UCT/GDAKI: Revert - Check for HCA reachability like rc_mlx5 - v1.23.x

[ofirfarjun7]

What?

Revert #11801

Why?

Reachability check of ib can be wrong for gda

[svc-nvidia-pr-review]

🤖 Starting review — findings will be posted here when done.

[svc-nvidia-pr-review]

why switch gda from uct_ib_iface_is_reachable_v2 to always-reachable? gda still uses IB device addresses, and the old callback rejected peers on a different link layer/subnet/pkey and enforced device-scope reachability (used by ucp wireup select). With this stub uct_iface_is_reachable_v2() reports any remote address as reachable, so a gda lane can be selected toward an unreachable peer and fail only at connect. pls add a comment explaining why dropping the check is safe here.

[svc-nvidia-pr-review]

test/gtest/uct/test_uct_iface.cc

this removes the only reachability coverage for gda. The is_reachable test expects some corrupted addresses to be unreachable, which can't hold once the callback always returns 1. is gda reachability / ep_is_connected device matching covered elsewhere? if not, pls keep a gda case instead of removing it.

[svc-ucx]

🤖 CI Triage Agent — UCX PR (AddressSanitizer BlueField on worker 1) · commit 2c66bd38

TL;DR: All 8627 gtests passed; the job failed only because LeakSanitizer reported a 131-byte "indirect leak" from glibc's __dlerror() buffer materialized by the pointless (void)dlerror() call in ucs_sys_get_lib_info() (src/ucs/sys/lib.c:18) during ucs_profile_cleanup. This is unrelated to the gdaki-reachability revert in PR #11839 — fix by removing that dlerror() call and/or adding a leak:__dlerror entry to contrib/lsan.supp.

Full analysis

Summary: make test in the AddressSanitizer/BlueField job exited with error 1 after the gtest run completed successfully, due to a LeakSanitizer report (SUMMARY: AddressSanitizer: 131 byte(s) leaked in 1 allocation(s)).

Root cause: The test run itself is clean — [ PASSED ] 8627 tests, no failures, no timeouts, continuous log activity to the end (largest gap ≈7 s, so no hang). The only error is at process teardown:

Indirect leak of 131 byte(s) in 1 object(s) allocated from:
  #1 __vasprintf_internal libio/vasprintf.c:71
  #2 ___asprintf stdio-common/asprintf.c:31
  #3 __dlerror dlfcn/dlerror.c:74
  #4 ucs_sys_get_lib_info  src/ucs/sys/lib.c:18
  #5 ucs_sys_get_lib_path  src/ucs/sys/lib.c:32
  #6 ucs_profile_write     src/ucs/profile/profile.c:322
  #7 ucs_profile_dump / #8 ucs_profile_cleanup
  #9 scoped_profile::~scoped_profile()  test/gtest/ucs/test_profile.cc:48
 #10 test_profile.perf_overhead          test/gtest/ucs/test_profile.cc:482

ucs_sys_get_lib_info() calls (void)dlerror() before dladdr(). That call is dead code — the result is discarded and dladdr() failure is detected via its return value — but it has a side effect: if any earlier dlopen()/dlsym() on the same thread left a pending error (very common on this platform, where optional plugins/CUDA/DOCA/GGA libraries are probed; note Suppressions used: 1 24 dlsym in the log), glibc asprintf()s the message into thread-local last_result storage. That buffer is only freed by the next dlerror() call or thread exit, so LSan sees it as leaked at exit().

contrib/lsan.supp suppresses dlopen and dlsym frames but has no entry for __dlerror, so this escapes suppression and fails the build. Because it depends on (a) a preceding failed dlopen/dlsym and (b) gtest's randomized test order putting test_profile.perf_overhead in the right place, it is intermittent — a flaky infrastructure failure, not a regression from this PR (the PR only reverts the GDAKI HCA-reachability check; src/ucs/sys/lib.c has not been touched since 2022 and nothing in the profile/dlerror path is related).

Implicated commit: Not a regression from 2c66bd38 (PR #11839). Pre-existing code: bca0bf5b / original 125542b1 (Leonid Genkin) for src/ucs/sys/lib.c; exposure increased by 49a0d4c1 "UCS/SYS: Added support for dynamically loaded external modules/plugins (#11206)" (Roie Danino), which added the dlopen/dlsym plugin probing and the corresponding lsan.supp entries.

File: src/ucs/sys/lib.c:18 (and contrib/lsan.supp:1-7)

Suggested fix:

1. Remove the no-op (void)dlerror(); at src/ucs/sys/lib.c:18 — it is never read and only forces glibc to allocate/format a stale pending error string:

   ucs_status_t ucs_sys_get_lib_info(Dl_info *dl_info)
   {
       if (dladdr(ucs_sys_get_lib_info, dl_info) == 0) {
           return UCS_ERR_NO_MEMORY;
       }
       return UCS_OK;
   }

2. Belt-and-braces: add leak:__dlerror (or leak:dlerror) to contrib/lsan.supp alongside the existing dlopen/dlsym entries, since any other dlerror() consumer can hit the same glibc artifact.
3. Meanwhile, this build should simply be re-run — the failure does not indicate a problem with the revert in PR UCT/GDAKI: Revert - Check for HCA reachability like rc_mlx5 - v1.23.x #11839.

Related: PR #11839 (this build), sibling reverts #11838 / #11819; suppression-list history: #11206 (added dlopen/dlsym suppressions), commit 311fa48b (added ibv_alloc_pd suppression for BF/GGA), f675ed50 (added the BlueField ASAN job).