Re: Oops in 2.6.13-rc5-git-current (0d317fb72fe3cf0f611608cf3a3015bbe6cd2a66)

[Alexander Nyberg <alexn@telia.com>]

> Unable to handle kernel paging request at virtual address 6b6b6b6b
>  printing eip:
> c0188d15
> *pde = 00000000
> Oops: 0000 [#1]
> PREEMPT 
> CPU:    0
> EIP:    0060:[inotify_inode_queue_event+85/336]    Not tainted VLI
> EFLAGS: 00010206   (2.6.13-rc5-g0d317fb7) 
> EIP is at inotify_inode_queue_event+0x55/0x150
> eax: 6b6b6b6b   ebx: 6b6b6b63   ecx: 00000000   edx: 00000066
> esi: c3effe34   edi: ce8c76ac   ebp: d4bb864c   esp: d8655eb0
> ds: 007b   es: 007b   ss: 0068
> Process nfsd (pid: 3750, threadinfo=d8654000 task=d6155020)
> Stack: 00000286 00000286 00000000 00000400 d4bb8760 d4bb8768 00000000 c3effe34 
>        ce8c76ac d4bb864c c0170626 00000000 c3effe34 d6608ad4 db74b17c c3effe34 
>        e0cfe9a4 00000013 e0d01b34 c0dd91b4 ce8c76ac ffffc000 d66092dc d66093c4 
> Call Trace:
>  [vfs_unlink+358/560] vfs_unlink+0x166/0x230
>  [pg0+544348580/1067586560] nfsd_unlink+0x104/0x230 [nfsd]
>  [pg0+544361268/1067586560] nfsd_cache_lookup+0x1c4/0x3c0 [nfsd]
>  [pg0+544371728/1067586560] nfsd3_proc_remove+0x80/0xc0 [nfsd]
>  [pg0+544381018/1067586560] nfs3svc_decode_diropargs+0x8a/0x100 [nfsd]
>  [pg0+544380880/1067586560] nfs3svc_decode_diropargs+0x0/0x100 [nfsd]
>  [pg0+544321698/1067586560] nfsd_dispatch+0x82/0x1f0 [nfsd]
>  [svc_authenticate+112/336] svc_authenticate+0x70/0x150
>  [svc_process+960/1648] svc_process+0x3c0/0x670
>  [pg0+544323105/1067586560] nfsd+0x1a1/0x350 [nfsd]
>  [ret_from_fork+6/20] ret_from_fork+0x6/0x14
>  [pg0+544322688/1067586560] nfsd+0x0/0x350 [nfsd]
>  [kernel_thread_helper+5/16] kernel_thread_helper+0x5/0x10

(akpm: a fix for this needs to go into 2.6.13, inotify + nfs 
trivially oopses otherwise, even if inotify isn't actively used)

It looks like the following sequence is done in the wrong order.
When vfs_unlink() is called from sys_unlink() it has taken a ref
on the inode and sys_unlink() does the last iput() but when called
from other callsites vfs_unlink() might do the last iput() and
free inode, so inotify_inode_queue_event() will receive an already
freed object and dereference an already freed object.

Signed-off-by: Alexander Nyberg <alexn@telia.com>

Index: mm/fs/namei.c
===================================================================
--- mm.orig/fs/namei.c	2005-08-07 12:06:16.000000000 +0200
+++ mm/fs/namei.c	2005-08-07 18:17:20.000000000 +0200
@@ -1869,8 +1869,8 @@
 	/* We don't d_delete() NFS sillyrenamed files--they still exist. */
 	if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
 		struct inode *inode = dentry->d_inode;
-		d_delete(dentry);
 		fsnotify_unlink(dentry, inode, dir);
+		d_delete(dentry);
 	}
 
 	return error;