Re: [PATCH] linux-user: Add emulation for MADV_WIPEONFORK and MADV_KEEPONFORK in madvise()

[Ilya Leoshkevich]

On Mon, Dec 12, 2022 at 10:49:24PM +0100, Helge Deller wrote:
> On 12/12/22 22:16, Ilya Leoshkevich wrote:
> > On Mon, Dec 12, 2022 at 08:00:45AM +0100, Helge Deller wrote:
> > > Both parameters have a different value on the parisc platform, so first
> > > translate the target value into a host value for usage in the native
> > > madvise() syscall.
> > > 
> > > Those parameters are often used by security sensitive applications (e.g.
> > > tor browser, boringssl, ...) which expect the call to return a proper
> > > return code on failure, so return -EINVAL if qemu fails to forward the
> > > syscall to the host OS.
> > > 
> > > Tested with testcase of tor browser when running hppa-linux guest on
> > > x86-64 host.
> > > 
> > > Signed-off-by: Helge Deller <deller@gmx.de>
> > > 
> > > diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> > > index 10f5079331..c75342108c 100644
> > > --- a/linux-user/mmap.c
> > > +++ b/linux-user/mmap.c
> > > @@ -901,11 +901,25 @@ abi_long target_madvise(abi_ulong start, abi_ulong 
> > > len_in, int advice)
> > >           return -TARGET_EINVAL;
> > >       }
> > > 
> > > +    /* Translate for some architectures which have different MADV_xxx 
> > > values */
> > > +    switch (advice) {
> > > +    case TARGET_MADV_DONTNEED:      /* alpha */
> > > +        advice = MADV_DONTNEED;
> > > +        break;
> > > +    case TARGET_MADV_WIPEONFORK:    /* parisc */
> > > +        advice = MADV_WIPEONFORK;
> > > +        break;
> > > +    case TARGET_MADV_KEEPONFORK:    /* parisc */
> > > +        advice = MADV_KEEPONFORK;
> > > +        break;
> > > +    /* we do not care about the other MADV_xxx values yet */
> > > +    }
> > > +
> > >       /*
> > >        * A straight passthrough may not be safe because qemu sometimes 
> > > turns
> > >        * private file-backed mappings into anonymous mappings.
> > >        *
> > > -     * This is a hint, so ignoring and returning success is ok.
> > > +     * For MADV_DONTNEED, which is a hint, ignoring and returning 
> > > success is ok.
> > 
> > Actually, MADV_DONTNEED is one of the few values, which is not always a
> > hint - it can be used to e.g. zero out pages.
> 
> Right, it _should_ zero out pages and return 0, or otherwise return failure.
> I think the problem is that some userspace apps will then sadly break if we
> change the current behaviour....
> 
> Anyway, in this patch I didn't wanted to touch MAD_DONTNEED.
> 
> > As the next paragraph states, strictly speaking, MADV_DONTNEED is
> > currently broken, because it can indeed be ignored without indication
> > in some cases, but it's still arguably better than not honoring it at
> > all.
> 
> Yep.
> 
> > >        *
> > >        * This breaks MADV_DONTNEED, completely implementing which is quite
> > >        * complicated. However, there is one low-hanging fruit: mappings 
> > > that are
> > > @@ -913,11 +927,17 @@ abi_long target_madvise(abi_ulong start, abi_ulong 
> > > len_in, int advice)
> > >        * passthrough is safe, so do it.
> > >        */
> > >       mmap_lock();
> > > -    if (advice == TARGET_MADV_DONTNEED &&
> > > -        can_passthrough_madv_dontneed(start, end)) {
> > > -        ret = get_errno(madvise(g2h_untagged(start), len, 
> > > MADV_DONTNEED));
> > > -        if (ret == 0) {
> > > -            page_reset_target_data(start, start + len);
> > > +    switch (advice) {
> > > +    case MADV_WIPEONFORK:
> > > +    case MADV_KEEPONFORK:
> > > +        ret = -EINVAL;
> > > +        /* fall through */
> > > +    case MADV_DONTNEED:
> > > +        if (can_passthrough_madv_dontneed(start, end)) {
> > > +            ret = get_errno(madvise(g2h_untagged(start), len, advice));
> > > +            if ((advice == MADV_DONTNEED) && (ret == 0)) {
> > > +                page_reset_target_data(start, start + len);
> > > +            }
> > >           }
> > >       }
> > >       mmap_unlock();
> > > 
> > 
> > Nit: maybe rename can_passthrough_madv_dontneed() to can_passthrough(),
> > since now it's used not only for MADV_DONTNEED?
> 
> Maybe can_passthrough_madvise() is better?

Sounds good to me as well. The idea with PAGE_PASSTHROUGH was that we
should be able to passthrough anything; but with this patch we still
use it only for madvise(), and indicating it in the name makes sense.

> > > With the MADV_DONTNEED comment change:
> 
> Just for me to understand correctly:
> You propose that I shouldn't touch that comment in my followup-patch, right?
> That's ok.

Either that, or maybe make it more precise - strictly speaking, it's
not correct at the moment anyway. Maybe something like this?

    Most advice values are hints, so ignoring and returning success is
    ok.

    However, this would break MADV_DONTNEED, MADV_WIPEONFORK and
    MADV_KEEPONFORK ...

> > Acked-by: Ilya Leoshkevich <iii@linux.ibm.com>
> 
> Thanks!
> Helge