Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> writes:
> On 02/25/2015 09:25 AM, Rusty Russell wrote:
>> Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> writes:
>>> On 02/23/2015 11:26 AM, Rusty Russell wrote:
>>>> Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> writes:
>>>>> On 02/20/2015 03:15 AM, Rusty Russell wrote:
>>>>>> Andrey Ryabinin <a.ryabinin@xxxxxxxxxxx> writes:
>>>>>>> On 02/19/2015 02:10 AM, Rusty Russell wrote:
>>>>>>>> This is not portable. Other archs don't use vmalloc, or don't use
>>>>>>>> (or define) MODULES_VADDR. If you really want to hook here, you'd
>>>>>>>> need a new flag (or maybe use PAGE_KERNEL_EXEC after an audit).
>>>>>>>>
>>>>>>>
>>>>>>> Well, instead of explicit (addr >= MODULES_VADDR && addr < MODULES_END)
>>>>>>> I could hide this into arch-specific function: 'kasan_need_to_allocate_shadow(const void *addr)'
>>>>>>> or make make all those functions weak and allow arch code to redefine them.
>>>>>>
>>>>>> That adds another layer of indirection. And how would the caller of
>>>>>> plain vmalloc() even know what to return?
>>>>>>
>>>>>
>>>>> I think I don't understand what do you mean here. vmalloc() callers shouldn't know
>>>>> anything about kasan/shadow.
>>>>
>>>> How else would kasan_need_to_allocate_shadow(const void *addr) work for
>>>> architectures which don't have a reserved vmalloc region for modules?
>>>>
>>>
>>>
>>> I think I need to clarify what I'm doing.
>>>
>>> Address sanitizer algorithm in short:
>>> -------------------------------------
>>> Every memory access is transformed by the compiler in the following way:
>>>
>>> Before:
>>> *address = ...;
>>>
>>> after:
>>>
>>> if (memory_is_poisoned(address)) {
>>> report_error(address, access_size);
>>> }
>>> *address = ...;
>>>
>>> where memory_is_poisoned():
>>> bool memory_is_poisoned(unsigned long addr)
>>> {
>>> s8 shadow_value = *(s8 *)kasan_mem_to_shadow((void *)addr);
>>> if (unlikely(shadow_value)) {
>>> s8 last_accessible_byte = addr & KASAN_SHADOW_MASK;
>>> return unlikely(last_accessible_byte >= shadow_value);
>>> }
>>> return false;
>>> }
>>> --------------------------------------
>>>
>>> So shadow memory should be present for every accessible address in kernel
>>> otherwise it will be unhandled page fault on reading shadow value.
>>>
>>> Shadow for vmalloc addresses (on x86_64) is readonly mapping of one zero page.
>>> Zero byte in shadow means that it's ok to access to that address.
>>> Currently we don't catch bugs in vmalloc because most of such bugs could be caught
>>> in more simple way with CONFIG_DEBUG_PAGEALLOC.
>>> That's why we don't need RW shadow for vmalloc, it just one zero page that readonly
>>> mapped early on boot for the whole [kasan_mem_to_shadow(VMALLOC_START, kasan_mem_to_shadow(VMALLOC_END)] range
>>> So every access to vmalloc range assumed to be valid.
>>>
>>> To catch out of bounds accesses in global variables we need to fill shadow corresponding
>>> to variable's redzone with non-zero (negative) values.
>>> So for kernel image and modules we need a writable shadow.
>>>
>>> If some arch don't have separate address range for modules and it uses general vmalloc()
>>> shadow for vmalloc should be writable, so it means that shadow has to be allocated
>>> for every vmalloc() call.
>>>
>>> In such arch kasan_need_to_allocate_shadow(const void *addr) should return true for every vmalloc address:
>>> bool kasan_need_to_allocate_shadow(const void *addr)
>>> {
>>> return addr >= VMALLOC_START && addr < VMALLOC_END;
>>> }
>>
>> Thanks for the explanation.
>>
>>> All above means that current code is not very portable.
>>> And 'kasan_module_alloc(p, size) after module alloc' approach is not portable
>>> too. This won't work for arches that use [VMALLOC_START, VMALLOC_END] addresses for modules,
>>> because now we need to handle all vmalloc() calls.
>>
>> I'm confused. That's what you do now, and it hasn't been a problem,
>> has it? The problem is on the freeing from interrupt context...
>>
>
> It's not problem now. It's only about portability.
Your first patch in this conversation says "Current approach in handling
shadow memory for modules is broken."
>> #define VM_KASAN 0x00000080 /* has shadow kasan map */
>>
>> Set that in kasan_module_alloc():
>>
>> if (ret) {
>> struct vm_struct *vma = find_vm_area(addr);
>>
>> BUG_ON(!vma);
>> /* Set VM_KASAN so vfree() can free up shadow. */
>> vma->flags |= VM_KASAN;
>> }
>>
>> And check that in __vunmap():
>>
>> if (area->flags & VM_KASAN)
>> kasan_module_free(addr);
>>
>> That is portable, and is actually a fairly small patch on what you
>> have at the moment.
>>
>> What am I missing?
>>
>
> That is not portable.
> Architectures that don't have separate region for modules should allocate shadow
> for every vmalloc() call, not only for modules.
OK, I didn't appreciate that. But couldn't you still use the "R/O
shared zero page shadow" for vmalloc, and have kasan_module_alloc()
simply replace the pages with r/w ones (and kasan_module_free()
would have to remove it again).
> Actually I'm fine with what you are proposing here. I think that portability issues could be fixed
> latter when this will become a real problem.
OK.
Thanks for your patience!
Rusty.
--
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>