lkml.org 
[lkml]   [2022]   [May]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [mm/page_alloc] f26b3fa046: netperf.Throughput_Mbps -18.0% regression
From
On 5/10/22 15:03, Linus Torvalds wrote:
> On Tue, May 10, 2022 at 11:47 AM Waiman Long <longman@redhat.com> wrote:>
>> Qspinlock still has one head waiter spinning on the lock. This is much
>> better than the original ticket spinlock where there will be n waiters
>> spinning on the lock.
> Oh, absolutely. I'm not saying we should look at going back. I'm more
> asking whether maybe we could go even further..

We can probably go a bit further, but there will be cost associated with it.

>> That is the cost of a cheap unlock. There is no way to eliminate all
>> lock spinning unless we use MCS lock directly which will require a
>> change in locking API as well as more expensive unlock.
> So there's no question that unlock would be more expensive for the
> contention case, since it would have to always not only clear the lock
> itself, as well as update the noce it points to.
>
> But does it actually require a change in the locking API?

Of course, it is not always necessary to change the locking API.


>
> The qspinlock slowpath already always allocates that mcs node (for
> some definition of "always" - I am obviously ignoring all the trylock
> cases both before and in the slowpath)
>
> But yes, clearly the simply store-release of the current
> queued_spin_unlock() wouldn't work as-is, and maybe the cost of
> replacing it with something else is much more expensive than any
> possible win.
At the minimum, we need to do a read to determine if the lock is
contended and also a write to clear the lock bit. An atomic write after
read is expensive. There is also a possibility of race with non-atomic
read-write. Moreover, the tail stored in the lock can tell you the queue
tail, not its current head. So there is no easy way to notify the
current head waiter to acquire the lock. The only thing I can think of
is to do some kind of proportional backoff for the head waiter by slowly
increasing the time gap with successive spinning attempts on the lock
over time with the cost of an increased lock acquisition latency.
> I think the PV case already basically does that - replacing the the
> "store release" with a much more complex sequence. No?

That is true for pvqspinlock as the unlock is an atomic operation with
increased cost. The original pv ticket spinlock that is now replaced by
pvqspinlock has similar expensive unlock due to the nature of reading
the lock to find out if the head waiter vCPU may have been preempted.

Cheers,
Longman

\
 
 \ /
  Last update: 2022-05-10 21:48    [W:0.270 / U:1.208 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site