[PATCH] drm/udl: Fix for the X server screen update

[Noralf Trønnes]

Den 18.09.2016 15:48, skrev poma:
> Fix for DisplayLink GPU USB2.0 device X server screen update
>
> Within X server on top of DisplayLink GPU USB2.0 device,
> screen content is not refreshed i.e. updated,
> which is the most basic functionality of the screen.
>
> This partially (udl_handle_damage()) reverts commit:
>
> - e375882406d0cc24030746638592004755ed4ae0
>    "drm/udl: Use drm_fb_helper deferred_io support"
>
> Thanks Noralf for the tip.
>
> $ modinfo udl
> filename:       /lib/modules/4.7.4-234.fc24.x86_64/updates/udl.ko
> license:        GPL
> alias:          usb:v17E9p*d*dc*dsc*dp*icFFisc00ip00in*
> depends:
> intree:         Y
> vermagic:       4.7.4-234.fc24.x86_64 SMP mod_unload
> parm:           fb_bpp:int
> parm:           fb_defio:int
>
> $ dmesg | grep udl
> [   41.888469] udl: module verification failed: signature and/or required key missing - tainting kernel
> [   42.156988] udl 1-2:1.0: fb1: udldrmfb frame buffer device
> [   42.158940] [drm] Initialized udl on minor 1
> [   42.159676] usbcore: registered new interface driver udl
>
> $ grep udl /var/log/Xorg.0.log
> [    71.194] (**) |   |-->Device "udl0"
>
> Tested-by: poma <poma at gmail.com>
> ---
>   drivers/gpu/drm/udl/udl_drv.h |  2 ++
>   drivers/gpu/drm/udl/udl_fb.c  | 39 +++++++++++++++++++++++++++++++++++++--
>   2 files changed, 39 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/udl/udl_drv.h b/drivers/gpu/drm/udl/udl_drv.h
> index 0b03d34..4a064ef 100644
> --- a/drivers/gpu/drm/udl/udl_drv.h
> +++ b/drivers/gpu/drm/udl/udl_drv.h
> @@ -81,6 +81,8 @@ struct udl_framebuffer {
>   	struct drm_framebuffer base;
>   	struct udl_gem_object *obj;
>   	bool active_16; /* active on the 16-bit channel */
> +	int x1, y1, x2, y2; /* dirty rect */
> +	spinlock_t dirty_lock;
>   };
>   
>   #define to_udl_fb(x) container_of(x, struct udl_framebuffer, base)
> diff --git a/drivers/gpu/drm/udl/udl_fb.c b/drivers/gpu/drm/udl/udl_fb.c
> index d5df555..b2b42d2 100644
> --- a/drivers/gpu/drm/udl/udl_fb.c
> +++ b/drivers/gpu/drm/udl/udl_fb.c
> @@ -90,6 +90,9 @@ int udl_handle_damage(struct udl_framebuffer *fb, int x, int y,
>   	struct urb *urb;
>   	int aligned_x;
>   	int bpp = (fb->base.bits_per_pixel / 8);
> +	int x2, y2;
> +	bool store_for_later = false;
> +	unsigned long flags;
>   
>   	if (!fb->active_16)
>   		return 0;
> @@ -115,6 +118,38 @@ int udl_handle_damage(struct udl_framebuffer *fb, int x, int y,
>   	    (y + height > fb->base.height))
>   		return -EINVAL;
>   
> +	/* if we are in atomic just store the info
> +	   can't test inside spin lock */
> +	if (in_atomic())
> +		store_for_later = true;
> +
> +	x2 = x + width - 1;
> +	y2 = y + height - 1;
> +
> +	spin_lock_irqsave(&fb->dirty_lock, flags);

You can drop the spinlock and store_for_later since it always runs in
process context.
Which means that there's no need to store x1/x2/y1/y2 in fb.

> +
> +	if (fb->y1 < y)
> +		y = fb->y1;
> +	if (fb->y2 > y2)
> +		y2 = fb->y2;
> +	if (fb->x1 < x)
> +		x = fb->x1;
> +	if (fb->x2 > x2)
> +		x2 = fb->x2;
> +
> +	if (store_for_later) {
> +		fb->x1 = x;
> +		fb->x2 = x2;
> +		fb->y1 = y;
> +		fb->y2 = y2;
> +		spin_unlock_irqrestore(&fb->dirty_lock, flags);
> +		return 0;
> +	}
> +
> +	fb->x1 = fb->y1 = INT_MAX;
> +	fb->x2 = fb->y2 = 0;
> +
> +	spin_unlock_irqrestore(&fb->dirty_lock, flags);
>   	start_cycles = get_cycles();
>   
>   	urb = udl_get_urb(dev);
> @@ -122,14 +157,14 @@ int udl_handle_damage(struct udl_framebuffer *fb, int x, int y,
>   		return 0;
>   	cmd = urb->transfer_buffer;
>   
> -	for (i = y; i < height ; i++) {
> +	for (i = y; i <= y2 ; i++) {
>   		const int line_offset = fb->base.pitches[0] * i;
>   		const int byte_offset = line_offset + (x * bpp);
>   		const int dev_byte_offset = (fb->base.width * bpp * i) + (x * bpp);
>   		if (udl_render_hline(dev, bpp, &urb,
>   				     (char *) fb->obj->vmapping,
>   				     &cmd, byte_offset, dev_byte_offset,
> -				     width * bpp,

There is obviously something wrong with the use of height/width here
as a substitute for y2 and the x1 formula.
If you add a printk here you can see how they differ.

Noralf.

> +				     (x2 - x + 1) * bpp,
>   				     &bytes_identical, &bytes_sent))
>   			goto error;
>   	}