tcp: Add TIME_WAIT sockets in bhash2.
authorKuniyuki Iwashima <kuniyu@amazon.com>
Mon, 26 Dec 2022 13:27:52 +0000 (22:27 +0900)
committerDavid S. Miller <davem@davemloft.net>
Fri, 30 Dec 2022 07:25:52 +0000 (07:25 +0000)
commit936a192f974018b4f6040f6f77b1cc1e75bd8666
treea69f32b36b71ddadb25fdffdba76d8ad380fc7f2
parent40cab44b9089a41f71bbd0eff753eb91d5dafd68
tcp: Add TIME_WAIT sockets in bhash2.

Jiri Slaby reported regression of bind() with a simple repro. [0]

The repro creates a TIME_WAIT socket and tries to bind() a new socket
with the same local address and port.  Before commit 28044fc1d495 ("net:
Add a bhash2 table hashed by port and address"), the bind() failed with
-EADDRINUSE, but now it succeeds.

The cited commit should have put TIME_WAIT sockets into bhash2; otherwise,
inet_bhash2_conflict() misses TIME_WAIT sockets when validating bind()
requests if the address is not a wildcard one.

The straight option is to move sk_bind2_node from struct sock to struct
sock_common to add twsk to bhash2 as implemented as RFC. [1]  However, the
binary layout change in the struct sock could affect performances moving
hot fields on different cachelines.

To avoid that, we add another TIME_WAIT list in inet_bind2_bucket and check
it while validating bind().

[0]: https://lore.kernel.org/netdev/6b971a4e-c7d8-411e-1f92-fda29b5b2fb9@kernel.org/
[1]: https://lore.kernel.org/netdev/20221221151258.25748-2-kuniyu@amazon.com/

Fixes: 28044fc1d495 ("net: Add a bhash2 table hashed by port and address")
Reported-by: Jiri Slaby <jirislaby@kernel.org>
Suggested-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Joanne Koong <joannelkoong@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/inet_hashtables.h
include/net/inet_timewait_sock.h
net/ipv4/inet_connection_sock.c
net/ipv4/inet_hashtables.c
net/ipv4/inet_timewait_sock.c