Re: [PATCH RESEND] mm/hugetlb: Don't call region_abort if region_chg fails

[Mike Kravetz]

On 04/10/2017 02:38 PM, Vegard Nossum wrote:
> On 29 March 2017 at 23:08, Mike Kravetz <mike.kravetz@xxxxxxxxxx> wrote:
>> Changes to hugetlbfs reservation maps is a two step process.  The first
>> step is a call to region_chg to determine what needs to be changed, and
>> prepare that change.  This should be followed by a call to call to
>> region_add to commit the change, or region_abort to abort the change.
>>
>> The error path in hugetlb_reserve_pages called region_abort after a
>> failed call to region_chg.  As a result, the adds_in_progress counter
>> in the reservation map is off by 1.  This is caught by a VM_BUG_ON
>> in resv_map_release when the reservation map is freed.
>>
>> syzkaller fuzzer found this bug, that resulted in the following:
>>
>>  kernel BUG at mm/hugetlb.c:742!
>>  Call Trace:
>>   hugetlbfs_evict_inode+0x7b/0xa0 fs/hugetlbfs/inode.c:493
>>   evict+0x481/0x920 fs/inode.c:553
>>   iput_final fs/inode.c:1515 [inline]
>>   iput+0x62b/0xa20 fs/inode.c:1542
>>   hugetlb_file_setup+0x593/0x9f0 fs/hugetlbfs/inode.c:1306
>>   newseg+0x422/0xd30 ipc/shm.c:575
>>   ipcget_new ipc/util.c:285 [inline]
>>   ipcget+0x21e/0x580 ipc/util.c:639
>>   SYSC_shmget ipc/shm.c:673 [inline]
>>   SyS_shmget+0x158/0x230 ipc/shm.c:657
>>   entry_SYSCALL_64_fastpath+0x1f/0xc2
>>  RIP: resv_map_release+0x265/0x330 mm/hugetlb.c:742
>>
>> Reported-by: Dmitry Vyukov <dvyukov@xxxxxxxxxx>
>> Signed-off-by: Mike Kravetz <mike.kravetz@xxxxxxxxxx>
>> Acked-by: Hillf Danton <hillf.zj@xxxxxxxxxxxxxxx>
>> ---
>>  mm/hugetlb.c | 4 +++-
>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>> index c7025c1..c65d45c 100644
>> --- a/mm/hugetlb.c
>> +++ b/mm/hugetlb.c
>> @@ -4233,7 +4233,9 @@ int hugetlb_reserve_pages(struct inode *inode,
>>         return 0;
>>  out_err:
>>         if (!vma || vma->vm_flags & VM_MAYSHARE)
>> -               region_abort(resv_map, from, to);
>> +               /* Don't call region_abort if region_chg failed */
>> +               if (chg >= 0)
>> +                       region_abort(resv_map, from, to);
>>         if (vma && is_vma_resv_set(vma, HPAGE_RESV_OWNER))
>>                 kref_put(&resv_map->refs, resv_map_release);
>>         return ret;
> 
> Hi guys,
> 
> I'm running into this on latest linus/master:
> 
> kernel BUG at mm/hugetlb.c:742!
> invalid opcode: 0000 [#1] SMP KASAN
> CPU: 3 PID: 20281 Comm: syz-executor0 Not tainted 4.11.0-rc6 #335
> Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS
> Ubuntu-1.8.2-1ubuntu1 04/01/2014
> task: ffff880064f30dc0 task.stack: ffff880065b38000
> RIP: 0010:resv_map_release+0x1cb/0x270
> RSP: 0018:ffff880065b3fc38 EFLAGS: 00010287
> RAX: 0000000000010000 RBX: ffff88006b5fe418 RCX: ffffc90001b52000
> RDX: 00000000000005de RSI: ffffffff8172026b RDI: ffff88006b5fe410
> RBP: ffff880065b3fc78 R08: ffff880065b3f958 R09: 0000000000000000
> R10: 0000000000000000 R11: 0000000000000000 R12: dffffc0000000000
> R13: ffff88006b5fe418 R14: ffff88006b5fe418 R15: ffff88006b5fe418
> FS:  00007f21647c5700(0000) GS:ffff88006d100000(0000) knlGS:0000000000000000
> CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> CR2: 0000000000460750 CR3: 000000005d123000 CR4: 00000000000006e0
> Call Trace:
>  hugetlbfs_evict_inode+0x80/0xa0
>  ? hugetlbfs_setattr+0x3c0/0x3c0
>  evict+0x24a/0x620
>  iput+0x48f/0x8c0
>  dentry_unlink_inode+0x31f/0x4d0
>  __dentry_kill+0x292/0x5e0
>  dput+0x730/0x830
>  __fput+0x438/0x720
>  ____fput+0x1a/0x20
>  task_work_run+0xfe/0x180
>  exit_to_usermode_loop+0x133/0x150
>  syscall_return_slowpath+0x184/0x1c0
>  entry_SYSCALL_64_fastpath+0xab/0xad
> 
> To reproduce:
> 
> mmap(0, 0x2000, 0, 0x40031, 0xffffffffffffffffULL, 0x8000000000000000ULL);
> 
> Curiously enough, it's the patch from this thread (i.e. commit
> ff8c0c53c47530ffea82c22a0a6df6332b56c957) that introduces it,
> according to git bisect. Reverting the commit from linus/master fixes
> the problem.

Thanks for finding this.

I do not think commit ff8c0c53 is the root cause of this BUG/issue.

Due to the very high offset (0x8000000000000000ULL) passed to mmap, there
is some overflow and/or truncation of values happening before getting to
the hugetlbfs reservation code.  The routine hugetlb_reserve_pages() is
passed a negative page offset value (from=4398046511104, to=-4398046511103).
Bad!!!  The routine region_chg() takes these values to determine how many
reservations are needed and calculates/returns a negative value.  This
appears as an error.  So, the code from commit ff8c0c53 prevents the call
to region_abort(), adds_in_progress does not get decremented and we hit the
BUG.

We should have never calculated and acted upon negative page offsets.  It
was just 'lucky' that things appeared to work before this commit.  I have
not yet determined all the things that could have gone wrong when passing
around these incorrect values.

I believe commit ff8c0c53 should remain.  I will start working on a fix
to this overflow and/or truncation of page offsets.

-- 
Mike Kravetz

> 
> Also found by syzcaller (no fault injections this time).
> 
> 
> Vegard
> 
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to majordomo@xxxxxxxxxx  For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>
>