WSL: Failing (and invisible in strace) clone syscall related to privileged namespaces and chrome's setuid sandbox
Build 16215
We already knew that chromium couldn’t be launched with its usual sandbox due to the fact that user namespaces aren’t implemented. I assumed privileged namespaces were not completely implemented either, though I had heard some of the devs say that privileged namespaces as a concept already exist in the underlying native NT context.
Chromium also includes a setuid sandbox implementation for kernels that don’t implement user namespaces (notably Arch, though I used a clean Ubuntu Xenial image to do this test), and you can launch chromium in this way with the flag --disable-namespace-sandbox, which disables the user namespace sandbox but enables the classic setuid sandbox.
In the user namespace sandbox case, we also get a line in between that actually shows the failing syscall:
clone(child_stack=0x7ffff8384930, flags=CLONE_NEWUSER|SIGCHLD) = -1 EINVAL (Invalid argument)
Now the weird part is this: Let’s try to run that without an strace first:
$ google-chrome --headless --disable-gpu --disable-namespace-sandbox
clone: Invalid argument
we get (among other errors, I know) that there is a clone syscall receiving an invalid parameter or option.
But when you look into the strace, there’s no EINVAL on a Clone syscall, so somewhere some clone syscall is failing on unimplemented surface but is also failing to be picked up in the strace, unlike in the namespace sandbox case.
Straces: for setuid sandbox: suidsandbox.txt
for user namespaces sandbox: unamespaces.txt
About this issue
- Original URL
- State: closed
- Created 9 years ago
- Comments: 38 (11 by maintainers)
@therealkenc - Well would you look at that…
The change to add CLONE_FS support is pretty small so I think I can squeeze it in for Fall Creator’s update. As always, thanks @therealkenc for all the help!
@therealkenc - I liked your idea of stubbing netlink support so I decided to see what would happen if I allowed creation of NETLINK_KOBJECT_UEVENT netlink sockets that will never receive any messages. With that change and the CLONE_FS change I above I’m able to launch chromium without any workarounds.
Unfortunately I can’t do a similar think for all netlink families. For example, allowing creation of NETLINK_AUDIT sockets but don’t actually support the messages opens worms (sudo, adduser, passwd all start failing).
I’m submitting a CR for this change now.
@therealkenc and @fpqc - Keep digging, you might find some surprises… 😃
@fpqc - no that clone call still fails but chrome appears to recover from the error.
Awesome. Fun, no? Now all you need to do is track why VSCode doesn’t launch (toy electron apps are fine). Give an inch, they’ll ask for a mile 😜.
@therealkenc - Dug into this a bit this afternoon. Once I implemented the CLONE_FS flag chromium makes it a bit further, but ultimately fails when trying to talk to udev over netlink sockets (I think).
CLONE_NEWUTS,CLONE_NEWIPC,CLONE_NEWPID,CLONE_NEWNSall work. I am actually a little embarrassed I didn’t notice. I have not triedCLONE_NEWNETbecause it is academic in the context of WSL.Need some more bits for
systemdthough. I did a short run at it last night but my heart wasn’t in it. But take that over to #994.From your first post, the issue here was and remains:
…full stop…
Yep. Funny thing, I think maybe it’s been there for a while. And to answer your next question, I’m looking into it.
Amusingly enough @benhillis’ work on #962 moved this forward. 16251 now gets over the
execve(). Actually if I was a little quicker on the uptake I would have thought of #962 when I saw theEPERM. Same mode of failure.Having missed that, I did get the “subsequent
clone()” failure hypothesis right.Child stack of zeroActually it wasCLONE_FS. Don’t think I have seen that before.So @fpqc now ya know.
Since I am here, I have to underline this ticket has nada to do with “namespaces”, even though the word is used 18 times in this thread so far including in the issue title (so says
ctrl-f). The whole point of--disable-namespace-sandboxis you are on a kernel without them. Eye on the ball.Hmmm. I assumed
CLONE_FSwas implemented (it shows up in quick issue search). You get over theudevnetlink socket problem with libudev-stub, natch 😉.I can’t underline how easy it is to stub the netlink socket part in the WSL kernel if you guys feel like it. My userspace stub just opens a AF_UNIX dgram socket here. Which is totally the wrong family, but it isn’t like userspace knows any different. The next thing userspace does with the socket descriptor is
select(), which never fires. You don’t need to actually implement the protocol.I ask because in the release notes for when they added mount namespaces, they did mention that at least initially, unsharing was not working properly.
Heh, don’t look at me. I don’t even run Chrome on WSL. 😉
I haven’t done an
unshare()test case or anything like that, but yes. It has to work, or it would be a resource leak.Oh I see what you mean. Yeah that counts, and “nada” is clearly overstated. It would be more correct to say “the problem right now is not related to namespaces”. So apologies for that.
[n.b. I just didn’t want to see this ticket go down the same path as #1962, which is a
mount -t drvfsbug/limitation. An otherwise useful ticket with an actionable problem was lost because an unrelated four letter acronym was repeated two dozen times.]Maybe. But probably not. That error message to
stderrdidn’t come from divine intervention. But either way that particularexecvesucceeds on Real Linux and fails on WSL.Note chrome goes out of it’s way to thwart
stracewhen running sandboxed (either kind), so what you could be seeing is a successfulexecvewhen not understraceon WSL, and then a subsequent clone fail, for reasons. As I mentioned, bog standard suid proper works on WSL, as far as it goes. The arguments passed toclonecalls preceding theexecvedon’t look that noteworthy from anEINVALstandpoint, AFAICT.All of which would be noteworthy if I thought Stephen actually put chrome in play.
Oh yeah, repro:
This is the fail:
I didn’t track it very far. Takes on Real Linux natch. Running suid binaries on WSL works as far as it goes.
In any case, the suid sandbox is nearly dead. Still be valuable (on its own merits) looking into why the
execvegoes south.