Couple fixes to fi_pingpong

[ghost]

First commit is cleanup. Second commit adds source address parameter to fix issue #9900, using fi_pingpong in multi-adapter environment.

[zachdworkin]

bot:aws:retest

[shijin-aws]

@zachdworkin AWS CI failed due to fi_pingpong failed on Windows platform

15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:283 :  * PingPong options:
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:285 :   - src_addr            : [None]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:286 :   - src_port            : [0]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:287 :   - dst_addr            : [None]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:288 :   - dst_port            : [0]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:289 :   - sizes_enabled       : default size = 64
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:290 :   - iterations          : default iterations: 10
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:293 :   - provider            : efa
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:2237: Selected endpoint: RDM
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:509 : Initializing control messages
15:13:10 socket(): C:\9971_1712094667\libfabric\util\pingpong.c:430 , ret=-22 (Invalid argument)
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:1876: Freeing resources of test suite
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:1910: Resources of test suite freed
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:283 :  * PingPong options:
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:285 :   - src_addr            : [None]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:286 :   - src_port            : [0]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:287 :   - dst_addr            : [127.0.0.1]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:288 :   - dst_port            : [0]
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:289 :   - sizes_enabled       : default size = 64
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:290 :   - iterations          : default iterations: 10
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:293 :   - provider            : efa
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:2237: Selected endpoint: RDM
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:509 : Initializing control messages
15:13:10 [error] C:\9971_1712094667\libfabric\util\pingpong.c:321 : getaddrinfo : E
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:1876: Freeing resources of test suite
15:13:10 [debug] C:\9971_1712094667\libfabric\util\pingpong.c:1910: Resources of test suite freed
15:13:10 2024-04-02 22:11:37,190 - INFO - test_runner_windows - Server process failed with return code 22
15:13:10 
15:13:10 Build step 'Execute shell' marked build as failure
15:13:10 Recording test results

[shijin-aws]

Did your change modify the default behavior for OOB exchange?

[zachdworkin]

@shijin-aws #9963 did modify the default behaviour for fabtests. The backlog went from 0 to 511. This prevents SYN Flooding on port 3000 when using oob exchange in fabtests. I did not change the behaviour outside of fabtests.

[shijin-aws]

Is fi_ini necessarily called on Windows? Saw the commit that ofi_osd_init() is removed from the test because it is called in fi_ini, but I find fi_ini has pthread calls which may not be compiled on windows at all

[j-xiong]

fi_ini is always called from fi_getinfo and fi_fabric. There are stub definitions for pthread under include/windows.

[shijin-aws]

Any ideas how can this change cause 15:13:10 socket(): C:\9971_1712094667\libfabric\util\pingpong.c:430 , ret=-22 (Invalid argument) ?

[ghost]

Any ideas how can this change cause 15:13:10 socket(): C:\9971_1712094667\libfabric\util\pingpong.c:430 , ret=-22 (Invalid argument) ?

It might be an ordering issue. ofi_socket is called before any calls to libfabric. I don't care that much about removing ofi_osd_init() call. I will drop that commit from this PR.

[shijin-aws]

Any ideas how can this change cause 15:13:10 socket(): C:\9971_1712094667\libfabric\util\pingpong.c:430 , ret=-22 (Invalid argument) ?

It might be an ordering issue. ofi_socket is called before any calls to libfabric. I don't care that much about removing ofi_osd_init() call. I will drop that commit from this PR.

Ah, that makes sense. Thank you!

[shijin-aws]

@chien-intel windows test passed for the latest push

[ghost]

Great, thanks!