Re: [RFC PATCH] mm: move pcp and lru-pcp drainging into vmstat_wq

[Michal Hocko]

On Wed 08-02-17 12:31:13, Mel Gorman wrote:
> On Wed, Feb 08, 2017 at 01:03:55PM +0100, Michal Hocko wrote:
> > > I don't object to it being actually moved. I have a slight concern that
> > > it could somehow starve a vmstat update while frequent drains happen
> > > during reclaim though which potentially compounds the problem. It could
> > > be offset by a variety of other factors but if it ever is an issue,
> > > it'll show up and the paths that really matter check the vmstats
> > > directly instead of waiting for an update.
> > 
> > vmstat updates can tolared delays, that's we we are using deferable
> > scheduling in the first place so I am not really worried about that. Any
> > user which needs a better precision should use *_snapshot API.
> > 
> 
> Agreed, we already had cases where deferred vmstat updates had problems
> and were resolved by using _snapshot. It's a slight concern only and I'd
> be surprised if the _snapshot usage didn't cover it.
> 
> > > The altering of the return value in setup_vmstat was mildly surprising as
> > > it increases the severity of registering the vmstat callback for memory
> > > hotplug so maybe split that out and appears unrelated.
> > 
> > not sure I understand. What do you mean?
> > 
> 
> This hunk
> 
> @@ -1763,9 +1762,11 @@ static int vmstat_cpu_dead(unsigned int cpu)
> 
>  static int __init setup_vmstat(void)
>  {
> -#ifdef CONFIG_SMP
> -       int ret;
> +       int ret = 0;
> +
> +       vmstat_wq = alloc_workqueue("vmstat", WQ_FREEZABLE|WQ_MEM_RECLAIM, 0);
> 
> +#ifdef CONFIG_SMP
>         ret = cpuhp_setup_state_nocalls(CPUHP_MM_VMSTAT_DEAD, "mm/vmstat:dead",
>                                         NULL, vmstat_cpu_dead);
>         if (ret < 0)
> @@ -1789,7 +1790,7 @@ static int __init setup_vmstat(void)
>         proc_create("vmstat", S_IRUGO, NULL, &proc_vmstat_file_operations);
>         proc_create("zoneinfo", S_IRUGO, NULL, &proc_zoneinfo_file_operations);
>  #endif
> -       return 0;
> +       return ret;
> 
> 
> A failed register of vmstat_cpu_dead is returning the failure code in an
> init function now. Chances are it'll never hit but it didn't seem related
> to the patches general intent.

Ohh, I see now. I will keep the original behavior.

> > > It also feels like vmstat is now a misleading name for something that
> > > handles vmstat, lru drains and per-cpu drains but that's cosmetic.
> > 
> > yeah a better name sounds like a good thing. mm_nonblock_wq?
> > 
> 
> it's not always non-blocking. Maybe mm_percpu_wq to describev a workqueue
> that handles a variety of MM-related per-cpu updates?

Why not, I do not have a strong preference. The WQ is already documented
for its requirements on workers so the name doesn't really have to be
explicit about blocking on allocations.

Thanks!

-- 
Michal Hocko
SUSE Labs