diff mbox

drm/exynos: use 4WORD dma burst length for small fbs

Message ID 1399461922-27131-1-git-send-email-rahul.sharma@samsung.com (mailing list archive)
State Not Applicable, archived
Headers show

Commit Message

Rahul Sharma May 7, 2014, 11:25 a.m. UTC
From: Rahul Sharma <Rahul.Sharma@samsung.com>

In case of exynos, setting dma-burst to 16Word causes permanent
tearing for very small buffers, e.g. cursor buffer. Burst Mode
switching, which is based on overlay size is not recommended as
overlay size varies a lot towards the end of the screen. This
causes unstable DMA which results into tearing again.

Rendering small buffers with lower burst size doesn't
cause any noticable performance overhead. 128 pixel width is
selected based on mulitple experiments with exynos5 SoCs.

Signed-off-by: Rahul Sharma <Rahul.Sharma@samsung.com>
Signed-off-by: Prathyush K <prathyush.k@samsung.com>
---
Based on Inki Dae's exynos-drm-next branch.

 drivers/gpu/drm/exynos/exynos_drm_fimd.c |   14 ++++++++++++++
 1 file changed, 14 insertions(+)

Comments

Inki Dae May 21, 2014, 11:13 a.m. UTC | #1
Hi Rahul,

On 2014? 05? 07? 20:25, Rahul Sharma wrote:
> From: Rahul Sharma <Rahul.Sharma@samsung.com>
> 
> In case of exynos, setting dma-burst to 16Word causes permanent
> tearing for very small buffers, e.g. cursor buffer. Burst Mode
> switching, which is based on overlay size is not recommended as
> overlay size varies a lot towards the end of the screen. This
> causes unstable DMA which results into tearing again.
> 
> Rendering small buffers with lower burst size doesn't
> cause any noticable performance overhead. 128 pixel width is
> selected based on mulitple experiments with exynos5 SoCs.
> 
> Signed-off-by: Rahul Sharma <Rahul.Sharma@samsung.com>
> Signed-off-by: Prathyush K <prathyush.k@samsung.com>
> ---
> Based on Inki Dae's exynos-drm-next branch.
> 
>  drivers/gpu/drm/exynos/exynos_drm_fimd.c |   14 ++++++++++++++
>  1 file changed, 14 insertions(+)
> 
> diff --git a/drivers/gpu/drm/exynos/exynos_drm_fimd.c b/drivers/gpu/drm/exynos/exynos_drm_fimd.c
> index 40fd6cc..ef56077 100644
> --- a/drivers/gpu/drm/exynos/exynos_drm_fimd.c
> +++ b/drivers/gpu/drm/exynos/exynos_drm_fimd.c
> @@ -38,6 +38,7 @@
>   */
>  
>  #define FIMD_DEFAULT_FRAMERATE 60
> +#define MIN_FB_WIDTH_FOR_16WORD_BURST 128
>  
>  /* position control register for hardware window 0, 2 ~ 4.*/
>  #define VIDOSD_A(win)		(VIDOSD_BASE + 0x00 + (win) * 16)
> @@ -446,6 +447,19 @@ static void fimd_win_set_pixfmt(struct fimd_context *ctx, unsigned int win)
>  
>  	DRM_DEBUG_KMS("bpp = %d\n", win_data->bpp);
>  
> +	/*
> +	 * In case of exynos, setting dma-burst to 16Word causes permanent
> +	 * tearing for very small buffers, e.g. cursor buffer. Burst Mode
> +	 * switching which is based on overlay size is not recommended as
> +	 * overlay size varies alot towards the end of the screen and rapid
> +	 * movement causes unstable DMA which results into iommu crash/tear.

FIMD has width limitation so width of hardware overlay may need to be
aligned to 16 pixels.
We had faced with similar issue and the issue had been resolved by
aligning it to 16 pixels.

So can you try to align it instead of changing burst len size?

Thanks,
Inki Dae

> +	 */
> +
> +	if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
> +		val &= ~WINCONx_BURSTLEN_MASK;
> +		val |= WINCONx_BURSTLEN_4WORD;
> +	}
> +
>  	writel(val, ctx->regs + WINCON(win));
>  }
>  
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Rahul Sharma May 22, 2014, 4:36 a.m. UTC | #2
Hi Inki,

On 21 May 2014 16:43, Inki Dae <inki.dae@samsung.com> wrote:
>
> Hi Rahul,
>
> On 2014? 05? 07? 20:25, Rahul Sharma wrote:
>> From: Rahul Sharma <Rahul.Sharma@samsung.com>
>>
>> In case of exynos, setting dma-burst to 16Word causes permanent
>> tearing for very small buffers, e.g. cursor buffer. Burst Mode
>> switching, which is based on overlay size is not recommended as
>> overlay size varies a lot towards the end of the screen. This
>> causes unstable DMA which results into tearing again.
[snip]
>> +     /*
>> +      * In case of exynos, setting dma-burst to 16Word causes permanent
>> +      * tearing for very small buffers, e.g. cursor buffer. Burst Mode
>> +      * switching which is based on overlay size is not recommended as
>> +      * overlay size varies alot towards the end of the screen and rapid
>> +      * movement causes unstable DMA which results into iommu crash/tear.
>
> FIMD has width limitation so width of hardware overlay may need to be
> aligned to 16 pixels.
> We had faced with similar issue and the issue had been resolved by
> aligning it to 16 pixels.
>
> So can you try to align it instead of changing burst len size?
>

This problem is only with the very small FBs which are
rendered towards corners. For large FBs like 1366x768,
we dont see this issue though 1366 is not 16 pixel
aligned.

But I still carried out following experiments just to
verify the fimd behaviour for small overlay.

1) Round DOWN:
+  overlay->crtc_width = ((overlay->crtc_width)/16)*16;

Cursor corruption from A to B
A: [  308.530000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
[  308.530000] ovl_width = 64, ovl_height = 64
[  308.540000] offset_x = 1302, offset_y = 335
[  308.540000] fb_width = 96, fb_height = 64

B: [  341.890000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
[  341.890000] ovl_width = 48, ovl_height = 64
[  341.900000] offset_x = 1303, offset_y = 335
[  341.900000] fb_width = 96, fb_height = 64


2) Round UP:

+ overlay->crtc_width = ((overlay->crtc_width + 16)/16)*16;
+ if(overlay->crtc_width > overlay->fb_width)
+         overlay->crtc_width = overlay->fb_width;

Cursor corruption from A to B
A: [   63.200000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
[   63.210000] ovl_width = 96, ovl_height = 64
[   63.210000] offset_x = 1271, offset_y = 346
[   63.220000] fb_width = 96, fb_height = 64

B: [   68.600000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
[   68.610000] ovl_width = 96, ovl_height = 64
[   68.610000] offset_x = 1272, offset_y = 346
[   68.620000] fb_width = 96, fb_height = 64

In both the scenarios ovl_width is always 16 Pixel aligned.
Please let me know if you have something else in mind. I
will check and share the observations.

The solution in the patch is proposed by hardware team
and work fine for exynos 5250, 5420 and 5800.

Regards,
Rahul Sharma

> Thanks,
> Inki Dae
>
>> +      */
>> +
>> +     if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
>> +             val &= ~WINCONx_BURSTLEN_MASK;
>> +             val |= WINCONx_BURSTLEN_4WORD;
>> +     }
>> +
>>       writel(val, ctx->regs + WINCON(win));
>>  }
>>
>>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Inki Dae May 22, 2014, 6:55 a.m. UTC | #3
On 2014? 05? 22? 13:36, Rahul Sharma wrote:
> Hi Inki,
> 
> On 21 May 2014 16:43, Inki Dae <inki.dae@samsung.com> wrote:
>>
>> Hi Rahul,
>>
>> On 2014? 05? 07? 20:25, Rahul Sharma wrote:
>>> From: Rahul Sharma <Rahul.Sharma@samsung.com>
>>>
>>> In case of exynos, setting dma-burst to 16Word causes permanent
>>> tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>> switching, which is based on overlay size is not recommended as
>>> overlay size varies a lot towards the end of the screen. This
>>> causes unstable DMA which results into tearing again.
> [snip]
>>> +     /*
>>> +      * In case of exynos, setting dma-burst to 16Word causes permanent
>>> +      * tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>> +      * switching which is based on overlay size is not recommended as
>>> +      * overlay size varies alot towards the end of the screen and rapid
>>> +      * movement causes unstable DMA which results into iommu crash/tear.
>>
>> FIMD has width limitation so width of hardware overlay may need to be
>> aligned to 16 pixels.
>> We had faced with similar issue and the issue had been resolved by
>> aligning it to 16 pixels.
>>
>> So can you try to align it instead of changing burst len size?
>>
> 
> This problem is only with the very small FBs which are
> rendered towards corners. For large FBs like 1366x768,
> we dont see this issue though 1366 is not 16 pixel
> aligned.

Right, we had test it with 8x64 pixels. the limitation would be decided
by memory bus length(ML), DMA burst length(DL) and pixel size in bytes(PS).
And that can be calculated like below, which was guided by hardware team,

Align { (ML * DI / PS) + (4bytes / PS), 2}

It seems too big value. Actually, 16 pixels works well and I don't see
why it woks well although out of limitation.

Thanks,
Inki Dae

> 
> But I still carried out following experiments just to
> verify the fimd behaviour for small overlay.
> 
> 1) Round DOWN:
> +  overlay->crtc_width = ((overlay->crtc_width)/16)*16;
> 
> Cursor corruption from A to B
> A: [  308.530000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
> [  308.530000] ovl_width = 64, ovl_height = 64
> [  308.540000] offset_x = 1302, offset_y = 335
> [  308.540000] fb_width = 96, fb_height = 64
> 
> B: [  341.890000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
> [  341.890000] ovl_width = 48, ovl_height = 64
> [  341.900000] offset_x = 1303, offset_y = 335
> [  341.900000] fb_width = 96, fb_height = 64
> 
> 
> 2) Round UP:
> 
> + overlay->crtc_width = ((overlay->crtc_width + 16)/16)*16;
> + if(overlay->crtc_width > overlay->fb_width)
> +         overlay->crtc_width = overlay->fb_width;
> 
> Cursor corruption from A to B
> A: [   63.200000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
> [   63.210000] ovl_width = 96, ovl_height = 64
> [   63.210000] offset_x = 1271, offset_y = 346
> [   63.220000] fb_width = 96, fb_height = 64
> 
> B: [   68.600000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
> [   68.610000] ovl_width = 96, ovl_height = 64
> [   68.610000] offset_x = 1272, offset_y = 346
> [   68.620000] fb_width = 96, fb_height = 64
> 
> In both the scenarios ovl_width is always 16 Pixel aligned.
> Please let me know if you have something else in mind. I
> will check and share the observations.
> 
> The solution in the patch is proposed by hardware team
> and work fine for exynos 5250, 5420 and 5800.
> 
> Regards,
> Rahul Sharma
> 
>> Thanks,
>> Inki Dae
>>
>>> +      */
>>> +
>>> +     if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
>>> +             val &= ~WINCONx_BURSTLEN_MASK;
>>> +             val |= WINCONx_BURSTLEN_4WORD;
>>> +     }
>>> +
>>>       writel(val, ctx->regs + WINCON(win));
>>>  }
>>>
>>>
>>
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Rahul Sharma May 22, 2014, 8:39 a.m. UTC | #4
On 22 May 2014 12:25, Inki Dae <inki.dae@samsung.com> wrote:
> On 2014? 05? 22? 13:36, Rahul Sharma wrote:
>> Hi Inki,
>>
>> On 21 May 2014 16:43, Inki Dae <inki.dae@samsung.com> wrote:
>>>
>>> Hi Rahul,
>>>
>>> On 2014? 05? 07? 20:25, Rahul Sharma wrote:
>>>> From: Rahul Sharma <Rahul.Sharma@samsung.com>
>>>>
>>>> In case of exynos, setting dma-burst to 16Word causes permanent
>>>> tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>>> switching, which is based on overlay size is not recommended as
>>>> overlay size varies a lot towards the end of the screen. This
>>>> causes unstable DMA which results into tearing again.
>> [snip]
>>>> +     /*
>>>> +      * In case of exynos, setting dma-burst to 16Word causes permanent
>>>> +      * tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>>> +      * switching which is based on overlay size is not recommended as
>>>> +      * overlay size varies alot towards the end of the screen and rapid
>>>> +      * movement causes unstable DMA which results into iommu crash/tear.
>>>
>>> FIMD has width limitation so width of hardware overlay may need to be
>>> aligned to 16 pixels.
>>> We had faced with similar issue and the issue had been resolved by
>>> aligning it to 16 pixels.
>>>
>>> So can you try to align it instead of changing burst len size?
>>>
>>
>> This problem is only with the very small FBs which are
>> rendered towards corners. For large FBs like 1366x768,
>> we dont see this issue though 1366 is not 16 pixel
>> aligned.
>
> Right, we had test it with 8x64 pixels. the limitation would be decided
> by memory bus length(ML), DMA burst length(DL) and pixel size in bytes(PS).
> And that can be calculated like below, which was guided by hardware team,
>
> Align { (ML * DI / PS) + (4bytes / PS), 2}

What will be the value for ML in Exynos? 4, I guess ?

>
> It seems too big value. Actually, 16 pixels works well and I don't see
> why it works well although out of limitation.

Actually Peach Pit and Peach Pi, native LCD resolution is 1366x768
which has no issues.

In above experiments, I rendered FB of size 96*64 which corrupts when
overlay width reduces beyond 48 Pixels for 5420, 5800 and 64 for 5250.

Probably we have not tested rendering the Small buffers (<96 Pixels) with
16 bit DMA burst and towards end of FB. As if now, this scenario only
applicable for Chrome which is having cursor.

Please let me know how to proceed further.

Regards,
Rahul Sharma.

>
> Thanks,
> Inki Dae
>
>>
>> But I still carried out following experiments just to
>> verify the fimd behaviour for small overlay.
>>
>> 1) Round DOWN:
>> +  overlay->crtc_width = ((overlay->crtc_width)/16)*16;
>>
>> Cursor corruption from A to B
>> A: [  308.530000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>> [  308.530000] ovl_width = 64, ovl_height = 64
>> [  308.540000] offset_x = 1302, offset_y = 335
>> [  308.540000] fb_width = 96, fb_height = 64
>>
>> B: [  341.890000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>> [  341.890000] ovl_width = 48, ovl_height = 64
>> [  341.900000] offset_x = 1303, offset_y = 335
>> [  341.900000] fb_width = 96, fb_height = 64
>>
>>
>> 2) Round UP:
>>
>> + overlay->crtc_width = ((overlay->crtc_width + 16)/16)*16;
>> + if(overlay->crtc_width > overlay->fb_width)
>> +         overlay->crtc_width = overlay->fb_width;
>>
>> Cursor corruption from A to B
>> A: [   63.200000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>> [   63.210000] ovl_width = 96, ovl_height = 64
>> [   63.210000] offset_x = 1271, offset_y = 346
>> [   63.220000] fb_width = 96, fb_height = 64
>>
>> B: [   68.600000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>> [   68.610000] ovl_width = 96, ovl_height = 64
>> [   68.610000] offset_x = 1272, offset_y = 346
>> [   68.620000] fb_width = 96, fb_height = 64
>>
>> In both the scenarios ovl_width is always 16 Pixel aligned.
>> Please let me know if you have something else in mind. I
>> will check and share the observations.
>>
>> The solution in the patch is proposed by hardware team
>> and work fine for exynos 5250, 5420 and 5800.
>>
>> Regards,
>> Rahul Sharma
>>
>>> Thanks,
>>> Inki Dae
>>>
>>>> +      */
>>>> +
>>>> +     if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
>>>> +             val &= ~WINCONx_BURSTLEN_MASK;
>>>> +             val |= WINCONx_BURSTLEN_4WORD;
>>>> +     }
>>>> +
>>>>       writel(val, ctx->regs + WINCON(win));
>>>>  }
>>>>
>>>>
>>>
>>
>
> _______________________________________________
> dri-devel mailing list
> dri-devel@lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Inki Dae May 22, 2014, 12:14 p.m. UTC | #5
On 2014? 05? 22? 17:39, Rahul Sharma wrote:
> On 22 May 2014 12:25, Inki Dae <inki.dae@samsung.com> wrote:
>> On 2014? 05? 22? 13:36, Rahul Sharma wrote:
>>> Hi Inki,
>>>
>>> On 21 May 2014 16:43, Inki Dae <inki.dae@samsung.com> wrote:
>>>>
>>>> Hi Rahul,
>>>>
>>>> On 2014? 05? 07? 20:25, Rahul Sharma wrote:
>>>>> From: Rahul Sharma <Rahul.Sharma@samsung.com>
>>>>>
>>>>> In case of exynos, setting dma-burst to 16Word causes permanent
>>>>> tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>>>> switching, which is based on overlay size is not recommended as
>>>>> overlay size varies a lot towards the end of the screen. This
>>>>> causes unstable DMA which results into tearing again.
>>> [snip]
>>>>> +     /*
>>>>> +      * In case of exynos, setting dma-burst to 16Word causes permanent
>>>>> +      * tearing for very small buffers, e.g. cursor buffer. Burst Mode
>>>>> +      * switching which is based on overlay size is not recommended as
>>>>> +      * overlay size varies alot towards the end of the screen and rapid
>>>>> +      * movement causes unstable DMA which results into iommu crash/tear.
>>>>
>>>> FIMD has width limitation so width of hardware overlay may need to be
>>>> aligned to 16 pixels.
>>>> We had faced with similar issue and the issue had been resolved by
>>>> aligning it to 16 pixels.
>>>>
>>>> So can you try to align it instead of changing burst len size?
>>>>
>>>
>>> This problem is only with the very small FBs which are
>>> rendered towards corners. For large FBs like 1366x768,
>>> we dont see this issue though 1366 is not 16 pixel
>>> aligned.
>>
>> Right, we had test it with 8x64 pixels. the limitation would be decided
>> by memory bus length(ML), DMA burst length(DL) and pixel size in bytes(PS).
>> And that can be calculated like below, which was guided by hardware team,
>>
>> Align { (ML * DI / PS) + (4bytes / PS), 2}
> 
> What will be the value for ML in Exynos? 4, I guess ?
> 
>>
>> It seems too big value. Actually, 16 pixels works well and I don't see
>> why it works well although out of limitation.
> 
> Actually Peach Pit and Peach Pi, native LCD resolution is 1366x768
> which has no issues.
> 
> In above experiments, I rendered FB of size 96*64 which corrupts when
> overlay width reduces beyond 48 Pixels for 5420, 5800 and 64 for 5250.
> 
> Probably we have not tested rendering the Small buffers (<96 Pixels) with
> 16 bit DMA burst and towards end of FB. As if now, this scenario only
> applicable for Chrome which is having cursor.
> 
> Please let me know how to proceed further.
> 

As we had a discussion about this, there is no right solution but one
issue. So merged it temporarily until we could find the right solution
that can consider memory bus length of all Exynos SoC: the limitation
would have dependency of memory bus length of SoC.

Thanks,
Inki Dae

> Regards,
> Rahul Sharma.
> 
>>
>> Thanks,
>> Inki Dae
>>
>>>
>>> But I still carried out following experiments just to
>>> verify the fimd behaviour for small overlay.
>>>
>>> 1) Round DOWN:
>>> +  overlay->crtc_width = ((overlay->crtc_width)/16)*16;
>>>
>>> Cursor corruption from A to B
>>> A: [  308.530000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>>> [  308.530000] ovl_width = 64, ovl_height = 64
>>> [  308.540000] offset_x = 1302, offset_y = 335
>>> [  308.540000] fb_width = 96, fb_height = 64
>>>
>>> B: [  341.890000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>>> [  341.890000] ovl_width = 48, ovl_height = 64
>>> [  341.900000] offset_x = 1303, offset_y = 335
>>> [  341.900000] fb_width = 96, fb_height = 64
>>>
>>>
>>> 2) Round UP:
>>>
>>> + overlay->crtc_width = ((overlay->crtc_width + 16)/16)*16;
>>> + if(overlay->crtc_width > overlay->fb_width)
>>> +         overlay->crtc_width = overlay->fb_width;
>>>
>>> Cursor corruption from A to B
>>> A: [   63.200000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>>> [   63.210000] ovl_width = 96, ovl_height = 64
>>> [   63.210000] offset_x = 1271, offset_y = 346
>>> [   63.220000] fb_width = 96, fb_height = 64
>>>
>>> B: [   68.600000] start addr = 0x20408000, end addr = 0x2040e000, size = 0x6000
>>> [   68.610000] ovl_width = 96, ovl_height = 64
>>> [   68.610000] offset_x = 1272, offset_y = 346
>>> [   68.620000] fb_width = 96, fb_height = 64
>>>
>>> In both the scenarios ovl_width is always 16 Pixel aligned.
>>> Please let me know if you have something else in mind. I
>>> will check and share the observations.
>>>
>>> The solution in the patch is proposed by hardware team
>>> and work fine for exynos 5250, 5420 and 5800.
>>>
>>> Regards,
>>> Rahul Sharma
>>>
>>>> Thanks,
>>>> Inki Dae
>>>>
>>>>> +      */
>>>>> +
>>>>> +     if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
>>>>> +             val &= ~WINCONx_BURSTLEN_MASK;
>>>>> +             val |= WINCONx_BURSTLEN_4WORD;
>>>>> +     }
>>>>> +
>>>>>       writel(val, ctx->regs + WINCON(win));
>>>>>  }
>>>>>
>>>>>
>>>>
>>>
>>
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel@lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-samsung-soc" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/gpu/drm/exynos/exynos_drm_fimd.c b/drivers/gpu/drm/exynos/exynos_drm_fimd.c
index 40fd6cc..ef56077 100644
--- a/drivers/gpu/drm/exynos/exynos_drm_fimd.c
+++ b/drivers/gpu/drm/exynos/exynos_drm_fimd.c
@@ -38,6 +38,7 @@ 
  */
 
 #define FIMD_DEFAULT_FRAMERATE 60
+#define MIN_FB_WIDTH_FOR_16WORD_BURST 128
 
 /* position control register for hardware window 0, 2 ~ 4.*/
 #define VIDOSD_A(win)		(VIDOSD_BASE + 0x00 + (win) * 16)
@@ -446,6 +447,19 @@  static void fimd_win_set_pixfmt(struct fimd_context *ctx, unsigned int win)
 
 	DRM_DEBUG_KMS("bpp = %d\n", win_data->bpp);
 
+	/*
+	 * In case of exynos, setting dma-burst to 16Word causes permanent
+	 * tearing for very small buffers, e.g. cursor buffer. Burst Mode
+	 * switching which is based on overlay size is not recommended as
+	 * overlay size varies alot towards the end of the screen and rapid
+	 * movement causes unstable DMA which results into iommu crash/tear.
+	 */
+
+	if (win_data->fb_width < MIN_FB_WIDTH_FOR_16WORD_BURST) {
+		val &= ~WINCONx_BURSTLEN_MASK;
+		val |= WINCONx_BURSTLEN_4WORD;
+	}
+
 	writel(val, ctx->regs + WINCON(win));
 }