From: Guo Ren <guoren@linux.alibaba.com>

The current cond_load primitive contains two parts (condition expression
and load value), but the usage of cond_load may require the sub-size
condition expression of the load size. That means hardware could utilize
a mask argument to optimize the wait condition. If the mask argument
size is less than the hardware minimum wait size, the hardware uses its
minimum size.

The patch contains a qspinlock example: When it is at the head of the
waitqueue, it waits for the owner & pending to go away. The forward
progress condition only cares locked_pending part, but it needs to load
the 32-bit lock value as a return.

That also means WFE-liked instruction would need a mask argument of the
load reservation set.

Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
Signed-off-by: Guo Ren <guoren@kernel.org>
Cc: Waiman Long <longman@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
---
 include/asm-generic/barrier.h | 22 ++++++++++++++++++++++
 include/linux/atomic.h        |  4 ++++
 kernel/locking/qspinlock.c    |  3 ++-
 3 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/include/asm-generic/barrier.h b/include/asm-generic/barrier.h
index 961f4d88f9ef..fec61629f769 100644
--- a/include/asm-generic/barrier.h
+++ b/include/asm-generic/barrier.h
@@ -275,6 +275,28 @@ do {									\
 })
 #endif
 
+/**
+ * smp_cond_load_mask_relaxed() - (Spin) wait for cond with ACQUIRE ordering
+ * @ptr: pointer to the variable to wait on
+ * @cond: boolean expression to wait for
+ * @mask: mask *ptr to wait for (effect of 0 is the same with -1)
+ */
+#ifndef smp_cond_load_mask_relaxed
+#define smp_cond_load_mask_relaxed(ptr, cond_expr, mask)	\
+	smp_cond_load_relaxed(ptr, cond_expr)
+#endif
+
+/**
+ * smp_cond_load_mask_acquire() - (Spin) wait for cond with ACQUIRE ordering
+ * @ptr: pointer to the variable to wait on
+ * @cond: boolean expression to wait for
+ * @mask: mask *ptr to wait for (effect of 0 is the same with -1)
+ */
+#ifndef smp_cond_load_mask_acquire
+#define smp_cond_load_mask_acquire(ptr, cond_expr, mask)	\
+	smp_cond_load_acquire(ptr, cond_expr)
+#endif
+
 /*
  * pmem_wmb() ensures that all stores for which the modification
  * are written to persistent storage by preceding instructions have
diff --git a/include/linux/atomic.h b/include/linux/atomic.h
index 8dd57c3a99e9..dc7351945f27 100644
--- a/include/linux/atomic.h
+++ b/include/linux/atomic.h
@@ -27,9 +27,13 @@
 
 #define atomic_cond_read_acquire(v, c) smp_cond_load_acquire(&(v)->counter, (c))
 #define atomic_cond_read_relaxed(v, c) smp_cond_load_relaxed(&(v)->counter, (c))
+#define atomic_cond_read_mask_acquire(v, c, m) smp_cond_load_mask_acquire(&(v)->counter, (c), (m))
+#define atomic_cond_read_mask_relaxed(v, c, m) smp_cond_load_mask_relaxed(&(v)->counter, (c), (m))
 
 #define atomic64_cond_read_acquire(v, c) smp_cond_load_acquire(&(v)->counter, (c))
 #define atomic64_cond_read_relaxed(v, c) smp_cond_load_relaxed(&(v)->counter, (c))
+#define atomic64_cond_read_mask_acquire(v, c, m) smp_cond_load_mask_acquire(&(v)->counter, (c), (m))
+#define atomic64_cond_read_mask_relaxed(v, c, m) smp_cond_load_mask_relaxed(&(v)->counter, (c), (m))
 
 /*
  * The idea here is to build acquire/release variants by adding explicit
diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
index ebe6b8ec7cb3..14fdd2ee752c 100644
--- a/kernel/locking/qspinlock.c
+++ b/kernel/locking/qspinlock.c
@@ -511,7 +511,8 @@ void __lockfunc queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
 	if ((val = pv_wait_head_or_lock(lock, node)))
 		goto locked;
 
-	val = atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK));
+	val = atomic_cond_read_mask_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK),
+					    _Q_LOCKED_PENDING_MASK);
 
 locked:
 	/*
-- 
2.36.1

On 12/25/22 06:55, guoren@kernel.org wrote:
> From: Guo Ren <guoren@linux.alibaba.com>
>
> The current cond_load primitive contains two parts (condition expression
> and load value), but the usage of cond_load may require the sub-size
> condition expression of the load size. That means hardware could utilize
> a mask argument to optimize the wait condition. If the mask argument
> size is less than the hardware minimum wait size, the hardware uses its
> minimum size.
>
> The patch contains a qspinlock example: When it is at the head of the
> waitqueue, it waits for the owner & pending to go away. The forward
> progress condition only cares locked_pending part, but it needs to load
> the 32-bit lock value as a return.
>
> That also means WFE-liked instruction would need a mask argument of the
> load reservation set.
>
> Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
> Signed-off-by: Guo Ren <guoren@kernel.org>
> Cc: Waiman Long <longman@redhat.com>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Boqun Feng <boqun.feng@gmail.com>
> Cc: Will Deacon <will@kernel.org>
> Cc: Ingo Molnar <mingo@redhat.com>
> ---
>   include/asm-generic/barrier.h | 22 ++++++++++++++++++++++
>   include/linux/atomic.h        |  4 ++++
>   kernel/locking/qspinlock.c    |  3 ++-
>   3 files changed, 28 insertions(+), 1 deletion(-)
>
> diff --git a/include/asm-generic/barrier.h b/include/asm-generic/barrier.h
> index 961f4d88f9ef..fec61629f769 100644
> --- a/include/asm-generic/barrier.h
> +++ b/include/asm-generic/barrier.h
> @@ -275,6 +275,28 @@ do {									\
>   })
>   #endif
>   
> +/**
> + * smp_cond_load_mask_relaxed() - (Spin) wait for cond with ACQUIRE ordering
> + * @ptr: pointer to the variable to wait on
> + * @cond: boolean expression to wait for
> + * @mask: mask *ptr to wait for (effect of 0 is the same with -1)
> + */
> +#ifndef smp_cond_load_mask_relaxed
> +#define smp_cond_load_mask_relaxed(ptr, cond_expr, mask)	\
> +	smp_cond_load_relaxed(ptr, cond_expr)
> +#endif
> +
> +/**
> + * smp_cond_load_mask_acquire() - (Spin) wait for cond with ACQUIRE ordering
> + * @ptr: pointer to the variable to wait on
> + * @cond: boolean expression to wait for
> + * @mask: mask *ptr to wait for (effect of 0 is the same with -1)
> + */
> +#ifndef smp_cond_load_mask_acquire
> +#define smp_cond_load_mask_acquire(ptr, cond_expr, mask)	\
> +	smp_cond_load_acquire(ptr, cond_expr)
> +#endif
> +
>   /*
>    * pmem_wmb() ensures that all stores for which the modification
>    * are written to persistent storage by preceding instructions have
> diff --git a/include/linux/atomic.h b/include/linux/atomic.h
> index 8dd57c3a99e9..dc7351945f27 100644
> --- a/include/linux/atomic.h
> +++ b/include/linux/atomic.h
> @@ -27,9 +27,13 @@
>   
>   #define atomic_cond_read_acquire(v, c) smp_cond_load_acquire(&(v)->counter, (c))
>   #define atomic_cond_read_relaxed(v, c) smp_cond_load_relaxed(&(v)->counter, (c))
> +#define atomic_cond_read_mask_acquire(v, c, m) smp_cond_load_mask_acquire(&(v)->counter, (c), (m))
> +#define atomic_cond_read_mask_relaxed(v, c, m) smp_cond_load_mask_relaxed(&(v)->counter, (c), (m))
>   
>   #define atomic64_cond_read_acquire(v, c) smp_cond_load_acquire(&(v)->counter, (c))
>   #define atomic64_cond_read_relaxed(v, c) smp_cond_load_relaxed(&(v)->counter, (c))
> +#define atomic64_cond_read_mask_acquire(v, c, m) smp_cond_load_mask_acquire(&(v)->counter, (c), (m))
> +#define atomic64_cond_read_mask_relaxed(v, c, m) smp_cond_load_mask_relaxed(&(v)->counter, (c), (m))
>   
>   /*
>    * The idea here is to build acquire/release variants by adding explicit
> diff --git a/kernel/locking/qspinlock.c b/kernel/locking/qspinlock.c
> index ebe6b8ec7cb3..14fdd2ee752c 100644
> --- a/kernel/locking/qspinlock.c
> +++ b/kernel/locking/qspinlock.c
> @@ -511,7 +511,8 @@ void __lockfunc queued_spin_lock_slowpath(struct qspinlock *lock, u32 val)
>   	if ((val = pv_wait_head_or_lock(lock, node)))
>   		goto locked;
>   
> -	val = atomic_cond_read_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK));
> +	val = atomic_cond_read_mask_acquire(&lock->val, !(VAL & _Q_LOCKED_PENDING_MASK),
> +					    _Q_LOCKED_PENDING_MASK);
>   
>   locked:
>   	/*

This patch is essentially a no-op. You will have to have at least one 
arch that has its own version of smp_cond_load_mask*() and get some 
benefit out of it. Otherwise, it is not likely to be merged.

Cheers,
Longman