Re: [RFC PATCH] mm/page_alloc.c: micro-optimization reduce oom critical section size

[Mateusz Nosek]

Hello,

Thank you for your comments.
I will modify the patch where necessary and resend v2, but first I will 
make 100% sure about the lack of synchronization problem, that might 
potentially be there as you mentioned in previous mail , and try to 
check some numbers for my support.

Sincerely yours,
Mateusz Nosek

On 9/15/2020 4:04 PM, Michal Hocko wrote:
> On Tue 15-09-20 15:09:59, Mateusz Nosek wrote:
> > On 9/14/2020 4:22 PM, Michal Hocko wrote:
> > > On Mon 14-09-20 12:06:54, mateusznosek0@xxxxxxxxx wrote:
> > > > From: Mateusz Nosek <mateusznosek0@xxxxxxxxx>
> > > >
> > > > Most operations from '__alloc_pages_may_oom' do not require oom_mutex hold.
> > > > Exception is 'out_of_memory'. The patch refactors '__alloc_pages_may_oom'
> > > > to reduce critical section size and improve overall system performance.
> > >
> > > This is a real slow path. What is the point of optimizing it? Do you
> > > have any numbers?
> >
> > I agree that as this is the slow path, then the hard, complicated
> > optimizations are not recommended. In my humble opinion introduced patch is
> > not complex and does not decrease code readability or maintainability. In a
> > nutshell I see no drawbacks of applying it.
>
> This is clearly a matter of taste. I do not see a good reason to apply
> it TBH. It is a claimed optimization without any numbers to back that
> claim. It is also a tricky area so I am usually very careful to touch
> this code without a strong reason.  Others might feel differently of
> course.
>
> [...]
>
> Anyway, I have only now looked closer at the patch...
>
> > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > > index b9bd75cacf02..b07f950a5825 100644
> > > > --- a/mm/page_alloc.c
> > > > +++ b/mm/page_alloc.c
> > > > @@ -3935,18 +3935,7 @@ __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
> > > >    		.order = order,
> > > >    	};
> > > >    	struct page *page;
> > > > -
> > > > -	*did_some_progress = 0;
> > > > -
> > > > -	/*
> > > > -	 * Acquire the oom lock.  If that fails, somebody else is
> > > > -	 * making progress for us.
> > > > -	 */
> > > > -	if (!mutex_trylock(&oom_lock)) {
> > > > -		*did_some_progress = 1;
> > > > -		schedule_timeout_uninterruptible(1);
> > > > -		return NULL;
> > > > -	}
> > > > +	bool success;
> > > >    	/*
> > > >    	 * Go through the zonelist yet one more time, keep very high watermark
> > > > @@ -3959,14 +3948,17 @@ __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
> > > >    				      ~__GFP_DIRECT_RECLAIM, order,
> > > >    				      ALLOC_WMARK_HIGH|ALLOC_CPUSET, ac);
> > > >    	if (page)
> > > > -		goto out;
> > > > +		return page;
> > > > +
> > > > +	/* Check if somebody else is making progress for us. */
> > > > +	*did_some_progress = mutex_is_locked(&oom_lock);
>
> This is not only quite ugly but wrong as well. In general checking for a
> lock state is racy unless the lock is taken somewhere up the call chain.
>
> In this particular case it wouldn't be a big deal because an additional
> retry (did_some_progress = 1) is not really critical. It would likely be
> nicer to be deterministic here and not retry on all the early bailouts
> regardless of the lock state.
>
> > > >    	/* Coredumps can quickly deplete all memory reserves */
> > > >    	if (current->flags & PF_DUMPCORE)
> > > > -		goto out;
> > > > +		return NULL;
> > > >    	/* The OOM killer will not help higher order allocs */
> > > >    	if (order > PAGE_ALLOC_COSTLY_ORDER)
> > > > -		goto out;
> > > > +		return NULL;
> > > >    	/*
> > > >    	 * We have already exhausted all our reclaim opportunities without any
> > > >    	 * success so it is time to admit defeat. We will skip the OOM killer
> > > > @@ -3976,12 +3968,12 @@ __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
> > > >    	 * The OOM killer may not free memory on a specific node.
> > > >    	 */
> > > >    	if (gfp_mask & (__GFP_RETRY_MAYFAIL | __GFP_THISNODE))
> > > > -		goto out;
> > > > +		return NULL;
> > > >    	/* The OOM killer does not needlessly kill tasks for lowmem */
> > > >    	if (ac->highest_zoneidx < ZONE_NORMAL)
> > > > -		goto out;
> > > > +		return NULL;
> > > >    	if (pm_suspended_storage())
> > > > -		goto out;
> > > > +		return NULL;
> > > >    	/*
> > > >    	 * XXX: GFP_NOFS allocations should rather fail than rely on
> > > >    	 * other request to make a forward progress.
> > > > @@ -3992,8 +3984,20 @@ __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
> > > >    	 * failures more gracefully we should just bail out here.
> > > >    	 */
> > > > +	/*
> > > > +	 * Acquire the oom lock.  If that fails, somebody else is
> > > > +	 * making progress for us.
> > > > +	 */
> > > > +	if (!mutex_trylock(&oom_lock)) {
> > > > +		*did_some_progress = 1;
> > > > +		schedule_timeout_uninterruptible(1);
> > > > +		return NULL;
> > > > +	}
> > > > +	success = out_of_memory(&oc);
> > > > +	mutex_unlock(&oom_lock);
> > > > +
> > > >    	/* Exhausted what can be done so it's blame time */
> > > > -	if (out_of_memory(&oc) || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
> > > > +	if (success || WARN_ON_ONCE(gfp_mask & __GFP_NOFAIL)) {
> > > >    		*did_some_progress = 1;
> > > >    		/*
> > > > @@ -4004,8 +4008,7 @@ __alloc_pages_may_oom(gfp_t gfp_mask, unsigned int order,
> > > >    			page = __alloc_pages_cpuset_fallback(gfp_mask, order,
> > > >    					ALLOC_NO_WATERMARKS, ac);
> > > >    	}
> > > > -out:
> > > > -	mutex_unlock(&oom_lock);
> > > > +
> > > >    	return page;
> > > >    }
> > > > --
> > > > 2.20.1
> > Sincerely yours,
> > Mateusz Nosek