On 2019/4/30 上午12:38, Cong Wang wrote:

On Sun, Apr 28, 2019 at 7:23 PM Jason Wang [off-list ref] wrote:

quoted

On 2019/4/29 上午1:59, Cong Wang wrote:

quoted

On Sun, Apr 28, 2019 at 12:51 AM Jason Wang [off-list ref] wrote:

quoted

quoted

tun_net_xmit() doesn't have the chance to
access the change because it holding the rcu_read_lock().

The problem is the following codes:


          --tun->numqueues;

          ...

          synchronize_net();

We need make sure the decrement of tun->numqueues be visible to readers
after synchronize_net(). And in tun_net_xmit():

It doesn't matter at all. Readers are okay to read it even they still use the
stale tun->numqueues, as long as the tfile is not freed readers can read
whatever they want...

This is only true if we set SOCK_RCU_FREE, isn't it?

Sure, this is how RCU is supposed to work.

quoted

quoted

The decrement of tun->numqueues is just how we unpublish the old
tfile, it is still valid for readers to read it _after_ unpublish, we only need
to worry about free, not about unpublish. This is the whole spirit of RCU.

The point is we don't convert tun->numqueues to RCU but use
synchronize_net().

Why tun->numqueues needs RCU? It is an integer, and reading a stale
value is _perfectly_ fine.

I meant we don't want e.g tun_net_xmit() to see the stale value after 
synchronize_net() in __tun_detach(), since it has various other steps 
with the assumption that no tfile dereference from data path. E.g one 
example is XDP rxq information un-registering which looks racy in the 
case of XDP_TX.

If you actually meant to say tun->tfiles[] itself, no, it is a fixed-size array,
it doesn't shrink or grow, so we don't need RCU for it. This is also why
a stale tun->numqueues is fine, as long as it never goes out-of-bound.

We do kind of shrinking or growing through tun->numqueues. That's why we 
check against it in various places. But, of course this is buggy.

quoted

quoted

You need to rethink about my SOCK_RCU_FREE patch.

The code is wrote before SOCK_RCU_FREE is introduced and assume no
de-reference from device after synchronize_net(). It doesn't harm to
figure out the root cause which may give us more confidence to the fix
(e.g like SOCK_RCU_FREE).

I believe SOCK_RCU_FREE is the fix for the root cause, not just a
cover-up.

quoted

I don't object to fix with SOCK_RCU_FREE, but then we should remove
the redundant synchronize_net(). But I still prefer to synchronize
everything explicitly like (completely untested):

I agree that synchronize_net() can be removed. However I don't
understand your untested patch at all, it looks like to fix a completely
different problem rather than this use-after-free.

As has been mentioned, the problem of current code is that we still 
leave pointers  to freed tfile in tfiles[] array in __tun_detach() and 
the check with tun->numqueues seems racy. So the patch just NULL out the 
detached tfile pointers and make sure no it can not be dereferenced from 
tfile after synchronize_net() by dereferencing tfile instead of checking 
tun->numqueues .


Thanks

Thanks.