Re: [PATCH FIX for-3.19] IB/ipoib: Fix failed multicast joins/sends

[Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 4827 bytes --]

On Wed, 2015-01-14 at 20:47 +0200, Erez Shitrit wrote:
> On 1/14/2015 6:09 PM, Doug Ledford wrote:
> > On Wed, 2015-01-14 at 18:02 +0200, Erez Shitrit wrote:
> >> Hi Doug,
> >>
> >> Perhaps I am missing something here, but ping6 still doesn't work for me
> >> in many cases.
> >>
> >> I think the reason is that your origin patch does the following:
> >> in function ipoib_mcast_join_task
> >>           if (test_bit(IPOIB_MCAST_FLAG_SENDONLY, &mcast->flags))
> >>               ipoib_mcast_sendonly_join(mcast);
> >>           else
> >>               ipoib_mcast_join(dev, mcast, 1);
> >>           return;
> >> The flow for sendonly_join doesn't include handling the mc_task, so only
> >> the first mc in the list (if it is sendonly mcg) will be sent, and no
> >> more mcg's that are in the ipoib mc list are going to be sent. (see how
> >> it is in ipoib_mcast_join flow)
> > Yes, I know what you are talking about.  However, my patches did not add
> > this bug, it was present in the original code.  Please check a plain
> > v3.18 kernel, which does not have my patches, and you will see that
> > ipoib_mcast_sendonly_join_complete also fails to restart the mcast join
> > thread there as well.
> Agree.
> but in 3.18 there was no call from mc_task to sendonly_join, just to the 
> full-member join, so no need at that point to handle the task. (the call 
> for sendonly-join was by demand whenever new packet to mcg was sent by 
> the kernel)
> only in 3.19 the sendonly join was called explicitly from the mc_task.

I just sent a patch set that fixes this.

> >
> >> I can demonstrate it with the log of ipoib:
> >> I am trying to ping6 fe80::202:c903:9f:3b0a via ib0
> >>
> >> The log is:
> >> ib0: restarting multicast task
> >> ib0: setting up send only multicast group for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016
> >> ib0: adding multicast entry for mgid ff12:601b:ffff:0000:0000:0001:ff43:3bf1
> >> ib0: no multicast record for ff12:601b:ffff:0000:0000:0000:0000:0016,
> >> starting sendonly join
> >> ib0: join completion for ff12:601b:ffff:0000:0000:0000:0000:0001 (status 0)
> >> ib0: MGID ff12:601b:ffff:0000:0000:0000:0000:0001 AV ffff88081afb5f40,
> >> LID 0xc015, SL 0
> >> ib0: join completion for ff12:401b:ffff:0000:0000:0000:0000:0001 (status 0)
> >> ib0: MGID ff12:401b:ffff:0000:0000:0000:0000:0001 AV ffff88081e1c42c0,
> >> LID 0xc014, SL 0
> >> ib0: sendonly multicast join failed for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016, status -22
> >> ib0: no multicast record for ff12:601b:ffff:0000:0000:0000:0000:0016,
> >> starting sendonly join
> >> ib0: sendonly multicast join failed for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016, status -22
> >> ib0: no multicast record for ff12:601b:ffff:0000:0000:0000:0000:0016,
> >> starting sendonly join
> >> ib0: sendonly multicast join failed for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016, status -22
> >> ib0: setting up send only multicast group for
> >> ff12:601b:ffff:0000:0000:0000:0000:0002
> >> ib0: no multicast record for ff12:601b:ffff:0000:0000:0000:0000:0016,
> >> starting sendonly join
> >> ib0: sendonly multicast join failed for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016, status -22
> >> ib0: setting up send only multicast group for
> >> ff12:601b:ffff:0000:0000:0001:ff9f:3b0a
> >>       >>>>>> here you can see that the ipv6 address is added and queued
> >> to the list
> >> ib0: no multicast record for ff12:601b:ffff:0000:0000:0000:0000:0016,
> >> starting sendonly join
> >> ib0: sendonly multicast join failed for
> >> ff12:601b:ffff:0000:0000:0000:0000:0016, status -22
> >>       >>>>>> the ipv6 mcg will not be sent because it is after some other
> >> sendonly, and no one in that flow re-queue the mc_task again.
> > This is a problem with the design of the original mcast task thread.
> > I'm looking at a fix now.  Currently the design only allows one join to
> > be outstanding at a time.  Is there a reason for that that I'm not aware
> > of?  Some historical context that I don't know about?
> IMHO, the reason for  only one mc on the air at a time was to make our 
> life easier, otherwise there are locks to take/manage, races between few 
> responses, etc. also, the multicast module in the core keeps all the 
> requests in serialize mode.
> perhaps, you can use the relevant code from the full-member join in the 
> sendonly joinin order to handle the mc_task, or to return the call to 
> send-only to the mcast_send instead of the mc_task.

I reworked things a bit, but yes, the send only task now does the right
thing.  Please review the latest patchset I posted.  It's working just
fine for me here.

-- 
Doug Ledford <dledford-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>
              GPG KeyID: 0E572FDD



[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 819 bytes --]