2013-03-22 11:59:08

by Zhu Yanhai

Subject: How does spin_unlock() in x86-64 align with the description in Documention/memory-barriers.txt?

Hi all,
In the documention it reads,

(2) UNLOCK operation implication:

Memory operations issued before the UNLOCK will be completed before the
UNLOCK operation has completed.

Memory operations issued after the UNLOCK may be completed before the
UNLOCK operation has completed.

However, on x86-64 __ticket_spin_unlock() merely does,

static __always_inline void __ticket_spin_unlock(raw_spinlock_t *lock)
{
asm volatile(
ALTERNATIVE(UNLOCK_LOCK_PREFIX"incb (%0);"ASM_NOP3,
UNLOCK_LOCK_ALT_PREFIX"movw $0, (%0)",
X86_FEATURE_UNFAIR_SPINLOCK)
:
: "Q" (&lock->slock)
: "memory", "cc");
}

While both UNLOCK_LOCK_PREFIX and UNLOCK_LOCK_ALT_PREFIX are empty
strings. So how such a function keeps the memory operations issued
before it completed?

--
Thanks,
Zhu Yanhai

2013-03-22 12:15:06

by Jan Beulich

Subject: Re: How does spin_unlock() in x86-64 align with the description in Documention/memory-barriers.txt?

>>> On 22.03.13 at 12:58, Zhu Yanhai <[email protected]> wrote:
> Hi all,
> In the documention it reads,
>
> (2) UNLOCK operation implication:
>
> Memory operations issued before the UNLOCK will be completed before the
> UNLOCK operation has completed.
>
> Memory operations issued after the UNLOCK may be completed before the
> UNLOCK operation has completed.
>
> However, on x86-64 __ticket_spin_unlock() merely does,
>
> static __always_inline void __ticket_spin_unlock(raw_spinlock_t *lock)
> {
> asm volatile(
> ALTERNATIVE(UNLOCK_LOCK_PREFIX"incb (%0);"ASM_NOP3,
> UNLOCK_LOCK_ALT_PREFIX"movw $0, (%0)",
> X86_FEATURE_UNFAIR_SPINLOCK)
> :
> : "Q" (&lock->slock)
> : "memory", "cc");
> }
>
> While both UNLOCK_LOCK_PREFIX and UNLOCK_LOCK_ALT_PREFIX are empty
> strings. So how such a function keeps the memory operations issued
> before it completed?

Please read the section "Memory Ordering in P6 and More Recent
Processor Families" in SDM Vol 3.

Jan