Re: [PATCH 15/18] tracing: Add string type for dynamic strings in function based events

[Steven Rostedt <rostedt@goodmis.org>]

On Fri, 9 Feb 2018 12:15:47 +0900
Namhyung Kim <namhyung@kernel.org> wrote:

> > @@ -124,6 +128,16 @@ enum {
> >  	FUNC_TYPE_MAX
> >  };
> >  
> > +#define MAX_STR		512
> > +
> > +/* Two contexts, normal and NMI, hence the " * 2" */
> > +struct func_string {
> > +	char		buf[MAX_STR * 2];
> > +};
> > +
> > +static struct func_string __percpu *str_buffer;
> > +static int nr_strings;  
> 
> What protects it?

Grumble, I was thinking that the entire create_function_event was under
the func_event_mutex, which it is not. So nr_strings is not fully
protected. I'll fix that thanks.

As for str_buffer, I should comment this as it is rather subtle.


+static int read_string(char *str, unsigned long addr)
+{
+	unsigned long flags;
+	struct func_string *strbuf;
+	char *ptr = (void *)addr;
+	char *buf;
+	int ret;
+
+	if (!str_buffer)
+		return 0;
+
+	strbuf = this_cpu_ptr(str_buffer);
+	buf = &strbuf->buf[0];
+
+	if (in_nmi())
+		buf += MAX_STR;
+
+	local_irq_save(flags);

Like I said, this is really subtle, and desperately needs a comment.

The str_buffer is per cpu and can only be access under irqs disabled.
If we are in NMI, then we move the starting position forward by MAX_STR.

I'll add comments and protect create_function_event with the mutex.

Thanks for pointing this out!

-- Steve


+	ret = strncpy_from_unsafe(buf, ptr, MAX_STR);
+	if (ret < 0)
+		ret = 0;
+	if (ret > 0 && str)
+		memcpy(str, buf, ret);
+	local_irq_restore(flags);
+
+	return ret;
+}