Re: [Gems-users] TBE allocation

[Mike Marty <mikem@xxxxxxxxxxx>]

It is possible.  MOESI_SMP_token-cache.sm includes the following
transitions that give a processor modify permission when a TBE is not
allocated:

  transition(O, Data_Shared_All_Tokens, M) {
    w_assertIncomingDataAndCacheDataMatch;
    q_updateTokensFromResponse;
    n_popResponseQueue;
  }

  transition(O, Ack_All_Tokens, M) {
    q_updateTokensFromResponse;
    n_popResponseQueue;
  }

  transition(S, Data_Owner_All_Tokens, M) {
    w_assertIncomingDataAndCacheDataMatch;
    q_updateTokensFromResponse;
    n_popResponseQueue;
  }

  transition(I, Data_Owner_All_Tokens, M) {
    u_writeDataToCache;
    q_updateTokensFromResponse;
    n_popResponseQueue;
  }


However these transitions should be quite rare in practice and only occur
under severe network delays/reorderings.  For example, consider this
weird situation:

P1 broadcasts GETX.  Message to P5 is severely delayed
P5 broadcasts GETX
P3 responds to P1's request, P1 gets all tokens
P1 receives P5's request and responds to P5
P5 finally receives delayed request from P1 and responds to P1
P1 receives data and all tokens, and has an Invalid tag allocated, so it
keeps them and now has exclusive permission.  no TBE was present.

--Mike


> I am currently modifying MOESI_SMP_token, and I was wondering if there are
> any situations in which the cache can receive modify permission without
> having a TBE allocated.  It seems like it cannot happen at first glance,
> but I have an error which as far as I can tell would be caused by
> something like this.
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>