hdl: refactor RX/TX DMA request prioritization

* fixed TX deadlock when RX requests are not executed
* simplified RX and TX request prioritization
This commit is contained in:
Ryan Thompson 2025-02-21 11:55:53 -06:00 • committed by Robert Ghilduta
commit cc6f329bb8

View file

@ -179,8 +179,8 @@ architecture sample_shuffler of fx3_gpif is
signal current, future : fsm_t := FSM_RESET_VALUE;
signal can_rx, should_rx : boolean;
signal can_tx, should_tx : boolean;
signal can_rx : boolean;
signal can_tx : boolean;
signal rx_fifo_enough : boolean;
signal tx_fifo_enough : boolean;
signal rx_fifo_critical : boolean;
@ -192,14 +192,10 @@ architecture sample_shuffler of fx3_gpif is
attribute preserve : boolean;
attribute preserve of can_rx : signal is true;
attribute preserve of can_tx : signal is true;
attribute preserve of should_rx : signal is true;
attribute preserve of should_tx : signal is true;
attribute keep : boolean;
attribute keep of can_rx : signal is true;
attribute keep of can_tx : signal is true;
attribute keep of should_rx : signal is true;
attribute keep of should_tx : signal is true;
-- acknowledge(dma_channel) returns a dma_handshake_t with all bits 0
-- except the bit corresponding to dma_channel
@ -337,22 +333,28 @@ begin
end if;
end process gpif_mux;
-- Arbitrate RX and TX requests
-- In the event the FX3 simultaneously requests RX and TX, we generally
-- prioritize the TX first; this helps avoid discontinuities on the TX
-- signal. However, if RX is about to overflow, we'll choose that.
can_and_should : process(pclk, reset)
variable rx_dma_req : boolean;
variable tx_dma_req : boolean;
variable rx_meta_ok : boolean;
begin
if (reset = '1') then
can_rx <= false;
can_tx <= false;
should_rx <= false;
should_tx <= false;
elsif (rising_edge(pclk)) then
can_rx <= dma_rx_enable = '1' and rx_fifo_enough and (dma_req.rx0 = '1' or dma_req.rx1 = '1');
can_tx <= dma_tx_enable = '1' and tx_fifo_enough and (dma_req.tx2 = '1' or dma_req.tx3 = '1');
should_rx <= can_rx and (not can_tx or rx_fifo_critical);
should_tx <= can_tx and not rx_fifo_critical;
rx_dma_req := dma_req.rx0 = '1' or dma_req.rx1 = '1';
rx_meta_ok := meta_enable = '1' nand rx_meta_fifo_empty = '1';
can_rx <= dma_rx_enable = '1' and
rx_fifo_enough and
rx_dma_req and
rx_meta_ok;
tx_dma_req := dma_req.tx2 = '1' or dma_req.tx3 = '1';
can_tx <= dma_tx_enable = '1' and
tx_fifo_enough and
tx_dma_req;
end if;
end process;
@ -408,21 +410,24 @@ begin
future.fini_downcount <= FINI_DOWNCOUNT_RESET;
future.finishing_rx <= '0';
-- Arbitrate RX and TX requests
-- In the event the FX3 simultaneously requests RX and TX, we generally
-- prioritize the TX first; this helps avoid discontinuities on the TX
-- signal. However, if RX is about to overflow, we'll choose that.
if (current.dma_idle = '1') then
if (should_rx and ( (rx_meta_fifo_empty = '0' and current.rx_meta_en = '1')
or (current.rx_meta_en = '0') ) ) then
-- There is an RX to perform (sending data to FX3).
future.dma_downcount <= gpif_buf_size-1;
if (can_rx and rx_fifo_critical) then
future.ack_downcount <= ACK_DOWNCOUNT_READ;
future.dma_downcount <= gpif_buf_size-1;
future.rx_current_dma <= RX0;
future.state <= SETUP_RD;
elsif (should_tx) then
-- There is a TX to perform (getting data from FX3).
elsif (can_tx) then
future.ack_downcount <= ACK_DOWNCOUNT_WRITE;
future.dma_downcount <= gpif_buf_size-1;
future.tx_current_dma <= TX3;
future.state <= SETUP_WR;
elsif (can_rx) then
future.ack_downcount <= ACK_DOWNCOUNT_READ;
future.rx_current_dma <= RX0;
future.state <= SETUP_RD;
end if;
end if;