In the current implementation of ILA, LWT is used to perform translation on both the input and output paths. This is functional, however there is a big performance hit in the receive path. Early demux occurs before the routing lookup (a hit actually obviates the route lookup). Therefore the stack currently performs early demux before translation so that a local connection with ILA addresses is never matched. Note that this issue is not just with ILA, but pretty much any translated or encapsulated packet handled by LWT would miss the opportunity for early demux. Solving the general problem seems non trivial since we would need to move the route lookup before early demx thereby mitigating the value.
This patch set addresses the issue for ILA by adding a fast locator lookup that occurs before early demux. This is done by creating an XFRM hook to perform address translation early in the receive path. For the backend we implement an rhashtable that contains identifier to locator to mappings. The table also allows more specific matches that include original locator and interface. This patch set: - Add an rhashtable function to atomically replace and element. This is useful to implement sub-trees from a table entry without needing to use a special anchor structure as the table entry. - Add a start callback for starting a netlink dump. - Creates an ila directory under net/ipv6 and moves ila.c to it. ila.c is split into ila_common.c and ila_lwt.c. - Implement a table to do identifier->locator mapping. This is an rhashtable. - Configuration for the table with netlink. - Add XFRM xlat_addr facility. This includes a callback registeration function and hook to call registered callbacks. - Call xfrm6_xlat_addr from ipv6_rcv before NF_HOOK and routing. Testing: Running 200 netperf TCP_RR streams No ILA, baseline 85.72% CPU utilization 1861945 tps 93/163/330 50/90/99% latencies ILA before fix (LWT on both input and output) 83.47 CPU utilization 16583186 tps (-11% from baseline) 107/183/338 50/90/99% latencies ILA after fix (hook for input) 84.97% CPU utilization 1833948 tps (-1.5% from baseline) 95/164/331 50/90/99% latencies Hacked DNPT to do ILA 80.94% CPU utilization 1683315 tps (-10% from baseline) 104/179/350 50/90/99% latencies Tom Herbert (6): ila: Create net/ipv6/ila directory rhashtable: add function to replace an element netlink: add a start callback for starting a netlink dump xfrm: Add xfrm6 address translation function ipv6: Call xfrm6_xlat_addr from ipv6_rcv ila: Add support for xfrm6_xlat_addr include/linux/netlink.h | 2 + include/linux/rhashtable.h | 82 ++++++ include/net/genetlink.h | 2 + include/net/xfrm.h | 25 ++ include/uapi/linux/ila.h | 22 ++ net/ipv6/Kconfig | 5 + net/ipv6/Makefile | 3 +- net/ipv6/ila.c | 229 ---------------- net/ipv6/ila/Makefile | 7 + net/ipv6/ila/ila.h | 48 ++++ net/ipv6/ila/ila_common.c | 103 ++++++++ net/ipv6/ila/ila_lwt.c | 152 +++++++++++ net/ipv6/ila/ila_xlat.c | 642 +++++++++++++++++++++++++++++++++++++++++++++ net/ipv6/ip6_input.c | 3 + net/ipv6/xfrm6_policy.c | 7 + net/ipv6/xfrm6_xlat_addr.c | 66 +++++ net/netlink/af_netlink.c | 4 + net/netlink/genetlink.c | 16 ++ 18 files changed, 1188 insertions(+), 230 deletions(-) delete mode 100644 net/ipv6/ila.c create mode 100644 net/ipv6/ila/Makefile create mode 100644 net/ipv6/ila/ila.h create mode 100644 net/ipv6/ila/ila_common.c create mode 100644 net/ipv6/ila/ila_lwt.c create mode 100644 net/ipv6/ila/ila_xlat.c create mode 100644 net/ipv6/xfrm6_xlat_addr.c -- 2.4.6 -- To unsubscribe from this list: send the line "unsubscribe netdev" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
