Hi Steven,

Thank you for comment.

> > Is your patch 2 of the next?
> > 
> >  * [Openais] [PATCH] When a failed to recv state happens,stop forwarding 
> > the token
> > 
> >  * [Openais] [PATCH] When a failed to recv state happens,stop forwarding 
> > the token(take 2)
> > 
> > Best Regards,
> > Hideo Yamauchi.
> > 
> 
> Yes the take 2 version.

All right.
Thanks!

Hideo Yamauch.

--- Steven Dake <[email protected]> wrote:

> On 02/08/2011 05:54 PM, [email protected] wrote:
> > Hi Steven,
> > 
> >> Have a try of the patch i have sent to this ml.  If the issue persists,
> >> we can look at more options.
> > 
> > Thank you for comment.
> > 
> > Is your patch 2 of the next?
> > 
> >  * [Openais] [PATCH] When a failed to recv state happens,stop forwarding 
> > the token
> > 
> >  * [Openais] [PATCH] When a failed to recv state happens,stop forwarding 
> > the token(take 2)
> > 
> > Best Regards,
> > Hideo Yamauchi.
> > 
> 
> Yes the take 2 version.
> 
> Regards
> -steve
> 
> 
> > 
> > 
> > --- Steven Dake <[email protected]> wrote:
> > 
> >> On 02/07/2011 11:11 PM, [email protected] wrote:
> >>> Hi Steven,
> >>>
> >>> I understood your opinion by mistake.
> >>>
> >>> We do not have simple test case.
> >>>
> >>> The phenomenon generated in our environment is the following thing.
> >>>
> >>> Step 1) corosync constitutes a cluster in 12 nodes.
> >>>  * begin communication in TOKEN
> >>>
> >>> Step 2) One node raises [FAILED TO RECEIVE].
> >>>
> >>> Step 3) 12 nodes begin the reconfiguration of the cluster again.
> >>>
> >>> Step 4) The node that occurred fails([FAILED TO RECEIVE]) in an consensus 
> >>> of the JOIN
> >> communication.
> >>>  * Because the node failed in an consensus, node make contents of 
> >>> faildlist and proclist
> same.
> >>>  * And this node compares faildlist with proclist and assert-fail 
> >>> happened.
> >>>
> >>>
> >>> When the node that made a cluster stood alone, I think that assert() is 
> >>> unnecessary.
> >>>
> >>> Because the reason is because there is the next processing.
> >>>
> >>>
> >>
> >> Have a try of the patch i have sent to this ml.  If the issue persists,
> >> we can look at more options.
> >>
> >> Thanks!
> >> -steve
> >>
> >>
> >>>
> >>> static void memb_join_process (
> >>>   struct totemsrp_instance *instance,
> >>>   const struct memb_join *memb_join)
> >>> {
> >>>   struct srp_addr *proc_list;
> >>>   struct srp_addr *failed_list;
> >>> (snip)
> >>>                           instance->failed_to_recv = 0;
> >>>                           srp_addr_copy (&instance->my_proc_list[0],
> >>>                                   &instance->my_id);
> >>>                           instance->my_proc_list_entries = 1;
> >>>                           instance->my_failed_list_entries = 0;
> >>>
> >>>                           memb_state_commit_token_create (instance);
> >>>
> >>>                           memb_state_commit_enter (instance);
> >>>                           return;
> >>>
> >>> (snip)
> >>>
> >>> Best Regards,
> >>> Hideo Yamauchi.
> >>>
> >>>
> >>>
> >>> --- [email protected] wrote:
> >>>
> >>>> Hi Steven,
> >>>>
> >>>>> Hideo,
> >>>>>
> >>>>> If you have a test case, I can make a patch for you to try.
> >>>>>
> >>>>
> >>>> All right.
> >>>>
> >>>> We use corosync.1.3.0.
> >>>>
> >>>> Please send me patch.
> >>>>
> >>>> Best Regards,
> >>>> Hideo Yamauchi.
> >>>>
> >>>> --- Steven Dake <[email protected]> wrote:
> >>>>
> >>>>> On 02/06/2011 09:16 PM, [email protected] wrote:
> >>>>>> Hi Steven,
> >>>>>> Hi Dejan,
> >>>>>>
> >>>>>>>>>> This code never got a chance to run because on failed_to_recv
> >>>>>>>>>> the two sets (my_process_list and my_failed_list) are equal which
> >>>>>>>>>> makes the assert fail in memb_consensus_agreed():
> >>>>>>
> >>>>>> The same problem occurs, and we are troubled, too. 
> >>>>>>
> >>>>>> How did this argument turn out?
> >>>>>>
> >>>>>> Best Regards,
> >>>>>> Hideo Yamauchi.
> >>>>>>
> >>>>>
> >>>>> Hideo,
> >>>>>
> >>>>> If you have a test case, I can make a patch for you to try.
> >>>>>
> >>>>> Regards
> >>>>> -steve
> >>>>>
> >>>>>>
> >>>>>> --- Dejan Muhamedagic <[email protected]> wrote:
> >>>>>>
> >>>>>>> nudge, nudge
> >>>>>>>
> >>>>>>> On Wed, Jan 05, 2011 at 02:05:55PM +0100, Dejan Muhamedagic wrote:
> >>>>>>>> On Tue, Jan 04, 2011 at 01:53:00PM -0700, Steven Dake wrote:
> >>>>>>>>> On 12/23/2010 06:14 AM, Dejan Muhamedagic wrote:
> >>>>>>>>>> Hi,
> >>>>>>>>>>
> >>>>>>>>>> On Wed, Dec 01, 2010 at 05:30:44PM +0200, Vladislav Bogdanov wrote:
> >>>>>>>>>>> 01.12.2010 16:32, Dejan Muhamedagic wrote:
> >>>>>>>>>>>> Hi,
> >>>>>>>>>>>>
> >>>>>>>>>>>> On Tue, Nov 23, 2010 at 12:53:42PM +0200, Vladislav Bogdanov 
> >>>>>>>>>>>> wrote:
> >>>>>>>>>>>>> Hi Steven, hi all.
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> I often see this assert on one of nodes after I stop corosync 
> >>>>>>>>>>>>> on some
> >>>>>>>>>>>>> another node in newly-setup 4-node cluster.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Does the assert happen on a node lost event? Or once new
> >>>>>>>>>>>> partition is formed?
> >>>>>>>>>>>
> >>>>>>>>>>> I first noticed it when I rebooted another node, just after 
> >>>>>>>>>>> console said
> >>>>>>>>>>> that OpenAIS is stopped.
> >>>>>>>>>>>
> >>>>>>>>>>> Can't say right now, what exactly event did it follow, I'm 
> >>>>>>>>>>> actually
> >>>>>>>>>>> fighting with several problems with corosync, pacemaker, NFS4 and
> >>>>>>>>>>> phantom uncorrectable ECC errors simultaneously and I'm a bit 
> >>>>>>>>>>> lost with
> >>>>>>>>>>> all of them.
> >>>>>>>>>>>
> >>>>>>>>>>>>
> >>>>>>>>>>>>> #0  0x00007f51953e49a5 in raise () from /lib64/libc.so.6
> >>>>>>>>>>>>> #1  0x00007f51953e6185 in abort () from /lib64/libc.so.6
> >>>>>>>>>>>>> #2  0x00007f51953dd935 in __assert_fail () from /lib64/libc.so.6
> >>>>>>>>>>>>> #3  0x00007f5196176406 in memb_consensus_agreed
> >>>>>>>>>>>>> (instance=0x7f5196554010) at totemsrp.c:1194
> >>>>>>>>>>>>> #4  0x00007f519617b2f3 in memb_join_process 
> >>>>>>>>>>>>> (instance=0x7f5196554010,
> >>>>>>>>>>>>> memb_join=0x262f628) at totemsrp.c:3918
> >>>>>>>>>>>>> #5  0x00007f519617b619 in message_handler_memb_join
> >>>>>>>>>>>>> (instance=0x7f5196554010, msg=<value optimized out>, 
> >>>>>>>>>>>>> msg_len=<value
> >>>>>>>>>>>>> optimized out>, endian_conversion_needed=<value optimized out>)
> >>>>>>>>>>>>>     at totemsrp.c:4161
> >>>>>>>>>>>>> #6  0x00007f5196173ba7 in passive_mcast_recv 
> >>>>>>>>>>>>> (rrp_instance=0x2603030,
> >>>>>>>>>>>>> iface_no=0, context=<value optimized out>, msg=<value optimized 
> >>>>>>>>>>>>> out>,
> >>>>>>>>>>>>> msg_len=<value optimized out>) at totemrrp.c:720
> >>>>>>>>>>>>> #7  0x00007f5196172b44 in rrp_deliver_fn (context=<value 
> >>>>>>>>>>>>> optimized out>,
> >>>>>>>>>>>>> msg=0x262f628, msg_len=420) at totemrrp.c:1404
> >>>>>>>>>>>>> #8  0x00007f5196171a76 in net_deliver_fn (handle=<value 
> >>>>>>>>>>>>> optimized out>,
> >>>>>>>>>>>>> fd=<value optimized out>, revents=<value optimized out>, 
> >>>>>>>>>>>>> data=0x262ef80)
> >>>>>>>>>>>>> at totemudp.c:1244
> >>>>>>>>>>>>> #9  0x00007f519616d7f2 in poll_run (handle=4858364909567606784) 
> >>>>>>>>>>>>> at
> >>>>>>>>>>>>> coropoll.c:510
> >>>>>>>>>>>>> #10 0x0000000000406add in main (argc=<value optimized out>, 
> >>>>>>>>>>>>> argv=<value
> >>>>>>>>>>>>> optimized out>, envp=<value optimized out>) at main.c:1680
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> Last fplay lines are:
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> rec=[36124] Log Message=Delivering MCAST message with seq 1366 
> >>>>>>>>>>>>> to
> >>>>>>>>>>>>> pending delivery queue
> >>>>>>>>>>>>> rec=[36125] Log Message=Delivering MCAST message with seq 1367 
> >>>>>>>>>>>>> to
> >>>>>>>>>>>>> pending delivery queue
> >>>>>>>>>>>>> rec=[36126] Log Message=Received ringid(10.5.4.52:12660) seq 
> >>>>>>>>>>>>> 1366
> >>>>>>>>>>>>> rec=[36127] Log Message=Received ringid(10.5.4.52:12660) seq 
> >>>>>>>>>>>>> 1367
> >>>>>>>>>>>>> rec=[36128] Log Message=Received ringid(10.5.4.52:12660) seq 
> >>>>>>>>>>>>> 1366
> >>>>>>>>>>>>> rec=[36129] Log Message=Received ringid(10.5.4.52:12660) seq 
> >>>>>>>>>>>>> 1367
> >>>>>>>>>>>>> rec=[36130] Log Message=releasing messages up to and including 
> >>>>>>>>>>>>> 1367
> >>>>>>>>>>>>> rec=[36131] Log Message=FAILED TO RECEIVE
> >>>>>>>>>>>>> rec=[36132] Log Message=entering GATHER state from 6.
> >>>>>>>>>>>>> rec=[36133] Log Message=entering GATHER state from 0.
> >>>>>>>>>>>>> Finishing replay: records found [33993]
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> What could be the reason for this? Bug, switches, memory errors?
> >>>>>>>>>>>>
> >>>>>>>>>>>> The assertion fails because corosync finds out that
> >>>>>>>>>>>> instance->my_proc_list and instance->my_failed_list are
> >>>>>>>>>>>> equal. That happens immediately after the "FAILED TO RECEIVE"
> >>>>>>>>>>>> message which is issued when fail_recv_const token rotations
> >>>>>>>>>>>> happened without any multicast packet received (defaults to 50).
> >>>>>>>>>>
> >>>>>>>>>> I took a look at the code and the protocol specification again
> >>>>>>>>>> and it seems like that assert is not valid since Steve patched
> >>>>>>>>>> the part dealing with the "FAILED TO RECEIVE" condition. The
> >>>>>>>>>> patch is from 2010-06-03 posted to the list here
> >>>>>>>>>> http://marc.info/?l=openais&m=127559807608484&w=2
> >>>>>>>>>>
> >>>>>>>>>> The last hunk of the patch contains this code (exec/totemsrp.c):
> >>>>>>>>>>
> >>>>>>>>>> 3933         if (memb_consensus_agreed (instance) && 
> >>>>>>>>>> instance->failed_to_recv == 1) {
>  
> >>
> >>>>  
> >>>>>>>
> >>>>>>>>>> 3934                 instance->failed_to_recv = 0;
> >>>>>>>>>> 3935                 srp_addr_copy (&instance->my_proc_list[0],
> >>>>>>>>>> 3936                     &instance->my_id);
> >>>>>>>>>> 3937                 instance->my_proc_list_entries = 1;
> >>>>>>>>>> 3938                 instance->my_failed_list_entries = 0;
> >>>>>>>>>> 3939            
> >>>>>>>>>> 3940                 memb_state_commit_token_create (instance);
> >>>>>>>>>> 3941            
> >>>>>>>>>> 3942                 memb_state_commit_enter (instance);
> >>>>>>>>>> 3943                 return;
> >>>>>>>>>> 3944         }
> >>>>>>>>>>
> >>>>>>>>>> This code never got a chance to run because on failed_to_recv
> >>>>>>>>>> the two sets (my_process_list and my_failed_list) are equal which
> >>>>>>>>>> makes the assert fail in memb_consensus_agreed():
> >>>>>>>>>>
> >>>>>>>>>> 1185     memb_set_subtract (token_memb, &token_memb_entries,
> >>>>>>>>>> 1186         instance->my_proc_list, 
> >>>>>>>>>> instance->my_proc_list_entries,
> >>>>>>>>>> 1187         instance->my_failed_list, 
> >>>>>>>>>> instance->my_failed_list_entries);
> >>>>>>>>>> ...
> >>>>>>>>>> 1195     assert (token_memb_entries >= 1);
> >>>>>>>>>>
> >>>>>>>>>> In other words, it's something like this:
> >>>>>>>>>>
> >>>>>>>>>>    if A:
> >>>>>>>>>>            if memb_consensus_agreed() and failed_to_recv:
> >>>>>>>>>>                    form a single node ring and try to recover
> >>>>>>>>>>
> >>>>>>>>>>    memb_consensus_agreed():
> >>>>>>>>>>            assert(!A)
> >>>>>>>>>>
> >>>>>>>>>> Steve, can you take a look and confirm that this holds.
> >>>>>>>>>>
> >>>>>>>>>> Cheers,
> >>>>>>>>>>
> >>>>>>>>>
> >>>>>>>>> Dejan,
> >>>>>>>>>
> >>>>>>>>> sorry for delay in response - big backlog which is mostly cleared 
> >>>>>>>>> out :)
> >>>>>>>>
> >>>>>>>> No problem.
> >>>>>>>>
> >>>>>>>>> The assert definitely isn't correct, but removing it without 
> >>>>>>>>> addressing
> >>>>>>>>> the contents of the proc and fail lists is also not right.  That 
> >>>>>>>>> would
> >>>>>>>>> cause the logic in the if statement at line 3933 not to be executed
> >>>>>>>>> (because the first part of the if would evaluate to false)
> >>>>>>>>
> >>>>>>>> Actually it wouldn't. The agreed variable is set to 1 and it
> >>>>>>>> is going to be returned unchanged.
> >>>>>>>>
> >>>>>>>>> I believe
> >>>>>>>>> what we should do is check the "failed_to_recv" value in
> >>>>>>>>> memb_consensus_agreed instead of at line 3933.
> >>>>>>>>>
> >>>>>>>>> The issue with this is memb_state_consensus_timeout_expired which 
> >>>>>>>>> also
> >>>>>>>>> executes some 'then' logic where we may not want to execute the
> >>>>>>>>> failed_to_recv logic.
> >>>>>>>>
> >>>>>>>> Perhaps we should just
> >>>>>>>>
> >>>>>>>> 3933         if (instance->failed_to_recv == 1) {
> >>>>>>>>
> >>>>>>>> ? In case failed_to_recv both proc and fail lists are equal so
> >>>>>>>> checking for memb_consensus_agreed won't make sense, right?
> >>>>>>>>
> >>>>>>>>> If anyone has a reliable reproducer and can forward to me, I'll 
> >>>>>>>>> test out
> >>>>>>>>> a change to address this problem.  Really hesitant to change 
> >>>>>>>>> anything in
> >>>>>>>>> totemsrp without a test case for this problem - its almost perfect 
> >>>>>>>>> ;-)
> >>>>>>>>
> >>>>>>>> Since the tester upgraded the switch firmware they couldn't
> >>>>>>>> reproduce it anymore.
> >>>>>>>>
> >>>>>>>> Would compiling with these help?
> >>>>>>>>
> >>>>>>>> /*
> >>>>>>>>  * These can be used to test the error recovery algorithms
> >>>>>>>>  * #define TEST_DROP_ORF_TOKEN_PERCENTAGE 30
> >>>>>>>>  * #define TEST_DROP_COMMIT_TOKEN_PERCENTAGE 30
> >>>>>>>>  * #define TEST_DROP_MCAST_PERCENTAGE 50
> >>>>>>>>  * #define TEST_RECOVERY_MSG_COUNT 300
> >>>>>>>>  */
> >>>>>>>>
> >>>>>>>> Cheers,
> >>>>>>>>
> >>>>>>>> Dejan
> >>>>>>>>
> >>>>>>>>> Regards
> >>>>>>>>> -steve
> >>>>>>>>>
> >>>>>>>>>> Dejan
> >>>>>>>>>> _______________________________________________
> >>>>>>>>>> Openais mailing list
> >>>>>>>>>> [email protected]
> >>>>>>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>>>>>>
> >>>>>>>>> _______________________________________________
> >>>>>>>>> Openais mailing list
> >>>>>>>>> [email protected]
> >>>>>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>>>>> _______________________________________________
> >>>>>>>> Openais mailing list
> >>>>>>>> [email protected]
> >>>>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>>>> _______________________________________________
> >>>>>>> Openais mailing list
> >>>>>>> [email protected]
> >>>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>>>>
> >>>>>>
> >>>>>> _______________________________________________
> >>>>>> Openais mailing list
> >>>>>> [email protected]
> >>>>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>>
> >>>>>
> >>>>
> >>>> _______________________________________________
> >>>> Openais mailing list
> >>>> [email protected]
> >>>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>>>
> >>>
> >>> _______________________________________________
> >>> Openais mailing list
> >>> [email protected]
> >>> https://lists.linux-foundation.org/mailman/listinfo/openais
> >>
> >>
> > 
> > _______________________________________________
> > Openais mailing list
> > [email protected]
> > https://lists.linux-foundation.org/mailman/listinfo/openais
> 
> 

_______________________________________________
Openais mailing list
[email protected]
https://lists.linux-foundation.org/mailman/listinfo/openais

Reply via email to