mm/mempolicy: failed to disable numa balancing

[PATCH] mm/mempolicy: failed to disable numa balancing

Posted by tzm 2 years, 9 months ago

It will be failed to  disable numa balancing policy permanently by passing
<numa_balancing=disable> to boot cmdline parameters.
The numabalancing_override variable is int and 1 for enable -1 for disable.
So, !enumabalancing_override will always be true, which cause this bug.

Signed-off-by: tzm <tcm1030@163.com>
---
 mm/mempolicy.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 61aa9aedb728..2789c0920293 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -2865,7 +2865,7 @@ static void __init check_numabalancing_enable(void)
 	if (numabalancing_override)
 		set_numabalancing_state(numabalancing_override == 1);
 
-	if (num_online_nodes() > 1 && !numabalancing_override) {
+	if (num_online_nodes() > 1 && (numabalancing_override == 1)) {
 		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
 			numabalancing_default ? "Enabling" : "Disabling");
 		set_numabalancing_state(numabalancing_default);
-- 
2.27.0

Re: [PATCH] mm/mempolicy: failed to disable numa balancing

Posted by Andrew Morton 2 years, 9 months ago

On Fri,  2 Dec 2022 22:16:30 +0800 tzm <tcm1030@163.com> wrote:

> It will be failed to  disable numa balancing policy permanently by passing
> <numa_balancing=disable> to boot cmdline parameters.
> The numabalancing_override variable is int and 1 for enable -1 for disable.
> So, !enumabalancing_override will always be true, which cause this bug.

That's really old code!

> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -2865,7 +2865,7 @@ static void __init check_numabalancing_enable(void)
>  	if (numabalancing_override)
>  		set_numabalancing_state(numabalancing_override == 1);
>  
> -	if (num_online_nodes() > 1 && !numabalancing_override) {
> +	if (num_online_nodes() > 1 && (numabalancing_override == 1)) {
>  		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
>  			numabalancing_default ? "Enabling" : "Disabling");
>  		set_numabalancing_state(numabalancing_default);

Looks right to me.  Mel?

After eight years, I wonder if we actually need this.

Re: [PATCH] mm/mempolicy: failed to disable numa balancing

Posted by Mel Gorman 2 years, 9 months ago

On Fri, Dec 02, 2022 at 11:59:54AM -0800, Andrew Morton wrote:
> On Fri,  2 Dec 2022 22:16:30 +0800 tzm <tcm1030@163.com> wrote:
> 
> > It will be failed to  disable numa balancing policy permanently by passing
> > <numa_balancing=disable> to boot cmdline parameters.
> > The numabalancing_override variable is int and 1 for enable -1 for disable.
> > So, !enumabalancing_override will always be true, which cause this bug.
> 
> That's really old code!
> 
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -2865,7 +2865,7 @@ static void __init check_numabalancing_enable(void)
> >  	if (numabalancing_override)
> >  		set_numabalancing_state(numabalancing_override == 1);
> >  
> > -	if (num_online_nodes() > 1 && !numabalancing_override) {
> > +	if (num_online_nodes() > 1 && (numabalancing_override == 1)) {
> >  		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
> >  			numabalancing_default ? "Enabling" : "Disabling");
> >  		set_numabalancing_state(numabalancing_default);
> 
> Looks right to me.  Mel?
> 
> After eight years, I wonder if we actually need this.

I don't think the patch is right aside from coding style issues such as
real names used in signed-off-by's.

The !numabalancing_override is checking "should the default be changed?",
itt's not checking if it should be enabled specifically. A better potential
fix would be something like this? (not actually tested)

diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 61aa9aedb728..fc649f8509f7 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -2862,10 +2862,12 @@ static void __init check_numabalancing_enable(void)
 		numabalancing_default = true;
 
 	/* Parsed by setup_numabalancing. override == 1 enables, -1 disables */
-	if (numabalancing_override)
+	if (numabalancing_override) {
 		set_numabalancing_state(numabalancing_override == 1);
+		return;
+	}
 
-	if (num_online_nodes() > 1 && !numabalancing_override) {
+	if (num_online_nodes() > 1) {
 		pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
 			numabalancing_default ? "Enabling" : "Disabling");
 		set_numabalancing_state(numabalancing_default);

Re: [PATCH] mm/mempolicy: failed to disable numa balancing

Posted by Yu Zhao 2 years, 9 months ago

On Fri, Dec 2, 2022 at 1:00 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Fri,  2 Dec 2022 22:16:30 +0800 tzm <tcm1030@163.com> wrote:
>
> > It will be failed to  disable numa balancing policy permanently by passing
> > <numa_balancing=disable> to boot cmdline parameters.
> > The numabalancing_override variable is int and 1 for enable -1 for disable.
> > So, !enumabalancing_override will always be true, which cause this bug.

!enumabalancing_override is false when enumabalancing_override = -1
(numa_balancing=disable).

> That's really old code!
>
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -2865,7 +2865,7 @@ static void __init check_numabalancing_enable(void)
> >       if (numabalancing_override)
> >               set_numabalancing_state(numabalancing_override == 1);
> >
> > -     if (num_online_nodes() > 1 && !numabalancing_override) {
> > +     if (num_online_nodes() > 1 && (numabalancing_override == 1)) {
> >               pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
> >                       numabalancing_default ? "Enabling" : "Disabling");
> >               set_numabalancing_state(numabalancing_default);
>
> Looks right to me.  Mel?
>
> After eight years, I wonder if we actually need this.

NAK.

The original code works as intended. This patch breaks my test with
CONFIG_NUMA_BALANCING_DEFAULT_ENABLED=n and numa_balancing=enable.

Re: [PATCH] mm/mempolicy: failed to disable numa balancing

Posted by Mina Almasry 2 years, 9 months ago

On Fri, Dec 2, 2022 at 12:00 PM Andrew Morton <akpm@linux-foundation.org> wrote:
>
> On Fri,  2 Dec 2022 22:16:30 +0800 tzm <tcm1030@163.com> wrote:
>
> > It will be failed to  disable numa balancing policy permanently by passing
> > <numa_balancing=disable> to boot cmdline parameters.
> > The numabalancing_override variable is int and 1 for enable -1 for disable.
> > So, !enumabalancing_override will always be true, which cause this bug.
>
> That's really old code!
>
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -2865,7 +2865,7 @@ static void __init check_numabalancing_enable(void)
> >       if (numabalancing_override)
> >               set_numabalancing_state(numabalancing_override == 1);
> >
> > -     if (num_online_nodes() > 1 && !numabalancing_override) {
> > +     if (num_online_nodes() > 1 && (numabalancing_override == 1)) {
> >               pr_info("%s automatic NUMA balancing. Configure with numa_balancing= or the kernel.numa_balancing sysctl\n",
> >                       numabalancing_default ? "Enabling" : "Disabling");
> >               set_numabalancing_state(numabalancing_default);
>
> Looks right to me.  Mel?
>

Maybe I'm missing something, but it looks wrong to me?

numabalancing_override is default initialized to 0, I think,
indicating that no override exists.
numabalancing_override == 1 indicates it has been overridden to true.
numabalancing_override == -1 indicates that it has been overridden to false.

The above code reads to me:

if (override_exists)
    set_numabalancing_state(override_value)

if (num_online_nodes() > ! && !override_exists)
    set_numabalancing_state(numabalancing_default)

A more clear fix for readability would be an early return between
these 2 if statements I think.

> After eight years, I wonder if we actually need this.
>