[ovs-dev] [PATCH v9 06/10] Add incremental proessing to lflow_run

Han Zhou zhouhan at gmail.com
Wed Mar 23 18:39:04 UTC 2016


On Tue, Mar 22, 2016 at 6:17 PM, Han Zhou <zhouhan at gmail.com> wrote:
>
>
>
> On Fri, Mar 11, 2016 at 1:06 PM, Ryan Moats <rmoats at us.ibm.com> wrote:
> >
> > From: RYAN D. MOATS <rmoats at us.ibm.com>
> >
> > This code changes lflow_run to do incremental process of the
> > logical flow table rather than processing the full table each run.
> >
> > Signed-off-by: RYAN D. MOATS <rmoats at us.ibm.com>
> > ---
> >  ovn/controller/binding.c        |    3 ++
> >  ovn/controller/lflow.c          |   53
+++++++++++++++++++++++++++++++++------
> >  ovn/controller/lflow.h          |    4 +-
> >  ovn/controller/ofctrl.c         |    4 +-
> >  ovn/controller/ofctrl.h         |    2 +
> >  ovn/controller/ovn-controller.c |    5 +++-
> >  6 files changed, 58 insertions(+), 13 deletions(-)
> >
> > diff --git a/ovn/controller/binding.c b/ovn/controller/binding.c
> > index 602a8fe..87cae99 100644
> > --- a/ovn/controller/binding.c
> > +++ b/ovn/controller/binding.c
> > @@ -15,6 +15,7 @@
> >
> >  #include <config.h>
> >  #include "binding.h"
> > +#include "lflow.h"
> >
> >  #include "lib/bitmap.h"
> >  #include "lib/hmap.h"
> > @@ -139,6 +140,7 @@ remove_local_datapath(struct hmap *local_datapaths,
unsigned int ins_seqno)
> >      if (ld) {
> >          hmap_remove(local_datapaths, &ld->hmap_node);
> >          hmap_remove(&local_datapaths_by_seqno, &ld->seqno_hmap_node);
> > +        reset_flow_processing();
> >      }
> >  }
> >
> > @@ -156,6 +158,7 @@ add_local_datapath(struct hmap *local_datapaths,
> >      hmap_insert(local_datapaths, &ld->hmap_node,
> >                  binding_rec->datapath->tunnel_key);
> >      hmap_insert(&local_datapaths_by_seqno, &ld->seqno_hmap_node,
ins_seqno);
> > +    reset_flow_processing();
> >  }
> >
> >  static void
> > diff --git a/ovn/controller/lflow.c b/ovn/controller/lflow.c
> > index 4856362..6d0d417 100644
> > --- a/ovn/controller/lflow.c
> > +++ b/ovn/controller/lflow.c
> > @@ -176,6 +176,20 @@ struct logical_datapath {
> >      enum ldp_type type;         /* Type of logical datapath */
> >  };
> >
> > +void reset_flow_processing(void);
> > +void ldp_port_create(uint32_t ins_seqno, char *name,
> > +                     struct logical_datapath *ldp);
> > +void ldp_port_update(uint32_t ins_seqno, char *name,
> > +                     struct logical_datapath *ldp);
> > +
> > +bool restart_flow_processing = false;
> > +
> > +void
> > +reset_flow_processing(void)
> > +{
> > +    restart_flow_processing = true;
> > +}
> > +
> >  /* Contains "struct logical_datapath"s. */
> >  static struct hmap logical_datapaths =
HMAP_INITIALIZER(&logical_datapaths);
> >
> > @@ -208,6 +222,7 @@ ldp_create(const struct sbrec_datapath_binding
*binding)
> >      const char *ls = smap_get(&binding->external_ids,
"logical-switch");
> >      ldp->type = ls ? LDP_TYPE_SWITCH : LDP_TYPE_ROUTER;
> >      simap_init(&ldp->ports);
> > +    reset_flow_processing();
> >      return ldp;
> >  }
> >
> > @@ -224,6 +239,7 @@ ldp_free(struct logical_datapath *ldp)
> >      simap_destroy(&ldp->ports);
> >      hmap_remove(&logical_datapaths, &ldp->hmap_node);
> >      free(ldp);
> > +    reset_flow_processing();
> >  }
> >
> >  /* Whether a particular port has been seen or not
> > @@ -319,6 +335,7 @@ ldp_run(struct controller_ctx *ctx)
> >                                              binding->logical_port);
> >          if (!old || old->data != binding->tunnel_key) {
> >              simap_put(&ldp->ports, binding->logical_port,
binding->tunnel_key);
> > +            reset_flow_processing();
> >          }
> >
> >          ldp_port_update(ins_seqno, binding->logical_port, ldp);
> > @@ -380,13 +397,21 @@ lflow_init(void)
> >
> >  /* Translates logical flows in the Logical_Flow table in the OVN_SB
database
> >   * into OpenFlow flows.  See ovn-architecture(7) for more information.
*/
> > -void
> > +unsigned int
> >  lflow_run(struct controller_ctx *ctx,
> >            const struct simap *ct_zones,
> > -          struct hmap *local_datapaths)
> > +          struct hmap *local_datapaths,
> > +          unsigned int seqno)
> >  {
> >      struct hmap flows = HMAP_INITIALIZER(&flows);
> >      uint32_t conj_id_ofs = 1;
> > +    unsigned int processed_seqno = seqno;
> > +
> > +    if (restart_flow_processing) {
> > +        seqno = 0;
> > +        ovn_flow_table_clear();
> > +        restart_flow_processing = false;
> > +    }
> >
> >      ldp_run(ctx);
> >
> > @@ -398,17 +423,29 @@ lflow_run(struct controller_ctx *ctx,
> >              OVSDB_IDL_CHANGE_MODIFY);
> >          unsigned int ins_seqno =
sbrec_logical_flow_row_get_seqno(lflow,
> >              OVSDB_IDL_CHANGE_INSERT);
> > -        // this offset is to protect the hard coded rules in physical.c
> > -        ins_seqno += 4;
> > -
> > +        if (del_seqno <= seqno && mod_seqno <= seqno && ins_seqno <=
seqno) {
> > +            continue;
> > +        }
> >          /* if the row has a del_seqno > 0, then trying to process the
> >           * row isn't going to work (as it has already been freed).
> > -         * Therefore all we can do is to pass the ins_seqno to
> > +         * Therefore all we can do is to pass the offset ins_seqno to
> >           * ofctrl_remove_flow() to remove the flow */
> >          if (del_seqno > 0) {
> > -            ofctrl_remove_flow(ins_seqno);
> > +            ofctrl_remove_flow(ins_seqno+4);
> > +            if (del_seqno > processed_seqno) {
> > +                processed_seqno = del_seqno;
> > +            }
> >              continue;
> >          }
> > +        if (mod_seqno > processed_seqno) {
> > +            processed_seqno = mod_seqno;
> > +        }
> > +        if (ins_seqno > processed_seqno) {
> > +            processed_seqno = ins_seqno;
> > +        }
> > +
> > +        // this offset is to protect the hard coded rules in physical.c
> > +        ins_seqno += 4;
> >
> >          /* Find the "struct logical_datapath" associated with this
> >           * Logical_Flow row.  If there's no such struct, that must be
because
> > @@ -544,12 +581,12 @@ lflow_run(struct controller_ctx *ctx,
> >                  ofpbuf_uninit(&conj);
> >              }
> >          }
> > -
> >          /* Clean up. */
> >          expr_matches_destroy(&matches);
> >          ofpbuf_uninit(&ofpacts);
> >          conj_id_ofs += n_conjs;
> >      }
> > +    return processed_seqno;
> >  }
> >
> >  void
> > diff --git a/ovn/controller/lflow.h b/ovn/controller/lflow.h
> > index e0e902c..31b187a 100644
> > --- a/ovn/controller/lflow.h
> > +++ b/ovn/controller/lflow.h
> > @@ -56,8 +56,8 @@ struct uuid;
> >  #define LOG_PIPELINE_LEN 16
> >
> >  void lflow_init(void);
> > -void lflow_run(struct controller_ctx *, const struct simap *ct_zones,
> > -               struct hmap *local_datapaths);
> > +unsigned int lflow_run(struct controller_ctx *, const struct simap
*ct_zones,
> > +                       struct hmap *local_datapaths, unsigned int
seqno);
> >  void lflow_destroy(void);
> >
> >  #endif /* ovn/lflow.h */
> > diff --git a/ovn/controller/ofctrl.c b/ovn/controller/ofctrl.c
> > index 2479ca1..5aa7044 100644
> > --- a/ovn/controller/ofctrl.c
> > +++ b/ovn/controller/ofctrl.c
> > @@ -106,7 +106,7 @@ static struct hmap installed_flows;
> >   * S_CLEAR_FLOWS or S_UPDATE_FLOWS, this is really the option we have.
*/
> >  static enum mf_field_id mff_ovn_geneve;
> >
> > -static void ovn_flow_table_clear(void);
> > +void ovn_flow_table_clear(void);
> >  static void ovn_flow_table_destroy(void);
> >
> >  static void ofctrl_recv(const struct ofp_header *, enum ofptype);
> > @@ -728,7 +728,7 @@ ovn_flow_destroy(struct ovn_flow *f)
> >
> >  /* Flow tables of struct ovn_flow. */
> >
> > -static void
> > +void
> >  ovn_flow_table_clear(void)
> >  {
> >      struct ovn_flow *f, *next;
> > diff --git a/ovn/controller/ofctrl.h b/ovn/controller/ofctrl.h
> > index 4ae0d42..ff870e6 100644
> > --- a/ovn/controller/ofctrl.h
> > +++ b/ovn/controller/ofctrl.h
> > @@ -43,4 +43,6 @@ void ofctrl_add_flow(uint8_t table_id, uint16_t
priority,
> >
> >  void ofctrl_remove_flow(unsigned int ins_seqno);
> >
> > +void ovn_flow_table_clear(void);
> > +
> >  #endif /* ovn/ofctrl.h */
> > diff --git a/ovn/controller/ovn-controller.c
b/ovn/controller/ovn-controller.c
> > index cb8536b..258d83e 100644
> > --- a/ovn/controller/ovn-controller.c
> > +++ b/ovn/controller/ovn-controller.c
> > @@ -208,6 +208,7 @@ main(int argc, char *argv[])
> >      struct unixctl_server *unixctl;
> >      bool exiting;
> >      int retval;
> > +    unsigned int ovnsb_last_lflow_seqno = 0;
> >
> >      ovs_cmdl_proctitle_init(argc, argv);
> >      set_program_name(argv[0]);
> > @@ -303,7 +304,9 @@ main(int argc, char *argv[])
> >
> >              pinctrl_run(&ctx, br_int);
> >
> > -            lflow_run(&ctx, &ct_zones, &local_datapaths);
> > +            ovnsb_last_lflow_seqno = lflow_run(&ctx, &ct_zones,
> > +                                               &local_datapaths,
> > +                                               ovnsb_last_lflow_seqno);
> >              if (chassis_id) {
> >                  physical_run(&ctx, mff_ovn_geneve,
> >                               br_int, chassis_id, &ct_zones,
> > --
> > 1.7.1
> >
> > _______________________________________________
> > dev mailing list
> > dev at openvswitch.org
> > http://openvswitch.org/mailman/listinfo/dev
>
> I just realized a problem here. We have an optimization already in
lflow_run to process egress flows only if it belongs to local-datapath,
which was added by Russell some time ago.
>
> With incremental lflow_run this can be a problem. Considering below
scenario:
>
> - HV-1 has lports for lswitch-A, so local-datapath contains only lswitch-A
> - Egress lflows for lswtich-B is added in SB, but ignored by HV-1 because
lswitch-B is not local.
> - A lport of lswitch-B attached to HV-1. Now lswitch-B becomes local for
HV-1. But there is no egress lflow updates for lswitch-B, so HV-1 still
will not generate physical flows for those egress lflows of lswitch-B,
because of the incremental processing logic.
>
> I don't think we should drop the local-datapath optimization. So the only
way I am thinking right now is to mark the lflows that we ignored, and when
a new lswitch becomes local, we need to recheck those lflows to see if it
belong to the newly added local-datapath. Any better ideas?
>
Marking the lflows that we ignored requires maintaining a lflow cache. A
simpler (but less optimized) approach would be resetting the seqno when a
new datapath is added to local_datapath so that we can fall back to
reprocess all lflows. This is just the idea, but I haven't worked out all
details. Ryan, what do you think?


--
Best regards,
Han



More information about the dev mailing list