arch/checksum: fix bug in carry propagation

This fixes a bug in our scalar x86 IP checksum implementation, where some
carries where missed. Really hard to find with random testing, added a
systematic test case for carry propagation.

src/arch/checksum.dasl

@@ -71,20 +71,23 @@ local function gen_checksum ()
       | jl >5                       -- Jump to branch '3'.
       | mov r9d, dword [rdi + r8]   -- Fetch 32-bit from data + r8 into r9d.
       | add rax, r9                 -- Sum acc with r9. Accumulate carry.
+      | adc rax, 0                  -- Sum carry-bit into acc.
       | sub rcx, 4                  -- Decrease index by 4.
       | add r8, 4                   -- Next 32-bit.
       | 5:
       | cmp rcx, 2                  -- If index is less than 2.
       | jl >6                       -- Jump to branch '4'.
       | movzx r9, word [rdi + r8]   -- Fetch 16-bit from data + r8 into r9.
       | add rax, r9                 -- Sum acc with r9. Accumulate carry.
+      | adc rax, 0                  -- Sum carry-bit into acc.
       | sub rcx, 2                  -- Decrease index by 2.
       | add r8, 2                   -- Next 16-bit.
       | 6:
       | cmp rcx, 1                  -- If index is less than 1.
       | jl >7                       -- Jump to branch '5'.
       | movzx r9, byte [rdi + r8]   -- Fetch 8-bit from data + r8 into r9.
       | add rax, r9                 -- Sum acc with r9. Accumulate carry.
+      | adc rax, 0                  -- Sum carry-bit into acc.
       -- Fold 64-bit into 16-bit.
       | 7:
       | mov r9, rax                 -- Assign acc to r9.
@@ -157,4 +160,11 @@ function selftest ()
       assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(ntohs(checksum_lua(pkt.data, pkt.length))))
       assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(C.cksum_generic(pkt.data, pkt.length, 0)))
    end
+   -- Test carry propagation
+   for l = 1, 63 do
+      local pkt = { data = ffi.new("uint8_t[?]", l), length = l }
+      for i = 0, l-2 do pkt.data[i]=0xff end; pkt.data[l-1] = 0x01
+      assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(ntohs(checksum_lua(pkt.data, pkt.length))))
+      assert(hex(checksum(pkt.data, pkt.length, 0)) == hex(C.cksum_generic(pkt.data, pkt.length, 0)))
+   end
 end