Re: [dpdk-dev] [dpdk-dev v9 1/4] cryptodev: add crypto data-path service APIs

["Zhang, Roy Fan" <roy.fan.zhang@intel.com>]

Hi Akhil,

> -----Original Message-----
> From: Akhil Goyal <akhil.goyal@nxp.com>
> Sent: Monday, September 21, 2020 1:00 PM
> To: Zhang, Roy Fan <roy.fan.zhang@intel.com>; dev@dpdk.org; Ananyev,
> Konstantin <konstantin.ananyev@intel.com>; Thomas Monjalon
> <thomas@monjalon.net>
> Cc: Trahe, Fiona <fiona.trahe@intel.com>; Kusztal, ArkadiuszX
> <arkadiuszx.kusztal@intel.com>; Dybkowski, AdamX
> <adamx.dybkowski@intel.com>; Bronowski, PiotrX
> <piotrx.bronowski@intel.com>; Anoob Joseph <anoobj@marvell.com>
> Subject: RE: [dpdk-dev v9 1/4] cryptodev: add crypto data-path service APIs
> 
> Hi Fan,
> 
> > > >
> > > > +/** Crypto data-path service types */
> > > > +enum rte_crypto_dp_service {
> > > > +	RTE_CRYPTO_DP_SYM_CIPHER_ONLY = 0,
> > > > +	RTE_CRYPTO_DP_SYM_AUTH_ONLY,
> > > > +	RTE_CRYPTO_DP_SYM_CHAIN,
> > > > +	RTE_CRYPTO_DP_SYM_AEAD,
> > > > +	RTE_CRYPTO_DP_N_SERVICE
> > > > +};
> > >
> > > Comments missing for this enum.
> > > Do we really need this enum?
> > > Can we not have this info in the driver from the xform list?
> > > And if we really want to add this, why to have it specific to raw data path
> APIs?
> > >
> > Will add comments to this enum.
> > Unless the driver will store xform data in certain way (in fact QAT has it) the
> > driver may not know which data-path to choose from.
> > The purpose of having this enum is that the driver knows to attach the
> correct
> > handler into the service data structure fast.
> >
> I believe all drivers are storing that information already in some way in the
> session private data.
> This enum is maintained inside driver as of current implementation. This is
> not specific to raw
> Data path APIs. If you are introducing this enum in library, then it should be
> generic for the legacy
> Case as well.
>
[Fan: I am not sure other drivers (that's part of the reason why it is here), but indeed QAT does, Ok] 
 
> 
> > >
> > > > +union rte_crypto_sym_additional_data {
> > > > +	struct {
> > > > +		void *cipher_iv_ptr;
> > > > +		rte_iova_t cipher_iv_iova;
> > > > +		void *auth_iv_ptr;
> > > > +		rte_iova_t auth_iv_iova;
> > > > +		void *digest_ptr;
> > > > +		rte_iova_t digest_iova;
> > > > +	} cipher_auth;
> > >
> > > Should be chain instead of cipher_auth
> > This field is used for cipher only, auth only, or chain use-cases so I believe
> this is
> > a better name for it.
> 
> Agreed that this struct will be used for all 3 cases, that is what is happening in
> Other crypto cases. We use chain for all these three cases in legacy codepath.
> Chain can be of one or two xforms and ordering can be anything -
> Cipher only, auth only, cipher auth and auth cipher.
> 
> 
> > >
> > > > +	struct {
> > > > +		void *iv_ptr;
> > > > +		rte_iova_t iv_iova;
> > > > +		void *digest_ptr;
> > > > +		rte_iova_t digest_iova;
> > > > +		void *aad_ptr;
> > > > +		rte_iova_t aad_iova;
> > > > +	} aead;
> > > > +};
> > > > +
> > > >  /**
> > > >   * Synchronous operation descriptor.
> > > >   * Supposed to be used with CPU crypto API call.
> > > > @@ -57,12 +81,25 @@ struct rte_crypto_sgl {
> > > >  struct rte_crypto_sym_vec {
> > > >  	/** array of SGL vectors */
> > > >  	struct rte_crypto_sgl *sgl;
> > > > -	/** array of pointers to IV */
> > > > -	void **iv;
> > > > -	/** array of pointers to AAD */
> > > > -	void **aad;
> > > > -	/** array of pointers to digest */
> > > > -	void **digest;
> > > > +
> > > > +	union {
> > > > +
> > > > +		/* Supposed to be used with CPU crypto API call. */
> > > > +		struct {
> > > > +			/** array of pointers to IV */
> > > > +			void **iv;
> > > > +			/** array of pointers to AAD */
> > > > +			void **aad;
> > > > +			/** array of pointers to digest */
> > > > +			void **digest;
> > > > +		};
> > >
> > > Can we also name this struct?
> > > Probably we should split this as a separate patch.
> > [Then this is an API break right?]
> 
> Since this an LTS release, I am ok to take this change.
> But others can comment on this.
> @Ananyev, Konstantin, @Thomas Monjalon
> Can you comment on this?
> 
> > >
> > > > +
> > > > +		/* Supposed to be used with
> > > > rte_cryptodev_dp_sym_submit_vec()
> > > > +		 * call.
> > > > +		 */
> > > > +		union rte_crypto_sym_additional_data *additional_data;
> > > > +	};
> > > > +
> > >
> > > Can we get rid of this unnecessary union
> rte_crypto_sym_additional_data
> > > And place chain and aead directly in the union? At any point, only one of
> the
> > > three
> > > would be used.
> > We have 2 main different uses cases, 1 for cpu crypto and 1 for data path
> APIs.
> > Within each main uses case there are 4 types of algo (cipher only/auth
> > only/aead/chain), one requiring HW address and virtual address, the other
> > doesn't.
> > It seems to causing too much confusion to include these many union into
> the
> > structure that initially was designed for cpu crypto only.
> > I suggest better to use different structure than squeeze all into a big union.
> >
> 
> IMO, the following union can clarify all doubts.
> @Ananyev, Konstantin: Any suggestions from your side?
> 
> /** IV and aad information for various use cases. */
> union {
>         /** Supposed to be used with CPU crypto API call. */
>         struct {
>                 /** array of pointers to IV */
>                 void **iv;
>                 /** array of pointers to AAD */
>                 void **aad;
>                 /** array of pointers to digest */
>                 void **digest;
>         } cpu_crypto;  < or any other useful name>
>         /* Supposed to be used with HW raw crypto API call. */
>         struct {
>                 void *cipher_iv_ptr;
>                 rte_iova_t cipher_iv_iova;
>                 void *auth_iv_ptr;
>                 rte_iova_t auth_iv_iova;
>                 void *digest_ptr;
>                 rte_iova_t digest_iova;
>         } hw_chain;
>         /* Supposed to be used with HW raw crypto API call. */
>         struct {
>                 void *iv_ptr;
>                 rte_iova_t iv_iova;
>                 void *digest_ptr;
>                 rte_iova_t digest_iova;
>                 void *aad_ptr;
>                 rte_iova_t aad_iova;
>         } hw_aead;
> };
> 
> 
[Structure looks good to me thanks!] 
> 
> > > > +/**
> > > > + * Context data for asynchronous crypto process.
> > > > + */
> > > > +struct rte_crypto_dp_service_ctx {
> > > > +	void *qp_data;
> > > > +
> > > > +	struct {
> > > > +		cryptodev_dp_submit_single_job_t submit_single_job;
> > > > +		cryptodev_dp_sym_submit_vec_t submit_vec;
> > > > +		cryptodev_dp_sym_operation_done_t submit_done;
> > > > +		cryptodev_dp_sym_dequeue_t dequeue_opaque;
> > > > +		cryptodev_dp_sym_dequeue_single_job_t dequeue_single;
> > > > +		cryptodev_dp_sym_operation_done_t dequeue_done;
> > > > +	};
> > > > +
> > > > +	/* Driver specific service data */
> > > > +	__extension__ uint8_t drv_service_data[];
> > > > +};
> > >
> > > Comments missing for structure params.
> > > Struct name can be rte_crypto_raw_dp_ctx.
> > >
> > > Who allocate and free this structure?
> > Same as crypto session, the user need to query the driver specific service
> data
> > Size and allocate the buffer accordingly. The difference is it does not have
> to
> > Be from mempool as it can be reused.
> 
> So this structure is saved and filled by the lib/driver and not the application.
> Right?
> This struct is opaque to application and will be part of session private data.
> Right?
> Assignment and calling appropriate driver's call backs will be hidden inside
> library
> and will be opaque to the application. In other words, the structure is not
> exposed
> to the application.
> Please add relevant comments on top of this structure.
> 
[Fan: will do] 
> 
> > > > +static __rte_always_inline int
> > > > +_cryptodev_dp_sym_dequeue_single_job(struct
> > > rte_crypto_dp_service_ctx
> > > > *ctx,
> > > > +		void **out_opaque)
> > > > +{
> > > > +	return (*ctx->dequeue_single)(ctx->qp_data, ctx->drv_service_data,
> > > > +		out_opaque);
> > > > +}
> > > > +
> > > > +/**
> > > > + * Submit single job into device queue but the driver will not start
> > > > + * processing until rte_cryptodev_dp_submit_done() is called. This is a
> > > > + * simplified
> 
> Comment not complete.
> 
> Regards,
> Akhil