Re: lib/test_overflow.c causes WARNING and tainted kernel

[Kees Cook <keescook@chromium.org>]

On Mon, May 27, 2019 at 09:53:33AM +0200, Rasmus Villemoes wrote:
> On 25/05/2019 17.33, Randy Dunlap wrote:
> > On 3/13/19 7:53 PM, Kees Cook wrote:
> >> Hi!
> >>
> >> On Wed, Mar 13, 2019 at 2:29 PM Randy Dunlap <rdunlap@infradead.org> wrote:
> >>>
> >>> This is v5.0-11053-gebc551f2b8f9, MAR-12 around 4:00pm PT.
> >>>
> >>> In the first test_kmalloc() in test_overflow_allocation():
> >>>
> >>> [54375.073895] test_overflow: ok: (s64)(0 << 63) == 0
> >>> [54375.074228] WARNING: CPU: 2 PID: 5462 at ../mm/page_alloc.c:4584 __alloc_pages_nodemask+0x33f/0x540
> >>> [...]
> >>> [54375.079236] ---[ end trace 754acb68d8d1a1cb ]---
> >>> [54375.079313] test_overflow: kmalloc detected saturation
> >>
> >> Yup! This is expected and operating as intended: it is exercising the
> >> allocator's detection of insane allocation sizes. :)
> >>
> >> If we want to make it less noisy, perhaps we could add a global flag
> >> the allocators could check before doing their WARNs?
> >>
> >> -Kees
> > 
> > I didn't like that global flag idea.  I also don't like the kernel becoming
> > tainted by this test.
> 
> Me neither. Can't we pass __GFP_NOWARN from the testcases, perhaps with
> a module parameter to opt-in to not pass that flag? That way one can
> make the overflow module built-in (and thus run at boot) without
> automatically tainting the kernel.
> 
> The vmalloc cases do not take gfp_t, would they still cause a warning?

They still warn, but they don't seem to taint. I.e. this patch:

diff --git a/lib/test_overflow.c b/lib/test_overflow.c
index fc680562d8b6..c922f0d86181 100644
--- a/lib/test_overflow.c
+++ b/lib/test_overflow.c
@@ -486,11 +486,12 @@ static int __init test_overflow_shift(void)
  * Deal with the various forms of allocator arguments. See comments above
  * the DEFINE_TEST_ALLOC() instances for mapping of the "bits".
  */
-#define alloc010(alloc, arg, sz) alloc(sz, GFP_KERNEL)
-#define alloc011(alloc, arg, sz) alloc(sz, GFP_KERNEL, NUMA_NO_NODE)
+#define alloc_GFP	(GFP_KERNEL | __GFP_NOWARN)
+#define alloc010(alloc, arg, sz) alloc(sz, alloc_GFP)
+#define alloc011(alloc, arg, sz) alloc(sz, alloc_GFP, NUMA_NO_NODE)
 #define alloc000(alloc, arg, sz) alloc(sz)
 #define alloc001(alloc, arg, sz) alloc(sz, NUMA_NO_NODE)
-#define alloc110(alloc, arg, sz) alloc(arg, sz, GFP_KERNEL)
+#define alloc110(alloc, arg, sz) alloc(arg, sz, alloc_GFP | __GFP_NOWARN)
 #define free0(free, arg, ptr)	 free(ptr)
 #define free1(free, arg, ptr)	 free(arg, ptr)
 
will remove the tainting behavior but is still a bit "noisy". I can't
find a way to pass __GFP_NOWARN to a vmalloc-based allocation, though.

Randy, is removing taint sufficient for you?

> BTW, I noticed that the 'wrap to 8K' depends on 64 bit and
> pagesize==4096; for 32 bit the result is 20K, while if the pagesize is
> 64K one gets 128K and 512K for 32/64 bit size_t, respectively. Don't
> know if that's a problem, but it's easy enough to make it independent of
> pagesize (just make it 9*4096 explicitly), and if we use 5 instead of 9
> it also becomes independent of sizeof(size_t) (wrapping to 16K).

Ah! Yes, all excellent points. I've adjusted that too now. I'll send
the result to Andrew.

Thanks!

-- 
Kees Cook