> -----Original Message-----
> From: Akhil Goyal <gak...@marvell.com>
> Sent: Monday, 17 March 2025 12:23
> To: Shani Peretz <shper...@nvidia.com>; dev@dpdk.org
> Cc: Suanming Mou <suanmi...@nvidia.com>; sta...@dpdk.org; Brian
> Dooley <brian.doo...@intel.com>; Pablo de Lara
> <pablo.de.lara.gua...@intel.com>
> Subject: RE: [EXTERNAL] [PATCH] app/crypto-perf: fix aad offset alignment
>
> External email: Use caution opening links or attachments
>
>
> Hi,
> > AAD offset in AES-GCM crypto test was calculated by adding 16-byte
> > alignment after the IV, which is only needed in AES-CCM.
>
> Agreed that CCM has a requirement for 16B alignment.
> But for GCM, does it break any protocol? Can we not align to byte boundary
> for performance?
> This is a performance application which mainly focus on getting the best
> throughput.
> Did you check if it is having some performance degradation?
>
> >
> > The patch correct the AAD offset calculation in AES-GCM algorithm tests.
> >
> > Fixes: 0b242422d385 ("app/crypto-perf: set AAD after the crypto
> > operation")
> > Cc: sta...@dpdk.org
> >
> > Signed-off-by: Shani Peretz <shper...@nvidia.com>
> > ---
> > app/test-crypto-perf/cperf_ops.c | 4 +++-
> > 1 file changed, 3 insertions(+), 1 deletion(-)
> >
> > diff --git a/app/test-crypto-perf/cperf_ops.c
> > b/app/test-crypto-perf/cperf_ops.c
> > index 6d5f510220..f9be51e17f 100644
> > --- a/app/test-crypto-perf/cperf_ops.c
> > +++ b/app/test-crypto-perf/cperf_ops.c
> > @@ -688,7 +688,9 @@ cperf_set_ops_aead(struct rte_crypto_op **ops,
> > uint16_t i;
> > /* AAD is placed after the IV */
> > uint16_t aad_offset = iv_offset +
> > - RTE_ALIGN_CEIL(test_vector->aead_iv.length, 16);
> > + ((options->aead_algo == RTE_CRYPTO_AEAD_AES_CCM)
> > ?
> > + RTE_ALIGN_CEIL(test_vector->aead_iv.length, 16) :
> > + test_vector->aead_iv.length);
> >
> > for (i = 0; i < nb_ops; i++) {
> > struct rte_crypto_sym_op *sym_op = ops[i]->sym;
> > --
> > 2.25.1
I checked the throughput test, and I haven't noticed any degradation compared
to upstream. I can share the results if needed.
Note that regardless of the performance it fixes several segmentation faults in
the test.
(The problem is that we allocate the crypto_op_private_size without alignment,
but we try to access as if it was 16 byte alignment)
