[dpdk-dev] [PATCH v2 1/2] examples/ipsec-secgw: fix 1st pkt dropped for inline crypto
bernard.iremonger at intel.com
Tue Mar 26 12:41:34 CET 2019
> >> Subject: Re: [dpdk-dev] [PATCH v2 1/2] examples/ipsec-secgw: fix 1st
> >> pkt dropped for inline crypto
> >> Hi Bernard,
> >> On 3/7/2019 8:27 PM, Bernard Iremonger wrote:
> >>> Inline crypto installs a flow rule in the NIC. This flow rule must
> >>> be installed before the first inbound packet is received.
> >>> The create_session() function installs the flow rule,
> >>> create_session() has been refactored into create_inline_session()
> >>> and create_lookaside_session(). The create_inline_session() function
> >>> uses the socket_ctx data and is now called at initialisation in
> >>> sa_add_rules().
> >> why do we need a separate function for session creation for inline
> >> and lookaside cases?
> >> Why can't we initialize the sessions on sa_init in both the cases?
> > For the create_inline_session(struct socket_ctx *skt_ctx, struct
> > ipsec_sa *sa) function, all of the required data is available in the in the
> skt_ctx variable.
> > The skt_ctx variable is already setup when sa_init() is called.
> > For the create_lookaside_session(struct ipsec_ctx *ipsec_ctx, struct
> > ipsec_sa *sa) function, the required data is available in the ipsec_ctx
> > The ipsec_ctx variable is not setup when sa_init() is called.
> > It is setup in the main_loop() function when the variable qconf is setup.
> > The main_loop() function is called after the sa_init() function is called.
> > I hope this answers your question
> Whatever information that is required for session creation is available before
> we call the main loop() in both the cases.
> My point is both the sessions(inline/lookaside) can be init at the same
> position, we do not need to have a separate path for them.
> If it is not possible in sa_init(), it may be somewhere else before the actual
> data path is started.
> The problem with inline processing is that, h/w need to know the SA before
> the first packet is received. So we cannot init the session on receive of first
> packet. However there is no such limitation in case of lookaside, it can be
> initialized anywhere.
This patch is intended to fix the bug in the inline processing, which is that the flow rule must be installed, before the first packet is received while leaving the lookaside processing as it was originally.
It is not intended to refactor the lookaside processing.
More information about the dev