Re: malloc probes for Systemtap

[Jeff Law <law at redhat dot com>]

On 10/02/2012 07:17 PM, Alexandre Oliva wrote:
> I guess myself and Siddhesh have been the most recent âmaintainersâ of
> malloc.  I've studied it a lot myself while preparing this patch, and
> Siddhesh did so even more deeply for other reasons.  Jeff Law is
> somewhat familiar with malloc and he's seen this patchset and, to some
> extent, guided the requirements for the probes I introduced.
>
> Any objections to my going ahead with the probes as they are, that
> internal reviews at Red Hat have IIUC already concluded would satisfy
> the requirements of exposing unusually expensive malloc behavior to
> probing?
So there's a Red Hat customer request behind the probes which looked 
something like add probes for unusual events such as new arena 
allocation, modification of dynamic tunables, etc.

I think this customer is at a stage where they want to better detect 
when malloc does something unexpected.  I would expect that after some 
working with these probes we'll have a follow-up request for more probes 
to start answering why malloc does something unexpected.


In the process of adding the initial probes, Alex, Siddhesh and myself 
came up with several others based on our experiences in that code.  I'll 
try to summarize the thought process behind each as I go through Alex's 
patches.



> > What we have in the way of
> > documentation for the existing probes is just
> > nptl/DESIGN-systemtap-probes.txt and elf/rtld-debugger-interface.txt.
>
> So I noticed.  I wasn't happy with that and decided it belonged in the
> manual, but...  I hadn't considered ABI stability aspects of probes.
> *blush* :-)
The ABI aspect of probing is a big open question for any project adding 
static probes.

> Anyhow, it wouldn't be hard a tall to reformat the texi input into a
> .txt file, or perhaps to have the newly-introduce docs unconditionally
> removed from the formatted manuals.
>
> > (There was some past discussion about this when the earlier probes
> > went in, but it only resolved to the point that we achieved consensus
> > that we didn't really need to resolve it right then.)
I think we're still in that boat (don't need to resolve ABI issues 
around probes).  However, for these probes to be useful we need some 
kind of docs on the probes.

I'm comfortable with docs in a file or with docs that are removed from 
the formatted manuals.
jeff