lkml.org 
[lkml]   [2018]   [Mar]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: netns: send uevent messages
From
Date
On 16.03.2018 02:46, Christian Brauner wrote:
> On Thu, Mar 15, 2018 at 05:14:13PM +0300, Kirill Tkhai wrote:
>> On 15.03.2018 16:39, Christian Brauner wrote:
>>> On Thu, Mar 15, 2018 at 12:47:30PM +0300, Kirill Tkhai wrote:
>>>> CC Andrey Vagin
>>>
>>> Hey Kirill,
>>>
>>> Thanks for CCing Andrey.
>>>
>>>>
>>>> On 15.03.2018 03:12, Christian Brauner wrote:
>>>>> This patch adds a receive method to NETLINK_KOBJECT_UEVENT netlink sockets
>>>>> to allow sending uevent messages into the network namespace the socket
>>>>> belongs to.
>>>>>
>>>>> Currently non-initial network namespaces are already isolated and don't
>>>>> receive uevents. There are a number of cases where it is beneficial for a
>>>>> sufficiently privileged userspace process to send a uevent into a network
>>>>> namespace.
>>>>>
>>>>> One such use case would be debugging and fuzzing of a piece of software
>>>>> which listens and reacts to uevents. By running a copy of that software
>>>>> inside a network namespace, specific uevents could then be presented to it.
>>>>> More concretely, this would allow for easy testing of udevd/ueventd.
>>>>>
>>>>> This will also allow some piece of software to run components inside a
>>>>> separate network namespace and then effectively filter what that software
>>>>> can receive. Some examples of software that do directly listen to uevents
>>>>> and that we have in the past attempted to run inside a network namespace
>>>>> are rbd (CEPH client) or the X server.
>>>>>
>>>>> Implementation:
>>>>> The implementation has been kept as simple as possible from the kernel's
>>>>> perspective. Specifically, a simple input method uevent_net_rcv() is added
>>>>> to NETLINK_KOBJECT_UEVENT sockets which completely reuses existing
>>>>> af_netlink infrastructure and does neither add an additional netlink family
>>>>> nor requires any user-visible changes.
>>>>>
>>>>> For example, by using netlink_rcv_skb() we can make use of existing netlink
>>>>> infrastructure to report back informative error messages to userspace.
>>>>>
>>>>> Furthermore, this implementation does not introduce any overhead for
>>>>> existing uevent generating codepaths. The struct netns gets a new uevent
>>>>> socket member that records the uevent socket associated with that network
>>>>> namespace. Since we record the uevent socket for each network namespace in
>>>>> struct net we don't have to walk the whole uevent socket list.
>>>>> Instead we can directly retrieve the relevant uevent socket and send the
>>>>> message. This keeps the codepath very performant without introducing
>>>>> needless overhead.
>>>>>
>>>>> Uevent sequence numbers are kept global. When a uevent message is sent to
>>>>> another network namespace the implementation will simply increment the
>>>>> global uevent sequence number and append it to the received uevent. This
>>>>> has the advantage that the kernel will never need to parse the received
>>>>> uevent message to replace any existing uevent sequence numbers. Instead it
>>>>> is up to the userspace process to remove any existing uevent sequence
>>>>> numbers in case the uevent message to be sent contains any.
>>>>>
>>>>> Security:
>>>>> In order for a caller to send uevent messages to a target network namespace
>>>>> the caller must have CAP_SYS_ADMIN in the owning user namespace of the
>>>>> target network namespace. Additionally, any received uevent message is
>>>>> verified to not exceed size UEVENT_BUFFER_SIZE. This includes the space
>>>>> needed to append the uevent sequence number.
>>>>>
>>>>> Testing:
>>>>> This patch has been tested and verified to work with the following udev
>>>>> implementations:
>>>>> 1. CentOS 6 with udevd version 147
>>>>> 2. Debian Sid with systemd-udevd version 237
>>>>> 3. Android 7.1.1 with ueventd
>>>>>
>>>>> Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com>
>>>>> ---
>>>>> include/net/net_namespace.h | 1 +
>>>>> lib/kobject_uevent.c | 88 ++++++++++++++++++++++++++++++++++++++++++++-
>>>>> 2 files changed, 88 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/include/net/net_namespace.h b/include/net/net_namespace.h
>>>>> index f306b2aa15a4..467bde763a9b 100644
>>>>> --- a/include/net/net_namespace.h
>>>>> +++ b/include/net/net_namespace.h
>>>>> @@ -78,6 +78,7 @@ struct net {
>>>>>
>>>>> struct sock *rtnl; /* rtnetlink socket */
>>>>> struct sock *genl_sock;
>>>>> + struct sock *uevent_sock; /* uevent socket */
>>>>
>>>> Since you add this per-net uevent_sock pointer, currently existing iterations in uevent_net_exit()
>>>> become to look confusing. There are:
>>>>
>>>> mutex_lock(&uevent_sock_mutex);
>>>> list_for_each_entry(ue_sk, &uevent_sock_list, list) {
>>>> if (sock_net(ue_sk->sk) == net)
>>>> goto found;
>>>> }
>>>>
>>>> Can't we make a small cleanup in lib/kobject_uevent.c after this change
>>>> and before the main part of the patch goes?
>>>
>>> Hm, not sure. It seems it makes sense to maintain them in a separate
>>> list. Looks like this lets us keep the locking simpler. Otherwise we
>>> would have to do something like for_each_net() and it seems that this
>>> would force us to use rntl_{un}lock().
>>
>> I'm about:
>>
>> mutex_lock();
>> list_del(net->ue_sk->list);
>> mutex_unlock();
>> kfree();
>>
>> Thus we avoid iterations in uevent_net_exit().
>
> Ah right, but I only added struct sock *uevent_sock to struct net, not
> struct uevent_sock. Thus there's no list member in there. I mainly only
> added struct sock to keep it clean an aligned with the other sock
> member. We can revisit this later though.

In my opinion, we shouldn't leave the code in inconsistent state just because
of it's not interesting for this patch. There are two logical changes:
1)add fast-access pointer to struct net
2)implement single-net uevent

It's easy to do 1), and this may help further reader not to be confused
by the inconsistency, so I don't see a reason why we shouldn't do that.

>>
>>>>
>>>>>
>>>>> struct list_head dev_base_head;
>>>>> struct hlist_head *dev_name_head;
>>>>> diff --git a/lib/kobject_uevent.c b/lib/kobject_uevent.c
>>>>> index 9fe6ec8fda28..10b2144b9fc3 100644
>>>>> --- a/lib/kobject_uevent.c
>>>>> +++ b/lib/kobject_uevent.c
>>>>> @@ -25,6 +25,7 @@
>>>>> #include <linux/uuid.h>
>>>>> #include <linux/ctype.h>
>>>>> #include <net/sock.h>
>>>>> +#include <net/netlink.h>
>>>>> #include <net/net_namespace.h>
>>>>>
>>>>>
>>>>> @@ -602,12 +603,94 @@ int add_uevent_var(struct kobj_uevent_env *env, const char *format, ...)
>>>>> EXPORT_SYMBOL_GPL(add_uevent_var);
>>>>>
>>>>> #if defined(CONFIG_NET)
>>>>> +static int uevent_net_send(struct sock *usk, struct sk_buff *skb,
>>>>> + struct netlink_ext_ack *extack)
>>>>> +{
>>>>> + int ret;
>>>>> + u64 seqnum;
>>>>> + /* u64 to chars: 2^64 - 1 = 21 chars */
>>>>
>>>> 18446744073709551615 -- 20 chars (+1 '\0'). Comment makes me think
>>>> we forgot +1 in buf declaration.
>>>
>>> sizeof("SEQNUM=") will include the '\0' pointer in contrast to
>>> strlen("SEQNUM=") so this is correct if I'm not completely mistaken.
>>
>> The code is OK, I'm worrying about comment. But I've missed this sizeof().
>> So, there is only 1 bytes excess allocated as 0xFFFFFFFFFFFFFFFF=18446744073709551615
>> Not so important...
>>
>>>>
>>>>> + char buf[sizeof("SEQNUM=") + 21];
>>>>> + struct sk_buff *skbc;
>>>>> +
>>>>> + /* bump sequence number */
>>>>> + mutex_lock(&uevent_sock_mutex);
>>>>> + seqnum = ++uevent_seqnum;
>>>>> + mutex_unlock(&uevent_sock_mutex);
>>>>
>>>> Commit 7b60a18da393 from Andrew Vagin says:
>>>>
>>>> uevent: send events in correct order according to seqnum (v3)
>>>>
>>>> The queue handling in the udev daemon assumes that the events are
>>>> ordered.
>>>>
>>>> Before this patch uevent_seqnum is incremented under sequence_lock,
>>>> than an event is send uner uevent_sock_mutex. I want to say that code
>>>> contained a window between incrementing seqnum and sending an event.
>>>>
>>>> This patch locks uevent_sock_mutex before incrementing uevent_seqnum.
>>>> Signed-off-by: Andrew Vagin <avagin@openvz.org>
>>>>
>>>> After this change the order will be lost. Also the rest of namespaces
>>>> will have holes in uevent numbers, as they won't receive a number sent
>>>> to specific namespace.
>>>
>>> Afaict from looking into udevd when I wrote the patch it only cares
>>> about numbers being ordered (which is also what Andrey's patch states)
>>> not that they are sequential so holes should be fine. udevd will use
>>> the DEVPATH to determine whether the sequence number of the current
>>> uevent should be used as "even->delaying_seqnum" number. All that
>>> matters is that it is greater than the previous one. About the ordering,
>>> if that's an issue then we should simply do what Andrey has been doing
>>> for kobject_uevent_env() and extend the lock being held until after we
>>> sent out the uevent. Since we're not serving all listeners but only
>>> ever one this is also way more lightweight then kobject_uevent_env().
>>
>> Yes, extending the lock to netlink_broadcast() should fix the problem.
>
> Yep, send another version of the patch but forgot to CC you, sorry:
> https://lkml.org/lkml/2018/3/15/1098
>
> Thanks!
> Christian
>
>>
>>>>
>>>> It seems we should made uevent_seqnum per-net to solve this problem.
>>>
>>> Yes, Eric and I have been discussing this already. The idea was to do
>>> this in a follow-up patch to keep this patch as simple as possible. If
>>> we agree that it would make sense right away then I will dig out the
>>> corresponding patch.
>>> It would basically just involve giving each struct net a uevent_seqnum
>>> member.
>>
>> pernet seqnum may have a sense if we introduce per-net locks to protect it.
>> If there is single mutex, it does not matter either they are pernet or not.
>> Per-net mutex may be useful only if we have many single-net messages like
>> you introduce in this patch, or if we are going to filter net in existing
>> kobject_uevent_net_broadcast() (by user_ns?!) in the future.
>>
>> Kirill

\
 
 \ /
  Last update: 2018-03-16 09:25    [W:0.044 / U:0.720 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site