Re: [PATCHv2 1/2] x86/mm: Do not allow non-MAP_FIXED mapping across DEFAULT_MAP_WINDOW border
From: Kirill A. Shutemov
Date: Wed Nov 15 2017 - 09:14:17 EST
On Wed, Nov 15, 2017 at 05:04:26PM +0300, Kirill A. Shutemov wrote:
> On Wed, Nov 15, 2017 at 03:10:42PM +0300, Kirill A. Shutemov wrote:
> > On Wed, Nov 15, 2017 at 12:39:40PM +0100, Thomas Gleixner wrote:
> > > On Wed, 15 Nov 2017, Kirill A. Shutemov wrote:
> > > > On Wed, Nov 15, 2017 at 12:00:46AM +0100, Thomas Gleixner wrote:
> > > > > On Wed, 15 Nov 2017, Kirill A. Shutemov wrote:
> > > > > > On Tue, Nov 14, 2017 at 09:54:52PM +0100, Thomas Gleixner wrote:
> > > > > > > On Tue, 14 Nov 2017, Kirill A. Shutemov wrote:
> > > > > > >
> > > > > > > > On Tue, Nov 14, 2017 at 05:01:50PM +0100, Thomas Gleixner wrote:
> > > > > > > > > @@ -198,11 +199,14 @@ arch_get_unmapped_area_topdown(struct fi
> > > > > > > > > /* requesting a specific address */
> > > > > > > > > if (addr) {
> > > > > > > > > addr = PAGE_ALIGN(addr);
> > > > > > > > > + if (!mmap_address_hint_valid(addr, len))
> > > > > > > > > + goto get_unmapped_area;
> > > > > > > > > +
> > > > > > > >
> > > > > > > > Here and in hugetlb_get_unmapped_area(), we should align the addr after
> > > > > > > > the check, not before. Otherwise the alignment itself can bring us over
> > > > > > > > the borderline as we align up.
> > > > > > >
> > > > > > > Hmm, then I wonder whether the next check against vm_start_gap() which
> > > > > > > checks against the aligned address is correct:
> > > > > > >
> > > > > > > addr = PAGE_ALIGN(addr);
> > > > > > > vma = find_vma(mm, addr);
> > > > > > >
> > > > > > > if (end - len >= addr &&
> > > > > > > (!vma || addr + len <= vm_start_gap(vma)))
> > > > > > > return addr;
> > > > > >
> > > > > > I think the check is correct. The check is against resulting addresses
> > > > > > that end up in vm_start/vm_end. In our case we want to figure out what
> > > > > > user asked for.
> > > > >
> > > > > Well, but then checking just against the user supplied addr is only half of
> > > > > the story.
> > > > >
> > > > > addr = boundary - PAGE_SIZE - PAGE_SIZE / 2;
> > > > > len = PAGE_SIZE - PAGE_SIZE / 2;
> > > > >
> > > > > That fits, but then after alignment we end up with
> > > > >
> > > > > addr = boudary - PAGE_SIZE;
> > > > >
> > > > > and due to len > PAGE_SIZE this will result in a mapping which crosses the
> > > > > boundary, right? So checking against the PAGE_ALIGN(addr) should be the
> > > > > right thing to do.
> > > >
> > > > IIUC, this is only the case if 'len' is not aligned, right?
> > > >
> > > > >From what I see we expect caller to align it (and mm/mmap.c does this, I
> > > > haven't checked other callers).
> > > >
> > > > And hugetlb would actively reject non-aligned len.
> > > >
> > > > I *think* we should be fine with checking unaligned 'addr'.
> > >
> > > I think we should keep it consistent for the normal and the huge case and
> > > just check aligned and be done with it.
> >
> > Aligned 'addr'? Or 'len'? Both?
> >
> > We would have problem with checking aligned addr. I steped it in hugetlb
> > case:
> >
> > - User asks for mmap((1UL << 47) - PAGE_SIZE, 2 << 20, MAP_HUGETLB);
> >
> > - On 4-level paging machine this gives us invalid hint address as
> > 'TASK_SIZE - len' is more than 'addr'. Goto get_unmapped_area.
> >
> > - On 5-level paging machine hint address gets rounded up to next 2MB
> > boundary that is exactly 1UL << 47 and we happily allocate from full
> > address space which may lead to trouble.
>
> Below is updated patch with self-test.
>
> Output on 5-level paging machine:
>
> mmap(NULL): 0x7fbbad1f3000 - OK
> mmap(LOW_ADDR): 0x40000000 - OK
> mmap(HIGH_ADDR): 0x4000000000000 - OK
> mmap(HIGH_ADDR) again: 0xffffbbad1fb000 - OK
> mmap(HIGH_ADDR, MAP_FIXED): 0x4000000000000 - OK
> mmap(-1): 0xffffbbad1f9000 - OK
> mmap(-1) again: 0xffffbbad1f7000 - OK
> mmap((1UL << 47), 2 * PAGE_SIZE): 0x7fbbad1f3000 - OK
> mmap((1UL << 47), 2 * PAGE_SIZE / 2): 0x7fbbad1f1000 - OK
> mmap((1UL << 47) - PAGE_SIZE, 2 * PAGE_SIZE, MAP_FIXED): 0x7ffffffff000 - OK
> mmap(NULL, MAP_HUGETLB): 0x7fbbac400000 - OK
> mmap(LOW_ADDR, MAP_HUGETLB): 0x40000000 - OK
> mmap(HIGH_ADDR, MAP_HUGETLB): 0x4000000000000 - OK
> mmap(HIGH_ADDR, MAP_HUGETLB) again: 0xffffbbace00000 - OK
> mmap(HIGH_ADDR, MAP_FIXED | MAP_HUGETLB): 0x4000000000000 - OK
> mmap(-1, MAP_HUGETLB): (nil) - OK
> mmap(-1, MAP_HUGETLB) again: 0x7fbbac400000 - OK
> mmap((1UL << 47), 2UL << 20, MAP_HUGETLB): 0x800000000000 - FAILED
> mmap((1UL << 47) - (2UL << 20), 4UL << 20, MAP_FIXED | MAP_HUGETLB): 0x7fffffe00000 - OK
>
> So, only hugetlb is problematic. mmap() aligns addr to PAGE_SIZE.
> See round_hint_to_min(). In this case we round address *down* and it works
> fine.
>
> Replacing 'addr = ALIGN(addr, huge_page_size(h))' in hugetlbpage.c with
> 'addr &= huge_page_mask(h)' fixes the issue.
What about this fixup:
diff --git a/arch/x86/kernel/sys_x86_64.c b/arch/x86/kernel/sys_x86_64.c
index 835b78720ca2..676774b9bb8d 100644
--- a/arch/x86/kernel/sys_x86_64.c
+++ b/arch/x86/kernel/sys_x86_64.c
@@ -198,7 +198,7 @@ arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
/* requesting a specific address */
if (addr) {
- addr = PAGE_ALIGN(addr);
+ addr &= PAGE_MASK;
if (!mmap_address_hint_valid(addr, len))
goto get_unmapped_area;
diff --git a/arch/x86/mm/hugetlbpage.c b/arch/x86/mm/hugetlbpage.c
index 92db903c3dad..00b296617ca4 100644
--- a/arch/x86/mm/hugetlbpage.c
+++ b/arch/x86/mm/hugetlbpage.c
@@ -166,7 +166,7 @@ hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
}
if (addr) {
- addr = ALIGN(addr, huge_page_size(h));
+ addr &= huge_page_mask(h);
if (!mmap_address_hint_valid(addr, len))
goto get_unmapped_area;
--
Kirill A. Shutemov