[PATCH bpf-next v2 0/4] kprobes: rethook: x86: Replace kretprobe trampoline with rethook

* [PATCH bpf-next v2 0/4] kprobes: rethook: x86: Replace kretprobe trampoline with rethook
@ 2022-03-25 14:22 Masami Hiramatsu
  2022-03-25 14:23 ` [PATCH bpf-next v2 1/4] kprobes: Use rethook for kretprobe if possible Masami Hiramatsu
                   ` (4 more replies)
  0 siblings, 5 replies; 11+ messages in thread
From: Masami Hiramatsu @ 2022-03-25 14:22 UTC (permalink / raw)
  To: Alexei Starovoitov, Andrii Nakryiko, x86, Peter Zijlstra
  Cc: Thomas Gleixner, Ingo Molnar, Borislav Petkov, Dave Hansen,
	Dan Carpenter, kernel-janitors, Steven Rostedt, Masami Hiramatsu,
	Jiri Olsa, bpf, linux-kernel

Hi,

Here are the 2nd version for generic kretprobe and kretprobe on x86 for
replacing the kretprobe trampoline with rethook. The previous version
is here[1]

[1] https://lore.kernel.org/all/164818251899.2252200.7306353689206167903.stgit@devnote2/T/#u

In this version I added completing pt_regs by saving regs->ss register
for rethook (from Peter) and optprobe.

Background:

This rethook came from Jiri's request of multiple kprobe for bpf[1].
He tried to solve an issue that starting bpf with multiple kprobe will
take a long time because bpf-kprobe will wait for RCU grace period for
sync rcu events.

Jiri wanted to attach a single bpf handler to multiple kprobes and
he tried to introduce multiple-probe interface to kprobe. So I asked
him to use ftrace and kretprobe-like hook if it is only for the
function entry and exit, instead of adding ad-hoc interface
to kprobes.
For this purpose, I introduced the fprobe (kprobe like interface for
ftrace) with the rethook (this is a generic return hook feature for
fprobe exit handler)[2].

[1] https://lore.kernel.org/all/20220104080943.113249-1-jolsa@kernel.org/T/#u
[2] https://lore.kernel.org/all/164191321766.806991.7930388561276940676.stgit@devnote2/T/#u

The rethook is basically same as the kretprobe trampoline. I just made
it decoupled from kprobes. Eventually, the all arch dependent kretprobe
trampolines will be replaced with the rethook trampoline instead of
cloning and set HAVE_RETHOOK=y.
When I port the rethook for all arch which supports kretprobe, the
legacy kretprobe specific code (which is for CONFIG_KRETPROBE_ON_RETHOOK=n)
will be removed eventually.

Worktree notice:

BTW, this patch can be applied to next-20220324, not the bpf-next tree
directly, because this depends on ANNOTATE_NOENDBR macro. However, since
the fprobe is merged in the bpf-next, I marked this for bpf-next.
So until merging the both of fprobes and ENDBR series, to compile this
you need below 2 lines in arch/x86/kernel/rethook.c.

#ifndef ANNOTATE_NOENDBR
#define ANNOTATE_NOENDBR

But after those are merged, these lines will be unneeded. How should I
handle this issue? (Just remove ANNOTATE_NOENDBR line in bpf-next?)

Thank you,

---

Masami Hiramatsu (3):
      kprobes: Use rethook for kretprobe if possible
      rethook: kprobes: x86: Replace kretprobe with rethook on x86
      x86,kprobes: Fix optprobe trampoline to generate complete pt_regs

Peter Zijlstra (1):
      Subject: x86,rethook: Fix arch_rethook_trampoline() to generate a complete pt_regs

 arch/Kconfig                     |    7 ++
 arch/x86/Kconfig                 |    1 
 arch/x86/include/asm/unwind.h    |   23 +++----
 arch/x86/kernel/Makefile         |    1 
 arch/x86/kernel/kprobes/common.h |    1 
 arch/x86/kernel/kprobes/core.c   |  107 ---------------------------------
 arch/x86/kernel/kprobes/opt.c    |   25 +++++---
 arch/x86/kernel/rethook.c        |  123 ++++++++++++++++++++++++++++++++++++++
 arch/x86/kernel/unwind_orc.c     |   10 ++-
 include/linux/kprobes.h          |   51 +++++++++++++++-
 kernel/kprobes.c                 |  124 ++++++++++++++++++++++++++++++++------
 kernel/trace/trace_kprobe.c      |    4 +
 12 files changed, 319 insertions(+), 158 deletions(-)
 create mode 100644 arch/x86/kernel/rethook.c

--
Masami Hiramatsu (Linaro) <mhiramat@kernel.org>

^ permalink raw reply	[flat|nested] 11+ messages in thread