Skip to contentSkip to conversation

csky: smp_mb__after_spinlock

PATCH [email protected] 2 participants 5 messages Aug 5, 2020 – Aug 7, 2020

Subsystems (2)
lore
Share & tools
lore

[email protected]

author maintainer
#0 lore raw

Hi,

While doing an audit of smp_mb__after_spinlock, I found that csky defines it, why?

CSKY only has smp_mb(), it doesn't override __atomic_acquire_fence or otherwise special cases it's atomic*_acquire() primitives. It has an explicit smp_mb() in its arch_spin_lock().

---

arch/csky/include/asm/spinlock.h +0 -3
index 7cf3f2b34cea..22a05caf2d18 100644
@@ -88,9 +88,6 @@ static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 
 #include <asm/qrwlock.h>
 
-/* See include/linux/spinlock.h */
-#define smp_mb__after_spinlock()	smp_mb()
-
 #else /* CONFIG_QUEUED_RWLOCKS */
 
 /*
author maintainer
#1 lore raw
Quoted 8 lines · On Wed, Aug 05, 2020 at 12:41:46PM +0200, [email protected] wrote:

Hi,

While doing an audit of smp_mb__after_spinlock, I found that csky defines it, why?

CSKY only has smp_mb(), it doesn't override __atomic_acquire_fence or otherwise special cases it's atomic*_acquire() primitives. It has an explicit smp_mb() in its arch_spin_lock().

Also, why have two implementations of all the locking ?

---

arch/csky/Kconfig +1 -1
index bd31ab12f77d..332738e93e57 100644
@@ -7,7 +7,7 @@ config CSKY
 	select ARCH_HAS_SYNC_DMA_FOR_CPU
 	select ARCH_HAS_SYNC_DMA_FOR_DEVICE
 	select ARCH_USE_BUILTIN_BSWAP
-	select ARCH_USE_QUEUED_RWLOCKS if NR_CPUS>2
+	select ARCH_USE_QUEUED_RWLOCKS
 	select ARCH_WANT_FRAME_POINTERS if !CPU_CK610
 	select COMMON_CLK
 	select CLKSRC_MMIO
arch/csky/include/asm/spinlock.h +0 -167

177 lines not on this page · raw message

maintainer Acked-by
#2 lore raw
Quoted 210 lines · On Thu, Aug 6, 2020 at 3:55 AM <[email protected]> wrote:
Quoted 8 lines · On Wed, Aug 05, 2020 at 12:41:46PM +0200, [email protected] wrote:

Hi,

While doing an audit of smp_mb__after_spinlock, I found that csky defines it, why?

CSKY only has smp_mb(), it doesn't override __atomic_acquire_fence or otherwise special cases it's atomic*_acquire() primitives. It has an explicit smp_mb() in its arch_spin_lock().

Also, why have two implementations of all the locking ?

---

arch/csky/Kconfig +1 -1
index bd31ab12f77d..332738e93e57 100644
@@ -7,7 +7,7 @@ config CSKY
        select ARCH_HAS_SYNC_DMA_FOR_CPU
        select ARCH_HAS_SYNC_DMA_FOR_DEVICE
        select ARCH_USE_BUILTIN_BSWAP
-       select ARCH_USE_QUEUED_RWLOCKS if NR_CPUS>2
+       select ARCH_USE_QUEUED_RWLOCKS
        select ARCH_WANT_FRAME_POINTERS if !CPU_CK610
        select COMMON_CLK
        select CLKSRC_MMIO
arch/csky/include/asm/spinlock.h +0 -167
index 7cf3f2b34cea..69f5aa249c5f 100644
@@ -6,8 +6,6 @@
 #include <linux/spinlock_types.h>
 #include <asm/barrier.h>
 
-#ifdef CONFIG_QUEUED_RWLOCKS
-
 /*
  * Ticket-based spin-locking.
  */
@@ -88,169 +86,4 @@ static inline int arch_spin_is_contended(arch_spinlock_t *lock)
 
 #include <asm/qrwlock.h>
 
-/* See include/linux/spinlock.h */
-#define smp_mb__after_spinlock()       smp_mb()
-
-#else /* CONFIG_QUEUED_RWLOCKS */
-
-/*
- * Test-and-set spin-locking.
- */
-static inline void arch_spin_lock(arch_spinlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       bnez            %0, 1b   \n"
-               "       movi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-       smp_mb();
-}
-
-static inline void arch_spin_unlock(arch_spinlock_t *lock)
-{
-       smp_mb();
-       WRITE_ONCE(lock->lock, 0);
-}
-
-static inline int arch_spin_trylock(arch_spinlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       bnez            %0, 2f   \n"
-               "       movi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               "       movi            %0, 0    \n"
-               "2:                              \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-
-       if (!tmp)
-               smp_mb();
-
-       return !tmp;
-}
-
-#define arch_spin_is_locked(x) (READ_ONCE((x)->lock) != 0)
-
-/*
- * read lock/unlock/trylock
- */
-static inline void arch_read_lock(arch_rwlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       blz             %0, 1b   \n"
-               "       addi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-       smp_mb();
-}
-
-static inline void arch_read_unlock(arch_rwlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       smp_mb();
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       subi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-}
-
-static inline int arch_read_trylock(arch_rwlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       blz             %0, 2f   \n"
-               "       addi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               "       movi            %0, 0    \n"
-               "2:                              \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-
-       if (!tmp)
-               smp_mb();
-
-       return !tmp;
-}
-
-/*
- * write lock/unlock/trylock
- */
-static inline void arch_write_lock(arch_rwlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       bnez            %0, 1b   \n"
-               "       subi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-       smp_mb();
-}
-
-static inline void arch_write_unlock(arch_rwlock_t *lock)
-{
-       smp_mb();
-       WRITE_ONCE(lock->lock, 0);
-}
-
-static inline int arch_write_trylock(arch_rwlock_t *lock)
-{
-       u32 *p = &lock->lock;
-       u32 tmp;
-
-       asm volatile (
-               "1:     ldex.w          %0, (%1) \n"
-               "       bnez            %0, 2f   \n"
-               "       subi            %0, 1    \n"
-               "       stex.w          %0, (%1) \n"
-               "       bez             %0, 1b   \n"
-               "       movi            %0, 0    \n"
-               "2:                              \n"
-               : "=&r" (tmp)
-               : "r"(p)
-               : "cc");
-
-       if (!tmp)
-               smp_mb();
-
-       return !tmp;
-}
-
-#endif /* CONFIG_QUEUED_RWLOCKS */
 #endif /* __ASM_CSKY_SPINLOCK_H */
Signature · 4 lines
Best Regards
 Guo Ren

ML: https://lore.kernel.org/linux-csky/
maintainer
#3 lore raw
In reply to [email protected]

Hi Peter,

Quoted 9 lines · On Thu, Aug 6, 2020 at 3:53 AM <[email protected]> wrote:

Hi,

While doing an audit of smp_mb__after_spinlock, I found that csky defines it, why?

CSKY only has smp_mb(), it doesn't override __atomic_acquire_fence or otherwise special cases it's atomic*_acquire() primitives. It has an explicit smp_mb() in its arch_spin_lock().

Yes, csky didn't implement ACQUIRE/RELEASE in spinlock.h. So it is a redundant and side-effect wrong macro, we should remove it to fixup.

TODO:
- implement csky's ACQUIRE/RELEASE barrier

Quoted 1 line

Also, why have two implementations of all the locking?

I just kept my baby's codes :P. Now, we could remove it and just keep the ticket's one.

BTW, I want to change the spinlock to qspinlock, but csky only has 32-bit atomic operation in hardware.

Any plan to deal with this in spinlock?

Maybe for the embedded scenario, qspinlock seems a bit heavy, are any tickets-like comm spinlock infrastructures in the plan?

--
Best Regards
Guo Ren

ML: https://lore.kernel.org/linux-csky/

author maintainer
#4 lore raw
Quoted 13 lines · Guo Ren

Hi Peter,

Quoted 9 lines · On Thu, Aug 6, 2020 at 3:53 AM <[email protected]> wrote:

Hi,

While doing an audit of smp_mb__after_spinlock, I found that csky defines it, why?

CSKY only has smp_mb(), it doesn't override __atomic_acquire_fence or otherwise special cases it's atomic*_acquire() primitives. It has an explicit smp_mb() in its arch_spin_lock().

Quoted 5 lines

Yes, csky didn't implement ACQUIRE/RELEASE in spinlock.h. So it is a redundant and side-effect wrong macro, we should remove it to fixup.

TODO:
- implement csky's ACQUIRE/RELEASE barrier

Fair enough; please Cc me when you get to it.

Quoted 7 lines
Quoted 1 line

Also, why have two implementations of all the locking?

I just kept my baby's codes :P. Now, we could remove it and just keep the ticket's one.

BTW, I want to change the spinlock to qspinlock, but csky only has 32-bit atomic operation in hardware.

Yeah, that's okay, you can do 'short' atomics using wider RmW.

Quoted 3 lines

Any plan to deal with this in spinlock?

Maybe for the embedded scenario, qspinlock seems a bit heavy,

I tend to agree, qspinlock only really makes sense when you have multiple cache domains and NUMA makes it shine.

Quoted 2 lines

are any
tickets-like comm spinlock infrastructures in the plan?

No. ticket locks are 'simple' and their implementation fairly straight forward. Also, since the generic code only has cmpxchg, a generic ticket lock would be sub-optimal on LL/SC (or even x86 for that matter).

It would look something like this, which I think is pretty terrible for just about any arch compared to what you can do with inline asm.

struct ticket_lock {
	union {
		atomic_t	val;
		struct {
#ifdef __LITTLE_ENDIAN
			u16	head;
			u16	tail;
#else
			u16	tail;
			u16	head;
#endif
		};
	};
};

void ticket_lock(struct ticket_lock *lock)

{
	unsigned int val = atomic_read(&lock->val);
	struct ticket_lock lock;

	do {
		lock.val = ATOMIC_INIT(val);
		lock.head++;
	} while (!atomic_try_cmpxchg_relaxed(&lock->lock, &val, lock.val));

	atomic_cond_read_acquire(&lock->val, (VAL >> 16) == lock.head);
}

void ticket_unlock(struct ticket_lock *lock)

{
	atomic_fetch_inc_release(&lock->val);
}