Hi SJ,
在 2022/9/13 下午11:17, SeongJae Park 写道:
Hi Xin,
On Tue, 13 Sep 2022 19:47:35 +0800 Xin Hao <xhao@xxxxxxxxxxxxxxxxx> wrote:
MADV_COLLAPSE has large different with MDVA_HUGE in dealing with huge
pages, it doesn't need to rely on 'khugepage' thread to create THP.
So combining with it will give damon users more options.
I think this is a simple and worthy change, but... could you put some more
detailed explanation of the scenario that the action could be used? Also, it
would be good to have some evaluation results you got with this change.
Yes, internally, we use this feature mainly focus on two reason:
the fist one,
prevent performance jitter casused by khugepage thread under memory
pressure.
the second one,
khugepage thread merge hugepages is slowness then MADV_COLLAPSE.
Als, to my understanding, MADV_COLLAPSE is not merged in the mainline yet,
right? I believe it will make it in v6.1-rc1. That said, to make Andrew's
burden less, how about delaying this change until MADV_COLLAPSE is mainlined?
If you have some numbers showing clear and great benefit of this change, I
think it would be worthy to add immediately, though.
Thanks,
SJ
Signed-off-by: Xin Hao <xhao@xxxxxxxxxxxxxxxxx>
---
include/linux/damon.h | 1 +
mm/damon/vaddr.c | 3 +++
2 files changed, 4 insertions(+)
diff --git a/include/linux/damon.h b/include/linux/damon.h
index 016b6c9c03d6..c44f9410d997 100644
--- a/include/linux/damon.h
+++ b/include/linux/damon.h
@@ -97,6 +97,7 @@ enum damos_action {
DAMOS_PAGEOUT,
DAMOS_HUGEPAGE,
DAMOS_NOHUGEPAGE,
+ DAMOS_COLLAPSE,
DAMOS_LRU_PRIO,
DAMOS_LRU_DEPRIO,
DAMOS_STAT, /* Do nothing but only record the stat */
diff --git a/mm/damon/vaddr.c b/mm/damon/vaddr.c
index c2c08c1b316b..58c0d068563d 100644
--- a/mm/damon/vaddr.c
+++ b/mm/damon/vaddr.c
@@ -655,6 +655,9 @@ static unsigned long damon_va_apply_scheme(struct damon_ctx *ctx,
case DAMOS_NOHUGEPAGE:
madv_action = MADV_NOHUGEPAGE;
break;
+ case DAMOS_COLLAPSE:
+ madv_action = MADV_COLLAPSE;
+ break;
case DAMOS_STAT:
return 0;
default:
--
2.31.0
