Re: WARNING: at mm/slub.c:3357, kernel BUG at mm/slub.c:3413

From: Eric W. Biederman
Date: Mon Dec 19 2011 - 04:05:23 EST


Markus Trippelsdorf <markus@xxxxxxxxxxxxxxx> writes:

> On 2011.12.18 at 19:21 -0800, Eric W. Biederman wrote:
>> Markus Trippelsdorf <markus@xxxxxxxxxxxxxxx> writes:
>>
>> > On 2011.11.21 at 17:34 +0100, Markus Trippelsdorf wrote:
>> >> On 2011.11.21 at 17:10 +0100, Markus Trippelsdorf wrote:
>> >> > On 2011.11.21 at 16:48 +0100, Eric Dumazet wrote:
>> >> > > Le lundi 21 novembre 2011 Ã 16:36 +0100, Markus Trippelsdorf a Ãcrit :
>> >> > > > On 2011.11.21 at 15:16 +0100, Eric Dumazet wrote:
>> >> > > > > Le lundi 21 novembre 2011 Ã 14:15 +0100, Markus Trippelsdorf a Ãcrit :
>> >> > > > >
>> >> > > > > > I've enabled CONFIG_SLUB_DEBUG_ON and this is what happend:
>> >> > > > > >
>> >> > > > >
>> >> > > > > Thanks
>> >> > > > >
>> >> > > > > Please continue to provide more samples.
>> >> > > > >
>> >> > > > > There is something wrong somewhere, but where exactly, its hard to say.
>> >> > > >
>> >> > > > New sample. This one points to lib/idr.c:
>> >> > > >
>> >> > > > =============================================================================
>> >> > > > BUG idr_layer_cache: Poison overwritten
>> >> > > > -----------------------------------------------------------------------------
>> >> > >
>> >> > > Thanks, could you now add "CONFIG_DEBUG_PAGEALLOC=y" in your config as
>> >> > > well ?
>> >> >
>> >> > Sure. This one happend with CONFIG_DEBUG_PAGEALLOC=y:
>> >> >
>> >> > =============================================================================
>> >> > BUG task_struct: Poison overwritten
>> >> > -----------------------------------------------------------------------------
>> >>
>> >> And sometimes this one that I've reported earlier already:
>> >>
>> >> (see: http://thread.gmane.org/gmane.linux.kernel/1215023 )
>> >>
>> >> ------------[ cut here ]------------
>> >> WARNING: at fs/sysfs/sysfs.h:195 sysfs_get_inode+0x136/0x140()
>> >> Hardware name: System Product Name
>> >> Pid: 1876, comm: slabinfo Not tainted 3.2.0-rc2-00274-g6fe4c6d #72
>> >> Call Trace:
>> >> [<ffffffff8106cac5>] warn_slowpath_common+0x75/0xb0
>> >> [<ffffffff8106cbc5>] warn_slowpath_null+0x15/0x20
>> >> [<ffffffff81163236>] sysfs_get_inode+0x136/0x140
>> >> [<ffffffff81164cef>] sysfs_lookup+0x6f/0x110
>> >> [<ffffffff811173f9>] d_alloc_and_lookup+0x39/0x80
>> >> [<ffffffff81118774>] do_lookup+0x294/0x3a0
>> >> [<ffffffff8111798a>] ? inode_permission+0x7a/0xb0
>> >> [<ffffffff8111a3f7>] do_last.isra.46+0x137/0x7f0
>> >> [<ffffffff8111ab76>] path_openat+0xc6/0x370
>> >> [<ffffffff81117606>] ? getname_flags+0x36/0x230
>> >> [<ffffffff810ec852>] ? handle_mm_fault+0x192/0x290
>> >> [<ffffffff8111ae5c>] do_filp_open+0x3c/0x90
>> >> [<ffffffff81127c8c>] ? alloc_fd+0xdc/0x120
>> >> [<ffffffff8110ce77>] do_sys_open+0xe7/0x1c0
>> >> [<ffffffff8110cf6b>] sys_open+0x1b/0x20
>> >> [<ffffffff814ccb7b>] system_call_fastpath+0x16/0x1b
>> >> ---[ end trace b1377eb8b131d37d ]---
>> >
>> > Hm, the "sysfs: use rb-tree" thing hit again during boot. Could this be
>> > the root cause of this all?
>> >
>> > I wrote down the following:
>> >
>> > RIP : rb_next
>> >
>> > Trace:
>> > sysfs_dir_pos
>> > sysfs_readdir
>> > ? sys_ioctl
>> > vfs_readdir
>> > sys_getdents
>>
>> Thanks for reporting this.
>>
>> Has this by any chance been resolved or stopped happening?
>
> Yes.
>
>> This looks for all of the world like something is stomping your sysfs
>> dirents. I haven't seen anyone else complaining so this seems like the
>> problem is unique to your configuration. Which suggests that it is not
>> sysfs itself that is wrong.
>>
>> I have been through the code a time or two and I haven't seen anything
>> obviously wrong. Everything that sysfs does is protected by the
>> sysfs_mutex so the locking is very very simple.
>>
>> My best guess of why now is that the rbtree code make a sysfs dirent
>> 48 bytes larger. And so it is much more exposed to these kinds of
>> problems.
>
> Sorry, but your subsystem was just accidentally hit by a bug in the
> Radeon driver, that sometimes randomly writes 0 dwords somewhere to
> memory after a kexec boot (see the rest of this huge thread).
> It's still not fixed in mainline, because Linus refused to take the fix
> this late in the series.

Awesome.

I guess that means I am only responsible for the Radeon driver having
the opportunity to take that code path. It is nice to see kexec being
used and being well known enough I didn't have to get involved.

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