[ovs-dev] [PATCH] ovn-controller: process lport bindings only when transaction is possible

[Ben Pfaff]

On Sat, Jul 02, 2016 at 04:11:14PM -0400, Lance Richardson wrote:
> 
> 
> ----- Original Message -----
> > From: "Ben Pfaff" <blp at ovn.org>
> > To: "Lance Richardson" <lrichard at redhat.com>
> > Cc: dev at openvswitch.org
> > Sent: Saturday, July 2, 2016 3:23:35 PM
> > Subject: Re: [ovs-dev] [PATCH] ovn-controller: process lport bindings only when transaction is possible
> > 
> > On Sat, Jul 02, 2016 at 01:56:21PM -0400, Lance Richardson wrote:
> > > As currently implemented, binding_run() normally updates the set of
> > > locally owned logical ports on each call.  When changes to the
> > > membership of this set are detected (i.e. when locally bound
> > > logical ports are added or deleted), additional processing to
> > > update the sb database with lport binding is performed.
> > > 
> > > However, the sb database can only be updated when a transaction to
> > > the sb database is possible (that is, when ctx->ovnsb_idl_txn is
> > > non-NULL).  If a new logical port is detected  while ctx->ovnsb_idl_txn
> > > happens to be NULL, its binding information will not be updated in
> > > the the sb database until another change to the set of locally-owned
> > > logical ports changes. If no such change ever occurs, the sb database
> > > is never updated with the appropriate binding information.
> > > 
> > > Eliminate this issue by only updating the set of locally owned logical
> > > ports when an sb database transaction is possible. This addresses
> > > a cause of occasional failures in the "3 HVs, 3 LS, 3 lports/LS, 1 LR"
> > > test case.
> > > 
> > > Signed-off-by: Lance Richardson <lrichard at redhat.com>
> > > ---
> > >  ovn/controller/binding.c | 2 +-
> > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > 
> > > diff --git a/ovn/controller/binding.c b/ovn/controller/binding.c
> > > index 8b439a6..369f8f2 100644
> > > --- a/ovn/controller/binding.c
> > > +++ b/ovn/controller/binding.c
> > > @@ -266,7 +266,7 @@ binding_run(struct controller_ctx *ctx, const struct
> > > ovsrec_bridge *br_int,
> > >      }
> > >  
> > >      if (br_int) {
> > > -        if (get_local_iface_ids(br_int, &lports)) {
> > > +        if (ctx->ovnsb_idl_txn && get_local_iface_ids(br_int, &lports)) {
> > >              process_full_binding = true;
> > >          }
> > >      } else {
> > 
> > At a glance, get_local_iface_ids() doesn't seem to directly modify the
> > southbound database, but it does modify data structures that later code
> > depends on.  Based on that, it's not obvious for me to see that this is
> > a correct change.  Can you point out anything relevant?
> > 
> > Thanks,
> > 
> > Ben.
> > 
> 
> Hi Ben,
> 
> The failing scenario goes like this:
>    1) Test case logical network setup is complete.
>    2) The last physical network port is added via
>       as hv3 ovs-vsctl --add-port ... --set Interface vif333 external-ids:iface-id=lp333
>    3) hv3 ovn-controller receives update from hv3 ovsdb-server with above mapping,
>       binding_run() is called, and ctx->ovnsb_idl_txn happens to be NULL.
>    4) binding_run() calls get_local_iface_ids(), which recognizes the new
>       local port as matching a logical port, so the lp333 is added to the
>       global ssets "lports" and "all_lports".  This means lp333 will not be treated
>       as a new logical port on subsequent calls. Because getLocal_iface_ids()
>       has discovered a new lport, it returns changed = true.
>    5) Because get_local_iface_ids() returned true, binding_run() sets process_full_binding
>       to true.
>    6) Because process_full_binding is true, binding_run() calls consider_local_datapath()
>       for each logical port in shash_lports (which now includes lp333).
>    7) consider_local_datapath() processing returns without calling
>       sbrec_port_binding_set_chassis() because ctx->ovnsb_idl_txn is NULL.
>    8) There are subsequent calls to binding_run() with non-NULL ctx->ovnsb_idl,
>       but because lp333 is already in the "lports" sset, get_local_iface_ids()
>       returns changed=false, so process_full_binding is false, which means
>       consider_local_datapath() is not called for lp333.
>    9) Because consider_local_datapath() is not called for lp333, the sb database
>       is not updated with the lport/chassis binding.
> 
> Hopefully the above is intelligible. Another way of looking at it would be
> to say the condition for calling consider_local_datapath() is an "edge trigger",
> this change suppresses the trigger until the necessary actions can be performed.

That is an amazingly detailed rationale.  I appended all of it to the
commit message and applied this to master.

> One thing I have a doubt about is whether there's any guarantee that, after a call
> with ctx->ovnsb_idl_txn==NULL, there will be another call with non-NULL ctx->ovnsb_idl_txn,
> but I suspect that other things would not work without this being the case.

This should be guaranteed.