lkml.org 
[lkml]   [2018]   [Jun]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/2] locking: Implement an algorithm choice for Wound-Wait mutexes
From
Date
On 06/13/2018 03:10 PM, Peter Zijlstra wrote:
> On Wed, Jun 13, 2018 at 12:40:29PM +0200, Thomas Hellstrom wrote:
>> On 06/13/2018 11:50 AM, Peter Zijlstra wrote:
>>>> +
>>>> + lockdep_assert_held(&lock->wait_lock);
>>>> +
>>>> + if (owner && hold_ctx && __ww_ctx_stamp_after(hold_ctx, ww_ctx) &&
>>>> + ww_ctx->acquired > 0) {
>>>> + WRITE_ONCE(hold_ctx->wounded, true);
>>>> + if (owner != current) {
>>>> + /*
>>>> + * wake_up_process() inserts a write memory barrier to
>>> It does no such thing. But yes, it does ensure the wakee sees all prior
>>> stores IFF the wakeup happened.
>>>
>>>> + * make sure owner sees it is wounded before
>>>> + * TASK_RUNNING in case it's sleeping on another
>>>> + * ww_mutex. Note that owner points to a valid
>>>> + * task_struct as long as we hold the wait_lock.
>>>> + */
>>> What exactly are you trying to say here ?
>>>
>>> I'm thinking this is the pairing barrier to the smp_mb() below, with
>>> your list_empty() thing? Might make sense to write a single coherent
>>> comment and refer to the other location.
>> So what I'm trying to say here is that wake_up_process() ensures that the
>> owner, if in !TASK_RUNNING, sees the write to hold_ctx->wounded before the
>> transition to TASK_RUNNING. This was how I interpreted "woken up" in the
>> wake up process documentation.
> There is documentation!? :-) Aaah, you mean that kerneldoc comment with
> wake_up_process() ? Yeah, that needs fixing. /me puts on endless todo
> list.
>
> Anyway, wakeup providing that ordering isn't something that needs a
> comment of that size; and I think the only comment here is that we care
> about the ordering and a reference to the site(s) that pairs with it.
>
> Maybe something like:
>
> /*
> * __ww_mutex_lock_check_stamp() will observe our wounded store.
> */

Yes.

Actually, I just found the set_current_state() kerneldoc which explains
the built-in barrier pairing with wake_up_xxx. Perhaps I also should
mention that as well. Looks like the use WRITE_ONCE() and READ_ONCE()
can be dropped as well.

>>>> - if (likely(!(atomic_long_read(&lock->base.owner) & MUTEX_FLAG_WAITERS)))
>>>> + if (likely(list_empty(&lock->base.wait_list)))
>>>> return;
>>>> /*
>>>> @@ -653,6 +695,17 @@ __ww_mutex_lock_check_stamp(struct mutex *lock, struct mutex_waiter *waiter,
>>>> struct ww_acquire_ctx *hold_ctx = READ_ONCE(ww->ctx);
>>>> struct mutex_waiter *cur;
>>>> + /*
>>>> + * If we miss a wounded == true here, we will have a pending
>>> Explain how we can miss that.
>> This is actually the pairing location of the wake_up_process() comment /
>> code discussed above. Here we should have !TASK_RUNNING, and let's say
>> ctx->wounded is set by another process immediately after we've read it (we
>> "miss" it). At that point there must be a pending wake-up-process() for us
>> and we'll pick up the set value of wounded on the next iteration after
>> returning from schedule().
> Right, so that's when the above wakeup isn't the one waking us.
>
>
>>> I can't say I'm a fan. I'm already cursing the ww_mutex stuff every time
>>> I have to look at it, and you just made it worse spagethi.
>> Well, I can't speak for the current ww implementation except I didn't think
>> it was too hard to understand for a first time reader.
>>
>> Admittedly the Wound-Wait path makes it worse since it's a preemptive
>> algorithm and we need to touch other processes a acquire contexts and worry
>> about ordering.
>>
>> So, assuming your review comments are fixed up, is that a solid NAK or do
>> you have any suggestion that would make you more comfortable with the code?
>> like splitting out ww-stuff to a separate file?
> Nah, not a NAK, but we should look at whan can be done to improve code.
> Maybe add a few more comments that explain why. Part of the problem with
> ww_mutex is always that I forget exactly how they work and mutex.c
> doesn't have much useful comments in (most of those are in ww_mutex.h
> and I always forget to look there).

Understood.

>
> Also; I'm not at all sure about the exact difference between what we
> have and what you propose. I did read the documentation part (I really
> should not have to) but it just doesn't jive.
>
> I suspect you're using preemption entirely different from what we
> usually call a preemption.

I think that perhaps requires a good understanding of the difference of
the algorithms in question before looking at the implementation. I put a
short explanation and some URLs to CS websites describing the two
algorithms and their pros and cons in the patch series introductory
message. I'll forward that.

In short, with Wait-Die (before the patch) it's the process _taking_ the
contended lock that backs off if necessary. No preemption required. With
Wound-Wait, it's the process _holding_ the contended lock that gets
wounded (preempted), and it needs to back off at its own discretion but
no later than when it's going to sleep on another ww mutex. That point
is where we intercept the preemption request. We're preempting the
transaction rather than the process.


>
>
> Also, __ww_ctx_stamp_after() is crap; did we want to write:
>
> return (signed long)(a->stamp - b->stamp) > 0;
>
> or something?
>
>
Hmm. Yes it indeed looks odd. Seems like the above code should do the trick.

/Thomas


\
 
 \ /
  Last update: 2018-06-13 16:07    [W:0.052 / U:6.816 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site