[dpdk-dev] [dpdk-stable] AVX512 bug on SkyLake

[Thomas Monjalon]

09/11/2018 14:17, Thomas Monjalon:
> 09/11/2018 11:03, Ferruh Yigit:
> > On 11/8/2018 11:01 PM, Yongseok Koh wrote:
> > > 
> > >> On Nov 8, 2018, at 9:21 AM, Ferruh Yigit <ferruh.yigit at intel.com> wrote:
> > >>
> > >> On 11/8/2018 3:59 PM, Thomas Monjalon wrote:
> > >>> Hi,
> > >>>
> > >>> We need to gather more information about this bug.
> > >>> More below.
> > >>>
> > >>> 07/11/2018 10:04, Wiles, Keith:
> > >>>>> On Nov 6, 2018, at 9:30 PM, Yongseok Koh <yskoh at mellanox.com> wrote:
> > >>>>>> On Nov 5, 2018, at 6:06 AM, Wiles, Keith <keith.wiles at intel.com> wrote:
> > >>>>>>> On Nov 2, 2018, at 9:04 PM, Yongseok Koh <yskoh at mellanox.com> wrote:
> > >>>>>>>
> > >>>>>>> This is a workaround to prevent a crash, which might be caused by
> > >>>>>>> optimization of newer gcc (7.3.0) on Intel Skylake.
> > >>>>>>
> > >>>>>> Should the code below not also test for the gcc version and
> > >>>>>> the Sky Lake processor, maybe I am wrong but it seems it is
> > >>>>>> turning AVX512 for all GCC builds
> > >>>>>
> > >>>>> I didn't want to check gcc version as 7.3.0 is very new. Only gcc 8 is newly up since then (gcc 8.2).
> > >>>>> Also, I wasn't able to test every gcc versions and I wanted to be a bit conservative for this crash.
> > >>>>> Performance drop (if any) by disabling a new (experimental) feature would be less risky than unaccountable crash.
> > >>>>> And, it does disable the feature only if CONFIG_RTE_ENABLE_AVX512=n. Please refer to v3.
> > >>>>
> > >>>> Are you not turning off all of the GCC versions for AVX512.
> > >>>> And you can test for range or greater then GCC version and
> > >>>> it just seems like we are turning off every gcc version, is that true?
> > >>>
> > >>> Do we know exactly which GCC versions are affected?
> > >>>
> > >>>>>> Also bug 97 seems a bit obscure reference, maybe you know
> > >>>>>> the bug report, but more details would be good?
> > >>>>>
> > >>>>> I sent out the report to dev list two month ago.
> > >>>>> And I created the Bug 97 in order to reference it
> > >>>>> in the commit message.
> > >>>>> I didn't want to repeat same message here and there,
> > >>>>> but it would've been better to have some sort of summary
> > >>>>> of the Bug, although v3 has a few more words.
> > >>>>> However, v3 has been merged.
> > >>>>
> > >>>> Still this is too obscure if nothing else give a link to
> > >>>> a specific bug not just 97.
> > >>>
> > >>> The URL is
> > >>> 	https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fbugs.dpdk.org%2Fshow_bug.cgi%3Fid%3D97&data=02%7C01%7Cyskoh%40mellanox.com%7C90ff6c361faf422b976108d6459eb490%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636772945282345908&sdata=2o%2Fg203aWrKCYg16S6oI4BcS41igpLu1DloS%2FrRnknc%3D&reserved=0
> > >>> The bug is also pointing to an email:
> > >>> 	https://emea01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fmails.dpdk.org%2Farchives%2Fdev%2F2018-September%2F111522.html&data=02%7C01%7Cyskoh%40mellanox.com%7C90ff6c361faf422b976108d6459eb490%7Ca652971c7d2e4d9ba6a4d149256f461b%7C0%7C0%7C636772945282345908&sdata=NCFKxaREd69iZ8eyFKg%2FWBP73CLTXkxrNQQeii%2Bbsao%3D&reserved=0
> > >>>
> > >>> Summary:
> > >>> 	- CPU: Intel Skylake
> > >>> 	- Linux environment: Ubuntu 18.04
> > >>> 	- Compiler: gcc-7.3 (Ubuntu 7.3.0-16ubuntu3)
> > >>
> > >> Is it possible to test a few other gcc versions to check if the issue is
> > >> specific to this compiler version?
> > > 
> > > Nothing's impossible but even with my quick search in gcc.gnu.org,
> > > I could find the following documents mention mavx512f support:
> > > 
> > > GCC 4.9.0
> > > April 22, 2014 (changes, documentation)
> > >  
> > > GCC 5.1
> > > April 22, 2015 (changes, documentation)
> > >  
> > > GCC 6.4
> > > July 4, 2017 (changes, documentation)
> > >  
> > > GCC 7.1
> > > May 2, 2017 (changes, documentation)
> > >  
> > > GCC 8.1
> > > May 2, 2018 (changes, documentation)
> > > 
> > > We altogether have to put quite large resource to verify all of the versions.
> > >  
> > > I assumed older than gcc 7 would have the same issue. I know it was a speculation
> > > but like I mentioned I wanted to be more conservative. I didn't mean this is a permanent fix.
> > > For two months, we couldn't have any tangible solution (actually nobody cared including myself),
> > > so I submitted the patch to temporarily disable mavx512f.
> > > 
> > > I'm still not sure what the best option is...
> > 
> > For permanent fix we need more information, currently we can't re-produce this
> > defect. Since you can reproduce it we need your support.
> > 
> > Right now we don't know if this is compiler issue or code defect in rte_memcpy()
> > or something else.
> > 
> > It is easy to disable mavx512f as temporarily solution but it is coming with the
> > cost of the performance drop, also without knowing the actual root cause I
> > wouldn't say this is being conservative, actual issue may be just hidden with
> > this change.
> > 
> > I think as first thing we need to find a way to reproduce this issue in any
> > other way than using mlx5 PMD. So that we can put more organized effort to fix this.
> > I attached a simple unit test for rte_memcpy(), if this is a rte_memcpy() with
> > avx512f defect as claimed, you should be able to see the issue with that, right?
> > Did you able to find a chance to test it? Do you observer any crash there?
> 
> I am able to connect to a machine where the issue is reproduced.
> So I have tested replacing rte_memcpy with memcpy,
> and the crash disappears when using memcpy.
> So it confirms that the issue is in rte_memcpy.

One workaround is to disable CONFIG_RTE_ENABLE_AVX,
but it is disabling AVX and AVX2 for all DPDK code.

A more limited fix (tested) can be to disable AVX2 version of rte_memcpy
and rely on the AVX version (which is not crashing):

--- a/lib/librte_eal/common/include/arch/x86/rte_memcpy.h
+++ b/lib/librte_eal/common/include/arch/x86/rte_memcpy.h
-#elif defined RTE_MACHINE_CPUFLAG_AVX2
+#elif defined RTE_MACHINE_CPUFLAG_AVX2_disable