Hi Michael, Sharwan and all folks, I finally found the correct API to do so in kernel space. Following Sharwan's tip of getting namespace from device, I did more grep through the device code, so it leads to this function call: static struct net *get_net_ns_by_pid(pid_t pid) in file net/core/rtnetlink.c, for 2.6.28.9. Indeed network namespace is stored in structure nsproxy. Thanks, all. Appreciate your feedback. Haibin On Wed, Apr 8, 2009 at 3:01 PM, Haibin Wang <haibin123us@xxxxxxxxx> wrote: > Hi Sharwan, > > Thank you very much for the answer. Yes, I read the network namespace > code, you could get namespace from sk by calling sock_net(sk), and it > is also available from device. > > For our case, it is how to get the namespace in the first place, so we > can match against open socks and/or devices. Since you mentioned that > process fills sk_buff, then how does the process know which namespace > it is in then? This is the answer I am looking for. > > In our case, the user space code tells kernel code what IP/port it is > interested, ( net namespace implied ?), so I am wondering where I > could get this namespace, implied by user space code, or available > from the process structure in kernel space. > > Thank you, guys. > Haibin > > On Wed, Apr 8, 2009 at 12:26 AM, Sharwan Joram <sharwan.joram@xxxxxxxxx> wrote: >> A namespace parameter is obtained in every function which creates the >> socket. This namespace field is present in the sk_buff structure which is >> filled up either by the process which is creating the socket or in second >> case from the device it is received. >>>> > > >>> >> >>> >> A question on how to get the current network namespace. >>> >> >>> >> I am porting a driver from 2.6.18 to 2.6.28. We do protocol analysis >>> >> by the 5 tuples: src IP, dst IP, src port, dst port and transportation >>> >> protocol. We have code in usr space to pass in these parameters, then >>> >> kernel space code do search through udp_hash and tcp_hash to locate >>> >> the exact socks. Since 2.6.18 doesn't support network namespace, this >>> >> works fine. >>> >> >>> >> But as 2.6.28 supports network namespace, search through udp_hash and >>> >> tcp_hash requires a sixth parameter -- network namespace as well. So >>> >> the question is, how to get current network namespace in this case? Is >>> >> it available from user space code and passed to kernel code, or there >>> >> is function to get it in kernel space? >>> >> >>> >> It seems I could go by by using the system init_net, but I want to be >>> >> new feature friendly at least. >>> > >>> > Netmark namespace means that different user space processes have a different >>> > view of the network (used e.g. by vservers). I do not know what will happen, >>> > if you pass the init_net (default namespace), if the socket is created >>> > in a >>> > different namespace. > -- To unsubscribe from this list: send an email with "unsubscribe kernelnewbies" to ecartis@xxxxxxxxxxxx Please read the FAQ at http://kernelnewbies.org/FAQ
