[Gluster-Maintainers] [Gluster-devel] Another regression in release-3.7 and master

[Kaushal M]

On Thu, Apr 7, 2016 at 9:06 PM, Kaushal M <kshlmster at gmail.com> wrote:
> On Thu, Apr 7, 2016 at 7:24 PM, Kaushal M <kshlmster at gmail.com> wrote:
>> On Thu, Apr 7, 2016 at 6:23 PM, Kaushal M <kshlmster at gmail.com> wrote:
>>> On Thu, Apr 7, 2016 at 6:00 PM, Atin Mukherjee <amukherj at redhat.com> wrote:
>>>>
>>>>
>>>> On 04/07/2016 05:37 PM, Kaushal M wrote:
>>>>>
>>>>> On 7 Apr 2016 5:36 p.m., "Niels de Vos" <ndevos at redhat.com
>>>>> <mailto:ndevos at redhat.com>> wrote:
>>>>>>
>>>>>> On Thu, Apr 07, 2016 at 05:13:54PM +0530, Kaushal M wrote:
>>>>>> > On Thu, Apr 7, 2016 at 5:11 PM, Kaushal M <kshlmster at gmail.com
>>>>> <mailto:kshlmster at gmail.com>> wrote:
>>>>>> > > We've hit another regression.
>>>>>> > >
>>>>>> > > With management encryption enabled, daemons like NFS and SHD don't
>>>>>> > > start on the current heads of release-3.7 and master branches.
>>>>>> > >
>>>>>> > > I still have no clear root cause for it, and would appreciate some
>>>>> help.
>>>>>> >
>>>>>> > This was working with 3.7.9 from what I've heard.
>>>>>>
>>>>>> Do we have a simple test-case for this? If someone write a script, we
>>>>>> should be able to "git bisect" it pretty quickly.
>>>>>
>>>>> I am doing this right now.
>>>> "b33f3c9 glusterd: Bug fixes for IPv6 support" has caused this
>>>> regression. I am yet to find the RCA though.
>>>
>>> git-bisect agrees with this as well.
>>>
>>> I initially thought it was because GlusterD didn't listen on IPv6
>>> (checked using `ss`).
>>> This change makes it so that connections to localhost use ::1 instead
>>> of 127.0.0.1, and so the connection failed.
>>> This should have caused all connection attempts to fail, irrespective
>>> of it being encrypted or not.
>>> But the failure only happens when management encryption is enabled.
>>> So this theory doesn't make sense.
>>
>> This is the part of the problem!
>>
>> The initial IPv6 connection to ::1 fails for non encrypted connections as well.
>> But these connections correctly retry connect with the next address
>> once the first connect attempt fails.
>> Since the next address is 127.0.0.1, the connection succeeds, volfile
>> is fetched and the daemon starts.
>>
>> Encrypted connections on the other hand, give up after the first
>> failure and don't attempt a reconnect.
>> This is somewhat surprising to me, as I'd recently fixed an issue
>> which caused crashes when encrypted connections attempted a reconnect
>> after a failure to connect.
>>
>> I'll diagnose this a little bit more and try to find a solution.
>
> Found the full problem. This is mainly a result of the fix I did, that
> I mentioned above.
> (A slight correction is that actually it wasn't crashes that it fixed,
> but a encrypted reconnect issue in GlusterD).
>
> I'm posting the root-cause as I described in the commit message for
> the fix for this.
> """
> With commit d117466 socket_poller() wasn't launched from
> socket_connect
> (for encrypted connections), if connect() failed. This was done to
> prevent the socket private data from being double unreffed, from the
> cleanups in both socket_poller() and socket_connect(). This allowed
> future reconnects to happen successfully.
>
> If a socket reconnects is sort of decided by the rpc notify function
> registered. The above change worked with glusterd, as the glusterd rpc
> notify function (glusterd_peer_rpc_notify()) continuously allowed
> reconnects on failure.
>
> mgmt_rpc_notify(), the rpc notify function in glusterfsd, behaves
> differently.
>
> For a DISCONNECT event, if more volfile servers are available or if
> more
> addresses are available in the dns cache, it allows reconnects. If not
> it terminates the program.
>
> For a CONNECT event, it attempts to do a volfile fetch rpc request. If
> sending this rpc fails, it immediately terminates the program.
>
> One side effect of commit d117466, was that the encrypted socket was
> registered with epoll, unintentionally, on a connect failure.  A weird
> thing happens because of this. The epoll notifier notifies
> mgmt_rpc_notify() of a CONNECT event, instead of a DISCONNECT as
> expected. This causes mgmt_rpc_notify() to attempt an unsuccessful
> volfile fetch rpc request, and terminate.
> (I still don't know why the epoll raises the CONNECT event)
>
> Commit 46bd29e fixed some issues with IPv6 in GlusterFS. This caused
> address resolution in GlusterFS to also request of IPv6 addresses
> (AF_UNSPEC) instead of just IPv4. On most systems, this causes the
> IPv6
> addresses to be returned first.
>
> GlusterD listens on 0.0.0.0:24007 by default. While this attaches to
> all
> interfaces, it only listens on IPv4 addresses. GlusterFS daemons and
> bricks are given 'localhost' as the volfile server. This resolves to
> '::1' as the first address.
>
> When using management encryption, the above reasons cause the daemon
> processes to fail to fetch volfiles and terminate.
>
> The solution to this is simple. Instead of cleaning up the encrypted
> socket in socket_connect(), launch socket_poller() and let it cleanup
> the socket instead. This prevents the unintentional registration with
> epoll, and socket_poller() sends the correct events to the rpc notify
> functions, which allows proper reconnects to happen.
> """
>
> I'll post the commit to r.g.o later, after doing a little more testing
> to verify it, but for now I'm off to have dinner.

Decided to post the change as an RFC review. I still need to open bug
reports for this.
https://review.gluster.org/13926

>
> ~kaushal
>
>>
>>>
>>> One other thing was on my laptop, even bricks failed to start when
>>> glusterd was started with management encryption.
>>> But on a VM, the bricks started, but other daemons failed.
>>>
>>>>>
>>>>>>
>>>>>> Niels
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> Gluster-devel mailing list
>>>>> Gluster-devel at gluster.org
>>>>> http://www.gluster.org/mailman/listinfo/gluster-devel
>>>>>