On 10.10.20 09:19, Andrei Vagin wrote:
> On Fri, Oct 09, 2020 at 03:28:15PM +0200, Christian Brauner wrote:
>> On Thu, Oct 08, 2020 at 07:39:42AM +0200, Michael Weiß wrote:
>>> getboottime64() provides the time stamp of system boot. In case of
>>> time namespaces, the offset to the boot time stamp was not applied
>>> earlier. However, getboottime64 is used e.g., in /proc/stat to print
>>> the system boot time to userspace. In container runtimes which utilize
>>> time namespaces to virtualize boottime of a container, this leaks
>>> information about the host system boot time.
>>>
>>> Therefore, we make getboottime64() to respect the time namespace offset
>>> for boottime by subtracting the boottime offset.
>>>
>>> Signed-off-by: Michael Weiß <michael.we...@aisec.fraunhofer.de>
>>> ---
>>> kernel/time/timekeeping.c | 3 +++
>>> 1 file changed, 3 insertions(+)
>>>
>>> diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c
>>> index 4c47f388a83f..67530cdb389e 100644
>>> --- a/kernel/time/timekeeping.c
>>> +++ b/kernel/time/timekeeping.c
>>> @@ -17,6 +17,7 @@
>>> #include <linux/clocksource.h>
>>> #include <linux/jiffies.h>
>>> #include <linux/time.h>
>>> +#include <linux/time_namespace.h>
>>> #include <linux/tick.h>
>>> #include <linux/stop_machine.h>
>>> #include <linux/pvclock_gtod.h>
>>> @@ -2154,6 +2155,8 @@ void getboottime64(struct timespec64 *ts)
>>> {
>>> struct timekeeper *tk = &tk_core.timekeeper;
>>> ktime_t t = ktime_sub(tk->offs_real, tk->offs_boot);
>>> + /* shift boot time stamp according to the timens offset */
>>> + t = timens_ktime_to_host(CLOCK_BOOTTIME, t);
>> Note that getbootime64() is mostly used in net/sunrpc and I don't know
>> if this change has any security implications for them.
> I would prefer to not patch kernel internal functions if they are used
> not only to expose time to the userspace.
>
> I think when kernel developers sees the getboottime64 function, they
> will expect that it returns the real time of kernel boot. They will
> not expect that it is aware of time namespaces and a returned time will
> depend on a task in which context it will be called.
>
> IMHO, as a minimum, we need to update the documentation for this function or
> even adjust a function name.
>
> And I think we need to consider an option to not change getbootime64 and
> apply a timens offset right in the show_stat(fs/proc/stat.c) function.
>
> Thanks,
> Andrei
Since the problems in softirq context mentioned from Thomas,
I would agree to Andrei's option to just patch proc/stat.c and leave
getboottime64 unchanged.
Digging around in the kernel tree, I just found /proc/stat as the only
place where boottime is exposed to userspace, thus it seems a valid
option.
What do you think? If you agree I'll come up with an updated patch-set.
Cheers,
Michael
