On Thu, 9 Nov 2023 at 21:57, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> So something like this should fix lockref. ENTIRELY UNTESTED, except
> now the code generation of lockref_put_return() looks much better,
> without a pointless flush to the stack, and now it has no pointless
> stack frame as a result.
Heh. And because I was looking at Al's tree, I didn't notice that
commit c6f4a9002252 ("asm-generic: ticket-lock: Optimize
arch_spin_value_unlocked()") had solved the ticket spinlock part of
this in this merge window in the meantime.
The qspinlock implementation - which is what x86 uses - is still
broken in mainline, though.
So that part of my patch still stands. Now attached just the small
one-liner part. Adding Ingo and Guo Ren, who did the ticket lock part
(and looks to have done it very similarly to my suggested patch.
Ingo?
Linus
include/asm-generic/qspinlock.h | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/include/asm-generic/qspinlock.h b/include/asm-generic/qspinlock.h
index 995513fa2690..0655aa5b57b2 100644
--- a/include/asm-generic/qspinlock.h
+++ b/include/asm-generic/qspinlock.h
@@ -70,7 +70,7 @@ static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
*/
static __always_inline int queued_spin_value_unlocked(struct qspinlock lock)
{
- return !atomic_read(&lock.val);
+ return !lock.val.counter;
}
/**
