On 04/02/2019 05:43 AM, Peter Zijlstra wrote:
> On Mon, Apr 01, 2019 at 10:36:19AM -0400, Waiman Long wrote:
>> On 03/29/2019 11:20 AM, Alex Kogan wrote:
>>> +config NUMA_AWARE_SPINLOCKS
>>> + bool "Numa-aware spinlocks"
>>> + depends on NUMA
>>> + default y
>>> + help
>>> + Introduce NUMA (Non Uniform Memory Access) awareness into
>>> + the slow path of spinlocks.
>>> +
>>> + The kernel will try to keep the lock on the same node,
>>> + thus reducing the number of remote cache misses, while
>>> + trading some of the short term fairness for better performance.
>>> +
>>> + Say N if you want absolute first come first serve fairness.
>>> +
>> The patch that I am looking for is to have a separate
>> numa_queued_spinlock_slowpath() that coexists with
>> native_queued_spinlock_slowpath() and
>> paravirt_queued_spinlock_slowpath(). At boot time, we select the most
>> appropriate one for the system at hand.
> Agreed; and until we have static_call, I think we can abuse the paravirt
> stuff for this.
I haven't checked Josh's patch to see if it is doing. The availability
of static_call will certainly make thing easier for this case.
> By the time we patch the paravirt stuff:
>
> check_bugs()
> alternative_instructions()
> apply_paravirt()
>
> we should already have enumerated the NODE topology and so nr_node_ids()
> should be set.
>
> So if we frob pv_ops.lock.queued_spin_lock_slowpath to
> numa_queued_spin_lock_slowpath before that, it should all get patched
> just right.
>
> That of course means the whole NUMA_AWARE_SPINLOCKS thing depends on
> PARAVIRT_SPINLOCK, which is a bit awkward...
Yes, this is one way of doing it. Another way to use static key to
switch between the native and numa version. So if PARAVIRT_SPINLOCK is
defined, we use the paravirt patching to point to the right function. If
PARAVIRT_SPINLOCK isn't enabled, we can do something like
static inline void queued_spin_lock_slowpath(struct qspinlock *lock, u32
val)
{
if (static_branch_unlikely(&use_numa_spinlock))
numa_queued_spin_lock_slowpath(lock, val);
else
native_queued_spin_lock_slowpath(lock, val);
}
Alternatively, we can also call numa_queued_spin_lock_slowpath() in
native_queued_spin_lock_slowpath() if we don't want to increase the code
size of spinlock call sites.
Cheers,
Longman
