Re:Re: Re:Re: Re:Re: Re: [PATCH net] ppp: Fix a scheduling-while-atomic bug in del_chan

[Gao Feng]

At 2017-08-10 05:00:19, "Cong Wang"  wrote:
>On Wed, Aug 9, 2017 at 12:17 AM, Gao Feng  wrote:
>> Hi Cong,
>>
>> Actually I have one question about the SOCK_RCU_FREE.
>> I don't think it could resolve the issue you raised even though it exists 
>> really.
>>
>> I checked the SOCK_RCU_FREE, it just defer the __sk_destruct after one rcu 
>> period.
>> But when it performs, someone still could find this sock by callid during 
>> the del_chan period and it may still deference the sock which may freed soon.
>>
>> The right flow should be following.
>> del_chan()
>> wait a rcu period
>> sock_put()  It is safe that someone gets the sock because it 
>> already hold sock refcnt.
>>
>> When using SOCK_RCU_FREE, its flow would be following.
>> wait a rcu period
>> del_chan()
>> free the sock directly  no sock refcnt check again.
>> Because the del_chan happens after rcu wait, not before, so it isn't helpful 
>> with SOCK_RCU_FREE.
>
>
>Yes, good point! With SOCK_RCU_FREE the sock_hold() should
>not be needed. For RCU, unpublish should indeed happen before
>grace period.

Sorry, I couldn't understand why sock_hold() isn't necessary with SOCK_RCU_FREE.
When lookup_chan finds the sock, it would return and reference it later.
If no refcnt, how to protect the sock ?

Best Regards
Feng

>
>
>>
>> I don't know if I misunderstand the SOCK_RCU_FREE usage.
>>
>> But it is a good news that the del_chan is only invoked in pptp_release 
>> actually and it would wait a rcu period before sock_put.
>>
>
>Looking at the code again, the reader lookup_chan() is actually
>invoked in BH context, but neither add_chan() nor del_chan()
>actually disables BH...

Re: Re:Re: Re:Re: Re: [PATCH net] ppp: Fix a scheduling-while-atomic bug in del_chan

[Cong Wang]

On Wed, Aug 9, 2017 at 12:17 AM, Gao Feng  wrote:
> Hi Cong,
>
> Actually I have one question about the SOCK_RCU_FREE.
> I don't think it could resolve the issue you raised even though it exists 
> really.
>
> I checked the SOCK_RCU_FREE, it just defer the __sk_destruct after one rcu 
> period.
> But when it performs, someone still could find this sock by callid during the 
> del_chan period and it may still deference the sock which may freed soon.
>
> The right flow should be following.
> del_chan()
> wait a rcu period
> sock_put()  It is safe that someone gets the sock because it 
> already hold sock refcnt.
>
> When using SOCK_RCU_FREE, its flow would be following.
> wait a rcu period
> del_chan()
> free the sock directly  no sock refcnt check again.
> Because the del_chan happens after rcu wait, not before, so it isn't helpful 
> with SOCK_RCU_FREE.


Yes, good point! With SOCK_RCU_FREE the sock_hold() should
not be needed. For RCU, unpublish should indeed happen before
grace period.


>
> I don't know if I misunderstand the SOCK_RCU_FREE usage.
>
> But it is a good news that the del_chan is only invoked in pptp_release 
> actually and it would wait a rcu period before sock_put.
>

Looking at the code again, the reader lookup_chan() is actually
invoked in BH context, but neither add_chan() nor del_chan()
actually disables BH...

Re:Re: Re:Re: Re: [PATCH net] ppp: Fix a scheduling-while-atomic bug in del_chan

[Gao Feng]

At 2017-08-09 03:45:53, "Cong Wang"  wrote:
>On Mon, Aug 7, 2017 at 6:10 PM, Gao Feng  wrote:
>>
>> Sorry, I don't get you clearly. Why the sock_hold() isn't helpful?
>
>I already told you, the dereference happends before sock_hold().
>
>sock = rcu_dereference(callid_sock[call_id]);
>if (sock) {
>opt = >proto.pptp;
>if (opt->dst_addr.sin_addr.s_addr != s_addr) <=== HERE
>sock = NULL;
>else
>sock_hold(sk_pppox(sock));
>}
>
>If we don't wait for readers properly, sock could be freed at the
>same time when deference it.

Maybe I didn't show my explanation clearly.
I think it won't happen as I mentioned in the last email.
Because the pptp_release invokes the synchronize_rcu to make sure it, and 
actually there is no one which would invoke del_chan except pptp_release.
It is guaranteed by that the pptp_release doesn't put the sock refcnt until 
complete all cleanup include marking sk_state as PPPOX_DEAD. 

In other words, even though the pptp_release is not the last user of this sock, 
the other one wouldn't invoke del_chan in pptp_sock_destruct.
Because the condition "!(sk->sk_state & PPPOX_DEAD)" must be false. 

As summary, the del_chan and pppox_unbind_sock in pptp_sock_destruct are 
unnecessary. 
And it even brings confusing.

Best Regards
Feng

>
>> The pptp_release invokes synchronize_rcu after del_chan, it could make sure 
>> the others has increased the sock refcnt if necessary
>> and the lookup is over.
>> There is no one could get the sock after synchronize_rcu in pptp_release.
>
>
>If this were true, then this code in pptp_sock_destruct()
>would be unneeded:
>
>if (!(sk->sk_state & PPPOX_DEAD)) {
>del_chan(pppox_sk(sk));
>pppox_unbind_sock(sk);
>}
>
>
>>
>>
>> But I think about another problem.
>> It seems the pptp_sock_destruct should not invoke del_chan and 
>> pppox_unbind_sock.
>> Because when the sock refcnt is 0, the pptp_release must have be invoked 
>> already.
>>
>
>
>I don't know. Looks like sock_orphan() is only called
>in pptp_release(), but I am not sure if there is a case
>we call sock destructor before release.
>
>Also note, this socket is very special, it doesn't support
>poll(), sendmsg() or recvmsg()..