Re: [PATCH] RISC-V: insn: Use a raw spinlock to protect TEXT_POKE*
From: Palmer Dabbelt
Date: Fri Apr 30 2021 - 15:44:41 EST
On Fri, 30 Apr 2021 04:34:31 PDT (-0700), rostedt@xxxxxxxxxxx wrote:
On Fri, 30 Apr 2021 04:06:35 +0000
Anup Patel <Anup.Patel@xxxxxxx> wrote:
This patch only takes care of ftrace path.
The RISC-V instruction patching is used by RISC-V jump label implementation
as well and will called from various critical parts of core kernel.
The RAW spinlock approach allows same instruction patching to be used
for kprobes, ftrace, and jump label.
So what path hits this outside of stop machine?
I didn't actually dig through all the usages of jump_label, I just saw a
handful in places where it's generally not sane to assume that sleeping
is safe -- for example, thoughout kernel/sched. If you think it's OK to
rely on users of the static branch stuff (IIUC the only jump_label user
in the kernel?) to know that it can sleep then I'm fine keeping the
text_mutex call in jump_label and adding one to ftrace (I'm fine with
something generic, but it's simple to do in arch/riscv).
IMO if the static branch stuff can be expected to sleep it'd be good to
call that out in the documentation, and I'd like to audit the uses
before committing to that. I'm happy to do that, we can just take the
lock in arch/riscv's frace code for now to get around the lockdep
assertion failure -- IIUC that's indicating a real bug, as nothing in
ftrace avoids concurrency with jump_label and kprobes.