On Mon, 2022-11-21 at 16:05 +0100, Greg Kroah-Hartman wrote: > On Mon, Nov 21, 2022 at 09:03:18AM -0500, James Bottomley wrote: > > On Mon, 2022-11-21 at 12:05 +0100, Greg Kroah-Hartman wrote: > > > On Sun, Nov 20, 2022 at 10:14:26PM -0500, James Bottomley wrote: [...] > > > > I already explained in the email that sysfs contains APIs like > > > > simple_pin_... which are completely inimical to namespacing. > > > > > > Then how does the networking code handle the namespace stuff in > > > sysfs? That seems to work today, or am I missing something? > > > > have you actually tried? > > > > jejb@lingrow:~> sudo unshare --net bash > > lingrow:/home/jejb # ls /sys/class/net/ > > lo tun0 tun10 wlan0 > > lingrow:/home/jejb # ip link show > > 1: lo: <LOOPBACK> mtu 65536 qdisc noop state DOWN mode DEFAULT > > group > > default qlen 1000 > > link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00 > > > > So, as you see, I've entered a network namespace and ip link shows > > me the only interface I can see in that namespace (a down loopback) > > but sysfs shows me every interface on the system outside the > > namespace. > > Then all of the code in include/kobject_ns.h is not being used? We > have a whole kobject namespace set up for networking, I just assumed > they were using it. If not, I'm all for ripping it out. Hm, looking at the implementation, it seems to trigger off the superblock (meaning you have to remount inside a mount namespace) and it only works to control visibility in label based namespaces, so this does actually work jejb@lingrow:~/git/linux> sudo unshare --net --mount bash lingrow:/home/jejb # mount -t sysfs none /sys lingrow:/home/jejb # ls /sys/class/net/ lo The label based approach means that any given file can be shown in one and only one namespace, which works for net, but not much else (although it probably could be adapted). James
