Date:   Wed, 2 Nov 2022 11:35:12 -0700
From:   Kuniyuki Iwashima <kuniyu@...zon.com>
To:     <joannelkoong@...il.com>
CC:     <davem@...emloft.net>, <edumazet@...gle.com>, <kuba@...nel.org>,
        <kuni1840@...il.com>, <kuniyu@...zon.com>, <martin.lau@...nel.org>,
        <martin.lau@...ux.dev>, <mathew.j.martineau@...ux.intel.com>,
        <netdev@...r.kernel.org>, <pabeni@...hat.com>
Subject: Re: [RFC] bhash2 and WARN_ON() for inconsistent sk saddr.

From:   Joanne Koong <joannelkoong@...il.com>
Date:   Wed, 2 Nov 2022 00:57:00 +0000
> On Tue, Nov 1, 2022 at 11:29 PM Martin KaFai Lau <martin.lau@...ux.dev> wrote:
> >
> > On 10/28/22 5:12 PM, Kuniyuki Iwashima wrote:
> > > Hi,
> > >
> > > I want to discuss bhash2 and WARN_ON() being fired every day this month
> > > on my syzkaller instance without repro.
> > >
> > >    WARNING: CPU: 0 PID: 209 at net/ipv4/inet_connection_sock.c:548 inet_csk_get_port (net/ipv4/inet_connection_sock.c:548 (discriminator 1))
> > >    ...
> > >    inet_csk_listen_start (net/ipv4/inet_connection_sock.c:1205)
> > >    inet_listen (net/ipv4/af_inet.c:228)
> > >    __sys_listen (net/socket.c:1810)
> > >    __x64_sys_listen (net/socket.c:1819 net/socket.c:1817 net/socket.c:1817)
> > >    do_syscall_64 (arch/x86/entry/common.c:50 arch/x86/entry/common.c:80)
> > >    entry_SYSCALL_64_after_hwframe (arch/x86/entry/entry_64.S:120)
> > >
> > > For the very first implementation of bhash2, there was a similar report
> > > hitting the same WARN_ON().  The fix was to update the bhash2 bucket when
> > > the kernel changes sk->sk_rcv_saddr from INADDR_ANY.  Then, syzbot has a
> > > repro, so we can indeed confirm that it no longer triggers the warning on
> > > the latest kernel.
> > >
> > >    https://lore.kernel.org/netdev/0000000000003f33bc05dfaf44fe@google.com/
> > >
> > > However, Mat reported at that time that there were at least two variants,
> > > the latter being the same as mine.
> > >
> > >    https://lore.kernel.org/netdev/4bae9df4-42c1-85c3-d350-119a151d29@linux.intel.com/
> > >    https://lore.kernel.org/netdev/23d8e9f4-016-7de1-9737-12c3146872ca@linux.intel.com/
> > >
> > > This week I started looking into this issue and finally figured out
> > > why we could not catch all cases with a single repro.
> > >
> > > Please see the source addresses of s2/s3 below after connect() fails.
> > > The s2 case is another variant of the first syzbot report, which has
> > > been already _fixed_.  And the s3 case is a repro for the issue that
> > > Mat and I saw.
> > >
> > >    # sysctl -w net.ipv4.tcp_syn_retries=1
> > >    net.ipv4.tcp_syn_retries = 1
> > >    # python3
> > >    >>> from socket import *
> > >    >>>
> > >    >>> s1 = socket()
> > >    >>> s1.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
> > >    >>> s1.bind(('0.0.0.0', 10000))
> > >    >>> s1.connect(('127.0.0.1', 10000))
> > >    >>>
> > >    >>> s2 = socket()
> > >    >>> s2.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
> > >    >>> s2.bind(('0.0.0.0', 10000))
> > >    >>> s2
> > >    <socket.socket fd=4, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>
> > >    >>>
> > >    >>> s2.connect(('127.0.0.1', 10000))
> > >    Traceback (most recent call last):
> > >      File "<stdin>", line 1, in <module>
> > >    OSError: [Errno 99] Cannot assign requested address
> > >    >>>
> > >    >>> s2
> > >    <socket.socket fd=4, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('127.0.0.1', 10000)>
> > >                                                                                                     > > >  ???
> > >    >>> s3 = socket()
> > >    >>> s3.setsockopt(SOL_SOCKET, SO_REUSEADDR, 1)
> > >    >>> s3.bind(('0.0.0.0', 10000))
> > >    >>> s3
> > >    <socket.socket fd=5, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>
> > >    >>>
> > >    >>> s3.connect(('0.0.0.1', 1))
> > >    Traceback (most recent call last):
> > >      File "<stdin>", line 1, in <module>
> > >    TimeoutError: [Errno 110] Connection timed out
> > >    >>>
> > >    >>> s3
> > >    <socket.socket fd=5, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>
> > >
> > > We can fire the WARN_ON() by s3.listen().
> > >
> > >    >>> s3.listen()
> > >    [ 1096.845905] ------------[ cut here ]------------
> > >    [ 1096.846290] WARNING: CPU: 0 PID: 209 at net/ipv4/inet_connection_sock.c:548 inet_csk_get_port+0x6bb/0x9e0
> > >
> > > In the s3 case, connect() resets sk->sk_rcv_saddr to INADDR_ANY without
> > > updating the bhash2 bucket; OTOH sk has the correct non-NULL bhash bucket.
> > > So, when we call listen() for s3, inet_csk_get_port() does not call
> > > inet_bind_hash() and the WARN_ON() for bhash2 fires.
> > >
> > >    if (!inet_csk(sk)->icsk_bind_hash)
> > >       inet_bind_hash(sk, tb, tb2, port);
> > >    WARN_ON(inet_csk(sk)->icsk_bind_hash != tb);
> > >    WARN_ON(inet_csk(sk)->icsk_bind2_hash != tb2);
> > >
> > > Here I think the s2 case also _should_ trigger WARN_ON().  The issue
> > > seems to be fixed, but it's just because we forgot to _fix_ the source
> > > address in error paths after inet6?_hash_connect() in tcp_v[46]_connect().
> > > (Of course, DCCP as well).
> > >
> > > In the s3 case, connect() falls into a different path.  We reach the
> > > sock_error label in __inet_stream_connect() and call sk_prot->disconnect(),
> > > which resets sk->sk_rcv_saddr.
> > >
> > > Then, there could be two subsequent issues.
> > >
> > >    * listen() leaks a newly allocated bhash2 bucket
> > >
> > >    * In inet_put_port(), inet_bhashfn_portaddr() computes a wrong hash for
> > >      inet_csk(sk)->icsk_bind2_hash, resulting in list corruption.
> > >
> > > We can fix these easily but it still leaves inconsistent sockets in bhash2,
> > > so we need to fix the root cause.
> > >
> > > As a side note, this issue only happens when we bind() only port before
> > > connect().  If we do not bind() port, tcp_set_state(sk, TCP_CLOSE) calls
> > > inet_put_port() and unhashes the sk from bhash2.
> > >
> > >
> > > At first, I applied the patch below so that both sk2 and sk3 trigger
> > > WARN_ON().  Then, I tried two approaches:
> > >
> > >    * Fix up the bhash2 entry when calling sk_rcv_saddr
> > >
> > >    * Change the bhash2 entry only when connect() succeeds
> > >
> > > The former does not work when we run out of memory.  When we change saddr
> > > from INADDR_ANY, inet_bhash2_update_saddr() could free (INADDR_ANY, port)
> > > bhash2 bucket.  Then, we possibly could not allocate it again when
> > > restoring saddr in the failure path.
> > >
> > > The latter does not work when a sk is in non-blocking mode.  In this case,
> > > a user might not call the second connect() to fix up the bhash2 bucket.
> > >
> > >    >>> s4 = socket()
> > >    >>> s4.bind(('0.0.0.0', 10000))
> > >    >>> s4.setblocking(False)
> > >    >>> s4
> > >    <socket.socket fd=3, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>
> > >
> > >    >>> s4.connect(('0.0.0.1', 1))
> > >    Traceback (most recent call last):
> > >      File "<stdin>", line 1, in <module>
> > >    BlockingIOError: [Errno 115] Operation now in progress
> > >    >>> s4
> > >    <socket.socket fd=3, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('10.0.2.15', 10000)>
> > >
> > > Also, the former approach does not work for the non-blocking case.  Let's
> > > say the second connect() fails.  What if we fail to allocate an INADDR_ANY
> > > bhash2 bucket?  We have to change saddr to INADDR_ANY but cannot.... but
> > > the connect() actually failed....??
> > >
> > >    >>> s4.connect(('0.0.0.1', 1))
> > >    Traceback (most recent call last):
> > >      File "<stdin>", line 1, in <module>
> > >    ConnectionRefusedError: [Errno 111] Connection refused
> > >    >>> s4
> > >    <socket.socket fd=3, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 10000)>
> > >
> > >
> > > Now, I'm thinking bhash2 bucket needs a refcnt not to be freed while
> > > refcnt is greater than 1.  And we need to change the conflict logic
> > > so that the kernel ignores empty bhash2 bucket.  Such changes could
> > > be big for the net tree, but the next LTS will likely be v6.1 which
> > > has bhash2.
> > >
> > > What do you think is the best way to fix the issue?
> >
> > Thanks for the repro script and the detailed analysis on the issue.  iiuc, this
> > is limited to the sk that was bind() to ADDR_ANY:port-1234 (or
> > !SOCK_BINDADDR_LOCK).

Yes, it's limited to !SOCK_BINDADDR_LOCK && SOCK_BINDPORT_LOCK, when
we don't reset snum but saddr.  We seem to need like ->put_addr() which
never fails even in the failure path.


> > Does it make sense to avoid adding the sk with
> > ADDR_ANY:port-1234 to bhash2 at all?  From inet_use_bhash2_on_bind(), it does
> > not seem ADDR_ANY will use bhash2.  or I have missed some cases?
> >
> 
> Thanks for your analysis and your repro instructions, Kuniyuki. And
> apologies for being absent from bhash2 these past ~6 weeks, I've been
> on PTO traveling and should have communicated that better. Thanks
> Martin for taking care of bhash2 during this time.

No worries :)


> For avoiding adding sockets with ADDR_ANY to the bhash2 hashtable, I
> think the issue is that other sockets need to detect whether there's a
> bind conflict against an ADDR_ANY socket on that port, so if ADDR_ANY
> is not hashed to bhash2, then on binds, we would have to iterate
> through the regular bhash table to check against ADDR_ANY, where the
> bhash table could be very long if there are many sockets bound to that
> port.

Right, inet_bhash2_addr_any_conflict() will not work then and it means
we cannot enjoy the very merit of bhash2.


> I intend to look deeply into this issue this weekend when I fly back
> home and have my dev environment accessible. Kuniyuki, you mentioned
> that the next LTS will likely be v6.1 which has bhash2 - do you know
> what the deadline for this is for getting the fix in by?

The bhash2 is already in the Linus' tree, so it will be released next
month until someone sends a revert, and we can send a fix anytime.

I think there is no deadline, but it should be better to send a fix
in this cycle.  My concern is that if the fix is big, it could be
queued for v6.2 and a bug could remain in LTS.

If bhash2 has the inconsistent state, we can bind a socket to a
would-conflict-address.

  >>> from socket import *
  >>> 
  >>> s1 = socket()
  >>> s1.bind(('0', 1000))
  >>> s1.connect(('localhost', 10))
  Traceback (most recent call last):
    File "<stdin>", line 1, in <module>
  ConnectionRefusedError: [Errno 111] Connection refused
  >>> s1
  <socket.socket fd=3, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('0.0.0.0', 1000)>

  ---
  0.0.0.0:1000   -> slot A -> tb A (should have 0.0.0.0:1000)

  10.0.2.15:1000 -> slot B -> tb B
                                `-> 0.0.0.0:1000
  ---

  >>> s2 = socket()
  >>> s2.bind(('10.0.2.16', 1000))
  >>> s2
  <socket.socket fd=4, family=AddressFamily.AF_INET, type=SocketKind.SOCK_STREAM, proto=0, laddr=('10.0.2.16', 1000)>

  ---
  0.0.0.0:1000   -> slot A -> tb A (no conflict)
  10.0.2.16:1000 -> slot C -> tb C (no conflict)
  ---

I don't know the real use cases for bind(INADDR_ANY, port) + connect(),
but I think this is not acceptable for LTS?

Powered by blists - more mailing lists