Re: [RFC bpf-next 00/16] bpf: Speed up trampoline attach

[Jiri Olsa <jolsa@redhat.com>]

On Wed, Oct 28, 2020 at 02:13:25PM -0700, Alexei Starovoitov wrote:
> On Tue, Oct 27, 2020 at 03:28:03PM +0100, Jiri Olsa wrote:
> > On Mon, Oct 26, 2020 at 09:30:14PM -0700, Alexei Starovoitov wrote:
> > > On Thu, Oct 22, 2020 at 10:42:05AM -0400, Steven Rostedt wrote:
> > > > On Thu, 22 Oct 2020 16:11:54 +0200
> > > > Jiri Olsa <jolsa@redhat.com> wrote:
> > > > 
> > > > > I understand direct calls as a way that bpf trampolines and ftrace can
> > > > > co-exist together - ebpf trampolines need that functionality of accessing
> > > > > parameters of a function as if it was called directly and at the same
> > > > > point we need to be able attach to any function and to as many functions
> > > > > as we want in a fast way
> > > > 
> > > > I was sold that bpf needed a quick and fast way to get the arguments of a
> > > > function, as the only way to do that with ftrace is to save all registers,
> > > > which, I was told was too much overhead, as if you only care about
> > > > arguments, there's much less that is needed to save.
> > > > 
> > > > Direct calls wasn't added so that bpf and ftrace could co-exist, it was
> > > > that for certain cases, bpf wanted a faster way to access arguments,
> > > > because it still worked with ftrace, but the saving of regs was too
> > > > strenuous.
> > > 
> > > Direct calls in ftrace were done so that ftrace and trampoline can co-exist.
> > > There is no other use for it.
> > > 
> > > Jiri,
> > > could you please redo your benchmarking hardcoding ftrace_managed=false ?
> > > If going through register_ftrace_direct() is indeed so much slower
> > > than arch_text_poke() then something gotta give.
> > > Either register_ftrace_direct() has to become faster or users
> > > have to give up on co-existing of bpf and ftrace.
> > > So far not a single user cared about using trampoline and ftrace together.
> > > So the latter is certainly an option.
> > 
> > I tried that, and IIRC it was not much faster, but I don't have details
> > on that.. but it should be quick check, I'll do it
> > 
> > anyway later I realized that for us we need ftrace to stay, so I abandoned
> > this idea ;-) and started to check on how to keep them both together and
> > just make it faster
> > 
> > also currently bpf trampolines will not work without ftrace being
> > enabled, because ftrace is doing the preparation work during compile,
> > and replaces all the fentry calls with nop instructions and the
> > replace code depends on those nops...  so if we go this way, we would
> > need to make this preparation code generic
> 
> I didn't mean that part.
> I was talking about register_ftrace_direct() only.
> Could you please still do ftrace_managed=false experiment?
> Sounds like the time to attach/detach will stay the same?
> If so, then don't touch ftrace internals then. What's the point?

actually, there's some speedup.. by running:

  # perf stat --table -e cycles:k,cycles:u -r 10 ./src/bpftrace -ve 'kfunc:__x64_sys_s* { } i:ms:10 { print("exit\n"); exit();}'

I've got following numbers on base:

     3,463,157,566      cycles:k                                                      ( +-  0.14% )
     1,164,026,270      cycles:u                                                      ( +-  0.29% )

             # Table of individual measurements:
             37.61 (-12.20) #######
             49.35 (-0.46) #
             54.03 (+4.22) ##
             50.82 (+1.01) #
             46.87 (-2.94) ##
             53.10 (+3.29) ##
             58.27 (+8.46) ###
             64.85 (+15.04) #####
             47.37 (-2.44) ##
             35.83 (-13.98) ########

             # Final result:
             49.81 +- 2.76 seconds time elapsed  ( +-  5.54% )


and following numbers with the patch below:

     2,037,364,413      cycles:k        ( +-  0.52% )
     1,164,769,939      cycles:u        ( +-  0.19% )

             # Table of individual measurements:
             30.52 (-8.54) ######
             43.43 (+4.37) ###
             43.72 (+4.66) ###
             35.70 (-3.36) ##
             40.70 (+1.63) #
             43.51 (+4.44) ###
             26.44 (-12.62) ##########
             40.21 (+1.14) #
             43.32 (+4.25) ##
             43.09 (+4.03) ##

             # Final result:
             39.06 +- 1.95 seconds time elapsed  ( +-  4.99% )


it looks like even ftrace_managed=false could be faster
with batch update, which is not used, but there's support
for it via text_poke_bp_batch function

jirka


---
diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index 35c5887d82ff..0a241e6eac7d 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c
@@ -111,6 +111,8 @@ static int is_ftrace_location(void *ip)
 {
 	long addr;
 
+	return 0;
+
 	addr = ftrace_location((long)ip);
 	if (!addr)
 		return 0;