copy_to_user speed from dma_alloc_coherent vs. kmalloc buffer

copy_to_user speed from dma_alloc_coherent vs. kmalloc buffer

[Orjan Friberg]

Hi,

I have a driver where I do memory to memory DMA between GPMC and SDRAM. 
  Adding a read function, I found that copy_to_user from a 
dma_alloc_coherent buffer is significantly slower than from a kmalloc'd one.

Looking at arch/arm/include/asm/pgtable.h I suspect this difference in 
speed is due to the fact that the dma_alloc_coherent buffer is unbuffered.

What are my options (besides using mmap)?

* Reserve a portion of memory at boot time to be used as the DMA 
destination buffer, use ioremap_cached + manual cache flush as needed?
* Turn on buffering for the DMA destination buffer for the duration of 
the copy_to_user call, then turn it off again (and flush it from the cache)?
* Something else entirely?


This is on a 3730, on Linux 2.6.32.

Thanks,
Orjan

-- 
Orjan Friberg
FlatFrog Laboratories AB

Re: copy_to_user speed from dma_alloc_coherent vs. kmalloc buffer

[Orjan Friberg]

On 2011-04-20 17:12, Orjan Friberg wrote:
> What are my options (besides using mmap)?

It looks like kmalloc + dma_map_single for the DMA destination buffer 
and then dma_sync_single_for_{cpu,device} around the call to 
copy_to_user pretty much does the trick.  At least the %sys load 
measured with mpstat goes from 13% to 2%.

-- 
Orjan Friberg
FlatFrog Laboratories AB