Question about enabling trave_events on module load

Question about enabling trave_events on module load

[Cezary Rojewski]

Hello Steven,

I bet that is not the first time said question is asked - that's for 
sure - but I failed to find a method for solving the issue, that is: not 
missing a single trace from the moment given module is loaded. Maybe I'm 
missing something or documentation wasn't clear enough and that's why 
I'm here.

If I am, please point to towards the right direction. Then you can slap 
me for not reading the documentation carefully.

"trace_event=" cmdline option seems to target built-in tracepoints 
_only_ so ain't much of a help to me. After digging the past for some 
time, I've found a very promising thread:
	tracing: Enable tracepoints via module parameters
	https://lore.kernel.org/patchwork/patch/240185/

Sadly, I wasn't able to find _that_ solution (or anything similar for 
that matter) implemented into the kernel.

So far, the only option I came with was separating traces into a 
built-in piece that declares all events upfront so "trace_event=" option 
has something to hook into. Said piece is of course made of a standard 
trace header file filled with macro usage and a .c file with handful of 
EXPORT_TRACEPOINT_SYMBOL(s).

While that solution could suffice, localization is the problem here - if 
a tree my module is built in is configured via -m, the built-in piece 
won't expose symbols at all and 'make' will leave me with bunch of 
"ERROR: <symbol> undefined" for every module my traces were used in. To 
fix the problem, I've relocated my trace .c file to /kernel/trace/. 
Finally it compiles and works as intended..

Not really satisfying, though. While there are some examples of 
subsystems keeping their trace .c in /kernel/trace (e.g.: 
power-traces.c), I don't believe that place is open for _every single 
driver_ to dump their trace sources into.

If indeed traces cannot be enabled on module load, then this is a gap.
While not everyone looked satisfied in the 9year old thread, I believe 
having the gap closed is important - and userspace can always be 
improved upon as time passes.


Thank you in advance for your input and time.

Czarek

Re: Question about enabling trave_events on module load

[Steven Rostedt]

On Fri, 28 Feb 2020 20:16:54 +0100
Cezary Rojewski <cezary.rojewski@intel.com> wrote:

> Hello Steven,
> 
> I bet that is not the first time said question is asked - that's for 
> sure - but I failed to find a method for solving the issue, that is: not 
> missing a single trace from the moment given module is loaded. Maybe I'm 
> missing something or documentation wasn't clear enough and that's why 
> I'm here.
> 
> If I am, please point to towards the right direction. Then you can slap 
> me for not reading the documentation carefully.
> 
> "trace_event=" cmdline option seems to target built-in tracepoints 
> _only_ so ain't much of a help to me. After digging the past for some 
> time, I've found a very promising thread:
> 	tracing: Enable tracepoints via module parameters
> 	https://lore.kernel.org/patchwork/patch/240185/

I find this email extremely amusing. Just the day before (Thursday), I
was trying to clean out my INBOX (it had emails back from 2008!) and I
came across this very thread, and said to myself. "Hmm, I wonder if I
should push this again?". What a coincidence that the next day, someone
would email me about that thread after 9 years!

> 
> Sadly, I wasn't able to find _that_ solution (or anything similar for 
> that matter) implemented into the kernel.

That's because it was dropped and forgotten about :-(  I left it in my
INBOX to remind myself to bring it back, but that didn't work out as
well as I expected, as my INBOX turned into more of a graveyard than a
TODO list.

> 
> So far, the only option I came with was separating traces into a 
> built-in piece that declares all events upfront so "trace_event=" option 
> has something to hook into. Said piece is of course made of a standard 
> trace header file filled with macro usage and a .c file with handful of 
> EXPORT_TRACEPOINT_SYMBOL(s).
> 
> While that solution could suffice, localization is the problem here - if 
> a tree my module is built in is configured via -m, the built-in piece 

I'm unfamiliar with "-m", what does that do?

> won't expose symbols at all and 'make' will leave me with bunch of 
> "ERROR: <symbol> undefined" for every module my traces were used in. To 
> fix the problem, I've relocated my trace .c file to /kernel/trace/. 
> Finally it compiles and works as intended..

Well, obviously (as you state below), that's not an answer.

> 
> Not really satisfying, though. While there are some examples of 
> subsystems keeping their trace .c in /kernel/trace (e.g.: 
> power-traces.c), I don't believe that place is open for _every single 
> driver_ to dump their trace sources into.
> 
> If indeed traces cannot be enabled on module load, then this is a gap.
> While not everyone looked satisfied in the 9year old thread, I believe 
> having the gap closed is important - and userspace can always be 
> improved upon as time passes.
> 
> 
> Thank you in advance for your input and time.

I think your email confirmed to me that I need to push this thread
again.

Thanks for reaching out!

-- Steve