[PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[Hoeun Ryu]

From: Hoeun Ryu <hoeun.ryu@lge.com>

 Many console device drivers hold the uart_port->lock spinlock with irq enabled
(using spin_lock()) while the device drivers are writing characters to their devices,
but the device drivers just try to hold the spin lock (using spin_trylock()) if
"oops_in_progress" is equal or greater than 1 to avoid deadlocks.

 There is a case ocurring a deadlock related to the lock and oops_in_progress. A CPU
could be stopped by smp_send_stop() while it was holding the port lock because irq was
enabled. Once a CPU stops, it doesn't respond interrupts anymore and the lock stays
locked forever.

 console_flush_on_panic() is called during panic() and it eventually holds the uart
lock but the lock is held by another stopped CPU and it is a deadlock. By moving
bust_spinlocks(0) after console_flush_on_panic(), let the console device drivers
think the Oops is still in progress to call spin_trylock() instead of spin_lock() to
avoid the deadlock.

Signed-off-by: Hoeun Ryu <hoeun.ryu@lge.com>
---
 kernel/panic.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/panic.c b/kernel/panic.c
index 42e4874..b4063b6 100644
--- a/kernel/panic.c
+++ b/kernel/panic.c
@@ -233,8 +233,6 @@ void panic(const char *fmt, ...)
 	if (_crash_kexec_post_notifiers)
 		__crash_kexec(NULL);
 
-	bust_spinlocks(0);
-
 	/*
 	 * We may have ended up stopping the CPU holding the lock (in
 	 * smp_send_stop()) while still having some valuable data in the console
@@ -246,6 +244,8 @@ void panic(const char *fmt, ...)
 	debug_locks_off();
 	console_flush_on_panic();
 
+	bust_spinlocks(0);
+
 	if (!panic_blink)
 		panic_blink = no_blink;
 
-- 
2.1.4

Re: [PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[Steven Rostedt]

On Mon,  4 Jun 2018 14:45:57 +0900
Hoeun Ryu <hoeun.ryu@lge.com.com> wrote:

> From: Hoeun Ryu <hoeun.ryu@lge.com>
> 
>  Many console device drivers hold the uart_port->lock spinlock with irq enabled
> (using spin_lock()) while the device drivers are writing characters to their devices,
> but the device drivers just try to hold the spin lock (using spin_trylock()) if
> "oops_in_progress" is equal or greater than 1 to avoid deadlocks.
> 
>  There is a case ocurring a deadlock related to the lock and oops_in_progress. A CPU
> could be stopped by smp_send_stop() while it was holding the port lock because irq was
> enabled. Once a CPU stops, it doesn't respond interrupts anymore and the lock stays
> locked forever.
> 
>  console_flush_on_panic() is called during panic() and it eventually holds the uart
> lock but the lock is held by another stopped CPU and it is a deadlock. By moving
> bust_spinlocks(0) after console_flush_on_panic(), let the console device drivers
> think the Oops is still in progress to call spin_trylock() instead of spin_lock() to
> avoid the deadlock.
> 
> Signed-off-by: Hoeun Ryu <hoeun.ryu@lge.com>
> ---
>  kernel/panic.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/kernel/panic.c b/kernel/panic.c
> index 42e4874..b4063b6 100644
> --- a/kernel/panic.c
> +++ b/kernel/panic.c
> @@ -233,8 +233,6 @@ void panic(const char *fmt, ...)
>  	if (_crash_kexec_post_notifiers)
>  		__crash_kexec(NULL);
>  
> -	bust_spinlocks(0);
> -
>  	/*
>  	 * We may have ended up stopping the CPU holding the lock (in
>  	 * smp_send_stop()) while still having some valuable data in the console
> @@ -246,6 +244,8 @@ void panic(const char *fmt, ...)
>  	debug_locks_off();
>  	console_flush_on_panic();
>  
> +	bust_spinlocks(0);

Added a few more to Cc. This looks like it could have subtle
side-effects. I'd like those that have been touching the code around
here to have a look.

-- Steve


> +
>  	if (!panic_blink)
>  		panic_blink = no_blink;
>  

RE: [PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[Hoeun Ryu]

I misunderstood the cause of a deadlock.
I sent v2 fixing the commit message about the reason of the deadlock.
Please ignore this and review v2.
Thank you.

> -----Original Message-----
> From: Steven Rostedt [mailto:rostedt@goodmis.org]
> Sent: Tuesday, June 05, 2018 10:44 AM
> To: Hoeun Ryu <hoeun.ryu@lge.com.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>; Kees Cook
> <keescook@chromium.org>; Borislav Petkov <bp@suse.de>; Andi Kleen
> <ak@linux.intel.com>; Josh Poimboeuf <jpoimboe@redhat.com>; Hoeun Ryu
> <hoeun.ryu@lge.com>; linux-kernel@vger.kernel.org; Tejun Heo
> <tj@kernel.org>; Vitaly Kuznetsov <vkuznets@redhat.com>
> Subject: Re: [PATCH] panic: move bust_spinlocks(0) after
> console_flush_on_panic() to avoid deadlocks
> 
> On Mon,  4 Jun 2018 14:45:57 +0900
> Hoeun Ryu <hoeun.ryu@lge.com.com> wrote:
> 
> > From: Hoeun Ryu <hoeun.ryu@lge.com>
> >
> >  Many console device drivers hold the uart_port->lock spinlock with irq
> enabled
> > (using spin_lock()) while the device drivers are writing characters to
> their devices,
> > but the device drivers just try to hold the spin lock (using
> spin_trylock()) if
> > "oops_in_progress" is equal or greater than 1 to avoid deadlocks.
> >
> >  There is a case ocurring a deadlock related to the lock and
> oops_in_progress. A CPU
> > could be stopped by smp_send_stop() while it was holding the port lock
> because irq was
> > enabled. Once a CPU stops, it doesn't respond interrupts anymore and the
> lock stays
> > locked forever.
> >
> >  console_flush_on_panic() is called during panic() and it eventually
> holds the uart
> > lock but the lock is held by another stopped CPU and it is a deadlock.
> By moving
> > bust_spinlocks(0) after console_flush_on_panic(), let the console device
> drivers
> > think the Oops is still in progress to call spin_trylock() instead of
> spin_lock() to
> > avoid the deadlock.
> >
> > Signed-off-by: Hoeun Ryu <hoeun.ryu@lge.com>
> > ---
> >  kernel/panic.c | 4 ++--
> >  1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/kernel/panic.c b/kernel/panic.c
> > index 42e4874..b4063b6 100644
> > --- a/kernel/panic.c
> > +++ b/kernel/panic.c
> > @@ -233,8 +233,6 @@ void panic(const char *fmt, ...)
> >  	if (_crash_kexec_post_notifiers)
> >  		__crash_kexec(NULL);
> >
> > -	bust_spinlocks(0);
> > -
> >  	/*
> >  	 * We may have ended up stopping the CPU holding the lock (in
> >  	 * smp_send_stop()) while still having some valuable data in the
> console
> > @@ -246,6 +244,8 @@ void panic(const char *fmt, ...)
> >  	debug_locks_off();
> >  	console_flush_on_panic();
> >
> > +	bust_spinlocks(0);
> 
> Added a few more to Cc. This looks like it could have subtle
> side-effects. I'd like those that have been touching the code around
> here to have a look.
> 
> -- Steve
> 
> 
> > +
> >  	if (!panic_blink)
> >  		panic_blink = no_blink;
> >

Re: [PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[Andrew Morton]

On Mon,  4 Jun 2018 14:45:57 +0900 Hoeun Ryu <hoeun.ryu@lge.com.com> wrote:

> From: Hoeun Ryu <hoeun.ryu@lge.com>
> 
>  Many console device drivers hold the uart_port->lock spinlock with irq enabled
> (using spin_lock()) while the device drivers are writing characters to their devices,
> but the device drivers just try to hold the spin lock (using spin_trylock()) if
> "oops_in_progress" is equal or greater than 1 to avoid deadlocks.
> 
>  There is a case ocurring a deadlock related to the lock and oops_in_progress. A CPU
> could be stopped by smp_send_stop() while it was holding the port lock because irq was
> enabled. Once a CPU stops, it doesn't respond interrupts anymore and the lock stays
> locked forever.
> 
>  console_flush_on_panic() is called during panic() and it eventually holds the uart
> lock but the lock is held by another stopped CPU and it is a deadlock. By moving
> bust_spinlocks(0) after console_flush_on_panic(), let the console device drivers
> think the Oops is still in progress to call spin_trylock() instead of spin_lock() to
> avoid the deadlock.

hm.  Sergey, is this at all related to the UART printk deadlock change
which you're presently discussing in
http://lkml.kernel.org/r/20180615093919.559-1-sergey.senozhatsky@gmail.com?

> --- a/kernel/panic.c
> +++ b/kernel/panic.c
> @@ -233,8 +233,6 @@ void panic(const char *fmt, ...)
>  	if (_crash_kexec_post_notifiers)
>  		__crash_kexec(NULL);
>  
> -	bust_spinlocks(0);
> -
>  	/*
>  	 * We may have ended up stopping the CPU holding the lock (in
>  	 * smp_send_stop()) while still having some valuable data in the console
> @@ -246,6 +244,8 @@ void panic(const char *fmt, ...)
>  	debug_locks_off();
>  	console_flush_on_panic();
>  
> +	bust_spinlocks(0);
> +
>  	if (!panic_blink)
>  		panic_blink = no_blink;
>  

Re: [PATCH] panic: move bust_spinlocks(0) after console_flush_on_panic() to avoid deadlocks

[Sergey Senozhatsky]

On (06/20/18 14:30), Andrew Morton wrote:
> > From: Hoeun Ryu <hoeun.ryu@lge.com>
> > 
> >  Many console device drivers hold the uart_port->lock spinlock with irq enabled
> > (using spin_lock()) while the device drivers are writing characters to their devices,
> > but the device drivers just try to hold the spin lock (using spin_trylock()) if
> > "oops_in_progress" is equal or greater than 1 to avoid deadlocks.
> > 
> >  There is a case ocurring a deadlock related to the lock and oops_in_progress. A CPU
> > could be stopped by smp_send_stop() while it was holding the port lock because irq was
> > enabled. Once a CPU stops, it doesn't respond interrupts anymore and the lock stays
> > locked forever.
> > 
> >  console_flush_on_panic() is called during panic() and it eventually holds the uart
> > lock but the lock is held by another stopped CPU and it is a deadlock. By moving
> > bust_spinlocks(0) after console_flush_on_panic(), let the console device drivers
> > think the Oops is still in progress to call spin_trylock() instead of spin_lock() to
> > avoid the deadlock.
> 
> hm.  Sergey, is this at all related to the UART printk deadlock change
> which you're presently discussing in
> http://lkml.kernel.org/r/20180615093919.559-1-sergey.senozhatsky@gmail.com?

Hi Andrew,

Not exactly. The change I'm discussing is a little different - it's about
re-entrant UART [+ circular locking in TTY], when UART deadlocks us because
of printk()-s issued by MM/tty/WQ/sched/other core kernel stuff/etc

Example:
	IRQ -> uart -> tty -> WQ -> printk -> uart

	-ss