Re: out of tree driver builds broken with C++

[Stephen Hemminger <stephen@networkplumber.org>]

On Mon, 14 Feb 2022 18:49:48 +0100
Thomas Monjalon <thomas@monjalon.net> wrote:

> 14/02/2022 18:t24, Stephen Hemminger:
> > On Mon, 14 Feb 2022 15:03:43 +0000
> > Bruce Richardson <bruce.richardson@intel.com> wrote:
> >   
> > > On Mon, Feb 14, 2022 at 11:43:35AM +0000, Bruce Richardson wrote:  
> > > > On Mon, Feb 14, 2022 at 12:19:24PM +0100, Thomas Monjalon wrote:    
> > > > > 14/02/2022 11:45, Bruce Richardson:    
> > > > > > On Mon, Feb 14, 2022 at 10:22:08AM +0100, Thomas Monjalon wrote:    
> > > > > > > 14/02/2022 10:13, Tyler Retzlaff:    
> > > > > > > > while the driver api is "internal" we agreed some time ago that drivers
> > > > > > > > could be built external to the dpdk tree. by enabling the meson setup
> > > > > > > > option -Denable_driver_sdk=true.
> > > > > > > > 
> > > > > > > > it was agreed that the driver api was internal and would attract no
> > > > > > > > binary compatibility support which was fine.  this change has now
> > > > > > > > imposed a further restriction that out of tree drivers have to be
> > > > > > > > authored in C only as non-C++ compatible code will invariably leak into
> > > > > > > > the internal structures.
> > > > > > > > 
> > > > > > > > you won't allow us to build C++ drivers in the dpdk tree and it seems
> > > > > > > > now you are preventing building of C++ drivers outside of the tree too.    
> > > > > > > 
> > > > > > > That's the problem of non-written assumptions, they are unknown or forgotten.
> > > > > > > Did we agree to support out-of-tree drivers in C++?
> > > > > > > 
> > > > > > > We really need to make things clear and written in documentation.
> > > > > > >     
> > > > > > > > could we please re-evaluate this.    
> > > > > > > 
> > > > > > > Yes we can re-evaluate.
> > > > > > > What is the list of impacted files?
> > > > > > >     
> > > > > > Hacking meson files a bit, the list of SDK header files is reported as below.
> > > > > > 
> > > > > > /Bruce
> > > > > > 
> > > > > > Message: SDK headers: 
> > > > > > Message: ethdev_driver.h
> > > > > > Message: ethdev_pci.h
> > > > > > Message: ethdev_vdev.h
> > > > > > Message: cryptodev_pmd.h
> > > > > > Message: eventdev_pmd.h
> > > > > > Message: eventdev_pmd_pci.h
> > > > > > Message: eventdev_pmd_vdev.h
> > > > > > Message: eventdev_trace.h
> > > > > > Message: event_timer_adapter_pmd.h
> > > > > > Message: rte_dmadev_pmd.h
> > > > > > Message: vdpa_driver.h    
> > > > > 
> > > > > I see no harm in supporting C++ include of these headers.
> > > > > Any objection?
> > > > > 
> > > > > Could we have a test in chkincs for the SDK headers?
> > > > >     
> > > > Yes. It may make things a little more complicated, though, as it seems
> > > > these headers also have a tendency to rely on some driver headers -
> > > > specifically bus driver headers.
> > > >     
> > > 
> > > Working on a patch to add these to checks. However, just running a c++
> > > compile does not check for valid 'extern "C"' blocks. Adding the following
> > > to chkincs/meson.build as a basic sanity check throws up a larger list of
> > > files to be looked at.
> > > 
> > > Diff:
> > > 
> > > +# check for extern C in files, since this is not detected as an error by the compiler
> > > +grep = find_program('grep', required: false)
> > > +if grep.found()
> > > +    errlist = run_command([grep, '--files-without-match', '^extern "C"', dpdk_chkinc_headers],
> > > +            check: true, capture: true).stdout().split()
> > > +    if errlist != []
> > > +        error('Files missing C++ \'extern "C"\' guards:\n- ' + '\n- '.join(errlist))
> > > +    endif
> > > +endif
> > > +
> > > 
> > > Output:
> > > ../buildtools/chkincs/meson.build:45:8: ERROR: Problem encountered: Files missing C++ 'extern "C"' guards:
> > > - /home/bruce/dpdk.org/lib/telemetry/rte_telemetry.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_bitops.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_branch_prediction.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_compat.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_hypervisor.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_keepalive.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_pci_dev_feature_defs.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_pci_dev_features.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_time.h
> > > - /home/bruce/dpdk.org/lib/eal/include/rte_trace_point_register.h
> > > - /home/bruce/dpdk.org/lib/eal/linux/include/rte_os.h
> > > - /home/bruce/dpdk.org/lib/ethdev/rte_dev_info.h
> > > - /home/bruce/dpdk.org/lib/ethdev/ethdev_driver.h
> > > - /home/bruce/dpdk.org/lib/ethdev/ethdev_pci.h
> > > - /home/bruce/dpdk.org/lib/ethdev/ethdev_vdev.h
> > > - /home/bruce/dpdk.org/lib/metrics/rte_metrics_telemetry.h
> > > - /home/bruce/dpdk.org/lib/acl/rte_acl_osdep.h
> > > - /home/bruce/dpdk.org/lib/bpf/bpf_def.h
> > > - /home/bruce/dpdk.org/lib/compressdev/rte_compressdev_internal.h
> > > - /home/bruce/dpdk.org/lib/cryptodev/cryptodev_pmd.h
> > > - /home/bruce/dpdk.org/lib/eventdev/rte_event_ring.h
> > > - /home/bruce/dpdk.org/lib/eventdev/eventdev_pmd.h
> > > - /home/bruce/dpdk.org/lib/eventdev/eventdev_pmd_pci.h
> > > - /home/bruce/dpdk.org/lib/eventdev/eventdev_pmd_vdev.h
> > > - /home/bruce/dpdk.org/lib/kni/rte_kni_common.h
> > > - /home/bruce/dpdk.org/lib/vhost/rte_vdpa.h
> > > - /home/bruce/dpdk.org/lib/vhost/rte_vhost_async.h
> > > - /home/bruce/dpdk.org/lib/vhost/rte_vhost_crypto.h
> > > - /home/bruce/dpdk.org/lib/vhost/vdpa_driver.h
> > > 
> > > Regards,
> > > /Bruce  
> > 
> > The actual C++ guards are small and have no impact therefore we
> > should accept patches as needed. Having more guards than needed is
> > just noise.
> > 
> > But the responsibility of telling DPDK project where they are required
> > should fall to any out-of-tree users who care.  
> 
> Sorry Stephen I don't understand.
> Please could you rephrase your comment?

Since there are no in-tree DPDK drivers that use C++ it is an extra
effort to add infrastructure to test if headers are broken. As new
features get added to the DPDK, it doesn't make sense for the upstream
DPDK to guarantee that no driver API changed in a way that breaks C++.
But if a breakage happens in C++ driver, the upstream DPDK is willing
to accept patches to fix it.