[#LU-1050] configure broken on centos6.2 with OFED-1.5.4

[LU-1050] configure broken on centos6.2 with OFED-1.5.4 Created: 28/Jan/12 Updated: 07/Jun/12 Resolved: 05/Mar/12
Status:	Resolved
Project:	Lustre
Component/s:	None
Affects Version/s:	Lustre 2.2.0
Fix Version/s:	Lustre 2.2.0, Lustre 2.1.2

Type:

Bug

Priority:

Minor

Reporter:

Shuichi Ihara (Inactive)

Assignee:

Minh Diep

Resolution:

Fixed

Votes:

Labels:

None

Environment:

2.1.55 on CentOS6.2 with OFED-1.5.4

Attachments:

configure-rhel6.2-ofed1.5.4.patch

Issue Links:

Duplicate
is duplicated by	~~LU-1108~~	'configure' incorrectly set HAVE_RDMA...	Resolved

Severity:

Rank (Obsolete):

4721

Description

I'm testing the master branch on CentOS6.2 with OFED-1.5.4, but "./configure" broken on this configuration.
From quick my analysis, the number of argument of rdma_create_id() is different between in-kerneltree-ofed on RHEL6.2 and OpenFabric OFED.
There is an fix in ~~LU-778~~ to handle 4 arguments for rdma_create_id() in rhel6.2's in-kenreltree-ofed, but when if we use OpenFabric OFED on RHEL6.2, they are still 3 arguments to rdma_create_id(), and "configure" doesn't recognize them. Attached is what I made a simple patch to avoid this issue.

Comments

Comment by Peter Jones [ 28/Jan/12 ]

Minh

Can you please look into this one.

Thanks

Peter

Comment by Minh Diep [ 31/Jan/12 ]

Looking into this

Comment by Minh Diep [ 02/Feb/12 ]

Hi Shuichi,

What the kernel version did you build?

Comment by James A Simmons [ 02/Feb/12 ]

Actually I seen this problem on RHEL6.2 kernel 2.6.32-220. The way I worked around it was by copying the OFED kernel headers in my kernel tree.

Comment by James A Simmons [ 02/Feb/12 ]

The reason for the problem is that the rdma_create_id arg count is done before we set up the O2IBPATHS. Thus that test is always done withe the $LINUX directory. We need the below before the rdma_create_id test.

O2IBCPPFLAGS="-I$O2IBPATH/include"
EXTRA_KCFLAGS_save="$EXTRA_KCFLAGS"
EXTRA_KCFLAGS="$EXTRA_KCFLAGS $O2IBCPPFLAGS"
EXTRA_LNET_INCLUDE="$EXTRA_LNET_INCLUDE $O2IBCPPFLAGS"

Comment by Shuichi Ihara (Inactive) [ 02/Feb/12 ]

Yes, but O2IBPATH is checked after counting of rdma_create_id's argument.

Comment by Shuichi Ihara (Inactive) [ 05/Feb/12 ]

submitted more better fixes. http://review.whamcloud.com/2097

here is current o2ib related ordering in lustre-lnet.m4.
1) checking supported arguments of rdma_create_id()
2) detection of O2IBPATHS if o2ib is enabled
3) check rdma_create_id() works or not

The problem is 1)checking rdma_create_id() arguments always compiled with in-kernel-tree OFED header.
If the number of arguments is four with rhel6.2 kernel, this only works if O2IBPATHS pointed to in-kernel-tree OFED.
If O2IBPATHS is pointed to OFED-1.5.4, it failed due to external OFED still only supports three arguments in rdma_create_id(), but four arguments are detected on 1).

I think checking supported arguments of rdma_create_id() should run after 2)detection of O2IBPATHS and find correct number of arguments.

Comment by Peter Jones [ 10/Feb/12 ]

Liang

Could you please advise on this one?

Thanks

Peter

Comment by Liang Zhen (Inactive) [ 10/Feb/12 ]

I've posted a comment on gerrit