[Tarantool-patches] [PATCH v4 4/4] crash: report crash data to the feedback server

Cyrill Gorcunov gorcunov at gmail.com
Sun Dec 20 22:16:54 MSK 2020


On Sun, Dec 20, 2020 at 07:41:47PM +0100, Vladislav Shpilevoy wrote:
> 
> The error above I checked on your branch. And it remains the same
> even with the latest commit which you didn't push yet. Because the
> feedback options still are not configured before box.cfg. It has
> nothing to do with the recursive crash.
> 
> It means, as I said above, that you can't send a crash before
> box.cfg is done. Therefore you can simply not even try if the
> host is not configured yet. It solves the error message, and alongside
> solves the crash chain issue. You don't need environment variables
> for this.
> 
> I would suggest not to rush responding to these emails, but give it
> some thought beforehand.
> 
> It is actually strange, that feedback daemon does not work before
> box.cfg. We don't see any reports from tarantool instances which are
> used as an application server. But it has nothing to do with your
> patch anyway.

The latest updates are in gorcunov/gh-5261-crash-report-4-rev2
---
...
static bool send_crashinfo = false;
...
static void
crash_signal_cb(int signo, siginfo_t *siginfo, void *context)
{
	static volatile sig_atomic_t in_cb = 0;
	struct crash_info *cinfo;

	if (in_cb == 0) {
		in_cb = 1;
		cinfo = crash_collect(signo, siginfo, context);
		crash_report_stderr(cinfo);
		if (send_crashinfo)
			crash_report_feedback_daemon(cinfo);
	} else {
		/* Got a signal while running the handler. */
		fprintf(stderr, "Fatal %d while backtracing", signo);
	}

	/* Try to dump a core */
	struct sigaction sa = {
		.sa_handler = SIG_DFL,
	};
	sigemptyset(&sa.sa_mask);
	sigaction(SIGABRT, &sa, NULL);
	abort();
}
...
void
crash_cfg_set_params(const char *host, bool is_enabled)
{
	if (host == NULL || !is_enabled) {
		if (send_crashinfo) {
			pr_debug("disable sending crashinfo feedback");
			send_crashinfo = false;
			feedback_host[0] = '\0';
		}
		return;
	}

	if (strcmp(feedback_host, host) != 0) {
		strlcpy_a(feedback_host, host);
		if (strlen(feedback_host) < strlen(host))
			pr_panic("feedback_host is too long");
	}

	if (!send_crashinfo) {
		pr_debug("enable sending crashinfo feedback");
		send_crashinfo = true;
	}
}
---

IOW, until explicitly configured via box.cfg{} the crash won't
be sent. I'll address all your previous comments in v5 of the branch.


More information about the Tarantool-patches mailing list