Re: add file system helpers that take kernel pointers for the init code v4

Re: add file system helpers that take kernel pointers for the init code v4

[Qian Cai]

On Tue, Jul 28, 2020 at 06:33:53PM +0200, Christoph Hellwig wrote:
> Hi Al and Linus,
> 
> currently a lot of the file system calls in the early in code (and the
> devtmpfs kthread) rely on the implicit set_fs(KERNEL_DS) during boot.
> This is one of the few last remaining places we need to deal with to kill
> off set_fs entirely, so this series adds new helpers that take kernel
> pointers.  These helpers are in init/ and marked __init and thus will
> be discarded after bootup.  A few also need to be duplicated in devtmpfs,
> though unfortunately.

Reverting this series from next-20200803 fixed the crash below on shutdown.

[ 7303.287890][    T1] systemd-shutdown[1]: All loop devices detached.
[ 7303.287930][    T1] systemd-shutdown[1]: Detaching DM devices.
[ 7303.441674][    T1] printk: shutdown: 9 output lines suppressed due to ratelimiting
[ 7303.443999][    T1] Kernel panic - not syncing: Attempted to kill init! exitcode=0x00000100
[ 7303.444027][    T1] CPU: 120 PID: 1 Comm: shutdown Not tainted 5.8.0-next-20200803 #2
[ 7303.444053][    T1] Call Trace:
[ 7303.444069][    T1] [c000000015d27b70] [c0000000006f3778] dump_stack+0xfc/0x174 (unreliable)
[ 7303.444103][    T1] [c000000015d27bc0] [c0000000000c9328] panic+0x214/0x4ac
[ 7303.444140][    T1] [c000000015d27c60] [c0000000000d4d28] do_exit+0xda8/0xee0
[ 7303.444178][    T1] [c000000015d27d60] [c0000000000d4f1c] do_group_exit+0x5c/0xd0
list_empty at include/linux/list.h:282
(inlined by) thread_group_empty at include/linux/sched/signal.h:671
(inlined by) do_group_exit at kernel/exit.c:888
[ 7303.444205][    T1] [c000000015d27da0] [c0000000000d4fac] sys_exit_group+0x1c/0x20
sys_exit_group at kernel/exit.c:914
[ 7303.444234][    T1] [c000000015d27dc0] [c00000000002c628] system_call_exception+0xf8/0x1d0
[ 7303.444262][    T1] [c000000015d27e20] [c00000000000d0a8] system_call_common+0xe8/0x218
[ 7304.936912][    T1] ---[ end Kernel panic - not syncing: Attempted to kill init! exitcode=0x00000100 ]---

> 
> The series sits on top of my previous
> 
>   "decruft the early init / initrd / initramfs code v2"
> 
> series.
> 
> 
> Git tree:
> 
>     git://git.infradead.org/users/hch/misc.git init_path
> 
> Gitweb:
> 
>     http://git.infradead.org/users/hch/misc.git/shortlog/refs/heads/init_path
> 
> 
> Changes since v3:
>  - rename fs/for_init.c to fs/init.c
>  - document the purpose of the routines in fs/init.c with a comment
>  - don't mark devtmpfs __init as that will cause it to get overwritten
>    by initmem poisoning
>  - add an init_dup helper to make Al more happy than with the version
>    commit to the "decruft the early init / initrd / initramfs code v2"
>    series
> 
> Changes since v2:
>  - move to fs/for_init.c
>  - reuse the init routines in devtmpfs after refactoring devtmpfsd
>    (and thus the broken error handling in the previous version)
>  - actually use kern_path in a place where user_path_at sneaked back in
> 
> Changes since v1:
>  - avoid most core VFS changes
>  - renamed the functions and move them to init/ and devtmpfs
>  - drop a bunch of cleanups that can be submitted independently now
> 
> 
> Diffstat:

Re: add file system helpers that take kernel pointers for the init code v4

[Christoph Hellwig]

On Mon, Aug 03, 2020 at 10:56:23AM -0400, Qian Cai wrote:
> On Tue, Jul 28, 2020 at 06:33:53PM +0200, Christoph Hellwig wrote:
> > Hi Al and Linus,
> > 
> > currently a lot of the file system calls in the early in code (and the
> > devtmpfs kthread) rely on the implicit set_fs(KERNEL_DS) during boot.
> > This is one of the few last remaining places we need to deal with to kill
> > off set_fs entirely, so this series adds new helpers that take kernel
> > pointers.  These helpers are in init/ and marked __init and thus will
> > be discarded after bootup.  A few also need to be duplicated in devtmpfs,
> > though unfortunately.
> 
> Reverting this series from next-20200803 fixed the crash below on shutdown.

Please try this patch:

---
From 6448eebe2fe7189cedc5136ab3464517956922b7 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Mon, 3 Aug 2020 15:56:18 +0200
Subject: init: fix init_dup

Don't allocate an unused fd for each call.  Also drop the extra
reference from filp_open after the init_dup calls while we're at it.

Fixes: 36e96b411649 ("init: add an init_dup helper")
Reported-by Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 fs/init.c   | 2 +-
 init/main.c | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/init.c b/fs/init.c
index 730e05acda2392..e9c320a48cf157 100644
--- a/fs/init.c
+++ b/fs/init.c
@@ -260,6 +260,6 @@ int __init init_dup(struct file *file)
 	fd = get_unused_fd_flags(0);
 	if (fd < 0)
 		return fd;
-	fd_install(get_unused_fd_flags(0), get_file(file));
+	fd_install(fd, get_file(file));
 	return 0;
 }
diff --git a/init/main.c b/init/main.c
index 089e21504b1fc1..9dae9c4f806bb9 100644
--- a/init/main.c
+++ b/init/main.c
@@ -1470,6 +1470,7 @@ void __init console_on_rootfs(void)
 	init_dup(file);
 	init_dup(file);
 	init_dup(file);
+	fput(file);
 }
 
 static noinline void __init kernel_init_freeable(void)
-- 
2.27.0