[PATCH 0/4] arch, mm: improve robustness of direct map manipulation
Mike Rapoport
rppt at kernel.org
Tue Oct 27 20:47:14 AEDT 2020
On Tue, Oct 27, 2020 at 09:46:35AM +0100, David Hildenbrand wrote:
> On 27.10.20 09:38, Mike Rapoport wrote:
> > On Mon, Oct 26, 2020 at 06:05:30PM +0000, Edgecombe, Rick P wrote:
> > > On Mon, 2020-10-26 at 11:05 +0200, Mike Rapoport wrote:
> > > > On Mon, Oct 26, 2020 at 01:13:52AM +0000, Edgecombe, Rick P wrote:
> > > > > On Sun, 2020-10-25 at 12:15 +0200, Mike Rapoport wrote:
> > > > > > Indeed, for architectures that define
> > > > > > CONFIG_ARCH_HAS_SET_DIRECT_MAP
> > > > > > it is
> > > > > > possible that __kernel_map_pages() would fail, but since this
> > > > > > function is
> > > > > > void, the failure will go unnoticed.
> > > > >
> > > > > Could you elaborate on how this could happen? Do you mean during
> > > > > runtime today or if something new was introduced?
> > > >
> > > > A failure in__kernel_map_pages() may happen today. For instance, on
> > > > x86
> > > > if the kernel is built with DEBUG_PAGEALLOC.
> > > >
> > > > __kernel_map_pages(page, 1, 0);
> > > >
> > > > will need to split, say, 2M page and during the split an allocation
> > > > of
> > > > page table could fail.
> > >
> > > On x86 at least, DEBUG_PAGEALLOC expects to never have to break a page
> > > on the direct map and even disables locking in cpa because it assumes
> > > this. If this is happening somehow anyway then we should probably fix
> > > that. Even if it's a debug feature, it will not be as useful if it is
> > > causing its own crashes.
> > >
> > > I'm still wondering if there is something I'm missing here. It seems
> > > like you are saying there is a bug in some arch's, so let's add a WARN
> > > in cross-arch code to log it as it crashes. A warn and making things
> > > clearer seem like good ideas, but if there is a bug we should fix it.
> > > The code around the callers still functionally assume re-mapping can't
> > > fail.
> >
> > Oh, I've meant x86 kernel *without* DEBUG_PAGEALLOC, and indeed the call
> > that unmaps pages back in safe_copy_page will just reset a 4K page to
> > NP because whatever made it NP at the first place already did the split.
> >
> > Still, on arm64 with DEBUG_PAGEALLOC=n there is a possibility of a race
> > between map/unmap dance in __vunmap() and safe_copy_page() that may
> > cause access to unmapped memory:
> >
> > __vunmap()
> > vm_remove_mappings()
> > set_direct_map_invalid()
> > safe_copy_page()
> > __kernel_map_pages()
> > return
> > do_copy_page() -> fault
> >
> > This is a theoretical bug, but it is still not nice :)
> >
> > > > Currently, the only user of __kernel_map_pages() outside
> > > > DEBUG_PAGEALLOC
> > > > is hibernation, but I think it would be safer to entirely prevent
> > > > usage
> > > > of __kernel_map_pages() when DEBUG_PAGEALLOC=n.
> > >
> > > I totally agree it's error prone FWIW. On x86, my mental model of how
> > > it is supposed to work is: If a page is 4k and NP it cannot fail to be
> > > remapped. set_direct_map_invalid_noflush() should result in 4k NP
> > > pages, and DEBUG_PAGEALLOC should result in all 4k pages on the direct
> > > map. Are you seeing this violated or do I have wrong assumptions?
> >
> > You are right, there is a set of assumptions about the remapping of the
> > direct map pages that make it all work, at least on x86.
> > But this is very subtle and it's not easy to wrap one's head around
> > this.
> >
> > That's why putting __kernel_map_pages() out of "common" use and
> > keep it only for DEBUG_PAGEALLOC would make things clearer.
> >
> > > Beyond whatever you are seeing, for the latter case of new things
> > > getting introduced to an interface with hidden dependencies... Another
> > > edge case could be a new caller to set_memory_np() could result in
> > > large NP pages. None of the callers today should cause this AFAICT, but
> > > it's not great to rely on the callers to know these details.
> > A caller of set_memory_*() or set_direct_map_*() should expect a failure
> > and be ready for that. So adding a WARN to safe_copy_page() is the first
> > step in that direction :)
> >
>
> I am probably missing something important, but why are we saving/restoring
> the content of pages that were explicitly removed from the identity mapping
> such that nobody will access them?
>
> Pages that are not allocated should contain garbage or be zero
> (init_on_free). That should be easy to handle without ever reading the page
> content.
I'm not familiar with hibernation to say anything smart here, but the
help text of DEBUG_PAGEALLOC in Kconfig says:
... this option cannot be enabled in combination with
hibernation as that would result in incorrect warnings of memory
corruption after a resume because free pages are not saved to
the suspend image.
Probably you are right and free pages need to be handled differently,
but it does not seem the case now.
> The other user seems to be vm_remove_mappings(), where we only *temporarily*
> remove the mapping - while hibernating, that code shouldn't be active
> anymore I guess - or we could protect it from happening.
Hmm, I _think_ vm_remove_mappings() shouldn't be active while
hibernating, but I'm not 100% sure.
> As I expressed in another mail, secretmem pages should rather not be saved
> when hibernating - hibernation should be rather be disabled.
Agree.
> What am I missing?
I think I miscommunicated the purpose of this set, which was to hide
__kernel_map_pages() under DEBUG_PAGEALLOC and make hibernation use
set_direct_map_*() explictly without major rework of free pages handling
during hibernation.
Does it help?
> --
> Thanks,
>
> David / dhildenb
>
--
Sincerely yours,
Mike.
More information about the Linuxppc-dev
mailing list