[PATCH v4 1/3] arm64: Improve kernel address detection of __is_lm_address()

[Vincenzo Frascino]

On 1/26/21 12:07 PM, Catalin Marinas wrote:
> On Tue, Jan 26, 2021 at 11:58:13AM +0000, Vincenzo Frascino wrote:
>> On 1/25/21 5:56 PM, Catalin Marinas wrote:
>>> On Mon, Jan 25, 2021 at 04:09:57PM +0000, Vincenzo Frascino wrote:
>>>> On 1/25/21 2:59 PM, Catalin Marinas wrote:
>>>>> On Mon, Jan 25, 2021 at 02:36:34PM +0000, Vincenzo Frascino wrote:
>>>>>> On 1/25/21 1:02 PM, Mark Rutland wrote:
>>>>>>> On Fri, Jan 22, 2021 at 03:56:40PM +0000, Vincenzo Frascino wrote:
>>>>>>>> Currently, the __is_lm_address() check just masks out the top 12 bits
>>>>>>>> of the address, but if they are 0, it still yields a true result.
>>>>>>>> This has as a side effect that virt_addr_valid() returns true even for
>>>>>>>> invalid virtual addresses (e.g. 0x0).
>>>>>>>>
>>>>>>>> Improve the detection checking that it's actually a kernel address
>>>>>>>> starting at PAGE_OFFSET.
>>>>>>>>
>>>>>>>> Cc: Catalin Marinas <catalin.marinas at arm.com>
>>>>>>>> Cc: Will Deacon <will at kernel.org>
>>>>>>>> Suggested-by: Catalin Marinas <catalin.marinas at arm.com>
>>>>>>>> Reviewed-by: Catalin Marinas <catalin.marinas at arm.com>
>>>>>>>> Signed-off-by: Vincenzo Frascino <vincenzo.frascino at arm.com>
>>>>>>>
>>>>>>> Looking around, it seems that there are some existing uses of
>>>>>>> virt_addr_valid() that expect it to reject addresses outside of the
>>>>>>> TTBR1 range. For example, check_mem_type() in drivers/tee/optee/call.c.
>>>>>>>
>>>>>>> Given that, I think we need something that's easy to backport to stable.
>>>>>>>
>>>>>>
>>>>>> I agree, I started looking at it this morning and I found cases even in the main
>>>>>> allocators (slub and page_alloc) either then the one you mentioned.
>>>>>>
>>>>>>> This patch itself looks fine, but it's not going to backport very far,
>>>>>>> so I suspect we might need to write a preparatory patch that adds an
>>>>>>> explicit range check to virt_addr_valid() which can be trivially
>>>>>>> backported.
>>>>>>>
>>>>>>
>>>>>> I checked the old releases and I agree this is not back-portable as it stands.
>>>>>> I propose therefore to add a preparatory patch with the check below:
>>>>>>
>>>>>> #define __is_ttrb1_address(addr)	((u64)(addr) >= PAGE_OFFSET && \
>>>>>> 					(u64)(addr) < PAGE_END)
>>>>>>
>>>>>> If it works for you I am happy to take care of it and post a new version of my
>>>>>> patches.
>>>>>
>>>>> I'm not entirely sure we need a preparatory patch. IIUC (it needs
>>>>> checking), virt_addr_valid() was fine until 5.4, broken by commit
>>>>> 14c127c957c1 ("arm64: mm: Flip kernel VA space"). Will addressed the
>>>>> flip case in 68dd8ef32162 ("arm64: memory: Fix virt_addr_valid() using
>>>>> __is_lm_address()") but this broke the <PAGE_OFFSET case. So in 5.4 a
>>>>> NULL address is considered valid.
>>>>>
>>>>> Ard's commit f4693c2716b3 ("arm64: mm: extend linear region for 52-bit
>>>>> VA configurations") changed the test to no longer rely on va_bits but
>>>>> did not change the broken semantics.
>>>>>
>>>>> If Ard's change plus the fix proposed in this test works on 5.4, I'd say
>>>>> we just merge this patch with the corresponding Cc stable and Fixes tags
>>>>> and tweak it slightly when doing the backports as it wouldn't apply
>>>>> cleanly. IOW, I wouldn't add another check to virt_addr_valid() as we
>>>>> did not need one prior to 5.4.
>>>>
>>>> Thank you for the detailed analysis. I checked on 5.4 and it seems that Ard
>>>> patch (not a clean backport) plus my proposed fix works correctly and solves the
>>>> issue.
>>>
>>> I didn't mean the backport of the whole commit f4693c2716b3 as it
>>> probably has other dependencies, just the __is_lm_address() change in
>>> that patch.
>>
>> Then call it preparatory patch ;)
> 
> It's preparatory only for the stable backports, not for current
> mainline. But I'd rather change the upstream patch when backporting to
> apply cleanly, no need for a preparatory stable patch.
> 

Thanks for the clarification.

-- 
Regards,
Vincenzo