UCT BTL fails to compile using UCX 1.7.0

[hppritcha]

The UCT BTL in both master and v4.0.x fail to compile against UCX 1.7.0 rc1.

  CC       btl_uct_tl.lo
In file included from btl_uct_device_context.h:16:0,
                 from btl_uct_endpoint.c:17:
btl_uct_rdma.h: In function 'mca_btl_uct_get_rkey':
btl_uct_rdma.h:61:5: error: too few arguments to function 'uct_rkey_unpack'
     ucs_status = uct_rkey_unpack ((void *) remote_handle, rkey);
     ^
In file included from btl_uct_device_context.h:16:0,
                 from btl_uct_rdma.c:12:
btl_uct_rdma.h: In function 'mca_btl_uct_get_rkey':
In file included from btl_uct_device_context.h:16:0,
                 from btl_uct_amo.c:12:
btl_uct_rdma.h: In function 'mca_btl_uct_get_rkey':
In file included from btl_uct_device_context.h:16:0,
                 from btl_uct_component.c:43:
btl_uct_rdma.h: In function 'mca_btl_uct_get_rkey':
btl_uct_rdma.h:61:5: error: too few arguments to function 'uct_rkey_unpack'
     ucs_status = uct_rkey_unpack ((void *) remote_handle, rkey);
     ^
btl_uct_rdma.h:61:5: error: too few arguments to function 'uct_rkey_unpack'
     ucs_status = uct_rkey_unpack ((void *) remote_handle, rkey);
     ^
btl_uct_rdma.h:61:5: error: too few arguments to function 'uct_rkey_unpack'
     ucs_status = uct_rkey_unpack ((void *) remote_handle, rkey);
     ^
In file included from btl_uct.h:41:0,
                 from btl_uct_endpoint.c:15:
/users/hpritchard/ucx_testing//ucx-1.7.0//test_install/include/uct/api/uct.h:2142:14: note: declared here
 ucs_status_t uct_rkey_unpack(uct_component_h component, const void *rkey_buffer,
              ^
In file included from btl_uct_device_context.h:16:0,
                 from btl_uct_endpoint.c:17:
btl_uct_rdma.h: In function 'mca_btl_uct_rkey_release':
btl_uct_rdma.h:72:5: warning: passing argument 1 of 'uct_rkey_release' from incompatible pointer type [enabled by default]
     uct_rkey_release (rkey);
     ^

This is master including the commits from #7016.

Here's the current UCT BTL build situation:

OMPI Version	UCX 1.6	UCX 1.7	UCX master
MASTER	✓	❌	✓
4.0.x	✓	❌	❌

[hppritcha]

we need this resolved for 4.0.x branch as not being able to build against UCX 1.7.x may impact ability to get UCX updates into certain linux distros.

[gpaulsen]

Lets talk on Tuesday about this.

[hjelmn]

I can resolve the 1.7 build failure. More than a little disappointed that the API freeze agreement was broken.

[jsquyres]

@hjelmn Thanks! We'll still need version checking in the configury, though (because otherwise, we could be in this position again in the future 😦).

[hjelmn]

Not sure I agree. It will just silently fail to build when I want to see an error message to indicate that UCX (again) decided to change an external API.

[hjelmn]

And really this comes down to a failure of communication. This sort of issue can be fixed in ~5 mins but I never get a message about the break until it comes down to "lets just no-build btl/uct".

[hjelmn]

And really the only reason I see anything ATM is because @jsquyres accidentally deleted the dev groups. Reset all my alert preferences 🤣. If there is a build failure for btl/uct please always tag me.

[jsquyres]

@hjelmn No, it will not silently fail -- that's the problem. If UCX changes their API, we can be in the following situation (which has happened twice, BTW):

1. Open MPI version ABC compiles/works fine with UCX version XYZ.
2. UCX comes out with version (XYZ+1) and changes their API.
3. Open MPI version ABC passes configure and decides to build UCX, but then later fails to compile.

So we need to either:

• Check every single UCX API in configure.m4 (because any of them can apparently change between UCX releases), or
• Put in UCX version checking in configure.m4.

Either is fine with me. I'm tired of getting blamed for UCX changing their API. ☹️

[hjelmn]

So, you want to fail configure? That would be ok with me. What I don't want is silently add btl-uct to the mca no build. UCX really should be blamed because they are changing an external API in every release now. I talked to them last December and they told me that was not going to happen. They said there were a couple of things they needed to update (which is fine) but then it would be stabilized. It has been a year and that has not happened.

[hjelmn]

I see the misunderstanding. With the check it will just silently fail btl/uct. That is what I do not want.