Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Tue, 10 Mar 2020 12:20:45 +0100
From:   Borislav Petkov <bp@alien8.de>
To:     Thomas Gleixner <tglx@linutronix.de>
Cc:     LKML <linux-kernel@vger.kernel.org>, x86@kernel.org,
        Steven Rostedt <rostedt@goodmis.org>,
        Brian Gerst <brgerst@gmail.com>,
        Juergen Gross <jgross@suse.com>,
        Frederic Weisbecker <frederic@kernel.org>,
        Alexandre Chartre <alexandre.chartre@oracle.com>
Subject: Re: [patch part-II V2 09/13] x86/entry/common: Split hardirq tracing
 into lockdep and ftrace parts
Message-ID: <20200310112045.GD29372@zn.tnic>
References: <20200308222359.370649591@linutronix.de>
 <20200308222609.825111830@linutronix.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Disposition: inline
In-Reply-To: <20200308222609.825111830@linutronix.de>
User-Agent: Mutt/1.10.1 (2018-07-13)
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sun, Mar 08, 2020 at 11:24:08PM +0100, Thomas Gleixner wrote:
> trace_hardirqs_off() is in fact a tracepoint which can be utilized by BPF,
> which is unsafe before calling enter_from_user_mode(), which in turn
> invokes context tracking. trace_hardirqs_off() also invokes
> lockdep_hardirqs_off() under the hood.
> 
> OTOH lockdep needs to know about the interrupts disabled state before
> enter_from_user_mode(). lockdep_hardirqs_off() is safe to call at this
> point.
> 
> Split it so lockdep knows about the state and invoke the tracepoint after
> the context is set straight.
> 
> Even if the functions attached to a tracepoint would all be safe to be
> called in rcuidle having it split up is still giving a performance
> advantage because rcu_read_lock_sched() is avoiding the whole dance of:
> 
>    scru_read_lock();
>    rcu_irq_enter_irqson();
>    ...
>    rcu_irq_exit_irqson();
>    scru_read_unlock();
>    
> around the tracepoint function invocation just to have the C entry points
> of syscalls call enter_from_user_mode() right after the above dance.
> 
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> ---
> V2: New patch
> ---
>  arch/x86/entry/common.c |   13 +++++++++++--
>  1 file changed, 11 insertions(+), 2 deletions(-)
> 
> --- a/arch/x86/entry/common.c
> +++ b/arch/x86/entry/common.c
> @@ -60,10 +60,19 @@ static __always_inline void syscall_entr
>  {
>  	/*
>  	 * Usermode is traced as interrupts enabled, but the syscall entry
> -	 * mechanisms disable interrupts. Tell the tracer.
> +	 * mechanisms disable interrupts. Tell lockdep before calling
> +	 * enter_from_user_mode(). This is safe vs. RCU while the
> +	 * tracepoint is not.
>  	 */
> -	trace_hardirqs_off();
> +	lockdep_hardirqs_on(CALLER_ADDR0);
> +
>  	enter_from_user_mode();
> +
> +	/*
> +	 * Tell the tracer about the irq state as well before enabling
> +	 * interrupts.
> +	 */
> +	__trace_hardirqs_off();

I wonder if those "__" variants should be named something else to
denote better the difference between __trace_hardirqs_{on,off} and
trace_hardirqs_{on,off}. Latter does the _rcuidle variant and lockdep
annotation but

	trace_hardirqs_{on,off}_rcuidle_lockdep()

sounds yuck.

Maybe lockdep_trace_hardirqs_{on,off}()...

Blergh, I can't think of a good name ATM.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette