[EXT] Re: [PATCH v2 2/3] ip_frag: improve reassembly lookup performance

[Pavan Nikhilesh Bhagavatula]

> On Tue, 23 May 2023 20:09:20 +0530
> <pbhagavatula at marvell.com> wrote:
> 
> > From: Pavan Nikhilesh <pbhagavatula at marvell.com>
> >
> > Improve reassembly lookup performance by using NEON intrinsics for
> > key validation.
> >
> > Signed-off-by: Pavan Nikhilesh <pbhagavatula at marvell.com>
> > ---
> >  lib/ip_frag/ip_frag_internal.c   | 224 +++++++++++++++++++++++++------
> >  lib/ip_frag/ip_reassembly.h      |   6 +
> >  lib/ip_frag/rte_ip_frag_common.c |  10 ++
> >  3 files changed, 196 insertions(+), 44 deletions(-)
> 
> 
> Using a function pointer for the lookup has some downsides.
> On Intel an indirect call is slower especially with SPECTRE mitigations.
> 

The patch doesn't use direct function pointers, it stores the function id and switches between them.
Function ID scheme doesn't break primary/secondary process scheme even with ASLR scheme.

> The bigger issue is that indirect call will break usage from primary/secondary
> process with ASLR. If primary sets up table and secondary uses it the function
> will be in different places.

I will be dropping this patch since the performance improvement with NEON is negligible, there is lot of bucket 
state that we unfortunately don't cache with the current implementation.