[PATCH] mm/vmpressure: scale the vmpressure window with machine size

[Tao Cui <cui.tao@linux.dev>]

From: Tao Cui <cuitao@kylinos.cn>

vmpressure_win -- the number of pages the reclaimer must scan before
socket pressure is re-evaluated -- has been a fixed 512 pages (2 MB) since
vmpressure was introduced.  The reclaimer scans more pages on a larger
machine, so the window is reached far more often there and socket pressure
is re-armed every handful of scanned pages.  The comment at the definition
has asked for the window to scale with machine size "as we do for vmstat
thresholds" for over a decade.

Scale it the same way calculate_normal_threshold() does: logarithmically
with memory (fls of memory in 128 MB units), computed once in a
subsys_initcall once totalram_pages() is known.  A machine under 128 MB
keeps the historical 512 pages; the window then grows by SWAP_CLUSTER_MAX
* 16 per doubling of memory.

Why this matters: the scanned/reclaimed ratio that drives socket pressure
is averaged over the window, and the window rate-limits the evaluation.
With a fixed 2 MB window the evaluation runs the same number of times
regardless of machine size, which is disproportionately many on a large
machine.  Measured by cold-booting one VM at each size and running the
same cgroup-bound reclaim workload (so the page count is identical across
sizes):

    config   scaled_win   pages scanned   512-win evals   scaled evals
    4 GB     3072         11.9 M          23267           3877
    8 GB     3584         11.9 M          23306           3329
    16 GB    4096         11.9 M          23281           2910
    32 GB    4608         11.9 M          23268           2585
    64 GB    5120         11.9 M          23281           2328

For the same reclaim work the fixed window evaluates ~23k times at every
machine size; the scaled window evaluates fewer times the larger the
machine -- a 6x reduction at 4 GB growing to 10x at 64 GB (and the
logarithmic growth continues: ~11x projected at 128 GB).  Each evaluation
takes the per-memcg sr_lock and may write the socket_pressure seqlock, so
on larger machines with more memcgs under pressure this is real overhead
the fixed window pays needlessly.

The default stays 512 until the initcall runs, so early-boot reclaim is
unchanged.

Signed-off-by: Tao Cui <cuitao@kylinos.cn>
---
 include/linux/vmpressure.h |  2 +-
 mm/vmpressure.c            | 20 +++++++++++++++++---
 2 files changed, 18 insertions(+), 4 deletions(-)

diff --git a/include/linux/vmpressure.h b/include/linux/vmpressure.h
index b4d13457bc2a..09111f5bdc88 100644
--- a/include/linux/vmpressure.h
+++ b/include/linux/vmpressure.h
@@ -51,7 +51,7 @@ extern struct vmpressure *memcg_to_vmpressure(struct mem_cgroup *memcg);
 extern struct mem_cgroup *vmpressure_to_memcg(struct vmpressure *vmpr);
 
 /* Shared with the v1 vmpressure block in mm/memcontrol-v1.c. */
-extern const unsigned long vmpressure_win;
+extern unsigned long vmpressure_win;
 extern enum vmpressure_levels vmpressure_calc_level(unsigned long scanned,
 						    unsigned long reclaimed);
 
diff --git a/mm/vmpressure.c b/mm/vmpressure.c
index 9629240d77ad..4c8671273c79 100644
--- a/mm/vmpressure.c
+++ b/mm/vmpressure.c
@@ -31,10 +31,24 @@
  * As the vmscan reclaimer logic works with chunks which are multiple of
  * SWAP_CLUSTER_MAX, it makes sense to use it for the window size as well.
  *
- * TODO: Make the window size depend on machine size, as we do for vmstat
- * thresholds. Currently we set it to 512 pages (2MB for 4KB pages).
+ * Scale the window with machine size, the way vmstat thresholds do: on a
+ * larger machine the reclaimer scans more pages, so a fixed window would
+ * re-evaluate pressure every handful of pages. The scaling is logarithmic
+ * (fls, like calculate_normal_threshold()), keeping the growth moderate.
+ * The default is the historical 512 pages; an early initcall applies the
+ * scaling once totalram_pages() is known.
  */
-const unsigned long vmpressure_win = SWAP_CLUSTER_MAX * 16;
+unsigned long vmpressure_win __read_mostly = SWAP_CLUSTER_MAX * 16;
+
+static int __init vmpressure_init_window(void)
+{
+	unsigned long mem128m;	/* machine memory in 128MB units */
+
+	mem128m = totalram_pages() >> (27 - PAGE_SHIFT);
+	vmpressure_win = SWAP_CLUSTER_MAX * 16 * (1 + fls(mem128m));
+	return 0;
+}
+subsys_initcall(vmpressure_init_window);
 
 /*
  * These thresholds are used when we account memory pressure through
-- 
2.43.0