We have a need of using favordynmods with cgroup v1, which doesn't support
changing mount flags during remount. Enabling CONFIG_CGROUP_FAVOR_DYNMODS at
build-time is not an option because we want to be able to selectively
enable it for certain systems.

This commit addresses this by introducing the cgroup_favordynmods=
command-line option. This option works for both cgroup v1 and v2 and also
allows for disabling favorynmods when the kernel built with
CONFIG_CGROUP_FAVOR_DYNMODS=y.

Also, note that when cgroup_favordynmods=true favordynmods is never
disabled in cgroup_destroy_root().

Signed-off-by: Luiz Capitulino <luizcap@amazon.com>
---
 .../admin-guide/kernel-parameters.txt          |  4 ++++
 kernel/cgroup/cgroup.c                         | 18 ++++++++++++++----
 2 files changed, 18 insertions(+), 4 deletions(-)

o v3
 - Handle destroy case [Michal]
 - Fix type in commit log [Michal]

o v2
 - Use __ro_after_init [Waiman]

Michal,

For the cgroup_destroy_root() case, I opted to keep disabling favordynmods
when cgroup_favordynmods=false. The rationale is that it should allow
for disabling favordynmods when/if all cgroups are gone if the user so wants.

diff --git a/Documentation/admin-guide/kernel-parameters.txt b/Documentation/admin-guide/kernel-parameters.txt
index 0a1731a0f0ef..8b744d39d393 100644
--- a/Documentation/admin-guide/kernel-parameters.txt
+++ b/Documentation/admin-guide/kernel-parameters.txt
@@ -580,6 +580,10 @@
 			named mounts. Specifying both "all" and "named" disables
 			all v1 hierarchies.
 
+	cgroup_favordynmods= [KNL] Enable or Disable favordynmods.
+			Format: { "true" | "false" }
+			Defaults to the value of CONFIG_CGROUP_FAVOR_DYNMODS.
+
 	cgroup.memory=	[KNL] Pass options to the cgroup memory controller.
 			Format: <string>
 			nosocket -- Disable socket memory accounting.
diff --git a/kernel/cgroup/cgroup.c b/kernel/cgroup/cgroup.c
index 1fb7f562289d..06515550e609 100644
--- a/kernel/cgroup/cgroup.c
+++ b/kernel/cgroup/cgroup.c
@@ -207,6 +207,8 @@ static u16 have_exit_callback __read_mostly;
 static u16 have_release_callback __read_mostly;
 static u16 have_canfork_callback __read_mostly;
 
+static bool have_favordynmods __ro_after_init = IS_ENABLED(CONFIG_CGROUP_FAVOR_DYNMODS);
+
 /* cgroup namespace for init task */
 struct cgroup_namespace init_cgroup_ns = {
 	.ns.count	= REFCOUNT_INIT(2),
@@ -1350,7 +1352,9 @@ static void cgroup_destroy_root(struct cgroup_root *root)
 		cgroup_root_count--;
 	}
 
-	cgroup_favor_dynmods(root, false);
+	if (!have_favordynmods)
+		cgroup_favor_dynmods(root, false);
+
 	cgroup_exit_root_id(root);
 
 	cgroup_unlock();
@@ -2243,9 +2247,9 @@ static int cgroup_init_fs_context(struct fs_context *fc)
 	fc->user_ns = get_user_ns(ctx->ns->user_ns);
 	fc->global = true;
 
-#ifdef CONFIG_CGROUP_FAVOR_DYNMODS
-	ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
-#endif
+	if (have_favordynmods)
+		ctx->flags |= CGRP_ROOT_FAVOR_DYNMODS;
+
 	return 0;
 }
 
@@ -6764,6 +6768,12 @@ static int __init enable_cgroup_debug(char *str)
 }
 __setup("cgroup_debug", enable_cgroup_debug);
 
+static int __init cgroup_favordynmods_setup(char *str)
+{
+	return (kstrtobool(str, &have_favordynmods) == 0);
+}
+__setup("cgroup_favordynmods=", cgroup_favordynmods_setup);
+
 /**
  * css_tryget_online_from_dir - get corresponding css from a cgroup dentry
  * @dentry: directory dentry of interest
-- 
2.40.1

On Wed, Sep 27, 2023 at 02:25:40PM +0000, Luiz Capitulino wrote:
> We have a need of using favordynmods with cgroup v1, which doesn't support
> changing mount flags during remount. Enabling CONFIG_CGROUP_FAVOR_DYNMODS at
> build-time is not an option because we want to be able to selectively
> enable it for certain systems.
> 
> This commit addresses this by introducing the cgroup_favordynmods=
> command-line option. This option works for both cgroup v1 and v2 and also
> allows for disabling favorynmods when the kernel built with
> CONFIG_CGROUP_FAVOR_DYNMODS=y.
> 
> Also, note that when cgroup_favordynmods=true favordynmods is never
> disabled in cgroup_destroy_root().
> 
> Signed-off-by: Luiz Capitulino <luizcap@amazon.com>

Applied to cgroup/for-6.7.

Thanks.

-- 
tejun

On Wed, Sep 27, 2023 at 02:25:40PM +0000, Luiz Capitulino <luizcap@amazon.com> wrote:
> We have a need of using favordynmods with cgroup v1, which doesn't support
> changing mount flags during remount. Enabling CONFIG_CGROUP_FAVOR_DYNMODS at
> build-time is not an option because we want to be able to selectively
> enable it for certain systems.
> 
> This commit addresses this by introducing the cgroup_favordynmods=
> command-line option. This option works for both cgroup v1 and v2 and also
> allows for disabling favorynmods when the kernel built with
> CONFIG_CGROUP_FAVOR_DYNMODS=y.
> 
> Also, note that when cgroup_favordynmods=true favordynmods is never
> disabled in cgroup_destroy_root().
> 
> Signed-off-by: Luiz Capitulino <luizcap@amazon.com>
> ---
>  .../admin-guide/kernel-parameters.txt          |  4 ++++
>  kernel/cgroup/cgroup.c                         | 18 ++++++++++++++----
>  2 files changed, 18 insertions(+), 4 deletions(-)

(resend too)
Reviewed-by: Michal Koutný <mkoutny@suse.com>