Re: firewire: CLOCK_MONOTONIC_RAW exposure

[Daniel Thompson <daniel.thompson@linaro.org>]

On 17/07/14 17:14, John Stultz wrote:
> On 07/17/2014 01:37 AM, Daniel Thompson wrote:
>> On 16/07/14 16:00, Mark Brown wrote:
>>> On Wed, Jul 16, 2014 at 04:16:35PM +0200, Clemens Ladisch wrote:
>>>> (I don't have numbers for the errors caused by NTP adjustments.  Daniel?)
>>> Right, the goal is to get a clock which is guaranteed to never have any
>>> adjustments that might cause discontinuities or rate changes applied to
>>> it.  My understanding is that the users are already doing their own rate
>>> matching and it's much more important to them to get a stable clock than
>>> it is to get a clock at a specific nominal rate, and given the set top
>>> box applications I expect they also need this from very soon after boot.
>> We are trying to match rates with a broadcast device that "shouts" the
>> current time many times per second (MPEG transport stream PCR packets).
>> These packets are timestamped on arrival with a local clock and the
>> resulting data is used to recover the broadcast clock. However due to
>> variable transmission delay of the packets we require very long
>> control loops to extract any useful information from this data (varies
>> between five minutes and half and hour).
>>
>> An NTP rate correction can change the rate of CLOCK_MONOTONIC
>> sufficiently to confuse our clock recovery algorithms so we use
>> CLOCK_MONOTONIC_RAW as the master view of time.
> 
> Just to further clarify on this point, the problem is that with NTP
> there are both frequency (ie: clock runs too fast) and offset (ie: we're
> out of sync by 10ms) corrections made to CLOCK_REALTIME. 
> 
> In the long-term when we've synced up with NTP, these are basically the
> same thing, so to keep things (relatively) simple we eventually combine
> these into one adjustment factor when steering the clock.  But in the
> short-term when we're trying to quickly get in sync with NTP, the offset
> correction can be fairly large.
> 
> The problem is that we want CLOCK_MONOTONIC to be frequency corrected,
> so that a second is really a second. But we don't really care about it
> being offset corrected. However, since its much simpler to define a
> fixed offset between _MONOTONIC and _REALTIME (which is only modified if
> _REALTIME is set or stepped).

Interesting. That certainly explains *why* our algorithm breaks!

I admit I was curious why having the clock tick more accurately part way
through the data gathering caused our sync algorithms to break (although
clearly not curious enough). However even a pretty gradual change
towards a new offset would certainly cause lots of problems for these
use cases.


> Ideally I guess we'd probably want to track the freq adjustment and
> offset adjustment separately and apply the freq offset to both
> _MONONTONIC and _REALTIME, but only apply offset corrections to
> _REALTIME. However, this would make the accounting much more complex and
> would break the fixed relationship between _MONOTONIC and _REALTIME.
> 
> Miroslav has discussed trying this previously, but I'm a bit skeptical
> its worth the extra effort and overhead.

Certainly the use case I presented is pretty niche and, intrinsically
non-portable.

In our case we make the "non-portable" assumption that the hardware
oscillator feeding CLOCK_MONOTONIC_RAW is of high enough quality for the
rest of the SoC to function correctly.

That's about the only assumption though.


> CLOCK_MONOTONIC_RAW provide just a nanosecond abstraction of a hardware
> counter. It was added because some folks who were doing time sync
> algorithms were using non-portable methods like rdtsc to measure
> corrections being made (as measuring correction with the time base being
> corrected is a bit circular). So in cases where the short-term
> adjustment is problematic, it can be a good choice, as long as accuracy
> needs are low (since a second may not be a real second).