When running LTP's cpuset_memory_pressure case, the following error occurs:

(1) mount a cgroup v1 hierarchy, enable cpuset, set memory limit
(2) write 1 to memory_pressure_enabled
(2) create a process in this group, allocate a large amount of memory
    to generate pressure
(3) write 0 to memory_pressure_enabled
(4) immediately read cpuset.memory_pressure:
    the kernel returns a non-zero value, and LTP determines it as FAIL

According to admin-guide/cgroup-v1/cpusets.rst,

"1.5 What is memory_pressure ?
 ...
  So only systems that enable this feature
  (i.e. memory_pressure_enabled) will compute the metric."

To be consistent with the documentation, when memory_pressure_enabled is 0,
cpuset.memory_pressure should always returns 0.

Signed-off-by: Jin Guojie <guojie.jin@gmail.com>
---
 kernel/cgroup/cpuset-v1.c | 11 +++++++++--
 1 file changed, 9 insertions(+), 2 deletions(-)

diff --git a/kernel/cgroup/cpuset-v1.c b/kernel/cgroup/cpuset-v1.c
index 25c1d7b77e2f..940beff47772 100644
--- a/kernel/cgroup/cpuset-v1.c
+++ b/kernel/cgroup/cpuset-v1.c
@@ -105,8 +105,15 @@ static int fmeter_getrate(struct fmeter *fmp)
        int val;

        spin_lock(&fmp->lock);
-       fmeter_update(fmp);
-       val = fmp->val;
+       if (cpuset_memory_pressure_enabled) {
+               fmeter_update(fmp);
+               val = fmp->val;
+       } else {
+               fmp->cnt = 0;
+               fmp->val = 0;
+               fmp->time = 0;
+               val = 0;
+       }
        spin_unlock(&fmp->lock);
        return val;
 }
--
2.34.1

On Thu, Dec 19, 2024 at 05:06:31PM +0800, Jin Guojie <guojie.jin@gmail.com> wrote:
> According to admin-guide/cgroup-v1/cpusets.rst,
> 
> "1.5 What is memory_pressure ?
>  ...
>   So only systems that enable this feature
>   (i.e. memory_pressure_enabled) will compute the metric."
> 
> To be consistent with the documentation, when memory_pressure_enabled is 0,
> cpuset.memory_pressure should always returns 0.

IMO a clearer fix would be to call fmeter_init() (again) when writing 1
to the memory_pressure_enabled.
At the same time, the docs interpretation is lose in what should be
shown (e.g. show last value, don't calculate updates), so it may need no
fix.

HTH,
Michal