Skip to contentSkip to conversation

reboot: use make_task_dead for the halt and power off fallback

PATCH In review Bradley Morgan 2 participants 4 messages Jul 13 – Jul 19 Stable branch

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 reboot: use make_task_dead for the halt and power off fallback

The reboot syscall calls do_exit(0) after kernel_halt() or kernel_power_off(). Those are expected to stop the machine and not return. When they do return, the shutdown path has already disabled interrupts and torn down state, and do_exit() then hits its WARN_ON(irqs_disabled()).

That is an error path, not a clean exit. Use make_task_dead() instead of do_exit(0): it is built for this, fixes up the irqs disabled and preempt state, and bounds repeated failure via oops_limit. This matches the make_task_dead pattern that exit.c already uses for the oops path.

Splat from syzbot:

ACPI: PM: Preparing to enter system sleep state S5
kvm: exiting hardware virtualization
reboot: Power down

  ------------[ cut here ]------------
  irqs_disabled()
  WARNING: kernel/exit.c:930 at do_exit+0x1cf7/0x2ae0 kernel/exit.c:930, CPU#0: init/6193
  Tainted: [L]=SOFTLOCKUP
  Call Trace:
   <TASK>
   __do_sys_reboot+0x36e/0x400 kernel/reboot.c:784
   do_syscall_64+0x115/0x840 arch/x86/entry/syscall_64.c:94
   entry_SYSCALL_64_after_hwframe+0x77/0x7f
   </TASK>
Fixes 001c28e57187 ("exit: Detect and fix irq disabled state in oops")
Reported-by [email protected]
---
 kernel/reboot.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

Changes since v1: v1 enabled IRQs and kept do_exit, fixing the symptom. Eric Biederman pointed out the real fix is to treat the returned halt or power off as the error path it is, and call make_task_dead instead.

kernel/reboot.c +5 -2
index bed6967bfa96..c10ac6a0200d 100644
@@ -14,6 +14,7 @@
 #include <linux/kmod.h>
 #include <linux/kmsg_dump.h>
 #include <linux/reboot.h>
+#include <linux/sched/task.h>
 #include <linux/suspend.h>
 #include <linux/syscalls.h>
 #include <linux/syscore_ops.h>
@@ -777,11 +778,13 @@ SYSCALL_DEFINE4(reboot, int, magic1, int, magic2, unsigned int, cmd,
 
 	case LINUX_REBOOT_CMD_HALT:
 		kernel_halt();
-		do_exit(0);
+		/* machine_halt() was expected to not return. */
+		make_task_dead(SIGKILL);
 
 	case LINUX_REBOOT_CMD_POWER_OFF:
 		kernel_power_off();
-		do_exit(0);
+		/* machine_power_off() was expected to not return. */
+		make_task_dead(SIGKILL);
 		break;
 
 	case LINUX_REBOOT_CMD_RESTART2:
Signature · 1 line
2.53.0
#1 lore raw
Quoted 5 lines · On Mon, 13 Jul 2026 06:23:32 +0000 Bradley Morgan <[email protected]> wrote:

The reboot syscall calls do_exit(0) after kernel_halt() or kernel_power_off(). Those are expected to stop the machine and not return. When they do return, the shutdown path has already disabled interrupts and torn down state, and do_exit() then hits its WARN_ON(irqs_disabled()).

Well... why are they returning? Is it both kernel_halt() and kernel_power_off()? The report seems to indicate that kernel_power_off() is returning.

So is there a flaw in x86 machine_power_off() which we should be addressing?

Quoted 5 lines

That is an error path, not a clean exit. Use make_task_dead() instead of do_exit(0): it is built for this, fixes up the irqs disabled and preempt state, and bounds repeated failure via oops_limit. This matches the make_task_dead pattern that exit.c already uses for the oops path.

And make_tsk_dead() is __noreturn.

#2 lore raw
Quoted 9 lines · Andrew Morton
Quoted 5 lines · On Mon, 13 Jul 2026 06:23:32 +0000 Bradley Morgan <[email protected]> wrote:

The reboot syscall calls do_exit(0) after kernel_halt() or kernel_power_off(). Those are expected to stop the machine and not return. When they do return, the shutdown path has already disabled interrupts and torn down state, and do_exit() then hits its WARN_ON(irqs_disabled()).

Quoted 3 lines

Well... why are they returning? Is it both kernel_halt() and kernel_power_off()? The report seems to indicate that kernel_power_off() is returning.

native_machine_power_off() has no __noreturn backstop.

(Only for kernel_power_off, kernel_halt() is theoretically safe, ends in stop_this_cpu)

Quoted 2 lines

So is there a flaw in x86 machine_power_off() which we should be addressing?

Well, yes, but it's really not that simple Andrew...

I did a quick dig at the other arches code and guess what?

x86, arm, arm64, (loongarch, riscv) power_off can all return,
(BUG!)

So just doing a fix arch specific isn't really right (unless we want a 3-5 patch series, which just adds more complication)

Quoted 7 lines
Quoted 5 lines

That is an error path, not a clean exit. Use make_task_dead() instead of do_exit(0): it is built for this, fixes up the irqs disabled and preempt state, and bounds repeated failure via oops_limit. This matches the make_task_dead pattern that exit.c already uses for the oops path.

And make_tsk_dead() is __noreturn.

Nice catch! let's just remove break;, it's dead code

Quoted 2 lines

Thanks!

#3 lore raw
Quoted 40 lines · Bradley Morgan
Quoted 9 lines · Andrew Morton
Quoted 5 lines · On Mon, 13 Jul 2026 06:23:32 +0000 Bradley Morgan <[email protected]> wrote:

The reboot syscall calls do_exit(0) after kernel_halt() or kernel_power_off(). Those are expected to stop the machine and not return. When they do return, the shutdown path has already disabled interrupts and torn down state, and do_exit() then hits its WARN_ON(irqs_disabled()).

Quoted 3 lines

Well... why are they returning? Is it both kernel_halt() and kernel_power_off()? The report seems to indicate that kernel_power_off() is returning.

native_machine_power_off() has no __noreturn backstop.

(Only for kernel_power_off, kernel_halt() is theoretically safe, ends in stop_this_cpu)

Quoted 2 lines

So is there a flaw in x86 machine_power_off() which we should be addressing?

Well, yes, but it's really not that simple Andrew...

I did a quick dig at the other arches code and guess what?

x86, arm, arm64, (loongarch, riscv) power_off can all return,
(BUG!)

So just doing a fix arch specific isn't really right (unless we want a 3-5 patch series, which just adds more complication)

Quoted 6 lines
Quoted 5 lines

That is an error path, not a clean exit. Use make_task_dead() instead of do_exit(0): it is built for this, fixes up the irqs disabled and preempt state, and bounds repeated failure via oops_limit. This matches the make_task_dead pattern that exit.c already uses for the oops path.

Quoted 8 lines
Quoted 1 line

And make_tsk_dead() is __noreturn.

Nice catch! let's just remove break;, it's dead code

Quoted 2 lines

Thanks!

Andrew? What do you think of this? (Sorry for the ping)

Thanks!