On Tue, 16 Jul 2019 16:20:05 +0200
Peter Zijlstra <[email protected]> wrote:

> On Tue, Jul 16, 2019 at 10:08:18PM +0800, Changbin Du wrote:
> > On Mon, Jul 15, 2019 at 12:12:31PM +0200, Peter Zijlstra wrote:  
> 
> > > 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.
> > >  
> > We can extend the recordmcount tool to search 'subprogram' tag in the DIE 
> > tree.
> > In below example, the 'DW_AT_type' is the type of function pidfd_create().
> > 
> > $ readelf -w kernel/pid.o
> >  [...]
> >  <1><1b914>: Abbrev Number: 232 (DW_TAG_subprogram)
> >     <1b916>   DW_AT_name        : (indirect string, offset: 0x415e): 
> > pidfd_create
> >     <1b91a>   DW_AT_decl_file   : 1
> >     <1b91b>   DW_AT_decl_line   : 471
> >     <1b91d>   DW_AT_decl_column : 12
> >     <1b91e>   DW_AT_prototyped  : 1
> >     <1b91e>   DW_AT_type        : <0xcc>
> >     <1b922>   DW_AT_low_pc      : 0x450
> >     <1b92a>   DW_AT_high_pc     : 0x50
> >     <1b932>   DW_AT_frame_base  : 1 byte block: 9c  (DW_OP_call_frame_cfa)
> >     <1b934>   DW_AT_GNU_all_call_sites: 1
> >     <1b934>   DW_AT_sibling     : <0x1b9d9>
> >  [...]
> > 
> > To that end, we need to introduce libdw library for recordmcount. I will 
> > have a
> > try this week.  
> 
> Right; but only when this config option is set.

Sure, and we can have fgraph support of return values depend on that
option ;-)

> 
> > And probably, we can also record the parameters?  
> 
> The 'fun' part is where to store all this information in the kernel and
> how fast you can find it while tracing.

This has been on my TODO list for a long time (I'm really happy if
someone else would do it!). My thought is that this information would
need to be able to be a module and loaded (like config.gz can be). And
then you can load the info, do the tracing, and then unload it.

For the speed part, we can add a way to hook the function with the
parameters, which shouldn't be an issue, as we already do that when
filtering for function graph. There's a function hash table that fgraph
users have that is tested to see if it should trace the function or
not. And the functions themselves are recorded in a mostly binary array
that can be looked up via a binary search from the ip address.

-- Steve

Reply via email to