On 8/20/25 4:27 PM, Dima Chumak wrote:
> On 2025-08-01 10:43 PM, Ilya Maximets wrote:
>> External email: Use caution opening links or attachments
>>
>>
>> On 7/23/25 3:12 PM, Dima Chumak via dev wrote:
>>> Introduce support for route lookup across several routing tables
>>> following a priority from highest to lowest.
>>>
>>> To properly prioritize route lookup order, the single router table in
>>> OVS is now split into three tables: local, main and default.
>>> Corresponding routing rules are created by default, where local table
>>> has the highest priority and main and default tables have two lowest
>>> priorities respectively. All non-standard tables can be added with a
>>> priority in between local and main.
>>>
>>> Non-standard routing tables are created by reading the Routing Policy
>>
>> Why only the non-standard tables?
>>
>>> Database (RPDB) from the kernel, on systems that support it, and
>>> importing only those tables which are referenced in the RPDB rules with
>>> a table lookup action. The table IDs and rule priority are copied from
>>> the kernel RPDB as is.
>>>
>>> Current implementation only supports RPDB rules with a source address
>>> selector, in form of '[not] from IP' match.
>>>
>>> Signed-off-by: Dima Chumak <dchu...@nvidia.com>
>>> ---
>>> lib/ovs-router.c | 231 +++++++++++++++++++++++++++++------
>>> lib/ovs-router.h | 7 ++
>>> lib/packets.c | 20 +++
>>> lib/packets.h | 2 +
>>> lib/route-table-bsd.c | 4 +-
>>> lib/route-table-stub.c | 4 +-
>>> lib/route-table.c | 225 ++++++++++++++++++++++++++++++++--
>>> lib/route-table.h | 18 ++-
>>> tests/test-lib-route-table.c | 5 +-
>>> 9 files changed, 465 insertions(+), 51 deletions(-)
>
> <snip>
>
>>> +bool
>>> +ovs_router_is_referenced(uint32_t table)
>>> +{
>>> + struct router_rule *rule;
>>> +
>>> + if (route_table_is_standard_id(table)) {
>>> + return true;
>>> + }
>>> +
>>> + PVECTOR_FOR_EACH (rule, &rules) {
>>> + if (rule->lookup_table == table) {
>>> + return true;
>>> + }
>>> + }
>>
>> A lookup in the hash map may be better here instead of iteration.
>>
>
> If I understand it right, this would require a dedicated hash table to
> store table IDs from the rule list. And the value stored would probably
> be refcounted because one table can appear multiple times in the rule
> list. I can add it as a follow up optimization change, what do you
> think?
Can be a ccmap (counting cmap), I guess. I'd not expect to have
a lot of rules, so maybe it's not needed after all. But let's see.
Can surely be optimized later, if necessary.
>
> Alternatively, if we try to consult the clsmap hash to check if table
> already exists it will prevent the logic in route_table_change() to work
> correctly - a new table could never be synchronized from the kernel
> because it would be considered as un-referenced.
OK.
>
> Thanks, Dima
>
> <snip>
_______________________________________________
dev mailing list
d...@openvswitch.org
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
