-----Original Message-----
From: Al Viro [mailto:viro@ftp.linux.org.uk] On Behalf Of Al Viro
Sent: Saturday, August 31, 2013 3:36 PM
To: Liu, Chuansheng
Cc: Eric Dumazet; linux-fsdevel@vger.kernel.org; linux-kernel@vger.kernel.org
Subject: Re: [PATCH] Fix the race between the fget() and close()

On Sat, Aug 31, 2013 at 07:01:33AM +0000, Liu, Chuansheng wrote:

quoted

My scenario is:
P1 files_struct refcount is 1, P2's is 1 also.
P1 get_files_struct(P2)
P1 install one file into P2's files_struct
P1 put_files_struct(P2)

Then P1 and P2's files_struct refcount are 1, then when P1 is doing ioctl() and

P2 is exiting

quoted

with put_files_struct(P2), the race will occur, my understanding is wrong?

First of all, this wouldn't have been a problem (so you get a new reference
to file inserted in P2's files_struct; file refcount had been bumped, so
destruction of P2's files_struct will undo that increment of file refcount
and we are still fine).  _Removal_ in a similar scenario would have been
a problem, with P2 doing fdget() while its table isn't shared, then P1
removing a reference from it and dropping a file - the last one, at that,
since fdget() assumed that the reference would've stayed in P2's descriptor
table.  HOWEVER, P1 does not do get_files_struct(P2) at all - it's only
done by P2 in binder_mmap().

Got it, thanks. In other process, the fget() + _fd_install() should be the same
as the process call open() directly, and the file reference count will be at least 2.

Again, the invariant to look for is this:
	* if current->files had not been shared at fdget() time, it won't
be shared at matching fdput() and no entries will have been removed in
between.

task_fd_install()/task_close_fd() are done on proc->files, which contributes
to descriptor table refcount.  All other modifications are done to
current->files, which also contributes to refcount.  If at fdget() time
current->files had refcount 1, we had no other processes with task->files
pointing to this descriptor table *and* no binder_proc had their ->files
pointint to it.  No new ones may appear, since new process could get
such a reference only from do_fork() called by us and new binder_proc could
get such a reference only from binder_mmap() called by us.  Neither is
called between fdget() and fdput().  So in that case the only reference
to this descriptor table will remain current->files and all removals
would have to be done by ourselves (and not via task_close_fd(), at that).

And AFAICS, binder_lock() prevents proc->files being dropped under
task_close_fd() and task_fd_install().  Hell knows...

How reproducible it is?  Do you have any more instances, or had that
been a one-off panic?

Just meet once yet.