RE: [PATCH] disable lock annotation with clang 3.4.2

[Raslan Darawsheh <rasland@nvidia.com>]

> -----Original Message-----
> From: Morten Brørup <mb@smartsharesystems.com>
> Sent: Monday, February 13, 2023 6:09 PM
> To: Bruce Richardson <bruce.richardson@intel.com>; David Marchand
> <david.marchand@redhat.com>
> Cc: dev@dpdk.org; NBU-Contact-Thomas Monjalon (EXTERNAL)
> <thomas@monjalon.net>; Raslan Darawsheh <rasland@nvidia.com>;
> Chenbo Xia <chenbo.xia@intel.com>; Maxime Coquelin
> <maxime.coquelin@redhat.com>
> Subject: RE: [PATCH] disable lock annotation with clang 3.4.2
> 
> > From: Bruce Richardson [mailto:bruce.richardson@intel.com]
> > Sent: Monday, 13 February 2023 15.55
> >
> > On Mon, Feb 13, 2023 at 03:44:55PM +0100, David Marchand wrote:
> > > Venerable RHEL7 clang 3.4.2 has (at least) two issues with lock
> > > annotations.
> > >
> > > A first one with regards to the attribute position:
> > > ../lib/vhost/vhost.h:518:2: error: GCC does not allow
> > > 	assert_exclusive_lock attribute in this position on a
> > > 	function definition [-Werror,-Wgcc-compat]
> > >         __rte_assert_exclusive_lock(&vq->access_lock)
> > >         ^
> > > ../lib/eal/include/rte_lock_annotations.h:29:38: note: expanded
> > > 	from macro '__rte_assert_exclusive_lock'
> > >         __attribute__((assert_exclusive_lock(__VA_ARGS__)))
> > >                                             ^
> > >
> > > This can be worked around by splitting and having the allocation on
> > the
> > > function declaration.
> > >
> > > But on the other hand, clang 3.4.2 does not seem to propagate those
> > > annotations in presence of a __builtin_expect (i.e. unlikely()), like
> > > for example when calling if (unlikely(rte_spinlock_trylock() == 0)).
> > >
> > > Those annotations were only working with clang in any case, so
> > restrict
> > > to clang versions newer than 3.5.0.
> > >
> > > Fixes: 657a98f38940 ("eal: annotate spinlock, rwlock and seqlock")
> > >
> > > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > > ---
> > >  drivers/meson.build | 2 +-
> > >  lib/meson.build     | 2 +-
> > >  2 files changed, 2 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/meson.build b/drivers/meson.build
> > > index bddc4a6cc4..0618c31a69 100644
> > > --- a/drivers/meson.build
> > > +++ b/drivers/meson.build
> > > @@ -168,7 +168,7 @@ foreach subpath:subdirs
> > >          enabled_drivers += name
> > >          lib_name = '_'.join(['rte', class, name])
> > >          cflags += '-DRTE_LOG_DEFAULT_LOGTYPE=' +
> > '.'.join([log_prefix, name])
> > > -        if annotate_locks and cc.has_argument('-Wthread-safety')
> > > +        if annotate_locks and cc.get_id() == 'clang' and
> > cc.version().version_compare('>=3.5.0')
> > >              cflags += '-DRTE_ANNOTATE_LOCKS'
> > >              cflags += '-Wthread-safety'
> > >          endif
> >
> > Are we likely to see any issues with this with any other compilers?
> > Should
> > we look to do a built-test in meson to determine feature support rather
> > than checking clang versions explicitly?
> >
> > On the plus side, checking clang version like this makes it clear when
> > we
> > can drop the conditional.
> 
> I prefer this over auto-detection in meson, for the reason mentioned by
> Bruce.
> 
> Furthermore, different compilers may use different syntax in the code, so it
> is impossible to detect generically; the auto-detection would be per compiler
> anyway.
> 
> Acked-by: Morten Brørup <mb@smartsharesystems.com>
Tested-by: Raslan Darawsheh <rasland@nvidia.com>

Kindest regards,
Raslan Darawsheh