Re: [RFC PATCH v5 1/3] trace: add lock-free stackmap for stack trace deduplication

[Pengfei Li]

On Wed, 2 Sep 2026 06:57:44 +0000 sashiko-bot@xxxxxxxxxx wrote:
> Is there a missing check here? It looks like ftrace_stackmap_reset() omits
> the tracing state verification entirely

> Should tracing_reset_all_cpus() be called before this memset?

Both observations are correct about the mismatch, but the code is the
side that is right. The stale part is the commit message.

The intended semantics, following Steven's feedback on the v4 thread
(https://lore.kernel.org/all/20260821235129.078dd489@fedora/), is
map-only reset: reset may run while tracing is active, and it does not
clear the ring buffer. So neither the tracer_tracing_is_on() check nor
the tracing_reset_all_cpus() call belongs in ftrace_stackmap_reset()
anymore. The v5 commit message still described the older, stricter
design that had already been dropped from the code.

On the resolution question: a trace can indeed still contain
<stack_id N> records after a reset. Such an id either no longer
resolves, or resolves to an unrelated stack once the slot is reused.
That is misleading userspace output rather than kernel corruption -
reset frees nothing and only clears storage the map still owns. The
guidance is to read the trace out before resetting if existing ids
must stay meaningful.

Fixed in v6: the commit message now describes map-only reset, and both
the kernel-doc and Documentation/trace/ftrace-stackmap.rst spell out
the id-reuse consequence.

Pengfei