continue on ahci_fis_transmit.v

utils/ahci_dma.v

@@ -46,6 +46,7 @@ module  ahci_dma (
     input                  [15:0] prdtl,        // number of entries in PRD table (valid at cmd_start)
     input                         dev_wr,       // write to device (valid at start)
     input                         cmd_start,     // start processing command table, reset prdbc
+    input                         prd_start,     // at or after cmd_start - enable reading PRD/data (if any)
     input                         cmd_abort,     // try to abort a command TODO: Implement
 
 // Optional control of the AXI cache mode, default will be set to 4'h3, 4'h3 at mrst
@@ -145,9 +146,16 @@ module  ahci_dma (
     reg     [31:7] ctba_r;
     reg     [15:0] prdtl_mclk;
     wire           cmd_start_hclk;
+    reg            prd_start_r;
+    wire           prd_start_hclk;
+    reg            prd_start_hclk_r; // to make sure it is with/after prd_start_hclk if in mclk they are in the same cycle
     wire           cmd_abort_hclk; // TODO: Implement as graceful as possible command abort
+    reg            prd_enabled;
+    reg      [1:0] ct_over_prd_enabled; // prd read and data r/w enabled, command table fetch done
+    
     reg     [31:4] ct_maddr; // granularity matches PRDT entry - 4xDWORD, 2xQWORD
     wire           ct_done;
+    wire           first_prd_fetch; // CT read done, prd enabled
     reg     [31:0] afi_addr; // common for afi_araddr and afi_awaddr
     wire           axi_set_raddr_ready = !(|afi_racount[2:1]) && (!axi_set_raddr_r || !afi_racount[0]); // What is the size of ra fifo - just 4? Latency?
 //    wire           axi_set_raddr_ready = !(|afi_racount) && !axi_set_raddr_r); // Most pessimistic
@@ -198,7 +206,9 @@ module  ahci_dma (
     
     reg            data_next_burst;
     
-    wire           raddr_prd_rq = (|prds_left) && (ct_done || prd_done);
+//    wire           raddr_prd_rq = (|prds_left) && (ct_done || prd_done);
+    wire           raddr_prd_rq = (|prds_left) && (first_prd_fetch || prd_done);
+    
     reg            raddr_prd_pend;
             
     wire           raddr_ct_rq = cmd_start_hclk;
@@ -222,6 +232,7 @@ module  ahci_dma (
     assign prd_irq = data_irq && prd_done;
     assign cmd_done_hclk = ((ct_busy_r==2'b10) && (prdtl_mclk == 0)) || done_flush || done_dev_rd;
     assign ct_done = (ct_busy_r == 2'b10);
+    assign first_prd_fetch = ct_over_prd_enabled == 2'b01;
     assign axi_set_raddr_w = axi_set_raddr_ready && (raddr_ct_pend || raddr_prd_pend || raddr_data_pend);    
     assign axi_set_waddr_w = axi_set_raddr_ready && raddr_data_pend;    
     assign axi_set_addr_data_w = (axi_set_raddr_ready && raddr_data_pend) || (axi_set_waddr_ready && waddr_data_pend);
@@ -278,12 +289,21 @@ module  ahci_dma (
 
         if      (mrst)                  afi_awcache <= 4'h3;
         else if (set_axi_wr_cache_mode) afi_awcache <= axi_wr_cache_mode;
+        
+        prd_start_r <= prd_start;
 
     end
        
         
         
     always @ (posedge hclk) begin
+        prd_start_hclk_r <= prd_start_hclk;
+        
+        if      (hrst || cmd_abort_hclk) prd_enabled <= 0;
+        else if (prd_start_hclk_r)       prd_enabled <= 1; // presedence over  cmd_start_hclk
+        else if (cmd_start_hclk)         prd_enabled <= 0;
+    
+    
         if (cmd_start_hclk)  ct_maddr[31:4] <= {ctba_r[31:7],3'b0};
         else if (ct_done)    ct_maddr[31:4] <= ct_maddr[31:4] + 16;
         else if (wcount_set) ct_maddr[31:4] <= ct_maddr[31:4] + 1;
@@ -364,6 +384,10 @@ module  ahci_dma (
         else if (afi_rd_ctl[0] && is_ct_addr && (&int_data_addr)) ct_busy_r[0] <= 0;
         ct_busy_r[1] <= ct_busy_r[0]; // delayed version to detect end of command
         
+        if (hrst || ct_busy_r[0])                   ct_over_prd_enabled[0] <= 0;
+        else if (prd_enabled)                       ct_over_prd_enabled[0] <= 1;
+        ct_over_prd_enabled[1] <= ct_over_prd_enabled[0];  // detecting 0->1 transition
+        
         // generate busy for PRD table entry read
         if      (hrst)        prd_rd_busy <= 0;
         else if (prd_rd_busy) prd_rd_busy <= 1;
@@ -475,6 +499,18 @@ module  ahci_dma (
         .out_pulse (cmd_abort_hclk),    // output
         .busy()                       // output
     );
+    pulse_cross_clock #(
+        .EXTRA_DLY(0)
+    ) prd_start_hclk_i (
+        .rst       (mrst),            // input
+        .src_clk   (mclk),            // input
+        .dst_clk   (hclk),            // input
+        .in_pulse  (prd_start_r),     // input
+        .out_pulse (prd_start_hclk),  // output
+        .busy()                       // output
+    );
+
+
     
     // hclk -> mclk;
     pulse_cross_clock #(

utils/ahci_dma_rd_stuff.v

@@ -45,7 +45,7 @@ module  ahci_dma_rd_stuff(
     output reg        flushed,  // flush (end of last PRD is finished - data left module)
     output reg [31:0] dout,     // output 32-bit data
     output            dout_vld, // output data valid
-    input             dout_re   // consumer reads output data (should be anded with dout_vld)
+    input             dout_re   // consumer reads output data (should be AND-ed with dout_vld)
 );
     reg  [15:0] hr; // holds 16-bit data from previous din_re if not consumed
     reg         hr_full;

utils/ahci_fis_receive.v

@@ -49,11 +49,17 @@ module  ahci_fis_receive#(
     output reg                    fis_ok,        // FIS done,  checksum OK reset by starting a new get FIS
     output reg                    fis_err,       // FIS done, checksum ERROR reset by starting a new get FIS
     output                        fis_ferr,      // FIS done, fatal error - FIS too long
+    // next commands use register address/data/we for 1 clock cycle - after next to command (commnd - t0, we - t2)
     input                         update_err_sts,// update PxTFD.STS and PxTFD.ERR from the last received regs d2h
+    input                         update_prdbc,  // update PRDBC in registers
     input                         clear_bsy_drq, // clear PxTFD.STS.BSY and PxTFD.STS.DRQ, update
     input                         set_bsy,       // set PxTFD.STS.BSY, update
     input                         set_sts_7f,    // set PxTFD.STS = 0x7f, update
     input                         set_sts_80,    // set PxTFD.STS = 0x80 (may be combined with set_sts_7f), update
+    
+    input                         decr_dwc,      // decrement DMA Xfer counter // need pulse to 'update_prdbc' to write to registers
+    input                  [11:2] decr_DXC_dw,   // decrement value (in DWORDs)
+    
     // TODO: Add writing PRDBC here?
     
     output                  [7:0] tfd_sts,       // Current PxTFD status field (updated after regFIS and SDB - certain fields)
@@ -66,7 +72,9 @@ module  ahci_fis_receive#(
     output reg                    pio_i,         // value of "I" field in received PIO Setup FIS
     output reg                    pio_d,         // value of "D" field in received PIO Setup FIS
     output reg              [7:0] pio_es,        // value of PIO E_Status
-    output reg             [31:1] xfer_cntr,     // transfer counter in words for both DMA (31 bit) and PIO (lower 15 bits)
+    // Using even word count (will be rounded up), partial DWORD (last) will be handled by PRD length if needed
+    output                 [31:2] xfer_cntr,     // transfer counter in words for both DMA (31 bit) and PIO (lower 15 bits), updated after decr_dwc
+    output reg                    xfer_cntr_zero,// valid next cycle                   
 
 // Registers interface
 // 2. HBA R/W registers, may be added external register layer
@@ -96,6 +104,7 @@ FB_OFFS =  0xc00 # Needs 0x100 bytes
 
 
 */
+localparam CLB_OFFS32 =        'h200; //  # In the second half of the register space (0x800..0xbff - 1KB)
 localparam HBA_OFFS32 =         0;
 localparam HBA_PORT0_OFFS32  = 'h40;
 localparam PXSIG_OFFS32 = HBA_OFFS32 + HBA_PORT0_OFFS32 + 'h9; 
@@ -153,10 +162,13 @@ localparam DATA_TYPE_ERR =      3;
     reg           [4:0] reg_ps;
     reg                 reg_d2h;    
     reg                 reg_sdb;    
+    reg          [31:2] xfer_cntr_r;
+    reg          [31:2] prdbc_r;
     
 
     reg          [15:0] tf_err_sts;
     reg                 update_err_sts_r;
+    reg                 update_prdbc_r;
     
     // Forward data to DMA (dev->mem) engine
     assign              dma_in_valid = dma_in_ready && (hda_data_in_type == DATA_TYPE_DMA) && data_in_ready && !too_long_err;
@@ -172,6 +184,9 @@ localparam DATA_TYPE_ERR =      3;
     
     assign tfd_sts = tf_err_sts[ 7:0];
     assign tfd_err = tf_err_sts[15:8];
+    
+    assign xfer_cntr = xfer_cntr_r[31:2];
+    
      
     always @ (posedge mclk) begin
         if (hba_rst || dma_in_stop) dma_in <= 0;
@@ -248,20 +263,23 @@ localparam DATA_TYPE_ERR =      3;
         if (reg_we_w)                         reg_data[31:8] <= hda_data_in[31:8];
         else if (update_sig[1])               reg_data[31:8] <= hda_data_in[23:0];
         else if (update_err_sts_r)            reg_data[31:8] <= {16'b0,tf_err_sts[15:8]};
+        else if (update_prdbc_r)              reg_data[31:8] <= {xfer_cntr_r[31:8]};
 
         if (reg_we_w)                         reg_data[ 7:0] <=  hda_data_in[ 7:0];
         else if (update_sig[3])               reg_data[ 7:0] <=  hda_data_in[ 7:0];
         else if (update_err_sts_r)            reg_data[ 7:0] <=  tf_err_sts [ 7:0];
+        else if (update_prdbc_r)              reg_data[ 7:0] <=  {xfer_cntr_r[ 7:2],2'b0};
         
         if (reg_d2h || update_sig[0])         tf_err_sts  <= hda_data_in[15:0];
         else if (reg_sdb)                     tf_err_sts  <= {hda_data_in[15:8], tf_err_sts[7], hda_data_in[6:4], tf_err_sts[3],hda_data_in[2:0]};
         else if (clear_bsy_drq || set_bsy)    tf_err_sts  <= tf_err_sts & {8'hff,clear_bsy_drq,3'h7,clear_bsy_drq,3'h7} | {8'h0,set_bsy,7'h0};
         else if (set_sts_7f || set_sts_80)    tf_err_sts  <= {tf_err_sts[15:8],set_sts_80,{7{set_sts_7f}}} ;
         
-        reg_we <= reg_we_w || update_sig[3] || update_err_sts_r;
+        reg_we <= reg_we_w || update_sig[3] || update_err_sts_r || update_prdbc_r;
         
         if (reg_we_w || update_sig[3])        reg_addr <=  reg_addr_r;
         else if (update_err_sts_r)            reg_addr <=  PXTFD_OFFS32;
+        else if (update_prdbc_r)              reg_addr <=  CLB_OFFS32 + 1; // location of PRDBC
 
         if (reg_d2h || reg_sdb || reg_ds[0])  fis_i <=           hda_data_in[14];
         if (reg_sdb)                          sdb_n <=           hda_data_in[15];
@@ -272,9 +290,19 @@ localparam DATA_TYPE_ERR =      3;
         if (hba_rst)                          pio_es  <=         0;
         else if (reg_ps[3])                   pio_es  <=         hda_data_in[31:24];
         
-        if (reg_ps[4] || reg_ds[5])           xfer_cntr[31:1] <= {reg_ds[5]?hda_data_in[31:16]:16'b0,hda_data_in[15:1]};
+        if (hba_rst || reg_sdb)               xfer_cntr_r[31:2] <= 0;
+        else if (reg_ps[4] || reg_ds[5])      xfer_cntr_r[31:2] <= {reg_ds[5]?hda_data_in[31:16]:16'b0, hda_data_in[15:2]} + hda_data_in[1]; // round up
+        else if (decr_dwc)                    xfer_cntr_r[31:2] <= {xfer_cntr_r[31:2]} - {20'b0, decr_DXC_dw[11:2]};
+        
+        if (hba_rst || reg_sdb || reg_ps[4] || reg_ds[5])  prdbc_r[31:2] <= 0;
+        else if (decr_dwc)                                 prdbc_r[31:2] <= {prdbc_r[31:2]} + {20'b0, decr_DXC_dw[11:2]};
+        
+        
+        
+        xfer_cntr_zero <=                     xfer_cntr_r[31:2] == 0;
         
         update_err_sts_r <= update_err_sts || clear_bsy_drq || set_bsy || set_sts_7f || set_sts_80;
+        update_prdbc_r <= update_prdbc; // same latency as update_err_sts
         
     end