Re: [PATCH 2/5] mm/mseal: move madvise() logic to mm/madvise.c

[Lorenzo Stoakes <lorenzo.stoakes@oracle.com>]

On Mon, Jul 14, 2025 at 05:52:56PM +0200, David Hildenbrand wrote:
> On 14.07.25 17:45, Lorenzo Stoakes wrote:
> > On Mon, Jul 14, 2025 at 05:41:45PM +0200, David Hildenbrand wrote:
> > > On 14.07.25 17:31, Pedro Falcato wrote:
> > > > On Mon, Jul 14, 2025 at 05:03:03PM +0200, David Hildenbrand wrote:
> > > > > [...]
> > > > >
> > > > > But now I wonder, why is it okay to discard anon pages in a MAP_PRIVATE file
> > > > > mapping?
> > > >
> > > > IIRC this was originally suggested by Linus, on one of the versions introducing
> > > > mseal. But the gist is that discarding pages is okay if you could already zero
> > > > them manually, using e.g memset. Hence the writeability checks.
> > >
> > > What you can do is
> > >
> > > a) mmap(PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, ...)
> > >
> > > b) modify content (write, whatever)
> > >
> > > c) mprotect(PROT_READ)
> > >
> > > d) mseal()
> > >
> > > But then still do
> > >
> > > madvise(MADV_DONTNEED)
> > >
> > > to discard.
> > >
> > >
> > > There is no writability anymore.
> >
> > Well, you can mprotect() writable it again :)
>
> Isn't that what sealing ... prohibits?

Doh! Haha. Yes :) very good point...

OK I'll send the fix once this refactoring lands.

I think the check should simply be replaced with a vma->vm_flags & VM_SHARED
check then?

I wonder if the argument was that the MAP_PRIVATE file mapping is somehow
ephemeral anyway so not so problematic to lose. But I'm sure there are users who
feel differently about that...

Note that a truncate operation on the file will immediately drop the MAP_PRIVATE
anon backing anyway so it's sort of 'droppable' the whole time.

However, since a user is explicitly asking to seal the memory, and would have
expectation of seal semantics, so I agree that we need to fix this.

>
> --
> Cheers,
>
> David / dhildenb
>