RE: [PATCH v7 4/4] eal: add nonnull and access function attributes

["Morten Brørup" <mb@smartsharesystems.com>]

> From: David Marchand [mailto:david.marchand@redhat.com]
> Sent: Tuesday, 31 January 2023 12.15
> 
> On Wed, Jan 18, 2023 at 9:31 AM Morten Brørup
> <mb@smartsharesystems.com> wrote:
> >
> > +To: Thomas & David, you probably have some opinions on this too!
> >
> > > From: Tyler Retzlaff [mailto:roretzla@linux.microsoft.com]
> > > Sent: Tuesday, 17 January 2023 22.17
> > >
> > > On Tue, Jan 17, 2023 at 09:19:22AM +0100, Morten Brørup wrote:
> > > > > From: Ferruh Yigit [mailto:ferruh.yigit@amd.com]
> > > > > Sent: Monday, 16 January 2023 18.02
> > > > >
> > > > > On 1/16/2023 1:07 PM, Morten Brørup wrote:
> > > > > > Add nonnull function attribute to help the compiler detect a
> NULL
> > > > > > pointer being passed to a function not accepting NULL
> pointers as
> > > an
> > > > > > argument at build time.
> > > > > >
> > > > > > Add access function attributes to tell the compiler how a
> > > function
> > > > > > accesses memory pointed to by its pointer arguments.
> > > > > >
> > > > > > Add these attributes to the rte_memcpy() function, as the
> first
> > > in
> > > > > > hopefully many to come.
> > > > > >
> > > > > > Signed-off-by: Morten Brørup <mb@smartsharesystems.com>
> > > > > > Acked-by: Tyler Retzlaff <roretzla@linux.microsoft.com>
> > > > > > Reviewed-by: Ruifeng Wang <ruifeng.wang@arm.com>
> > > > > > ---
> > > >
> > > > [...]
> > > >
> > > > > > +/**
> > > > > > + * Tells compiler that the pointer arguments must be non-
> null.
> > > > > > + *
> > > > > > + * @param ...
> > > > > > + *    Comma separated list of parameter indexes of pointer
> > > > > arguments.
> > > > > > + */
> > > > > > +#if defined(RTE_CC_GCC) || defined(RTE_CC_CLANG)
> > > > > > +#define __rte_nonnull_params(...) \
> > > > > > +       __attribute__((nonnull(__VA_ARGS__)))
> > > > > > +#else
> > > > > > +#define __rte_nonnull_params(...)
> > > > > > +#endif
> > > > > > +
> > > > >
> > > > > What do you think to have a namespace for macros like
> > > > > '__rte_param_xxx',
> > > > > so name macros as:
> > > > > __rte_param_nonull
> > > > > __rte_param_read_only
> > > > > __rte_param_write_only
> > > > >
> > > > > No strong opinion, it just feels tidier this way
> > > >
> > > > Being a proponent of the world_country_city naming scheme myself,
> I
> > > would usually agree with this proposal.
> > > >
> > > > However, in the future, we might add macros without _param for
> use
> > > along with the function parameters, e.g.:
> > > >
> > > > int foo(int bar __rte_nonnull __rte_read_only);
> > > >
> > > > So I decided for this order in the names (treating
> > > nonnull/access_mode as "country" and param/params as "city"), also
> > > somewhat looking at the __rte_deprecated and
> __rte_deprecated_msg(msg)
> > > macros.
> > > >
> > > > I have no strong preference either, so if anyone does, please
> speak
> > > up.
> > > >
> > > > Slightly related, also note this:
> > > >
> > > > The nonnull macro is plural (_params), because it can take
> multiple
> > > pointer parameter indexes.
> > > > The access mode macros are singular (_param), because they only
> take
> > > one pointer parameter index, and the optional size parameter index.
> > > >
> > > > I considered splitting up the access mode macros even more,
> making
> > > two variants of each, e.g. __rte_read_only_param(ptr_index) and
> > > __rte_read_only_param_size(ptr_index, size_index), but concluded
> that
> > > it would be excruciatingly verbose. The only purpose would be to
> reduce
> > > the risk of using them incorrectly. I decided against it, thinking
> that
> > > any developer clever enough to use these macros is also clever
> enough
> > > to understand how to use them (or at least read their parameter
> > > descriptions to learn how).
> > > >
> > >
> > > microsoft also has a tool & annotation vehicle for this type of
> stuff.
> > > this discussion has caused me to wonder what happens if we would
> like
> > > to
> > > add additional annotations for other tools. just load on the
> > > annotations
> > > and expand them empty conditionally?
> > >
> > > https://learn.microsoft.com/en-us/cpp/code-quality/using-sal-
> > > annotations-to-reduce-c-cpp-code-defects?view=msvc-170
> > >
> > > anyway, just a thought. no serious response required here.
> >
> > Excellent input, Tyler!
> >
> > If we want DPDK to be considered truly cross-platform, and not treat
> non-Linux/non-GCC as second class citizens, we need to discuss this.
> >
> > Microsoft's Source Code Annotation Language (SAL) seems very good,
> based on its finer granularity than GCC's attributes (which in
> comparison seem added as an afterthought, not cleanly structured like
> SAL). I have only skimmed the documentation, but that is my immediate
> impression of it.
> >
> > SAL uses a completely different syntax than GCC attributes, and
> Microsoft happens to also use memcpy() as an example in the
> documentation referred to:
> >
> > void * memcpy(
> >    _Out_writes_bytes_all_(count) void *dest,
> >    _In_reads_bytes_(count) const void *src,
> >    size_t count
> > );

Stephen had bad experiences with SAL, so let's just consider the SAL memcpy() example a reference only, showing how the syntax of annotations can differ very much between build systems.

> >
> > Going back to how we can handle this in DPDK, we can either:
> >
> > 1. Not annotate the functions at all, and miss out on finding the
> errors for us.
> 
> Seeing how clang safety checks helped me catch bugs, that would be a
> pity.
> 
> >
> > 2. Invent our own language (or find something existing) for function
> headers, and use a parser to convert them to compiler specific C/C++
> headers when building the code.
> 
> Argh, no.
> 
> >
> > 3a. Keep loading on attributes, with empty macros for unsupported
> compilers.
> >
> > 3b. Keep loading on attributes, with empty macros for unsupported
> compilers. But limit ourselves to GCC/Clang style attributes.
> >
> > 3c. Keep loading on attributes, with empty macros for unsupported
> compilers. But limit ourselves to Microsoft SAL style attributes.
> >
> > 3d. Keep loading on attributes, with empty macros for unsupported
> compilers. But limit ourselves to the most relevant attributes, using
> performance and/or bug detection as criteria when considering
> relevance.
> >
> > I am strongly against both 1 and 2.
> >
> > If bug detection is the primary driver, we could stick with either 3b
> or 3c (i.e. only target one specific build environment) and rely on the
> DPDK CI for detecting bugs. But then application developers would not
> benefit, because they don't run their code through the DPDK CI. So I am
> also against this.
> >
> > I think 3d (keep loading on attributes, but only the most relevant
> ones) is the best choice.
> >
> > GCC/Clang style attributes are already supported as macros prefixed
> by __rte, so let's not change the way we do that.
> >
> > Regarding the Microsoft SAL, I suppose Microsoft already chose
> annotation names to avoid collisions, so we could consider using those
> exact names (i.e. without __rte prefix), and define empty macros for
> non-Microsoft compilers. This would allow using Microsoft SAL
> annotations directly in the DPDK code.
> >
> 
> I have a bit of trouble understanding the difference between 3a and
> 3d.. 3a would be about accepting any annotation?

Correct.

> 
> 3d is the best option as it is not changing anything to what we were
> doing so far: we evaluate the pros and cons of each annotations/tools,
> case per case.

Agree!