On Thu, Dec 05, 2019 at 10:32:29AM +0100, Peter Zijlstra wrote:
> On Wed, Dec 04, 2019 at 09:42:47PM +0300, Dan Carpenter wrote:
>
> > > The current code has this:
> > >
> > > static int __init syscall_enter_define_fields(struct trace_event_call *call)
> > > {
> > > struct syscall_trace_enter trace;
> > > struct syscall_metadata *meta = call->data;
> > > int ret;
> > > int i;
> > > int offset = offsetof(typeof(trace), args);
> > >
> > > ret = trace_define_field(call, SYSCALL_FIELD(int, nr, __syscall_nr),
> > > FILTER_OTHER);
> >
> > In linux-next this ret = trace_define_field() assignment is removed.
> > That was commit 60fdad00827c ("ftrace: Rework event_create_dir()").
>
> Yep, mea culpa.
>
> > > if (ret)
> > > return ret;
> > >
> > > for (i = 0; i < meta->nb_args; i++) {
> > > ret = trace_define_field(call, meta->types[i],
> > > meta->args[i], offset,
> > > sizeof(unsigned long), 0,
> > > FILTER_OTHER);
> > > offset += sizeof(unsigned long);
> > > }
> > >
> > > return ret;
> > > }
> > >
> > >
> > > How can ret possibly be uninitialized?
> >
> > I should have written this commit more carefully and verified whether
> > meta->nb_args can actually be zero instead of just assuming it was a
> > false positive...
>
> Right, I'm thinking this is in fact possible. We have syscalls without
> arguments (sys_sched_yield for exmaple).
Well, it would have triggered a run time bug because of that thing with
GCC where it sometimes initializes variables to zero.
Let me resend properly with a Fixes tag.
regards,
dan carpenter
