Hi On Wed, Aug 22, 2018 at 6:08 PM Daniel P. Berrangé <berra...@redhat.com> wrote: > > On Wed, Aug 22, 2018 at 05:58:46PM +0200, Marc-André Lureau wrote: > > On Wed, Aug 22, 2018 at 5:46 PM, Daniel P. Berrangé <berra...@redhat.com> > > wrote: > > > On Wed, Aug 22, 2018 at 04:29:56PM +0200, Marc-André Lureau wrote: > > > > > > At this point you might as well not bother using seccomp at all. The > > > thread that is confined merely needs to scribble something into the > > > stack of the unconfined thread and now it can do whatever it wants. > > > > Actually, that message is incorrect, it should rather be "not all > > threads will be filtered" (as described in commit message). > > > > > IMHO we need to find a way to get the policy to apply to those other > > > threads. > > > > That's what the patch is about ;) > > It only does it in some scenarios, leaving other unfixed. We need > a solution (or choice of multiple solutions) that works all the time > > > > > > The RCU thread is tricky as it is spawned from a __constructor__ > > > function, which means it'll be active way before we setup seccomp. > > > > > > I think we need to figure out a way todo synchronization between > > > the RCU thread and the seccomp setup code. Could we have a global > > > variable 'int seccomp_initialized' that we check from the RCU > > > thread loop - when that toggles to non-zero, the RCU thread can > > > then call into the seccomp_start() method to activate policy in > > > its thread. We'd need a synchronous feedback mechansim back to > > > the main thread, as it must block startup until all the threads > > > have activated the seccomp filter. > > > > That's a bit like TSYNC, except we do it ourself with RCU thread. But > > what about other threads? For examples one that could be created by > > external libraries (like mesa) > > Does mesa create threads from library constructors too, or somewhere > else *before* we do -seccomp setup ?
That was an example, I don't think mesa creates threads before -seccomp. But what about the other 100 dependencies, or if we introduce other threads without the seccomp sync by mistake? I think we are better off using tsync. > > > >> diff --git a/qemu-options.hx b/qemu-options.hx > > >> index 5515dfaba5..dafacb60c6 100644 > > >> --- a/qemu-options.hx > > >> +++ b/qemu-options.hx > > >> @@ -3864,6 +3864,8 @@ Disable set*uid|gid system calls > > >> Disable *fork and execve > > >> @item resourcecontrol=@var{string} > > >> Disable process affinity and schedular priority > > >> +@item tsync=@var{bool} > > >> +Apply seccomp filter to all threads (default is auto, and will warn if > > >> fail) > > > > > > IMHO this should never exist, as setting "tsync" to anything other > > > than "yes", is akin to just running without any sandbox. > > > > Then we should just fail -sandbox on those systems. > > We would have to make libvirt probe for tsync support too, because it > now unconditionally uses -sandbox for new enough QEMU. sigh :( that's where the -sandbox tsync option could have been helpful keeping the compatibility. -- Marc-André Lureau