On Thu, Sep 18, 2025, Michael S. Tsirkin wrote: > On Thu, Sep 18, 2025 at 09:52:19AM -0700, Sean Christopherson wrote: > > On Thu, Sep 18, 2025, Michael S. Tsirkin wrote: > > > On Thu, Sep 18, 2025 at 09:04:07AM -0700, Sean Christopherson wrote: > > > > On Thu, Sep 18, 2025, Sebastian Andrzej Siewior wrote: > > > > > On 2025-09-18 11:09:05 [-0400], Michael S. Tsirkin wrote: > > > > > > So how about switching to this approach then? > > > > > > Instead of piling up fixes like we seem to do now ... > > > > > > > > I don't have a strong preference for 6.17, beyond landing a fix of some > > > > kind. > > > > I think there are three options for 6.17, in order of "least like to > > > > break > > > > something": > > > > > > > > 1. Sebastian's get_task_struct() fix > > > > > > > > > I am just a bit apprehensive that we don't create a situation > > > where we leak the task struct somehow, given the limited > > > testing time. Can you help me get convinced that risk is 0? > > > > I doubt it, I share same similar concerns about lack of testing. So I guess > > thinking about this again, #2 is probably safer since it'd only impact KVM? > > I can't say I understand completely how we get that state though? > Why did the warning trigger if it's not a UAF?
It's purely a flaw in the sanity check itself due to the ordering in vhost_task_fn(). As is, vhost_task_fn() marks the task KILLED before invoking ->handle_sigkill(), i.e. before vhost_worker_killed() is guaranteed to complete, and thus before worker->killed is set. As a result, vhost can keep waking workers that have KILLED set, but haven't actually exited. That's perfectly fine as UAF won't occur until do_exit() is called, and that won't happen until ->handle_sigkill() completes. > > > > 2. This series, without the KILLED sanity check in __vhost_task_wake() > > > > 3. This series, with my fixup (with which syzbot was happy) >
