Hi Andreas, 

The session hash table entry is cleaned up by session_delete, unless the 
transport has somehow been altered and the lookup fails. Obviously the 
transport should be still intact in closed state, hence the delete 
notification. Do you see the warning message that the hash cannot be deleted? 

I assume you do, in which case you need to figure out how the transport got 
altered (for instance, in the transport cleanup notify you do something to the 
transport that’s about to be removed). If not, then we need to understand why 
that would happen, i.e., hash claims it deleted something that you later find 
it’s still valid.   

We’ve been regularly testing with >1M entries recently but because of the 
specific circumstances, without using the session hash. So, at least the 
session delete logic seems to be fine, if not the way the hash is cleaned up. 

Regards,
Florin

> On May 29, 2020, at 5:32 AM, Andreas Schultz <andreas.schu...@travelping.com> 
> wrote:
> 
> Hi,
> 
> It seems that the {v4,v6}_session_hash is leaking entries. When creating 
> large amounts of TCP session (incoming and outgoing) we are observing entries 
> in the session_hash pointing to already deleted sessions.
> 
> I have verified with a somewhat hackish check in session_delete that 
> sometimes a session is deleted that still has an active entry in the session 
> hash. The entry does not get deleted later. Under the right circumstances, 
> that invalid entry is then picked up by session_lookup_connection_wt4, 
> leading to crash.
> 
> So far the cause seems to be the session_delete in 
> session_transport_delete_notify when the status is STATE_CLOSED. I have no 
> idea how the session ends up there.
> 
> Regards,
> Andreas
> 
> -- 
> Andreas Schultz
> 

-=-=-=-=-=-=-=-=-=-=-=-
Links: You receive all messages sent to this group.

View/Reply Online (#16573): https://lists.fd.io/g/vpp-dev/message/16573
Mute This Topic: https://lists.fd.io/mt/74542352/21656
Group Owner: vpp-dev+ow...@lists.fd.io
Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub  [arch...@mail-archive.com]
-=-=-=-=-=-=-=-=-=-=-=-

Reply via email to