[dpdk-dev] [RFC PATCH 0/1] IPSec Inline and look aside crypto offload

[Jerin Jacob]

-----Original Message-----
> Date: Fri, 8 Sep 2017 16:42:56 +0530
> From: Akhil Goyal <akhil.goyal at nxp.com>
> To: Jerin Jacob <jerin.jacob at caviumnetworks.com>, Radu Nicolau
>  <radu.nicolau at intel.com>
> CC: Thomas Monjalon <thomas at monjalon.net>, dev at dpdk.org,
>  borisp at mellanox.com, declan.doherty at intel.com, aviadye at mellanox.com,
>  sandeep.malik at nxp.com, hemant.agrawal at nxp.com,
>  pablo.de.lara.guarch at intel.com, pathreya at caviumnetworks.com,
>  andriy.berestovskyy at cavium.com, sunil.kulkarni at cavium.com,
>  balasubramanian.manoharan at cavium.com, suheil.chandran at cavium.com
> Subject: Re: [RFC PATCH 0/1] IPSec Inline and look aside crypto offload
> User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:52.0) Gecko/20100101
>  Thunderbird/52.3.0
> 
> Hi  Jerin,

Hi Akhil,

> 
> On 9/6/2017 9:23 PM, Jerin Jacob wrote:
> > -----Original Message-----
> > > Date: Thu, 31 Aug 2017 15:09:45 +0100
> > > From: Radu Nicolau <radu.nicolau at intel.com>
> > > To: Thomas Monjalon <thomas at monjalon.net>, Akhil Goyal <akhil.goyal at nxp.com>
> > > CC: dev at dpdk.org, borisp at mellanox.com, declan.doherty at intel.com,
> > >   aviadye at mellanox.com, sandeep.malik at nxp.com, hemant.agrawal at nxp.com,
> > >   pablo.de.lara.guarch at intel.com
> > > Subject: Re: [dpdk-dev] [RFC PATCH 0/1] IPSec Inline and look aside crypto
> > >   offload
> > > User-Agent: Mozilla/5.0 (Windows NT 10.0; WOW64; rv:52.0) Gecko/20100101
> > >   Thunderbird/52.1.0
> > > 
> > > 
> > > On 8/31/2017 2:14 PM, Thomas Monjalon wrote:
> > > > 31/08/2017 12:52, Akhil Goyal:
> > > > > On 8/31/2017 3:36 PM, Thomas Monjalon wrote:
> > > > > > 31/08/2017 11:37, Akhil Goyal:
> > > > > > > On 8/29/2017 8:19 PM, Thomas Monjalon wrote:
> > > > > > > > 25/07/2017 13:21, Akhil Goyal:
> > > > > > > 2. Ipsec inline(RTE_SECURITY_SESS_ETH_INLINE_CRYPTO) - This is when the
> > > > > > > crypto operations are performed by ethernet device instead of crypto
> > > > > > > device. This is also without protocol knowledge inside the ethernet device
> > > > > > If the ethernet device can act as a crypto device, this function
> > > > > > should be offered via the cryptodev interface.
> > > > > yes this could be thought of but the intent was to keep cryptodev and
> > > > > ethdev separate, as this would create confusion and will become
> > > > > difficult to manage.
> > > > I think the reverse: it is confusing to do crypto operations through
> > > > ethdev interface.
> > > > If a device can do "standalone crypto" and networking, it should appear as
> > > > 2 different ports in my opinion.
> > > > 
> > > > > > How is it different from mode RTE_SECURITY_SESS_NONE?
> > > > > In RTE_SECURITY_SESS_NONE - crypto device is used for crypto operations.
> > > > > In RTE_SECURITY_SESS_ETH_INLINE_CRYPTO - ethernet device is used for
> > > > > crypto operations.
> > > > > For details of the data path of this mode, refer to the covernote of RFC
> > > > > patch from Boris.
> > > > > http://dpdk.org/ml/archives/dev/2017-July/070793.html
> > > > > 
> > > > > For implementation of this mode, see patches from Radu,
> > > > > http://dpdk.org/ml/archives/dev/2017-August/073587.html
> > > > Boris RFC uses rte_flow.
> > > > Radu implementation does not use rte_flow.
> > > > So I still don't understand the big picture.
> > > > Boris asked the question and had no answer.
> > > I'll answer here: it was an omission from my side; v2 of the will include
> > > rte_flow usage, derived from Boris RFC.
> > 
> > 
> > Cavium would like to contribute to the definition of this specification
> > as our HW supports the IPSec offload.
> > 
> > I was trying to review the latest patch. But it looks like there are
> > multiple versions of the header file floating around. like,
> > 
> > http://dpdk.org/ml/archives/dev/2017-August/073587.html
> > http://dpdk.org/ml/archives/dev/2017-August/073738.html
> > 
> > Can some one tell which one is latest one to review?
> > 
> > Previously for rte_flow, rte_eventdev specification, etc we had some
> > header file sign off before jumping to the RFC implementation. IMO, That
> > model was useful where all the vendors could make inline comments on the
> > proposal instead of maintaining in the draft repo.  So it possible for
> > sending the latest revision of the header file patch on the mailing list
> > for the inline comments.
> > 
> 
> The RFC remained for some time, there were not many comments. so we all
> agreed moved to implementation. That is the point we requested for the repo.

Yes. Nothing much happened on mailing list, All we saw a few comments
from Thomas and which ended up as NACK.

> 
> The Cavium comments came bit late.
> 
> Currently I have just consolidated the patches in the draft repo and I am
> going rebase it and post to mailing list as well in next 1-2 days.

OK. We will review it.

> 
> Since, the implementation is started, we will request any subsequent
> comments as an incremental patches.
> 
> > Akhil,
> > 
> > Based on your v2 version, we could map a lot with our HW. However, there
> > are three top level quires for the further review.
> > 
> > 1) Some HW cannot offload all types of packets(like IP fragmented
> > packets) and/or there may have back pressure momentarily from IPSec offload
> > engine (like Queue is full) etc. So in that case what is the expected behavior
> > a) Is it an offload driver responsibility to take care of that or
> > b) Is it passed to application as encrypted packets(in case of inbound)
> > and the application has to take or of them.
> > 
> 
> It will depend on the HW capability. If the HW is not supporting the
> fragmented etc packets, they will come as an encrypted packed to the
> application and application need to take care of them.

OK

> 
> > 2) In case of inbound traffic, What is the packet format from offload
> > driver. i.e
> > a) Will ESP header will be removed from the packet area after the
> > decryption.
> > 
> It depend on the session action type. e.g. for inline crypto, the header
> will be intact. for inline proto, the headers will be removed.
> In any case, we need to improve the documentation.

I thought, we need to keep the header in both cases otherwise the
application may not able to check anti-replay if ESP header removed.

> 
> > 3) We have a few feature like, anti-replay check, SA expiry((byte/time)
> > notification, etc from HW/FW. So it is not clear from the specification
> > on the contract between between offload driver vs application
> > responsibility? Can you give some insight on that? Especially
> > the error notification scheme if it is an offload driver responsibility.
> > 
> 
> Anti-replay, SA expiry management is still in my todo  list.
> The responsibilities will depend on the amount of offloading the HW/FW is
> offering. The current intent is that SA management and expiry is being
> managed by the applicaiton. However, SA expiry event for byte based SA will
> be passed by the HW/FW to application.
> 
> In short, the current focus is covering the basic support  only. Rest will
> be incremental.


Makes sense. This is the hard part to solve in inline HW IPSec implementation.
I suggest to keep API experimental till we solve this hard problems which are
tightly coupled with HW capabilities.