Re: Linux 2.6.38-rc4 (target_core: rmmod GP fault)

["Nicholas A. Bellinger" <nab@linux-iscsi.org>]

On Wed, 2011-02-09 at 11:00 -0800, Linus Torvalds wrote:
> On Wed, Feb 9, 2011 at 9:28 AM, Randy Dunlap <randy.dunlap@oracle.com> wrote:
> > x86_64, nearly allmodconfig.  No target hardware.
> >
> >
> > [  144.508473] general protection fault: 0000 [#1] SMP DEBUG_PAGEALLOC
> > [  144.509901] last sysfs file: /sys/devices/pci0000:00/0000:00:1d.1/usb6/6-1/6-1.3/devnum
> > [  144.512026] CPU 1
> > [  144.512026]
> > [  144.512026] Pid: 2597, comm: rmmod Not tainted 2.6.38-rc4 #1 0TY565/OptiPlex 745
> > [  144.512026] RIP: 0010:[<ffffffff810c3e5f>]  [<ffffffff810c3e5f>] __lock_acquire+0xd8/0x4e8
> > [  144.512026] RSP: 0018:ffff88006df1bb78  EFLAGS: 00010006
> > [  144.512026] RAX: 0000000000000002 RBX: 6b6b6b6b6b6b6be3 RCX: 0000000000000000
> 
> The code disassembles to
> 
>    0:	8d 01                	lea    (%rcx),%eax
>    2:	e8 6c b1 fb ff       	callq  0xfffffffffffbb173
>    7:	48 ff 05 8b 32 8d 01 	incq   0x18d328b(%rip)        # 0x18d3299
>    e:	48 ff 05 8c 32 8d 01 	incq   0x18d328c(%rip)        # 0x18d32a1
>   15:	48 ff 05 95 32 8d 01 	incq   0x18d3295(%rip)        # 0x18d32b1
>   1c:	e9 e3 03 00 00       	jmpq   0x404
>   21:	48 ff 05 81 32 8d 01 	incq   0x18d3281(%rip)        # 0x18d32a9
>   28:*	48 81 3b 40 5f 26 82 	cmpq   $0xffffffff82265f40,(%rbx)     <--
> trapping instruction
>   2f:	75 07                	jne    0x38
>   31:	48 ff 05 81 32 8d 01 	incq   0x18d3281(%rip)        # 0x18d32b9
>   38:	83 fe 01             	cmp    $0x1,%esi
> 
> and %rbx (and %rdi) contains the poison pattern for free'd memory (0x6b6b6b..).
> 
> > [  144.512026] Process rmmod (pid: 2597, threadinfo ffff88006df1a000, task ffff88006dec3000)
> 
> .. and that's likely not a very commonly tested case.
> 
> > [  144.512026]  [<ffffffffa06ace26>] configfs_unregister_subsystem+0x105/0x194 [configfs]
> > [  144.512026]  [<ffffffffa06baf55>] target_core_exit_configfs+0x185/0x1eb [target_core_mod]
> > [  144.512026]  [<ffffffff810d46a8>] sys_delete_module+0x2d6/0x368
> 
> The target_core_exit_configfs() code looks _very_ broken. It looks
> broken for two reasons:
> 
>  - it's very different from the cleanup code for the "failed to init"
> case in target_core_init_configfs, which does a lot less (see the
> "out:" code there)
> 

When registering a top level struct configfs_subsystem to appear under

	/sys/kernel/config/$SUBSYSTEM

the releasing of the top-level default group via
configfs_unregister_subsystem() during a failure in
target_core_init_configfs() is done for us, but we are still missing the
extra config_item_put()'s on the sub top-level groups (Joel, please
correct me)

The original 'out:' failure path code does not call config_item_put() on
these default groups, because config_group_init_type_name() has only
initialized struct config_group until configfs_register_subsystem() is
called to register the top level struct config_subsystem.

With the current 'out:' path being broken, to address the first point I
think moving the following code chunk in target_core_init_configfs to
before the configfs_register_subsystem() would make sense so that
configfs_register_subsystem() will fail last:

        /*
         * Register built-in RAMDISK subsystem logic for virtual LUN 0
         */
        ret = rd_module_init();
        if (ret < 0)
                goto out;

        if (core_dev_setup_virtual_lun0() < 0)
                goto out;

        return 0;

However looking at fs/configfs/dir.c:configfs_register_subsystem(), I
think the caller is still expected to release any sub top-level struct
config_group->default_groups[] w/ config_item_put() even though
unlink_group() is called from the configfs_attach_group() failure path..
(Joel..?)

>  - it seems to do a lot of manual freeing of the
> "su_group.default_groups" stuff etc, which is all internal configfs
> stuff, and seems to be used by the register/unregister phases.
> 

The specific issue rmmod with SLUB poisioning had been reported by Fubo
Chen to linux-scsi in the last weeks.  The patch to address the proper
release of the top-level + sub top-level struct configfs_subsystem's
default_groups in target_core_exit_configfs() has been committed into
the upstream tree in lio-core-2.6.git/linus-38-rc3 and sent out to
linux-scsi here:

[PATCH] target: Fix top-level configfs_subsystem default_group shutdown breakage
http://marc.info/?l=linux-scsi&m=129662389218924&w=2

> So somebody show knows configfs better should really check that
> cleanup, but it looks like target-core is just totally broken for the
> rmmod case.
> 
> Added more people to the cc. Nicholas, Joel and James. Guys: please
> check the insmod/rmmod case with
>  (a) spinlock debugging and lockdep enabled
>  (b) SLUB poisoning enabled.
> ie all of these should be on:
> 
>   CONFIG_SLUB_DEBUG_ON=y
>   CONFIG_DEBUG_SPINLOCK=y
>   CONFIG_DEBUG_MUTEXES=y
>   CONFIG_DEBUG_LOCK_ALLOC=y
>   CONFIG_PROVE_LOCKING=y
>   CONFIG_LOCKDEP=y
>   CONFIG_DEBUG_LOCKDEP=y
>   CONFIG_TRACE_IRQFLAGS=y
>   CONFIG_DEBUG_SPINLOCK_SLEEP=y
>   CONFIG_STACKTRACE=y
> 
> and you might also want to add CONFIG_DEBUG_PAGEALLOC to the mix.
> 

<nod>  I believe the above patch resolves the specific rmmod issue.
However, during SLUB poisioning testing we also came across errors with
the incorrect use of struct config_item_operations->release() in
target_core_configfs.c and target_core_fabric_configfs.c code.  The
series to address these was included in the last series to James here:

[PATCH 00/12] target: Updates for .38-rc4
http://marc.info/?l=linux-scsi&m=129680191624837&w=2

Note that this series for-38 mainline needs to be applied on top of the
original update series after the drivers/target/ mainline merge:

[PATCH 00/24] target updates for .38-rc3 (v2)
http://marc.info/?l=linux-scsi&m=129632617326015&w=2

The entire series is available from
     
   git://git.kernel.org/pub/scm/linux/kernel/git/nab/scsi-post-merge-2.6.git for-38-rc4

James, please review + sign-off so we can get these updates into mainline.