Re: [PATCH V2 11/19] csky: Atomic operations

[Guo Ren]

On Thu, Jul 05, 2018 at 07:59:02PM +0200, Peter Zijlstra wrote:
> On Mon, Jul 02, 2018 at 01:30:14AM +0800, Guo Ren wrote:
> 
> > +static inline void arch_spin_lock(arch_spinlock_t *lock)
> > +{
> > +	unsigned int *p = &lock->lock;
> > +	unsigned int 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)
> > +		: "memory");
> > +	smp_mb();
> > +}
> 
> Test-and-set with MB acting as ACQUIRE, ok.
Em ... Ok, I'll try to use test-and-set function instead of it.

> > +static inline void arch_spin_unlock(arch_spinlock_t *lock)
> > +{
> > +	unsigned int *p = &lock->lock;
> > +	unsigned int tmp;
> > +
> > +	smp_mb();
> > +	asm volatile (
> > +		"1:	ldex.w		%0, (%1) \n"
> > +		"	movi		%0, 0    \n"
> > +		"	stex.w		%0, (%1) \n"
> > +		"	bez		%0, 1b   \n"
> > +		: "=&r" (tmp)
> > +		: "r"(p)
> > +		: "memory");
> > +}
> 
> MB acting for RELEASE, but _why_ are you using a LDEX/STEX to clear the
> lock word? Would not a normal store work?
Normal store is enough, I'll fixup it in next version patch.
 
> Also, the fact that you need MB for release implies your LDEX does not
> in fact imply anything and your xchg/cmpxchg implementation is broken.
xchg/cmxchg broken without 1th smp_mb()? Why we need protect the
instructions flow before the ldex.w?

> > +static inline int arch_spin_trylock(arch_spinlock_t *lock)
> > +{
> > +	unsigned int *p = &lock->lock;
> > +	unsigned int 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)
> > +		: "memory");
> > +	smp_mb();
> > +
> > +	return !tmp;
> > +}
> 
> Strictly speaking you can avoid the MB on failure. You only need to
> provide ACQUIRE semantics on success.
> 
> That said, I would really suggest you implement a ticket lock instead of
> a test-and-set lock. They're not really all that complicated and do
> provide better worst case behaviour.
Ok, I'll try to implement ticket lock in next version patch.

> 
> 
> > +/****** read lock/unlock/trylock ******/
> 
> Please have a look at using qrwlock -- esp. if you implement a ticket
> lock, then the rwlock comes for 'free'.
Ok, I'll try it.

 Guo Ren