[PATCH] If init dies, log a signal which killed it, if any.

[PATCH] If init dies, log a signal which killed it, if any.

[Denys Vlasenko]

I just received another user's pleas for help when their
init mystriously dies. I again explained that they need to check
whether it dies because of bad instruction, a segv, or something else.

Which prompted me to make kernel do this first step automatically.
We can easily detect when the death is from e.g. SIGILL,
and let user know that.

The code is fairly self-explanatory. Compile-tested.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
---
 kernel/exit.c |   23 ++++++++++++++++++++++-
 1 files changed, 22 insertions(+), 1 deletions(-)

diff --git a/kernel/exit.c b/kernel/exit.c
index 294b170..89d0892 100644
--- a/kernel/exit.c
+++ b/kernel/exit.c
@@ -710,8 +710,29 @@ static struct task_struct *find_new_reaper(struct task_struct *father)
 
 	if (unlikely(pid_ns->child_reaper == father)) {
 		write_unlock_irq(&tasklist_lock);
-		if (unlikely(pid_ns == &init_pid_ns))
+		if (unlikely(pid_ns == &init_pid_ns)) {
+			/*
+			 * The situation when init segfaults is rather typical.
+			 * Give some useful diagnostics: do we die on signal?
+			 */
+			if (fatal_signal_pending(father)) {
+				const char *msg = "";
+				sigset_t *mask = &father->pending.signal;
+				/* Only force_sig()ned signals kill init */
+				if (sigismember(mask, SIGSEGV))
+					msg = " SIGSEGV";
+				if (sigismember(mask, SIGBUS))
+					msg = " SIGBUS";
+				if (sigismember(mask, SIGILL))
+					msg = " SIGILL";
+				if (sigismember(mask, SIGFPE))
+					msg = " SIGFPE";
+				/* (do we want to check SIGTRAP too?) */
+				printk(KERN_ERR
+					"init received fatal signal%s\n", msg);
+			}
 			panic("Attempted to kill init!");
+		}
 
 		zap_pid_ns_processes(pid_ns);
 		write_lock_irq(&tasklist_lock);
-- 
1.7.7.5

Re: [PATCH] If init dies, log a signal which killed it, if any.

[Oleg Nesterov]

On 01/20, Denys Vlasenko wrote:
>
> I just received another user's pleas for help when their
> init mystriously dies. I again explained that they need to check
> whether it dies because of bad instruction, a segv, or something else.
>
> Which prompted me to make kernel do this first step automatically.
> We can easily detect when the death is from e.g. SIGILL,
> and let user know that.
>
> The code is fairly self-explanatory. Compile-tested.
>
> Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
> ---
>  kernel/exit.c |   23 ++++++++++++++++++++++-
>  1 files changed, 22 insertions(+), 1 deletions(-)
>
> diff --git a/kernel/exit.c b/kernel/exit.c
> index 294b170..89d0892 100644
> --- a/kernel/exit.c
> +++ b/kernel/exit.c
> @@ -710,8 +710,29 @@ static struct task_struct *find_new_reaper(struct task_struct *father)
>
>  	if (unlikely(pid_ns->child_reaper == father)) {
>  		write_unlock_irq(&tasklist_lock);
> -		if (unlikely(pid_ns == &init_pid_ns))
> +		if (unlikely(pid_ns == &init_pid_ns)) {
> +			/*
> +			 * The situation when init segfaults is rather typical.
> +			 * Give some useful diagnostics: do we die on signal?
> +			 */
> +			if (fatal_signal_pending(father)) {

The fatal signal can be already dequeued. Although mostly this works.

> +				const char *msg = "";
> +				sigset_t *mask = &father->pending.signal;
> +				/* Only force_sig()ned signals kill init */
> +				if (sigismember(mask, SIGSEGV))
> +					msg = " SIGSEGV";
> +				if (sigismember(mask, SIGBUS))
> +					msg = " SIGBUS";
> +				if (sigismember(mask, SIGILL))
> +					msg = " SIGILL";
> +				if (sigismember(mask, SIGFPE))
> +					msg = " SIGFPE";

This doesn't look right too. Again, if it was killed by SIGSEGV
this signal can be dequeued and not pending.

> +				/* (do we want to check SIGTRAP too?) */
> +				printk(KERN_ERR
> +					"init received fatal signal%s\n", msg);
> +			}

I'd suggest this trivial change instead,

	-	panic("Attempted to kill init!");
	+	panic("Attempted to kill init! code=%08x\n",
	+		father->signal->group_exit_code ?: father->exit_code);

Oleg.