Re: [PATCH v3 3/5] locking/qspinlock: Introduce CNA into the slow path of qspinlock

From: Waiman Long
Date: Wed Jul 17 2019 - 13:58:48 EST


On 7/17/19 1:44 PM, Alex Kogan wrote:
>> On Jul 16, 2019, at 10:50 AM, Waiman Long <longman@xxxxxxxxxx> wrote:
>>
>> On 7/16/19 10:29 AM, Alex Kogan wrote:
>>>> On Jul 15, 2019, at 7:22 PM, Waiman Long <longman@xxxxxxxxxx
>>>> <mailto:longman@xxxxxxxxxx>> wrote:
>>>>
>>>> On 7/15/19 5:30 PM, Waiman Long wrote:
>>>>>> -#ifndef _GEN_PV_LOCK_SLOWPATH
>>>>>> +#if !defined(_GEN_PV_LOCK_SLOWPATH) && !defined(_GEN_CNA_LOCK_SLOWPATH)
>>>>>>
>>>>>> #include <linux/smp.h>
>>>>>> #include <linux/bug.h>
>>>>>> @@ -77,18 +77,14 @@
>>>>>> #define MAX_NODES 4
>>>>>>
>>>>>> /*
>>>>>> - * On 64-bit architectures, the mcs_spinlock structure will be 16 bytes in
>>>>>> - * size and four of them will fit nicely in one 64-byte cacheline. For
>>>>>> - * pvqspinlock, however, we need more space for extra data. To accommodate
>>>>>> - * that, we insert two more long words to pad it up to 32 bytes. IOW, only
>>>>>> - * two of them can fit in a cacheline in this case. That is OK as it is rare
>>>>>> - * to have more than 2 levels of slowpath nesting in actual use. We don't
>>>>>> - * want to penalize pvqspinlocks to optimize for a rare case in native
>>>>>> - * qspinlocks.
>>>>>> + * On 64-bit architectures, the mcs_spinlock structure will be 20 bytes in
>>>>>> + * size. For pvqspinlock or the NUMA-aware variant, however, we need more
>>>>>> + * space for extra data. To accommodate that, we insert two more long words
>>>>>> + * to pad it up to 36 bytes.
>>>>>> */
>>>>> The 20 bytes figure is wrong. It is actually 24 bytes for 64-bit as the
>>>>> mcs_spinlock structure is 8-byte aligned. For better cacheline
>>>>> alignment, I will like to keep mcs_spinlock to 16 bytes as before.
>>>>> Instead, you can use encode_tail() to store the CNA node pointer in
>>>>> "locked". For instance, use (encode_tail() << 1) in locked to
>>>>> distinguish it from the regular locked=1 value.
>>>> Actually, the encoded tail value is already shift left either 16 bits
>>>> or 9 bits. So there is no need to shift it. You can assigned it directly:
>>>>
>>>> mcs->locked = cna->encoded_tail;
>>>>
>>>> You do need to change the type of locked to "unsigned int", though,
>>>> for proper comparison with "1".
>>>>
>>> Got it, thanks.
>>>
>> I forgot to mention that I would like to see a boot command line option
>> to force off and maybe on as well the numa qspinlock code. This can help
>> in testing as you don't need to build 2 separate kernels, one with
>> NUMA_AWARE_SPINLOCKS on and one with it off.
> IIUC it should be easy to add a boot option to force off the NUMA-aware spinlock
> even if it is enabled though config, but the other way around would require
> compiling in the NUMA-aware spinlock stuff even if the config option is disabled.
> Is that ok?

That is not what I am looking for. If the config option is disabled, the
boot command line option is disabled also. For the on case, one possible
usage scenario is with a VM guest where all the vcpus are pinned to a
physical CPUs with no over-commit and correct numa information. In that
case, one may want to use numa-qspinlock instead of pv-qspinlock.

> Also, what should the option name be?
> "numa_spinlock=on/offâ if we want both ways, or âno_numa_spinlock" if we want just the âforce offâ option?

I think "numa_spinlock=on/off" will be good. The default is "auto" where
it will be turned on when there is more than one numa nodes in the system.

>> For small 2-socket systems,
>> numa qspinlock may not help much.
> It actually helps quite a bit (e.g., speedup of up to 42-57% for will-it-scale on a dual-socket x86 system).
> We have numbers and plots in our paper on arxiv.

I am talking about older 2-socket systems where each socket may have
just a few cpus. Also some Intel CPU can be configured to have 2 numa
nodes per socket. For AMD EPYC, it can be configured to have 4 numa
nodes per socket.

Cheers,
Longman