[dm-devel] dm: submit stacked requests in irq enabled context

[Keith Busch]

On Tue, May 09, 2017 at 03:19:31PM +0530, Neeraj Soni wrote:
>    + Alasdair and dm-devel for awareness and inputs.
> 
>    On 5/9/2017 12:26 PM, Neeraj Soni wrote:
> 
>      Hi Keith/Snitzer,
> 
>      I have recently started using kernel 4.4 on a Android device and ran
>      Androbench to check storage read/write performance. I found that the
>      Random Read (RR) and Random write(RW) performance with Full Disk
>      Encryption is degraded compared to no Disk Encryption. Initially i
>      thought this must the issue with the storage part used and i compared
>      the performance of similar storage part on a device that was using
>      Android with kernel 3.18. I found that with no Disk Encryption the
>      performance was equivalent to the device which was using 4.4 but with
>      Disk Encryption there was degradation in RR (~20%) and RW(10%).
> 
>      I then tried to compare the changes that was brought in kernel 4.4 in
>      Full Disk Encryption path. I came across the patch mentioned in subject
>      and found that now a worker thread is scheduled in dm_request_fn() to
>      process the requests instead of directly invoking map_request() as was
>      in kernel 3.18.
> 
>      I reverted this patch and found that the RR and RW performance was now
>      closer to what we have without Disk Encryption. From the commit message
>      i understand that this change is significant and will be required for
>      blk-mq support but have you came across such degradation issue with your
>      patch and do we have any fix for this degradation available?

Just for comparison, could you check performance on a more recent stable
kernel?