[PATCH] time: Fix get_ticks being non-monotonic

[Sean Anderson]

On 9/7/20 9:57 AM, Simon Glass wrote:
> Hi Sean,
> 
> On Sun, 6 Sep 2020 at 20:02, Sean Anderson <seanga2 at gmail.com> wrote:
>>
>> On 9/6/20 9:43 PM, Simon Glass wrote:
>>> Hi Sean,
>>>
>>> On Tue, 1 Sep 2020 at 13:56, Sean Anderson <seanga2 at gmail.com> wrote:
>>>>
>>>> get_ticks does not always succeed. Sometimes it can be called before the
>>>> timer has been initialized. If it does, it returns a negative errno.
>>>> This causes the timer to appear non-monotonic, because the value will
>>>> become much smaller after the timer is initialized.
>>>>
>>>> No users of get_ticks which I checked handle errors of this kind. Further,
>>>> functions like tick_to_time mangle the result of get_ticks, making it very
>>>> unlikely that one could check for an error without suggesting a patch such
>>>> as this one.
>>>>
>>>> This patch changes get_ticks to always return 0 when there is an error.
>>>> 0 is the least unsigned integer, ensuring get_ticks appears monotonic. This
>>>> has the side effect of time apparently not passing until the timer is
>>>> initialized. However, without this patch, time does not pass anyway,
>>>> because the error value is likely to be the same.
>>>>
>>>> Fixes: c8a7ba9e6a5
>>>> Signed-off-by: Sean Anderson <seanga2 at gmail.com>
>>>> ---
>>>>
>>>>  lib/time.c | 4 ++--
>>>>  1 file changed, 2 insertions(+), 2 deletions(-)
>>>
>>> Would it be better to panic so people can fix the bug?
>>
>> I thought this was expected behavior. It's only a bug if you do
>> something like udelay before any timers are created. We just can't
>> report errors through get_ticks, because its users assume that it always
>> returns a time of some kind.
> 
> I think it indicates a bug. If you use a device before it is ready you
> don't really know what it will do. I worry that this patch is just
> going to cause confusion, since the behaviour depends on when you call
> it. If we panic, people can figure out why the timer is being inited
> too late, or being used too early.

Hm, maybe. I don't think it's as clear cut as "us[ing] a device before
it is ready," because get_ticks tries to initialize the timer if it
isn't already initialized. Unless someone else does it first, the first
call to get_ticks will always be before the timer is initialized.

The specific problem I ran into was that after relocation, the watchdog
may be initialized before the timer. This occurs on RISC-V because
without [1] a timer only exists after arch_early_init_r. So, for the
first few calls to watchdog_reset there is no timer.
 
The second return could probably be turned into a panic. I checked, and
all current timer drivers always succeed in getting the time (except for
the RISC-V timer, which is fixed in [1]), so the only way for
timer_get_count to fail is if timer_ops.get_count doesn't exist. That is
almost certainly an error on the driver author's part, so I think
panicking there is the only reasonable option.

(Does get_count even need to have a return value? I think it's
reasonable to always expect the timer to return a value.)

--Sean

[1] https://patchwork.ozlabs.org/project/uboot/list/?series=198797