[PATCH] test-lib-functions: fix logic error in test_must_fail

[PATCH] test-lib-functions: fix logic error in test_must_fail

[Ramsay Jones]

After commit 710eb805 ("implement test_might_fail using a refactored
test_must_fail", 19-11-2015) several tests now unexpectedly pass:

  $ tail -21 ptest-out
  [17:25:53]
  All tests successful.

  Test Summary Report
  -------------------
  t1060-object-corruption.sh                       (Wstat: 0 Tests: 13 Failed: 0)
    TODO passed:   13
  t3510-cherry-pick-sequence.sh                    (Wstat: 0 Tests: 39 Failed: 0)
    TODO passed:   31
  t5707-clone-detached.sh                          (Wstat: 0 Tests: 13 Failed: 0)
    TODO passed:   4
  t6042-merge-rename-corner-cases.sh               (Wstat: 0 Tests: 26 Failed: 0)
    TODO passed:   8
  Files=727, Tests=13139, 407 wallclock secs ( 3.66 usr  0.54 sys + 85.10 cusr 258.06 csys = 347.36 CPU)
  Result: PASS
  make clean-except-prove-cache
  make[2]: Entering directory `/home/ramsay/git/t'
  rm -f -r 'trash directory'.* 'test-results'
  rm -f -r valgrind/bin
  make[2]: Leaving directory `/home/ramsay/git/t'
  make[1]: Leaving directory `/home/ramsay/git/t'
  $

Fix the logic bug in the refactored test_must_fail. Also, use the
integer comparison operator -eq when comparing the exit code, rather
than the string = operator.

Signed-off-by: Ramsay Jones <ramsay@ramsayjones.plus.com>
---

Hi Lars,

This patch is on top of the pu branch, but it relates to the two
patches on your 'ls/test-must-fail-sigpipe' branch. Could you
please squash the relevant parts of this patch into your patches.

So, in addition to the logic error in commit 710eb805, I modified
the test introduced by commit 653a48e56 ("add "ok=sigpipe" to
test_must_fail and use it to fix flaky tests", 19-11-2015).

Since I cannot test this second change (t5516 and t5504 don't
fail for me), I don't know if this change is correct - please
test and confirm. (No, it's not clear to me exactly what this
commit is supposed to do! :-D ).

[I didn't have time to go look what value would be returned by
a case statement where there is no 'default' limb - I suspect
that it is undefined behaviour. Even if it is defined, do all
shells behave properly? In any event, it is much simpler to
compare the strings directly!]

I have to say, I'm not keen on either of these commits, but Jeff
and Junio seem OK with it, so ... (the tests being flaky implies
that the git client is flaky - we should fix that).

Thanks.

ATB,
Ramsay Jones

 t/test-lib-functions.sh | 12 +++++-------
 1 file changed, 5 insertions(+), 7 deletions(-)

diff --git a/t/test-lib-functions.sh b/t/test-lib-functions.sh
index 1fdc58c..9061742 100644
--- a/t/test-lib-functions.sh
+++ b/t/test-lib-functions.sh
@@ -593,24 +593,22 @@ test_must_fail () {
 	esac
 	"$@"
 	exit_code=$?
-	if ! case ",$_test_ok," in *,success,*) false;; esac &&
-		test $exit_code = 0
+	if test $exit_code -eq 0 && test x$_test_ok != xsuccess
 	then
 		echo >&2 "test_must_fail: command succeeded: $*"
-		return 0
-	elif ! case ",$_test_ok," in *,sigpipe,*) false;; esac &&
-		test $exit_code = 141
+		return 1
+	elif test $exit_code -eq 141 && test x$_test_ok = xsigpipe
 	then
 		return 0
 	elif test $exit_code -gt 129 && test $exit_code -le 192
 	then
 		echo >&2 "test_must_fail: died by signal: $*"
 		return 1
-	elif test $exit_code = 127
+	elif test $exit_code -eq 127
 	then
 		echo >&2 "test_must_fail: command not found: $*"
 		return 1
-	elif test $exit_code = 126
+	elif test $exit_code -eq 126
 	then
 		echo >&2 "test_must_fail: valgrind error: $*"
 		return 1
-- 
2.6.0

Re: [PATCH] test-lib-functions: fix logic error in test_must_fail

[Jeff King]

On Tue, Nov 24, 2015 at 03:59:24PM +0000, Ramsay Jones wrote:

> After commit 710eb805 ("implement test_might_fail using a refactored
> test_must_fail", 19-11-2015) several tests now unexpectedly pass:

Thanks. I noticed there were some new passes, but I hadn't investigated
yet (I assumed it was "yay, we fixed a bug" not "oops, we broke the test
suite).

> diff --git a/t/test-lib-functions.sh b/t/test-lib-functions.sh
> index 1fdc58c..9061742 100644
> --- a/t/test-lib-functions.sh
> +++ b/t/test-lib-functions.sh
> @@ -593,24 +593,22 @@ test_must_fail () {
>  	esac
>  	"$@"
>  	exit_code=$?
> -	if ! case ",$_test_ok," in *,success,*) false;; esac &&
> -		test $exit_code = 0
> +	if test $exit_code -eq 0 && test x$_test_ok != xsuccess

I don't think this is quite right. I had the impression the original was
trying to allow something like:

  test_must_fail ok=success,sigpipe

And that's what the commas were for in the case statement.

If I understand the logic bug correctly, we simply need to flip the "!"
at the start of the case statement. But we could do something like:

  list_contains () {
	case ",$1," in
	*,$2,*)
		return 0
		;;
	esac
	return 1
  }
  ...
  if ! list_contains "$_test_ok" success && test "$exit_code" -eq 0
  then
	return 0
  fi

which is perhaps a bit more clear, as it encapsulates the funny
negative.

> Since I cannot test this second change (t5516 and t5504 don't
> fail for me), I don't know if this change is correct - please
> test and confirm. (No, it's not clear to me exactly what this
> commit is supposed to do! :-D ).

They don't fail consistently. It's a SIGPIPE race.

> [I didn't have time to go look what value would be returned by
> a case statement where there is no 'default' limb - I suspect
> that it is undefined behaviour. Even if it is defined, do all
> shells behave properly? In any event, it is much simpler to
> compare the strings directly!]

Yeah, I wondered that. We wouldn't depend on it in the example I gave
above.

> I have to say, I'm not keen on either of these commits, but Jeff
> and Junio seem OK with it, so ... (the tests being flaky implies
> that the git client is flaky - we should fix that).

I was actually reasonably happy with just having test_must_fail ignore
SIGPIPE, for the second one. But the infrastructure added by the first
patch does fix real issues in test_might_fail (e.g., that it should
complain of a valgrind failure but doesn't). And once you have that
infrastructure, the second patch becomes trivial.

I don't think the git client is actually flaky here in a way that we can
fix. If I understand the issue correctly, it is that the server side
hangs up in an error case while the client is still writing. So we might
get one of two outcomes:

  1. If the client has finished writing, it will do a read, see the
     hang up, and die().

  2. If the client is still writing, it will get SIGPIPE and die.

There's no solution outside of ignoring SIGPIPE and handling the write()
error ourselves. We've been loathe to do that in the past because
SIGPIPE applies globally, and missing a write() anywhere in the program
means we may spin on bogus writes longer than we need to.

-Peff

Re: [PATCH] test-lib-functions: fix logic error in test_must_fail

[Ramsay Jones]

On 24/11/15 21:08, Jeff King wrote:
> On Tue, Nov 24, 2015 at 03:59:24PM +0000, Ramsay Jones wrote:
> 
>> After commit 710eb805 ("implement test_might_fail using a refactored
>> test_must_fail", 19-11-2015) several tests now unexpectedly pass:
> 
> Thanks. I noticed there were some new passes, but I hadn't investigated
> yet (I assumed it was "yay, we fixed a bug" not "oops, we broke the test
> suite).
> 
>> diff --git a/t/test-lib-functions.sh b/t/test-lib-functions.sh
>> index 1fdc58c..9061742 100644
>> --- a/t/test-lib-functions.sh
>> +++ b/t/test-lib-functions.sh
>> @@ -593,24 +593,22 @@ test_must_fail () {
>>  	esac
>>  	"$@"
>>  	exit_code=$?
>> -	if ! case ",$_test_ok," in *,success,*) false;; esac &&
>> -		test $exit_code = 0
>> +	if test $exit_code -eq 0 && test x$_test_ok != xsuccess
> 
> I don't think this is quite right. I had the impression the original was
> trying to allow something like:
> 
>   test_must_fail ok=success,sigpipe
> 
> And that's what the commas were for in the case statement.

Ah, OK, that makes a bit more sense. (perhaps the commit message
could mention this).

> 
> If I understand the logic bug correctly, we simply need to flip the "!"
> at the start of the case statement. But we could do something like:
> 
>   list_contains () {
> 	case ",$1," in
> 	*,$2,*)
> 		return 0
> 		;;
> 	esac
> 	return 1
>   }

OK.

>   ...
>   if ! list_contains "$_test_ok" success && test "$exit_code" -eq 0
>   then
> 	return 0
>   fi
    ^^
Is this intended to be part of the if..elif.. chain, or a separate
initial conditional? Hmm, actually it doesn't matter since it has
exactly the same logic error as the original patch ... :-D

> 
> which is perhaps a bit more clear, as it encapsulates the funny
> negative.
> 
>> Since I cannot test this second change (t5516 and t5504 don't
>> fail for me), I don't know if this change is correct - please
>> test and confirm. (No, it's not clear to me exactly what this
>> commit is supposed to do! :-D ).
> 
> They don't fail consistently. It's a SIGPIPE race.

Yes, and unfortunately I can't get it to happen for me.
Which platforms has it been observed on?

> 
>> [I didn't have time to go look what value would be returned by
>> a case statement where there is no 'default' limb - I suspect
>> that it is undefined behaviour. Even if it is defined, do all
>> shells behave properly? In any event, it is much simpler to
>> compare the strings directly!]
> 
> Yeah, I wondered that. We wouldn't depend on it in the example I gave
> above.

Indeed.

> 
>> I have to say, I'm not keen on either of these commits, but Jeff
>> and Junio seem OK with it, so ... (the tests being flaky implies
>> that the git client is flaky - we should fix that).
> 
> I was actually reasonably happy with just having test_must_fail ignore
> SIGPIPE, for the second one. But the infrastructure added by the first
> patch does fix real issues in test_might_fail (e.g., that it should
> complain of a valgrind failure but doesn't). And once you have that
> infrastructure, the second patch becomes trivial.

OK

> 
> I don't think the git client is actually flaky here in a way that we can
> fix. If I understand the issue correctly, it is that the server side
> hangs up in an error case while the client is still writing. So we might
> get one of two outcomes:
> 
>   1. If the client has finished writing, it will do a read, see the
>      hang up, and die().
> 
>   2. If the client is still writing, it will get SIGPIPE and die.
> 
> There's no solution outside of ignoring SIGPIPE and handling the write()
> error ourselves.

This is what I had in mind. Of course, I haven't spent any time looking
into how much work is involved (or what effect it would have on code
maintenance) ... [So, it may not be worth doing, but "we can't fix" is
perhaps a bit strong.]

>                   We've been loathe to do that in the past because
> SIGPIPE applies globally, and missing a write() anywhere in the program
> means we may spin on bogus writes longer than we need to.

Hmm, OK.

ATB,
Ramsay Jones

Re: [PATCH] test-lib-functions: fix logic error in test_must_fail

[Jeff King]

On Wed, Nov 25, 2015 at 12:05:36AM +0000, Ramsay Jones wrote:

> >   ...
> >   if ! list_contains "$_test_ok" success && test "$exit_code" -eq 0
> >   then
> > 	return 0
> >   fi
>     ^^
> Is this intended to be part of the if..elif.. chain, or a separate
> initial conditional? Hmm, actually it doesn't matter since it has
> exactly the same logic error as the original patch ... :-D

I meant it as part of the same one, but yeah, it has the wrong "return"
(I noticed it when writing up the real patch as a SQUASH on top of
Lars's topics).

> > They don't fail consistently. It's a SIGPIPE race.
> 
> Yes, and unfortunately I can't get it to happen for me.
> Which platforms has it been observed on?

I haven't tried to reproduce it, but I think it was seen originally in
the Travis build.

-Peff