Re: [RFC PATCH 3/4] mm: madvise: implement lightweight guard page mechanism

[Jann Horn]

On Mon, Oct 14, 2024 at 7:02 PM Lorenzo Stoakes
<lorenzo.stoakes@xxxxxxxxxx> wrote:
> On Mon, Oct 14, 2024 at 05:56:50PM +0200, Jann Horn wrote:
> > On Mon, Oct 14, 2024 at 1:09 PM Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx> wrote:
> > > On Fri, Oct 11, 2024 at 08:11:36PM +0200, Jann Horn wrote:
> > > > On Fri, Sep 27, 2024 at 2:51 PM Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx> wrote:
> > > By being optimistic and simply having the user having to handle looping
> > > which seems reasonable (again, it's weird if you're installing poison
> > > markers and another thread could be racing you) we avoid all of that.
> >
> > I guess one case in which that could happen legitimately is if you
> > race a MADV_POISON on the area 0x1ff000-0x200100 (first page is
> > populated, second page is not, pmd entry corresponding to 0x200000 is
> > clear) with a page fault at 0x200200? So you could have a scenario
> > like:
> >
> > 1. MADV_POISON starts walk_page_range()
> > 2. MADV_POISON sees non-zero, non-poison PTE at 0x1ff000, stops the walk
> > 3. MADV_POISON does zap_page_range_single()
> > 4. pagefault at 0x200200 happens and populates with a hugepage
> > 5. MADV_POISON enters walk_page_range()
> > 6. MADV_POISON splits the THP
> > 7. MADV_POISON sees a populated PTE
>
> You really shouldn't be seeing page faults in the range you are setting up
> poison markers for _at all_ :) it's something you'd do ahead of time.

But that's not what happens in my example - the address where the
fault happens (0x200200) *is not* in the address range that
MADV_POISON is called on (0x1ff000-0x200100). The fault and the
MADV_POISON are in different 4KiB pages. What causes the conflict is
that the fault and the MADV_POISON overlap the same *2MiB region*
(both are in the region 0x200000-0x400000), and so THP stuff can
effectively cause "page faults in the range you are setting up poison
markers for".

> But of course it's possible some scenario could arise like that, that's
> what the EAGAIN is for.
>
> I just really don't want to get into a realm of trying to prove absolutely
> under all circumstances that we can't go on forever in a loop like that.

We can have a bailout on signal_pending() or something like that, and
a cond_resched(). Then as far as I know, it won't really make a
difference in behavior whether the loop is in the kernel or in
userspace code that's following what the manpage tells it to do -
either way, the program will loop until it either finishes its work or
is interrupted by a signal, and either way it can get preempted.
(Well, except under PREEMPT_NONE, but that is basically asking for
long scheduling delays.)

And we do have other codepaths that have to loop endlessly if they
keep racing with page table updates the wrong way, though I guess
those loops are not going to always scan over a large address range
over and over again...

Maybe something like this would be good enough, and mirror what you'd
otherwise tell userspace to do?


@@ -1598,6 +1598,7 @@ int do_madvise(struct mm_struct *mm, unsigned
long start, size_t len_in, int beh
                return madvise_inject_error(behavior, start, start + len_in);
 #endif

+retry:
        write = madvise_need_mmap_write(behavior);
        if (write) {
                if (mmap_write_lock_killable(mm))
@@ -1627,6 +1628,12 @@ int do_madvise(struct mm_struct *mm, unsigned
long start, size_t len_in, int beh
        else
                mmap_read_unlock(mm);

+       if (error == <<<some special value>>>) {
+               if (!signal_pending(current))
+                       goto retry;
+               error = -ERESTARTNOINTR;
+       }
+
        return error;
 }

Buuut, heh, actually, I just realized: You could even omit this and
simply replace -EINTR with -ERESTARTNOINTR in your code as the error
value, and then the kernel would automatically go back into the
syscall for you after going through signal handing and such, without
userspace noticing.
https://lore.kernel.org/all/20121206220955.GZ4939@xxxxxxxxxxxxxxxxxx/
has some explanation on how this works. Basically it tells the
architecture's syscall entry code to move the userspace instruction
pointer back to the syscall instruction, so as soon as execution
returns to userspace, the first userspace instruction that executes
will immediately re-do the syscall. That might be the easiest way,
even if it is maybe a *little* bit of an API abuse to use this thing
without having a pending signal...


> If you drop the lock for contention then you up the risk of that, it just
> feels dangerous.
>
> A userland program can however live with a 'if EAGAIN try again' situation.
>
> An alternative approach to this might be to try to take the VMA lock, but
> given the fraught situation with locking elsewhere I wonder if we should.
>
> Also, you have to be realy unlucky with timing for this to happen, even in
> the scenario you mention (where you'd have to be unlucky with alignment
> too), unless you're _heavily_ page faulting in the range, either way a
> userland loop checking EAGAIN doesn't seem unreasonable.

Yes, we could do -EINTR and document that for userspace, and as long
as everyone using this properly reads the documentation, it will be
fine. Though I imagine that from the userspace programmer perspective
that's a weird API design - as in, if this error code always means I
have to try again, why can't the kernel do that internally. It's kind
of leaking an implementation detail into the UAPI.