Re: [PATCH] tracing/fgraph: support recording function return values

[Peter Zijlstra]

On Mon, Jul 15, 2019 at 09:29:30AM +0100, Will Deacon wrote:
> On Sat, Jul 13, 2019 at 08:10:26PM +0800, Changbin Du wrote:
> > This patch adds a new trace option 'funcgraph-retval' and is disabled by
> > default. When this option is enabled, fgraph tracer will show the return
> > value of each function. This is useful to find/analyze a original error
> > source in a call graph.
> > 
> > One limitation is that the kernel doesn't know the prototype of functions.
> > So fgraph assumes all functions have a retvalue of type int. You must ignore
> > the value of *void* function. And if the retvalue looks like an error code
> > then both hexadecimal and decimal number are displayed.
> 
> This seems like quite a significant drawback and I think it could be pretty
> confusing if you have to filter out bogus return values from the trace.
> 
> For example, in your snippet:
> 
> >  3)               |  kvm_vm_ioctl() {
> >  3)               |    mutex_lock() {
> >  3)               |      _cond_resched() {
> >  3)   0.234 us    |        rcu_all_qs(); /* ret=0x80000000 */
> >  3)   0.704 us    |      } /* ret=0x0 */
> >  3)   1.226 us    |    } /* ret=0x0 */
> >  3)   0.247 us    |    mutex_unlock(); /* ret=0xffff8880738ed040 */
> 
> mutex_unlock() is wrongly listed as returning something.
> 
> How much of this could be achieved from userspace by placing kretprobes on
> non-void functions instead?

Alternatively, we can have recordmcount (or objtool) mark all functions
with a return value when the build has DEBUG_INFO on. The dwarves know
the function signature.