Skip to contentSkip to conversation

tracing: annotate emit_log_char() notrace

PATCH In review Thomas Gleixner 3 participants 6 messages May 27, 2009 – May 28, 2009

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 tracing: annotate emit_log_char() notrace

printk calls emit_log_char() in a loop which is cluttering the trace buffer. Make it notrace.

---
 kernel/printk.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
kernel/printk.c +1 -1
Index: linux-2.6-tip/kernel/printk.c
===================================================================
@@ -499,7 +499,7 @@ static void call_console_drivers(unsigne
 	_call_console_drivers(start_print, end, msg_level);
 }
 
-static void emit_log_char(char c)
+static void notrace emit_log_char(char c)
 {
 	LOG_BUF(log_end) = c;
 	log_end++;
#1 lore raw
Quoted 5 lines · Thomas Gleixner

printk calls emit_log_char() in a loop which is cluttering the trace buffer. Make it notrace.

---

Does it produce that much annoying traces?
I mean, printk() are usually rare events.

Do you have a particular debugging workflow that makes this function invasive in the trace?

May be it's because you are debugging using ftrace and other debugging options that use a lot of printk()
In such case, it would indeed be good to apply this patch.

Thanks,
Frederic.

Quoted 16 lines
 kernel/printk.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Index: linux-2.6-tip/kernel/printk.c
===================================================================

kernel/printk.c +1 -1
@@ -499,7 +499,7 @@ static void call_console_drivers(unsigne
 	_call_console_drivers(start_print, end, msg_level);
 }
 
-static void emit_log_char(char c)
+static void notrace emit_log_char(char c)
 {
 	LOG_BUF(log_end) = c;
 	log_end++;
#2 lore raw
Quoted 11 lines · Frederic Weisbecker
Quoted 5 lines · Thomas Gleixner

printk calls emit_log_char() in a loop which is cluttering the trace buffer. Make it notrace.

---

Does it produce that much annoying traces?
I mean, printk() are usually rare events.

Quoted 6 lines

Do you have a particular debugging workflow that makes this function invasive in the trace?

May be it's because you are debugging using ftrace and other debugging options that use a lot of printk()
In such case, it would indeed be good to apply this patch.

I noticed when I chased a boot up problem, half of the trace was full of

0.540550: emit_log_char <-vprintk
0.540550: emit_log_char <-vprintk
...

though there is another annoyance in the same code path:

0.540557: vgacon_set_cursor_size <-vgacon_cursor 0.540558: __rcu_read_lock <-__atomic_notifier_call_chain 0.540558: __rcu_read_unlock <-__atomic_notifier_call_chain 0.540558: __rcu_read_lock <-__atomic_notifier_call_chain
...

It's actually the same number of times as we have emit_log_char() there. __atomic_notifier_call_chain() itself is marked notrace but the rcu calls inside are not. :(

IIRC the same happens with a serial console, but that can be worse as you busy wait during the character output.

It'd be nice if we could inherit the notrace from the caller :)

Thanks,

tglx

#3 lore raw
Quoted 34 lines · Thomas Gleixner
Quoted 11 lines · Frederic Weisbecker
Quoted 5 lines · Thomas Gleixner

printk calls emit_log_char() in a loop which is cluttering the trace buffer. Make it notrace.

---

Does it produce that much annoying traces?
I mean, printk() are usually rare events.

Quoted 6 lines

Do you have a particular debugging workflow that makes this function invasive in the trace?

May be it's because you are debugging using ftrace and other debugging options that use a lot of printk()
In such case, it would indeed be good to apply this patch.

I noticed when I chased a boot up problem, half of the trace was full of

   0.540550: emit_log_char <-vprintk
   0.540550: emit_log_char <-vprintk
   ...

though there is another annoyance in the same code path:

   0.540557: vgacon_set_cursor_size <-vgacon_cursor
   0.540558: __rcu_read_lock <-__atomic_notifier_call_chain
   0.540558: __rcu_read_unlock <-__atomic_notifier_call_chain
   0.540558: __rcu_read_lock <-__atomic_notifier_call_chain
   ...
Quoted 9 lines

It's actually the same number of times as we have emit_log_char() there. __atomic_notifier_call_chain() itself is marked notrace but the rcu calls inside are not. :(

IIRC the same happens with a serial console, but that can be worse as you busy wait during the character output.

It'd be nice if we could inherit the notrace from the caller :)

I'd hate to make rcu_read_*lock be a notrace. I've done some debugging with wanting to know where the rcu locks were taken.

I'm sure there's lots of functions that people will want to add notrace for. Really, the notrace on a function should only be there to prevent crashes (functions used by the tracer or early boot up code for some archs).

That said, there are times when you want to avoid tracing things on early boot up. Thus, if someone adds to the kernel command line "ftrace=function" you have no way to filter out these functions you don't care about.

What if I add a "ftrace_notrace=<function list>" kernel command line, that has <function list> be a comma separated list of functions you want to go into the "set_ftrace_notrace" on boot up? Would that satisfy you. Then you can keep adding to the list dynamically every boot. And keep it there in a grub file.

This will also keep people from constantly asking for notrace annotations just because they are annoyed by the functions that always fill up their logs.

-- Steve

#4 lore raw
Quoted 8 lines · Steven Rostedt
Quoted 5 lines

IIRC the same happens with a serial console, but that can be worse as you busy wait during the character output.

It'd be nice if we could inherit the notrace from the caller :)

I'd hate to make rcu_read_*lock be a notrace. I've done some debugging with wanting to know where the rcu locks were taken.

Sure.

Quoted 4 lines

I'm sure there's lots of functions that people will want to add notrace for. Really, the notrace on a function should only be there to prevent crashes (functions used by the tracer or early boot up code for some archs).

Well, lets look at the atomic notifier then. Why is it notrace ? And if there is a reason why does it trace the calls which are done inside of it. We should really think about inheriting the notrace down the call chain at least for those calls which must be annotated to avoid wreckage. That could then be made optional (via commandline) for all notrace calls.

Quoted 10 lines

That said, there are times when you want to avoid tracing things on early boot up. Thus, if someone adds to the kernel command line "ftrace=function" you have no way to filter out these functions you don't care about.

What if I add a "ftrace_notrace=<function list>" kernel command line, that has <function list> be a comma separated list of functions you want to go into the "set_ftrace_notrace" on boot up? Would that satisfy you. Then you can keep adding to the list dynamically every boot. And keep it there in a grub file.

That list might grow fast beyond the command line lenght limit :)

But yeah, I can see your point. At least it would be better than cursing and adding ad hoc notrace patches all over the place.

Thanks,

tglx

#5 lore raw
Quoted 17 lines · Thomas Gleixner
Quoted 8 lines · Steven Rostedt
Quoted 5 lines

IIRC the same happens with a serial console, but that can be worse as you busy wait during the character output.

It'd be nice if we could inherit the notrace from the caller :)

I'd hate to make rcu_read_*lock be a notrace. I've done some debugging with wanting to know where the rcu locks were taken.

Sure.

Quoted 4 lines

I'm sure there's lots of functions that people will want to add notrace for. Really, the notrace on a function should only be there to prevent crashes (functions used by the tracer or early boot up code for some archs).

Quoted 6 lines

Well, lets look at the atomic notifier then. Why is it notrace ? And if there is a reason why does it trace the calls which are done inside of it. We should really think about inheriting the notrace down the call chain at least for those calls which must be annotated to avoid wreckage. That could then be made optional (via commandline) for all notrace calls.

Every function marked with __kprobes becomes notrace, which atomic notifier is.

We probably can even remove that now. Ingo added it when we had the issue with the NMI causing lockups. See 36dcd67ae994fece615b7c700958d215e884b9ae But we have since fixed the NMI issues.

Also, there's no real way to inherit notrace automatically. Well we could write some really nasty scripts to try it at compile time. notrace is simply an attribute in gcc to prevent it from calling mcount.

Quoted 12 lines
Quoted 11 lines

That said, there are times when you want to avoid tracing things on early boot up. Thus, if someone adds to the kernel command line "ftrace=function" you have no way to filter out these functions you don't care about.

What if I add a "ftrace_notrace=<function list>" kernel command line, that has <function list> be a comma separated list of functions you want to go into the "set_ftrace_notrace" on boot up? Would that satisfy you. Then you can keep adding to the list dynamically every boot. And keep it there in a grub file.

Quoted 5 lines

That list might grow fast beyond the command line lenght limit :)

But yeah, I can see your point. At least it would be better than cursing and adding ad hoc notrace patches all over the place.

Well, the boot options still allow for simple wild cards. So hopefully the list will not grow too much. You can also just add a list of functions that you want to trace (filter as suppose to notrace).

If you are tracing after boot up, then you can add the list to a start up script and just echo into the set_ftrace_notrace file.

-- Steve