Re: [PATCH] virtio: don't zero out memory region cache for indirect descriptors

[Stefan Hajnoczi]

On Tue, Aug 08, 2023 at 12:28:47AM +0200, Ilya Maximets wrote:
> Lots of virtio functions that are on a hot path in data transmission
> are initializing indirect descriptor cache at the point of stack
> allocation.  It's a 112 byte structure that is getting zeroed out on
> each call adding unnecessary overhead.  It's going to be correctly
> initialized later via special init function.  The only reason to
> actually initialize right away is the ability to safely destruct it.
> However, we only need to destruct it when it was used, i.e. when a
> desc_cache points to it.
> 
> Removing these unnecessary stack initializations improves throughput
> of virtio-net devices in terms of 64B packets per second by 6-14 %
> depending on the case.  Tested with a proposed af-xdp network backend
> and a dpdk testpmd application in the guest, but should be beneficial
> for other virtio devices as well.
> 
> Signed-off-by: Ilya Maximets <i.maximets@ovn.org>
> ---
>  hw/virtio/virtio.c | 42 +++++++++++++++++++++++++++---------------
>  1 file changed, 27 insertions(+), 15 deletions(-)

Another option is to create an address_space_cache_init_invalid()
function that only assigns mrs.mr = NULL instead of touching all bytes
of the struct like = MEMORY_REGION_CACHE_INVALID. There would be less
code and the existing mrs.mr check in address_space_cache_destroy()
would serve the same function as the desc_cache == &indirect_desc_cache
check added by this patch.

I'm fine with your approach too:

Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>

signature.asc
Description: PGP signature