DMARC-Filter: OpenDMARC Filter v1.3.2 smtp.codeaurora.org D31A46071B
Date: Tue, 26 Dec 2017 18:21:54 -0800
From: Stephen Boyd <sboyd@codeaurora.org>
To: David Lechner <david@lechnology.com>
Cc: linux-clk@vger.kernel.org, Michael Turquette <mturquette@baylibre.com>,
        linux-kernel@vger.kernel.org
Subject: Re: [PATCH] clk: fix reentrancy of clk_enable() on UP systems
Message-ID: <20171227022154.GA7997@codeaurora.org>
References: <1514309062-1768-1-git-send-email-david@lechnology.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1514309062-1768-1-git-send-email-david@lechnology.com>
User-Agent: Mutt/1.5.21 (2010-09-15)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2985
Lines: 82

On 12/26, David Lechner wrote:
> Reentrant calls to clk_enable() are not working on UP systems. This is
> caused by the fact spin_trylock_irqsave() always returns true when
> CONFIG_SMP=n (and CONFIG_DEBUG_SPINLOCK=n) which causes the reference
> counting to not work correctly when clk_enable_lock() is called twice
> before clk_enable_unlock() is called (this happens when clk_enable()
> is called from within another clk_enable()).
> 
> This introduces a new set of clk_enable_lock() and clk_enable_unlock()
> functions for UP systems that doesn't use spinlocks but effectively does
> the same thing as the SMP version of the functions.
> 
> Signed-off-by: David Lechner <david@lechnology.com>
> ---
> 
> Previous discussion of this issue for reference:
> * https://patchwork.kernel.org/patch/10108437/
> * https://patchwork.kernel.org/patch/10115483/
> 
> 
>  drivers/clk/clk.c | 39 +++++++++++++++++++++++++++++++++++++++
>  1 file changed, 39 insertions(+)
> 
> diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
> index bb1b1f9..259a77f 100644
> --- a/drivers/clk/clk.c
> +++ b/drivers/clk/clk.c
> @@ -136,6 +136,8 @@ static void clk_prepare_unlock(void)
>  	mutex_unlock(&prepare_lock);
>  }
>  
> +#ifdef CONFIG_SMP
> +
>  static unsigned long clk_enable_lock(void)
>  	__acquires(enable_lock)
>  {
> @@ -170,6 +172,43 @@ static void clk_enable_unlock(unsigned long flags)
>  	spin_unlock_irqrestore(&enable_lock, flags);
>  }
>  
> +#else
> +
> +static unsigned long clk_enable_lock(void)
> +	__acquires(enable_lock)
> +{
> +	unsigned long flags;
> +
> +	local_irq_save(flags);
> +	preempt_disable();

Well we certainly don't want to do both irq save and preemption
disabling. Really all we need to do is save away the current
flags to restore them them later. On UP systems, we would do that
on each call to this function. On SMP systems, we would actually
test the enable_owner and up the enable_refcnt outside of the
locked section because if spin_trylock_irqsave() returns a 0, we
have restored the irq flags to whatever they were before, and
then we test enable_owner. So I suppose if it was safe on SMP to
test enable_owner and up the refcnt without holding any sort of
lock then it's also safe on UP, because current is a "special"
variable.

Does this ugly hack do the trick? I really dislike the semi-colon
barf, but I fail to see another way out of this right now.

diff --git a/drivers/clk/clk.c b/drivers/clk/clk.c
index b56c11f51baf..77b2c5bbba68 100644
--- a/drivers/clk/clk.c
+++ b/drivers/clk/clk.c
@@ -141,7 +141,8 @@ static unsigned long clk_enable_lock(void)
 {
 	unsigned long flags;
 
-	if (!spin_trylock_irqsave(&enable_lock, flags)) {
+	if ((!IS_ENABLED(CONFIG_SMP) && ({local_save_flags(flags); 1;})) ||
+	    !spin_trylock_irqsave(&enable_lock, flags)) {
 		if (enable_owner == current) {
 			enable_refcnt++;
 			__acquire(enable_lock);
-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project