[DTrace-devel] semantics of "timestamp"

[Kris Van Hees]

On Thu, Jul 16, 2020 at 12:57:17PM -0700, Eugene Loh wrote:
> What are the semantics of the built-in variable "timestamp"? 
> Specifically, does it record a fresh time stamp each time it is called?  
> Or it is measured once upon probe (or clause) entry and then reused?
> 
> The documentation isn't real specific, but it seems to imply a fresh 
> time on each call.  Meanwhile, DTv1 seems to get a fresh call for each 
> clause and then reuse that value during that clause.  In contrast, DTv2 
> seems to get a fresh value with each clause.  Is the change from DTv1 to 
> DTv2 intentional?  Should the legacy behavior be preserved?

The correct behaviour based on the reference implementation is that the value
of 'timestamp' is consistent within the scope of a clause (and its predicate).
Therefore, yes, we need to fix this.

What needs to happen is that the mstate structure needs to get a new field
to hold the timestamp value, and it should be initialized to 0 in the
prologue code generated in dt_cg_prologue().  Then, in get_bvar() we need to
check to see if mst->timestamp is 0.  If it is, we use the BPF helper to obtain
the timestamp value and we store that in mst->timestamp.  Either way, we then
return mst->timestamp as value.

That will ensure that the value is consistent for multiple timestamp references
within a clause (and its predicate) while ensuring that between clauses we get
a new value.

Want to tackle that?