Re: [PATCH v6 00/11] Kernel address sanitizer - runtime memory debugger.

From: Andrey Ryabinin
Date: Fri Nov 21 2014 - 06:19:26 EST

On 11/21/2014 10:32 AM, Dmitry Vyukov wrote:
> On Fri, Nov 21, 2014 at 2:00 AM, Andrew Morton
> <akpm@xxxxxxxxxxxxxxxxxxxx> wrote:
>> On Thu, 20 Nov 2014 20:32:30 +0400 Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>>> Let me provide some background first.
>> Well that was useful. Andrey, please slurp Dmitry's info into the 0/n
>> changelog?
>> Also, some quantitative info about the kmemleak overhead would be
>> useful.
>> In this discussion you've mentioned a few planned kasan enhancements.
>> Please also list those and attempt to describe the amount of effort and
>> complexity levels. Partly so other can understand the plans and partly
>> so we can see what we're semi-committing ourselves to if we merge this
>> stuff.
> The enhancements are:
> 1. Detection of stack out-of-bounds. This is done mostly in the
> compiler. Kernel only needs adjustments in reporting.

Not so easy.
- Because of redzones stack size needs enlarging.
- We also need to populate shadow for addresses where kernel .data section mapped
because we need shadow memory for init task's stack.

> 2. Detection of global out-of-bounds. Kernel will need to process
> compiler-generated list of globals during bootstrap. Complexity is
> very low and it is isolated in Asan code.

One easy thing to do here is adding support for .init.array.* constructors.
Kernel already supports .init.array constructors, but for address sanitizer,
GCC puts constructors into .init.array.00099 section.

Just as for stack redzones, shadow needs to be populated for kernel .data addresses.
Plus shadow memory for module mapping space is also needed.

> 3. Heap quarantine (delayed reuse of heap blocks). We will need to
> hook into slub, queue freed blocks in an efficient/scalable way and
> integrate with memory shrinker (register_shrinker). This will be
> somewhat complex and touch production kernel code. Konstantin
> Khlebnikov wants to make the quarantine available independently of
> Asan, as part of slub debug that can be enabled at runtime.

If someone wants to try quarantine for slub: git:// --branch=quarantine

It has some problems with switching it on/off in runtime, besides that, it works.

> 4. Port Asan to slAb.
> 5. Do various tuning of allocator integration, redzones sizes,
> speeding up what is currently considered debug-only paths in
> malloc/free, etc.
> 6. Some people also expressed interest in ARM port.

7. Compiler can't instrument assembler code, so it would be nice to have
checks in most frequently used parts of inline assembly. Something like

static inline void atomic_inc(atomic_t *v)
kasan_check _memory(v, sizeof(*v), WRITE);
asm volatile(LOCK_PREFIX "incl %0"
: "+m" (v->counter));

8. With asan's inline instrumentation bugs like NULL-ptr derefs or access to user space
turn into General protection faults. I will add a hint message into GPF handler to
indicate that GPF could be caused by NULL-ptr dereference or user memory access.
It's trivial, so I'll do this in v7.

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at