Re: LTTng + dynamick tick

Hi Thomas,

* Thomas Gleixner ([email protected]) wrote:
> > 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.
> 
> These tweaks are horrible without having seen them.
> 

They look like :
static inline u32 ltt_get_timestamp32(void)
{
        unsigned long seq;
        unsigned long try = 5;
        u32 ret;

        do {
                seq = read_seqbegin(&xtime_lock);
                ret = (jiffies << LTTNG_LOGICAL_SHIFT)
                        | (atomic_add_return(1, &lttng_logical_clock));
        } while(read_seqretry(&xtime_lock, seq) && (--try) > 0);

        if(try == 0)
                return 0;
        else
                return ret;
}

> The only way you can deadlock is when you are inside a section which
> holds xtime_lock. So in your locking path all you need is:
> 
> 	if (!is_writelocked_seqlock(&xtime_lock))
> 		now = ktime_get();
> 	else
> 		now = (ktime_t) {.tv64 = 0 };
> 
> and in seqlock.h
> 
> static inline int is_writelocked_seqlock(struct seqlock *sl)
> {
> 	return spin_is_locked(&sl->lock);
> }
> 
> You may get some false positives, but this should not matter much.
> 

The problem with your approach is that there will be a lot of false positives
on SMP systems. The problem of NMI nested over a write lock is rare and we
should not create false positives by the way we handle it.

The other way around would be to rework the xtime and wall_to_monotonic values
so they would use a pointer to an array and make the update atomic (a preempt
disable is needed for the reader). I have done such tricks in ltt-heartbeat.c.
It would have to be investigated to see if it could be applied to get rid of
the seqlock.

Mathieu


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