<snip>

> Hi Jon,
> 
> Frankly speaking, I more like to Erik's proposal than yours as Erik's 
> method is more common in networking subsystem.

Ok. I have no strong opinions on this, so I accept being outvoted.

[Ying] That's great!

Erik, please submit your patch to netdev with my acked-by and tested-by.
It sounds like that Jon also acknowledged your patch :)

> 
> Regarding the backtrace of the issue, it seems that we should not 
> deliver message without holding "owner" flag in BH. But we still don't 
> hold "owner" flag in your solution when sending message in socket timeout 
> function.

There is no need to, as we are finished with the socket from a sender's 
viewpoint at this moment. When the call comes back to tipc_sk_rcv(), it is no 
different than any other incoming BH call, i.e., the "owned" flag is set if 
there is a user context call in progress, and only then.

[Ying]Yes, you are right.  In recent days I also spent some time further 
analyzing that original stack backtrace provided by Guna. But unfortunately, I 
am still unable to understand what exact scenario caused that soft lockup. 

Now I would like to have to have some feedback from GUNA if this really solved 
the problem.
[Ying] Me too.

Regards,
Ying

///jon

> 
> Regards,
> Ying
> 
> -----Original Message-----
> From: Jon Maloy [mailto:[email protected]]
> Sent: 2016年6月2日 21:13
> To: Erik Hugne; Xue, Ying
> Cc: Richard Alpe; Parthasarathy Bhuvaragan; tipc- 
> [email protected]
> Subject: RE: [RFC PATCH] tipc: fix timer handling when socket is owned
> 
> 
> 
> From: Erik Hugne [mailto:[email protected]]
> Sent: Thursday, 02 June, 2016 14:11
> To: Ying Xue
> Cc: Richard Alpe; Parthasarathy Bhuvaragan; Jon Maloy; tipc- 
> [email protected]
> Subject: RE: [RFC PATCH] tipc: fix timer handling when socket is owned
> 
> 
> On Jun 2, 2016 1:03 PM, "Xue, Ying" <[email protected]> wrote:
> >>
> >> Acked-by: Ying Xue <[email protected]>
> >>
> >> Jon, whatever the patch can fix Guna's issue or not, I think the 
> >> change is right
> because there is an obvious error that we deliver message through
> tipc_node_xmit_skb() when "owner" flag is set.
> >> So, I suggest that the patch should be submitted to upstream as 
> >> soon as
> possible.
> >>
> >If you think the change is reasonable, could you please help me test it?
> >My crappy AMD A10 laptop does not handle VM's well..
> >I have one concern about it, locking policy have changed slightly 
> >since
> tipc_node_xmit_skb is now called with the socket spinlock held. And i 
> dont know if this introduces a new race..
> 
> Logically it should be ok, since it only means that the response 
> message always ends up in the backlog queue. But it is still an unnecessary 
> change.
> Personally I liked better the first proposal, which I think was simpler.
> 
> Why not
> 1) grab the lock and the mutex, if necessary after several attempts.
> 2) fetch all relevant socket fields to the stack.
> 3) release the lock and the mutex
> 4) create the message
> 5) send the message
> 
> This would give only one timeout function and minimal time spent 
> inside the lock/mutex.
> I  you do so you should also extend the wait period for retrial to 
> more e.g., 1 second until you retry. The risk of finding the socket 
> busy again after a longer period is lower.
> 
> ///jon
> 
> 
> >//E
------------------------------------------------------------------------------
What NetFlow Analyzer can do for you? Monitors network bandwidth and traffic
patterns at an interface-level. Reveals which users, apps, and protocols are 
consuming the most bandwidth. Provides multi-vendor support for NetFlow, 
J-Flow, sFlow and other flows. Make informed decisions using capacity 
planning reports. https://ad.doubleclick.net/ddm/clk/305295220;132659582;e
_______________________________________________
tipc-discussion mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/tipc-discussion

Reply via email to