Should tests make sure that providers respect FI_MSG_PREFIX when the -k option is used?
bturrubiates opened this issue · comments
I noticed looking at our Jenkins output that usNIC is passing the -k option on fi_msg_pingpong. The -k option sets the FI_MSG_PREFIX message on input to fi_getinfo, but the test never checks that the provider didn't clear the bit on output. This means that the user requets -k at the command line, and it may or may not take effect depending on if the provider supports it.
The test can handle non-prefix mode, but it's a bit deceiving. Should the test be checking if the provider cleared the bit and returning an error?
I would rather drop the -k option completely and let the app indicate support for MSG_PREFIX and use it or not based on the output.
I would rather drop the -k option completely and let the app indicate support for MSG_PREFIX and use it or not based on the output.
Well, I think it's useful to be able to enable and disable app support. This is one of the main ways we test both code paths.
In that case, I'm fine with the test aborting with ENODATA or similar if the bit has been cleared. I would change the -k text to indicate that it forces MSG_PREFIX or aborts if not needed.