Support killing sockets using SOCK_DESTROY.
This gives netd the ability to close sockets on a particular
source IP address using SOCK_DESTROY. It does not yet enable
this behaviour.
The microbenchmark is able to close 500 IPv6 sockets in ~30ms on
my angler. Specifically:
- Scanning 500 socketpairs: ~5ms
- Scanning 500 socketpairs and killing one half of each: ~30ms
- Scanning 500 socketpairs and killing both halves of each: ~40ms
This is about ~2.5x-3.5x slower than SIOCKILLADDR:
- For 500 sockets, it's 9.5ms vs. 22.9ms.
- For 4000 sockets, it's ~40ms vs ~135ms.
A large part of that is due to sending RST packets, which
SIOCKILLADDR does not do. If the kernel is modified so that
SOCK_DESTROY does not send RSTs, the time taken to kill 4000
sockets goes down to ~70ms
Batching the destroy operations does not help much. It saves
5-10%, but it complicates error handling.
Bug: 26976388
(cherry picked from commit 8464e1ed13a30ca91ae44dd2e334f63de7ade0f3)
Change-Id: Ib85d573ee2af6f6ffc4ce50a31cfbb6e9a720ca1
5 files changed
