-- probe.lua -- watch the dwt/tcpip payloads from inside MAME.
--
-- Samples marker cells and the CPU PC once per frame.  This is independent of
-- the terminal's network stack, so it still reports after a payload has wedged
-- rdebug or TCP.  PCs inside the payload or the socket/TFTP code are printed,
-- which is how a hang is localised.
--
--   <mame> vt1200 ... -debug -debugger none \
--       -autoboot_script dwt/tcpip/probe.lua -autoboot_delay 0
--
-- Stop with PROBE_END=<frame> (default 20000).

local machine = manager.machine
local cpu = machine.devices[':maincpu']
local prog = cpu.spaces['program']

local function hex(v) return string.format("%08X", v & 0xFFFFFFFF) end

local M = {
    -- tfetch.asm
    stage = 0x01871000, fd = 0x01871020, buf = 0x01871040,
    total = 0x01871060, last = 0x01871080, done = 0x018710A0,
    close = 0x018710C0, sp = 0x01871200, a14 = 0x01871220,
    -- launcher.asm
    lstage = 0x01872500, pcb = 0x01872520,
    -- socktest.asm
    sstage = 0x01874000, sfd = 0x01874020,
    -- tftpc.asm
    cstage = 0x0187C000, cfd = 0x0187C020, cbuf = 0x0187C040,
    ctotal = 0x0187C060, clast = 0x0187C080, cblock = 0x0187C0A0,
    cdone = 0x0187C0C0,
}

local END_AT = tonumber(os.getenv("PROBE_END") or "1000000")
local last = {}
local frames = 0
local pc_lines = 0
local checked = false

-- Ground truth for the fetch: read the malloc'd buffer from inside MAME and
-- compare it with the file the server holds.  This avoids trusting rdebug peek.
local function check_buffer(buf, total)
    local f = io.open("dwt/tcpip/probe.bin", "rb")
    if not f then print("PROBE: no probe.bin to compare") return end
    local want = f:read("*a")
    f:close()
    local n = math.min(total, #want)
    local bad, first = 0, -1
    for i = 0, n - 1 do
        local got = prog:read_u8(buf + i * 8)
        if got ~= want:byte(i + 1) then
            bad = bad + 1
            if first < 0 then first = i end
        end
    end
    print(string.format("PROBE BUFFER buf=%s total=%d file=%d bad=%d first_bad=%d",
        hex(buf), total, #want, bad, first))
end

local function in_range(pc, lo, hi) return pc >= lo and pc < hi end

local function sample()
    frames = frames + 1
    local vals = {}
    for k, a in pairs(M) do vals[k] = prog:read_u32(a) end
    local pc = cpu.state['PC'].value
    local changed = false
    for k, v in pairs(vals) do
        if last[k] ~= v then changed = true end
    end
    local interesting = in_range(pc, 0x01870000, 0x01874000)
        or in_range(pc, 0x3F5C0000, 0x3F5F0000)
        or in_range(pc, 0x3F700000, 0x3F710000)
    if changed or interesting or frames % 120 == 1 then
        if interesting then pc_lines = pc_lines + 1 end
        if not interesting or pc_lines <= 400 then
            print(string.format("PROBE f=%d pc=%s stage=%s fd=%s buf=%s total=%s last=%s done=%s close=%s sp=%s a14=%s lstage=%s pcb=%s sstage=%s sfd=%s",
                frames, hex(pc), hex(vals.stage), hex(vals.fd), hex(vals.buf),
                hex(vals.total), hex(vals.last), hex(vals.done),
                hex(vals.close), hex(vals.sp), hex(vals.a14),
                hex(vals.lstage), hex(vals.pcb), hex(vals.sstage), hex(vals.sfd)))
            print(string.format("TCPC  cstage=%s cfd=%s cbuf=%s ctotal=%s clast=%s cblock=%s cdone=%s",
                hex(vals.cstage), hex(vals.cfd), hex(vals.cbuf), hex(vals.ctotal),
                hex(vals.clast), hex(vals.cblock), hex(vals.cdone)))
        end
    end
    if vals.cdone == 0x05A5A5A5 and not checked then
        checked = true
        check_buffer(vals.cbuf, vals.ctotal)
    end
    last = vals
    if frames >= END_AT then
        print("PROBE end")
        machine:exit()
    end
end

if not machine.debugger then
    print("PROBE ERROR: pass -debug")
    machine:exit()
    return
end

emu.register_frame_done(sample, "probe")
print("PROBE armed")