Re: mm: BUG in resv_map_release

[Dmitry Vyukov]

On Thu, Mar 23, 2017 at 7:02 PM, Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
> On 03/23/2017 10:25 AM, Mike Kravetz wrote:
>> On 03/23/2017 03:19 AM, Dmitry Vyukov wrote:
>>> Hello,
>>>
>>> I've got the following BUG while running syzkaller fuzzer.
>>> Note the injected kmalloc failure, most likely it's the root cause.
>>
>> Thanks  Dmitry,
>>
>> The BUG indicates someone called region_chg() in the process of adding
>> a hugetlbfs page reservation, but did not complete this 'two step'
>> process with a call to region_add() or region_abort().  Most likely a
>> missed call in an error path somewhere.  :(
>>
>> I'll try to track this down.  The hint of 'injected kmalloc failure'
>> should help.
>
> Actually, in this case I believe the bug is in hugetlb_reserve_pages.
> It calls region_chg(), but gets an error due to the injected kmalloc
> failure.  At this point, the resv_map->adds_in_progress is 0 as it
> should be.  However, the error path for hugetlb_reserve_pages calls
> region_abort() which will unconditionally decrement adds_in_progress.
> So, adds_in_progress goes negative and we eventually BUG.  :(
>
> I'll look for other misuses of region_chg()/region_add()/region_abort()
> and put together a patch.
>
> Dmitry, is there some way to run the fuzzer with kmalloc failure injection
> and target the hugetlbfs code?  I'm suspect we could flush out other bugs.
> I noticed one other you discovered, and will look at that next.

syzkaller systematically targets all of the kernel code. So far I've
seen only these 2 involving hugetlbfs code. I don't think we need to
do anything special for hugetlbfs.