Segmentation fault when calling ucp_listener_create using IB-enabled library on non-IB machine

[krakowski]

Describe the bug

Calling ucp_listener_create with a IB-enabled library on a machine without InfiniBand adapters leads to a segmentation fault due to a null pointer dereference in uct_listener_create (cm is null here). Debugging this issue I found out that this happens because the Worker's Connection Manager array contains a null pointer.

Disabling all IB-related features during configuration fixes this problem.

--with-verbs=no \
--with-rc=no \
--with-ud=no \
--with-dc=no \
--with-mlx5-dv=no \
--with-ib-hw-tm=no \
--with-dm=no \
--with-cm=no \
--with-rdmacm=no \

Steps to Reproduce

• UCX version 1.10.0
• Configured with contrib/configure-release

Setup and versions

• Linux 5.4.85-1-MANJARO (x86_64)
• rdma-core 30.0

[krakowski]

This may also be the cause of issue #6244

[alinask]

Hi @krakowski ,
can you please try setting the following environment parameter on both sides (client and server) to see it helps?
UCX_SOCKADDR_TLS_PRIORITY=tcp
(given that you are not setting any other UCX environment parameters)

[krakowski]

Hi @alinask,

yes, I can confirm that it works by setting UCX_SOCKADDR_TLS_PRIORITY=tcp.

[alinask]

Can you please use this environment parameter for UCX v1.10?
We will fix this issue for the next UCX release.
Thanks!

[abellina]

I just hit this also, and was about to file an issue @alinask thanks for jumping on it. UCX_SOCKADDR_TLS_PRIORITY=tcp also fixes it for me.

[abellina]

@alinask @yosefe do you think it's possible to get this into 1.10.1?

@petro-rudenko fyi.

[alinask]

@alinask @yosefe do you think it's possible to get this into 1.10.1?

@petro-rudenko fyi.

@abellina The fix was merged into the UCX v1.10.x branch.

[alinask]

@krakowski @abellina Following the fix in the UCX v1.10.x branch, can we close this ticket?

[krakowski]

From my side it can be closed. Thanks for the quick fix! 🙂

[abellina]

Same for me. Thanks @alinask. Verified a TCP-only setup with v1.10.x is not segfaulting.