Re: [RFC][PATCH 03/14 v2] function_graph: Allow multiple users to attach to function graph

From: Peter Zijlstra
Date: Fri May 24 2019 - 08:30:53 EST


On Fri, May 24, 2019 at 08:12:19AM -0400, Steven Rostedt wrote:
> On Fri, 24 May 2019 13:26:08 +0200
> Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> > But but but but.. why not add all the required bits to the shadow stack
> > in the first place and do away with the array entirely?
>
> What required bits would that be? The pointer to the fgraph_ops,
> because we need that to pass to the calling function.

I was thinking a smaller structure comprising of {func,callback}, which
you pop, if func matches, run callback.

> > So on ret, just keep POP'ing until either the stack is empty or the
> > entry is for another function.
>
> When we hit a fgraph_ops, how do we know if it was freed or not? We
> can't just blindly reference it.
>
> The idea of the array, is that we can maintain state in a single
> location of when the fgraph_ops is freed. If we return from a function,
> we have an index and a counter, and if the counter doesn't match with
> what's in the array, then we know that the fgraph_ops is no longer
> around and we just drop it.
>
> The reason for the array, is to keep track of if the fgraph_ops has
> been freed or not. Otherwise, when we unregister the fgraph_ops, we
> would need to search all shadow stacks, looking for it to unreference
> it.
>
> Believe me, I rather not have that array, but I couldn't come up with a
> better solution to handle freeing of fgraph_ops.

The trivial answer would be to refcount the thing, but can't we make
rcu_tasks do this?

And delay the unreg until all active users are gone -- who gives a crap
that can take a while.