[LU-9118] Failed to build MOFED 3.4-1.0.0

• Type: Bug
• Resolution: Fixed
• Priority: Minor
• Fix Version/s: Lustre 2.10.0
• Affects Version/s: Lustre 2.10.0
• Severity: 3
• Assignee: James A Simmons
• Reporter: Minh Diep

Description

+ make -j4 -s
Making all in .
In file included from /tmp/rpmbuild-lustre-jenkins-2HDIMGOt/BUILD/lustre-2.9.52_82_gb2d45e0/libcfs/include/libcfs/linux/libcfs.h:85:0,
                 from /tmp/rpmbuild-lustre-jenkins-2HDIMGOt/BUILD/lustre-2.9.52_82_gb2d45e0/libcfs/include/libcfs/libcfs.h:37,
                 from /tmp/rpmbuild-lustre-jenkins-2HDIMGOt/BUILD/lustre-2.9.52_82_gb2d45e0/lnet/klnds/o2iblnd/o2iblnd.h:77,
                 from /tmp/rpmbuild-lustre-jenkins-2HDIMGOt/BUILD/lustre-2.9.52_82_gb2d45e0/lnet/klnds/o2iblnd/o2iblnd.c:38:
/tmp/rpmbuild-lustre-jenkins-2HDIMGOt/BUILD/lustre-2.9.52_82_gb2d45e0/libcfs/include/libcfs/linux/linux-time.h:149:19: error: redefinition of 'ktime_get_real_ns'
 static inline u64 ktime_get_real_ns(void)
                   ^
In file included from /var/lib/jenkins/workspace/lustre-reviews-ib/arch/x86_64/build_type/client/distro/sles12/ib_stack/mlx/BUILD/mlnx-ofa_kernel-devel/usr/src/ofa_kernel/default/include/linux/compat-2.6.h:82:0,
                 from <command-line>:0:
/var/lib/jenkins/workspace/lustre-reviews-ib/arch/x86_64/build_type/client/distro/sles12/ib_stack/mlx/BUILD/mlnx-ofa_kernel-devel/usr/src/ofa_kernel/default/include/linux/compat-3.17.h:12:19: note: previous definition of 'ktime_get_real_ns' was here
 static inline u64 ktime_get_real_ns(void) {
                   ^

Comments

[Dmitry Eremin (Inactive)]

James, I didn't reproduce this locally yet. I just got an issue report from our customer. I'm going to reproduce this and check this patch soon.

[James A Simmons]

Does my patch help you Dmitry

[Dmitry Eremin (Inactive)]

It looks this also affect the Open OFED 3.18-3.

[James A Simmons]

Okay I figured out a solution. So both OFED and libcfs was setting HAVE_KTIME_GET_REAL_NS. So I inverted the logic of detection in libcfs, now NEED_KTIME_GET_REAL_NS and unset it in o2iblnd.h.  The logic is in the commit. Try the latest patch. Tested on RHEL7.2 + MOFED 3.3 and RHEL7.3 + MOFED 4.0 and SLES11SP3 + OFED12

[Minh Diep]

Yes, I confirmed that it failed in el6.8 too

 

[James A Simmons]

I tested it on MOFED 4 + RHEL7.3 but then I went to my Power8 RHEL7.2 + MOFED 3.3 which didn't work. I also have a Cray system with OFED 3.12 + SLES11SP3. The man problem is both MOFED and libcfs are setting HAVE_KTIME_GET_REAL_NS  So I need to rethink it.

[Bob Glossman (Inactive)]

I hesitate to give this a +review. I can see the proposed solution works when building without MOFED. Can't see evidence of it working building with MOFED.

James, have you tested build with MOFED on both el6 and el7 to cover both cases of finding and not finding ktime_get_real_ns in the kernel?

[Gerrit Updater]

James Simmons ([e-mail]) uploaded a new patch: https://review.whamcloud.com/25564
Subject: LU-9118 o2iblnd: handle MOFED libcfs time api collision
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: 25f64d4aeb292d42ccafbd8568476df4f32b9e7c

[James A Simmons]

I'm installing a MOFED 4 stack on our system. Will push something soon.

[Minh Diep]

you need to fix ktime_get_real_ns first

[James A Simmons]

I just wanted to see if ktime_get_ns() is the only problem.