[PATCH] ntp: Fix integer overflow when setting time

[PATCH] ntp: Fix integer overflow when setting time

[Sasha Levin]

On 64bit machines, the difference between 'time_reftime' and current time
is 8 bits, this variable is used as the divisor in div_s64() to calculate
the frequency, but div_s64() assumes the divisor is 32bits.

This means that we are able to trigger a division by zero if the difference
has 0's in it's lower bytes. This way it would skip the sanity checks before
the div_64s() but when it gets to that div it would get truncated and
become 0.

This causes the following error:

[  135.947035] divide error: 0000 [#1] PREEMPT SMP DEBUG_PAGEALLOC
[  135.947095] CPU 3
[  135.947095] Pid: 28166, comm: trinity Tainted: G        W    3.3.0-rc6-next-20120309-sasha-00001-g10cf0bc-dirty #46
[  135.947095] RIP: 0010:[<ffffffff811066b0>]  [<ffffffff811066b0>] ntp_update_offset+0xb0/0x160
[  135.947095] RSP: 0018:ffff88001baa3e08  EFLAGS: 00010046
[  135.947095] RAX: 007d000000000000 RBX: 0000000001f40000 RCX: 0000000000000000
[  135.947095] RDX: 0000000000000000 RSI: ffffffff83240fd8 RDI: 0000000001f40000
[  135.947095] RBP: ffff88001baa3e18 R08: 0000000000000000 R09: 0000000000000001
[  135.947095] R10: 0000000000000000 R11: 0000000000000000 R12: 0000019800000000
[  135.947095] R13: ffffffff83b7b480 R14: 0000000001154600 R15: ffff88001baa3e78
[  135.947095] FS:  00007f11348c6700(0000) GS:ffff88003e400000(0000) knlGS:0000000000000000
[  135.947095] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  135.947095] CR2: 0000000001155608 CR3: 00000000050b2000 CR4: 00000000000406e0
[  135.947095] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  135.947095] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
[  135.947095] Process trinity (pid: 28166, threadinfo ffff88001baa2000, task ffff88001b7cb000)
[  135.947095] Stack:
[  135.947095]  ffff88001baa3e78 00000000ffffffff ffff88001baa3e58 ffffffff81106e19
[  135.947095]  0000000001154600 2c6d37333b315b1b 0000019800000017 0000000001e08c17
[  135.947095]  0000000000000000 fffffffffffffff2 ffff88001baa3e68 ffffffff810d14dc
[  135.947095] Call Trace:
[  135.947095]  [<ffffffff81106e19>] do_adjtimex+0x3c9/0x6d0
[  135.947095]  [<ffffffff810d14dc>] posix_clock_realtime_adj+0xc/0x10
[  135.947095]  [<ffffffff810d29a6>] sys_clock_adjtime+0xc6/0xf0
[  135.947095]  [<ffffffff8270073d>] system_call_fastpath+0x1a/0x1f
[  135.947095] Code: 02 7e 2b 49 81 fc 00 08 00 00 7f 05 83 e1 08 74 1d 80 ce 40 48 89 d8 89 15 6e a9 13 02 48 c1 e0 1e 49 63 cc 48 89 c2 48 c1 fa 3f <48> f7 f9 48 03 05 46 b4 a7 02 48 8b 35 67 a9 13 02 ba 01 00 00
[  135.947095] RIP  [<ffffffff811066b0>] ntp_update_offset+0xb0/0x160
[  135.947095]  RSP <ffff88001baa3e08>
[  135.947095] ---[ end trace f995700f304a6fef ]---

Signed-off-by: Sasha Levin <levinsasha928@gmail.com>
---
 kernel/time/ntp.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)

diff --git a/kernel/time/ntp.c b/kernel/time/ntp.c
index 17fb1b9..efe894a 100644
--- a/kernel/time/ntp.c
+++ b/kernel/time/ntp.c
@@ -289,7 +289,7 @@ static inline s64 ntp_update_offset_fll(s64 offset64, long secs)
 
 	time_status |= STA_MODE;
 
-	return div_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);
+	return div64_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);
 }
 
 static void ntp_update_offset(long offset)
-- 
1.7.8.4

Re: [PATCH] ntp: Fix integer overflow when setting time

[Thomas Gleixner]

On Wed, 14 Mar 2012, Sasha Levin wrote:

> On 64bit machines, the difference between 'time_reftime' and current time
> is 8 bits, this variable is used as the divisor in div_s64() to calculate
> the frequency, but div_s64() assumes the divisor is 32bits.

Why is that delta 8 bits???? 

> This means that we are able to trigger a division by zero if the difference
> has 0's in it's lower bytes. This way it would skip the sanity checks before
> the div_64s() but when it gets to that div it would get truncated and
> become 0.

This does not make sense at all.

So what I assume you are talking about is, that the divisor (i.e. the
delta) is greater than (1 << 32) - 1 and all 32 lower bits are 0, right ?

> diff --git a/kernel/time/ntp.c b/kernel/time/ntp.c
> index 17fb1b9..efe894a 100644
> --- a/kernel/time/ntp.c
> +++ b/kernel/time/ntp.c
> @@ -289,7 +289,7 @@ static inline s64 ntp_update_offset_fll(s64 offset64, long secs)
>  
>  	time_status |= STA_MODE;
> 
> -	return div_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);
> +	return div64_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);

This needs a comment and it would be nice to avoid the 64 division on
32bit machines. Something like:

#if BITS_PER_LONG == 64 
# define div64_long(x,y)	div64_s64(x,y)
#else
# define div64_long(x,y)	div_s64(x,y)
#endif

Nice catch, though it took me a while to grok the changelog, as the 8
bit case is catched by the MINSEC check. Please be more careful about
that.

Thanks,

	tglx

Re: [PATCH] ntp: Fix integer overflow when setting time

[Sasha Levin]

On Wed, Mar 14, 2012 at 11:16 PM, Thomas Gleixner <tglx@linutronix.de> wrote:
> On Wed, 14 Mar 2012, Sasha Levin wrote:
>
>> On 64bit machines, the difference between 'time_reftime' and current time
>> is 8 bits, this variable is used as the divisor in div_s64() to calculate
>> the frequency, but div_s64() assumes the divisor is 32bits.
>
> Why is that delta 8 bits????
>
>> This means that we are able to trigger a division by zero if the difference
>> has 0's in it's lower bytes. This way it would skip the sanity checks before
>> the div_64s() but when it gets to that div it would get truncated and
>> become 0.
>
> This does not make sense at all.
>
> So what I assume you are talking about is, that the divisor (i.e. the
> delta) is greater than (1 << 32) - 1 and all 32 lower bits are 0, right ?

Yup, exactly.

>> diff --git a/kernel/time/ntp.c b/kernel/time/ntp.c
>> index 17fb1b9..efe894a 100644
>> --- a/kernel/time/ntp.c
>> +++ b/kernel/time/ntp.c
>> @@ -289,7 +289,7 @@ static inline s64 ntp_update_offset_fll(s64 offset64, long secs)
>>
>>       time_status |= STA_MODE;
>>
>> -     return div_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);
>> +     return div64_s64(offset64 << (NTP_SCALE_SHIFT - SHIFT_FLL), secs);
>
> This needs a comment and it would be nice to avoid the 64 division on
> 32bit machines. Something like:
>
> #if BITS_PER_LONG == 64
> # define div64_long(x,y)        div64_s64(x,y)
> #else
> # define div64_long(x,y)        div_s64(x,y)
> #endif

I'll change that.

> Nice catch, though it took me a while to grok the changelog, as the 8
> bit case is catched by the MINSEC check. Please be more careful about
> that.

Sorry about that, I suck at writing changelogs :(

> Thanks,
>
>        tglx