's packet processing
+-- functions in order to help LuaJIT do its optimizations
+--
+-- Specifically, it avoids a struct value return by creating wrappers
+-- that pass a struct pointer in instead. LuaJIT can handle that
+-- better than a struct return (which is not supported).
+
+module(..., package.seeall)
+
+local dasm = require("dasm")
+local ffi  = require("ffi")
+local ndpi = require("ndpi.c").lib
+
+local debug = false
+
+|.arch x64
+|.actionlist actions
+
+-- the definitions here (anchor, assemble, gen) are borrowed from lwaftr
+-- (see multi_copy.lua)
+__anchor = {}
+
+local function assemble (name, prototype, generator)
+   local Dst = dasm.new(actions)
+   generator(Dst)
+   local mcode, size = Dst:build()
+   table.insert(__anchor, mcode)
+   if debug then
+      print("mcode dump: "..name)
+      dasm.dump(mcode, size)
+   end
+   return ffi.cast(prototype, mcode)
+end
+
+local function gen_nddpw(orig_f)
+   local function gen(Dst)
+      -- pass the first stack argument onto the original function
+      | mov rax, [rsp+8]
+      | push rax
+
+      -- call the original function, do stack cleanup
+      | mov64 rax, orig_f
+      | call rax
+      | add rsp, 8
+
+      -- at this point, rax and rdx have struct
+      -- fields in them, which we want to write into
+      -- the struct pointer (2nd stack arg)
+      | mov rcx, [rsp+16]
+      | mov [rcx], rax
+      | mov [rcx+4], rdx
+
+      | ret
+   end
+
+   return gen
+end
+
+local function gen_nupw(orig_f)
+   local function gen(Dst)
+      -- call the original function, aligning on 16
+      | sub rsp, 8
+      | mov64 rax, orig_f
+      | call rax
+      | add rsp, 8
+
+      -- like above, write into struct
+      | mov rcx, [rsp+8]
+      | mov [rcx], rax
+      | mov [rcx+4], rdx
+
+      | ret
+   end
+
+   return gen
+end
+
+-- see ljndpi/ndpi/c.lua for the corresponding headers for these functions
+-- these have an extra void* argument at the end for the struct pointer
+local function make_nddp_wrapper(f)
+   local wrap = assemble("ndpi_detection_process_packet_wrapper",
+                         ffi.typeof("void (*)(void*, void*, void*, unsigned short, uint64_t, void*, void*, void*)"),
+                         gen_nddpw(f))
+   return function(self, flow, data, len, tick, src, dst)
+             local proto = ffi.new("ndpi_protocol_t")
+             wrap(self, flow, data, len, tick, src, dst, proto)
+             return proto.master_protocol, proto.protocol
+          end
+end
+
+local function make_ngup_wrapper(f)
+   local wrap = assemble("ndpi_detection_process_packet_wrapper",
+                         ffi.typeof("void (*)(void*, uint8_t, uint32_t, uint16_t, uint32_t, uint16_t, void*)"),
+                         gen_nupw(f))
+   return function(self, prot_n, src_h, src_p, dst_h, dst_p)
+             local proto = ffi.new("ndpi_protocol_t")
+             wrap(self, prot_n, src_h, src_p, dst_h, dst_p, proto)
+             return proto.master_protocol, proto.protocol
+          end
+end
+
+process_packet = make_nddp_wrapper(ndpi.ndpi_detection_process_packet)
+guess_undetected_protocol = make_ngup_wrapper(ndpi.ndpi_guess_undetected_protocol)