On 02/16/2015 05:58 AM, Rusty Russell wrote:
> Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> writes:
>> This feature let us to detect accesses out of bounds of
>> global variables. This will work as for globals in kernel
>> image, so for globals in modules. Currently this won't work
>> for symbols in user-specified sections (e.g. __init, __read_mostly, ...)
>>
>> The idea of this is simple. Compiler increases each global variable
>> by redzone size and add constructors invoking __asan_register_globals()
>> function. Information about global variable (address, size,
>> size with redzone ...) passed to __asan_register_globals() so we could
>> poison variable's redzone.
>>
>> This patch also forces module_alloc() to return 8*PAGE_SIZE aligned
>> address making shadow memory handling ( kasan_module_alloc()/kasan_module_free() )
>> more simple. Such alignment guarantees that each shadow page backing
>> modules address space correspond to only one module_alloc() allocation.
>
> Hmm, I understand why you only fixed x86, but it's weird.
>
> I think MODULE_ALIGN belongs in linux/moduleloader.h, and every arch
> should be fixed up to use it (though you could leave that for later).
>
> Might as well fix the default implementation at least.
>
>> @@ -49,8 +49,15 @@ void kasan_krealloc(const void *object, size_t new_size);
>> void kasan_slab_alloc(struct kmem_cache *s, void *object);
>> void kasan_slab_free(struct kmem_cache *s, void *object);
>>
>> +#define MODULE_ALIGN (PAGE_SIZE << KASAN_SHADOW_SCALE_SHIFT)
>> +
>> +int kasan_module_alloc(void *addr, size_t size);
>> +void kasan_module_free(void *addr);
>> +
>> #else /* CONFIG_KASAN */
>>
>> +#define MODULE_ALIGN 1
>
> Hmm, that should be PAGE_SIZE (we assume that in several places).
>
>> @@ -1807,6 +1808,7 @@ static void unset_module_init_ro_nx(struct module *mod) { }
>> void __weak module_memfree(void *module_region)
>> {
>> vfree(module_region);
>> + kasan_module_free(module_region);
>> }
>
> This looks racy (memory reuse?). Perhaps try other order?
>
You are right, it's racy. Concurrent kasan_module_alloc() could fail because
kasan_module_free() wasn't called/finished yet, so whole module_alloc() will fail
and module loading will fail.
However, I just find out that this race is not the worst problem here.
When vfree(addr) called in interrupt context, memory at addr will be reused for
storing 'struct llist_node':
void vfree(const void *addr)
{
...
if (unlikely(in_interrupt())) {
struct vfree_deferred *p = this_cpu_ptr(&vfree_deferred);
if (llist_add((struct llist_node *)addr, &p->list))
schedule_work(&p->wq);
In this case we have to free shadow *after* freeing 'module_region', because 'module_region'
is still used in llist_add() and in free_work() latter.
free_work() (in mm/vmalloc.c) processes list in LIFO order, so to free shadow after freeing
'module_region' kasan_module_free(module_region); should be called before vfree(module_region);
It will be racy still, but this is not so bad as potential crash that we have now.
Honestly, I have no idea how to fix this race nicely. Any suggestions?
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx. For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>