On 13-May-14 6:38 PM, minhchau wrote:
> Hi Praveen,
>
> My patch has problem about the way npi component is retrieved, it
> should be retrieved from csi as currently.
> I just load your patch and test it. I can see components are incurred
> many extra restarts than expected.
This will happen because of previous recovery policy "component restart"
a component is restarted. But due to surestart policy every component
restarted again and
some components may get restarted again and again. But a component needs
to be restarted in the context of surestart atleast once even if it has
been already restarted in component restart recovery. The reason is
there may be dependency among components.
So we are getting all the components instantiated at the expense of
multiple restart of some components and this is due to parallel execution
of recovery policies component restart and surestart. ideally the fix
should differentiate if the component is restarted in the context of
component restart policy
or surestart policy. In this way when cleanup of component because of
comp restart recovery succeeds and AMFND sees that already surestart is
going on then
it should not instantiate this component again as it will be taken care
in surestart. At the same time Surestart should use this successful
cleanup to just instantiate the component instead of doing further
restart (cleanup + instantiated) because in both the cases cleanup will
be done same way.
I will also see if this can be achieved.
> Let's look at this case:
> kill comp #15 #10 #8 #2 -> componentRestart
> kill comp #5 -> suRestart
> If restart comp #2 finishes, it will continue restart #1
> If restart comp #5 finishes, it will restart the #4, and then #3, then
> #2 (again) because presence state of comp #2 is INSTANTIATED now
> Similarly, the other restart iterations:
> #15 -> #14 -> #13 -> #12 -> #11 -> #10 -> #9 ... -> #1
> #10 -> #9 -> #8 -> ... -> #1
> #8 -> #7 ... -> #1
> So many components have to be restarted many times because it is at
> the end of restart iteration which originally is initiated from other
> components.
>
> I probably keep my approach which marks the comp FAILED (or any better
> flag?) (in order that it restarts only one time per component), revise
> the way of looping through npi components and send it out again.
>
I could not get much benefit of marking all the components failed.
Please add some comment/explanation for making all components failed
when patch will be floated again.
Thanks,
Praveen
> Thanks/
> Minh
> On 5/13/2014 7:34 PM, praveen malviya wrote:
>> Please see inline some comments and at the end a simple patch that
>> fixes the problem with less code.
>>
>> Thanks
>> Praveen
>> On 13-May-14 1:27 PM, Minh Hon Chau wrote:
>>> osaf/services/saf/amf/amfnd/err.cc | 13 ++++++++++---
>>> osaf/services/saf/amf/amfnd/susm.cc | 24 ++++++++++++++++++------
>>> 2 files changed, 28 insertions(+), 9 deletions(-)
>>>
>>>
>>> Problem: In case of npi su restart recovery, the condition to change
>>> su presence
>>> state is not sufficient. If the component of the last csi in
>>> csi_list has been
>>> restarted first due to componentRestart (before suRestart), amfnd
>>> can not find
>>> any next csi and changes the su presence state to INSTANTIATED, this
>>> will miss
>>> the restart for the rest of components.
>>>
>>> Change: Mark all components are FAILED during suRestart, and su
>>> presence state
>>> is changed only if all components are unmarked FAILED.
>>>
>>> diff --git a/osaf/services/saf/amf/amfnd/err.cc
>>> b/osaf/services/saf/amf/amfnd/err.cc
>>> --- a/osaf/services/saf/amf/amfnd/err.cc
>>> +++ b/osaf/services/saf/amf/amfnd/err.cc
>>> @@ -616,13 +616,20 @@ static uint32_t avnd_err_rcvr_comp_resta
>>> uint32_t avnd_err_rcvr_su_restart(AVND_CB *cb, AVND_SU *su,
>>> AVND_COMP *failed_comp)
>>> {
>>> uint32_t rc = NCSCC_RC_SUCCESS;
>>> + AVND_COMP *comp = NULL;
>>> TRACE_ENTER();
>>> - /* mark the su & comp failed */
>>> + /* mark the su & its all comps are failed if not marked */
>>> m_AVND_SU_FAILED_SET(su);
>>> m_AVND_SEND_CKPT_UPDT_ASYNC_UPDT(cb, su,
>>> AVND_CKPT_SU_FLAG_CHANGE);
>>> - m_AVND_COMP_FAILED_SET(failed_comp);
>>> - m_AVND_SEND_CKPT_UPDT_ASYNC_UPDT(cb, failed_comp,
>>> AVND_CKPT_COMP_FLAG_CHANGE);
>>> +
>>> + for (comp =
>>> m_AVND_COMP_FROM_SU_DLL_NODE_GET(m_NCS_DBLIST_FIND_FIRST(&su->comp_list));
>>> + comp; comp =
>>> m_AVND_COMP_FROM_SU_DLL_NODE_GET(m_NCS_DBLIST_FIND_NEXT(&comp->su_dll_node)))
>>>
>>> {
>>> + if (!m_AVND_COMP_IS_FAILED(comp)) {
>>> + m_AVND_COMP_FAILED_SET(comp);
>>> + m_AVND_SEND_CKPT_UPDT_ASYNC_UPDT(cb, comp,
>>> AVND_CKPT_COMP_FLAG_CHANGE);
>>> + }
>>> + }
>> Why there is no differentiation between a PI and NPI SU in the loop.
>> The problem is reported in NPI SU case only and does not exist in PI
>> SU case.
>>> /* delete su current info */
>>> rc = avnd_su_curr_info_del(cb, su);
>>> diff --git a/osaf/services/saf/amf/amfnd/susm.cc
>>> b/osaf/services/saf/amf/amfnd/susm.cc
>>> --- a/osaf/services/saf/amf/amfnd/susm.cc
>>> +++ b/osaf/services/saf/amf/amfnd/susm.cc
>>> @@ -2638,12 +2638,24 @@ uint32_t avnd_su_pres_restart_compinst_h
>>> /* get the next csi */
>>> curr_csi = (AVND_COMP_CSI_REC
>>> *)m_NCS_DBLIST_FIND_NEXT(&curr_csi->si_dll_node);
>>> if (curr_csi) {
>>> - /* we have another csi. trigger the comp fsm with
>>> RestartEv */
>>> - rc = avnd_comp_clc_fsm_trigger(cb, curr_csi->comp,
>>> AVND_COMP_CLC_PRES_FSM_EV_RESTART);
>>> - if (NCSCC_RC_SUCCESS != rc)
>>> - goto done;
>>> - } else {
>>> - /* => si assignment done */
>>> + /* we have another csi. trigger the comp fsm with
>>> RestartEv if this comp is marked failed*/
>>> + if (curr_csi->comp->pres ==
>>> SA_AMF_PRESENCE_INSTANTIATED &&
>>> m_AVND_COMP_IS_FAILED(curr_csi->comp)) {
>>> + rc = avnd_comp_clc_fsm_trigger(cb, curr_csi->comp,
>>> AVND_COMP_CLC_PRES_FSM_EV_RESTART);
>>> + if (NCSCC_RC_SUCCESS != rc)
>>> + goto done;
>>> + }
>>> + }
>>> +
>>> + /* check whether all comp's are instantiated */
>>> + for (curr_comp =
>>> m_AVND_COMP_FROM_SU_DLL_NODE_GET(m_NCS_DBLIST_FIND_FIRST(&su->comp_list));
>>> + curr_comp && all_inst == true;
>>> + curr_comp =
>>> m_AVND_COMP_FROM_SU_DLL_NODE_GET(m_NCS_DBLIST_FIND_NEXT(&curr_comp->su_dll_node)))
>>>
>>> {
>>> + if (curr_comp->pres != SA_AMF_PRESENCE_INSTANTIATED ||
>>> m_AVND_COMP_IS_FAILED(curr_comp))
>>> + all_inst = false;
>>> + }
>>> +
>>> + /* OK, all are instantiated */
>>> + if (all_inst == true) {
>>> avnd_su_pres_state_set(su, SA_AMF_PRESENCE_INSTANTIATED);
>>> }
>> In case of NPI SU, loops cannot be iterated over components as a user
>> can configure spare components also. In that case SU will get stuck
>> in Restarting state as this check will never pass.
>> There is much simpler approach to solve this issue.It takes care of
>> this spare components problem also. Here is the simple patch :
>>
>> diff --git a/osaf/services/saf/amf/amfnd/susm.cc
>> b/osaf/services/saf/amf/amfnd/susm.cc
>> --- a/osaf/services/saf/amf/amfnd/susm.cc
>> +++ b/osaf/services/saf/amf/amfnd/susm.cc
>> @@ -2642,8 +2642,7 @@ uint32_t avnd_su_pres_restart_compinst_h
>> rc = avnd_comp_clc_fsm_trigger(cb,
>> curr_csi->comp, AVND_COMP_CLC_PRES_FSM_EV_RESTART);
>> if (NCSCC_RC_SUCCESS != rc)
>> goto done;
>> - } else {
>> - /* => si assignment done */
>> + } else if (all_csis_in_assigned_state(su)) {
>> avnd_su_pres_state_set(su,
>> SA_AMF_PRESENCE_INSTANTIATED);
>> }
>> }
>>
>> What do you think about this approach?
>> Isn't simple and much less code.
>> I have tested it for 2N model and it works fine with 15 csi
>> configuration.
>>
>>
>>
>>
>>> }
>>
>>
>
------------------------------------------------------------------------------
"Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE
Instantly run your Selenium tests across 300+ browser/OS combos.
Get unparalleled scalability from the best Selenium testing platform available
Simple to use. Nothing to install. Get started now for free."
http://p.sf.net/sfu/SauceLabs
_______________________________________________
Opensaf-devel mailing list
Opensaf-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/opensaf-devel
