Re: [PATCH 2/4] tracing/hist: Call hist functions directly via a switch statement

[Steven Rostedt <rostedt@goodmis.org>]

On Mon, 5 Sep 2022 22:48:49 +0900
Masami Hiramatsu (Google) <mhiramat@kernel.org> wrote:

> Hi Steve,
> 
> I found this crashed the kernel if I ran 
> 
>  # ftracetest test.d/trigger/trigger-hist-expressions.tc
> 
> More specifically, here is the minimum reproduce command.
> 
>  # echo "hist:keys=common_pid:x=16-8-4" > events/sched/sched_process_fork/trigger 
> [   75.874402] BUG: kernel NULL pointer dereference, address: 0000000000000038
> [   75.875378] #PF: supervisor read access in kernel mode
> 
> I investigated it and found you missed to set correct fn_num for
> combined constant field;
> 
> > @@ -2725,7 +2748,8 @@ static struct hist_field *parse_expr(struct hist_trigger_data *hist_data,
> >  		if (var2)
> >  			expr->operands[1] = var2;
> >  
> > -		expr->constant = op_fn(expr, NULL, NULL, NULL, NULL);
> > +		expr->fn_num = op_fn;
> > +		expr->constant = hist_fn_call(expr, NULL, NULL, NULL, NULL);
> >  
> >  		expr->operands[0] = NULL;
> >  		expr->operands[1] = NULL;  
> 
> Here, we need this;
> 
> 		expr->fn_num = HIST_FIELD_FN_CONST;
> 
> Without this fix, the binary op_fn is kept on this const field and it
> causes the problem because the const field doesn't have operands.
> 

Thanks Masami,

Today is a US holiday, and I'm only cleaning out some email (and working on
my LPC presentation for tomorrow's speaker's training), but not doing any
other work. I'll look at this tomorrow.

-- Steve