On 1 October 2014 06:56, Pravin Shelar <pshe...@nicira.com> wrote:
> On Mon, Sep 29, 2014 at 3:59 PM, Joe Stringer <joestrin...@nicira.com>
> wrote:
> >
> >
> > On 30 September 2014 10:10, Ben Pfaff <b...@nicira.com> wrote:
> >>
> >> On Fri, Sep 26, 2014 at 09:28:15PM +1200, Joe Stringer wrote:
> >> > If a datapath is created with the flag OVS_DP_F_INDEX_BY_UID, then an
> >> > additional table_instance is added to the flow_table, which is indexed
> >> > by unique identifiers ("UID"). Userspace implementations can specify a
> >> > UID of up to 128 bits along with a flow operation as shorthand for the
> >> > key. This allows revalidation performance improvements of up to 50%.
> >> >
> >> > If a datapath is created using OVS_DP_F_INDEX_BY_UID and a UID is not
> >> > specified at flow setup time, then that operation will fail. If
> >> > OVS_UID_F_* flags are specified for an operation, then they will
> modify
> >> > what is returned through the operation. For instance,
> OVS_UID_F_SKIP_KEY
> >> > allows the datapath to skip returning the key (eg, during dump to
> reduce
> >> > memory copy).
> >> >
> >> > Signed-off-by: Joe Stringer <joestrin...@nicira.com>
> >> > ---
> >> > v6: Fix documentation for supporting UIDs between 32-128 bits.
> >> > Minor style fixes.
> >> > Rebase.
> >> > v5: No change.
> >> > v4: Fix memory leaks.
> >> > Log when triggering the older userspace issue above.
> >> > v3: Initial post.
> >>
> >> This review is from an ABI standpoint only; it's not a review of the
> >> kernel code itself.
> >>
> >> OVS_UID_ATTR_ID is marked as 32-128 bits long. For the "userdata"
> >> attribute of the userspace action, we originally had it fixed at 64
> >> bits, then later we decided that it was more flexible to allow it to
> >> be any size. Is there an advantage to fixing it within this range?
> >
> >
> > I'm not sure there's any advantage, that's just the way it's written
> right
> > now. Perhaps with a bit of tweaking, we could get rid of MAX_UID_BUFSIZE
> and
> > have no restrictions on the size of this.
> >
> >>
> >>
> >> I'm a little surprised that OVS_DP_F_INDEX_BY_UID is necessary. In
> >> the past we've only added flags for features that otherwise required a
> >> backward-incompatible change to the datapath interface. Is adding a
> >> UID such a change?
> >
> >
> > Pravin had some preferences on this during the original drafting, but I
> > can't find a direct requirement for this. The alternative means that
> flows
> > might not be present in both of the hastables (indexed by UID vs. exact
> > flow_key), although they would always need to be in the exact flow_key
> > table. Might be worth bouncing this off Pravin to see if I'm on the mark
> > with how I've used it here.
>
> I looked into the patch and I think we can get rid of
> OVS_DP_F_INDEX_BY_UID.
> On flow insert we can use flow-id provided by userspace, if it is not
> passed we can generate in kernel and use it to insert it in the UID
> hash table. sw_flow can have a flag set for kernel generated flow-uid,
> this can be used along with OVS_UID_F_SKIP_KEY in flow dump operation
> to return key to userspace or not. On flow dump can always iterate UID
> hash table where flow iteration should be relatively stable.
>
OK, so in this case generally when userspace doesn't support UIDs, the
datapath will generate them mainly to keep the flow-id and flow-key
hashtables in sync. (Currently, we just disable the flow-id hashtable).
Currently, OVS_UID_F_SKIP_KEY is a request flag that means "omit the flow
key". Are you suggesting that it should only omit flow keys for flows which
have a userspace-specified flow-id?
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
