Bug 920032

(In reply to comment #0)
> Description of problem:
> 
> From vaughan <vaughan.cao@xxxxxxxxxx>
> 
> I guess I found a bug -- hypervkvpd running alone without hv_utils
> loaded encounters segfault when service cgred start on RHEL6.4. It
> occurs with both 0.8 and 0.9, regardless of i686 or x86_64.
> 
> I read in hv_kvp_daemon.c that the user mode componet should first
> registers with the kernel component.
> But in my test, the hand shake phase has been ignored.
> Things happens like this:
> hv_utils.ko and hv_vmbus.ko is not loaded, start hypervkvpd is fine.
> Then, I start cgred with the default configuration. cgroup also use
> NETLINK_CONNECTOR protocol and send messages with cb_id{1,1}. Hypervkvpd
> receive messages without checking their source. Some messages with
> cb_id{1,1} were receviced and blindly interpreted as hv_kvp_msg.
> Since the hand_shake check is as below:
> if ((in_hand_shake) && (op == KVP_OP_REGISTER1)) {
> ...
> continue;
> }
> //handle kvp messages
> switch (op) { ... }
> Register phase is also skipped.
> Everytime the KVP_OP_SET opcode is reached, kvp_key_add_or_modify() is
> invoked with an very large key_size. After several iterations, segfault
> occurs in memcpy(record[i].key, key, key_size) (key_size is negative now).
> 
> I'm not very familiar with connector. But I ran the sample in
> Documentation/connector/ and found that a NETLINK_CONNECTOR socket would
> always some messages with cb_id{1,1}. So blindly suppose all messages
> are kvp_msg is not correct. hypervkvpd should check the source of
> messages and perhaps even check nlmsg_type in the nlmsghdr.

I did some testing. It doesn't matter if hv_utils is unloaded or not.
Handling of incoming NetLink messages is not very fortunate and should
be changed for sure.

The registration phase seems to be mandatory. Kernel part of the hyper-v kvp
functionality is depending on user-space daemon to register with it. What's
strange is that the kernel module registers with the daemon only for the first
time. If daemon is restarted, then the registration with kernel module is
never completed. I think this might have something to do with Bug #886781.

I have some ideas, but I need to test them first.

The problem was in setting sockaddr_nl.nl_groups to CN_KVP_IDX of value "9".
Since nl_groups is a bit mask, hypervkvpd was getting Netlink messages also
from group number "1".

I changed sockaddr_nl.nl_groups to "0" (not subscribe to any multicast group)
and rather subscribe to CN_KVP_IDX group using setsockopt().

I also added check for nlmsg_type in received Netlink message header so now
daemon processes the message only if it's type is NLMSG_DONE. Otherwise the
message is ignored.

I tested the solution on Windows Server 2012 Guest with RHEL 6.4 Guest and
everything worked fine. Host was able to get information from Guest and
when starting/restarting cgred service for couple of times hypervkvpd kept
running just fine.

I sent patches to upstream:
http://driverdev.linuxdriverproject.org/pipermail/devel/2013-March/036284.html
http://driverdev.linuxdriverproject.org/pipermail/devel/2013-March/036285.html
http://driverdev.linuxdriverproject.org/pipermail/devel/2013-March/036286.html

Patches were accepted by upstream.

verify this bug in x86_64 and i386 guest, no segfault occur,  the kernel of which is 2.6.32-415.0.1.el6.x86_64/2.6.32-415.0.1.el6.i686 , the package of hypervkvpd is hypervkvpd-0-0.12.el6.x86_64/hypervkvpd-0-0.12.el6.i686

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

http://rhn.redhat.com/errata/RHBA-2013-1539.html