Re: [PATCH] RISC-V: insn: Use a raw spinlock to protect TEXT_POKE*

From: Changbin Du <changbin.du@gmail.com>
To: Steven Rostedt <rostedt@goodmis.org>
Cc: Palmer Dabbelt <palmer@dabbelt.com>,
	linux-riscv@lists.infradead.org,
	Paul Walmsley <paul.walmsley@sifive.com>,
	aou@eecs.berkeley.edu, peterz@infradead.org, jpoimboe@redhat.com,
	jbaron@akamai.com, ardb@kernel.org,
	Atish Patra <Atish.Patra@wdc.com>,
	Anup Patel <Anup.Patel@wdc.com>,
	akpm@linux-foundation.org, rppt@kernel.org, mhiramat@kernel.org,
	zong.li@sifive.com, guoren@linux.alibaba.com,
	wangkefeng.wang@huawei.com, 0x7f454c46@gmail.com,
	chenhuang5@huawei.com, linux-kernel@vger.kernel.org,
	kernel-team@android.com,
	Palmer Dabbelt <palmerdabbelt@google.com>,
	Changbin Du <changbin.du@gmail.com>
Subject: Re: [PATCH] RISC-V: insn: Use a raw spinlock to protect TEXT_POKE*
Date: Fri, 30 Apr 2021 05:54:51 +0800	[thread overview]
Message-ID: <20210429215451.yuey5gzmfh2dkzp5@mail.google.com> (raw)
In-Reply-To: <20210429123007.5144fc0d@gandalf.local.home>

On Thu, Apr 29, 2021 at 12:30:07PM -0400, Steven Rostedt wrote:
> On Wed, 28 Apr 2021 23:17:13 -0700
> Palmer Dabbelt <palmer@dabbelt.com> wrote:
> 
> > From: Palmer Dabbelt <palmerdabbelt@google.com>
> > 
> > We currently use text_mutex to protect the fixmap sections from
> > concurrent callers.  This is convienent for kprobes as the generic code
> > already holds text_mutex, but ftrace doesn't which triggers a lockdep
> > assertion.  We could take text_mutex for ftrace, but the jump label
> > implementation (which is currently taking text_mutex) isn't explicitly
> > listed as being sleepable and it's called from enough places it seems
> > safer to just avoid sleeping.
> > 
> > arm64 and parisc, the other two TEXT_POKE-style patching
> > implemnetations, already use raw spinlocks.  abffa6f3b157 ("arm64:
> > convert patch_lock to raw lock") lays out the case for a raw spinlock as
> > opposed to a regular spinlock, and while I don't know of anyone using rt
> > on RISC-V I'm sure it'll eventually show up and I don't see any reason
> > to wait.
> 
> On x86 we use text_mutex for jump label and ftrace. I don't understand the
> issue here. The arm64 update was already using spin locks in the
> insn_write() function itself. riscv just makes sure that text_mutex is held.
> 
> It also looks like ftrace on riscv should also have text_mutex held
> whenever it modifies the code. Because I see this in
> arch/riscv/kernel/ftrace.c:
> 
> 
> int ftrace_arch_code_modify_prepare(void) __acquires(&text_mutex)
> {
>         mutex_lock(&text_mutex);
>         return 0;
> }
> 
> int ftrace_arch_code_modify_post_process(void) __releases(&text_mutex)
> {
>         mutex_unlock(&text_mutex);
>         return 0;
> }
> 
> Which should be getting called before and after respectively from when
> ftrace does its updates.
> 
> Can you show me the back trace of that lockdep splat?
> 
The problem is that lockdep cannot handle locks across tasks since we use
stopmachine to patch code for risc-v. So there's a false positive report.
See privious disscussion here:
https://lkml.org/lkml/2021/4/29/63

> -- Steve

-- 
Cheers,
Changbin Du