Re: [Xen-devel] [PATCHv3 3/4] xen: use ticket locks for spin locks

["Jan Beulich" <JBeulich@xxxxxxxx>]

>>> On 21.04.15 at 12:11, <david.vrabel@xxxxxxxxxx> wrote:
> @@ -213,27 +211,32 @@ int _spin_trylock(spinlock_t *lock)
>  
>  void _spin_barrier(spinlock_t *lock)
>  {
> +    spinlock_tickets_t sample;
>  #ifdef LOCK_PROFILE
>      s_time_t block = NOW();
> -    u64      loop = 0;
> +#endif
>  
>      check_barrier(&lock->debug);
> -    do { smp_mb(); loop++;} while ( _raw_spin_is_locked(&lock->raw) );
> -    if ((loop > 1) && lock->profile)
> +    sample = observe_lock(&lock->tickets);
> +    if (sample.head != sample.tail)

Even if the old code didn't conform to coding style, please make sure
the new code does (a few more further down).

> @@ -127,8 +125,16 @@ struct lock_profile_qhead { };
>  
>  #endif
>  
> +typedef union {
> +    u32 head_tail;
> +    struct {
> +        u16 head;
> +        u16 tail;
> +    };
> +} spinlock_tickets_t;
> +
>  typedef struct spinlock {
> -    raw_spinlock_t raw;
> +    spinlock_tickets_t tickets;

At least for x86 this means a growth of this and hence various
other structures - did you examine the effects thereof? Of
course otoh getting the lock size uniform across architectures
is a good thing.

Jan


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel