On Mon, Nov 21, 2022 at 12:33:55PM -0500, James Bottomley wrote: > 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). Great, thanks for verifying it works properly. No other subsystem other than networking has cared about adding support for namespaces to their sysfs representations. But the base logic is all there if they want to do so. thanks, greg k-h
