RE: [PATCH] Fix the race between the fget() and close()

[Liu, Chuansheng]

Hello Al,

> -----Original Message-----
> From: Al Viro [mailto:viro@xxxxxxxxxxxxxxxx] On Behalf Of Al Viro
> Sent: Tuesday, August 27, 2013 8:49 AM
> To: Liu, Chuansheng
> Cc: Eric Dumazet; linux-fsdevel@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH] Fix the race between the fget() and close()
> 
> On Tue, Aug 27, 2013 at 01:42:47AM +0100, Al Viro wrote:
> > Might be buggered refcounting on struct file somewhere (i.e. extra fput()
> done,
> > getting the file freed *before* close(), leaving a dangling pointer in
> > descriptor table).  Might be memory corruption of some kind, slapping junk
> > pointer into descriptor table.  Might be buggered refcounting on struct
> > dentry - if extra dput() is done somewhere, dentry might get freed under
> > us or become negative.
> >
> > Hell, might be buggered refcounting on descriptor table - binder is playing
> > interesting games there.  Try to reproduce that with
> CONFIG_DEBUG_KMEMLEAK
> > and slab debugging turned on, see if you hit anything from those; if it's
> > more or less readily reproducible, I would start with that - too many
> > scenarios involve broken refcounting of one sort or another.
> 
> Nevermind dentry refcounting - you get NULL dentry, not NULL inode.
> Other scenarios still remain, so I'd really recommend slab/kmemleak
> debugging turned on.

I think I found one of possible race here(two processes P1 and P2):
P1 has the the files_struct pointer FILES1, P2 has the files_struct pointer FILES2,

When P1 open file, the new struct file pointer SHARE_FILE will be installed into FILES1,
and file refcount is 1;

And in P1, we can get P2's files_struct FILES2, and thru _fd_install(), we can add SHARE_FILE
into P2's FILES2.

Then the same file pointer SHARE_FILE stayed in both P1 and P2's files_struct, and the panic case
will happen:
P1                                                            P2
Open the SHARE_FILE
Installed SHARE_FILE into P2's file_struct FILES2

Ioctl(SHARE_FILE)                                                When P2 exiting,
 fget_light()
   due to FILES1->refcount is 1,                                     put_files_struct will be called,
   there will be no RCU and SHARE_FILE refcount increasing                will close all files including SHARE_FILE

But at this time, P1 is still operate SHARE_FILE without the refcount safety.

Then the panic will happen at vfs_ioctl() due to the SHARE_FILE has been freed.

Is it allowable that installing one file pointer into another FILES_STRUCT? Seems binder is doing the similar things.
In fact, if in ioctl function, we can call fget() instead of fget_light(), this panic can be avoided.

Is it making sense?





--
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/