init: move usermodehelper_enable() to populate_rootfs()

Currently, usermodehelper is enabled right before PID1 starts going
through the initcalls. However, any call of a usermodehelper from a
pure_, core_, postcore_, arch_, subsys_ or fs_ initcall is futile, as
there is no filesystem contents yet.

Up until commit e7cb072 ("init/initramfs.c: do unpacking
asynchronously"), such calls, whether via some request_module(), a
legacy uevent "/sbin/hotplug" notification or something else, would
just fail silently with (presumably) -ENOENT from
kernel_execve(). However, that commit introduced the
wait_for_initramfs() synchronization hook which must be called from
the usermodehelper exec path right before the kernel_execve, in order
that request_module() et al done from *after* rootfs_initcall()
time (i.e. device_ and late_ initcalls) would continue to find a
populated initramfs as they used to.

Any call of wait_for_initramfs() done before the unpacking has been
scheduled (i.e. before rootfs_initcall time) must just return
immediately [and let the caller find an empty file system] in order
not to deadlock the machine. I mistakenly thought, and my limited
testing confirmed, that there were no such calls, so I added a
pr_warn_once() in wait_for_initramfs(). It turns out that one can
indeed hit request_module() as well as kobject_uevent_env() during
those early init calls, leading to a user-visible warning in the
kernel log emitted consistently for certain configurations.

We could just remove the pr_warn_once(), but I think it's better to
postpone enabling the usermodehelper framework until there is at least
some chance of finding the executable. That is also a little more
efficient in that a lot of work done in umh.c will be elided. However,
it does change the error seen by those early callers from -ENOENT to
-EBUSY, so there is a risk of a regression if any caller care about
the exact error value.

Link: https://lkml.kernel.org/r/20210728134638.329060-1-linux@rasmusvillemoes.dk
Fixes: e7cb072 ("init/initramfs.c: do unpacking asynchronously")
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Reported-by: Alexander Egorenkov <egorenar@linux.ibm.com>
Reported-by: Bruno Goncalves <bgoncalv@redhat.com>
Reported-by: Heiner Kallweit <hkallweit1@gmail.com>
Cc: Luis Chamberlain <mcgrof@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

Author: Villemoes

init/initramfs.c

@@ -15,6 +15,7 @@
 #include <linux/mm.h>
 #include <linux/namei.h>
 #include <linux/init_syscalls.h>
+#include <linux/umh.h>
 
 static ssize_t __init xwrite(struct file *file, const char *p, size_t count,
 		loff_t *pos)
@@ -727,6 +728,7 @@ static int __init populate_rootfs(void)
 {
 	initramfs_cookie = async_schedule_domain(do_populate_rootfs, NULL,
 						 &initramfs_domain);
+	usermodehelper_enable();
 	if (!initramfs_async)
 		wait_for_initramfs();
 	return 0;

init/main.c

@@ -1392,7 +1392,6 @@ static void __init do_basic_setup(void)
 	driver_init();
 	init_irq_proc();
 	do_ctors();
-	usermodehelper_enable();
 	do_initcalls();
 }
 

init/noinitramfs.c

@@ -10,6 +10,7 @@
 #include <linux/kdev_t.h>
 #include <linux/syscalls.h>
 #include <linux/init_syscalls.h>
+#include <linux/umh.h>
 
 /*
  * Create a simple rootfs that is similar to the default initramfs
@@ -18,6 +19,7 @@ static int __init default_rootfs(void)
 {
 	int err;
 
+	usermodehelper_enable();
 	err = init_mkdir("/dev", 0755);
 	if (err < 0)
 		goto out;