Re: [PATCH v2 2/3] mm/mglru: add tracepoint for scan_folios()
From: Ridong Chen
Date: Wed Sep 16 2026 - 23:12:40 EST
On 9/17/2026 9:07 AM, Baoquan He wrote:
On 09/16/26 at 08:17pm, Ridong Chen wrote:
^~~~ typo, s/not/no/
On 9/16/2026 10:14 AM, Baoquan He wrote:
On 09/15/26 at 08:17pm, Ridong Chen wrote:
On 9/15/2026 7:14 PM, Baoquan He wrote:
On 09/11/26 at 06:29pm, Ridong Chen wrote:
From: Ridong Chen <chenridong@xxxxxxxxxx>...snip...
---
include/trace/events/vmscan.h | 63 +++++++++++++++++++++++++++++++++++
mm/vmscan.c | 6 ++++
2 files changed, 69 insertions(+)
diff --git a/include/trace/events/vmscan.h b/include/trace/events/vmscan.h
index 8a872990b4be..5defa8f6719c 100644
--- a/include/trace/events/vmscan.h
+++ b/include/trace/events/vmscan.h
@@ -392,6 +392,69 @@ TRACE_EVENT(mm_vmscan_lru_isolate,
__print_symbolic(__entry->lru, LRU_NAMES))
);
+TRACE_EVENT(mm_mglru_scan_folios,
+
+ TP_PROTO(struct lruvec *lruvec,
+ int highest_zoneidx,
+ int order,
+ unsigned long nr_requested,
+ unsigned long nr_scanned,
+ unsigned long nr_sorted,
+ unsigned long nr_skipped,
+ unsigned long nr_taken,
+ int lru,
+ unsigned long max_seq,
+ int tier,
+ unsigned long min_seq),
Thank you for your review.
Do we need show nid? Guess phone vendor doesn't care about multi-NUMA
system.
Eh, nid is not an mglru-specific key. Since there are 12 elements here, it
can't be added here.
Hmm, we print memcg_id but not nid, it may not be so clear. How do you
differentiate several lruvec-s on one memcg? I am not very familiar with
trace mechanism, do we have a element number limit?
Yes, it will fail to compile if there are more than 12 elements.
With AI's help, see
include/trace/bpf_probe.h
```
30 #define __CAST1(a,...) __CAST_TO_U64(a)
31 #define __CAST2(a,...) __CAST_TO_U64(a), __CAST1(__VA_ARGS__)
32 #define __CAST3(a,...) __CAST_TO_U64(a), __CAST2(__VA_ARGS__)
33 #define __CAST4(a,...) __CAST_TO_U64(a), __CAST3(__VA_ARGS__)
34 #define __CAST5(a,...) __CAST_TO_U64(a), __CAST4(__VA_ARGS__)
35 #define __CAST6(a,...) __CAST_TO_U64(a), __CAST5(__VA_ARGS__)
36 #define __CAST7(a,...) __CAST_TO_U64(a), __CAST6(__VA_ARGS__)
37 #define __CAST8(a,...) __CAST_TO_U64(a), __CAST7(__VA_ARGS__)
38 #define __CAST9(a,...) __CAST_TO_U64(a), __CAST8(__VA_ARGS__)
39 #define __CAST10(a,...) __CAST_TO_U64(a), __CAST9(__VA_ARGS__)
40 #define __CAST11(a,...) __CAST_TO_U64(a), __CAST10(__VA_ARGS__)
41 #define __CAST12(a,...) __CAST_TO_U64(a), __CAST11(__VA_ARGS__)
/* tracepoints with more than 12 arguments will hit build error */
```
Imagine I only want to trace this function, now I can see memcg_id,
zone_idex, while don't know node/lruvec, that's a little pitty. Just my
personal opinion, I could miss anything.
Yeah, you are right. In my personal experience, I always enable
trace_mm_vmscan_lru_isolate and trace_mm_vmscan_lru_shrink_inactive, which
contain node_id. So we can see how the lruvec is scanned and shrunk, and we
can get node_id from trace_mm_vmscan_lru_shrink_inactive. I am not arguing
that this is exactly the 'correct' way to use these tracepoints, but I think
it is a common usage.
I am open to that. If we want to record node_id, we have to remove an
element. Maybe tier? But I would strongly recommend record memcg_id, we
frequently need to check the reclaim behavior of a memcg, and most of our
Sorry, that's my typo which could cause misunderstanding. I meant "we print
memcg_id but no nid, it may not be so clear". I agree with you that
memcg_id is undispensable.
scenarios are non-NUMA.
Understood, mobile phone mostly is non-NUMA, while server is different.
About adding nid, I cousulted AI too, it says it's doable.
It says The 12-argument limit only counts the TP_PROTO arguments (that's
the BPF __CAST12/bpf_trace_run12 path), not the __entry fields in
TP_STRUCT__entry. And nid doesn't have to be a 13th argument at all. We
already pass lruvec, so nid can be derived from it in TP_fast_assign with
lruvec_pgdat(lruvec)->node_id.
e.g. for mm_mglru_scan_folios:
TP_STRUCT__entry(
__field(u64, memcg_id)
__field(int, nid)
__field(int, highest_zoneidx)
__field(int, order)
__field(unsigned long, nr_requested)
__field(unsigned long, nr_scanned)
__field(unsigned long, nr_sorted)
__field(unsigned long, nr_skipped)
__field(unsigned long, nr_taken)
__field(int, lru)
__field(unsigned long, max_seq)
__field(int, tier)
__field(unsigned long, min_seq)
),
TP_fast_assign(
__entry->memcg_id = mem_cgroup_id(lruvec_memcg(lruvec));
__entry->nid = lruvec_pgdat(lruvec)->node_id;
__entry->highest_zoneidx = highest_zoneidx;
...
),
TP_printk("memcg_id=%llu nid=%d classzone=%d order=%d ...",
__entry->memcg_id,
__entry->nid,
__entry->highest_zoneidx,
...)
lruvec_pgdat() comes from include/linux/mmzone.h, which vmscan.h already
pulls in via linux/mm.h, so no extra include is needed. The same pattern
applies to mm_mglru_inc_max_seq in patch 3 if you want nid there too.
I didn't try it though, please feel free to give it a shot.
Good. It works, I just tested it. Will add nid.
And I checked v1 thread, seems Steven suggested to move the two int to
avoid extra padding.
Yeah, I will update it. I updated v2 to fix Sashiko's issue, before Steven
had any comment.
+
+ TP_ARGS(lruvec, highest_zoneidx, order, nr_requested, nr_scanned,
+ nr_sorted, nr_skipped, nr_taken, lru, max_seq, tier, min_seq),
+
+ TP_STRUCT__entry(
+ __field(u64, memcg_id)
+ __field(int, highest_zoneidx)
+ __field(int, order)
+ __field(unsigned long, nr_requested)
+ __field(unsigned long, nr_scanned)
+ __field(unsigned long, nr_sorted)
+ __field(unsigned long, nr_skipped)
+ __field(unsigned long, nr_taken)
+ __field(int, lru)
+ __field(unsigned long, max_seq)
+ __field(int, tier)
+ __field(unsigned long, min_seq)
+ ),
+
+ TP_fast_assign(
+ __entry->memcg_id = mem_cgroup_id(lruvec_memcg(lruvec));
+ __entry->highest_zoneidx = highest_zoneidx;
+ __entry->order = order;
+ __entry->nr_requested = nr_requested;
+ __entry->nr_scanned = nr_scanned;
+ __entry->nr_sorted = nr_sorted;
+ __entry->nr_skipped = nr_skipped;
+ __entry->nr_taken = nr_taken;
+ __entry->lru = lru;
+ __entry->max_seq = max_seq;
+ __entry->tier = tier;
+ __entry->min_seq = min_seq;
+ ),
+
+ TP_printk("memcg_id=%llu classzone=%d order=%d nr_requested=%lu nr_scanned=%lu nr_sorted=%lu nr_skipped=%lu nr_taken=%lu lru=%s max_seq=%lu tier=%d min_seq=%lu",
+ __entry->memcg_id,
+ __entry->highest_zoneidx,
+ __entry->order,
+ __entry->nr_requested,
+ __entry->nr_scanned,
+ __entry->nr_sorted,
+ __entry->nr_skipped,
+ __entry->nr_taken,
+ __print_symbolic(__entry->lru, LRU_NAMES),
+ __entry->max_seq,
+ __entry->tier,
+ __entry->min_seq)
+);
+
TRACE_EVENT(mm_vmscan_write_folio,
TP_PROTO(struct folio *folio),
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 2554a6513aa8..67f59aa73fb9 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -4876,6 +4876,12 @@ static int scan_folios(unsigned long nr_to_scan, struct lruvec *lruvec,
trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan,
scanned, skipped, isolated,
type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON);
+ trace_mm_mglru_scan_folios(lruvec,
+ sc->reclaim_idx, sc->order, nr_to_scan,
+ scanned, sorted, skipped, isolated,
+ type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON,
+ lrugen->max_seq, tier,
+ lrugen->min_seq[type]);
And seems Steven also suggested to calculate lru, max_seq, min_seq in
TP_fast_assign() in v1 which could be missed.
Thanks, I will update it following Steven's suggestions.
TP_fast_assign(
__entry->memcg_id = mem_cgroup_id(lruvec_memcg(lruvec));
__entry->highest_zoneidx = sc->reclaim_idx;
__entry->order = sc->order;
__entry->nr_requested = nr_requested;
__entry->nr_scanned = nr_scanned;
__entry->nr_sorted = nr_sorted;
__entry->nr_skipped = nr_skipped;
__entry->nr_taken = nr_taken;
__entry->lru = type ? LRU_INACTIVE_FILE : LRU_INACTIVE_ANON;
__entry->max_seq = lrugen->max_seq;
__entry->tier = tier;
__entry->min_seq = lrugen->min_seq;
),
However, 'sc' is an internal structure, so sc->order cannot be accessed in
TP_fast_assign(), and I don't think it is worth making scan_control an
external structure just for this.
Agree.
*isolatedp = isolated;
return scanned;
--
2.34.1
--
Best regards
Ridong
--
Best regards
Ridong
--
Best regards
Ridong