Messages in this thread Patch in this message | | | Subject | Re: [PATCH v3 1/2] init/initramfs.c: do unpacking asynchronously | From | Rasmus Villemoes <> | Date | Wed, 28 Jul 2021 13:49:16 +0200 |
| |
On 28/07/2021 12.36, Alexander Egorenkov wrote: > Rasmus Villemoes <linux@rasmusvillemoes.dk> writes: > >> On 24/07/2021 09.46, Alexander Egorenkov wrote: >>> Hello, >>> >>> since e7cb072eb988 ("init/initramfs.c: do unpacking asynchronously"), we >>> started seeing the following problem on s390 arch regularly: >>> >>> [ 5.039734] wait_for_initramfs() called before rootfs_initcalls >> >> While that message was added as part of the same patch, it's a red >> herring: It merely means that something ends up calling usermodehelper >> (perhaps a request_module) before the init sequence has come around to >> rootfs_initcalls. At that point, the rootfs is (with or without my async >> patch) entirely empty, so those usermodehelper calls have always failed >> with -ENOENT. >> >> If you have CONFIG_UEVENT_HELPER=y and CONFIG_UEVENT_HELPER_PATH set to >> a non-empty string, you can try setting the latter to the empty string. >> But the message won't go away if it's really a request_module() and not >> a uevent notification. >> > > Thanks for the helpful explanation. I disabled UEVENT on my test machine and > as you said, the message is being triggered by request_module() now.
Yes. But as I've said, the "called before rootfs_initcalls" message is entirely harmless and not in any way related to whatever problems is later encountered when the initramfs is actually being decompressed and extracted.
I have this so far untested patch that I plan on sending, which should remove that message
diff --git a/init/initramfs.c b/init/initramfs.c index af27abc59643..51a686a8c929 100644 --- a/init/initramfs.c +++ b/init/initramfs.c @@ -727,6 +727,7 @@ 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; diff --git a/init/main.c b/init/main.c index f5b8246e8aa1..d5c5542fe142 100644 --- a/init/main.c +++ b/init/main.c @@ -1387,7 +1387,6 @@ static void __init do_basic_setup(void) driver_init(); init_irq_proc(); do_ctors(); - usermodehelper_enable(); do_initcalls(); }
diff --git a/init/noinitramfs.c b/init/noinitramfs.c index 3d62b07f3bb9..f1d9e5495cc1 100644 --- a/init/noinitramfs.c +++ b/init/noinitramfs.c @@ -18,6 +18,7 @@ static int __init default_rootfs(void) { int err;
+ usermodehelper_enable(); err = init_mkdir("/dev", 0755); if (err < 0) goto out; because any call of a usermodehelper (be it a uevent hotplug notification or a request_module) would just return -EBUSY during all pure_, core_, postcore_, arch_, subsys_ and fs_ initcalls. (It is really beyond me why rootfs_ initcalls are hidden between fs_ and device_ initcalls).
Currently (with or without my async patch) umh calls from those initcalls probably return -ENOENT or whatever kernel_execve() returns when there's no such binary. So if something actually looks at the return value, the change from -ENOENT to -EBUSY might cause a regression. But I doubt it.
Rasmus
| |