Re: LTTng + dynamick tick

* Deepak Saxena ([email protected]) wrote:
> 
> Hi,
> 
> I'm working on integrating LTTng into the MontaVista kernel which is
> basically 2.6.18 + the RT patch. One of the thing this patch adds and
> that we plan on supporting is dynamic ticks. Does the existing code 
> rely on having a timer tick? If so, how difficult of a task is it to
> not do so?
> 
> Tnx,
> ~Deepak
>

Hi Deepak,

I have put Thomas in CC, as he may think this is relevant to a wider range of
users.

We could try to use a modified version of 
include/linux/ktime.h : ktime_get_ts

It gets a monotonic time, which is good.
timespec is in ns, which is good too.
Now, about locking : 

kernel/hrtimer.c :

void ktime_get_ts(struct timespec *ts)
{
        struct timespec tomono;
        unsigned long seq;

        do {
                seq = read_seqbegin(&xtime_lock);
                getnstimeofday(ts);
                tomono = wall_to_monotonic;

        } while (read_seqretry(&xtime_lock, seq));

        set_normalized_timespec(ts, ts->tv_sec + tomono.tv_sec,
                                ts->tv_nsec + tomono.tv_nsec);
}

The problem here comes if the logging code is nested on the write lock. We need
to use a hack like the one found in include/asm-arm/ltt.h to fail after a few
loops.

In getnstimeofday :

void getnstimeofday (struct timespec *tv)
{
        unsigned long seq,sec,nsec;

        do {
                seq = read_seqbegin(&xtime_lock);
                sec = xtime.tv_sec;
                nsec = xtime.tv_nsec+time_interpolator_get_offset();
        } while (unlikely(read_seqretry(&xtime_lock, seq)));

        while (unlikely(nsec >= NSEC_PER_SEC)) {
                nsec -= NSEC_PER_SEC;
                ++sec;
        }
        tv->tv_sec = sec;
        tv->tv_nsec = nsec;
}

Another seqlock... same thing. There should be a _getnstimeofday which does not
take a read seqlock, as it is already taken by ktime_get_ts.

Now for timer.c : time_interpolator_get_offset :
It calls time_interpolator_get_counter(0), which uses a cmpxchg (this is good
for reentrancy!) :)

So, over all, I think we could switch to hrtimers for architectures which lack
proper synchronized TSC support, but with the seqlock tweaks I just mentioned.

Mathieu



OpenPGP public key:              http://krystal.dyndns.org:8080/key/compudj.gpg
Key fingerprint:     8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68