t7518: fix flaky grep invocation

From: Elijah Newren <[email protected]>

t7518.1 added in commit 862e80a413 ("ident: handle NULL email when complaining of empty name", 2017-02-23), was trying to make sure that the test with an empty ident did not segfault and did not result in glibc quiety translating a NULL pointer into a name of "(null)". It did the latter by ensuring that a grep for "null" didn't appear in the output, but on one automatic CI run I observed the following output:

fatal: empty ident name (for <runner@fv-az128-670.gcliasfzo2nullsdbrimjtbyhg.cx.internal.cloudapp.net>) not allowed

Note that 'null' appears as a substring of the domain name, found within 'gcliasfzo2nullsdbrimjtbyhg'. Tighten the test by searching for "(null)" rather than "null".

---
t7518: fix flaky grep invocation

Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-884%2Fnewren%2Ffix-flaky-t7518-grep-v1
Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-884/newren/fix-flaky-t7518-grep-v1
Pull-Request: https://github.com/git/git/pull/884

 t/t7518-ident-corner-cases.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

index dc3e9c8c88..905957bd0a 100755
@@ -13,7 +13,7 @@ test_expect_success 'empty name and missing email' '
 		sane_unset GIT_AUTHOR_EMAIL &&
 		GIT_AUTHOR_NAME= &&
 		test_must_fail git commit --allow-empty -m foo 2>err &&
-		test_i18ngrep ! null err
+		test_i18ngrep ! "(null)" err
 	)
 '
 

base-commit: d4a392452e292ff924e79ec8458611c0f679d6d4

gitgitgadget

Hi Elijah,

:-). I have always been surprised by glibc's behavior here; I'd almost always rather have a segfault than silently turning a NULL pointer into "(null)" when formatted with "%s".

Yep; that sounds like a straightforward fix that will make this test non-flaky.

Thanks,
Taylor

Why do we even need grep again? If we were to segfault, wouldn't test_must_fail catch it for us?

... ahh, OK, your libc may allow NULL left in email and that is what we are trying to catch?

Honestly, I am not sure if it even makes sense to test it like this. The code that the test tries to protect against future breakages roughly look like this these days.

	if (!email) {
		... one way to assign non-NULL to email ...
	}
	if (!email) {
		... do other things to assign non-NULL to email or die ...
	}

	if (want_name) {
		... here we require email to be set because we show
		... it in an error message
	}

The original problem was that the code had "if (want_name)" part first before email==NULL condition has been dealt with, and used email==NULL in one of the error messages. The fix was to move the part that deals with email==NULL up, as it does not need any of the effects that happen in "if (want_name)" block.

Now, what is this particular test protecting against? We may again move the "if (want_name)" block up and trigger the error message in it that uses e-mail while having NULL in email again? If so, would it make more sense to do something along this line, lose the "grep" and keep "test_must_fail"? We expect and require at this point in the code (the patch is inside that "if (want_name)" block) email must have been set up already, so it would be a BUG() otherwise.

Hmm?

 ident.c | 2 ++
 1 file changed, 2 insertions(+)

index 6aba4b5cb6..9b71563d95 100644
@@ -430,6 +430,8 @@ const char *fmt_ident(const char *name, const char *email,
 			if (strict) {
 				if (using_default)
 					ident_env_hint(whose_ident);
+				if (!email)
+					BUG("NULL email when we need to complain???");
 				die(_("empty ident name (for <%s>) not allowed"), email);
 			}
 			pw = xgetpwuid_self(NULL);

Yes, as mentioned in the commit message:

...was trying to make sure that
the test with an empty ident did not segfault and did not result in glibc quiety translating a NULL pointer into a name of "(null)".

This was also taken from the comment in the file just above the code in question (though it was just barely outside the context region).

It may come down to the angle you view it from; "the current code doesn't have such a problem and can't within its newer design, so why test it?" vs. "when you add a regression test at the time you make a bugfix, the regression test should demonstrate the failure with the old version of the code". Without the grep, the test would not have failed on the original old version of the code, at least not on some platforms.
I'm guessing that's why Peff put it there, and still had a preference to leave the grep in (see his comments on this issue at
https://lore.kernel.org/git/[email protected]/;
this is a case where it'd be nice if gitgitgadget accepted an in-reply-to flag of some sort).

However, although I have a slight preference to fixing what grep is searching for, my primary concern was with avoiding flakiness in the test. I'm happy if someone wants to make additional code changes or tweak the testcase here so long as we avoid the flakiness.