Re: [PATCH v8 03/23] mm: Check against orig_pte for finish_fault()

[Peter Xu]

 On Wed, Apr 13, 2022 at 04:03:28PM +0200, Marek Szyprowski wrote:
> Hi,

Hi, Marek,

> 
> On 05.04.2022 03:48, Peter Xu wrote:
> > We used to check against none pte in finish_fault(), with the assumption
> > that the orig_pte is always none pte.
> >
> > This change prepares us to be able to call do_fault() on !none ptes.  For
> > example, we should allow that to happen for pte marker so that we can restore
> > information out of the pte markers.
> >
> > Let's change the "pte_none" check into detecting changes since we fetched
> > orig_pte.  One trivial thing to take care of here is, when pmd==NULL for
> > the pgtable we may not initialize orig_pte at all in handle_pte_fault().
> >
> > By default orig_pte will be all zeros however the problem is not all
> > architectures are using all-zeros for a none pte.  pte_clear() will be the
> > right thing to use here so that we'll always have a valid orig_pte value
> > for the whole handle_pte_fault() call.
> >
> > Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
> 
> This patch landed in today's linux next-202204213 as commit fa6009949163 
> ("mm: check against orig_pte for finish_fault()"). Unfortunately it 
> causes serious system instability on some ARM 32bit machines. I've 
> observed it on all tested boards (various Samsung Exynos based, 
> Raspberry Pi 3b and 4b, even QEMU's virt 32bit machine) when kernel was 
> compiled from multi_v7_defconfig.

Thanks for the report.

> 
> Here is a crash log from QEMU's ARM 32bit virt machine:
> 
> 8<--- cut here ---
> Unable to handle kernel paging request at virtual address e093263c
> [e093263c] *pgd=42083811, *pte=00000000, *ppte=00000000
> Internal error: Oops: 807 [#1] SMP ARM
> Modules linked in:
> CPU: 1 PID: 37 Comm: kworker/u4:0 Not tainted 
> 5.18.0-rc2-00176-gfa6009949163 #11684
> Hardware name: Generic DT based system
> PC is at cpu_ca15_set_pte_ext+0x4c/0x58
> LR is at handle_mm_fault+0x46c/0xbb0

I had a feeling that for some reason the pte_clear() isn't working right
there when it's applying to a kernel stack variable for arm32.  I'm totally
newbie to arm32, so what I'm reading is this:

https://people.kernel.org/linusw/arm32-page-tables

Especially:

https://dflund.se/~triad/images/classic-mmu-page-table.jpg

It does match with what I read from arm32's proc-v7-2level.S of it, where
from the comment above cpu_v7_set_pte_ext:

  * - ptep  - pointer to level 2 translation table entry
  *    (hardware version is stored at +2048 bytes)        <----------

So it seems to me that arm32 needs to store some metadata at offset 0x800
of any pte_t* pointer passed over to pte_clear(), then it must be a real
pgtable or it'll write to random places in the kernel, am I correct?

Does it mean that all pte_*() operations upon a kernel stack var will be
wrong?  I thought it could happen easily in the rest of mm too but I didn't
yet check much.  The fact shows that it's mostly possible the current code
just work well with arm32 and no such violation occured yet.

That does sound a bit tricky, IMHO.  But I don't have an immediate solution
to make it less tricky.. though I have a thought of workaround, by simply
not calling pte_clear() on the stack var.

Would you try the attached patch to replace this problematic patch? So we
need to revert commit fa6009949163 and apply the new one.  Please let me
know whether it'll solve the problem, so far I only compile tested it, but
I'll run some more test to make sure the uffd-wp scenarios will be working
right with the new version.

Thanks,

-- 
Peter Xu