// retroDE_ps2 — tb_sif_dma_overflow
//
// Full-stop backpressure proof. Instantiates sif_dma_stub with DEPTH=2 and
// issues a QWC=4 DMAC transfer. After beats 0 and 1 land, the receive
// buffer is full; `in_ready` must drop and the DMAC must stall waiting
// for space that never comes (this stub has no consume path yet).
//
// Pass criteria (per Codex capacity-hardening spec):
//   - exactly 2 beats accepted (rx_count == 2)
//   - full_o asserts
//   - no additional beats accepted while full
//   - no DMA_DONE
//   - two stored qwords intact and ordered
//   - DMAC is still trying to emit the next beat at the expected source
//     address (MADR + 32) — verified by internal state peek

`timescale 1ns/1ps

module tb_sif_dma_overflow;

    logic clk;
    logic rst_n;

    initial clk = 1'b0;
    always #5 clk = ~clk;

    localparam int QWC_VAL = 4;
    localparam int RAM_SIZE_BYTES = 4 * 1024;
    localparam int RAM_ADDR_W     = $clog2(RAM_SIZE_BYTES);
    localparam int SIF_DEPTH      = 2;                 // deliberately small
    localparam int SIF_IDX_W      = $clog2(SIF_DEPTH);

    localparam logic [31:0] PAYLOAD_MADR = 32'h0000_0400;

    // ------------------------------------------------------------------
    // ee_ram_stub
    // ------------------------------------------------------------------

    logic                  ram_rd_en;
    logic [RAM_ADDR_W-1:0] ram_rd_addr;
    logic [127:0]          ram_rd_data;
    logic                  ram_rd_valid;

    logic                  ram_wr_en;
    logic [RAM_ADDR_W-1:0] ram_wr_addr;
    logic [127:0]          ram_wr_data;
    logic [15:0]           ram_wr_be;

    logic [7:0]            ram_master_id;

    logic                ram_ev_valid;
    trace_pkg::subsys_e  ram_ev_subsys;
    trace_pkg::event_e   ram_ev_event;
    logic [63:0]         ram_ev_arg0, ram_ev_arg1, ram_ev_arg2, ram_ev_arg3;
    logic [31:0]         ram_ev_flags;

    ee_ram_stub #(.SIZE_BYTES(RAM_SIZE_BYTES)) u_ram (
        .clk(clk), .rst_n(rst_n),
        .rd_en(ram_rd_en), .rd_addr(ram_rd_addr),
        .rd_data(ram_rd_data), .rd_valid(ram_rd_valid),
        .wr_en(ram_wr_en), .wr_addr(ram_wr_addr),
        .wr_data(ram_wr_data), .wr_be(ram_wr_be),
        .master_id(ram_master_id),
        .ev_valid(ram_ev_valid), .ev_subsys(ram_ev_subsys),
        .ev_event(ram_ev_event),
        .ev_arg0(ram_ev_arg0), .ev_arg1(ram_ev_arg1),
        .ev_arg2(ram_ev_arg2), .ev_arg3(ram_ev_arg3),
        .ev_flags(ram_ev_flags)
    );

    assign ram_master_id = ram_rd_en ? 8'd1 : 8'd0;

    // ------------------------------------------------------------------
    // DMAC (channel 5)
    // ------------------------------------------------------------------

    logic        dmac_reg_wr_en;
    logic [7:0]  dmac_reg_offset;
    logic [31:0] dmac_reg_wr_data;

    logic         dmac_mem_rd_en;
    logic [31:0]  dmac_mem_rd_addr;
    logic [127:0] map_to_dmac_rd_data;
    logic         map_to_dmac_rd_valid;

    logic         dmac_ep_valid;
    logic [127:0] dmac_ep_data;
    logic         dmac_ep_last;
    logic         dmac_ep_ready;

    logic                dmac_ev_valid;
    trace_pkg::subsys_e  dmac_ev_subsys;
    trace_pkg::event_e   dmac_ev_event;
    logic [63:0]         dmac_ev_arg0, dmac_ev_arg1, dmac_ev_arg2, dmac_ev_arg3;
    logic [31:0]         dmac_ev_flags;

    dmac_reg_stub #(.CHANNEL(4'd5), .PATH_ID(4'd5)) u_dmac (
        .clk(clk), .rst_n(rst_n),
        .reg_wr_en(dmac_reg_wr_en), .reg_offset(dmac_reg_offset),
        .reg_wr_data(dmac_reg_wr_data),
        .reg_rd_en(1'b0), .reg_rd_data(), .reg_rd_valid(),
        .mem_rd_en(dmac_mem_rd_en), .mem_rd_addr(dmac_mem_rd_addr),
        .mem_rd_data(map_to_dmac_rd_data), .mem_rd_valid(map_to_dmac_rd_valid),
        .ep_valid(dmac_ep_valid), .ep_data(dmac_ep_data),
        .ep_last(dmac_ep_last), .ep_ready(dmac_ep_ready),
        .irq_completion_o(),
        .ev_valid(dmac_ev_valid), .ev_subsys(dmac_ev_subsys),
        .ev_event(dmac_ev_event),
        .ev_arg0(dmac_ev_arg0), .ev_arg1(dmac_ev_arg1),
        .ev_arg2(dmac_ev_arg2), .ev_arg3(dmac_ev_arg3),
        .ev_flags(dmac_ev_flags)
    );

    // ------------------------------------------------------------------
    // ee_memory_map_stub
    // ------------------------------------------------------------------

    logic        map_ram_rd_en;
    logic [24:0] map_ram_rd_addr;

    logic                map_ev_valid;
    trace_pkg::subsys_e  map_ev_subsys;
    trace_pkg::event_e   map_ev_event;
    logic [63:0]         map_ev_arg0, map_ev_arg1, map_ev_arg2, map_ev_arg3;
    logic [31:0]         map_ev_flags;

    ee_memory_map_stub u_map (
        .clk(clk), .rst_n(rst_n),
        .ee_rd_en(1'b0), .ee_rd_addr(32'd0),
        .ee_rd_data(), .ee_rd_valid(),
        .ee_wr_en(1'b0), .ee_wr_addr(32'd0),
        .ee_wr_data(32'd0), .ee_wr_be(4'd0),
        .dmac_rd_en(dmac_mem_rd_en), .dmac_rd_addr(dmac_mem_rd_addr),
        .dmac_rd_data(map_to_dmac_rd_data),
        .dmac_rd_valid(map_to_dmac_rd_valid),
        .bios_rd_en(), .bios_rd_addr(),
        .bios_rd_data(32'd0), .bios_rd_valid(1'b0),
        .ram_rd_en(map_ram_rd_en), .ram_rd_addr(map_ram_rd_addr),
        .ram_rd_data(ram_rd_data), .ram_rd_valid(ram_rd_valid),
        .bridge_wr_en(1'b0), .bridge_wr_addr(32'd0),
        .bridge_wr_data(128'd0), .bridge_wr_be(16'd0),
        .bridge_master_id(8'd0),
        .ram_wr_en(), .ram_wr_addr(), .ram_wr_data(),
        .ram_wr_be(), .ram_master_id(),
        .ee_dmac_ch2_wr_en(), .ee_dmac_ch2_wr_addr(), .ee_dmac_ch2_wr_data(),
        .ee_dmac_ch2_rd_en(), .ee_dmac_ch2_rd_addr(),
        .ee_dmac_ch2_rd_data(32'd0), .ee_dmac_ch2_rd_valid(1'b0),
        .ee_intc_wr_en(), .ee_intc_wr_addr(), .ee_intc_wr_data(),
        .ee_intc_rd_en(), .ee_intc_rd_addr(),
        .ee_intc_rd_data(32'd0), .ee_intc_rd_valid(1'b0),
        .ee_misc_mmio_wr_en(), .ee_misc_mmio_wr_addr(), .ee_misc_mmio_wr_data(), .ee_misc_mmio_wr_be(),
        .ee_misc_mmio_rd_en(), .ee_misc_mmio_rd_addr(),
        .ee_misc_mmio_rd_data(32'd0), .ee_misc_mmio_rd_valid(1'b0),
        .ee_biu_wr_en(), .ee_biu_wr_addr(), .ee_biu_wr_data(), .ee_biu_wr_be(),
        .ee_biu_rd_en(), .ee_biu_rd_addr(),
        .ee_biu_rd_data(32'd0), .ee_biu_rd_valid(1'b0),
        .ev_valid(map_ev_valid), .ev_subsys(map_ev_subsys),
        .ev_event(map_ev_event),
        .ev_arg0(map_ev_arg0), .ev_arg1(map_ev_arg1),
        .ev_arg2(map_ev_arg2), .ev_arg3(map_ev_arg3),
        .ev_flags(map_ev_flags)
    );

    assign ram_rd_en   = map_ram_rd_en;
    assign ram_rd_addr = map_ram_rd_addr[RAM_ADDR_W-1:0];

    // ------------------------------------------------------------------
    // sif_dma_stub (DEPTH=2)
    // ------------------------------------------------------------------

    logic                  sif_rd_en;
    logic [SIF_IDX_W-1:0]  sif_rd_idx;
    logic [127:0]          sif_rd_data;
    logic                  sif_rd_valid;

    logic                  sif_stall;

    logic [31:0]           sif_rx_count;
    logic                  sif_last_seen;
    logic                  sif_full;

    logic                sif_ev_valid;
    trace_pkg::subsys_e  sif_ev_subsys;
    trace_pkg::event_e   sif_ev_event;
    logic [63:0]         sif_ev_arg0, sif_ev_arg1, sif_ev_arg2, sif_ev_arg3;
    logic [31:0]         sif_ev_flags;

    sif_dma_stub #(.DEPTH(SIF_DEPTH)) u_sif (
        .clk(clk), .rst_n(rst_n),
        .in_valid(dmac_ep_valid), .in_data(dmac_ep_data),
        .in_last(dmac_ep_last),   .in_ready(dmac_ep_ready),
        .rd_en(sif_rd_en), .rd_idx(sif_rd_idx),
        .rd_data(sif_rd_data), .rd_valid(sif_rd_valid),
        .stall_in(sif_stall),
        .rx_count(sif_rx_count), .last_seen(sif_last_seen),
        .full_o(sif_full),
        .ev_valid(sif_ev_valid), .ev_subsys(sif_ev_subsys),
        .ev_event(sif_ev_event),
        .ev_arg0(sif_ev_arg0), .ev_arg1(sif_ev_arg1),
        .ev_arg2(sif_ev_arg2), .ev_arg3(sif_ev_arg3),
        .ev_flags(sif_ev_flags)
    );

    // ------------------------------------------------------------------
    // Trace sinks (single combined file for brevity)
    // ------------------------------------------------------------------

    trace_sink_stub #(.FILENAME("sif_dma_overflow_dmac.trace"), .SINK_LABEL("dmac"))
        u_trace_dmac (.clk(clk), .rst_n(rst_n),
            .ev_valid(dmac_ev_valid), .ev_subsys(dmac_ev_subsys),
            .ev_event(dmac_ev_event), .ev_arg0(dmac_ev_arg0),
            .ev_arg1(dmac_ev_arg1), .ev_arg2(dmac_ev_arg2),
            .ev_arg3(dmac_ev_arg3), .ev_flags(dmac_ev_flags));

    trace_sink_stub #(.FILENAME("sif_dma_overflow_sif.trace"), .SINK_LABEL("sif"))
        u_trace_sif (.clk(clk), .rst_n(rst_n),
            .ev_valid(sif_ev_valid), .ev_subsys(sif_ev_subsys),
            .ev_event(sif_ev_event), .ev_arg0(sif_ev_arg0),
            .ev_arg1(sif_ev_arg1), .ev_arg2(sif_ev_arg2),
            .ev_arg3(sif_ev_arg3), .ev_flags(sif_ev_flags));

    // ------------------------------------------------------------------
    // Counters
    // ------------------------------------------------------------------

    int   dma_beat_count;
    int   dma_done_count;
    logic dma_done_seen;
    int   errors;

    initial begin
        dma_beat_count = 0;
        dma_done_count = 0;
        dma_done_seen  = 1'b0;
        errors         = 0;
    end

    always_ff @(posedge clk) begin
        if (rst_n && dmac_ev_valid) begin
            if (dmac_ev_event == trace_pkg::EV_DMA_BEAT)
                dma_beat_count <= dma_beat_count + 1;
            if (dmac_ev_event == trace_pkg::EV_DMA_DONE) begin
                dma_done_count <= dma_done_count + 1;
                dma_done_seen  <= 1'b1;
            end
        end
    end

    // ------------------------------------------------------------------
    // Payload table
    // ------------------------------------------------------------------

    logic [127:0] expected_payload [0:QWC_VAL-1];

    initial begin
        expected_payload[0] = 128'h1111_0000_1111_0000_1111_0000_1111_0000;
        expected_payload[1] = 128'h2222_0000_2222_0000_2222_0000_2222_0000;
        expected_payload[2] = 128'h3333_0000_3333_0000_3333_0000_3333_0000;
        expected_payload[3] = 128'h4444_0000_4444_0000_4444_0000_4444_0000;
    end

    // ------------------------------------------------------------------
    // Helpers
    // ------------------------------------------------------------------

    task automatic write_dmac(input logic [7:0] offset, input logic [31:0] data);
        @(negedge clk);
        dmac_reg_wr_en   = 1'b1;
        dmac_reg_offset  = offset;
        dmac_reg_wr_data = data;
        @(negedge clk);
        dmac_reg_wr_en   = 1'b0;
        dmac_reg_offset  = 8'd0;
        dmac_reg_wr_data = 32'd0;
    endtask

    task automatic preload_ram_qword(input logic [RAM_ADDR_W-1:0] addr,
                                     input logic [127:0] data);
        @(negedge clk);
        ram_wr_en   = 1'b1;
        ram_wr_addr = addr;
        ram_wr_data = data;
        ram_wr_be   = 16'hFFFF;
        @(negedge clk);
        ram_wr_en   = 1'b0;
        ram_wr_addr = '0;
        ram_wr_data = 128'd0;
        ram_wr_be   = 16'd0;
    endtask

    task automatic sif_read_slot(input logic [SIF_IDX_W-1:0] idx,
                                 output logic [127:0] data);
        @(negedge clk);
        sif_rd_en  = 1'b1;
        sif_rd_idx = idx;
        @(negedge clk);
        sif_rd_en  = 1'b0;
        sif_rd_idx = '0;
        data = sif_rd_data;
    endtask

    // ------------------------------------------------------------------
    // Stimulus
    // ------------------------------------------------------------------

    int           settle_beats;
    int           settle_rx;
    logic         settle_full;
    logic         settle_done;
    int           beat_index_peek;
    logic [127:0] beat_payload_peek;
    logic [127:0] rb;

    initial begin
        rst_n            = 1'b0;
        dmac_reg_wr_en   = 1'b0;
        dmac_reg_offset  = 8'd0;
        dmac_reg_wr_data = 32'd0;
        ram_wr_en        = 1'b0;
        ram_wr_addr      = '0;
        ram_wr_data      = 128'd0;
        ram_wr_be        = 16'd0;
        sif_rd_en        = 1'b0;
        sif_rd_idx       = '0;
        sif_stall        = 1'b0;

        repeat (4) @(posedge clk);
        rst_n = 1'b1;
        repeat (2) @(posedge clk);

        begin : do_preload
            logic [31:0] a;
            for (int i = 0; i < QWC_VAL; i++) begin
                a = PAYLOAD_MADR + (i * 32'd16);
                preload_ram_qword(a[RAM_ADDR_W-1:0], expected_payload[i]);
            end
        end

        // Program DMAC for a QWC=4 transfer. With DEPTH=2, the receiver
        // should accept exactly 2 beats and then backpressure indefinitely.
        write_dmac(8'h10, PAYLOAD_MADR);
        write_dmac(8'h20, 32'd4);
        write_dmac(8'h00, 32'h0000_0001);

        // Let the DMAC exhaust buffer capacity and settle into stall.
        repeat (80) @(posedge clk);

        settle_beats = dma_beat_count;
        settle_rx    = sif_rx_count;
        settle_full  = sif_full;
        settle_done  = dma_done_seen;

        // Whitebox peeks into the DMAC to verify it is still trying to
        // emit beat 2 at MADR + 32, not skipping ahead.
        beat_index_peek    = u_dmac.beat_index;
        beat_payload_peek  = u_dmac.beat_payload;

        if (settle_beats !== 2) begin
            $error("[tb_sif_dma_overflow] expected exactly 2 beats accepted, got %0d",
                   settle_beats);
            errors = errors + 1;
        end
        if (settle_rx !== 32'd2) begin
            $error("[tb_sif_dma_overflow] rx_count=%0d, expected 2", settle_rx);
            errors = errors + 1;
        end
        if (!settle_full) begin
            $error("[tb_sif_dma_overflow] full_o never asserted");
            errors = errors + 1;
        end
        if (settle_done) begin
            $error("[tb_sif_dma_overflow] DMA_DONE fired while buffer full");
            errors = errors + 1;
        end
        if (beat_index_peek !== 2) begin
            $error("[tb_sif_dma_overflow] DMAC beat_index peek=%0d, expected 2 (stuck on beat 2)",
                   beat_index_peek);
            errors = errors + 1;
        end
        if (beat_payload_peek !== expected_payload[2]) begin
            $error("[tb_sif_dma_overflow] DMAC beat_payload peek mismatch");
            $error("  got:      0x%032h", beat_payload_peek);
            $error("  expected: 0x%032h (= payload[2], fetched from MADR+32)",
                   expected_payload[2]);
            errors = errors + 1;
        end

        // Spin another window — nothing should advance.
        repeat (40) @(posedge clk);
        if (dma_beat_count !== 2 || sif_rx_count !== 32'd2 || dma_done_seen) begin
            $error("[tb_sif_dma_overflow] state advanced during second stall window");
            errors = errors + 1;
        end

        // Verify the two stored qwords are intact.
        for (int i = 0; i < SIF_DEPTH; i++) begin
            sif_read_slot(SIF_IDX_W'(i), rb);
            if (rb !== expected_payload[i]) begin
                $error("[tb_sif_dma_overflow] slot[%0d]: got 0x%032h expected 0x%032h",
                       i, rb, expected_payload[i]);
                errors = errors + 1;
            end
        end

        repeat (4) @(posedge clk);

        // ------------------------------------------------------------------
        $display("[tb_sif_dma_overflow] settle_beats=%0d settle_rx=%0d full=%0b done=%0b beat_idx=%0d errors=%0d",
                 settle_beats, settle_rx, settle_full, settle_done,
                 beat_index_peek, errors);

        if (errors == 0 &&
            settle_beats == 2 && settle_rx === 32'd2 &&
            settle_full && !settle_done &&
            beat_index_peek == 2 &&
            beat_payload_peek === expected_payload[2])
            $display("[tb_sif_dma_overflow] PASS");
        else
            $display("[tb_sif_dma_overflow] FAIL");

        $finish;
    end

    initial begin
        #400000;
        $error("[tb_sif_dma_overflow] timeout");
        $finish;
    end

endmodule : tb_sif_dma_overflow