Re: [PATCH v10 06/14] unwind_user/deferred: Add deferred unwinding interface

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jun 18, 2025 at 11:37:06AM -0400, Steven Rostedt wrote:
> On Wed, 18 Jun 2025 16:20:00 +0200
> Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > > The timestamp is passed to the caller on request, and when the stacktrace is
> > > generated upon returning to user space, it call the requester's callback
> > > with the timestamp as well as the stacktrace.  
> > 
> > This whole story hinges on there being a high resolution time-stamp
> > available... Good thing we killed x86 !TSC support when we did. You sure
> > there's no other architectures you're interested in that lack a high res
> > time source?
> > 
> > What about two CPUs managing to request an unwind at exactly the same
> > time?
> 
> It's mapped to a task. As long as each timestamp is unique for a task it
> should be fine. As the trace can record the current->pid along with the
> timestamp to map to the unique user space stack trace.
> 
> As for resolution, as long as there can't be two system calls back to back
> within the same time stamp. Otherwise, yeah, we have an issue.

Well, up until very recent, jiffies was the fallback clock on x86. You
can get PID reuse in a jiffy if you push things hard.

Most archs seem to have some higher res clock these days, but I would
not be surprised if among the museum pieces we still support there's
some crap lying in wait.

If you want to rely on consecutive system calls never seeing the same
timestamp, let alone PID reuse in the same timestamp -- for some generic
infrastructure -- you need to go audit all the arch code.