On 2/28/19 9:14 AM, Yafang Shao wrote:
> In the page alloc fast path, it may do node reclaim, which may cause
> latency spike.
> We should add tracepoint for this event, and also mesure the latency
> it causes.
>
> So bellow two tracepoints are introduced,
> mm_vmscan_node_reclaim_begin
> mm_vmscan_node_reclaim_end
>
> Signed-off-by: Yafang Shao <laoar.shao@xxxxxxxxx>
> ---
> include/trace/events/vmscan.h | 48 +++++++++++++++++++++++++++++++++++++++++++
> mm/vmscan.c | 13 +++++++++++-
> 2 files changed, 60 insertions(+), 1 deletion(-)
>
> diff --git a/include/trace/events/vmscan.h b/include/trace/events/vmscan.h
> index a1cb913..9310d5b 100644
> --- a/include/trace/events/vmscan.h
> +++ b/include/trace/events/vmscan.h
> @@ -465,6 +465,54 @@
> __entry->ratio,
> show_reclaim_flags(__entry->reclaim_flags))
> );
> +
> +TRACE_EVENT(mm_vmscan_node_reclaim_begin,
> +
> + TP_PROTO(int nid, int order, int may_writepage,
> + gfp_t gfp_flags, int zid),
> +
> + TP_ARGS(nid, order, may_writepage, gfp_flags, zid),
> +
> + TP_STRUCT__entry(
> + __field(int, nid)
> + __field(int, order)
> + __field(int, may_writepage)
For node reclaim may_writepage is statically set in node_reclaim_mode,
so I'm not sure it's worth including it.
> + __field(gfp_t, gfp_flags)
> + __field(int, zid)
zid seems wasteful and misleading as it's simply derived by
gfp_zone(gfp_mask), so I would drop it.
> + ),
> +
> + TP_fast_assign(
> + __entry->nid = nid;
> + __entry->order = order;
> + __entry->may_writepage = may_writepage;
> + __entry->gfp_flags = gfp_flags;
> + __entry->zid = zid;
> + ),
> +
> + TP_printk("nid=%d zid=%d order=%d may_writepage=%d gfp_flags=%s",
> + __entry->nid,
> + __entry->zid,
> + __entry->order,
> + __entry->may_writepage,
> + show_gfp_flags(__entry->gfp_flags))
> +);
> +
> +TRACE_EVENT(mm_vmscan_node_reclaim_end,
> +
> + TP_PROTO(int result),
> +
> + TP_ARGS(result),
> +
> + TP_STRUCT__entry(
> + __field(int, result)
Reporting sc.nr_reclaimed sounds more useful and in line with other
reclaim tracepoints. Result (sc.nr_reclaimed >= nr_pages) can then be
derived by postprocessing as the beginning tracepoint contains 'order'
thus we know nr_pages?
> + ),
> +
> + TP_fast_assign(
> + __entry->result = result;
> + ),
> +
> + TP_printk("result=%d", __entry->result)
> +);
> #endif /* _TRACE_VMSCAN_H */
>
> /* This part must be outside protection */
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index ac4806f..01a0401 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -4240,6 +4240,12 @@ static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned in
> .may_swap = 1,
> .reclaim_idx = gfp_zone(gfp_mask),
> };
> + int result;
> +
> + trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order,
> + sc.may_writepage,
> + sc.gfp_mask,
> + sc.reclaim_idx);
>
> cond_resched();
> fs_reclaim_acquire(sc.gfp_mask);
> @@ -4267,7 +4273,12 @@ static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned in
> current->flags &= ~PF_SWAPWRITE;
> memalloc_noreclaim_restore(noreclaim_flag);
> fs_reclaim_release(sc.gfp_mask);
> - return sc.nr_reclaimed >= nr_pages;
> +
> + result = sc.nr_reclaimed >= nr_pages;
> +
> + trace_mm_vmscan_node_reclaim_end(result);
> +
> + return result;
> }
>
> int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
>
