On 1/24/22 19:37, Peter Xu wrote:
Alex reported invalid page pointer returned with pin_user_pages_remote() from
vfio after upstream commit 4b6c33b32296 ("vfio/type1: Prepare for batched
pinning with struct vfio_batch"). This problem breaks NVIDIA vfio mdev.
It turns out that it's not the fault of the vfio commit; however after vfio
switches to a full page buffer to store the page pointers it starts to expose
the problem easier.
The problem is for VM_PFNMAP vmas we should normally fail with an -EFAULT then
vfio will carry on to handle the MMIO regions. However when the bug triggered,
follow_page_mask() returned -EEXIST for such a page, which will jump over the
current page, leaving that entry in **pages untouched. However the caller is
not aware of it, hence the caller will reference the page as usual even if the
pointer data can be anything.
We had that -EEXIST logic since commit 1027e4436b6a ("mm: make GUP handle pfn
mapping unless FOLL_GET is requested") which seems very reasonable. It could
be that when we reworked GUP with FOLL_PIN we could have overlooked that
special path in commit 3faa52c03f44 ("mm/gup: track FOLL_PIN pages"), even if
that commit rightfully touched up follow_devmap_pud() on checking FOLL_PIN when
it needs to return an -EEXIST.
Since at it, add another WARN_ON_ONCE() at the -EEXIST handling to make sure we
mustn't have **pages set when reaching there, because otherwise it means the
caller will try to read a garbage right after __get_user_pages() returns.
Attaching the Fixes to the FOLL_PIN rework commit, as it happened later than
1027e4436b6a.
Cc: Andrea Arcangeli <aarcange@xxxxxxxxxx>
Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
Cc: Jan Kara <jack@xxxxxxx>
Cc: Jérôme Glisse <jglisse@xxxxxxxxxx>
Cc: John Hubbard <jhubbard@xxxxxxxxxx>
Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
Fixes: 3faa52c03f44 ("mm/gup: track FOLL_PIN pages")
Reported-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
Debugged-by: Alex Williamson <alex.williamson@xxxxxxxxxx>
Signed-off-by: Peter Xu <peterx@xxxxxxxxxx>
---
mm/gup.c | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/mm/gup.c b/mm/gup.c
index f0af462ac1e2..8ebc04058e97 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -440,7 +440,7 @@ static int follow_pfn_pte(struct vm_area_struct *vma, unsigned long address,
pte_t *pte, unsigned int flags)
{
/* No page to get reference */
- if (flags & FOLL_GET)
+ if (flags & (FOLL_GET | FOLL_PIN))
return -EFAULT;
Yes. This clearly fixes the problem that the patch describes, and also
clearly matches up with the Fixes tag. So that's correct.
if (flags & FOLL_TOUCH) {
@@ -1181,7 +1181,13 @@ static long __get_user_pages(struct mm_struct *mm,
/*
* Proper page table entry exists, but no corresponding
* struct page.
+ *
+ * Warn if we jumped over even with a valid **pages.
+ * It shouldn't trigger in practise, but when there's
+ * buggy returns on -EEXIST we'll warn before returning
+ * an invalid page pointer in the array.
*/
+ WARN_ON_ONCE(pages);
Here, however, I think we need to consider this a little more carefully,
and attempt to actually fix up this case. It is never going to be OK
here, to return a **pages array that has these little landmines of
potentially uninitialized pointers. And so continuing on *at all* seems
very wrong.
Can we bail out at this point, without breaking the world? I think we can...
Also: this part, even if it remains as is, should be a separate fix and a
separate patch, IMHO.
thanks,
--
John Hubbard
NVIDIA
goto next_page;
} else if (IS_ERR(page)) {
ret = PTR_ERR(page);
