Re: [PATCH 2/6] t1406: prepare for the refs code to fail with BUG()

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Hi,

On Wed, 2 May 2018, Junio C Hamano wrote:

> Duy Nguyen <pclouds@xxxxxxxxx> writes:
> 
> > On Mon, Apr 30, 2018 at 12:17 AM, Johannes Schindelin
> > <johannes.schindelin@xxxxxx> wrote:
> >> t1406 specifically verifies that certain code paths fail with a BUG: ...
> >> message.
> >>
> >> In the upcoming commit, we will convert that message to be generated via
> >> BUG() instead of die("BUG: ..."), which implies SIGABRT instead of a
> >> regular exit code.
> >
> > On the other hand, SIGABRT on linux creates core dumps. And on some
> > setup (like mine) core dumps may be redirected to some central place
> > via /proc/sys/kernel/core_pattern. I think systemd does it too but I
> > didn't check.
> >
> > This moving to SIGABRT when we know it _will_ happen when running the
> > test suite will accumulate core dumps over time and not cleaned up by
> > the test suite. Maybe keeping die("BUG: here is a good compromise.
> 
> I do not think it is.  At regular runtime, we _do_ want Git to dump
> core if it triggers BUG() condition, whose point is to mark
> conditions that should never happen.

Indeed.

> As discussed in this thread, tests that use t/helper/ executables
> that try to trickle BUG() codepath to ensure that these "should
> never happen" conditions are caught do need to deal with it.  If
> dumping core is undesirable, tweaking BUG() implementation so that
> it becomes die("BUG: ...") *ONLY* when the caller knows what it is
> doing (e.g. running t/helper/ commands) is probably a good idea.
> Perhaps GIT_TEST_OPTS can gain one feature "--bug-no-abort" and set
> an environment variable so that implementation of BUG() can notice,
> or something.

I think we can do even better than that. t/helper/*.c could set a global
variable that no other code is supposed to set, to trigger an alternative
to SIGABRT. Something like

-- snip --
diff --git a/t/helper/test-tool.c b/t/helper/test-tool.c
index 87066ced62a..5176f9f20ae 100644
--- a/t/helper/test-tool.c
+++ b/t/helper/test-tool.c
@@ -47,7 +47,9 @@ static struct test_cmd cmds[] = {
 int cmd_main(int argc, const char **argv)
 {
 	int i;
+	extern int BUG_exit_code;
 
+	BUG_exit_code = 99;
 	if (argc < 2)
 		die("I need a test name!");
 
diff --git a/usage.c b/usage.c
index cdd534c9dfc..9c84dccfa97 100644
--- a/usage.c
+++ b/usage.c
@@ -210,6 +210,9 @@ void warning(const char *warn, ...)
 	va_end(params);
 }
 
+/* Only set this, ever, from t/helper/, when verifying that bugs are
caught. */
+int BUG_exit_code;
+
 static NORETURN void BUG_vfl(const char *file, int line, const char *fmt,
va_list params)
 {
 	char prefix[256];
@@ -221,6 +224,8 @@ static NORETURN void BUG_vfl(const char *file, int
line, const char *fmt, va_lis
 		snprintf(prefix, sizeof(prefix), "BUG: ");
 
 	vreportf(prefix, fmt, params);
+	if (BUG_exit_code)
+		exit(BUG_exit_code);
 	abort();
 }
 
-- snap --

I'll try to find some time to play with this.

Ciao,
Dscho
> 
> When we are testing normal parts of Git outside t/helper/, we never
> want to hit BUG().  Aborting and dumping core when that happens is
> an desirable outcome.  From that point of view, the idea in 1/6 of
> this patch series to annotate test_must_fail and say "we know this
> one is going to hit BUG()" is a sound one.  The implementation in
> 1/6 to treat SIGABRT as an acceptable failure needs to be replaced
> to instead use the above mechanism you would use to tell BUG() not
> to abort but die with message to arrange that to happen before
> running the git command (most likely something from t/helper/) under
> test_must_fail ok=sigabrt; and then those who regularly break their
> Git being tested (read: us devs) and hit BUG() could instead set the
> environment variable (i.e. internal implementation detail) manually
> in their environment to turn these BUG()s into die("BUG:...)s while
> testing their early mistakes if they do not want core (of course,
> you could just do "ulimit -c", and that may be simpler solution of
> your "testing Git contaminates central core depot" issue).
> 
> 
> 
> 
> 
> 
> 



[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]

  Powered by Linux