Virtualization - Apr 2014 - [PATCH v9 05/19] qspinlock: Optimize for smaller NR_CPUS

On Thu, Apr 17, 2014 at 11:03:57AM -0400, Waiman Long
wrote:
> +struct __qspinlock {
> +	union {
> +		atomic_t val;
> +		struct {
> +#ifdef __LITTLE_ENDIAN
> +			u16	locked_pending;
> +			u16	tail;
> +#else
> +			u16	tail;
> +			u16	locked_pending;
> +#endif
> +		};
> +	};
> +};
> +
> +/**
> + * clear_pending_set_locked - take ownership and clear the pending bit.
> + * @lock: Pointer to queue spinlock structure
> + * @val : Current value of the queue spinlock 32-bit word
> + *
> + * *,1,0 -> *,0,1
> + */
> +static __always_inline void
> +clear_pending_set_locked(struct qspinlock *lock, u32 val)
> +{
> +	struct __qspinlock *l = (void *)lock;
> +
> +	ACCESS_ONCE(l->locked_pending) = 1;
You lost the __constant_le16_to_cpu(_Q_LOCKED_VAL) there. The
unconditional 1 is wrong. You also have to flip the bytes in
locked_pending.

On 04/17/2014 11:56 AM, Peter Zijlstra wrote:
> On Thu, Apr 17, 2014 at 11:03:57AM -0400, Waiman Long wrote:
>> +struct __qspinlock {
>> +	union {
>> +		atomic_t val;
>> +		struct {
>> +#ifdef __LITTLE_ENDIAN
>> +			u16	locked_pending;
>> +			u16	tail;
>> +#else
>> +			u16	tail;
>> +			u16	locked_pending;
>> +#endif
>> +		};
>> +	};
>> +};
>> +
>> +/**
>> + * clear_pending_set_locked - take ownership and clear the pending
bit.
>> + * @lock: Pointer to queue spinlock structure
>> + * @val : Current value of the queue spinlock 32-bit word
>> + *
>> + * *,1,0 ->  *,0,1
>> + */
>> +static __always_inline void
>> +clear_pending_set_locked(struct qspinlock *lock, u32 val)
>> +{
>> +	struct __qspinlock *l = (void *)lock;
>> +
>> +	ACCESS_ONCE(l->locked_pending) = 1;
> You lost the __constant_le16_to_cpu(_Q_LOCKED_VAL) there. The
> unconditional 1 is wrong. You also have to flip the bytes in
> locked_pending.
I don't think that is wrong. The lock byte is in the least significant 8 
bits and the pending byte is the next higher significant 8 bits 
irrespective of the endian-ness. So a value of 1 in a 16-bit context 
means the lock byte is set, but the pending byte is cleared. The name 
"locked_pending" doesn't mean that locked variable is in a lower
address
than pending.

-Longman

On Thu, Apr 17, 2014 at 05:46:27PM -0400, Waiman Long
wrote:
> On 04/17/2014 11:56 AM, Peter Zijlstra wrote:
> >On Thu, Apr 17, 2014 at 11:03:57AM -0400, Waiman Long wrote:
> >>+struct __qspinlock {
> >>+	union {
> >>+		atomic_t val;
		char bytes[4];
> >>+		struct {
> >>+#ifdef __LITTLE_ENDIAN
> >>+			u16	locked_pending;
> >>+			u16	tail;
> >>+#else
> >>+			u16	tail;
> >>+			u16	locked_pending;
> >>+#endif
> >>+		};
		struct {
#ifdef __LITTLE_ENDIAN
			u8	locked;
#else
			u8	res[3];
			u8	locked;
#endif
		};
> >>+	};
> >>+};
> >>+
> >>+/**
> >>+ * clear_pending_set_locked - take ownership and clear the pending
bit.
> >>+ * @lock: Pointer to queue spinlock structure
> >>+ * @val : Current value of the queue spinlock 32-bit word
> >>+ *
> >>+ * *,1,0 ->  *,0,1
> >>+ */
> >>+static __always_inline void
> >>+clear_pending_set_locked(struct qspinlock *lock, u32 val)
> >>+{
> >>+	struct __qspinlock *l = (void *)lock;
> >>+
> >>+	ACCESS_ONCE(l->locked_pending) = 1;
> >You lost the __constant_le16_to_cpu(_Q_LOCKED_VAL) there. The
> >unconditional 1 is wrong. You also have to flip the bytes in
> >locked_pending.
> 
> I don't think that is wrong. The lock byte is in the least significant
8
> bits and the pending byte is the next higher significant 8 bits
irrespective
> of the endian-ness. So a value of 1 in a 16-bit context means the lock byte
> is set, but the pending byte is cleared. The name
"locked_pending" doesn't
> mean that locked variable is in a lower address than pending.
val            is LE bytes[0,1,2,3] BE [3,2,1,0]
locked_pending is LE bytes[0,1]     BE     [1,0]
locked            LE bytes[0]       BE       [0]

That does mean that the LSB of BE locked_pending is bytes[1].
So if you do BE: locked_pending = 1, you set bytes[1], not bytes[0].