Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 13 Aug 2020 10:56:44 -0400
From:   Steven Rostedt <rostedt@goodmis.org>
To:     Jiafei Pan <jiafei.pan@nxp.com>
Cc:     "peterz@infradead.org" <peterz@infradead.org>,
        "mingo@kernel.org" <mingo@kernel.org>,
        "tglx@linutronix.de" <tglx@linutronix.de>,
        "romain.perier@gmail.com" <romain.perier@gmail.com>,
        "will@kernel.org" <will@kernel.org>,
        "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
        "linux-rt-users@vger.kernel.org" <linux-rt-users@vger.kernel.org>,
        Leo Li <leoyang.li@nxp.com>,
        Vladimir Oltean <vladimir.oltean@nxp.com>
Subject: Re: [PATCH] softirq: add irq off checking for
 __raise_softirq_irqoff
Message-ID: <20200813105644.1eb6f2cd@oasis.local.home>
In-Reply-To: <AM0PR04MB4772B1B6D3391BAEAD4411808A430@AM0PR04MB4772.eurprd04.prod.outlook.com>
References: <20200806040729.39186-1-Jiafei.Pan@nxp.com>
        <AM0PR04MB4772B1B6D3391BAEAD4411808A430@AM0PR04MB4772.eurprd04.prod.outlook.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Thu, 13 Aug 2020 03:03:46 +0000
Jiafei Pan <jiafei.pan@nxp.com> wrote:

> Any comments? Thanks.
> 
> @Steven Rostedt, I thinks irq off checking is necessary especially

This is probably more for Thomas Gleixner.

> for Preempt-RT kernel, because some context may be changed from irq
> off to irq on when enable Preempt RT, I once met a issue that hrtimer
> soft irq is lost when enabled Preempt RT, finally I found
> napi_schedule_irqoff is called in hardware interrupt handler, there
> maybe no issue for non RT kernel, but for Preempt RT, interrupt is
> threaded, so irq is on in interrupt handler, the result is
> __raise_softirq_irqoff is called in irq on context, so that per-CPU
> softirq masking is corrupted because of the process of updating of
> soft irq masking is interrupted and not a atomic operation , and then
> caused hrtimer soft irq is lost. So I think adding irq status
> checking in __raise_softirq_irqoff can report such issue directly and
> help us to find the root cause of such issue.
> 
> I know that there may be performance impaction to add extra checking
> here, if it is the case, how about to include it in some debug
> configuration items? Such as CONFIG_DEBUG_PREEMPT or other debug
> items?
> 


> Best Regards,
> Jiafei.
> 
> -----Original Message-----
> From: Jiafei Pan <Jiafei.Pan@nxp.com> 
> Sent: Thursday, August 6, 2020 12:07 PM
> To: peterz@infradead.org; mingo@kernel.org; tglx@linutronix.de; rostedt@goodmis.org; romain.perier@gmail.com; will@kernel.org
> Cc: linux-kernel@vger.kernel.org; linux-rt-users@vger.kernel.org; Jiafei Pan <jiafei.pan@nxp.com>; Leo Li <leoyang.li@nxp.com>; Vladimir Oltean <vladimir.oltean@nxp.com>; Jiafei Pan <jiafei.pan@nxp.com>
> Subject: [PATCH] softirq: add irq off checking for __raise_softirq_irqoff
> 
> __raise_softirq_irqoff will update per-CPU mask of pending softirqs, it need to be called in irq disabled context in order to keep it atomic operation, otherwise it will be interrupted by hardware interrupt, and per-CPU softirqs pending mask will be corrupted, the result is there will be unexpected issue, for example hrtimer soft irq will be losed and soft hrtimer will never be expire and handled.

Please wrap your change logs.

> 
> Adding irqs disabled checking here to provide warning in irqs enabled context.
> 
> Signed-off-by: Jiafei Pan <Jiafei.Pan@nxp.com>
> ---
>  kernel/softirq.c | 5 +++++
>  1 file changed, 5 insertions(+)
> 
> diff --git a/kernel/softirq.c b/kernel/softirq.c index bf88d7f62433..11f61e54a3ae 100644
> --- a/kernel/softirq.c
> +++ b/kernel/softirq.c
> @@ -481,6 +481,11 @@ void raise_softirq(unsigned int nr)
>  
>  void __raise_softirq_irqoff(unsigned int nr)  {
> +	/* This function can only be called in irq disabled context,
> +	 * otherwise or_softirq_pending will be interrupted by hardware
> +	 * interrupt, so that there will be unexpected issue.
> +	 */
> +	WARN_ON_ONCE(!irqs_disabled());

Perhaps: lockdep_assert_irqs_disabled() is more appropriate, and
doesn't add extra overhead on production systems.

-- Steve


>  	trace_softirq_raise(nr);
>  	or_softirq_pending(1UL << nr);
>  }
> --
> 2.17.1