On Fri, Mar 3, 2017 at 3:30 PM, Arnd Bergmann [off-list ref] wrote:
On Fri, Mar 3, 2017 at 2:55 PM, Alexander Potapenko [off-list ref] wrote:
quoted
On Fri, Mar 3, 2017 at 2:50 PM, Andrey Ryabinin [off-list ref] wrote:
quoted
quoted
quoted
@@ -416,6 +416,17 @@ static __always_inline void __write_once_size(volatile void *p, void *res, int s
*/
#define noinline_for_stack noinline
+/*
+ * CONFIG_KASAN can lead to extreme stack usage with certain patterns when
+ * one function gets inlined many times and each instance requires a stack
+ * ckeck.
+ */
+#ifdef CONFIG_KASAN
+#define noinline_for_kasan noinline __maybe_unused
noinline_iff_kasan might be a better name. noinline_for_kasan gives the impression
that we always noinline function for the sake of kasan, while noinline_iff_kasan
clearly indicates that function is noinline only if kasan is used.
Fine with me. I actually tried to come up with a name that implies that the
symbol is actually "inline" (or even __always_inline_ without KASAN, but
couldn't think of any good name for it.
quoted
FWIW we may be facing the same problem with other compiler-based
tools, e.g. KMSAN (which isn't there yet).
So it might be better to choose a macro name that doesn't use the name "KASAN".
E.g. noinline_iff_memtool (or noinline_iff_memory_tool if that's not too long).
WDYT?
Would KMSAN also force local variables to be non-overlapping the way that
asan-stack=1 and -fsanitize-address-use-after-scope do? As I understood it,
KMSAN would add extra code for maintaining the uninit bits, but in an example
like this
The thing is that KMSAN (and other tools that insert heavyweight
instrumentation) may cause heavy register spilling which will also
blow up the stack frames.
int f(int *);
static inline __attribute__((always_inline)) int g(void)
{
int i;
f(&i);
return i;
}
int f(void)
{
return g()+g()+g()+g();
}
each of the four copies of 'i' could have the same location on the stack
and get marked uninitialized again before calling f(). We only need
noinline_for_kasan (whatever we end up calling that) for compiler
features that force each instance of 'i' to have its own stack redzone.
Arnd
--
Alexander Potapenko
Software Engineer
Google Germany GmbH
Erika-Mann-Straße, 33
80636 München
Geschäftsführer: Matthew Scott Sucherman, Paul Terence Manicle
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg