Re: [PATCH] Fix tracing infrastructure to support multiple includes when defining CREATE_TRACE_POINTS

[Neil Horman <nhorman@tuxdriver.com>]

On Tue, Dec 15, 2009 at 11:13:49PM -0500, Steven Rostedt wrote:
> On Tue, 2009-12-15 at 11:08 -0500, Neil Horman wrote:
> 
> Hi Neil,
> 
> I was playing with this, and I got really nasty errors in the trace
> parsing tools. Then I noticed why:
> 
> > -DECLARE_TRACE(napi_poll,
> > +TRACE_EVENT(napi_poll,
> > +
> >  	TP_PROTO(struct napi_struct *napi),
> > -	TP_ARGS(napi));
> > +
> > +	TP_ARGS(napi),
> > +
> > +	TP_STRUCT__entry(
> > +		__field(        struct napi_struct *,	napi)
> > +        ),
> > +
> > +        TP_fast_assign(
> > +                __entry->napi = napi;
> > +        ),
> > +
> > +        TP_printk("napi poll on napi struct %p for device %s",
> > +		__entry->napi, __entry->napi->dev->name)
> 
> You can't trust this! That "__entry" happens to reside on the ring
> buffer. If for some reason the device goes away, this blows up when you
> read the trace.
> 
> If you need to save the name of the device, then store it in the ring
> buffer. You can do it with a dynamic array:
> 
> 	TP_STRUCT__entry(
> 		__field(	struct napi_struct *, napi)
> 		__string(	dev_name, napi->dev->name)
> 	),
> 
> 	TP_fast_assign(
> 		__entry->napi = napi;
> 		__assign_str(dev_name, napi->dev->name);
> 	),
> 
> 	TP_printk("napi poll on napi struct %p for device %s",
> 		__entry->napi, __get_string(dev_name))
> 
> -- Steve
> 
> > +);
> >  
> >  #endif
> 
> 
> 
Ok, thanks, I'll resubmit shortly.
Neil