Fix lwAFTR trace performance issues (for release Alpina)

src/apps/lwaftr/binding_table.lua

@@ -78,26 +78,29 @@ psid_map_value_t = ffi.typeof[[
 
 BTLookupQueue = {}
 
+local BTLookupQueue_size = 128
+assert(BTLookupQueue_size >= engine.pull_npackets)
+
 -- BTLookupQueue needs a binding table to get softwires and PSID lookup.
 function BTLookupQueue.new(binding_table)
    local ret = {
       binding_table = assert(binding_table),
    }
-   ret.streamer = binding_table.softwires:make_lookup_streamer(32)
-   ret.packet_queue = ffi.new("struct packet * [32]")
+   ret.streamer = binding_table.softwires:make_lookup_streamer(BTLookupQueue_size)
+   ret.packet_queue = ffi.new("struct packet * [?]", BTLookupQueue_size)
    ret.length = 0
    return setmetatable(ret, {__index=BTLookupQueue})
 end
 
 function BTLookupQueue:enqueue_lookup(pkt, ipv4, port)
+   assert(self.length < BTLookupQueue_size, "BTLookupQueue overflow")
    local n = self.length
    local streamer = self.streamer
    streamer.entries[n].key.ipv4 = ipv4
    streamer.entries[n].key.psid = port
    self.packet_queue[n] = pkt
    n = n + 1
    self.length = n
-   return n == 32
 end
 
 function BTLookupQueue:process_queue()

src/apps/lwaftr/lwaftr.lua

@@ -780,16 +780,10 @@ end
 
 function LwAftr:enqueue_encapsulation(pkt, ipv4, port, pkt_src_link)
    if pkt_src_link == PKT_FROM_INET then
-      if self.inet_lookup_queue:enqueue_lookup(pkt, ipv4, port) then
-         -- Flush the queue right away if enough packets are queued up already.
-         self:flush_encapsulation()
-      end
+      self.inet_lookup_queue:enqueue_lookup(pkt, ipv4, port)
    else
       assert(pkt_src_link == PKT_HAIRPINNED)
-      if self.hairpin_lookup_queue:enqueue_lookup(pkt, ipv4, port) then
-         -- Flush the queue right away if enough packets are queued up already.
-         self:flush_hairpin()
-      end
+      self.hairpin_lookup_queue:enqueue_lookup(pkt, ipv4, port)
    end
 end
 
@@ -977,10 +971,7 @@ function LwAftr:flush_decapsulation()
 end
 
 function LwAftr:enqueue_decapsulation(pkt, ipv4, port)
-   if self.inet_lookup_queue:enqueue_lookup(pkt, ipv4, port) then
-      -- Flush the queue right away if enough packets are queued up already.
-      self:flush_decapsulation()
-   end
+   self.inet_lookup_queue:enqueue_lookup(pkt, ipv4, port)
 end
 
 -- FIXME: Verify that the packet length is big enough?

src/lib/ctable.lua

@@ -509,7 +509,6 @@ end
 function LookupStreamer:stream()
    local width = self.width
    local entries = self.entries
-   local keys = self.keys
    local pointers = self.pointers
    local stream_entries = self.stream_entries
    local entries_per_lookup = self.entries_per_lookup

src/lib/hash/siphash.dasl

@@ -576,8 +576,6 @@ local function make_hash2(opts)
    local hash = make_hash1(opts)
    local stride = opts.stride or opts.size
    return function(ptr, result)
-      ptr = ffi.cast('uint8_t*', ptr)
-      result = ffi.cast('uint32_t*', result)
       result[0] = hash(ptr)
       result[1] = hash(ptr + stride)
    end
@@ -587,8 +585,6 @@ local function make_hash4(opts)
    local hash = make_hash2(opts)
    local stride = opts.stride or opts.size
    return function(ptr, result)
-      ptr = ffi.cast('uint8_t*', ptr)
-      result = ffi.cast('uint32_t*', result)
       hash(ptr, result)
       hash(ptr + stride*2, result + 2)
    end
@@ -612,19 +608,13 @@ function make_multi_hash(opts)
 
    if width % 4 == 0 then
       return function(input, output)
-         input = ffi.cast('uint8_t*', input)
-         output = ffi.cast('uint32_t*', output)
-         for i=1,width,4 do
-            hash4(input, output)
-            input = input + stride*4
-            output = output + 4
+         for i=0,width-1,4 do
+            hash4(input + stride*i, output + i)
          end
       end
    end
 
    return function(input, output)
-      input = ffi.cast('uint8_t*', input)
-      output = ffi.cast('uint32_t*', output)
       for i=1,bit.rshift(width, 2) do
          hash4(input, output)
          input = input + stride*4