Age | Commit message (Collapse) | Author | Files | Lines |
|
Type: improvement
Remove rwlock contention on timestamps. ~10% pps with
10k sessions. Use fixed-size-pools of increasing sizes
starting with 4K, and with a x2 step each time.
We don't free/shrink allocated pools.
Change-Id: I5fea51faba40430106c823275a6356e81709d118
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
|
|
This replace the cnat ip4/ip6 to client
lookups previously done with a regular
hash, by a bihash lookup.
Type: improvement
Do the client lookup in a bihash instead of
a hash.
Change-Id: I730c1893525c002b44ada8e290a36802835e88e9
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
|
|
Type: fix
This fixes two issues :
- We used a hash to throttle RPC for adding fib entries,
but as we rely on a refcount, we cannot accept loosing an
entry, which could happen in case of a collision.
- On client cleanup we weren't freeing the fib entry correctly
which resulted in crashes when recreating an entry.
Added a test that ensures proper cleanup
Change-Id: Ie6660b0b02241f75092737410ae2299f8710d6b9
Signed-off-by: Nathan Skrzypczak <nathan.skrzypczak@gmail.com>
|
|
Type: feature
Signed-off-by: Neale Ranns <nranns@cisco.com>
Change-Id: I64a99a4fbc674212944247793fd5c1fb701408cb
|