Re: [RFC PATCH] spinlock_debug: Save stacktrace at lock acquisition

[Waiman Long]

On 6/29/21 9:28 PM, Guru Das Srinagesh wrote:
> In case of spinlock recursion bugs, knowing which entity acquired the
> lock initially is key to debugging. Therefore, save the stack frames
> when the lock is acquired and print them when spinlock recursion is
> detected.
>
> Signed-off-by: Guru Das Srinagesh <gurus@xxxxxxxxxxxxxx>
> ---
>   include/linux/spinlock_types.h  | 3 +++
>   kernel/locking/spinlock_debug.c | 9 +++++++++
>   2 files changed, 12 insertions(+)
>
> diff --git a/include/linux/spinlock_types.h b/include/linux/spinlock_types.h
> index b981caa..8a68d55 100644
> --- a/include/linux/spinlock_types.h
> +++ b/include/linux/spinlock_types.h
> @@ -22,6 +22,9 @@ typedef struct raw_spinlock {
>   #ifdef CONFIG_DEBUG_SPINLOCK
>   	unsigned int magic, owner_cpu;
>   	void *owner;
> +#define MAX_STACK_LEN 16
> +	int stack_len;
> +	unsigned long stack_trace[MAX_STACK_LEN];
>   #endif

Nak

Locking problem like this should be detected by the lockdep code. Adding 
136 bytes (64-bit archs) to the size of a spinlock is just too big an 
overhead.

Regards,
Longman