Re: [PATCH] drm/mediatek: dp: Wait for the complete AUX reply before reading it

From: CK Hu (胡俊光)

Date: Tue Nov 17 2026 - 22:06:38 EST


On Thu, 2026-11-12 at 14:17 +0800, Yidi Lin wrote:
> mtk_dp_aux_wait_for_completion() returns for a read as soon as the RX
> FIFO write pointer is non-zero, i.e. when the first byte of the reply
> has arrived. The rest of the reply is still on the AUX line, so
> mtk_dp_aux_read_rx_fifo() sleeps 500-1000 us before reading each byte
> of an I2C-over-AUX reply and 100-200 us before each byte of a native
> reply. The FIFO is drained this way even for DEFER replies, which carry
> no data.
>
> A 16-byte EDID read therefore takes about 16 ms, although the reply is
> complete within about 300 us of the request.
>
> Wait for the AUX receive-complete (or 400 us timeout) interrupt status
> instead, which the write path already relies on, and drain the RX FIFO
> without delays. Poll every 20 us instead of every 100-500 us.
>
> Reading the 256-byte EDID through the DP AUX DDC adapter drops from 266
> to 6.5 ms on an MT8189 Chromebook, and from 547 to 23 ms on an MT8196
> Chromebook whose panel defers every 16-byte read once. A 16-byte DPCD
> read drops from about 4.7 to 0.5 ms.
>
> Signed-off-by: Yidi Lin <yidilin@xxxxxxxxxx>
> ---
> drivers/gpu/drm/mediatek/mtk_dp.c | 64 ++++++++++++-------------------
> 1 file changed, 25 insertions(+), 39 deletions(-)
>
> diff --git a/drivers/gpu/drm/mediatek/mtk_dp.c b/drivers/gpu/drm/mediatek/mtk_dp.c
> index f656b85b84212..b2f7f42cbe715 100644
> --- a/drivers/gpu/drm/mediatek/mtk_dp.c
> +++ b/drivers/gpu/drm/mediatek/mtk_dp.c
> @@ -45,7 +45,8 @@
> #define MTK_DP_4P1T 4
> #define MTK_DP_HDE 2
> #define MTK_DP_PIX_PER_ADDR 2
> -#define MTK_DP_AUX_WAIT_REPLY_COUNT 20
> +#define MTK_DP_AUX_POLL_US 20
> +#define MTK_DP_AUX_TIMEOUT_US 10000
> #define MTK_DP_TBC_BUF_READ_START_ADDR 0x8
> #define MTK_DP_TRAIN_VOLTAGE_LEVEL_RETRY 5
> #define MTK_DP_TRAIN_DOWNSCALE_RETRY 10
> @@ -825,20 +826,18 @@ static void mtk_dp_aux_fill_write_fifo(struct mtk_dp *mtk_dp, u8 *buf,
> }
>
> static void mtk_dp_aux_read_rx_fifo(struct mtk_dp *mtk_dp, u8 *buf,
> - size_t length, int read_delay)
> + size_t length)
> {
> int read_pos;
>
> mtk_dp_update_bits(mtk_dp, MTK_DP_AUX_P0_3620,
> 0, AUX_RD_MODE_AUX_TX_P0_MASK);
>
> + /* The complete reply is in the RX FIFO, so read it without delays. */
> for (read_pos = 0; read_pos < length; read_pos++) {
> mtk_dp_update_bits(mtk_dp, MTK_DP_AUX_P0_3620,
> AUX_RX_FIFO_READ_PULSE_TX_P0,
> AUX_RX_FIFO_READ_PULSE_TX_P0);
> -
> - /* Hardware needs time to update the data */

We assume the delay is just for AUX_RX_AUX_RECV_COMPLETE_IRQ_AUX_TX_P0 and no the other hardware delay.
It seems this assumption is right, so

Reviewed-by: CK Hu <ck.hu@xxxxxxxxxxxx>

> - usleep_range(read_delay, read_delay * 2);
> buf[read_pos] = (u8)(mtk_dp_read(mtk_dp, MTK_DP_AUX_P0_3620) &
> AUX_RX_FIFO_READ_DATA_AUX_TX_P0_MASK);
> }
> @@ -864,33 +863,28 @@ static void mtk_dp_aux_set_length(struct mtk_dp *mtk_dp, size_t length)
> }
> }
>
> -static int mtk_dp_aux_wait_for_completion(struct mtk_dp *mtk_dp, bool is_read)
> +static int mtk_dp_aux_wait_for_completion(struct mtk_dp *mtk_dp)
> {
> - int wait_reply = MTK_DP_AUX_WAIT_REPLY_COUNT;
> -
> - while (--wait_reply) {
> - u32 aux_irq_status;
> -
> - if (is_read) {
> - u32 fifo_status = mtk_dp_read(mtk_dp, MTK_DP_AUX_P0_3618);
> -
> - if (fifo_status &
> - (AUX_RX_FIFO_WRITE_POINTER_AUX_TX_P0_MASK |
> - AUX_RX_FIFO_FULL_AUX_TX_P0_MASK)) {
> - return 0;
> - }
> - }
> -
> - aux_irq_status = mtk_dp_read(mtk_dp, MTK_DP_AUX_P0_3640);
> - if (aux_irq_status & AUX_RX_AUX_RECV_COMPLETE_IRQ_AUX_TX_P0)
> - return 0;
> + u32 aux_irq_status;
> + int ret;
>
> - if (aux_irq_status & AUX_400US_TIMEOUT_IRQ_AUX_TX_P0)
> - return -ETIMEDOUT;
> + /*
> + * Wait until the whole reply has been received. The RX FIFO write
> + * pointer becomes non-zero as soon as the first byte of the reply
> + * arrives, while the rest of the reply is still on the AUX line, so
> + * it cannot be used to tell that the reply is complete.
> + */
> + ret = regmap_read_poll_timeout(mtk_dp->regs, MTK_DP_AUX_P0_3640,
> + aux_irq_status,
> + aux_irq_status &
> + (AUX_RX_AUX_RECV_COMPLETE_IRQ_AUX_TX_P0 |
> + AUX_400US_TIMEOUT_IRQ_AUX_TX_P0),
> + MTK_DP_AUX_POLL_US, MTK_DP_AUX_TIMEOUT_US);
> + if (ret)
> + return ret;
>
> - /* Give the hardware a chance to reach completion before retrying */
> - usleep_range(100, 500);
> - }
> + if (aux_irq_status & AUX_RX_AUX_RECV_COMPLETE_IRQ_AUX_TX_P0)
> + return 0;
>
> return -ETIMEDOUT;
> }
> @@ -928,7 +922,7 @@ static int mtk_dp_aux_do_transfer(struct mtk_dp *mtk_dp, bool is_read, u8 cmd,
> mtk_dp_aux_request_ready(mtk_dp);
>
> /* Wait for feedback from sink device. */
> - ret = mtk_dp_aux_wait_for_completion(mtk_dp, is_read);
> + ret = mtk_dp_aux_wait_for_completion(mtk_dp);
>
> *reply_cmd = mtk_dp_read(mtk_dp, MTK_DP_AUX_P0_3624) &
> AUX_RX_REPLY_COMMAND_AUX_TX_P0_MASK;
> @@ -952,15 +946,7 @@ static int mtk_dp_aux_do_transfer(struct mtk_dp *mtk_dp, bool is_read, u8 cmd,
> AUX_TX_AUXTX_OV_EN_AUX_TX_P0_MASK |
> AUX_RESERVED_RW_0_AUX_TX_P0_MASK);
> } else if (is_read) {
> - int read_delay;
> -
> - if (cmd == (DP_AUX_I2C_READ | DP_AUX_I2C_MOT) ||
> - cmd == DP_AUX_I2C_READ)
> - read_delay = 500;
> - else
> - read_delay = 100;
> -
> - mtk_dp_aux_read_rx_fifo(mtk_dp, buf, length, read_delay);
> + mtk_dp_aux_read_rx_fifo(mtk_dp, buf, length);
> }
>
> return 0;