Re: [GIT pull] perf/urgent for 5.7-rc2

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Apr 22, 2020 at 09:45:12AM +0200, Ingo Molnar wrote:
> 
> * Peter Zijlstra <peterz@infradead.org> wrote:
> 
> > On Mon, Apr 20, 2020 at 09:48:45AM +0200, Ingo Molnar wrote:
> > > Fortunately, much of what objtool does against vmlinux.o can be 
> > > parallelized in a rather straightforward fashion I believe, if we build 
> > > with -ffunction-sections.
> > 
> > So that FGKASLR is going to get us -ffunction-sections, but
> > parallelizing objtool isn't going to be trivial, it's data structures
> > aren't really build for that, esp. decode_instructions() which actively
> > generates data.
> > 
> > Still, it's probably doable.
> 
> So AFAICS in the narrow code section I identified as the main overhead, 
> only the instruction hash needs threading, i.e. this step:
> 
>                         hash_add(file->insn_hash, &insn->hash, insn->offset);
>                         list_add_tail(&insn->list, &file->insn_list);
> 
> Objtool can still be single-threaded before and after generating the 
> instruction hash.
> 
> 99% of the overhead within decode_instructions() is in 
> arch_decode_instruction(), which is fully thread-safe AFAICS.

Correct; I suppose you can farm out the sections to N threads for
arch_decode_instruction() and then have the main thread collect decoded
sections and frob them in the global data structures.

Another pass you can probably parallize fairly easily is
validate_functions() / validate_unwind_hints(). While that modifies
state, the state it modifies should be local to the section at hand.

That needs an audit of course, but it should be entirely doable.