Re: [PATCHv3 6/8] x86/mm: Provide ARCH_GET_UNTAG_MASK and ARCH_ENABLE_TAGGED_ADDR

["Andy Lutomirski" <luto@kernel.org>]

On Fri, Jun 10, 2022, at 3:18 PM, Edgecombe, Rick P wrote:
> On Fri, 2022-06-10 at 11:08 -0700, Edgecombe, Richard P wrote:
>> On Fri, 2022-06-10 at 21:06 +0300, Kirill A. Shutemov wrote:
>> > On Fri, Jun 10, 2022 at 04:16:01PM +0000, Edgecombe, Rick P wrote:
>> > > On Fri, 2022-06-10 at 17:35 +0300, Kirill A. Shutemov wrote:
>> > > > +static int prctl_enable_tagged_addr(unsigned long nr_bits)
>> > > > +{
>> > > > +       struct mm_struct *mm = current->mm;
>> > > > +
>> > > > +       /* Already enabled? */
>> > > > +       if (mm->context.lam_cr3_mask)
>> > > > +               return -EBUSY;
>> > > > +
>> > > > +       /* LAM has to be enabled before spawning threads */
>> > > > +       if (get_nr_threads(current) > 1)
>> > > > +               return -EBUSY;
>> > > 
>> > > Does this work for vfork()? I guess the idea is that locking is
>> > > not
>> > > needed below because there is only one thread with the MM, but
>> > > with
>> > > vfork() another task could operate on the MM, call fork(), etc.
>> > > I'm
>> > > not
>> > > sure...
>> > 
>> > I'm not sure I follow. vfork() blocks parent process until child
>> > exit
>> > or
>> > execve(). I don't see how it is a problem.
>> 
>> Oh yea, you're right.
>
> Actually, I guess vfork() only suspends the calling thread. So what if
> you had:
> 1. Parent spawns a bunch of threads
> 2. vforks()
> 3. Child enables LAM (it only has one thread, so succeeds)
> 4. Child exits()
> 5. Parent has some threads with LAM, and some not
>
> It's some weird userspace that doesn't deserve to have things work for
> it, but I wonder if it could open up little races around untagging. As
> an example, KVM might have a super narrow race where it checks for tags
> in memslots using addr != untagged_addr(addr) before checking
> access_ok(addr, ...). See __kvm_set_memory_region(). If mm-
>>context.untag_mask got set in the middle, tagged memslots could be
> added.

get_nr_threads() is the wrong thing.  Either look at mm->mm_users or find a way to get rid of this restriction entirely.

IMO it would not be insane to have a way to iterate over all tasks using an mm.  But doing this for io_uring, etc might be interesting.