PPPoE subscriber disconnected due to 'lcp-keepalive-failure' on MPC5E PIC0 interfaces.

Some PPPoE subscribers may be disconnected by MX due to "lcp-keepalive-failure".

In general, if the client does not respond to BRAS PPP LCP echo request, the BRAS will terminate this PPP session. In this case, the problem clients are always on MPC5E PIC0 interfaces. When these problem clients were moved to PIC1 interfaces, the PPPoE subscribers worked stably.

Since MX uses inline-ka to maintain PPP sessions, we cannot use tcpdump to capture PPP keepalive messages on the MX side. We did capture on the switch, between the problem PPPoE subscriber and MX, then confirmed MX did NOT send out the PPP LCP echo request message.
 

user@mx> show network-access aaa terminate-code brief

Terminate-code:
  RADIUS     Custom Usage-Count Type Code
  10         no     1303        ppp  lcp-keepalive-failure

Inline-ka infra sends the PPP keepalive packet to the wrong XM and causes the PPP keepalive packet to drop.

The 'disp_cookie.xm_num' did not reset before starting transmission of a new inline-ka entry. If the first inline-ka entry transmitted on XM1 and the second ka entry is expected to send on XM0, then the 'disp_cookie.xm_num' value remains XM1 during the transmission of the second entry and causes the inline-ka infra to send the second KA entry to XM1, not to XM0. XM1 drops PPP keepalive packets as not destined to it.

This issue has been fixed via PR1343687 - Some PPPoE subscribers disconnected due to lcp-keepalive-failure on MPC5E/MPC6E PIC0 interfaces
Please upgrade to the resolved Junos verson.