Merge pull request #6578 from grondo/issue#6576

shell: fix lost log messages during initialization

src/shell/log.c

@@ -249,7 +249,7 @@ void flux_shell_raise (const char *type,
     char buf [4096];
     va_list ap;
 
-    if (!shell || !shell->h || logger.exception_logged)
+    if (!shell || !shell->h || !shell->info || logger.exception_logged)
         return;
 
     va_start (ap, fmt);

src/shell/shell.c

@@ -1918,6 +1918,8 @@ int main (int argc, char *argv[])
 {
     flux_shell_t shell;
 
+    memset (&shell, 0, sizeof (shell));
+
     /* Initialize locale from environment
      */
     setlocale (LC_ALL, "");
@@ -1948,10 +1950,10 @@ int main (int argc, char *argv[])
      * Fetches missing info from shell handle if set.
      */
     if (!(shell.info = shell_info_create (&shell)))
-        exit (1);
+        shell_die (1, "failed to initialize shell info");
 
     if (shell_export_environment_from_job (&shell) < 0)
-        exit (1);
+        shell_die (1, "failed to initialize shell environment");
 
     /* Set verbose flag if set in attributes.system.shell.verbose */
     if (flux_shell_getopt_unpack (&shell, "verbose", "i", &shell.verbose) < 0)

t/t2608-job-shell-log.t

@@ -220,4 +220,18 @@ test_expect_success 'flux-shell: stdout/err from task.exec works' '
 	grep "^this is stderr" print.err &&
 	grep "^this is stdout" print.out
 '
+# Note: the following test uses alloc-bypass.so to bypass the scheduler in
+# order to introduce a jobspec error, to be caught by the job shell early
+# in initialization.
+test_expect_success 'flux-shell: early fatal errors are captured in eventlog' '
+	flux jobtap load alloc-bypass.so &&
+	id=$(flux run --dry-run -Salloc-bypass.R="$(flux resource R -i0)" true | \
+	     jq .version=2 | \
+	     flux job submit --flags=novalidate) &&
+	echo jobid=$id &&
+	flux job wait-event -Hvt 30 $id clean &&
+	test_expect_code 1 flux job attach $id >early-errors.out 2>&1 &&
+	grep "error parsing jobspec" early-errors.out &&
+	flux jobtap remove alloc-bypass.so
+'
 test_done