Re: [PATCH 1/8] time: Add persistent clock support

[John Stultz]

On Sat, Jun 23, 2018 at 5:14 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> On Wed, 13 Jun 2018, Baolin Wang wrote:
>> Moreover we can register the clocksource with CLOCK_SOURCE_SUSPEND_NONSTOP
>> to be one persistent clock, then we can simplify the suspend/resume
>> accounting by removing CLOCK_SOURCE_SUSPEND_NONSTOP timing. After that
>> we can only compensate the OS time by persistent clock or RTC.
>
> That makes sense because it adds a gazillion lines of code and removes 5?
> Not really,
>
>> +/**
>> + * persistent_clock_read_data - data required to read persistent clock
>> + * @read: Returns a cycle value from persistent clock.
>> + * @last_cycles: Clock cycle value at last update.
>> + * @last_ns: Time value (nanoseconds) at last update.
>> + * @mask: Bitmask for two's complement subtraction of non 64bit clocks.
>> + * @mult: Cycle to nanosecond multiplier.
>> + * @shift: Cycle to nanosecond divisor.
>> + */
>> +struct persistent_clock_read_data {
>> +     u64 (*read)(void);
>> +     u64 last_cycles;
>> +     u64 last_ns;
>> +     u64 mask;
>> +     u32 mult;
>> +     u32 shift;
>> +};
>> +/**
>> + * persistent_clock - represent the persistent clock
>> + * @read_data: Data required to read from persistent clock.
>> + * @seq: Sequence counter for protecting updates.
>> + * @freq: The frequency of the persistent clock.
>> + * @wrap: Duration for persistent clock can run before wrapping.
>> + * @alarm: Update timeout for persistent clock wrap.
>> + * @alarm_inited: Indicate if the alarm has been initialized.
>> + */
>> +struct persistent_clock {
>> +     struct persistent_clock_read_data read_data;
>> +     seqcount_t seq;
>> +     u32 freq;
>> +     ktime_t wrap;
>> +     struct alarm alarm;
>> +     bool alarm_inited;
>> +};
>
> NAK!
>
> There is no reason to invent yet another set of data structures and yet
> more read functions with a sequence counter. which are just a bad and
> broken copy of the existing timekeeping/clocksource code. And of course the
> stuff is not serialized against multiple registrations, etc. etc.
>
> Plus the utter nonsense that any call site has to do the same thing over
> and over:
>
>     register():
>     start_alarm_timer();
>
> Why is this required in the first place? It's not at all. The only place
> where such an alarm timer will be required is when the system actually goes
> to suspend. Starting it at registration time is pointless and even counter
> productive. Assume the clocksource wraps every 2 hours. So you start it at
> boot time and after 119 minutes uptime the system suspends. So it will
> wakeup one minute later to update the clocksource. Heck no. If the timer is
> started when the machine actually suspends it will wake up earliest in 120
> minutes.
>
> And you even add that to the TSC which does not need it at all. It will
> wrap in about 400 years on a 2GHZ machine. So you degrade the functionality
> instead of improving it.
>
> So no, this is not going anywhere.
>
> Let's look at the problem itself:
>
>    You want to use one clocksource for timekeeping during runtime which is
>    fast and accurate and another one for suspend time injection which is
>    slower and/or less accurate because the fast one stops in suspend.
>
>    Plus you need an alarmtimer which makes sure that the clocksource does
>    not wrap around during suspend.
>
> Now lets look what we have already:
>
>    Both clocksources already exist and are registered as clocksources with
>    all required data in the clocksource core.
>
> Ergo the only sane and logical conclusion is to expand the existing
> infrastructure to handle that.
>
> When a clocksource is registered, then the registration function already
> makes decisions about using it as timekeeping clocksource. So add a few
> lines of code to check whether the newly registered clocksource is suitable
> and preferred for suspend.
>
>         if (!stops_in_suspend(newcs)) {
>                 if (!suspend_cs || is_preferred_suspend_cs(newcs))
>                         suspend_cs = newcs;
>         }
>
> The is_preferred_suspend_cs() can be based on rating, the maximum suspend
> length which can be achieved or whatever is sensible. It should start of as
> a very simple decision function based on rating and not an prematurely
> overengineered monstrosity.
>
> The suspend/resume() code needs a few very simple changes:
>
> xxx_suspend():
>         clocksource_prepare_suspend();
>
>   Note, this is _NOT_ timekeeping_suspend() because that is invoked _AFTER_
>   alarmtimer_suspend(). So if an alarm timer is required it needs to be
>   armed before that. A trivial solution might be to just call it from
>   alarmtimer_suspend(), but that a minor detail to worry about.
>
> timekeeping_suspend()
> {
>         clocksource_enter_suspend();
>         ...
>
> timekeeping_resume()
> {
> ...
>         if (clocksource_leave_suspend(&nsec)) {
>                 ts_delta = ns_to_timespec64(nsec);
>                 sleeptime_injected = true;
>         } else if (......
>
>
> Now lets look at the new functions:
>
> void clocksource_prepare_suspend(void)
> {
>         if (!suspend_cs)
>                 return;
>
>         if (needs_alarmtimer(suspend_cs))
>                 start_suspend_alarm(suspend_cs);
> }
>
> void clocksource_enter_suspend(void)
> {
>         if (!suspend_cs)
>                 return;
>         suspend_start = suspend_cs->read();
> }
>
> bool clocksource_leave_suspend(u64 *nsec)
> {
>         u64 now, delta;
>
>         if (!suspend_cs)
>                 return false;
>
>         now = suspend_cs->read();
>         delta = clocksource_delta(now, suspend_start, suspend_cs->mask);
>         *nsec = mul_u64_u32_shr(delta, suspend_cs->mult, suspend_cs->shift);
>         return true;
> }
>
> See? It does not need any of this totally nonsensical stuff in your
> registration function and not any new read functions and whatever, because
> it simply can use the bog standard mult/shift values. Why? Because the
> conversion above can cope with a full 64 bit * 32 bit multiply without
> falling apart. It's already there in timekeeping_resume() otherwise
> resuming with a NONSTOP TSC would result in bogus sleep times after a few
> minutes. It's slower than the normal clocksource conversion which is
> optimized for performance, but thats completely irrelevant on resume. This
> whole blurb about requiring separate mult/shift values is just plain
> drivel.
>
> Plus any reasonably broad clocksource will not need an alarmtimer at
> all. Because the only reason it is needed is when the clocksource itself
> wraps around. And that has absolutely nothing to do with mult/shift
> values. That just depends on the frequency and the bitwidth of the counter,
>
> So it does not need an update function either because in case of broad
> enough clocksources there is absolutely no need for update and in case of
> wrapping ones the alarmtimer brings it out of suspend on time. And because
> the only interesting thing is the delta between suspend and resume this is
> all a complete non issue.
>
> The clocksource core already has all the registration/unregistration
> functionality plus an interface to reconfigure the frequency, so
> clocksources can come and go and be reconfigured and all of this just
> works.
>
> Once the extra few lines of code are in place, then you can go and cleanup
> the existing mess of homebrewn interfaces and claim that this is
> consolidation and simplification.
>
> <rant>
>
> What's wrong with you people? Didn't they teach you in school that the
> first thing to do is proper problem and code analysis? If they did not, go
> back to them and ask your money back,
>
> I'm really tired of these overengineered trainwrecks which are then
> advertised with bullshit marketing like the best invention since sliced
> bread. This might work in random corporates, but not here.
>
> </rant>

Thomas, what is wrong with *you*?  This is completely unnecessary.

First of all, I sent Baolin on this approach, as I was getting sick of
seeing the suspend/resume timing continually adding extra warts and
complexity to handle odd hardware that needs some solution. So I
suggested he clean that up in a way that lets these multiple ways of
solving the problem be done in device specific code w/o adding more
complexity to the core logic - in a fashion how the clocksources
allowed us to support nice fast/well-behaving hardware w/o preventing
ugly-but-neccessary solutions like the pit clocksource.

Now, I've also been quite a poor maintainer of late, and haven't done
much to help with pre-review of Baloin's code.  So that's on me, not
him.

Admittedly, my taste isn't always great, so there is likely to be a
better approach, and having your guidance here is *really* valued. I
just wish we could get it without all this unnecessary vinegar.

I was really lucky to have your guidance and support when I was
starting in the community. Your helping bring me up as a co-maintainer
(only a a relatively small set of responsibility compared to your much
wider maintainership), does let me see that having the responsibility
of billions of devices running your code is immense and comes with no
small amount of stress. Juggling the infinite incoming stream of
review requests on naieve or otherwise lacking code with other
important development priorities is a major burden, so *I really get
how frustrating it is*.

And its super annoying to have lots of short-term development requests
being thrown at you when you're the one who has to maintain it for the
long term.  But long-long-term, no one is going to be a maintainer
forever, and we are not going to develop more olympic swimmers if we
keep peeing in the pool.


Baolin: My apologies for leading you poorly here. Work on something
else for a day to let the flames burn off you, then review Thomas'
technical feedback, ignoring the harsh style, and try to address his
technical comments with the approach. I'll try to do better in finding
time to review your work.

-john