Re: [dpdk-dev] [PATCH 2/3] mbuf: avoid cast-align warning in pktmbuf mtod offset macro

[Eli Britstein <elibr@nvidia.com>]

Hi Olivier,

On 8/1/2021 11:06 AM, Eli Britstein wrote:
>
> On 7/30/2021 2:10 PM, Olivier Matz wrote:
>> External email: Use caution opening links or attachments
>>
>>
>> Hi Eli,
>>
>> On Thu, Jul 29, 2021 at 10:13:45AM +0300, Eli Britstein wrote:
>>> On 7/28/2021 6:28 PM, Olivier Matz wrote:
>>>> External email: Use caution opening links or attachments
>>>>
>>>>
>>>> On Tue, Jul 13, 2021 at 09:49:09AM +0300, Eli Britstein wrote:
>>>>> In rte_pktmbuf_mtod_offset macro, there is a casting from char * 
>>>>> to type
>>>>> 't', which may cause cast-align warning when using gcc flags
>>>>> '-Werror -Wcast-align':
>>>>>
>>>>> .../include/rte_mbuf_core.h:723:3: error: cast increases required 
>>>>> alignment
>>>>>       of target type [-Werror=cast-align]
>>>>>     723 |  ((t)((char *)(m)->buf_addr + (m)->data_off + (o)))
>>>>>         |   ^
>>>>>
>>>>> As the code assumes correct alignment, add first a (void *) 
>>>>> casting, to
>>>>> avoid the warning.
>>>>>
>>>>> Fixes: af75078fece3 ("first public release")
>>>>> Cc: stable@dpdk.org
>>>>>
>>>>> Signed-off-by: Eli Britstein <elibr@nvidia.com>
>>>> My initial thinking was that it's the problem of the application: if
>>>> -Werror=cast-align is used, it is up to the application to cast the
>>>> return value of rte_pktmbuf_mtod_offset() to (void *) before 
>>>> casting it
>>>> to the network type.
>>>>
>>>> But, if I understand correctly, the problem is not about the 
>>>> application
>>>> code itself, but about inlined code in the header files of dpdk
>>>> (i.e. compiling an empty C file that just includes the dpdk headers 
>>>> with
>>>> -Werror=cast-align). Is it correct? If yes I think it should be
>>>> highlighted in the commit log.
>>> I think yes, though in this specific patch it is not even an inline
>>> function, but a macro.
>>>
>>> However, I don't have a synthetic application example to show those
>>> warnings, thus didn't put such in the commit msg.
>> For this patch, I think it would be useful to have a way to reproduce
>> the issue first, so we can check whether it is the proper place to fix
>> the problem.
> --- a/examples/l2fwd/Makefile
> +++ b/examples/l2fwd/Makefile
> @@ -22,6 +22,7 @@ static: build/$(APP)-static
>         ln -sf $(APP)-static build/$(APP)
>
>  PC_FILE := $(shell $(PKGCONF) --path libdpdk 2>/dev/null)
> +CFLAGS += -Wcast-align=strict
>  CFLAGS += -O3 $(shell $(PKGCONF) --cflags libdpdk)
>
> gcc (Ubuntu 9.3.0-10ubuntu2) 9.3.0
> Copyright (C) 2019 Free Software Foundation, Inc.
> This is free software; see the source for copying conditions. There is NO
> warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR 
> PURPOSE.
>
> make -C examples/l2fwd clean static

To reproduce locally with DPDK only, no need to change any file. Only run:

CFLAGS="-Wcast-align=strict" make V=1 -C examples/l2fwd clean static

How would you like to proceed?

Thanks,

Eli

>
>>
>> To me, it is assumed in the DPDK project that we can mmap a network
>> structure on mbuf data (maybe I'm wrong?). If an external application
>> like OVS wants to use -Werror=cast-align, it has to cast the result of
>> calls to rte_pktmbuf_mtod() family.
>>
>> The only corner cases are DPDK header files which have static inline
>> functions or macro that forces the use of rte_pktmbuf_mtod() family
>> without a cast (like for your patch 1/3), because it cannot be fixed in
>> the external project.
>>
>> I think we have to make our header files compliant to projects that want
>> to use -Werror=cast-align, like we do to make our header files compliant
>> to C++.
>>
>> What you suggest in this patch forces the cast to (void *) for all users
>> of rte_pktmbuf_mtod() family. This could be a problem for projects that
>> want to see these warnings.
>>
>> Would it be possible instead to add a cast in DPDK headers, in inline
>> functions that make use of these mtod functions?
>>
>> Regards,
>> Olivier
>>
>>
>>
>>>> Out of curiosity, how did you find the errors? I mean, is it possible
>>>> that some casts are missing some other headers, or is this patchset
>>>> exhaustive?
>>> Currently OVS-DPDK is compiled only with -Wno-cast-align.
>>>
>>> Following complaint that a recent commit introduced a degradation in 
>>> OVS
>>> [1], I compiled OVS without this warning deprecation.
>>> The fixes in OVS are [2] and [3] (already merged). The fixes in DPDK 
>>> are in
>>> this patch-set.
>>>
>>> [1] 
>>> https://mail.openvswitch.org/pipermail/ovs-dev/2021-July/385084.html
>>> [2] 
>>> https://mail.openvswitch.org/pipermail/ovs-dev/2021-July/386278.html
>>>      e8cccd3a3589 ("netdev-offload-dpdk: Fix IPv6 rewrite cast-align
>>> warning.")
>>> [3] 
>>> https://mail.openvswitch.org/pipermail/ovs-dev/2021-July/386279.html
>>>      1f7f557603a5 ("netdev-offload-dpdk: Fix vxlan vni cast-align 
>>> warnings.")
>>>> Thanks,
>>>> Olivier
>>>>
>>>>
>>>>> ---
>>>>>    lib/mbuf/rte_mbuf_core.h | 2 +-
>>>>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/lib/mbuf/rte_mbuf_core.h b/lib/mbuf/rte_mbuf_core.h
>>>>> index bb38d7f581..dabdeee604 100644
>>>>> --- a/lib/mbuf/rte_mbuf_core.h
>>>>> +++ b/lib/mbuf/rte_mbuf_core.h
>>>>> @@ -720,7 +720,7 @@ struct rte_mbuf_ext_shared_info {
>>>>>     *   The type to cast the result into.
>>>>>     */
>>>>>    #define rte_pktmbuf_mtod_offset(m, t, o)     \
>>>>> -     ((t)((char *)(m)->buf_addr + (m)->data_off + (o)))
>>>>> +     ((t)(void *)((char *)(m)->buf_addr + (m)->data_off + (o)))
>>>>>
>>>>>    /**
>>>>>     * A macro that points to the start of the data in the mbuf.
>>>>> -- 
>>>>> 2.28.0.2311.g225365fb51
>>>>>