Date:	Thu, 7 May 2015 10:03:19 -0700
From:	josh@...htriplett.org
To:	Peter Hurley <peter@...leysoftware.com>
Cc:	Fengguang Wu <fengguang.wu@...el.com>,
	Iulia Manda <iulia.manda21@...il.com>,
	"Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Linux Memory Management List <linux-mm@...ck.org>,
	LKP <lkp@...org>, linux-kernel@...r.kernel.org
Subject: Re: [CONFIG_MULTIUSER] BUG: unable to handle kernel paging request
 at ffffffee

On Thu, May 07, 2015 at 12:24:07PM -0400, Peter Hurley wrote:
> On 05/07/2015 11:56 AM, josh@...htriplett.org wrote:
> > On Wed, May 06, 2015 at 08:39:22PM -0400, Peter Hurley wrote:
> >> On 05/06/2015 07:59 PM, josh@...htriplett.org wrote:
> >>> On Wed, May 06, 2015 at 08:44:29AM -0700, Josh Triplett wrote:
> >>>> On Wed, May 06, 2015 at 05:08:50PM +0800, Fengguang Wu wrote:
> >>>>> FYI, the reported bug is still not fixed in linux-next 20150506.
> >>>>
> >>>> This isn't the same bug.  The previous one you mentioned was a userspace
> >>>> assertion failure in libnih, likely caused because some part of upstart
> >>>> didn't have appropriate error handling for some syscall returning
> >>>> ENOSYS; that one wasn't an issue, since CONFIG_MULTIUSER=n is not
> >>>> expected to boot a standard Linux distribution.
> >>>>
> >>>> This one, on the other hand, is a kernel panic, and does need fixing.
> >>>>
> >>>>> commit 2813893f8b197a14f1e1ddb04d99bce46817c84a
> >>>>>
> >>>>> +-----------------------------------------------------------+------------+------------+------------+
> >>>>> |                                                           | c79574abe2 | 2813893f8b | cbdacaf0c1 |
> >>>>> +-----------------------------------------------------------+------------+------------+------------+
> >>>>> | boot_successes                                            | 60         | 0          | 0          |
> >>>>> | boot_failures                                             | 0          | 22         | 1064       |
> >>>>> | BUG:unable_to_handle_kernel                               | 0          | 22         | 1032       |
> >>>>> | Oops                                                      | 0          | 22         | 1032       |
> >>>>> | EIP_is_at_devpts_new_index                                | 0          | 22         | 1032       |
> >>>>> | Kernel_panic-not_syncing:Fatal_exception                  | 0          | 22         | 1032       |
> >>>>> | backtrace:do_sys_open                                     | 0          | 22         | 1032       |
> >>>>> | backtrace:SyS_open                                        | 0          | 22         | 1032       |
> >>>>> | WARNING:at_arch/x86/kernel/fpu/core.c:#fpu__clear()       | 0          | 0          | 32         |
> >>>>> | Kernel_panic-not_syncing:Attempted_to_kill_init!exitcode= | 0          | 0          | 32         |
> >>>>> +-----------------------------------------------------------+------------+------------+------------+
> >>>>
> >>>> Is this table saying the number of times the type of error in the first
> >>>> column occurred in each commit?
> >>>>
> >>>> In any case, investigating.  Iulia, can you look at this as well?
> >>>>
> >>>> I'm digging through the call stack, and I'm having a hard time seeing
> >>>> how the CONFIG_MULTIUSER patch could affect anything here.
> >>>
> >>> Update: it looks like init_devpts_fs is getting ERR_PTR(-EINVAL) back
> >>> from kern_mount and storing that in devpts_mnt; later, devpts_new_index
> >>> pokes at devpts_mnt and explodes.
> >>>
> >>> So, there are two separate bugs here.  On the one hand, CONFIG_MULTIUSER
> >>> should not be causing kern_mount to fail with -EINVAL; tracking that
> >>> down now.
> >>
> >> The mount failure is probably from the devpts mount options specifying
> >> gid= for devpts nodes:
> >>
> >> devpts /dev/pts devpts rw,nosuid,noexec,relatime,gid=5,mode=620,ptmxmode=000 0 0
> >>
> >> The relevant code is fs/devpts/inode.c:parse_mount_options().
> >> devpts also supports specifying the uid.
> >>
> >> To me, kern_mount() appropriately fails with -EINVAL, since the mount
> >> options failed.
> > 
> > Except that init_devpts_fs is called at module_init time, long before
> > the actual mount syscall; it appears to be creating a kernel-internal
> > mount, and I don't see how mount options provided by userspace much
> > later would cause the earlier kern_mount to fail.
> 
> Yeah, I realized that later; that the userspace mount is really a rebind
> to that initial root kernel mount.
>  
> > Also, I don't see anything in parse_mount_options that should actually
> > fail with CONFIG_MULTIUSER unset.
> 
> I didn't look deeper than that, but it seemed likely that it stemmed from
> that. Maybe it's related to CONFIG_DEVPTS_MULTIPLE_INSTANCES (documented
> in Documentation/fs/devpts.txt) and FS_USERNS_MOUNT?

Looks like it's actually mknod_ptmx that's failing; it's returning
EINVAL from the uid_valid/gid_valid checks, which shouldn't happen.

> >>> On the other hand, devpts and ptmx should handle the failure
> >>> better, without crashing; ptmx_open should have gracefully failed back
> >>> to userspace with -ENODEV or something, since ptmx doesn't make sense
> >>> without devpts.  I'll send a patch for that too.
> >>
> >> Yeah, crashing is bad, but I don't think we should even be init'ing
> >> either BSD or SysV pty drivers if there is no devpts.
> > 
> > Can you review the patch I sent to fix the crash, and see if it looks
> > reasonable to you?
> 
> On my todo list for today.

Thanks!

- Josh Triplett
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists