RE: [EXTERNAL] [PATCH v2 1/2] crypto: fix build issues on unsetting crypto callbacks macro

[Akhil Goyal <gakhil@marvell.com>]

> > > From: Kundapura, Ganapati [mailto:ganapati.kundapura@intel.com]
> > > Sent: Thursday, 30 May 2024 16.22
> > >
> > > Hi,
> > >
> > > > From: Akhil Goyal <gakhil@marvell.com>
> > > > Sent: Thursday, May 30, 2024 5:17 PM
> > > >
> > > > > > > #if may not be needed in application.
> > > > > > > Test should be skipped if API is not available/supported.
> > > > > > >
> > > > > It's needed otherwise application developer has to check the
> > > > > implementation for supported/not supported or else run the application
> > > > > to get to know whether api is supported or not.
> > > > >
> > > >
> > > > Application is always required to check the return value or else it will
> > > miss the
> > > > other errors that the API can return.
> > > Currently RTE_CRYPTO_CALLBACKS is enabled by default and test application
> > > checks the
> > > return value of the APIs. This patch fixes build issues on compiling the DPDK
> > > with unsetting
> > > RTE_CRYPTO_CALLBACKS.
> > > >
> > > > > > > > diff --git a/lib/cryptodev/rte_cryptodev.c
> > > > > > > > b/lib/cryptodev/rte_cryptodev.c index 886eb7a..2e0890f 100644
> > > > > > > > --- a/lib/cryptodev/rte_cryptodev.c
> > > > > > > > +++ b/lib/cryptodev/rte_cryptodev.c
> > > > > > > > @@ -628,6 +628,7 @@
> > > > > > rte_cryptodev_asym_xform_capability_check_hash(
> > > > > > > >  	return ret;
> > > > > > > >  }
> > > > > > > >
> > > > > > > > +#if RTE_CRYPTO_CALLBACKS
> > > > > > > >  /* spinlock for crypto device enq callbacks */  static
> > > > > > > > rte_spinlock_t rte_cryptodev_callback_lock =
> > > > > > > RTE_SPINLOCK_INITIALIZER;
> > > > > > > >
> > > > > > > > @@ -744,6 +745,7 @@ cryptodev_cb_init(struct rte_cryptodev
> *dev)
> > > > > > > >  	cryptodev_cb_cleanup(dev);
> > > > > > > >  	return -ENOMEM;
> > > > > > > >  }
> > > > > > > > +#endif /* RTE_CRYPTO_CALLBACKS */
> > > > > > >
> > > > > > >
> > > > > > > > @@ -1485,6 +1491,7 @@ rte_cryptodev_queue_pair_setup(uint8_t
> > > > > > dev_id,
> > > > > > > > uint16_t queue_pair_id,
> > > > > > > >  			socket_id);
> > > > > > > >  }
> > > > > > > >
> > > > > > > > +#if RTE_CRYPTO_CALLBACKS
> > > > > > > >  struct rte_cryptodev_cb *
> > > > > > > >  rte_cryptodev_add_enq_callback(uint8_t dev_id,
> > > > > > > >  			       uint16_t qp_id,
> > > > > > > > @@ -1763,6 +1770,7 @@
> > > > rte_cryptodev_remove_deq_callback(uint8_t
> > > > > > dev_id,
> > > > > > > >  	rte_spinlock_unlock(&rte_cryptodev_callback_lock);
> > > > > > > >  	return ret;
> > > > > > > >  }
> > > > > > > > +#endif /* RTE_CRYPTO_CALLBACKS */
> > > > > > >
> > > > > > > There is an issue here.
> > > > > > > The APIs are visible in .h file and are available for application to
> > > use.
> > > > > > > But the API implementation is compiled out.
> > > > > > > Rather, you should add a return ENOTSUP from the beginning of the
> > > > > > > APIs if RTE_CRYPTO_CALLBACKS  is enabled.
> > > > > > > With this approach application will not need to put #if in its code.
> > > > > API declarations wrapped under the macro changes in next patch.
> > > >
> > > > No, that is not the correct way. Application should check the return value.
> > > > And we cannot force it to add ifdefs.
> > > Test application is indeed checking the return value. Ifdefs are added to
> > > avoid build issues on compiling with RTE_CRYPTO_CALLBACKS is turned off
> > > Which is on by default.
> >
> > The test application should be able to build and run, regardless if the DPDK
> library
> > was built with RTE_CRYPTO_CALLBACKS defined or not.
> >
> > The test application should not assume that the DPDK library was built with the
> > same RTE_CRYPTO_CALLBACKS configuration (i.e. defined or not) as the test
> > application.
> >
> > > Even ethdev callbacks also doesn't return -ENOTSUP
> > > on setting/unsetting RTE_ETHDEV_RXTX_CALLBACKS config.
> >
> > That would be a bug in the ethdev library.
> > I just checked the ethdev source code (/source/lib/ethdev/rte_ethdev.c), and
> all
> > the add/remove rx/tx callback functions fail with ENOTSUP if
> > RTE_ETHDEV_RXTX_CALLBACKS is not defined.
> > Please note that some ethdev callbacks are not rx/tx callbacks, and thus are not
> > gated by RTE_ETHDEV_RXTX_CALLBACKS.
> 
> Hi Ganapati,
> Can you send a new version incorporating above comments and
> work on similar lines as ethdev is currently doing.
> 
> I believe as Morten pointed out, use of ifdef is as per DPDK convention,
> So better move it that way.
> We can discuss later if we can incorporate these in meson options.
> 
Any update on this?