prov/psm3: --disable-shared causes "No HALs registered" error

[raffenet]

Describe the bug
Error message from psm3 provider during initialization.

To Reproduce

./autogen.sh
./configure --prefix=$PWD/i --disable-shared
make -j; make install
FI_PROVIDER=psm3 i/bin/fi_info

compute-386-08% FI_PROVIDER=psm3 i/bin/fi_info
compute-386-08:pid2800529.fi_info: No HALs registered
fi_getinfo: -61 (No data available)

Expected behavior
No error message. If psm3 cannot work with --disable-shared, then configure should disable the provider.

Environment:
Linux (Ubuntu 22.04)

[acgoldma]

I have been able to reproduce. Oddly, looks like the constructors that are included in other files are working, but the ones in the hals are not.

[acgoldma]

Comment from #8972:

looks like the __psmi_hal_verbs_constructor is grepable in libfabric.a, but not in fi_info.

It looks like constructors are not always included when linking a static lib to an executable.

I was able to fix this for fi_info by adding this:

 util_fi_info_SOURCES = \
        util/info.c
 util_fi_info_LDADD = $(linkback)
+util_fi_info_LDFLAGS = -Wl,--whole-archive,src/.libs/libfabric.a,--no-whole-archive

 util_fi_strerror_SOURCES = \
        util/strerror.c

This will link the whole archive instead of just the sections fi_info calls.

Not sure why some constructors worked and others did not.
Specifically all psm3 constructors, excpet the ones to register the HALs, worked fine.

Structurally, __psmi_gethostname_lock_constructor() works the same way as the hals, but seems to work fine.

As a work around for now, using the above patch should work.

[github-actions]

This issue is stale because it has been open 360 days with no activity. Remove stale label or comment, otherwise it will be closed in 7 days.