[dpdk-dev] [PATCH v2] mbuf: replace zero-length marker with unnamed union

[Bruce Richardson]

On Wed, Mar 11, 2020 at 10:04:33AM +0100, Morten Brørup wrote:
> > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Gavin Hu
> > Sent: Wednesday, March 11, 2020 8:50 AM
> > 
> > Hi Morten,
> > 
> > > From: Morten Brørup <mb at smartsharesystems.com>
> > > Sent: Monday, March 9, 2020 9:31 PM
> > >
> > > > From: dev [mailto:dev-bounces at dpdk.org] On Behalf Of Ferruh Yigit
> > > > Sent: Monday, March 9, 2020 12:30 PM
> > > >
> > > > On 3/9/2020 9:45 AM, Gavin Hu wrote:
> > > > > Hi Ferruh,
> > > > >
> > > > >> -----Original Message-----
> > > > >> From: Ferruh Yigit <ferruh.yigit at intel.com>
> > > > >> Sent: Monday, March 9, 2020 4:55 PM
> > > > >>
> > > > >> On 3/7/2020 3:56 PM, Gavin Hu wrote:
> > > > >>> Declaring zero-length arrays in other contexts, including as
> > > > interior
> > > > >>> members of structure objects or as non-member objects, is
> > > > discouraged.
> > > > >>> Accessing elements of zero-length arrays declared in such
> > contexts
> > > > is
> > > > >>> undefined and may be diagnosed.[1]
> > > > >>>
> > > > >>> Fix by using unnamed union and struct.
> > > > >>>
> > > > >>> https://bugs.dpdk.org/show_bug.cgi?id=396
> > > > >>>
> > > > >>> Bugzilla ID: 396
> > > > >>>
> > > > >>> [1] https://gcc.gnu.org/onlinedocs/gcc/Zero-Length.html
> > > > >>>
> > > > >>> Fixes: 3e6181b07038 ("mbuf: use structure marker from EAL")
> > > > >>> Cc: stable at dpdk.org
> > > > >>>
> > > > >>> Signed-off-by: Gavin Hu <gavin.hu at arm.com>
> > > > >>> ---
> > > > >>> v2:
> > > > >>> * change 'uint64_t rearm_data' to 'uint_64_t rearm_data[1]' to
> > fix
> > > > >>>   the SFC PMD compiling error on x86. <Kevin Traynor>
> > > > >>> ---
> > > > >>>  lib/librte_mbuf/rte_mbuf_core.h | 54 +++++++++++++++++++------
> > ----
> > > > ----
> > > > >>>  1 file changed, 32 insertions(+), 22 deletions(-)
> > > > >>>
> > > > >>> diff --git a/lib/librte_mbuf/rte_mbuf_core.h
> > > > >> b/lib/librte_mbuf/rte_mbuf_core.h
> > > > >>> index b9a59c879..34cb152e2 100644
> > > > >>> --- a/lib/librte_mbuf/rte_mbuf_core.h
> > > > >>> +++ b/lib/librte_mbuf/rte_mbuf_core.h
> > > > >>> @@ -480,31 +480,41 @@ struct rte_mbuf {
> > > > >>>  		rte_iova_t buf_physaddr; /**< deprecated */
> > > > >>>  	} __rte_aligned(sizeof(rte_iova_t));
> > > > >>>
> > > > >>> -	/* next 8 bytes are initialised on RX descriptor rearm */
> > > > >>> -	RTE_MARKER64 rearm_data;
> > > > >>> -	uint16_t data_off;
> > > > >>> -
> > > > >>> -	/**
> > > > >>> -	 * Reference counter. Its size should at least equal to the
> > size
> > > > >>> -	 * of port field (16 bits), to support zero-copy broadcast.
> > > > >>> -	 * It should only be accessed using the following
> > functions:
> > > > >>> -	 * rte_mbuf_refcnt_update(), rte_mbuf_refcnt_read(), and
> > > > >>> -	 * rte_mbuf_refcnt_set(). The functionality of these
> > functions
> > > > (atomic,
> > > > >>> -	 * or non-atomic) is controlled by the
> > > > >> CONFIG_RTE_MBUF_REFCNT_ATOMIC
> > > > >>> -	 * config option.
> > > > >>> -	 */
> > > > >>>  	RTE_STD_C11
> > > > >>>  	union {
> > > > >>> -		rte_atomic16_t refcnt_atomic; /**< Atomically
> > accessed
> > > > >> refcnt */
> > > > >>> -		/** Non-atomically accessed refcnt */
> > > > >>> -		uint16_t refcnt;
> > > > >>> -	};
> > > > >>> -	uint16_t nb_segs;         /**< Number of segments. */
> > > > >>> +		/* next 8 bytes are initialised on RX descriptor
> > rearm */
> > > > >>> +		uint64_t rearm_data[1];
> > > > >> We are using zero length array as markers only and know what we
> > are
> > > > doing
> > > > >> with them,
> > > > >> what would you think disabling the warning instead of increasing
> > the
> > > > >> complexity
> > > > >> in mbuf struct?
> > > > > Okay, I will add -Wno-zero-length-bounds to the compiler
> > toolchain
> > > > flags.
> > > >
> > > > This would be my preference but I would like to get more input, can
> > you
> > > > please
> > > > for more comments before changing the implementation in case there
> > are
> > > > some
> > > > strong opinion on it?
> > > >
> > >
> > > I have some input to this discussion.
> > >
> > > Let me repeat what Gavin's GCC reference states: Declaring zero-
> > length
> > > arrays [...] as interior members of structure objects [...] is
> > discouraged.
> > >
> > > Why would we do something that the compiler documentation says is
> > > discouraged? I think the problem (i.e. using discouraged techniques)
> > should
> > > be fixed, not the symptom (i.e. getting warnings about using
> > discouraged
> > > techniques).
> > >
> > > Compiler warnings are here to help, and in my experience they are
> > actually
> > > very helpful, although avoiding them often requires somewhat more
> > > verbose source code. Disabling this warning not only affects this
> > file, but
> > > disables warnings about potential bugs in other source code too.
> > >
> > > Generally, disabling compiler warnings is a slippery slope. It would
> > be
> > > optimal if DPDK could be compiled with -Wall, and it would probably
> > reduce
> > > the number of released bugs too.
> > >
> > > With that said, sometimes the optimal solution has to give way for
> > the
> > > practical solution. And this is a core file, so we should thread
> > lightly.
> > >
> > >
> > > As for an alternative solution, perhaps we can get rid of the MARKERs
> > in the
> > > struct and #define them instead. Not as elegant as Gavin's suggested
> > union
> > > based solution, but it might bring inspiration...
> > >
> > > struct rte_mbuf {
> > >     ...
> > >     } __rte_aligned(sizeof(rte_iova_t));
> > >
> > >     uint16_t data_off;
> > >     ...
> > > }
> > >
> > > #define rte_mbuf_rearm_data(m) ((uint64_t *)m->data_off)
> > 
> > This does not work out, it generates new errors:
> > /root/dpdk/build/include/rte_mbuf_core.h:485:33: error: dereferencing
> > type-punned pointer will break strict-aliasing rules [-Werror=strict-
> > aliasing]
> >   485 | #define rte_mbuf_rearm_data(m) ((uint64_t *)&m->data_off)
> > 
> 
> OK. Then Bruce's suggestion probably won't work either.
> 
> I found this article about strict aliasing: https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8
> 
> The article basically says that the union based method (i.e. your original suggestion) is valid C (but not C++) and is the common solution.
> 
> Alternatives have now been discussed and tested, so we should all support your original suggestion, which seems to be the only correct and viable solution.
> 
> Please go ahead with that, and then someone should update the SFC PMD accordingly.
> 
> Furthermore, I think that Stephen's suggestion about getting rid of the markers all together is good thinking, but it would require updating a lot of PMDs accordingly. So please also consider removing other markers that can be removed without affecting a whole bunch of other files.
> 

Does it still give errors if we don't have the cast in the macro?