lkml.org 
[lkml]   [2020]   [Nov]   [19]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH 2/2] mm/vmalloc: rework the drain logic
On Thu, Nov 19, 2020 at 09:40:29AM +0800, Huang, Ying wrote:
> Uladzislau Rezki <urezki@gmail.com> writes:
>
> > On Wed, Nov 18, 2020 at 10:44:13AM +0800, huang ying wrote:
> >> On Tue, Nov 17, 2020 at 9:04 PM Uladzislau Rezki <urezki@gmail.com> wrote:
> >> >
> >> > On Tue, Nov 17, 2020 at 10:37:34AM +0800, huang ying wrote:
> >> > > On Tue, Nov 17, 2020 at 6:00 AM Uladzislau Rezki (Sony)
> >> > > <urezki@gmail.com> wrote:
> >> > > >
> >> > > > A current "lazy drain" model suffers from at least two issues.
> >> > > >
> >> > > > First one is related to the unsorted list of vmap areas, thus
> >> > > > in order to identify the [min:max] range of areas to be drained,
> >> > > > it requires a full list scan. What is a time consuming if the
> >> > > > list is too long.
> >> > > >
> >> > > > Second one and as a next step is about merging all fragments
> >> > > > with a free space. What is also a time consuming because it
> >> > > > has to iterate over entire list which holds outstanding lazy
> >> > > > areas.
> >> > > >
> >> > > > See below the "preemptirqsoff" tracer that illustrates a high
> >> > > > latency. It is ~24 676us. Our workloads like audio and video
> >> > > > are effected by such long latency:
> >> > >
> >> > > This seems like a real problem. But I found there's long latency
> >> > > avoidance mechanism in the loop in __purge_vmap_area_lazy() as
> >> > > follows,
> >> > >
> >> > > if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
> >> > > cond_resched_lock(&free_vmap_area_lock);
> >> > >
> >> > I have added that "resched threshold" because of on my tests i could
> >> > simply hit out of memory, due to the fact that a drain work is not up
> >> > to speed to process such long outstanding list of vmap areas.
> >>
> >> OK. Now I think I understand the problem. For free area purging,
> >> there are multiple "producers" but one "consumer", and it lacks enough
> >> mechanism to slow down the "producers" if "consumer" can not catch up.
> >> And your patch tries to resolve the problem via accelerating the
> >> "consumer".
> >>
> > Seems, correct. But just in case one more time:
> >
> > the cond_resched_lock was added once upon a time to get rid of long
> > preemption off time. Due to dropping the lock, "producers" can start
> > generate further vmap area, so "consumer" can not catch up. Seems
>
> Yes. And in theory there are vfree() storm, that is, thousands vfree()
> can be called in short time. But I don't think that's practical use
> case.
>
> > Later on, a resched threshold was added. It is just a simple protection
> > threshold, passing which, a freeing is prioritized back over allocation,
> > so we guarantee that we do not hit out of memory.
>
> Yes. That can accelerate freeing if necessary.
>
> >>
> >> That isn't perfect, but I think we may have quite some opportunities
> >> to merge the free areas, so it should just work.
> >>
> > Yes, merging opportunity should do the work. But of course there are
> > exceptions.
> >
> >> And I found the long latency avoidance logic in
> >> __purge_vmap_area_lazy() appears problematic,
> >>
> >> if (atomic_long_read(&vmap_lazy_nr) < resched_threshold)
> >> cond_resched_lock(&free_vmap_area_lock);
> >>
> >> Shouldn't it be something as follows?
> >>
> >> if (i >= BATCH && atomic_long_read(&vmap_lazy_nr) <
> >> resched_threshold) {
> >> cond_resched_lock(&free_vmap_area_lock);
> >> i = 0;
> >> } else
> >> i++;
> >>
> >> This will accelerate the purging via batching and slow down vmalloc()
> >> via holding free_vmap_area_lock. If it makes sense, can we try this?
> >>
> > Probably we can switch to just using "batch" methodology:
> >
> > <snip>
> > if (!(i++ % batch_threshold))
> > cond_resched_lock(&free_vmap_area_lock);
> > <snip>
>
> That's the typical long latency avoidance method.
>
> > The question is, which value we should use as a batch_threshold: 100, 1000, etc.
>
> I think we can do some measurement to determine it?
>
Hmm.. looking at it one more time i do not see what batching solves.
Anyway we need to have some threshold(what we do have), that regulates
a priority between vmalloc()/vfree().

What we can do more with it are:

- purging should be just performed asynchronously in workqueue context.
Giving the fact, that now we also do a merge of outstanding areas, the
data structure(rb-tree) will not be so fragmented.

- lazy_max_pages() can slightly be decreased. If there are existing
workloads which suffer from such long value. It would be good to get
real complains and evidence.

> > Apart of it and in regard to CONFIG_KASAN_VMALLOC, it seems that we are not
> > allowed to drop the free_vmap_area_lock at all. Because any simultaneous
> > allocations are not allowed within a drain region, so it should occur in
> > disjoint regions. But i need to double check it.
> >
> >>
> >> And, can we reduce lazy_max_pages() to control the length of the
> >> purging list? It could be > 8K if the vmalloc/vfree size is small.
> >>
> > We can adjust it for sure. But it will influence on number of global
> > TLB flushes that must be performed.
>
> Em... For example, if we set it to 100, then the number of the TLB
> flushes can be reduced to 1% of the un-optimized implementation
> already. Do you think so?
>
If we set lazy_max_pages() to vague value such as 100, the performance
will be just destroyed.

Thanks!

--
Vlad Rezki

\
 
 \ /
  Last update: 2020-11-19 18:39    [W:1.089 / U:0.764 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site