Re: [PATCH] memcg: add tracing for memcg stat updates

[Yosry Ahmed <yosryahmed@google.com>]

On Wed, Oct 9, 2024 at 5:36 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
>
> The memcg stats are maintained in rstat infrastructure which provides
> very fast updates side and reasonable read side. However memcg added
> plethora of stats and made the read side, which is cgroup rstat flush,
> very slow. To solve that, threshold was added in the memcg stats read
> side i.e. no need to flush the stats if updates are within the
> threshold.
>
> This threshold based improvement worked for sometime but more stats were
> added to memcg and also the read codepath was getting triggered in the
> performance sensitive paths which made threshold based ratelimiting
> ineffective. We need more visibility into the hot and cold stats i.e.
> stats with a lot of updates. Let's add trace to get that visibility.
>
> Signed-off-by: Shakeel Butt <shakeel.butt@linux.dev>

One question below, otherwise:

Reviewed-by: Yosry Ahmed <yosryahmed@google.com>

> ---
>  include/trace/events/memcg.h | 59 ++++++++++++++++++++++++++++++++++++
>  mm/memcontrol.c              | 13 ++++++--
>  2 files changed, 70 insertions(+), 2 deletions(-)
>  create mode 100644 include/trace/events/memcg.h
>
> diff --git a/include/trace/events/memcg.h b/include/trace/events/memcg.h
> new file mode 100644
> index 000000000000..913db9aba580
> --- /dev/null
> +++ b/include/trace/events/memcg.h
> @@ -0,0 +1,59 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#undef TRACE_SYSTEM
> +#define TRACE_SYSTEM memcg
> +
> +#if !defined(_TRACE_MEMCG_H) || defined(TRACE_HEADER_MULTI_READ)
> +#define _TRACE_MEMCG_H
> +
> +#include <linux/memcontrol.h>
> +#include <linux/tracepoint.h>
> +
> +
> +DECLARE_EVENT_CLASS(memcg_rstat,
> +
> +       TP_PROTO(struct mem_cgroup *memcg, int item, int val),
> +
> +       TP_ARGS(memcg, item, val),
> +
> +       TP_STRUCT__entry(
> +               __field(u64, id)
> +               __field(int, item)
> +               __field(int, val)
> +       ),
> +
> +       TP_fast_assign(
> +               __entry->id = cgroup_id(memcg->css.cgroup);
> +               __entry->item = item;
> +               __entry->val = val;
> +       ),
> +
> +       TP_printk("memcg_id=%llu item=%d val=%d",
> +                 __entry->id, __entry->item, __entry->val)
> +);
> +
> +DEFINE_EVENT(memcg_rstat, mod_memcg_state,
> +
> +       TP_PROTO(struct mem_cgroup *memcg, int item, int val),
> +
> +       TP_ARGS(memcg, item, val)
> +);
> +
> +DEFINE_EVENT(memcg_rstat, mod_memcg_lruvec_state,
> +
> +       TP_PROTO(struct mem_cgroup *memcg, int item, int val),
> +
> +       TP_ARGS(memcg, item, val)
> +);
> +
> +DEFINE_EVENT(memcg_rstat, count_memcg_events,
> +
> +       TP_PROTO(struct mem_cgroup *memcg, int item, int val),
> +
> +       TP_ARGS(memcg, item, val)
> +);
> +
> +
> +#endif /* _TRACE_MEMCG_H */
> +
> +/* This part must be outside protection */
> +#include <trace/define_trace.h>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index c098fd7f5c5e..17af08367c68 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -71,6 +71,10 @@
>
>  #include <linux/uaccess.h>
>
> +#define CREATE_TRACE_POINTS
> +#include <trace/events/memcg.h>
> +#undef CREATE_TRACE_POINTS
> +
>  #include <trace/events/vmscan.h>
>
>  struct cgroup_subsys memory_cgrp_subsys __read_mostly;
> @@ -682,7 +686,9 @@ void __mod_memcg_state(struct mem_cgroup *memcg, enum memcg_stat_item idx,
>                 return;
>
>         __this_cpu_add(memcg->vmstats_percpu->state[i], val);
> -       memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
> +       val = memcg_state_val_in_pages(idx, val);
> +       memcg_rstat_updated(memcg, val);
> +       trace_mod_memcg_state(memcg, idx, val);
>  }
>
>  /* idx can be of type enum memcg_stat_item or node_stat_item. */
> @@ -741,7 +747,9 @@ static void __mod_memcg_lruvec_state(struct lruvec *lruvec,
>         /* Update lruvec */
>         __this_cpu_add(pn->lruvec_stats_percpu->state[i], val);
>
> -       memcg_rstat_updated(memcg, memcg_state_val_in_pages(idx, val));
> +       val = memcg_state_val_in_pages(idx, val);
> +       memcg_rstat_updated(memcg, val);
> +       trace_mod_memcg_lruvec_state(memcg, idx, val);
>         memcg_stats_unlock();
>  }
>
> @@ -832,6 +840,7 @@ void __count_memcg_events(struct mem_cgroup *memcg, enum vm_event_item idx,
>         memcg_stats_lock();
>         __this_cpu_add(memcg->vmstats_percpu->events[i], count);
>         memcg_rstat_updated(memcg, count);
> +       trace_count_memcg_events(memcg, idx, count);

count here is an unsigned long, and we are casting it to int, right?

Would it be slightly better if the tracepoint uses a long instead of
int? It's still not ideal but probably better than int.

>         memcg_stats_unlock();
>  }
>
> --
> 2.43.5
>