Need to bump main branch libfabric version

[shijin-aws]

https://github.com/ofiwg/libfabric/blob/main/configure.ac#L11

Should it be updated to 1.19.0 or newer? I am asking because I was seeing a bug caused by this PR #8826 which only populate hmem_data for FI_VERSION >= 1.19, but the current libfabric version is 1.18.0rc1, which is even lower than the v1.18.x branch.

[shijin-aws]

@shefty

[shefty]

Yes, we can update the version in preparation for the next release.

[shijin-aws]

Yes, we can update the version in preparation for the next release.

Does that mean the version in main branch is always lower than released branch before preparing a release? That sounds odd to me.

The bug in #8826 is that cur_abi_attr->hmem_data = FI_VERSION_GE(user_version, FI_VERSION(1, 19)) is only populated for libfabric version >=1.19, but the mr attr flags OFI_HMEM_DATA_GDRCOPY_HANDLE was added in main branch in https://github.com/ofiwg/libfabric/blob/main/prov/efa/src/efa_mr.c#L268. And here we have an inconsistency.

So we either need to add a FI_VERSION check in https://github.com/ofiwg/libfabric/blob/main/prov/efa/src/efa_mr.c#L268, which is done in #8902,
Or we bump the libfabric version in main branch right now to 1.19.

Which solution do you prefer? I would prefer the latter one as having a main branch of version 1.18* is kind of strange.

[shefty]

We updated the main branch to v1.18.0rc1, then created the v1.18.x branch. The main branch simply hasn't been updated with a new version since then. Updating the defines is trivial and can be done. Updating the version in configure.ac is less trivial and either means waiting closer to the actual release or predicting whether there will be any changes in the public header files.

[shijin-aws]

or predicting whether there will be any changes in the public header files.

I see, I think we should have updated the version as part of this PR 49d1a76, which changed the public header files.

[shijin-aws]

Should we update the version to 1.19.0 or 1.19.0rc1? It looks to me 1.19.0rc1 makes FI_VERSION_GE(user_version, FI_VERSION(1, 19)) returns false as well.

[shefty]

The version should be MAJOR 1, MINOR 19, REVISION 0. rc1 is a package number and part of configure.ac.

[shefty]

Btw, don't update configure to rc1. You can update to a1 if needed, but I usually update configure as part of the release process.

[shijin-aws]

Btw, don't update configure to rc1. You can update to a1 if needed, but I usually update configure as part of the release process.

I see, so I can just bump this line https://github.com/ofiwg/libfabric/blob/main/include/rdma/fabric.h#L87, is that good for you?

[shijin-aws]

#8905

[shijin-aws]

PR is merged