Re: [PATCH RFC net 1/2] tcp: Add TIME_WAIT sockets in bhash2.

[Kuniyuki Iwashima <kuniyu@amazon.com>]

From:   Joanne Koong <joannelkoong@gmail.com>
Date:   Fri, 23 Dec 2022 11:34:40 -0800
> On Thu, Dec 22, 2022 at 5:55 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> >
> > From:   Joanne Koong <joannelkoong@gmail.com>
> > Date:   Thu, 22 Dec 2022 16:25:10 -0800
> > > On Thu, Dec 22, 2022 at 3:27 PM Kuniyuki Iwashima <kuniyu@amazon.com> wrote:
> > > >
> > > > From:   Joanne Koong <joannelkoong@gmail.com>
> > > > Date:   Thu, 22 Dec 2022 13:46:57 -0800
> > > > > On Thu, Dec 22, 2022 at 7:06 AM Paolo Abeni <pabeni@redhat.com> wrote:
> > > > > >
> > > > > > On Thu, 2022-12-22 at 00:12 +0900, Kuniyuki Iwashima wrote:
> > > > > > > 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.
> > > > >
> > > > > (resending my reply because it wasn't in plaintext mode)
> > > > >
> > > > > Thanks for adding this! I hadn't realized TIME_WAIT sockets also are
> > > > > considered when checking against inet bind conflicts.
> > > > >
> > > > > >
> > > > > > How does keeping the timewait sockets inside bhash2 affect the bind
> > > > > > loopup performance? I fear that could defeat completely the goal of
> > > > > > 28044fc1d495, on quite busy server we could have quite a bit of tw with
> > > > > > the same address/port. If so, we could even consider reverting
> > > > > > 28044fc1d495.
> > > >
> > > > It will slow down along the number of twsk, but I think it's still faster
> > > > than bhash if we listen() on multiple IP.  If we don't, bhash is always
> > > > faster because of bhash2's additional locking.  However, this is the
> > > > nature of bhash2 from the beginning.
> > > >
> > > >
> > > > > >
> > > > >
> > > > > Can you clarify what you mean by bind loopup?
> > > >
> > > > I think it means just bhash2 traversal.  (s/loopup/lookup/)
> > > >
> > > > >
> > > > > > > [0]: https://lore.kernel.org/netdev/6b971a4e-c7d8-411e-1f92-fda29b5b2fb9@kernel.org/
> > > > > > >
> > > > > > > Fixes: 28044fc1d495 ("net: Add a bhash2 table hashed by port and address")
> > > > > > > Reported-by: Jiri Slaby <jirislaby@kernel.org>
> > > > > > > Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
> > > > > > > ---
> > > > > > >  include/net/inet_timewait_sock.h |  2 ++
> > > > > > >  include/net/sock.h               |  5 +++--
> > > > > > >  net/ipv4/inet_hashtables.c       |  5 +++--
> > > > > > >  net/ipv4/inet_timewait_sock.c    | 31 +++++++++++++++++++++++++++++--
> > > > > > >  4 files changed, 37 insertions(+), 6 deletions(-)
> > > > > > >
> > > > > > > diff --git a/include/net/inet_timewait_sock.h b/include/net/inet_timewait_sock.h
> > > > > > > index 5b47545f22d3..c46ed239ad9a 100644
> > > > > > > --- a/include/net/inet_timewait_sock.h
> > > > > > > +++ b/include/net/inet_timewait_sock.h
> > > > > > > @@ -44,6 +44,7 @@ struct inet_timewait_sock {
> > > > > > >  #define tw_bound_dev_if              __tw_common.skc_bound_dev_if
> > > > > > >  #define tw_node                      __tw_common.skc_nulls_node
> > > > > > >  #define tw_bind_node         __tw_common.skc_bind_node
> > > > > > > +#define tw_bind2_node                __tw_common.skc_bind2_node
> > > > > > >  #define tw_refcnt            __tw_common.skc_refcnt
> > > > > > >  #define tw_hash                      __tw_common.skc_hash
> > > > > > >  #define tw_prot                      __tw_common.skc_prot
> > > > > > > @@ -73,6 +74,7 @@ struct inet_timewait_sock {
> > > > > > >       u32                     tw_priority;
> > > > > > >       struct timer_list       tw_timer;
> > > > > > >       struct inet_bind_bucket *tw_tb;
> > > > > > > +     struct inet_bind2_bucket        *tw_tb2;
> > > > > > >  };
> > > > > > >  #define tw_tclass tw_tos
> > > > > > >
> > > > > > > diff --git a/include/net/sock.h b/include/net/sock.h
> > > > > > > index dcd72e6285b2..aaec985c1b5b 100644
> > > > > > > --- a/include/net/sock.h
> > > > > > > +++ b/include/net/sock.h
> > > > > > > @@ -156,6 +156,7 @@ typedef __u64 __bitwise __addrpair;
> > > > > > >   *   @skc_tw_rcv_nxt: (aka tw_rcv_nxt) TCP window next expected seq number
> > > > > > >   *           [union with @skc_incoming_cpu]
> > > > > > >   *   @skc_refcnt: reference count
> > > > > > > + *   @skc_bind2_node: bind node in the bhash2 table
> > > > > > >   *
> > > > > > >   *   This is the minimal network layer representation of sockets, the header
> > > > > > >   *   for struct sock and struct inet_timewait_sock.
> > > > > > > @@ -241,6 +242,7 @@ struct sock_common {
> > > > > > >               u32             skc_window_clamp;
> > > > > > >               u32             skc_tw_snd_nxt; /* struct tcp_timewait_sock */
> > > > > > >       };
> > > > > > > +     struct hlist_node       skc_bind2_node;
> > > > > >
> > > > > > I *think* it would be better adding a tw_bind2_node field to the
> > > > > > inet_timewait_sock struct, so that we leave unmodified the request
> > > > > > socket and we don't change the struct sock binary layout. That could
> > > > > > affect performances moving hot fields on different cachelines.
> > > > > >
> > > > > +1. The rest of this patch LGTM.
> > > >
> > > > Then we can't use sk_for_each_bound_bhash2(), or we have to guarantee this.
> > > >
> > > >   BUILD_BUG_ON(offsetof(struct sock, sk_bind2_node),
> > > >                offsetof(struct inet_timewait_sock, tw_bind2_node))
> > > >
> > > > Considering the number of members in struct sock, at least we have
> > > > to move sk_bind2_node forward.
> > > >
> > > > Another option is to have another TIME_WAIT list in inet_bind2_bucket like
> > > > tb2->deathrow or something.  sk_for_each_bound_bhash2() is used only in
> > > > inet_bhash2_conflict(), so I think this is feasible.
> > >
> > > Oh I see, thanks for clarifying!
> > >
> > > I think we could also check sk_state (which is in __sk_common already)
> > > and if it's TCP_TIME_WAIT, then we know sk is at offsetof(struct
> > > inet_timewait_sock, tw_bind2_node), whereas otherwise it's at
> > > offsetof(struct sock, sk_bind2_node). This seems simpler/cleaner to me
> > > than the other approaches. What are your thoughts?
> >
> > Sorry, I don't get it.  You mean we can check sk_state first and change
> > how we traverse ?  But then we cannot know the offset of sk_state if we
> > don't know if the socket is TIME_WAIT ... ?
> 
> I think the offset of sk_state is the same for both sockets because
> sk_state is in "struct sock_common" (__sk_common.skc_state) that both
> share.

Yes, but the problem is we don't know the address of the first member
of struct sock_common.  The hlist_head/hlist_node point to the next
hlist_node, not the head of sock_common.  In the loop, we calculate
a pointer address like this.

  ptr = (struct sock_common *)(list_node - offsetof(struct X, sk_bind2_node))

We have to know which struct the next node is, otherwise, the ptr will
points to an invalid address.