[PATCH] powerpc/watchdog: disable watchdog when watchdog_thresh=0

[PATCH] powerpc/watchdog: disable watchdog when watchdog_thresh=0

[Luiz Capitulino (Red Hat)]

On powerpc, trying to disable the hardlockup detector with:

 # echo 0 > /proc/sys/kernel/watchdog_thresh

causes an immediate and endless hardlockup detection, eventually leading
to an unusable system (splat example below). This happens because
watchdog_calc_timeouts() sets wd_timer_period_ms, wd_panic_timeout_tb
and wd_smp_panic_timeout_tb to zero, but start_watchdog() still starts
the timer (with expiry=0) in this code path:

proc_watchdog_thresh()
  proc_watchdog_update()
    __lockup_detector_reconfigure()
      watchdog_hardlockup_start()
        start_watchdog_on_cpu()
          start_watchdog()
            hrtimer_start() /* wd_timer_period_ms=0 */

However, according to Documentation/admin-guide/sysctl/kernel.rst,
setting watchdog_thresh=0 should disable both the soft and hard lockup
detectors.

This condition is already checked in __lockup_detector_reconfigure() for
the soft lockup detector and the arch-independent hardlockup detector,
what's missing is powerpc performing the same check in start_watchdog().

[  278.110085] watchdog: CPU 0 detected hard LOCKUP on other CPUs 1
[  278.110086] watchdog: CPU 1 became unstuck TB:33050716089564575
[  278.110097] CPU: 1 UID: 0 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: G            E       7.4.0-rc1 #1 PREEMPT(full)
[  278.110103] watchdog: CPU 0 TB:33050716089564185, last SMP heartbeat TB:33050716089563885 (0ms ago)
[  278.110109] Tainted: [E]=UNSIGNED_MODULE
[  278.110126] Hardware name: IBM,8375-42A POWER9 (architected) 0x4e0202 0xf000005 of:IBM,FW940.00 (VL940_027) hv:phyp pSeries
[  278.110136] NIP:  c0000000001bd518 LR: c0000000014dedf8 CTR: 0000000000000000
[  278.110145] REGS: c000000004b17a80 TRAP: 0900   Tainted: G            E        (7.4.0-rc1)
[  278.110155] MSR:  800000000280b033 <SF,VEC,VSX,EE,FP,ME,IR,DR,RI,LE>  CR: 22000224  XER: 20040004
[  278.110189] CFAR: 0000000000000000 IRQMASK: 0
GPR00: 0000000000000000 c000000004b17d20 c000000001c38100 0000000000000000
GPR04: 000000000000ffff 0000000000000000 0000000000000000 0000000000000000
GPR08: 0000000000000000 0000000000000080 0000000000000080 ffffffffffffffff
GPR12: 0000000000000000 c000000017fff300 0000000000000000 000000002ef3d820
GPR16: 0000000000000000 0000000000000000 0000000000000000 0000000000000000
GPR20: 0000000000000000 0000000000000000 0000000000000000 0000000000000001
GPR24: 0000000000000001 00000040c0a6075e 0000000000000000 0000000000000001
GPR28: c0000001fdb5b140 0000000000000001 c000000002291968 c000000002291970
[  278.110306] NIP [c0000000001bd518] plpar_hcall_norets_notrace+0x18/0x2c
[  278.110319] LR [c0000000014dedf8] check_and_cede_processor+0x48/0x60
[  278.110333] Call Trace:
[  278.110338] [c000000004b17d20] [c000000004b17dd0] 0xc000000004b17dd0 (unreliable)
[  278.110354] [c000000004b17d80] [c0000000014df020] shared_cede_loop+0x70/0x170
[  278.110367] [c000000004b17dc0] [c0000000014de460] cpuidle_enter_state+0x2f0/0x748
[  278.110381] [c000000004b17e70] [c000000000fe43b8] cpuidle_enter+0x48/0x70
[  278.110397] [c000000004b17eb0] [c0000000002cdcf8] call_cpuidle+0x48/0x90
[  278.110409] [c000000004b17ed0] [c0000000002d6300] cpuidle_idle_call+0x160/0x2d0
[  278.110422] [c000000004b17f20] [c0000000002d6578] do_idle+0x108/0x1f0
[  278.110434] [c000000004b17f60] [c0000000002d6938] cpu_startup_entry+0x48/0x50
[  278.110446] [c000000004b17f90] [c0000000000615c0] start_secondary+0x290/0x2a0
[  278.110459] [c000000004b17fe0] [c00000000000e258] start_secondary_prolog+0x10/0x14

Fixes: 2104180a5369 ("powerpc/64s: implement arch-specific hardlockup watchdog")
Cc: stable@vger.kernel.org
Signed-off-by: Luiz Capitulino (Red Hat) <luiz.capitulino@linux.dev>
---
Notes:
- This issue was not discovered by AI. We have a new test-case in our CI that
  sets watchdog_thresh=0, when running on powerpc we get the described issue
- This issue was initially reproduced on Power10, but I developed the fix
  on Power9 (which is what I have access now)

 arch/powerpc/kernel/watchdog.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
index 582ee3fd71c1..44f178e07531 100644
--- a/arch/powerpc/kernel/watchdog.c
+++ b/arch/powerpc/kernel/watchdog.c
@@ -487,6 +487,9 @@ static void start_watchdog(void *arg)
 	if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
 		return;
 
+	if (!watchdog_thresh)
+		return;
+
 	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
 		return;
 
-- 
2.56.0

Re: [PATCH] powerpc/watchdog: disable watchdog when watchdog_thresh=0

[Michael Ellerman]

On 4/11/26 2:40 pm, Luiz Capitulino (Red Hat) wrote:
> On powerpc, trying to disable the hardlockup detector with:
> 
>   # echo 0 > /proc/sys/kernel/watchdog_thresh
> 
> causes an immediate and endless hardlockup detection, eventually leading
> to an unusable system (splat example below).

...
> This condition is already checked in __lockup_detector_reconfigure() for
> the soft lockup detector and the arch-independent hardlockup detector,
> what's missing is powerpc performing the same check in start_watchdog().

Arguably setting watchdog_thresh to 0 should mean watchdog_enabled becomes false.
Then there wouldn't need to be extra checks for thresh == 0 anywhere.

But if the existing code already checks for thresh == 0 then this is fine.

> Fixes: 2104180a5369 ("powerpc/64s: implement arch-specific hardlockup watchdog")
> Cc: stable@vger.kernel.org
> Signed-off-by: Luiz Capitulino (Red Hat) <luiz.capitulino@linux.dev>
> ---
> Notes:
> - This issue was not discovered by AI. We have a new test-case in our CI that
>    sets watchdog_thresh=0, when running on powerpc we get the described issue
> - This issue was initially reproduced on Power10, but I developed the fix
>    on Power9 (which is what I have access now)
> 
>   arch/powerpc/kernel/watchdog.c | 3 +++
>   1 file changed, 3 insertions(+)
> 
> diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
> index 582ee3fd71c1..44f178e07531 100644
> --- a/arch/powerpc/kernel/watchdog.c
> +++ b/arch/powerpc/kernel/watchdog.c
> @@ -487,6 +487,9 @@ static void start_watchdog(void *arg)
>   	if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
>   		return;
>   
> +	if (!watchdog_thresh)
> +		return;
> +
>   	if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
>   		return;


Fix looks good to me.

Reviewed-by: Michael Ellerman <mpe@kernel.org>

cheers

Re: [PATCH] powerpc/watchdog: disable watchdog when watchdog_thresh=0

[Luiz Capitulino (Red Hat)]

On 11/6/26 5:26 AM, Michael Ellerman wrote:
> On 4/11/26 2:40 pm, Luiz Capitulino (Red Hat) wrote:
>> On powerpc, trying to disable the hardlockup detector with:
>>
>>   # echo 0 > /proc/sys/kernel/watchdog_thresh
>>
>> causes an immediate and endless hardlockup detection, eventually leading
>> to an unusable system (splat example below).
> 
> ...
>> This condition is already checked in __lockup_detector_reconfigure() for
>> the soft lockup detector and the arch-independent hardlockup detector,
>> what's missing is powerpc performing the same check in start_watchdog().
> 
> Arguably setting watchdog_thresh to 0 should mean watchdog_enabled becomes false.
> Then there wouldn't need to be extra checks for thresh == 0 anywhere.
> 
> But if the existing code already checks for thresh == 0 then this is fine.

Yes, the common code in kernel/watchdog.c actually checks for both.

Thanks for the review.

> 
>> Fixes: 2104180a5369 ("powerpc/64s: implement arch-specific hardlockup watchdog")
>> Cc: stable@vger.kernel.org
>> Signed-off-by: Luiz Capitulino (Red Hat) <luiz.capitulino@linux.dev>
>> ---
>> Notes:
>> - This issue was not discovered by AI. We have a new test-case in our CI that
>>    sets watchdog_thresh=0, when running on powerpc we get the described issue
>> - This issue was initially reproduced on Power10, but I developed the fix
>>    on Power9 (which is what I have access now)
>>
>>   arch/powerpc/kernel/watchdog.c | 3 +++
>>   1 file changed, 3 insertions(+)
>>
>> diff --git a/arch/powerpc/kernel/watchdog.c b/arch/powerpc/kernel/watchdog.c
>> index 582ee3fd71c1..44f178e07531 100644
>> --- a/arch/powerpc/kernel/watchdog.c
>> +++ b/arch/powerpc/kernel/watchdog.c
>> @@ -487,6 +487,9 @@ static void start_watchdog(void *arg)
>>       if (!(watchdog_enabled & WATCHDOG_HARDLOCKUP_ENABLED))
>>           return;
>> +    if (!watchdog_thresh)
>> +        return;
>> +
>>       if (!cpumask_test_cpu(cpu, &watchdog_cpumask))
>>           return;
> 
> 
> Fix looks good to me.
> 
> Reviewed-by: Michael Ellerman <mpe@kernel.org>
> 
> cheers