On Thu, Jun 19, 2025 at 04:37:33AM -0400, Steven Rostedt wrote:

On Thu, 19 Jun 2025 10:34:15 +0200
Peter Zijlstra [off-list ref] wrote:

quoted

Why can't we cmpxchg_local() the thing and avoid this horrible stuff?

static u64 get_timestamp(struct unwind_task_info *info)
{
	u64 new, old = info->timestamp;

	if (old)
		return old;
	
	new = local_clock();
	old = cmpxchg_local(&info->timestamp, old, new);
	if (old)
		return old;
	return new;
}

Seems simple enough; what's wrong with it?

It's a 64 bit number where most 32 bit architectures don't have any
decent cmpxchg on 64 bit values. That's given me hell in the ring
buffer code :-p

Do we really have to support 32bit?

But IIRC a previous version of all this had a syscall counter. If you
make this a per task syscall counter, unsigned long is plenty.

I suppose that was dropped because adding that counter increment to all
syscalls blows. But if you really want to support 32bit, that might be a
fallback.

Luckily, x86 dropped support for !CMPXCHG8B right along with !TSC. So on
x86 we good with timestamps, even on 32bit.