Re: [PATCH v3 11/19] unwind: Add deferred user space unwinding API

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Oct 30, 2024 at 02:17:22AM -0400, Steven Rostedt wrote:
> On Tue, 29 Oct 2024 19:20:32 +0100
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > The 48:16 bit split gives you uniqueness for around 78 hours at 1GHz.
> 
> Are you saying that there will be one system call per nanosecond? This
> number is incremented only when a task enters the kernel from user
> spaces *and* requests a stack trace. If that happens 1000 times a
> second, that would still be around 9000 years.

We used to be able to do well over a million syscalls a second. I'm not
exactly sure where we are now, the whole speculation shit-show hurt
things quite badly.

> > 
> > But seriously, perf doesn't need this. It really only needs a sequence
> > number if you care to stitch over a LOST packet (and I can't say I care
> > about that case much) -- and doing that right doesn't really take much
> > at all.
> 
> Perf may not care because it has a unique descriptor per task, right?
> Where it can already know what events are associated to a task. But
> that's just a unique characteristic of perf. The unwinder should give a
> identifier for every user space stack trace that it will produce and
> pass that back to the tracer when it requests a stack trace but it
> cannot yet be performed. This identifier is what we are calling a
> context cookie. Then when it wants the stack trace, the unwinder will
> give the tracer the stack trace along with the identifier
> (context-cookie) that this stack trace was for in the past.

You're designing things inside out again. You should add functionality
by adding layers.

Pass a void * into the 'request-unwind' and have the 'do-unwind'
callback get that same pointer. Then anybody that needs identifiers to
figure out where things came from can stuff something in there.