Re: [PATCH] writeback: sum memcg dirty counters as needed

[Roman Gushchin <guro@fb.com>]

On Thu, Mar 07, 2019 at 08:56:32AM -0800, Greg Thelen wrote:
> Since commit a983b5ebee57 ("mm: memcontrol: fix excessive complexity in
> memory.stat reporting") memcg dirty and writeback counters are managed
> as:
> 1) per-memcg per-cpu values in range of [-32..32]
> 2) per-memcg atomic counter
> When a per-cpu counter cannot fit in [-32..32] it's flushed to the
> atomic.  Stat readers only check the atomic.
> Thus readers such as balance_dirty_pages() may see a nontrivial error
> margin: 32 pages per cpu.
> Assuming 100 cpus:
>    4k x86 page_size:  13 MiB error per memcg
>   64k ppc page_size: 200 MiB error per memcg
> Considering that dirty+writeback are used together for some decisions
> the errors double.
> 
> This inaccuracy can lead to undeserved oom kills.  One nasty case is
> when all per-cpu counters hold positive values offsetting an atomic
> negative value (i.e. per_cpu[*]=32, atomic=n_cpu*-32).
> balance_dirty_pages() only consults the atomic and does not consider
> throttling the next n_cpu*32 dirty pages.  If the file_lru is in the
> 13..200 MiB range then there's absolutely no dirty throttling, which
> burdens vmscan with only dirty+writeback pages thus resorting to oom
> kill.
> 
> It could be argued that tiny containers are not supported, but it's more
> subtle.  It's the amount the space available for file lru that matters.
> If a container has memory.max-200MiB of non reclaimable memory, then it
> will also suffer such oom kills on a 100 cpu machine.
> 
> The following test reliably ooms without this patch.  This patch avoids
> oom kills.
>
> ...
> 
> Make balance_dirty_pages() and wb_over_bg_thresh() work harder to
> collect exact per memcg counters when a memcg is close to the
> throttling/writeback threshold.  This avoids the aforementioned oom
> kills.
> 
> This does not affect the overhead of memory.stat, which still reads the
> single atomic counter.
> 
> Why not use percpu_counter?  memcg already handles cpus going offline,
> so no need for that overhead from percpu_counter.  And the
> percpu_counter spinlocks are more heavyweight than is required.
> 
> It probably also makes sense to include exact dirty and writeback
> counters in memcg oom reports.  But that is saved for later.
> 
> Signed-off-by: Greg Thelen <gthelen@google.com>
> ---
>  include/linux/memcontrol.h | 33 +++++++++++++++++++++++++--------
>  mm/memcontrol.c            | 26 ++++++++++++++++++++------
>  mm/page-writeback.c        | 27 +++++++++++++++++++++------
>  3 files changed, 66 insertions(+), 20 deletions(-)
> 
> diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> index 83ae11cbd12c..6a133c90138c 100644
> --- a/include/linux/memcontrol.h
> +++ b/include/linux/memcontrol.h
> @@ -573,6 +573,22 @@ static inline unsigned long memcg_page_state(struct mem_cgroup *memcg,
>  	return x;
>  }

Hi Greg!

Thank you for the patch, definitely a good problem to be fixed!

>  
> +/* idx can be of type enum memcg_stat_item or node_stat_item */
> +static inline unsigned long
> +memcg_exact_page_state(struct mem_cgroup *memcg, int idx)
> +{
> +	long x = atomic_long_read(&memcg->stat[idx]);
> +#ifdef CONFIG_SMP

I doubt that this #ifdef is correct without corresponding changes
in __mod_memcg_state(). As now, we do use per-cpu buffer which spills
to an atomic value event if !CONFIG_SMP. It's probably something
that we want to change, but as now, #ifdef CONFIG_SMP should protect
only "if (x < 0)" part.


> +	int cpu;
> +
> +	for_each_online_cpu(cpu)
> +		x += per_cpu_ptr(memcg->stat_cpu, cpu)->count[idx];
> +	if (x < 0)
> +		x = 0;
> +#endif
> +	return x;
> +}

Also, isn't it worth it to generalize memcg_page_state() instead?
By adding an bool exact argument? I believe dirty balance is not
the only place, where we need a better accuracy.

Thanks!