Re: [RFC PATCH 0/7] Improve swiotlb performance by using physical addresses

[Alexander Duyck <alexander.h.duyck@intel.com>]

On 10/05/2012 01:02 PM, Andi Kleen wrote:
>> I was thinking the issue was all of the calls to relatively small
>> functions occurring in quick succession.  The way most of this code is
>> setup it seems like it is one small function call in turn calling
>> another, and then another, and I would imagine the code fragmentation
>> can have a significant negative impact.
> Maybe. Can you just inline everything and see if it it's faster then?
>
> This was out of line when the "text cost at all costs" drive was still
> envogue, but luckily we're not doing  that anymore.
>
> -Andiu
>

Inlining everything did speed things up a bit, but I still didn't reach
the same speed I achieved using the patch set.  However I did notice the
resulting swiotlb code was considerably larger.

I did a bit more digging and the issue may actually be simple repetition
of the calls.  By my math it would seem we would end up calling
is_swiotlb_buffer 3 times per packet in the routing test case, once in
sync_for_cpu and once for sync_for_device in the Rx cleanup path, and
once in unmap_page in the Tx cleanup path.  Each call to
is_swiotlb_buffer will result in 2 calls to __phys_addr.  In freeing the
skb we end up doing a call to virt_to_head_page which will call
__phys_addr.  In addition we end up mapping the skb using map_single so
we end up using __phys_addr to do a virt_to_page translation in the
xmit_frame_ring path, and then call __phys_addr when we check
dma_mapping_error.  So in total that ends up being 3 calls to
is_swiotlb_buffer, and 9 calls to __phys_addr per packet routed.

With the patches the is_swiotlb_buffer function, which was 25 lines of
assembly, is replaced with 8 lines of assembly and becomes inline.  In
addition we drop the number of calls to __phys_addr from 9 to 2 by
dropping them all from swiotlb.  By my math I am probably saving about
120 instructions per packet.  I suspect all of that would probably be
cutting the number of instructions per packet enough to probably account
for a 5% difference when you consider I am running at about 1.5Mpps per
core on a 2.7Ghz processor.

Thanks,

Alex