Re: 2.6.22-rc1-mm1 cifs_mount oops

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 23 May 2007 00:50:13 +0000 "young dave" <hidave.darkstar@gmail.com> wrote:

> Hi,
> when I use mount -t cifs , the kernel oops, seems break at
> kthread_stop, I'm not sure.
> 
> But if I add the CONFIG_CIFS_DEBUG2=y to config file, rebuild kernel,
> then the oops disappeared.
> 
> Below is the oops message:
> 
> BUG: unable to handle kernel NULL pointer dereference at virtual
> address 00000008
>  printing eip:
> c012e910
> *pde = 00000000
> Oops: 0002 [#1]
> PREEMPT
> Modules linked in: cifs smbfs radeon drm ipv6 snd_seq_dummy
> snd_seq_oss snd_seq_midi_event snd_seq snd_seq_device snd_pcm_oss
> snd_mixer_oss capability commoncap e100 mii psmouse sg evdev serio_raw
> snd_hda_intel snd_pcm snd_timer snd soundcore snd_page_alloc intel_agp
> agpgart i2c_i801 pcspkr
> CPU:    0
> EIP:    0060:[<c012e910>]    Not tainted VLI
> EFLAGS: 00210246   (2.6.22-rc1-mm1 #3)
> EIP is at kthread_stop+0x10/0x90
> eax: c051bde0   ebx: 00000000   ecx: c1fba000   edx: c1fef040
> esi: 00000000   edi: 00000064   ebp: c2a36c80   esp: c1fbbd58
> ds: 007b   es: 007b   fs: 0000  gs: 0033  ss: 0068
> Process mount.cifs (pid: 3955, ti=c1fba000 task=c2b38540 task.ti=c1fba000)
> Stack: c1fef040 ffffff90 ffffff90 f8a7a328 c285a504 f8a9a9fb 00000083 000000cf
>        000000dc 0000000b c2b38540 c2af5740 c292c540 00000000 00000000 c285a4c0
>        00000000 c411b400 c3a4f500 c3cec200 c1fef052 c291c1e0 c1fef037 c291c940
> Call Trace:
>  [<f8a7a328>] cifs_mount+0xbe8/0xf10 [cifs]
>  [<c02633fe>] idr_get_new_above_int+0x3e/0x50
>  [<f8a6d04e>] cifs_read_super+0x4e/0x160 [cifs]
>  [<c0166fc0>] set_anon_super+0x0/0xd0
>  [<f8a6d6c0>] cifs_get_sb+0x60/0xd0 [cifs]
>  [<c0167491>] vfs_kern_mount+0x91/0x130
>  [<c017d648>] permit_mount+0x28/0xa0
>  [<c017e09a>] do_new_mount+0x8a/0x140
>  [<c017e95e>] do_mount+0x25e/0x280
>  [<c0440000>] schedule+0x2e0/0x680
>  [<c017e602>] exact_copy_from_user+0x32/0x70
>  [<c017e69a>] copy_mount_options+0x5a/0xc0
>  [<c017ec19>] sys_mount+0x79/0xc0
>  [<c01040bc>] syscall_call+0x7/0xb
>  =======================
> Code: 88 d1 d3 e0 89 43 5c 83 c4 18 5b c3 eb 0d 90 90 90 90 90 90 90
> 90 90 90 90 90 90 53 83 ec 08 89 c3 b8 e0 bd 51 c0 e8 90 26 31 00 <ff>
> 43 08 31 c9 b8 f0 c1 58 c0 89 0d ec c1 58 c0 e8 3b 01 00 00
> EIP: [<c012e910>] kthread_stop+0x10/0x90 SS:ESP 0068:c1fbbd58
> 

I assume cifs_demultiplex_thread() took the SIGKILL, zeroed server->tsk
then exitted.  Then, cifs_mount() did a kthread_stop() on the now-NULL
pointer.

I don't see a non-racy way of fixing this as the code stands at present. 
This:

--- a/fs/cifs/connect.c~cifs-oops-fix
+++ a/fs/cifs/connect.c
@@ -2086,7 +2086,6 @@ cifs_mount(struct super_block *sb, struc
 					if ((temp_rc == -ESHUTDOWN) &&
 					   (pSesInfo->server) && (pSesInfo->server->tsk)) {
 						send_sig(SIGKILL,pSesInfo->server->tsk,1);
-						kthread_stop(pSesInfo->server->tsk);
 					}
 				} else
 					cFYI(1, ("No session or bad tcon"));
_

has a decent chance of fixing it.  But it's now racy against thread
*startup*: if we send SIGKILL to that task before it has done its
allow_signal(), it will presumably never get shut down.

Steve, can we just pull all the signal stuff out of there and use the
kthread machinery alone?