Re: [breakage] panic() does not halt arm64 systems under certain conditions

Re: [breakage] panic() does not halt arm64 systems under certain conditions

[Will Deacon]

Hi,

[Expanding CC list; original message is here:
 https://lore.kernel.org/linux-arm-kernel/BX1W47JXPMR8.58IYW53H6M5N@dragonstone/]

On Mon, Sep 16, 2019 at 09:35:36PM -0400, Xogium wrote:
> On arm64 in some situations userspace will continue running even after a
> panic. This means any userspace watchdog daemon will continue pinging,
> that service managers will keep running and displaying messages in certain
> cases, and that it is possible to enter via ssh in the now unstable system
> and to do almost anything except reboot/power off and etc. If
> CONFIG_PREEMPT=n is set in the kernel's configuration, the issue is fixed.
> I have reproduced the very same behavior with linux 4.19, 5.2 and 5.3. On
> x86/x86_64 the issue does not seem to be present at all.

I've managed to reproduce this under both 32-bit and 64-bit ARM kernels.
The issue is that the infinite loop at the end of panic() can run with
preemption enabled (particularly when invoking by echoing 'c' to
/proc/sysrq-trigger), so we end up rescheduling user tasks. On x86, this
doesn't happen because smp_send_stop() disables the local APIC in
native_stop_other_cpus() and so interrupts are effectively masked while
spinning.

A straightforward fix is to disable preemption explicitly on the panic()
path (diff below), but I've expanded the cc list to see both what others
think, but also in case smp_send_stop() is supposed to have the side-effect
of disabling interrupt delivery for the local CPU.

Will

--->8

diff --git a/kernel/panic.c b/kernel/panic.c
index 057540b6eee9..02d0de31c42d 100644
--- a/kernel/panic.c
+++ b/kernel/panic.c
@@ -179,6 +179,7 @@ void panic(const char *fmt, ...)
	 * after setting panic_cpu) from invoking panic() again.
	 */
	local_irq_disable();
+	preempt_disable_notrace();
 
	/*
	 * It's possible to come here directly from a panic-assertion and

Re: [breakage] panic() does not halt arm64 systems under certain conditions

[Russell King - ARM Linux admin]

On Tue, Sep 17, 2019 at 11:45:19AM +0100, Will Deacon wrote:
> Hi,
> 
> [Expanding CC list; original message is here:
>  https://lore.kernel.org/linux-arm-kernel/BX1W47JXPMR8.58IYW53H6M5N@dragonstone/]
> 
> On Mon, Sep 16, 2019 at 09:35:36PM -0400, Xogium wrote:
> > On arm64 in some situations userspace will continue running even after a
> > panic. This means any userspace watchdog daemon will continue pinging,
> > that service managers will keep running and displaying messages in certain
> > cases, and that it is possible to enter via ssh in the now unstable system
> > and to do almost anything except reboot/power off and etc. If
> > CONFIG_PREEMPT=n is set in the kernel's configuration, the issue is fixed.
> > I have reproduced the very same behavior with linux 4.19, 5.2 and 5.3. On
> > x86/x86_64 the issue does not seem to be present at all.
> 
> I've managed to reproduce this under both 32-bit and 64-bit ARM kernels.
> The issue is that the infinite loop at the end of panic() can run with
> preemption enabled (particularly when invoking by echoing 'c' to
> /proc/sysrq-trigger), so we end up rescheduling user tasks. On x86, this
> doesn't happen because smp_send_stop() disables the local APIC in
> native_stop_other_cpus() and so interrupts are effectively masked while
> spinning.
> 
> A straightforward fix is to disable preemption explicitly on the panic()
> path (diff below), but I've expanded the cc list to see both what others
> think,

Yep, and it looks like this bug goes back into the dim and distant past.
At least to the start of modern git history, 2.6.12-rc2.

> but also in case smp_send_stop() is supposed to have the side-effect
> of disabling interrupt delivery for the local CPU.

That can't fix it.  Consider a preemptive non-SMP kernel.
smp_send_stop() becomes a no-op there.

I'd suggest that a preemptive UP kernel on x86 hardware will suffer
this same issue - it will be able to preempt out of this loop and
continue running userspace.

-- 
RMK's Patch system: https://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 12.1Mbps down 622kbps up
According to speedtest.net: 11.9Mbps down 500kbps up

Re: [breakage] panic() does not halt arm64 systems under certain conditions

[Will Deacon]

On Tue, Sep 17, 2019 at 11:51:36AM +0100, Russell King - ARM Linux admin wrote:
> On Tue, Sep 17, 2019 at 11:45:19AM +0100, Will Deacon wrote:
> > [Expanding CC list; original message is here:
> >  https://lore.kernel.org/linux-arm-kernel/BX1W47JXPMR8.58IYW53H6M5N@dragonstone/]
> > 
> > On Mon, Sep 16, 2019 at 09:35:36PM -0400, Xogium wrote:
> > > On arm64 in some situations userspace will continue running even after a
> > > panic. This means any userspace watchdog daemon will continue pinging,
> > > that service managers will keep running and displaying messages in certain
> > > cases, and that it is possible to enter via ssh in the now unstable system
> > > and to do almost anything except reboot/power off and etc. If
> > > CONFIG_PREEMPT=n is set in the kernel's configuration, the issue is fixed.
> > > I have reproduced the very same behavior with linux 4.19, 5.2 and 5.3. On
> > > x86/x86_64 the issue does not seem to be present at all.
> > 
> > I've managed to reproduce this under both 32-bit and 64-bit ARM kernels.
> > The issue is that the infinite loop at the end of panic() can run with
> > preemption enabled (particularly when invoking by echoing 'c' to
> > /proc/sysrq-trigger), so we end up rescheduling user tasks. On x86, this
> > doesn't happen because smp_send_stop() disables the local APIC in
> > native_stop_other_cpus() and so interrupts are effectively masked while
> > spinning.
> > 
> > A straightforward fix is to disable preemption explicitly on the panic()
> > path (diff below), but I've expanded the cc list to see both what others
> > think,
> 
> Yep, and it looks like this bug goes back into the dim and distant past.
> At least to the start of modern git history, 2.6.12-rc2.
> 
> > but also in case smp_send_stop() is supposed to have the side-effect
> > of disabling interrupt delivery for the local CPU.
> 
> That can't fix it.  Consider a preemptive non-SMP kernel.
> smp_send_stop() becomes a no-op there.
> 
> I'd suggest that a preemptive UP kernel on x86 hardware will suffer
> this same issue - it will be able to preempt out of this loop and
> continue running userspace.

You're right; I managed to reproduce this locally on my xeon box.

Will

Re: [breakage] panic() does not halt arm64 systems under certain conditions

[Jookia]

On Tue, Sep 17, 2019 at 11:45:19AM +0100, Will Deacon wrote:
> Hi,
> 
> [Expanding CC list; original message is here:
>  https://lore.kernel.org/linux-arm-kernel/BX1W47JXPMR8.58IYW53H6M5N@dragonstone/]
> 
> On Mon, Sep 16, 2019 at 09:35:36PM -0400, Xogium wrote:
> > On arm64 in some situations userspace will continue running even after a
> > panic. This means any userspace watchdog daemon will continue pinging,
> > that service managers will keep running and displaying messages in certain
> > cases, and that it is possible to enter via ssh in the now unstable system
> > and to do almost anything except reboot/power off and etc. If
> > CONFIG_PREEMPT=n is set in the kernel's configuration, the issue is fixed.
> > I have reproduced the very same behavior with linux 4.19, 5.2 and 5.3. On
> > x86/x86_64 the issue does not seem to be present at all.
> 
> I've managed to reproduce this under both 32-bit and 64-bit ARM kernels.
> The issue is that the infinite loop at the end of panic() can run with
> preemption enabled (particularly when invoking by echoing 'c' to
> /proc/sysrq-trigger), so we end up rescheduling user tasks. On x86, this
> doesn't happen because smp_send_stop() disables the local APIC in
> native_stop_other_cpus() and so interrupts are effectively masked while
> spinning.
> 
> A straightforward fix is to disable preemption explicitly on the panic()
> path (diff below), but I've expanded the cc list to see both what others
> think, but also in case smp_send_stop() is supposed to have the side-effect
> of disabling interrupt delivery for the local CPU.
> 
> Will
> 
> --->8
> 
> diff --git a/kernel/panic.c b/kernel/panic.c
> index 057540b6eee9..02d0de31c42d 100644
> --- a/kernel/panic.c
> +++ b/kernel/panic.c
> @@ -179,6 +179,7 @@ void panic(const char *fmt, ...)
> 	 * after setting panic_cpu) from invoking panic() again.
> 	 */
> 	local_irq_disable();
> +	preempt_disable_notrace();
>  
> 	/*
> 	 * It's possible to come here directly from a panic-assertion and
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel

When you run with panic=... it will send you to a loop earlier in the
panic code before local_irq_disable() is hit, working around the bug.
A patch like this would make the behaviour the same:

diff --git a/kernel/panic.c b/kernel/panic.c
index 4d9f55bf7d38..92abbb5f8d38 100644
--- a/kernel/panic.c
+++ b/kernel/panic.c
@@ -331,7 +331,6 @@ void panic(const char *fmt, ...)

        /* Do not scroll important messages printed above */
        suppress_printk = 1;
-       local_irq_enable();
        for (i = 0; ; i += PANIC_TIMER_STEP) {
                touch_softlockup_watchdog();
                if (i >= i_next) {

Re: [breakage] panic() does not halt arm64 systems under certain conditions

[Will Deacon]

On Fri, Sep 20, 2019 at 02:25:01PM +1000, Jookia wrote:
> On Tue, Sep 17, 2019 at 11:45:19AM +0100, Will Deacon wrote:
> > A straightforward fix is to disable preemption explicitly on the panic()
> > path (diff below), but I've expanded the cc list to see both what others
> > think, but also in case smp_send_stop() is supposed to have the side-effect
> > of disabling interrupt delivery for the local CPU.
> > 
> > diff --git a/kernel/panic.c b/kernel/panic.c
> > index 057540b6eee9..02d0de31c42d 100644
> > --- a/kernel/panic.c
> > +++ b/kernel/panic.c
> > @@ -179,6 +179,7 @@ void panic(const char *fmt, ...)
> > 	 * after setting panic_cpu) from invoking panic() again.
> > 	 */
> > 	local_irq_disable();
> > +	preempt_disable_notrace();
> >  
> > 	/*
> > 	 * It's possible to come here directly from a panic-assertion and
> > 
> When you run with panic=... it will send you to a loop earlier in the
> panic code before local_irq_disable() is hit, working around the bug.
> A patch like this would make the behaviour the same:
> 
> diff --git a/kernel/panic.c b/kernel/panic.c
> index 4d9f55bf7d38..92abbb5f8d38 100644
> --- a/kernel/panic.c
> +++ b/kernel/panic.c
> @@ -331,7 +331,6 @@ void panic(const char *fmt, ...)
> 
>         /* Do not scroll important messages printed above */
>         suppress_printk = 1;
> -       local_irq_enable();
>         for (i = 0; ; i += PANIC_TIMER_STEP) {
>                 touch_softlockup_watchdog();
>                 if (i >= i_next) {

The reason I kept irqs enabled is because I figured they might be useful
for magic sysrq keyboard interrupts (e.g. if you wanted to reboot the box).

With 'panic=', the reboot happens automatically, so there's no issue there
afaict.

Will