Merge branch 'divider' of https://github.com/paulusmack/microwatt

[microwatt.git] / dmi_dtm_xilinx.vhdl

-- Xilinx internal JTAG to DMI interface
--
-- DMI bus
--
--  req : ____/------------\_____
--  addr: xxxx<            >xxxxx
--  dout: xxxx<            >xxxxx
--  wr  : xxxx<            >xxxxx
--  din : xxxxxxxxxxxx<      >xxx
--  ack : ____________/------\___
--
--  * addr/dout set along with req, can be latched on same cycle by slave
--  * ack & din remain up until req is dropped by master, the slave must
--    provide a stable output on din on reads during that time.
--  * req remains low at until at least one sysclk after ack seen down.
--
--   JTAG (tck)                    DMI (sys_clk)
--
--   * jtag_req = 1
--        (jtag_req_0)             *
--          (jtag_req_1) ->        * dmi_req = 1 >
--                                 *.../...
--                                 * dmi_ack = 1 <
--   *                         (dmi_ack_0)
--   *                   <-  (dmi_ack_1)
--   * jtag_req = 0 (and latch dmi_din)
--        (jtag_req_0)             *
--          (jtag_req_1) ->        * dmi_req = 0 >
--                                 * dmi_ack = 0 <
--  *                          (dmi_ack_0)
--  *                    <-  (dmi_ack_1)
--
--  jtag_req can go back to 1 when jtag_rsp_1 is 0
--
--  Questions/TODO:
--    - I use 2 flip fops for sync, is that enough ?
--    - I treat the jtag_reset as an async reset, is that necessary ?
--    - Dbl check reset situation since we have two different resets
--      each only resetting part of the logic...
--    - Look at optionally removing the synchronizer on the ack path,
--      assuming JTAG is always slow enough that ack will have been
--      stable long enough by the time CAPTURE comes in.
--    - We could avoid the latched request by not shifting while a
--      request is in progress (and force TDO to 1 to return a busy
--      status).
--
--  WARNING: This isn't the real DMI JTAG protocol (at least not yet).
--           a command while busy will be ignored. A response of "11"
--           means the previous command is still going, try again.
--           As such We don't implement the DMI "error" status, and
--           we don't implement DTMCS yet... This may still all change
--           but for now it's easier that way as the real DMI protocol
--           requires for a command to work properly that enough TCK
--           are sent while IDLE and I'm having trouble getting that
--           working with UrJtag and the Xilinx BSCAN2 for now.

library ieee;
use ieee.std_logic_1164.all;
use ieee.math_real.all;

library work;
use work.wishbone_types.all;

library unisim;
use unisim.vcomponents.all;

entity dmi_dtm is
    generic(ABITS : INTEGER:=8;
            DBITS : INTEGER:=32);

    port(sys_clk        : in std_ulogic;
         sys_reset      : in std_ulogic;
         dmi_addr       : out std_ulogic_vector(ABITS - 1 downto 0);
         dmi_din        : in std_ulogic_vector(DBITS - 1 downto 0);
         dmi_dout       : out std_ulogic_vector(DBITS - 1 downto 0);
         dmi_req        : out std_ulogic;
         dmi_wr         : out std_ulogic;
         dmi_ack        : in std_ulogic
--       dmi_err        : in std_ulogic TODO: Add error response
         );
end entity dmi_dtm;

architecture behaviour of dmi_dtm is

    -- Signals coming out of the BSCANE2 block
    signal jtag_reset           : std_ulogic;
    signal capture              : std_ulogic;
    signal update               : std_ulogic;
    signal drck                 : std_ulogic;
    signal jtag_clk             : std_ulogic;
    signal sel                  : std_ulogic;
    signal shift                : std_ulogic;
    signal tdi                  : std_ulogic;
    signal tdo                  : std_ulogic;
    signal tck                  : std_ulogic;

    -- ** JTAG clock domain **

    -- Shift register
    signal shiftr       : std_ulogic_vector(ABITS + DBITS + 1 downto 0);

    -- Latched request
    signal request      : std_ulogic_vector(ABITS + DBITS + 1 downto 0);

    -- A request is present
    signal jtag_req     : std_ulogic;

    -- Synchronizer for jtag_rsp (sys clk -> jtag_clk)
    signal dmi_ack_0    : std_ulogic;
    signal dmi_ack_1    : std_ulogic;

    -- ** sys clock domain **

    -- Synchronizer for jtag_req (jtag clk -> sys clk)
    signal jtag_req_0   : std_ulogic;
    signal jtag_req_1   : std_ulogic;

    -- ** combination signals
    signal jtag_bsy     : std_ulogic;
    signal op_valid     : std_ulogic;
    signal rsp_op       : std_ulogic_vector(1 downto 0);

    -- ** Constants **
    constant DMI_REQ_NOP : std_ulogic_vector(1 downto 0) := "00";
    constant DMI_REQ_RD  : std_ulogic_vector(1 downto 0) := "01";
    constant DMI_REQ_WR  : std_ulogic_vector(1 downto 0) := "10";
    constant DMI_RSP_OK  : std_ulogic_vector(1 downto 0) := "00";
    constant DMI_RSP_BSY : std_ulogic_vector(1 downto 0) := "11";

begin

    -- Implement the Xilinx bscan2 for series 7 devices (TODO: use PoC to
    -- wrap this if compatibility is required with older devices).
    bscan : BSCANE2
        generic map (
            JTAG_CHAIN          => 2
            )
        port map (
            CAPTURE             => capture,
            DRCK                => drck,
            RESET               => jtag_reset,
            RUNTEST             => open,
            SEL                 => sel,
            SHIFT               => shift,
            TCK                 => tck,
            TDI                 => tdi,
            TMS                 => open,
            UPDATE              => update,
            TDO                 => tdo
            );

    -- Some examples out there suggest buffering the clock so it's
    -- treated as a proper clock net. This is probably needed when using
    -- drck (the gated clock) but I'm using the real tck here to avoid
    -- missing the update phase so maybe not...
    --
    clkbuf : BUFG
        port map (
--          I => drck,
            I => tck,
            O => jtag_clk
            );


    -- dmi_req synchronization
    dmi_req_sync : process(sys_clk)
    begin
        -- sys_reset is synchronous
        if rising_edge(sys_clk) then
            if (sys_reset = '1') then
                jtag_req_0 <= '0';
                jtag_req_1 <= '0';
            else
                jtag_req_0 <= jtag_req;
                jtag_req_1 <= jtag_req_0;
            end if;
        end if;
    end process;
    dmi_req <= jtag_req_1;

    -- dmi_ack synchronization
    dmi_ack_sync: process(jtag_clk, jtag_reset)
    begin
        -- jtag_reset is async (see comments)
        if jtag_reset = '1' then
            dmi_ack_0 <= '0';
            dmi_ack_1 <= '0';
        elsif rising_edge(jtag_clk) then
            dmi_ack_0 <= dmi_ack;
            dmi_ack_1 <= dmi_ack_0;
        end if;
    end process;
   
    -- jtag_bsy indicates whether we can start a new request, we can when
    -- we aren't already processing one (jtag_req) and the synchronized ack
    -- of the previous one is 0.
    --
    jtag_bsy <= jtag_req or dmi_ack_1;

    -- decode request type in shift register
    with shiftr(1 downto 0) select op_valid <=
        '1' when DMI_REQ_RD,
        '1' when DMI_REQ_WR,
        '0' when others;

    -- encode response op
    rsp_op <= DMI_RSP_BSY when jtag_bsy = '1' else DMI_RSP_OK;

    -- Some DMI out signals are directly driven from the request register
    dmi_addr <= request(ABITS + DBITS + 1 downto DBITS + 2);
    dmi_dout <= request(DBITS + 1 downto 2);
    dmi_wr   <= '1' when request(1 downto 0) = DMI_REQ_WR else '0';

    -- TDO is wired to shift register bit 0
    tdo <= shiftr(0);

    -- Main state machine. Handles shift registers, request latch and
    -- jtag_req latch. Could be split into 3 processes but it's probably
    -- not worthwhile.
    --
    shifter: process(jtag_clk, jtag_reset)
    begin
        if jtag_reset = '1' then
            shiftr <= (others => '0');
            request <= (others => '0');
            jtag_req <= '0';
        elsif rising_edge(jtag_clk) then

            -- Handle jtag "commands" when sel is 1
            if sel = '1' then
                -- Shift state, rotate the register
                if shift = '1' then
                    shiftr <= tdi & shiftr(ABITS + DBITS + 1 downto 1);
                end if;

                -- Update state (trigger)
                --
                -- Latch the request if we aren't already processing one and
                -- it has a valid command opcode.
                --
                if update = '1' and op_valid = '1' then
                    if jtag_bsy = '0' then
                        request <= shiftr;
                        jtag_req <= '1';
                    end if;
                    -- Set the shift register "op" to "busy". This will prevent
                    -- us from re-starting the command on the next update if
                    -- the command completes before that.
                    shiftr(1 downto 0) <= DMI_RSP_BSY;
                end if;

                -- Request completion.
                --
                -- Capture the response data for reads and clear request flag.
                --
                -- Note: We clear req (and thus dmi_req) here which relies on tck
                -- ticking and sel set. This means we are stuck with dmi_req up if
                -- the jtag interface stops. Slaves must be resilient to this.
                --
                if jtag_req = '1' and dmi_ack_1 = '1' then
                    jtag_req <= '0';
                    if request(1 downto 0) = DMI_REQ_RD then
                        request(DBITS + 1 downto 2) <= dmi_din;
                    end if;
                end if;

                -- Capture state, grab latch content with updated status
                if capture = '1' then
                    shiftr <= request(ABITS + DBITS + 1 downto 2) & rsp_op;
                end if;

            end if;
        end if;
    end process;
end architecture behaviour;