Re: [PATCH] locking/qrwlock: fix write unlock issue in big endian

From: Waiman Long
Date: Fri Jun 03 2016 - 16:58:39 EST


On 06/03/2016 03:17 AM, xinhui wrote:

On 2016å06æ02æ 19:02, Peter Zijlstra wrote:
On Thu, Jun 02, 2016 at 12:44:51PM +0200, Arnd Bergmann wrote:
On Thursday, June 2, 2016 6:09:08 PM CEST Pan Xinhui wrote:
diff --git a/include/asm-generic/qrwlock.h b/include/asm-generic/qrwlock.h
index 54a8e65..eadd7a3 100644
--- a/include/asm-generic/qrwlock.h
+++ b/include/asm-generic/qrwlock.h
@@ -139,7 +139,7 @@ static inline void queued_read_unlock(struct qrwlock *lock)
*/
static inline void queued_write_unlock(struct qrwlock *lock)
{
- smp_store_release((u8 *)&lock->cnts, 0);
+ (void)atomic_sub_return_release(_QW_LOCKED, &lock->cnts);
}

Isn't this more expensive than the existing version?

Yes, loads. And while this might be a suitable fix for asm-generic, it
will introduce a fairly large regression on x86 (which is currently the
only user of this).

well, to show respect to struct __qrwlock private field.
We can keep smp_store_release((u8 *)&lock->cnts, 0) in little_endian machine.
as this should be quick and no performance issue to all other archs(although there is only 1 now)

BUT, We need use (void)atomic_sub_return_release(_QW_LOCKED, &lock->cnts) in big_endian machine.
because it's bad to export struct __qrwlock and set its private field to NULL.

How about code like below.

static inline void queued_write_unlock(struct qrwlock *lock)
{
#ifdef __BIG_ENDIAN
(void)atomic_sub_return_release(_QW_LOCKED, &lock->cnts);
#else
smp_store_release((u8 *)&lock->cnts, 0);
#endif
}

BUT I think that would make thing a little complex to understand. :(
So at last, in my opinion, I suggest my patch :)
any thoughts?

Another alternative is to make queued_write_unlock() overrideable from asm/qrwlock.h, just like what we did with queued_spin_unlock().

Cheers,
Longman