[PATCH] ipsec: use sym_session_opaque_data for RTE_SECURITY_TYPE_CPU_CRYPTO

[Konstantin Ananyev]

Hi Garry,

> Hi Konstantin, Akhil,
> 
> The patch is based on an issue I encountered when using the CPU_CRYPTO
> support - I was having problems where the ipsec session lookup was
> failing / was inconsistent.
> 
> Examining the code in DPDK and looking for the use of
> RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO I could see a reasonably
> consistent pattern where if TYPE_NONE or TYPE_CPU_CRYPTO was set -
> then the code was making use of ss->crypto.ses instead of
> ss->security.ses.
> 
> For example - see examples/ipsec-secgw.c where the one_session_free
> function has the following code:
> 
>     if (ips->type == RTE_SECURITY_ACTION_TYPE_NONE ||
>         ips->type == RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO) {
>         /* Session has not been created */
>         if (ips->crypto.ses == NULL)
>             return 0;
> 
>         ret = rte_cryptodev_sym_session_free(ips->crypto.dev_id,
>                 ips->crypto.ses);
>     } else {
>         /* Session has not been created */
>         if (ips->security.ctx == NULL || ips->security.ses == NULL)
>             return 0;
> 
>         ret = rte_security_session_destroy(ips->security.ctx,
>                            ips->security.ses);
>     }
> 
> And similarly - if we look at the session_check function in lib/ipsec/ses.c:
> 
>     if (ss->type == RTE_SECURITY_ACTION_TYPE_NONE ||
>         ss->type == RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO) {
>         if (ss->crypto.ses == NULL)
>             return -EINVAL;
>     } else {
>         if (ss->security.ses == NULL)
>             return -EINVAL;
>         if ((ss->type == RTE_SECURITY_ACTION_TYPE_INLINE_CRYPTO ||
>                 ss->type ==
>                 RTE_SECURITY_ACTION_TYPE_INLINE_PROTOCOL) &&
>                 ss->security.ctx == NULL)
>             return -EINVAL;
>     }

Thanks for explanation.
Yes, I agree that TYPE_NONE and TYPE_CPU_CRYPTO both use crypto session
to keep/propagate crypto related pamaters.
What is not clear to me why for  and TYPE_CPU_CRYPTO we need to store
pointer to rte_ipsec_session as opaque user data for crypto session.
As I remember, for lookaside crypto we need to do that to extract
related rte_ipsec_session pointer from crypto_op, after lookaside crypto device
finished the processing and sending sym-ops back to user.
But for CPU_CRYPTO it is not necessary, as all processing is synchronous and
user already has a pointer for  related rte_ipsec_session.
We probably still can, but what is the benefit, who will use it?

Actually looking at the rte_ipsec_session_prepare() once again,
you probably right - it is a bug here, as we shouldn’t call  rte_security_session_opaque_data_set()
for TYPE_CPU_CRYPTO.
So shouldn't it be like that:

        ss->pkt_func = fp;

        if (ss->type == RTE_SECURITY_ACTION_TYPE_NONE)
                rte_cryptodev_sym_session_opaque_data_set(ss->crypto.ses,
                        (uintptr_t)ss);
-       else
+      else if (ss->type != RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO)
                rte_security_session_opaque_data_set(ss->security.ses, (uintptr_t)ss);
 
> Without the patch in rte_ipsec_session_prepare - for the
> RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO type, then ss->crypto.ses will not
> be set.

Hmm... not clear why?
AFAIK, ss->crypto.ses supposed to be set by user *before* calling rte_ipsec_session_prepare().
From lib/ipsec/rte_ipsec.h:
/**
 * Checks that inside given rte_ipsec_session crypto/security fields
 * are filled correctly and setups function pointers based on these values.
 * Expects that all fields except IPsec processing function pointers
 * (*pkt_func*) will be filled correctly by caller.
 * @param ss
 *   Pointer to the *rte_ipsec_session* object
 * @return
 *   - Zero if operation completed successfully.
 *   - -EINVAL if the parameters are invalid.
 */
int
rte_ipsec_session_prepare(struct rte_ipsec_session *ss);

> 
> Regards,
> 
> Garry.
> 
> 
> On Tue, Oct 31, 2023 at 1:09 AM Konstantin Ananyev
> <konstantin.v.ananyev at yandex.ru> wrote:
> >
> > >
> > >
> > > ipsec related processing in dpdk makes use of the crypto.ses opaque
> > > data pointer.  This patch updates rte_ipsec_session_prepare to set
> > > ss->crypto.ses in the RTE_SECURITY_TYPE_CPU_CRYPTO case.
> >
> > Hmm.. not sure why we need to do that for CPU_CRYPTO?
> > As I remember CPU_CRYPTO is synchronous operation and before calling
> > rte_ipsec_pkt_cpu_prepare() should already know ipsec session these
> > packets belong to.
> > Can you probably explain the logic behind this patch a bit more?
> > Konstantin
> >
> > >
> > > Signed-off-by: Garry Marshall <gazmarsh at meaningfulname.net>
> > > ---
> > >  lib/ipsec/ses.c | 3 ++-
> > >  1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/lib/ipsec/ses.c b/lib/ipsec/ses.c
> > > index d9ab1e6d2b..29eb5ff6ca 100644
> > > --- a/lib/ipsec/ses.c
> > > +++ b/lib/ipsec/ses.c
> > > @@ -44,7 +44,8 @@ rte_ipsec_session_prepare(struct rte_ipsec_session *ss)
> > >
> > >       ss->pkt_func = fp;
> > >
> > > -     if (ss->type == RTE_SECURITY_ACTION_TYPE_NONE)
> > > +     if (ss->type == RTE_SECURITY_ACTION_TYPE_NONE ||
> > > +             ss->type == RTE_SECURITY_ACTION_TYPE_CPU_CRYPTO)
> > >               rte_cryptodev_sym_session_opaque_data_set(ss->crypto.ses,
> > >                       (uintptr_t)ss);
> > >       else
> > > --
> > > 2.39.2