On Thu, Mar 30, 2017 at 8:53 AM, Ingo Molnar [off-list ref] wrote:

* Masami Hiramatsu [off-list ref] wrote:

quoted

quoted

So this is something I missed while the original code was merged, but the concept
looks a bit weird: why do we do any "allocation" while a handler is executing?

That's fundamentally fragile. What's the maximum number of parallel
'kretprobe_instance' required per kretprobe - one per CPU?

It depends on the place where we put the probe. If the probed function will be
blocked (yield to other tasks), then we need a same number of threads on
the system which can invoke the function. So, ultimately, it is same
as function_graph tracer, we need it for each thread.

So then put it into task_struct (assuming there's no kretprobe-inside-kretprobe
nesting allowed). There's just no way in hell we should be calling any complex
kernel function from kernel probes!

Some kprobes are called from an interruption context. We have a kprobe
on tcp_set_state() and this is sometimes called when the network card
receives a packet.

I mean, think about it, a kretprobe can be installed in a lot of places, and now
we want to call get_free_pages() from it?? This would add a massive amount of
fragility.

Instrumentation must be _simple_, every patch that adds more complexity to the
most fundamental code path of it should raise a red flag ...

So let's make this more robust, ok?

Thanks,

        Ingo

Thanks,
Alban