On Tue, Jul 10, 2018 at 12:26:34PM +0800, Huacai Chen wrote:
> Hi, Paul and Peter,
>
> I think we find the real root cause, READ_ONCE() doesn't need any
> barriers, the problematic code is queued_spin_lock_slowpath() in
> kernel/locking/qspinlock.c:
>
> if (old & _Q_TAIL_MASK) {
> prev = decode_tail(old);
>
> /* Link @node into the waitqueue. */
> WRITE_ONCE(prev->next, node);
>
> pv_wait_node(node, prev);
> arch_mcs_spin_lock_contended(&node->locked);
>
> /*
> * While waiting for the MCS lock, the next pointer may have
> * been set by another lock waiter. We optimistically load
> * the next pointer & prefetch the cacheline for writing
> * to reduce latency in the upcoming MCS unlock operation.
> */
> next = READ_ONCE(node->next);
> if (next)
> prefetchw(next);
> }
>
> After WRITE_ONCE(prev->next, node); arch_mcs_spin_lock_contended()
> enter a READ_ONCE() loop, so the effect of WRITE_ONCE() is invisible
> by other cores because of the write buffer.
And _that_ is a hardware bug. Also please explain how that is different
from the ARM bug mentioned elsewhere.
> As a result,
> arch_mcs_spin_lock_contended() will wait for ever because the waiters
> of prev->next will wait for ever. I think the right way to fix this is
> flush SFB after this WRITE_ONCE(), but I don't have a good solution:
> 1, MIPS has wbflush() which can be used to flush SFB, but other archs
> don't have;
Sane archs don't need this.
> 2, Every arch has mb(), and add mb() after WRITE_ONCE() can actually
> solve Loongson's problem, but in syntax, mb() is different from
> wbflush();
Still wrong, because any non-broken arch doesn't need that flush to
begin with.
> 3, Maybe we can define a Loongson-specific WRITE_ONCE(), but not every
> WRITE_ONCE() need wbflush(), we only need wbflush() between
> WRITE_ONCE() and a READ_ONCE() loop.
No no no no ...
So now explain why the cpu_relax() hack that arm did doesn't work for
you?
