param_sysfs_init(), and in particular param_sysfs_builtin() is rather
time-consuming; for my board, it currently takes about 30ms.

That amounts to about 3% of the time budget I have from U-Boot hands
over control to linux and linux must assume responsibility for keeping
the external watchdog happy.

We must still continue to initialize module_kset at subsys_initcall
time, since otherwise any request_module() would fail in
mod_sysfs_init(). However, the bulk of the work in
param_sysfs_builtin(), namely populating /sys/module/*/version and/or
/sys/module/*/parameters/ for builtin modules, can be deferred to
late_initcall time - there's no userspace yet anyway to observe
contents of /sys or the lack thereof.

Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
---

This is on top of
https://lore.kernel.org/lkml/20220929121039.702873-1-linux@rasmusvillemoes.dk/
which I think has been applied somewhere, but I can't yet find it in
linux-next.

 kernel/params.c | 17 +++++++++++++++--
 1 file changed, 15 insertions(+), 2 deletions(-)

diff --git a/kernel/params.c b/kernel/params.c
index 8d4e9a3f0df2..db524cf51231 100644
--- a/kernel/params.c
+++ b/kernel/params.c
@@ -953,7 +953,7 @@ struct kobj_type module_ktype = {
 };
 
 /*
- * param_sysfs_init - wrapper for built-in params support
+ * param_sysfs_init - create "module" kset
  */
 static int __init param_sysfs_init(void)
 {
@@ -964,11 +964,24 @@ static int __init param_sysfs_init(void)
 		return -ENOMEM;
 	}
 
+	return 0;
+}
+subsys_initcall(param_sysfs_init);
+
+/*
+ * param_sysfs_builtin_init - add sysfs version and parameter
+ * attributes for built-in modules
+ */
+static int __init param_sysfs_builtin_init(void)
+{
+	if (!module_kset)
+		return -ENOMEM;
+
 	version_sysfs_builtin();
 	param_sysfs_builtin();
 
 	return 0;
 }
-subsys_initcall(param_sysfs_init);
+late_initcall(param_sysfs_builtin_init);
 
 #endif /* CONFIG_SYSFS */
-- 
2.37.2

On Tue, Oct 25, 2022 at 03:00:02PM +0200, Rasmus Villemoes wrote:
> param_sysfs_init(), and in particular param_sysfs_builtin() is rather
> time-consuming; for my board, it currently takes about 30ms.
> 
> That amounts to about 3% of the time budget I have from U-Boot hands
> over control to linux and linux must assume responsibility for keeping
> the external watchdog happy.
> 
> We must still continue to initialize module_kset at subsys_initcall
> time, since otherwise any request_module() would fail in
> mod_sysfs_init().

It would be good to document this through a comment.

> However, the bulk of the work in
> param_sysfs_builtin(), namely populating /sys/module/*/version and/or
> /sys/module/*/parameters/ for builtin modules, can be deferred to
> late_initcall time - there's no userspace yet anyway to observe
> contents of /sys or the lack thereof.
> 

Other than that, this looks good and looks cautious enough.

Acked-by: Luis Chamberlain <mcgrof@kernel.org>

Happy to take this in through moduels and give this a spin on linux-next
to see what blows up early. Can you send a v2 with a small code comment for
the above?

  Luis

param_sysfs_init(), and in particular param_sysfs_builtin() is rather
time-consuming; for my board, it currently takes about 30ms.

That amounts to about 3% of the time budget I have from U-Boot hands
over control to linux and linux must assume responsibility for keeping
the external watchdog happy.

We must still continue to initialize module_kset at subsys_initcall
time, since otherwise any request_module() would fail in
mod_sysfs_init(). However, the bulk of the work in
param_sysfs_builtin(), namely populating /sys/module/*/version and/or
/sys/module/*/parameters/ for builtin modules, can be deferred to
late_initcall time - there's no userspace yet anyway to observe
contents of /sys or the lack thereof.

Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
---
 kernel/params.c | 21 +++++++++++++++++++--
 1 file changed, 19 insertions(+), 2 deletions(-)

diff --git a/kernel/params.c b/kernel/params.c
index 8d4e9a3f0df2..a06f80c56f19 100644
--- a/kernel/params.c
+++ b/kernel/params.c
@@ -953,7 +953,11 @@ struct kobj_type module_ktype = {
 };
 
 /*
- * param_sysfs_init - wrapper for built-in params support
+ * param_sysfs_init - create "module" kset
+ *
+ * This must be done before the initramfs is unpacked and
+ * request_module() thus becomes possible, because otherwise the
+ * module load would fail in mod_sysfs_init.
  */
 static int __init param_sysfs_init(void)
 {
@@ -964,11 +968,24 @@ static int __init param_sysfs_init(void)
 		return -ENOMEM;
 	}
 
+	return 0;
+}
+subsys_initcall(param_sysfs_init);
+
+/*
+ * param_sysfs_builtin_init - add sysfs version and parameter
+ * attributes for built-in modules
+ */
+static int __init param_sysfs_builtin_init(void)
+{
+	if (!module_kset)
+		return -ENOMEM;
+
 	version_sysfs_builtin();
 	param_sysfs_builtin();
 
 	return 0;
 }
-subsys_initcall(param_sysfs_init);
+late_initcall(param_sysfs_builtin_init);
 
 #endif /* CONFIG_SYSFS */
-- 
2.37.2

On Wed, Oct 26, 2022 at 12:13:56AM +0200, Rasmus Villemoes wrote:
> param_sysfs_init(), and in particular param_sysfs_builtin() is rather
> time-consuming; for my board, it currently takes about 30ms.
> 
> That amounts to about 3% of the time budget I have from U-Boot hands
> over control to linux and linux must assume responsibility for keeping
> the external watchdog happy.
> 
> We must still continue to initialize module_kset at subsys_initcall
> time, since otherwise any request_module() would fail in
> mod_sysfs_init(). However, the bulk of the work in
> param_sysfs_builtin(), namely populating /sys/module/*/version and/or
> /sys/module/*/parameters/ for builtin modules, can be deferred to
> late_initcall time - there's no userspace yet anyway to observe
> contents of /sys or the lack thereof.
> 
> Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>

Thanks, queued onto modules-next.

  Luis