On Thu, Aug 02, 2018 at 09:50:58AM -0700, Dave Watson wrote: > On 08/02/18 09:50 PM, Vakul Garg wrote: > > Function decrypt_skb() made a bad assumption that number of sg entries > > required for mapping skb to be decrypted would always be less than > > MAX_SKB_FRAGS. The required count of sg entries for skb should always be > > calculated. If they cannot fit in local array sgin_arr[], allocate them > > from heap irrespective whether it is zero-copy case or otherwise. The > > change also benefits the non-zero copy case as we could use sgin_arr[] > > instead of always allocating sg entries from heap. > > > > Signed-off-by: Vakul Garg <vakul.g...@nxp.com> > > Looks great, thanks. > > Reviewed-by: Dave Waston <davejwat...@fb.com>
I agree that this is a problem, but I'm not sure that this is the best way to fix it. Calling skb_cow_data unconditionally is expensive. In my benchmarks this patch cause a 5% CPU regression, all from memcpy from skb_cow_data, and a 15% regression in encrypted NBD IOPS. It is possible to calculate the number of scatterlist elements required to map the skb without invoking skb_cow_data. I'll have a patch up shortly.
