Re: [PATCH hotfix 6.12 v2 4/8] mm: resolve faulty mmap_region() error path behaviour

["Liam R. Howlett" <Liam.Howlett@oracle.com>]

* Liam R. Howlett <Liam.Howlett@oracle.com> [241028 15:50]:
> * Lorenzo Stoakes <lorenzo.stoakes@oracle.com> [241028 15:14]:
> > On Mon, Oct 28, 2024 at 09:05:44AM -1000, Linus Torvalds wrote:
> > > On Mon, 28 Oct 2024 at 08:57, Lorenzo Stoakes
> > > <lorenzo.stoakes@oracle.com> wrote:
> > > >
> > > > So likely hook on your mapping changes flags to set VM_MTE | VM_MTE_ALLOWED and
> > > > expects this to be checked after (ugh).
> > >
> > > Gaah. Yes. mm/shmem.c: shmem_mmap() does
> > >
> > >         /* arm64 - allow memory tagging on RAM-based files */
> > >         vm_flags_set(vma, VM_MTE_ALLOWED);
> > >
> > > and while I found the equivalent hack for the VM_SPARC_ADI case, I
> > > hadn't noticed that MTE thing.
> > >
> > > How very annoying.
> > >
> > > So the arch_validate_flags() case does need to be done after the ->mmap() call.
> > >
> > > How about just finalizing everything, and then doing a regular
> > > munmap() afterwards and returning an error (all still holding the mmap
> > > semaphore, of course).
> > >
> > > That still avoids the whole "partially completed mmap" case.
> > >
> > >              Linus
> > 
> > Yeah I was thinking the same... just bite the bullet, go through the whole damn
> > process and revert if arch_validate_flags() chokes. It also removes the ugly
> > #ifdef CONFIG_SPARC64 hack...
> > 
> > This will litearlly only be applicable for these two cases and (hopefully) most
> > of the time you'd not fail it.
> > 
> > I mean by then it'll be added into the rmap and such but nothing will be
> > populated yet and we shouldn't be able to fault as vma_start_write() should have
> > incremented the vma lock seqnum.
> > 
> > Any issues from the RCU visibility stuff Liam?
> 
> It is probably fine?  We would see a mapping appear then disappear.
> We'd have a (benign) race with rmap for truncating the PTEs (but it's
> safe).  Page faults would be stopped though.
> 
> Unfortunately, we'd have to write to the vma tree so that we could...
> write to the vma tree.  We'd have to somehow ensure munmap() is done
> with a gfp flag to ensure no failures as well...
> 
> Maybe we should just call close on the vma again (and do whatever
> call_mmap() needs to undo)?

I take it back, that won't work.

> 
> > 
> > Any security problems Jann...?
> > 
> > It'd suck to have to bring back a partial complete case. Though I do note we
> > handle errors from mmap_file() ok so we could still potentially handle that
> > there, but would sort of semi-undo some of the point of the series.