diff mbox series

[mptcp-net] mptcp: prevent tcp diag from closing listener subflows

Message ID f6371f65b1ebabc0f308528494ec39c9bdf786fd.1702631517.git.pabeni@redhat.com (mailing list archive)
State Accepted, archived
Commit 5913b5d75d1f8a286057f25abe2a51ff84109178
Delegated to: Matthieu Baerts
Headers show
Series [mptcp-net] mptcp: prevent tcp diag from closing listener subflows | expand

Checks

Context Check Description
matttbe/KVM_Validation__normal__except_selftest_mptcp_join_ success Success! ✅
matttbe/KVM_Validation__normal__only_selftest_mptcp_join_ success Success! ✅
matttbe/KVM_Validation__debug__except_selftest_mptcp_join_ success Success! ✅
matttbe/KVM_Validation__debug__only_selftest_mptcp_join_ warning Unstable: 1 failed test(s): selftest_mptcp_join
matttbe/build success Build and static analysis OK
matttbe/checkpatch warning total: 0 errors, 1 warnings, 0 checks, 31 lines checked

Commit Message

Paolo Abeni Dec. 15, 2023, 9:12 a.m. UTC
The MPTCP protocol does not expect that any other entity could change
the first subflow status when such socket is listening.
Unfortunately the TCP diag interface allows aborting any TCP socket,
including MPTCP listeners subflows. As reported by syzbot, that trigger
a WARN() and could lead to later bigger trouble.

The MPTCP protocol needs to do some MPTCP-level cleanup actions to
properly shutdown the listener. To keep the fix simple, prevent
entirely the diag interface from stopping such listeners.

We could refine the diag callback in a later, larger patch targeting
net-next.

Fixes: 57fc0f1ceaa4 ("mptcp: ensure listener is unhashed before updating the sk status")
Closes: https://lore.kernel.org/netdev/0000000000004f4579060c68431b@google.com/
Reported-by: syzbot+5a01c3a666e726bc8752@syzkaller.appspotmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
---
 net/mptcp/subflow.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

Comments

MPTCP CI Dec. 15, 2023, 11:10 a.m. UTC | #1
Hi Paolo,

Thank you for your modifications, that's great!

Our CI did some validations and here is its report:

- KVM Validation: normal (except selftest_mptcp_join):
  - Success! ✅:
  - Task: https://cirrus-ci.com/task/6646535530741760
  - Summary: https://api.cirrus-ci.com/v1/artifact/task/6646535530741760/summary/summary.txt

- KVM Validation: normal (only selftest_mptcp_join):
  - Success! ✅:
  - Task: https://cirrus-ci.com/task/5666210654715904
  - Summary: https://api.cirrus-ci.com/v1/artifact/task/5666210654715904/summary/summary.txt

- KVM Validation: debug (except selftest_mptcp_join):
  - Success! ✅:
  - Task: https://cirrus-ci.com/task/5731741856432128
  - Summary: https://api.cirrus-ci.com/v1/artifact/task/5731741856432128/summary/summary.txt

- KVM Validation: debug (only selftest_mptcp_join):
  - Unstable: 1 failed test(s): selftest_mptcp_join 
Mat Martineau Dec. 16, 2023, 12:47 a.m. UTC | #2
On Fri, 15 Dec 2023, Paolo Abeni wrote:

> The MPTCP protocol does not expect that any other entity could change
> the first subflow status when such socket is listening.
> Unfortunately the TCP diag interface allows aborting any TCP socket,
> including MPTCP listeners subflows. As reported by syzbot, that trigger
> a WARN() and could lead to later bigger trouble.
>
> The MPTCP protocol needs to do some MPTCP-level cleanup actions to
> properly shutdown the listener. To keep the fix simple, prevent
> entirely the diag interface from stopping such listeners.
>
> We could refine the diag callback in a later, larger patch targeting
> net-next.
>
> Fixes: 57fc0f1ceaa4 ("mptcp: ensure listener is unhashed before updating the sk status")
> Closes: https://lore.kernel.org/netdev/0000000000004f4579060c68431b@google.com/
> Reported-by: syzbot+5a01c3a666e726bc8752@syzkaller.appspotmail.com
> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> ---
> net/mptcp/subflow.c | 13 +++++++++++++
> 1 file changed, 13 insertions(+)
>
> diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
> index 64302a1ac306..1f98bec264a6 100644
> --- a/net/mptcp/subflow.c
> +++ b/net/mptcp/subflow.c
> @@ -1981,6 +1981,17 @@ static void tcp_release_cb_override(struct sock *ssk)
> 	tcp_release_cb(ssk);
> }
>
> +static int tcp_abort_override(struct sock *ssk, int err)
> +{
> +	/* closing a listener subflow requires a great deal of care.
> +	 * keep it simple and just prevent such operation
> +	 */
> +	if (inet_sk_state_load(ssk) == TCP_LISTEN)
> +		return -EINVAL;
> +
> +	return tcp_abort(ssk, err);

Hi Paolo -

I looked around the code to see if there might be issues with an 
unexpected tcp_abort on non-listening subflow sockets. Closest code I 
found was mptcp_subflow_reset(), and the main difference seems to be 
giving the msk a chance to clean up:

 	if (!test_and_set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &mptcp_sk(sk)->flags))
 		mptcp_schedule_work(sk);

Worthwhile to add that in this override function?

- Mat

> +}
> +
> static struct tcp_ulp_ops subflow_ulp_ops __read_mostly = {
> 	.name		= "mptcp",
> 	.owner		= THIS_MODULE,
> @@ -2025,6 +2036,7 @@ void __init mptcp_subflow_init(void)
>
> 	tcp_prot_override = tcp_prot;
> 	tcp_prot_override.release_cb = tcp_release_cb_override;
> +	tcp_prot_override.diag_destroy = tcp_abort_override;
>
> #if IS_ENABLED(CONFIG_MPTCP_IPV6)
> 	/* In struct mptcp_subflow_request_sock, we assume the TCP request sock
> @@ -2060,6 +2072,7 @@ void __init mptcp_subflow_init(void)
>
> 	tcpv6_prot_override = tcpv6_prot;
> 	tcpv6_prot_override.release_cb = tcp_release_cb_override;
> +	tcpv6_prot_override.diag_destroy = tcp_abort_override;
> #endif
>
> 	mptcp_diag_subflow_init(&subflow_ulp_ops);
> -- 
> 2.42.0
>
>
>
Paolo Abeni Dec. 21, 2023, 10 a.m. UTC | #3
On Fri, 2023-12-15 at 16:47 -0800, Mat Martineau wrote:
> On Fri, 15 Dec 2023, Paolo Abeni wrote:
> 
> > The MPTCP protocol does not expect that any other entity could change
> > the first subflow status when such socket is listening.
> > Unfortunately the TCP diag interface allows aborting any TCP socket,
> > including MPTCP listeners subflows. As reported by syzbot, that trigger
> > a WARN() and could lead to later bigger trouble.
> > 
> > The MPTCP protocol needs to do some MPTCP-level cleanup actions to
> > properly shutdown the listener. To keep the fix simple, prevent
> > entirely the diag interface from stopping such listeners.
> > 
> > We could refine the diag callback in a later, larger patch targeting
> > net-next.
> > 
> > Fixes: 57fc0f1ceaa4 ("mptcp: ensure listener is unhashed before updating the sk status")
> > Closes: https://lore.kernel.org/netdev/0000000000004f4579060c68431b@google.com/
> > Reported-by: syzbot+5a01c3a666e726bc8752@syzkaller.appspotmail.com
> > Signed-off-by: Paolo Abeni <pabeni@redhat.com>
> > ---
> > net/mptcp/subflow.c | 13 +++++++++++++
> > 1 file changed, 13 insertions(+)
> > 
> > diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
> > index 64302a1ac306..1f98bec264a6 100644
> > --- a/net/mptcp/subflow.c
> > +++ b/net/mptcp/subflow.c
> > @@ -1981,6 +1981,17 @@ static void tcp_release_cb_override(struct sock *ssk)
> > 	tcp_release_cb(ssk);
> > }
> > 
> > +static int tcp_abort_override(struct sock *ssk, int err)
> > +{
> > +	/* closing a listener subflow requires a great deal of care.
> > +	 * keep it simple and just prevent such operation
> > +	 */
> > +	if (inet_sk_state_load(ssk) == TCP_LISTEN)
> > +		return -EINVAL;
> > +
> > +	return tcp_abort(ssk, err);
> 
> Hi Paolo -
> 
> I looked around the code to see if there might be issues with an 
> unexpected tcp_abort on non-listening subflow sockets. Closest code I 
> found was mptcp_subflow_reset(), and the main difference seems to be 
> giving the msk a chance to clean up:
> 
>  	if (!test_and_set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &mptcp_sk(sk)->flags))
>  		mptcp_schedule_work(sk);
> 
> Worthwhile to add that in this override function?

Reporting there the discussion from the last mtg for completeness.
Async closing of mptcp listener via the worker has proven to be quite
complex and bug prone. In this case it would probably be safe, but
given the troubled git history of mptcp_check_listen_stop(), I would
keep this version for net, and do the complete thing for net-next, with
no rush.

Cheers,

Paolo
Mat Martineau Dec. 21, 2023, 5:25 p.m. UTC | #4
On Thu, 21 Dec 2023, Paolo Abeni wrote:

> On Fri, 2023-12-15 at 16:47 -0800, Mat Martineau wrote:
>> On Fri, 15 Dec 2023, Paolo Abeni wrote:
>>
>>> The MPTCP protocol does not expect that any other entity could change
>>> the first subflow status when such socket is listening.
>>> Unfortunately the TCP diag interface allows aborting any TCP socket,
>>> including MPTCP listeners subflows. As reported by syzbot, that trigger
>>> a WARN() and could lead to later bigger trouble.
>>>
>>> The MPTCP protocol needs to do some MPTCP-level cleanup actions to
>>> properly shutdown the listener. To keep the fix simple, prevent
>>> entirely the diag interface from stopping such listeners.
>>>
>>> We could refine the diag callback in a later, larger patch targeting
>>> net-next.
>>>
>>> Fixes: 57fc0f1ceaa4 ("mptcp: ensure listener is unhashed before updating the sk status")
>>> Closes: https://lore.kernel.org/netdev/0000000000004f4579060c68431b@google.com/
>>> Reported-by: syzbot+5a01c3a666e726bc8752@syzkaller.appspotmail.com
>>> Signed-off-by: Paolo Abeni <pabeni@redhat.com>
>>> ---
>>> net/mptcp/subflow.c | 13 +++++++++++++
>>> 1 file changed, 13 insertions(+)
>>>
>>> diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
>>> index 64302a1ac306..1f98bec264a6 100644
>>> --- a/net/mptcp/subflow.c
>>> +++ b/net/mptcp/subflow.c
>>> @@ -1981,6 +1981,17 @@ static void tcp_release_cb_override(struct sock *ssk)
>>> 	tcp_release_cb(ssk);
>>> }
>>>
>>> +static int tcp_abort_override(struct sock *ssk, int err)
>>> +{
>>> +	/* closing a listener subflow requires a great deal of care.
>>> +	 * keep it simple and just prevent such operation
>>> +	 */
>>> +	if (inet_sk_state_load(ssk) == TCP_LISTEN)
>>> +		return -EINVAL;
>>> +
>>> +	return tcp_abort(ssk, err);
>>
>> Hi Paolo -
>>
>> I looked around the code to see if there might be issues with an
>> unexpected tcp_abort on non-listening subflow sockets. Closest code I
>> found was mptcp_subflow_reset(), and the main difference seems to be
>> giving the msk a chance to clean up:
>>
>>  	if (!test_and_set_bit(MPTCP_WORK_CLOSE_SUBFLOW, &mptcp_sk(sk)->flags))
>>  		mptcp_schedule_work(sk);
>>
>> Worthwhile to add that in this override function?
>
> Reporting there the discussion from the last mtg for completeness.
> Async closing of mptcp listener via the worker has proven to be quite
> complex and bug prone. In this case it would probably be safe, but
> given the troubled git history of mptcp_check_listen_stop(), I would
> keep this version for net, and do the complete thing for net-next, with
> no rush.
>

Thanks Paolo, sounds like a good plan.

Reviewed-by: Mat Martineau <martineau@kernel.org>
Matthieu Baerts Dec. 22, 2023, 10:44 a.m. UTC | #5
Hi Paolo, Mat,

On 15/12/2023 10:12, Paolo Abeni wrote:
> The MPTCP protocol does not expect that any other entity could change
> the first subflow status when such socket is listening.
> Unfortunately the TCP diag interface allows aborting any TCP socket,
> including MPTCP listeners subflows. As reported by syzbot, that trigger
> a WARN() and could lead to later bigger trouble.
> 
> The MPTCP protocol needs to do some MPTCP-level cleanup actions to
> properly shutdown the listener. To keep the fix simple, prevent
> entirely the diag interface from stopping such listeners.
> 
> We could refine the diag callback in a later, larger patch targeting
> net-next.

Thank you for this patch and the review!

> Fixes: 57fc0f1ceaa4 ("mptcp: ensure listener is unhashed before updating the sk status")
> Closes: https://lore.kernel.org/netdev/0000000000004f4579060c68431b@google.com/
> Reported-by: syzbot+5a01c3a666e726bc8752@syzkaller.appspotmail.com

(I just inverted these two lines to make checkpatch happy ;) )
(and it looks like b4 automatically adds '<>' around the email address)

Now in our tree (fixes for -net):

New patches for t/upstream-net and t/upstream:
- 5913b5d75d1f: mptcp: prevent tcp diag from closing listener subflows
- Results: 9293d696010b..b7f484870b63 (export-net)
- Results: 4720030d95d3..b52e904674ac (export)

(Tests are not in progress)

Cheers,
Matt
diff mbox series

Patch

diff --git a/net/mptcp/subflow.c b/net/mptcp/subflow.c
index 64302a1ac306..1f98bec264a6 100644
--- a/net/mptcp/subflow.c
+++ b/net/mptcp/subflow.c
@@ -1981,6 +1981,17 @@  static void tcp_release_cb_override(struct sock *ssk)
 	tcp_release_cb(ssk);
 }
 
+static int tcp_abort_override(struct sock *ssk, int err)
+{
+	/* closing a listener subflow requires a great deal of care.
+	 * keep it simple and just prevent such operation
+	 */
+	if (inet_sk_state_load(ssk) == TCP_LISTEN)
+		return -EINVAL;
+
+	return tcp_abort(ssk, err);
+}
+
 static struct tcp_ulp_ops subflow_ulp_ops __read_mostly = {
 	.name		= "mptcp",
 	.owner		= THIS_MODULE,
@@ -2025,6 +2036,7 @@  void __init mptcp_subflow_init(void)
 
 	tcp_prot_override = tcp_prot;
 	tcp_prot_override.release_cb = tcp_release_cb_override;
+	tcp_prot_override.diag_destroy = tcp_abort_override;
 
 #if IS_ENABLED(CONFIG_MPTCP_IPV6)
 	/* In struct mptcp_subflow_request_sock, we assume the TCP request sock
@@ -2060,6 +2072,7 @@  void __init mptcp_subflow_init(void)
 
 	tcpv6_prot_override = tcpv6_prot;
 	tcpv6_prot_override.release_cb = tcp_release_cb_override;
+	tcpv6_prot_override.diag_destroy = tcp_abort_override;
 #endif
 
 	mptcp_diag_subflow_init(&subflow_ulp_ops);