[PATCH] export softirq_to_name symbol

[PATCH] export softirq_to_name symbol

[Jason Baron]

hi,

In response to a user request:


Export 'softirq_to_name' so that modules using the softirq tracepoints
can resolve the # to a name.

Signed-off-by: Jason Baron <jbaron@redhat.com>
---
 kernel/softirq.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/kernel/softirq.c b/kernel/softirq.c
index b76064b..1841ab7 100644
--- a/kernel/softirq.c
+++ b/kernel/softirq.c
@@ -60,6 +60,7 @@ char *softirq_to_name[NR_SOFTIRQS] = {
 	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK",
 	"TASKLET", "SCHED", "HRTIMER",	"RCU"
 };
+EXPORT_SYMBOL_GPL(softirq_to_name);
 
 /*
  * we cannot loop indefinitely here to avoid userspace starvation,
-- 
1.6.0.6

Re: [PATCH] export softirq_to_name symbol

[Christoph Hellwig]

On Fri, May 01, 2009 at 03:30:22PM -0400, Jason Baron wrote:
> 
> hi,
> 
> In response to a user request:
> 
> 
> Export 'softirq_to_name' so that modules using the softirq tracepoints
> can resolve the # to a name.

Please submit it together with the module that uses it.

Re: [PATCH] export softirq_to_name symbol

[Josh Stone]

On 05/01/2009 12:50 PM, Christoph Hellwig wrote:
> On Fri, May 01, 2009 at 03:30:22PM -0400, Jason Baron wrote:
>> hi,
>>
>> In response to a user request:
>>
>>
>> Export 'softirq_to_name' so that modules using the softirq tracepoints
>> can resolve the # to a name.
> 
> Please submit it together with the module that uses it.

I'm adding SystemTap support for the format strings in the TRACE_EVENT
macros.  The softirq's TP_fast_assign uses softirq_to_name, and so that
array is needed to prepare the softirq's trace string.

So, there's no explicit reference to the symbol in SystemTap -- it's
just indirectly referenced by the tracepoint declaration.

Thanks,

Josh

Re: [PATCH] export softirq_to_name symbol

[Christoph Hellwig]

On Fri, May 01, 2009 at 01:07:34PM -0700, Josh Stone wrote:
> I'm adding SystemTap support for the format strings in the TRACE_EVENT
> macros.  The softirq's TP_fast_assign uses softirq_to_name, and so that
> array is needed to prepare the softirq's trace string.
> 
> So, there's no explicit reference to the symbol in SystemTap -- it's
> just indirectly referenced by the tracepoint declaration.

TP_fast_assign should only be called by core code, it's the piece that
copies the trace into the ring buffer.  If systemtap copies events into
the ring buffer from modular code something is deeply wrong in it's
design.

Re: [PATCH] export softirq_to_name symbol

[Josh Stone]

On 05/01/2009 01:10 PM, Christoph Hellwig wrote:
> On Fri, May 01, 2009 at 01:07:34PM -0700, Josh Stone wrote:
>> I'm adding SystemTap support for the format strings in the TRACE_EVENT
>> macros.  The softirq's TP_fast_assign uses softirq_to_name, and so that
>> array is needed to prepare the softirq's trace string.
>>
>> So, there's no explicit reference to the symbol in SystemTap -- it's
>> just indirectly referenced by the tracepoint declaration.
> 
> TP_fast_assign should only be called by core code, it's the piece that
> copies the trace into the ring buffer.  If systemtap copies events into
> the ring buffer from modular code something is deeply wrong in it's
> design.

More precisely, it's the piece of code that copies the trace into an
__entry pointer.  In my case, __entry has nothing to do with the ring
buffer.

Josh

Re: [PATCH] export softirq_to_name symbol

[Christoph Hellwig]

On Fri, May 01, 2009 at 01:17:59PM -0700, Josh Stone wrote:
> More precisely, it's the piece of code that copies the trace into an
> __entry pointer.  In my case, __entry has nothing to do with the ring
> buffer.

Well, it should be the ring buffer.  If you want to do something more
fancy rewrite it in proper C code and submit a patch for discussion to
lkml.

Re: [PATCH] export softirq_to_name symbol

[Josh Stone]

On 05/01/2009 01:22 PM, Christoph Hellwig wrote:
> On Fri, May 01, 2009 at 01:17:59PM -0700, Josh Stone wrote:
>> More precisely, it's the piece of code that copies the trace into an
>> __entry pointer.  In my case, __entry has nothing to do with the ring
>> buffer.
> 
> Well, it should be the ring buffer.  If you want to do something more
> fancy rewrite it in proper C code and submit a patch for discussion to
> lkml.

It's not that fancy -- this is basically it:

#define __field(type, item)         type item;
#define __array(type, item, len)    type item[len];
#define __string(item, src)         const char * __str_##item;
#define __assign_str(item, src)     __entry->__str_##item = src
#define __get_str(item)             __entry->__str_##item
#define TRACE_EVENT(name, proto, args, tstruct, assign, print)  \
  static inline int                                             \
  trace_snprintf_##name(char * __buf, size_t __size, proto) {   \
    struct { tstruct } __value, *__entry = &__value;            \
    assign;                                                     \
    return snprintf(__buf, __size, print);                      \
  }

I don't know if anyone would need this besides stap, but I can post it
as a patch if there's interest.

Josh

Re: [PATCH] export softirq_to_name symbol

[Ingo Molnar]

* Jason Baron <jbaron@redhat.com> wrote:

> hi,
> 
> In response to a user request:
> 
> Export 'softirq_to_name' so that modules using the softirq 
> tracepoints can resolve the # to a name.
> 
> Signed-off-by: Jason Baron <jbaron@redhat.com>
> ---
>  kernel/softirq.c |    1 +
>  1 files changed, 1 insertions(+), 0 deletions(-)
> 
> diff --git a/kernel/softirq.c b/kernel/softirq.c
> index b76064b..1841ab7 100644
> --- a/kernel/softirq.c
> +++ b/kernel/softirq.c
> @@ -60,6 +60,7 @@ char *softirq_to_name[NR_SOFTIRQS] = {
>  	"HI", "TIMER", "NET_TX", "NET_RX", "BLOCK",
>  	"TASKLET", "SCHED", "HRTIMER",	"RCU"
>  };
> +EXPORT_SYMBOL_GPL(softirq_to_name);

Hm, that very much looks like as if someone was writing kernel 
tracing code without the intention to submitting it upstream. Not 
having exported that symbol is thus a subtle "dont do that please" 
warning sign ;-)

	Ingo

Re: [PATCH] export softirq_to_name symbol

[Frederic Weisbecker]

On Fri, May 01, 2009 at 04:10:28PM -0400, Christoph Hellwig wrote:
> On Fri, May 01, 2009 at 01:07:34PM -0700, Josh Stone wrote:
> > I'm adding SystemTap support for the format strings in the TRACE_EVENT
> > macros.  The softirq's TP_fast_assign uses softirq_to_name, and so that
> > array is needed to prepare the softirq's trace string.
> > 
> > So, there's no explicit reference to the symbol in SystemTap -- it's
> > just indirectly referenced by the tracepoint declaration.
> 
> TP_fast_assign should only be called by core code, it's the piece that
> copies the trace into the ring buffer.  If systemtap copies events into
> the ring buffer from modular code something is deeply wrong in it's
> design.


In itself, the use of TRACE_EVENT from a module is fine, otherwise
Steve wouldn't have written the module support.

For example the mac80211 subsystem can be built as a module, and
if a tracing code is merged for this subsystem, we want it to be
usable whenever it is build as a module or not.

Frederic.