Ideas/suggestions to avoid repeated locking and reducing too many lists with dmaengine?
joelf at ti.com
Mon Feb 24 17:38:48 EST 2014
Firstly, thanks for your quick reply.
On 02/24/2014 01:21 PM, Russell King - ARM Linux wrote:
> Wrapping... (I've had to manually edit this.)
> On Mon, Feb 24, 2014 at 01:03:32PM -0600, Joel Fernandes wrote:
>> Just wanted your thoughts/suggestions on how we can avoid overhead in
>> the EDMA dmaengine driver. I am seeing a lots of performance drop
>> specially for small transfers with EDMA versus before raw EDMA was
>> moved to DMAEngine framework (atleast 25%).
>> One of the things I am thinking about is the repeated (spin)
>> locking/unlocking of the virt_dma_chan->lock or vc->lock. In many
>> cases, there's only 1 user or thread requiring to do a DMA, so I
>> feel the locking is unnecessary and potential overhead. If there's
>> a sane way to detect this an avoid locking altogether, that
>> would be great.
> For the case where there's no contention, spinlocks /should/ be light.
> What will make them more expensive is if you have things like lockdep
> enabled, which adds much more code into those paths to do state tracking.
> It's a known side effect of using that debug.
> So, if you're developing, then you should always have turned lockdep on.
> If you're testing for performance, you should have lockdep and spinlock
> debugging turned off.
Thanks, indeed I had to turn them off.
>> Also with respect to virt_dma (which is used by edma to manage all the
>> descriptors and lists) there are too many lists: submitted, issued,
>> completed etc and the descriptor moves from one to the other. I am
>> thinking if there is a way we can avoid using so many lists and just
>> have 2 lists and move the desc from one list to the other, That could
>> avoid using the intermediate list altogether and classify dma requests
>> as "done" or "not done".
> The reason I created separate submitted and issued lists is that it's
> much easier to manage than having everything on a single list.
> We could deal with the submitted vs issued list, and that's to have the
> channel store the cookie for the last issued descriptor - but I wonder
> if it's worth the effort.
> What I'd suggest is to try some profiling, and post some profiling
> results which show where the problems are, rather than pointing at
> bits of code you might not particularly like.
Actually I did do some tracing earlier before I posted this thread- and
notice there was excessive traces of locking/unlocking. It is very light
though as you pointed and lighter without debug options. The only other
notable difference is the fact that we are now going through the dmaengine
framework in the newer kernel vs the faster one.
One more thing in my trace is omap_dma_sync repeatedly call in memcpy_to_io
for every barrier call which is not necessary. I am working on a fix this.
On turning off DEBUG_KERNEL and running more tests, I do see some
improvements however the throughput reduction is still =~ 10%
With a modified openssl speed test app, I sent 16-byte sized block
repeatedly to the AES crypto hardware accelerator using EDMA:
On v3.13.5 kernel:
root at am335x-evm:~# openssl speed -evp aes-128-cbc -engine cryptodev
engine "cryptodev" set.
Doing aes-128-cbc for 3s on 16 size blocks: 79902 aes-128-cbc's
With v3.2 kernel,
Doing aes-128-cbc for 3s on 16 size blocks: 92314 aes-128-cbc's
So we're able to encrypt around 13k more ops, or around 4.5k ops/second
As such, I do see this as a problem but difference is much lesser for
larger blocks so its not a very big alarm. We should just be using PIO mode
for small blocks.
More information about the linux-arm-kernel