On 2025-07-02 15:40, Steven Rostedt wrote:
On Wed, 2 Jul 2025 15:36:00 -0400 Steven Rostedt <[email protected]> wrote:union unwind_task_id { struct { u32 task_id; u32 cnt; } u64 id; }; static u64 get_cookie(struct unwind_task_info *info) { u32 cnt = READ_ONCE(info->id.cnt); u32 new_cnt; if (cnt & 1) return info->id; if (unlikely(!info->id.task_id)) { u32 task_id = local_clock(); cnt = 0; if (try_cmpxchg(&info->id.task_id, &cnt, task_id)) task_id = cnt; } new_cnt = cnt + 3; if (try_cmpxchg(&info->id, &cnt, new_cnt)) new_cnt = cnt; // try_cmpxchg() expects something return info->id; }Honestly I think this is way overkill. What I would do, is to have the cookie saved in the event be 64 bit, but we can start with the simple 32 bit solution keeping the top 32 bits zeros. If this does indeed become an issue in the future, we could fix it with a 64 bit number. By making sure all the exposed "cookies" are 64 bit, it should not break anything. The cookie is just supposed to be a random unique number that associates a request with its deferred user space stack trace. With any exposed cookies to user space being 64 bits, this should not be an issue to address in the future.
FWIW, I liked your idea of making the cookie 64-bit with: - 32-bit cpu number, - 32-bit per-CPU free running counter. This is simple, works on 32-bit archs, and it does not overflow as often as time LSB because it counts execution contexts. Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. https://www.efficios.com
