Re: [Dprobes] dprobes --getvars --reset

SourceForge Headquarters 225 Broadway Suite 1600 San Diego, CA 92101 +1 (858) 454-5900

Hi,

Not taking the write_lock while resetting the variables is a bug. We will
fix that.

We have designed/implemented the getvars command such that when the
variables are being reset, we don't return their previous contents. There
really is no reason not to return their previous values, though.

It looks like one would expect the application to return the present values
and then reset them in one atomic operation. If it seems more natural we
could do it that way. It is not hard at all, there is no issue with kmalloc
either, we handle that already. We will include it in 3.4.

Regards.. Vamsi.

Vamsi Krishna S.
Linux Technology Center,
IBM Software Lab, Bangalore.
Ph: +91 80 5044959
Internet: vam...@in...

                    david oleszkiewicz                                                                                                             
                    <da...@us...>       To:     Dprobes Mailing List <dp...@ww...>                                 
                    Sent by:                  cc:                                                                                                  
                    dprobes-admin@www-1       Subject:     [Dprobes] dprobes --getvars --reset                                                     
                    24.ibm.com                                                                                                                     

                    02/27/02 03:05 AM                                                                                                              

So I expected using --getvars and --reset together would have allowed me
to get the value of a variable and reset it in one operation.  This would,
of
course, be in such a manner that I could not lose any data.  For example,
if I was using dprobes to do some sort of statistics gathering, I would
like to be able to reset any set of variables with the assurance that
I do not lose any data.  I can then be smart about grouping variables
under 'name=' clauses and giving them contiguous numbers so that i can
easily
zero and read them in one "atomic" operation.

A briefly looked at the code and it would seem like there is no great
obstacle to implementing this feature.

looking in dp_get_globalvars() we have
if (cmd & DP_GETVARS_RESET) {
                memset(dp_gv + req->from, 0, num_vars*sizeof(unsigned
long));
                read_unlock_irqrestore(&dp_gv_lock, eflags);
                goto done;
        }

I imagine you don't get a write lock since you are not concerned about
multiple user processes reseting variables; that's the user's
problem (or more specifically only root).  and the write lock obtained in
the interrupt code will ensure that no variable will be changed while any
of the variables are being reset...so resetting the group is already
atomic for us, right?

if we wanted to allow my described operation we would need to get a write
lock
when we wanted to reset the vars and read them with the same lock held.
This
does make sense, since we are writing to the variables after all...so the
code
wouldn't look too funny.
I guess the only issue is where to place the kmalloc since we don't want
to do
that with lock held.

does such operation make sense, or is there some theoretical reasoning
behind
not allowing this.

dave

--

_______________________________________________
Dprobes mailing list
Dp...@ww...
http://www-124.ibm.com/developerworks/oss/mailman/listinfo/dprobes