On ons, mar 06, 2024 at 16:46, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: > On Wed, 06 Mar 2024 21:02:06 +0100 > Tobias Waldekranz <tobias@xxxxxxxxxxxxxx> wrote: > >> On ons, mar 06, 2024 at 10:15, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: >> > On Mon, 04 Mar 2024 23:40:49 +0100 >> > Tobias Waldekranz <tobias@xxxxxxxxxxxxxx> wrote: >> > >> >> On ons, feb 28, 2024 at 09:56, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote: >> >> > On Wed, 28 Feb 2024 11:47:24 +0100 >> >> > Tobias Waldekranz <tobias@xxxxxxxxxxxxxx> wrote: >> >> > >> >> > The "trace_seq p" is a pointer to trace_seq descriptor that can build >> >> > strings, and then you can use it to print a custom string in the trace >> >> > output. >> >> >> >> Yes I managed to decode the hidden variable :) I also found >> >> trace_seq_acquire() (and its macro alter ego __get_buf()), which would >> >> let me keep the generic stringer functions. So far, so good. >> >> >> >> I think the foundational problem remains though: TP_printk() is not >> >> executed until a user reads from the trace_pipe; at which point the >> >> object referenced by __entry->info may already be dead and >> >> buried. Right? >> > >> > Correct. You would need to load all the information into the event data >> > itself, at the time of the event is triggered, that is needed to determine >> > how to display it. >> >> Given that that is quite gnarly to do for the events I'm trying to >> trace, because of the complex object graph, would it be acceptable to >> format the message in the assign phase and store it as dynamic data? >> I.e., what (I think) you suggested at the end of your first response. > > It's really up to what you want to do ;-) Alright. I'll interpret that as "there's a >0% chance that I'll give you an Acked-by on something like that" :) >> >> My thinking is: >> >> - Managing a duplicate (flattened) object graph, exclusively for use by >> these tracepoints, increases the effort to keep the tracing in sync >> with new additions to switchdev; which I think will result in >> developers simply avoiding it altogether. In other words: I'd rather >> have somewhat inefficient but simple flashlight, rather than a very >> efficient one that no one knows how to change the batteries in. >> >> - This is typically not a very hot path. Most events are triggered by >> user configuration. Otherwise when new neighbors are discovered. >> >> - __entry->info is still there for use by raw tracepoint consumers from >> userspace. > > How big is this info? The common struct (switchdev_notifier_info) is 24B at the moment. Depending on __entry->val, the size of the enclosing notification (e.g. switchdev_notifier_port_obj_info) is between 40-64B. This pattern may then repeat again inside the concrete notifier, where you have a pointer to a common object (e.g. switchdev_obj, 48B) whose outer size (e.g. switchdev_obj_port_vlan, 56B) is determined by an accompanying enum.
