[ovs-dev] [PATCH] flex-array: allow arrays of unions with flexible members.

[Luc Van Oostenryck]

On Thu, Oct 08, 2020 at 12:04:55PM +0200, Ilya Maximets wrote:
> On 10/8/20 11:13 AM, Ilya Maximets wrote:
> > On 10/8/20 1:15 AM, Luc Van Oostenryck wrote:
> >> On Wed, Oct 07, 2020 at 01:52:34PM +0200, Ilya Maximets wrote:
> >>> The actual code in question that makes sparse fail
> >>> OVS build could be found here:
> >>>   https://github.com/openvswitch/ovs/blob/39fbd2c3f0392811689ec780f09baf90faceb877/lib/netdev-linux.c#L1238
> >>
> >> I'm impressed and surprised you're using of includes just for Sparse.
> >> I also see that this is since 2011. Just for my curiosity, have
> >> you an idea for why exactly this was needed and if it is still
> >> really needed?
> > 
> > There are some Sparse-related headers that could be safely removed now
> > because required functionality is already supported by Sparse.  We need
> > to clean this up someday, e.g. OVS builds fine without
> > include/sparse/threads.h since Sparse 0.5.1.
> > 
> > However, there are headers that are necessary for successful build.
> > There are few classes of issues that these headers are targeted on:
> > 
> > 1. Missing functionality in Sparse.
> > For example, it doesn't know about __builtin_ia32_pause, so we have to have
> > include/sparse/xmmintrin.h.
> > Sparse also doesn't know __atomic_load_n that comes from some DPDK headers.
> > OVS itself avoids using builtin atomics if __CHECKER__ defined.
> > DPDK library also has some issues with types in __sync_add_and_fetch, but I
> > do not remember exact problem.  We have include/sparse/rte_atomic.h for it.
> > These are from the top of my head.  I could go through our specific headers
> > and make a list of missing features someday if you're interested.

OK, I see. I'll add __builtin_ia32_pause() now and I'll look at
__atomic_load_n(), __sync_add_and_fetch() and friends soon.

> > 2. Sparse complains on standard libraries.
> > Complains on PTHREAD_MUTEX_INITIALIZER: 
> >     error: Using plain integer as NULL pointer
> > So, we have to have include/sparse/pthread.h.
> > Maybe some other examples, but I do not remember right now.
> > 
> > 3. Issues inside external libraries.
> > Ex. numa.h header from libnuma contains non-ANSI function declarations.
> > So we have include/sparse/numa.h.

It's always possible to use -Wno-non-pointer-null and
-Wno-old_-style-definition but ...

> > 4. Issues with restricted types (these are heaviest).
> > OVS uses restricted types like 'ovs_be32' inside (with __attribute__((bitwise))),
> > but standard functions like htonl() operates with usual 'uint32_t'.  While it's
> > safe to use these exact functions, Sparse complains about type mismatch.  One
> > option here is to explicitly cast arguments with (OVS_FORCE ovs_be32) each time,
> > but that is impractical (too many occurrences, ~4K lines of code only for hton/ntoh
> > conversions) and will make code less readable.  Much easier to override these
> > functions just for Sparse.  Ex. include/sparse/netinet/in.h.
> > 
> > Similar issue with some data types that goes from external libraries and system
> > headers. e.g. DPDK library operates with its own types like rte_be32_t.  While
> > it's completely safe to mix them with ovs_be32, Sparse doesn't know about that,
> > because DPDK doesn't mark them as bitwise.  This issue might be fixed on DPDK
> > side, I guess.  But Sparse will complain about different types even if these types
> > defined in exactly same way.  e.g. following test fails:
> > 
> > diff --git a/validation/bitwise-cast.c b/validation/bitwise-cast.c
> > index 0583461c..9284bd05 100644
> > --- a/validation/bitwise-cast.c
> > +++ b/validation/bitwise-cast.c
> > @@ -35,6 +35,18 @@ static __be32 quuy(void)
> >         return (__attribute__((force)) __be32) 1730;
> >  }
> >  
> > +/* Implicit casts of equally defined types, should be legal? */
> 
> I do understand why this is illegal.  Otherwise it will be not possible to
> create le and be types.  However, since the main purpose of 'bitwise' attribute,
> AFAIU, is to create le and be types, maybe it make sense to have 2 different
> attributes? e.g. 'bitwise_le' and 'bitwise_be'.  This way Sparse will be able
> to detect that 2 different types ovs_be32 and rte_be32_t could be safely mixed,
> if both defined with attribute 'bitwise_be' and has same base type uint32_t.

Well, one of the key characteristic of __bitwise is that it creates new,
distinct, incompatible types and it is very much used as such. For example,
in the kernel the 6 {be,le}{16,32,64} are maybe the most used but there
are about 100 other bitwise types that are defined (like poll_t or
pci_power_t) just to have a distinct type.

So, for the problem with rte_be32_t & ovs_be32, the solutions should be
to have a common definition, either via a macro or a typedef. Not an
easy thing with external libraries.
 
Best regards,
-- Luc