On Tue, 2012-03-27 at 11:09 -0700, Arun Sharma wrote:

On 3/24/12 7:14 PM, Frederic Weisbecker wrote:

quoted

quoted

The other problem in branch stacks/LBR is that they're
sampled branches. Just because I got a sample with:

a ->  b
b ->  c

doesn't necessarily mean that the callchain was a ->  b ->  c.

Not sure what you mean. If you have a ->  b, b ->  c in single
LBR sample it means you got a ->  b ->  c.

I was going by Stephane's commit message here:

http://article.gmane.org/gmane.linux.kernel/1236999

 > Statistical sampling of taken branch should not be confused
 > for branch tracing. Not all branches are necessarily captured

Stephane, could you please explain if the 16 filtered branches in LBR 
are guaranteed to be from a given callchain to the leaf function? My 
understanding is that it's not.

callchain1: a -> b -> d -> e (sample a->b)
callchain2: a -> c -> b -> f (sample b->f)

on PMU interrupt can we end up with:

b -> f <- top of stack
a -> b
...

even though a -> b -> f can never happen in the actual program flow?

Right, so the LBR is a queue not a stack. A program like:

foo() {
	bar1();
	bar2();
}

will, using the lbr, look like: foo->bar1->bar2 (if you filter returns),
or foo->bar1->foo+x->bar2 if you include returns.

A callchain is a pure stack, a return pops the top most entry, the above
program can only give 3 possible callchains:

a) foo
b) foo, bar1
c) foo, bar2

Furthermore, the LBR is about any branch, callchains are about function
calls.