Hi Vu,
I raised the concern for termination in a separate thread because IMCN is
blocked in one of the CCB callbacks while sending notification
NotificationSend() API. This is happening because of circular depedency as
standby NTFS is blocked. We need to remember that when IMCN is blocked, the
test case is still continuing and making configuration changes and because of
this IMCN is accumulating those callbacks (IMM is not blocked and sending those
callbacks to IMCN). These callbacks plus already in the list will be processed
once IMCN gets unlocked. Now moving termination in a separate thread will
surely allow standby NTFS to proceed and will unblock IMCN, but IMCN component
after getting unblocked will still be executing those remaining callbacks.
Earliar NTFS used to abort IMCN when terminates gets timed out. So we need
gaurantee that thread will surely terminated IMCN.
One more thing, Will the temination by NTFS in a separate thread not lead to
accumulation of terminations itself? Suppose NTFS started the thread for
termination and termination thread has still not exited and another switchver
happens. In this way one more thread will get spawned for termination.
If both these problems are resolved and gauranteed, I do not see any problem
with V3 because my gues is the problem of notifcation being lost has been
always existed and it will remain so irrespective of v1 or v3 (refined version)
approach. The reasons are:
1) NTFS responds to AMF in csi set callback before termination of IMCN
component. Because of this during switchover, a time comes when active NTFS has
just responded for active assignments and is terminating standby IMCN process.
While this is happening quiesced NTFS has got the standby CSI callback and it
also has started termination of active IMCN process. Since there is no
syncronization between these two terminations though on different nodes,
configuration changes may get missed and hence the notification.
2) IMCN is not role aware. So NTFS currently terminates IMCN when it gets
standby role andMCN exits without any confirmation that it has cleared its
notification list and CCB callback list. Neither patch v1 not v3 handles this.
Note: These is a ticket fot these two (#157 NTF: Improve HA handling in IMCN).
Thanks,
Praveen
---
** [tickets:#2219] ntfd: circular dependency with osafntfimcnd**
**Status:** assigned
**Milestone:** 5.0.2
**Created:** Thu Dec 08, 2016 05:14 AM UTC by Gary Lee
**Last Updated:** Mon Dec 12, 2016 03:03 AM UTC
**Owner:** Praveen
A circular dependency can be seen when performing a si-swap of
safSi=SC-2N,safApp=OpenSAF:
1. Active NTFD is trying to sync with Standby using MBC
2. Standby NTFD is the process of terminating its local osafntfimcnd. It is
stuck in timedwait_imcn_exit() and cannot reply to the Active.
3. osafntfimcnd [on standby] is trying to send a notfication to Active NTFD
So we have (1) depending on (2) depending on (3) depending on (1)
This results in a temporary deadlock that dramatically slows down NTFD's
ability to process its main dispatch loop. The deadlock only lasts for approx.
1 second, when mbcsv_mds_send_msg() times out. But since there could be lots of
MBC messages to send, sometimes osafntfimcnd is killed with SIGABRT generating
a coredump. The si-swap operation will also timeout.
steps to reproduce
- Run loop of ntftest 32
root@SC-1:~# for i in {1..10}; do ntftest 32; done
- On another terminal, keep swapping 2N Opensaf SI, got coredump after couples
of swaps
root@SC-1:~# amf-adm si-swap safSi=SC-2N,safApp=OpenSAF
...
root@SC-1:~# amf-adm si-swap safSi=SC-2N,safApp=OpenSAF
~~~
SC-2 (active)
There are a lot of send failures. Each taking approx. 1 second to timeout.
During these 1 second timeouts, NTFD cannot process the main dispatch loop.
Dec 7 11:01:37.531772 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:37.531781 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:38.537307 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:38.537758 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:38.537766 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:39.543180 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:39.543695 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:39.543698 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:40.545252 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:40.545719 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:40.545726 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:41.551328 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:41.551971 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:41.551979 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:42.557594 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:42.558171 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:42.558179 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:43.564051 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:43.564874 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:43.564883 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:44.572407 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:44.573262 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
Dec 7 11:01:44.573271 osafntfd [452:mbcsv_mds.c:0209] TR send type
MDS_SENDTYPE_REDRSP:
Dec 7 11:01:45.575091 osafntfd [452:mbcsv_mds.c:0247] <<
mbcsv_mds_send_msg: failure
Dec 7 11:01:47.083548 osafntfd [452:mbcsv_mds.c:0185] >>
mbcsv_mds_send_msg: sending to vdest:e
~~~
~~~
SC-1 (standby)
NTFD is trying to terminate osafntfimcnd. While it is doing that, it cannot
reply to NTFD on SC-2. Meanwhile, osafntfimcnd is sending NTF notifications to
NTFD on SC-1.
Dec 7 11:01:35.453151 osafntfd [464:ntfs_imcnutil.c:0316] TR
handle_state_ntfimcn: Terminating osafntfimcnd process
Dec 7 11:01:45.474313 osafntfd [464:ntfs_imcnutil.c:0124] TR Termination
timeout
Dec 7 11:01:45.474375 osafntfd [464:ntfs_imcnutil.c:0130] <<
wait_imcnproc_termination: rc = -1, retry_cnt = 101
Dec 7 11:01:45.474387 osafntfd [464:ntfs_imcnutil.c:0168] TR Normal
termination failed. Escalate to abort
Dec 7 11:01:45.574703 osafntfd [464:ntfs_imcnutil.c:0172] TR Imcn
successfully aborted
Dec 7 11:01:45.574712 osafntfd [464:ntfs_imcnutil.c:0187] <<
timedwait_imcn_exit
~~~
---
Sent from sourceforge.net because [email protected] is
subscribed to https://sourceforge.net/p/opensaf/tickets/
To unsubscribe from further messages, a project admin can change settings at
https://sourceforge.net/p/opensaf/admin/tickets/options. Or, if this is a
mailing list, you can unsubscribe from the mailing list.------------------------------------------------------------------------------
Developer Access Program for Intel Xeon Phi Processors
Access to Intel Xeon Phi processor-based developer platforms.
With one year of Intel Parallel Studio XE.
Training and support from Colfax.
Order your platform today.http://sdm.link/xeonphi
_______________________________________________
Opensaf-tickets mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/opensaf-tickets
