Re: [PATCH 2/2] mm: adds NOSIGBUS extension for out-of-band shmem read

On Thu, 3 Jun 2021, Linus Torvalds wrote:

On Wed, Jun 2, 2021 at 5:46 PM Hugh Dickins [off-list ref] wrote:

quoted

Ideally you can simply call do_anonymous_page() from __do_fault()
in the VM_FAULT_SIGBUS on VM_NOSIGBUS case.

Heh.

We're actually then back to my original patch.

That one doesn't handle shared mappings (even read-only ones), for the
simple reason that do_anonymous_page() refuses to insert anonymous
pages into a shared mapping, and has

        /* File mapping without ->vm_ops ? */
        if (vma->vm_flags & VM_SHARED)
                return VM_FAULT_SIGBUS;

at the very top.

But yes, if we just remove that check, I think my original patch
should actually "JustWork(tm)".

But no!

Sorry, I don't have time for this at present, so haven't looked at
your original patch.

But the point that we've arrived at, that I'm actually now fairly
happy with, is do *not* permit MAP_NOSIGBUS on MAP_SHARED mappings.

I didn't check the placement yet, easy to get wrong, but I believe
Ming Lin is now enforcing that over at the mmap() end.

On a MAP_PRIVATE mapping, the nasty opaque blob of zeroes can
claim some precedent in what already happens with COW'ed pages.

Which leaves MAP_NOSIGBUS on MAP_SHARED as currently unsupported,
perhaps never supported on anything, perhaps one day supported on
shmem; but if it's ever supported then that one will naturally be
transparent to future changes in page cache - we call that "shared".

Of course, internally, there's the in-between case of MAP_SHARED
without PROT_WRITE and without writable fd: VM_MAYSHARE without
VM_SHARED or VM_MAYWRITE.  We *could* let that one accept
MAP_NOSIGBUS, but who wants to write the manpage for it?

Please stick to MAP_PRIVATE: that's good enough.

I'm attaching it again, with old name and old commentary (ie that

    /* FIXME! We don't have a VM_NOFAULT bit */

should just be replaced with that VM_NOSIGBUS bit instead, and the
#if'ed out region should be enabled.

Oh, and we need to think hard about one more case: mprotect().

In particular, I think the attached patch fails horribly for the case
of a shared mapping that starts out read-only, then inserts a zero
page, then somebody does mprotect(MAP_WRITE), and then writes to the
page. I haven't checked what the write protect fault handler does, but
I think that for a shared mapping it will just make the page dirty and
writable.

Obviously the finished patch will need to be scrutinized carefully, but
I think the mprotect() questions vanish when restricted to MAP_PRIVATE.

Which would be horribly wrong for VM_NOSIGBUS.

So that support infrastructure that adds MAP_NOSIGBUS, and checks that
it is only done on a read-only mapping, also has to make sure that it
clears the VM_MAYWRITE bit when it sets VM_NOSIGBUS.

That way mprotect can't then later make it writable.

Hugh, comments on this approach?

Comments above, just stick to MAP_PRIVATE.

Hugh

Again: this patch is my *OLD* one, I didn't try to update it to the
new world order. It requires

 - Ming's MAP_NOSIGBUS ccode

 - removal of that "File mapping without ->vm_ops" case

 - that FIXME fixed and name updated

 - and that VM_MAYWRITE clearing if VM_NOSIGBUS is set, to avoid the
mprotect issue.

Hmm?

                  Linus