Set an applier to a class. Then exit immapplier to detach the applier.
root@SC1:~# immapplier -a @whatever Test
Let another node join the cluster.
Create a CCB which is active on an object of 'Test' class. Don't commit the CCB.
root@SC1:~# immcfg > immcfg -c Test test=1 >
Try to set applier again.
root@SC1:/srv/shared# immapplier -a @whatever Test Implementer: @whatever ImmVersion: A 2 16 error - saImmOiImplementerSet FAILED: SA_AIS_ERR_TRY_AGAIN (6)
SC-1
osafimmnd [419:immsv_evt.c:5414] T8 Received: IMMND_EVT_D2ND_IMPLSET_RSP (60) from 0 osafimmnd [419:immnd_evt.c:9527] T2 originated here?:1 nodeId:2010f conn: 225 osafimmnd [419:ImmModel.cc:12967] >> implementerSet osafimmnd [419:ImmModel.cc:13008] T7 Re-using implementer for @whatever osafimmnd [419:ImmModel.cc:13040] TR TRY_AGAIN: ccb 2 is active on object 'test=1' bound to class applier '@whatever'. Can not re-attach applier osafimmnd [419:ImmModel.cc:13156] << implementerSet
PL-3
osafimmnd [392:immsv_evt.c:5414] T8 Received: IMMND_EVT_D2ND_IMPLSET_RSP (60) from 0 osafimmnd [392:immnd_evt.c:9527] T2 originated here?:0 nodeId:2010f conn: 225 osafimmnd [392:ImmModel.cc:12967] >> implementerSet osafimmnd [392:ImmModel.cc:13008] T7 Re-using implementer for @whatever osafimmnd [392:ImmModel.cc:13087] NO Implementer (applier) connected: 5 (@whatever) <0, 2010f> osafimmnd [392:ImmModel.cc:13156] << implementerSet
IMMND on SC-1 rejected the implSet request but IMMND on PL-3 accepted it.
The applier was not synced to PL-3 (mAppliers.empty() returned true) so the implSet request passed the ccb check.
if( ! obj->mClassInfo->mAppliers.empty()) { ImplementerSet::iterator ii = obj->mClassInfo->mAppliers.begin(); for(; ii != obj->mClassInfo->mAppliers.end(); ++ii) { if((*ii) == info) { TRACE("TRY_AGAIN: ccb %u is active on object '%s' " "bound to class applier '%s'. Can not re-attach applier", ccb->mId, omit->first.c_str(), implName.c_str()); err = SA_AIS_ERR_TRY_AGAIN; goto done; } } }
Now commit the CCB and try to set the applier again.
SC-1
osafimmnd [419:immsv_evt.c:5414] T8 Received: IMMND_EVT_D2ND_IMPLSET_RSP (60) from 0 osafimmnd [419:immnd_evt.c:9527] T2 originated here?:1 nodeId:2010f conn: 226 osafimmnd [419:ImmModel.cc:12967] >> implementerSet osafimmnd [419:ImmModel.cc:13008] T7 Re-using implementer for @whatever osafimmnd [419:ImmModel.cc:13087] NO Implementer (applier) connected: 6 (@whatever) <226, 2010f> osafimmnd [419:ImmModel.cc:13156] << implementerSet
PL-3
osafimmnd [392:immsv_evt.c:5414] T8 Received: IMMND_EVT_D2ND_IMPLSET_RSP (60) from 0 osafimmnd [392:immnd_evt.c:9527] T2 originated here?:0 nodeId:2010f conn: 226 osafimmnd [392:ImmModel.cc:12967] >> implementerSet osafimmnd [392:ImmModel.cc:13003] T7 ERR_EXIST: Registered implementer already exists: @whatever osafimmnd [392:ImmModel.cc:13005] << implementerSet
The applier had diferent ids on SC-1 and PL-3. When a new node joins the cluster, IMMND on PL-3 will crash when verifying the implementers.
PL3 osafimmnd[392]: ER Sync-verify: Established node has different Implementer-id: 5 for name: @whatever, sync says 6.
Not only class appliers, but also object appliers also have this problem.
sObjAppliersMap is not built up on sync-client.
I have a problem with this ticket.
Appliers are intentionally not synced.
They should not need to be synced.
The question here is how you manage to execute a sync with a ccb being active.
Non empty Ccbs are terminated before the actual sync can start.
So there seems to have been introduced a bug somewhere.
The CCB was created after the sync.
The applier-names are synced but the class/object-applier data is not sync-ed.
That is intentional and I dont want a solution that tries to sync all applier information to all nodes.
The class-applier and object-applier mechanism is inherrently local, i.e. only used at the node where
the applier exists. Remeber that an applier is a listener and not a true particiapnt in CCbs, so
its existence should only matter locally. The only thing thatg is global is the existence of an
applier with a certain name and the current location if any for that exact applier with that name.
Having said that, it is still important that the local class/object applier is not allowed to attach in such
a way that it can see an incomplete ccb.
Iam thinking about what the best approach foir a fix would be.
Dont start doing some complex implementation of this yet.
The problem is the feature of implicit class-implementer-set and implicit object-implementer-set.
Ironically this feature is parctically useless for appliers.
One possible (and relatively simple) solution would be to only do the ccb interference checks for
appliers at the node where the applier is actually attaching. That would almost be in
fevs_local_checks, except that implementer-set is not a regular fevs message at the sending side.
So instad it would be in immnd_evt_proc_impl_set in immnd_evt.c.
If the check fails then the local IMMND simply rejects the request with TRY_AGAIN (or ERR_BUSY
would in reality be better here since the immsv has no control over how long the wait will be).
The current applier check at the fevs receiving side for implementer-set is simply removed.
With the above solution there is the issue that the check is then not done in fevs order.
By the time the implementer-set arrives over fevs at all nodes, there may have been creaed a
ccb-operation that interferes, resulting in the implementer-set having to be aborted anyway.
The local immnd thus has to run the applier checks again in the receiving fevs for implementer-set.
If that check fails, it rejects the operation, replies with error to the client and broadcast an implementer_clear over fevs.
The actual local check done for the locall attaching applier is correct. The problem is that the same check may fail at other nodes (where the applier is not currently attaching) resulting in the
applier attach being handled as rejected at some other node. This will not impact the correct
behavior of the applier mechanism where it attached. That logic is purely node local.
A detrimental effect that can happen is that a future sync may detect an inconstency on
applier-id for a given applier name and escalate this to a restart of the node director(s) that detects this inconsistency. That is a serious side effect so the problem needs to be fixed.
One quick solution would be to disarm the check, or at least reduce its severity to a warning.
That would of course not remove the inconsistency.
Worst case scenario resulting from he inconsistency on applier existence/identity is that two appliers could be allowed to attach under the same applier name, at two or more different nodes.
That in itself would not be a problem for the applier mechanism. A problem would only ocur
if a third party tries to communicate with the supposed unique named applier using a direct
admin-operation with the applier-name as the target. Such an admin-operation would end up
arbitrarily reaching one of the appliers under that name.
https://sourceforge.net/p/opensaf/mailman/message/34518910/
https://sourceforge.net/p/opensaf/mailman/message/34522118/
opensaf-4.5.x:
changeset: 7001:32075f5c5570
branch: opensaf-4.5.x
parent: 6998:bebc2783183f
user: Zoran Milinkovic zoran.milinkovic@ericsson.com
date: Tue Oct 13 10:56:50 2015 +0200
summary: imm: synchronize applier set on all nodes [#1504]
opensaf-4.6.x:
changeset: 7002:fa5da6b01f61
branch: opensaf-4.6.x
tag: tip
parent: 6997:ae65b0ffa596
user: Zoran Milinkovic zoran.milinkovic@ericsson.com
date: Tue Oct 13 11:00:39 2015 +0200
summary: imm: synchronize applier set on all nodes [#1504]
opensaf-4.7.x:
changeset: 6999:2767588ed092
branch: opensaf-4.7.x
parent: 6996:e5bb3f7120eb
user: Zoran Milinkovic zoran.milinkovic@ericsson.com
date: Wed Oct 07 16:35:59 2015 +0200
summary: imm: synchronize applier set on all nodes [#1504]
default(4.7):
changeset: 7000:9d30fa46a7a5
parent: 6995:fe634f270a98
user: Zoran Milinkovic zoran.milinkovic@ericsson.com
date: Wed Oct 07 16:35:59 2015 +0200
summary: imm: synchronize applier set on all nodes [#1504]
Related
Tickets:
#1504