Re: [v3 6/7] drm/mediatek: change the dsi phytiming calculate method

[Guenter Roeck]

On Sun, May 19, 2019 at 05:25:36PM +0800, Jitao Shi wrote:
> Change the method of frame rate calc which can get more accurate
> frame rate.
> 
> data rate = pixel_clock * bit_per_pixel / lanes
> Adjust hfp_wc to adapt the additional phy_data
> 
> if MIPI_DSI_MODE_VIDEO_BURST
> 	hfp_wc = hfp * bpp - data_phy_cycles * lanes - 12 - 6;
> else
> 	hfp_wc = hfp * bpp - data_phy_cycles * lanes - 12;
> 
> Note:
> //(2: 1 for sync, 1 for phy idle)
> data_phy_cycles = T_hs_exit + T_lpx + T_hs_prepare + T_hs_zero + 2;
> 
> bpp: bit per pixel
> 
> Signed-off-by: Jitao Shi <jitao.shi@xxxxxxxxxxxx>
> Tested-by: Ryan Case <ryandcase@xxxxxxxxxxxx>
> ---
>  drivers/gpu/drm/mediatek/mtk_dsi.c | 119 +++++++++++++++++++++--------
>  1 file changed, 86 insertions(+), 33 deletions(-)
> 
> diff --git a/drivers/gpu/drm/mediatek/mtk_dsi.c b/drivers/gpu/drm/mediatek/mtk_dsi.c
> index 1165ff944889..3f51b2000c68 100644
> --- a/drivers/gpu/drm/mediatek/mtk_dsi.c
> +++ b/drivers/gpu/drm/mediatek/mtk_dsi.c
> @@ -158,6 +158,25 @@
>  	(type == MIPI_DSI_GENERIC_READ_REQUEST_2_PARAM) || \
>  	(type == MIPI_DSI_DCS_READ))
>  
> +struct mtk_phy_timing {
> +	u32 lpx;
> +	u32 da_hs_prepare;
> +	u32 da_hs_zero;
> +	u32 da_hs_trail;
> +
> +	u32 ta_go;
> +	u32 ta_sure;
> +	u32 ta_get;
> +	u32 da_hs_exit;
> +
> +	u32 clk_hs_zero;
> +	u32 clk_hs_trail;
> +
> +	u32 clk_hs_prepare;
> +	u32 clk_hs_post;
> +	u32 clk_hs_exit;
> +};
> +
>  struct phy;
>  
>  struct mtk_dsi_driver_data {
> @@ -182,12 +201,13 @@ struct mtk_dsi {
>  	struct clk *digital_clk;
>  	struct clk *hs_clk;
>  
> -	u32 data_rate;
> +	u64 data_rate;

This results in 64-bit divide operations and thus build failures
with 32-bit builds. More on that below.

>  
>  	unsigned long mode_flags;
>  	enum mipi_dsi_pixel_format format;
>  	unsigned int lanes;
>  	struct videomode vm;
> +	struct mtk_phy_timing phy_timing;
>  	int refcount;
>  	bool enabled;
>  	u32 irq_data;
> @@ -221,17 +241,39 @@ static void mtk_dsi_phy_timconfig(struct mtk_dsi *dsi)
>  {
>  	u32 timcon0, timcon1, timcon2, timcon3;
>  	u32 ui, cycle_time;
> +	struct mtk_phy_timing *timing = &dsi->phy_timing;
> +
> +	ui = 1000000000 / dsi->data_rate;
> +	cycle_time = 8000000000 / dsi->data_rate;

This results in 64-bit divide operations. On top of that, 8000000000
is larger than 0xffffffff, resulting in an integer overflow on 32-bit
systems; it should be provided as 8000000000ULL. 

> +
> +	timing->lpx = NS_TO_CYCLE(60, cycle_time);
> +	timing->da_hs_prepare = NS_TO_CYCLE((40 + 5 * ui), cycle_time);
> +	timing->da_hs_zero = NS_TO_CYCLE((110 + 6 * ui), cycle_time);
> +	timing->da_hs_trail = NS_TO_CYCLE(((0x4 * ui) + 80), cycle_time);
> +
> +	if (timing->da_hs_zero > timing->da_hs_prepare)
> +		timing->da_hs_zero -= timing->da_hs_prepare;
> +
> +	timing->ta_go = 4 * timing->lpx;
> +	timing->ta_sure = 3 * timing->lpx / 2;
> +	timing->ta_get = 5 * timing->lpx;
> +	timing->da_hs_exit = 2 * timing->lpx;
> +
> +	timing->clk_hs_zero = NS_TO_CYCLE(0x150, cycle_time);
> +	timing->clk_hs_trail = NS_TO_CYCLE(0x64, cycle_time) + 0xa;
>  
> -	ui = 1000 / dsi->data_rate + 0x01;
> -	cycle_time = 8000 / dsi->data_rate + 0x01;
> +	timing->clk_hs_prepare = NS_TO_CYCLE(0x40, cycle_time);
> +	timing->clk_hs_post = NS_TO_CYCLE(80 + 52 * ui, cycle_time);
> +	timing->clk_hs_exit = 2 * timing->lpx;
>  
> -	timcon0 = T_LPX | T_HS_PREP << 8 | T_HS_ZERO << 16 | T_HS_TRAIL << 24;
> -	timcon1 = 4 * T_LPX | (3 * T_LPX / 2) << 8 | 5 * T_LPX << 16 |
> -		  T_HS_EXIT << 24;
> -	timcon2 = ((NS_TO_CYCLE(0x64, cycle_time) + 0xa) << 24) |
> -		  (NS_TO_CYCLE(0x150, cycle_time) << 16);
> -	timcon3 = NS_TO_CYCLE(0x40, cycle_time) | (2 * T_LPX) << 16 |
> -		  NS_TO_CYCLE(80 + 52 * ui, cycle_time) << 8;
> +	timcon0 = timing->lpx | timing->da_hs_prepare << 8 |
> +		  timing->da_hs_zero << 16 | timing->da_hs_trail << 24;
> +	timcon1 = timing->ta_go | timing->ta_sure << 8 |
> +		  timing->ta_get << 16 | timing->da_hs_exit << 24;
> +	timcon2 = 1 << 8 | timing->clk_hs_zero << 16 |
> +		  timing->clk_hs_trail << 24;
> +	timcon3 = timing->clk_hs_prepare | timing->clk_hs_post << 8 |
> +		  timing->clk_hs_exit << 16;
>  
>  	writel(timcon0, dsi->regs + DSI_PHY_TIMECON0);
>  	writel(timcon1, dsi->regs + DSI_PHY_TIMECON1);
> @@ -418,7 +460,8 @@ static void mtk_dsi_config_vdo_timing(struct mtk_dsi *dsi)
>  	u32 horizontal_sync_active_byte;
>  	u32 horizontal_backporch_byte;
>  	u32 horizontal_frontporch_byte;
> -	u32 dsi_tmp_buf_bpp;
> +	u32 dsi_tmp_buf_bpp, data_phy_cycles;
> +	struct mtk_phy_timing *timing = &dsi->phy_timing;
>  
>  	struct videomode *vm = &dsi->vm;
>  
> @@ -433,7 +476,8 @@ static void mtk_dsi_config_vdo_timing(struct mtk_dsi *dsi)
>  	writel(vm->vactive, dsi->regs + DSI_VACT_NL);
>  
>  	if (dsi->driver_data->has_size_ctl)
> -		writel(vm->vactive << 16 | vm->hactive, dsi->regs + DSI_SIZE_CON);
> +		writel(vm->vactive << 16 | vm->hactive,
> +		       dsi->regs + DSI_SIZE_CON);
>  
>  	horizontal_sync_active_byte = (vm->hsync_len * dsi_tmp_buf_bpp - 10);
>  
> @@ -444,7 +488,34 @@ static void mtk_dsi_config_vdo_timing(struct mtk_dsi *dsi)
>  		horizontal_backporch_byte = ((vm->hback_porch + vm->hsync_len) *
>  			dsi_tmp_buf_bpp - 10);
>  
> -	horizontal_frontporch_byte = (vm->hfront_porch * dsi_tmp_buf_bpp - 12);
> +	data_phy_cycles = timing->lpx + timing->da_hs_prepare +
> +				  timing->da_hs_zero + timing->da_hs_exit + 2;
> +
> +	if (dsi->mode_flags & MIPI_DSI_MODE_VIDEO_BURST) {
> +		if (vm->hfront_porch * dsi_tmp_buf_bpp >
> +		    data_phy_cycles * dsi->lanes + 18) {
> +			horizontal_frontporch_byte = vm->hfront_porch *
> +						     dsi_tmp_buf_bpp -
> +						     data_phy_cycles *
> +						     dsi->lanes - 18;
> +		} else {
> +			DRM_WARN("HFP less than d-phy, FPS will under 60Hz\n");
> +			horizontal_frontporch_byte = vm->hfront_porch *
> +						     dsi_tmp_buf_bpp;
> +		}
> +	} else {
> +		if (vm->hfront_porch * dsi_tmp_buf_bpp >
> +		    data_phy_cycles * dsi->lanes + 12) {
> +			horizontal_frontporch_byte = vm->hfront_porch *
> +						     dsi_tmp_buf_bpp -
> +						     data_phy_cycles *
> +						     dsi->lanes - 12;
> +		} else {
> +			DRM_WARN("HFP less than d-phy, FPS will under 60Hz\n");
> +			horizontal_frontporch_byte = vm->hfront_porch *
> +						     dsi_tmp_buf_bpp;
> +		}
> +	}
>  
>  	writel(horizontal_sync_active_byte, dsi->regs + DSI_HSA_WC);
>  	writel(horizontal_backporch_byte, dsi->regs + DSI_HBP_WC);
> @@ -544,8 +615,7 @@ static int mtk_dsi_poweron(struct mtk_dsi *dsi)
>  {
>  	struct device *dev = dsi->dev;
>  	int ret;
> -	u64 pixel_clock, total_bits;
> -	u32 htotal, htotal_bits, bit_per_pixel, overhead_cycles, overhead_bits;
> +	u32 bit_per_pixel;
>  
>  	if (++dsi->refcount != 1)
>  		return 0;
> @@ -564,24 +634,7 @@ static int mtk_dsi_poweron(struct mtk_dsi *dsi)
>  		break;
>  	}
>  
> -	/**
> -	 * htotal_time = htotal * byte_per_pixel / num_lanes
> -	 * overhead_time = lpx + hs_prepare + hs_zero + hs_trail + hs_exit
> -	 * mipi_ratio = (htotal_time + overhead_time) / htotal_time
> -	 * data_rate = pixel_clock * bit_per_pixel * mipi_ratio / num_lanes;
> -	 */
> -	pixel_clock = dsi->vm.pixelclock;
> -	htotal = dsi->vm.hactive + dsi->vm.hback_porch + dsi->vm.hfront_porch +
> -			dsi->vm.hsync_len;
> -	htotal_bits = htotal * bit_per_pixel;
> -
> -	overhead_cycles = T_LPX + T_HS_PREP + T_HS_ZERO + T_HS_TRAIL +
> -			T_HS_EXIT;
> -	overhead_bits = overhead_cycles * dsi->lanes * 8;
> -	total_bits = htotal_bits + overhead_bits;
> -
> -	dsi->data_rate = DIV_ROUND_UP_ULL(pixel_clock * total_bits,
> -					  htotal * dsi->lanes);
> +	dsi->data_rate = dsi->vm.pixelclock * bit_per_pixel / dsi->lanes;

pixelclock and bit_per_pixel are not u64, and neither is dsi->lanes.
The above will thus be a 32-bit operation on 32-bit systems, and never
really assign a true 64-bit value to data_rate. On top of that,
clk_set_rate() expects an unsigned long argument. Declaring data_rate
as anything but unsigned long has therefore no value.

Note that the old code took possible overflow conditions due to interim
results larger than u32 / unsigned long into account. This is no longer
the case. I don't know if this is a concern, but it might be worth watching
out for. If it is not a concern because the code is not anymore expected
to run on 32-bit systems, it should be restricted to 64-bit builds.

Guenter

>  
>  	ret = clk_set_rate(dsi->hs_clk, dsi->data_rate);
>  	if (ret < 0) {