[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[Rameshwar Prasad Sahu]

The DMA engine supports memory copy, RAID5 XOR, RAID6 PQ, and other
computations. But the bandwidth of the entire DMA engine is shared
among all channels. This patch re-configures operations availability
such that one can achieve maximum performance for XOR and PQ
computation by removing the memory offload operations.

Signed-off-by: Rameshwar Prasad Sahu <rsahu@apm.com>
---
 drivers/dma/xgene-dma.c | 1 -
 1 file changed, 1 deletion(-)

diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
index 8d57b1b..cd38022 100644
--- a/drivers/dma/xgene-dma.c
+++ b/drivers/dma/xgene-dma.c
@@ -1707,7 +1707,6 @@ static void xgene_dma_set_caps(struct xgene_dma_chan *chan,
 	dma_cap_zero(dma_dev->cap_mask);

 	/* Set DMA device capability */
-	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
 	dma_cap_set(DMA_SG, dma_dev->cap_mask);

 	/* Basically here, the X-Gene SoC DMA engine channel 0 supports XOR
--
1.8.2.1

[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[Vinod Koul]

On Thu, Oct 08, 2015 at 02:36:57PM +0530, Rameshwar Prasad Sahu wrote:
> The DMA engine supports memory copy, RAID5 XOR, RAID6 PQ, and other
> computations. But the bandwidth of the entire DMA engine is shared
> among all channels. This patch re-configures operations availability
> such that one can achieve maximum performance for XOR and PQ
> computation by removing the memory offload operations.
> 
> Signed-off-by: Rameshwar Prasad Sahu <rsahu@apm.com>
> ---
>  drivers/dma/xgene-dma.c | 1 -
>  1 file changed, 1 deletion(-)
> 
> diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
> index 8d57b1b..cd38022 100644
> --- a/drivers/dma/xgene-dma.c
> +++ b/drivers/dma/xgene-dma.c
> @@ -1707,7 +1707,6 @@ static void xgene_dma_set_caps(struct xgene_dma_chan *chan,
>  	dma_cap_zero(dma_dev->cap_mask);
> 
>  	/* Set DMA device capability */
> -	dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
>  	dma_cap_set(DMA_SG, dma_dev->cap_mask);

This will not work. Still clients can invoke memcpy operation. You need to
remove the callback

-- 
~Vinod

[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[Rameshwar Sahu]

Hi Vinod,

On Fri, Oct 9, 2015 at 9:42 PM, Vinod Koul <vinod.koul@intel.com> wrote:
> On Thu, Oct 08, 2015 at 02:36:57PM +0530, Rameshwar Prasad Sahu wrote:
>> The DMA engine supports memory copy, RAID5 XOR, RAID6 PQ, and other
>> computations. But the bandwidth of the entire DMA engine is shared
>> among all channels. This patch re-configures operations availability
>> such that one can achieve maximum performance for XOR and PQ
>> computation by removing the memory offload operations.
>>
>> Signed-off-by: Rameshwar Prasad Sahu <rsahu@apm.com>
>> ---
>>  drivers/dma/xgene-dma.c | 1 -
>>  1 file changed, 1 deletion(-)
>>
>> diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
>> index 8d57b1b..cd38022 100644
>> --- a/drivers/dma/xgene-dma.c
>> +++ b/drivers/dma/xgene-dma.c
>> @@ -1707,7 +1707,6 @@ static void xgene_dma_set_caps(struct xgene_dma_chan *chan,
>>       dma_cap_zero(dma_dev->cap_mask);
>>
>>       /* Set DMA device capability */
>> -     dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
>>       dma_cap_set(DMA_SG, dma_dev->cap_mask);
>
> This will not work. Still clients can invoke memcpy operation. You need to
> remove the callback

Got it, instead of removing the callback registration can we keep it
inside the condition like..

if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
            dma_dev->device_prep_dma_memcpy = xgene_dma_prep_memcpy;


>
> --
> ~Vinod

[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[Vinod Koul]

On Fri, Oct 09, 2015 at 10:08:32PM +0530, Rameshwar Sahu wrote:
> Hi Vinod,
> 
> On Fri, Oct 9, 2015 at 9:42 PM, Vinod Koul <vinod.koul@intel.com> wrote:
> > On Thu, Oct 08, 2015 at 02:36:57PM +0530, Rameshwar Prasad Sahu wrote:
> >> The DMA engine supports memory copy, RAID5 XOR, RAID6 PQ, and other
> >> computations. But the bandwidth of the entire DMA engine is shared
> >> among all channels. This patch re-configures operations availability
> >> such that one can achieve maximum performance for XOR and PQ
> >> computation by removing the memory offload operations.
> >>
> >> Signed-off-by: Rameshwar Prasad Sahu <rsahu@apm.com>
> >> ---
> >>  drivers/dma/xgene-dma.c | 1 -
> >>  1 file changed, 1 deletion(-)
> >>
> >> diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
> >> index 8d57b1b..cd38022 100644
> >> --- a/drivers/dma/xgene-dma.c
> >> +++ b/drivers/dma/xgene-dma.c
> >> @@ -1707,7 +1707,6 @@ static void xgene_dma_set_caps(struct xgene_dma_chan *chan,
> >>       dma_cap_zero(dma_dev->cap_mask);
> >>
> >>       /* Set DMA device capability */
> >> -     dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
> >>       dma_cap_set(DMA_SG, dma_dev->cap_mask);
> >
> > This will not work. Still clients can invoke memcpy operation. You need to
> > remove the callback
> 
> Got it, instead of removing the callback registration can we keep it
> inside the condition like..
> 
> if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
>             dma_dev->device_prep_dma_memcpy = xgene_dma_prep_memcpy;

But the code is dead and will not be tested and hence not maintained. I do
not see a reason to keep dead code in kernel

-- 
~Vinod

[PATCH v2] dmaengine: xgene-dma: Disable memcpy operation due to performance drop

[Rameshwar Sahu]

Hi Vinod,

On Mon, Oct 12, 2015 at 7:53 PM, Vinod Koul <vinod.koul@intel.com> wrote:
> On Fri, Oct 09, 2015 at 10:08:32PM +0530, Rameshwar Sahu wrote:
>> Hi Vinod,
>>
>> On Fri, Oct 9, 2015 at 9:42 PM, Vinod Koul <vinod.koul@intel.com> wrote:
>> > On Thu, Oct 08, 2015 at 02:36:57PM +0530, Rameshwar Prasad Sahu wrote:
>> >> The DMA engine supports memory copy, RAID5 XOR, RAID6 PQ, and other
>> >> computations. But the bandwidth of the entire DMA engine is shared
>> >> among all channels. This patch re-configures operations availability
>> >> such that one can achieve maximum performance for XOR and PQ
>> >> computation by removing the memory offload operations.
>> >>
>> >> Signed-off-by: Rameshwar Prasad Sahu <rsahu@apm.com>
>> >> ---
>> >>  drivers/dma/xgene-dma.c | 1 -
>> >>  1 file changed, 1 deletion(-)
>> >>
>> >> diff --git a/drivers/dma/xgene-dma.c b/drivers/dma/xgene-dma.c
>> >> index 8d57b1b..cd38022 100644
>> >> --- a/drivers/dma/xgene-dma.c
>> >> +++ b/drivers/dma/xgene-dma.c
>> >> @@ -1707,7 +1707,6 @@ static void xgene_dma_set_caps(struct xgene_dma_chan *chan,
>> >>       dma_cap_zero(dma_dev->cap_mask);
>> >>
>> >>       /* Set DMA device capability */
>> >> -     dma_cap_set(DMA_MEMCPY, dma_dev->cap_mask);
>> >>       dma_cap_set(DMA_SG, dma_dev->cap_mask);
>> >
>> > This will not work. Still clients can invoke memcpy operation. You need to
>> > remove the callback
>>
>> Got it, instead of removing the callback registration can we keep it
>> inside the condition like..
>>
>> if (dma_has_cap(DMA_MEMCPY, dma_dev->cap_mask))
>>             dma_dev->device_prep_dma_memcpy = xgene_dma_prep_memcpy;
>
> But the code is dead and will not be tested and hence not maintained. I do
> not see a reason to keep dead code in kernel

Got it, I will remove the memcpy offload support and will send another
patch for review.
>
> --
> ~Vinod