Re: [PATCH] mm/vmalloc: reduce the number of lazy_max_pages to reduce latency

[Jisheng Zhang <jszhang@marvell.com>]

On Thu, 29 Sep 2016 12:07:14 +0100 Chris Wilson wrote:

> On Thu, Sep 29, 2016 at 04:28:08PM +0800, Jisheng Zhang wrote:
> > On Thu, 29 Sep 2016 09:18:18 +0100 Chris Wilson wrote:
> >   
> > > On Thu, Sep 29, 2016 at 03:34:11PM +0800, Jisheng Zhang wrote:  
> > > > On Marvell berlin arm64 platforms, I see the preemptoff tracer report
> > > > a max 26543 us latency at __purge_vmap_area_lazy, this latency is an
> > > > awfully bad for STB. And the ftrace log also shows __free_vmap_area
> > > > contributes most latency now. I noticed that Joel mentioned the same
> > > > issue[1] on x86 platform and gave two solutions, but it seems no patch
> > > > is sent out for this purpose.
> > > > 
> > > > This patch adopts Joel's first solution, but I use 16MB per core
> > > > rather than 8MB per core for the number of lazy_max_pages. After this
> > > > patch, the preemptoff tracer reports a max 6455us latency, reduced to
> > > > 1/4 of original result.    
> > > 
> > > My understanding is that
> > > 
> > > diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> > > index 91f44e78c516..3f7c6d6969ac 100644
> > > --- a/mm/vmalloc.c
> > > +++ b/mm/vmalloc.c
> > > @@ -626,7 +626,6 @@ void set_iounmap_nonlazy(void)
> > >  static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
> > >                                         int sync, int force_flush)
> > >  {
> > > -       static DEFINE_SPINLOCK(purge_lock);
> > >         struct llist_node *valist;
> > >         struct vmap_area *va;
> > >         struct vmap_area *n_va;
> > > @@ -637,12 +636,6 @@ static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
> > >          * should not expect such behaviour. This just simplifies locking for
> > >          * the case that isn't actually used at the moment anyway.
> > >          */
> > > -       if (!sync && !force_flush) {
> > > -               if (!spin_trylock(&purge_lock))
> > > -                       return;
> > > -       } else
> > > -               spin_lock(&purge_lock);
> > > -
> > >         if (sync)
> > >                 purge_fragmented_blocks_allcpus();
> > >  
> > > @@ -667,7 +660,6 @@ static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
> > >                         __free_vmap_area(va);
> > >                 spin_unlock(&vmap_area_lock);  
> > 
> > Hi Chris,
> > 
> > Per my test, the bottleneck now is __free_vmap_area() over the valist, the
> > iteration is protected with spinlock vmap_area_lock. So the larger lazy max
> > pages, the longer valist, the bigger the latency.
> > 
> > So besides above patch, we still need to remove vmap_are_lock or replace with
> > mutex.  
> 
> Or follow up with
> 
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 3f7c6d6969ac..67b5475f0b0a 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -656,8 +656,10 @@ static void __purge_vmap_area_lazy(unsigned long *start, unsigned long *end,
>  
>         if (nr) {
>                 spin_lock(&vmap_area_lock);
> -               llist_for_each_entry_safe(va, n_va, valist, purge_list)
> +               llist_for_each_entry_safe(va, n_va, valist, purge_list) {
>                         __free_vmap_area(va);
> +                       cond_resched_lock(&vmap_area_lock);

oh, great! This seems works fine. I'm not sure there's any side effect or
performance regression, but this patch plus previous purge_lock removing do
addressed my problem.

Thanks,
Jisheng

> +               }
>                 spin_unlock(&vmap_area_lock);
>         }
>  }
> 
> ?
> -Chris
>