Re: [PATCH] mm, kasan: introduce a special shadow value for allocator metadata

[Andrey Ryabinin]

On 06/02/2016 03:02 PM, Alexander Potapenko wrote:
> On Wed, Jun 1, 2016 at 6:31 PM, Alexander Potapenko <glider@xxxxxxxxxx> wrote:
>> On Wed, Jun 1, 2016 at 5:23 PM, Andrey Ryabinin <aryabinin@xxxxxxxxxxxxx> wrote:
>>> On 05/31/2016 08:49 PM, Alexander Potapenko wrote:
>>>> On Tue, May 31, 2016 at 1:52 PM, Andrey Ryabinin
>>>> <aryabinin@xxxxxxxxxxxxx> wrote:
>>>>>
>>>>>
>>>>> On 05/31/2016 01:44 PM, Alexander Potapenko wrote:
>>>>>> Add a special shadow value to distinguish accesses to KASAN-specific
>>>>>> allocator metadata.
>>>>>>
>>>>>> Unlike AddressSanitizer in the userspace, KASAN lets the kernel proceed
>>>>>> after a memory error. However a write to the kmalloc metadata may cause
>>>>>> memory corruptions that will make the tool itself unreliable and induce
>>>>>> crashes later on. Warning about such corruptions will ease the
>>>>>> debugging.
>>>>>
>>>>> It will not. Whether out-of-bounds hits metadata or not is absolutely irrelevant
>>>>> to the bug itself. This information doesn't help to understand, analyze or fix the bug.
>>>>>
>>>> Here's the example that made me think the opposite.
>>>>
>>>> I've been reworking KASAN hooks for mempool and added a test that did
>>>> a write-after-free to an object allocated from a mempool.
>>>> This resulted in flaky kernel crashes somewhere in quarantine
>>>> shrinking after several attempts to `insmod test_kasan.ko`.
>>>> Because there already were numerous KASAN errors in the test, it
>>>> wasn't evident that the crashes were related to the new test, so I
>>>> thought the problem was in the buggy quarantine implementation.
>>>> However the problem was indeed in the new test, which corrupted the
>>>> quarantine pointer in the object and caused a crash while traversing
>>>> the quarantine list.
>>>>
>>>> My previous experience with userspace ASan shows that crashes in the
>>>> tool code itself puzzle the developers.
>>>> As a result, the users think that the tool is broken and don't believe
>>>> its reports.
>>>>
>>>> I first thought about hardening the quarantine list by checksumming
>>>> the pointers and validating them on each traversal.
>>>> This prevents the crashes, but doesn't give the users any idea about
>>>> what went wrong.
>>>> On the other hand, reporting the pointer corruption right when it happens does.
>>>> Distinguishing between a regular UAF and a quarantine corruption
>>>> (which is what the patch in question is about) helps to prioritize the
>>>> KASAN reports and give the developers better understanding of the
>>>> consequences.
>>>>
>>>
>>> After the first report we have memory in a corrupted state, so we are done here.
>> This is theoretically true, that's why we crash after the first report
>> in the userspace ASan.
>> But since the kernel proceeds after the first KASAN report, it's
>> possible that we see several different reports, and they are sometimes
>> worth looking at.
>>
>>> Anything that happens after the first report can't be trusted since it can be an after-effect,
>>> just like in your case. Such crashes are not worthy to look at.
>>> Out-of-bounds that doesn't hit metadata as any other memory corruption also can lead to after-effects crashes,
>>> thus distinguishing such bugs doesn't make a lot of sense.
>> Unlike the crashes in the kernel itself, crashes with KASAN functions
>> in the stack trace may make the developer think the tool is broken.
>>>
>>> test_kasan module is just a quick hack, made only to make sure that KASAN works.
>>> It does some crappy thing, and may lead to crash as well. So I would recommend an immediate
>>> reboot even after single attempt to load it.
>> Agreed. However a plain write into the first byte of the freed object
>> will cause similar problems.
> 
> On a second thought, we could do without the additional shadow byte
> value, by just comparing the address to the metadata offset.
> 

We could. But still, there is no point in doing anything like that.