Received-SPF: pass (google.com: domain of linux-kernel+bounces-77873-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.199.223 as permitted sender) client-ip=147.75.199.223;
From: Thomas Gleixner <tglx@linutronix.de>
To: Leonardo Bras <leobras@redhat.com>
Cc: Leonardo Bras <leobras@redhat.com>, Greg Kroah-Hartman
 <gregkh@linuxfoundation.org>, Jiri Slaby <jirislaby@kernel.org>, Tony
 Lindgren <tony@atomide.com>, Andy Shevchenko
 <andriy.shevchenko@linux.intel.com>, John Ogness
 <john.ogness@linutronix.de>, Ilpo =?utf-8?Q?J=C3=A4rvinen?=
 <ilpo.jarvinen@linux.intel.com>, Uwe =?utf-8?Q?Kleine-K=C3=B6nig?=
 <u.kleine-koenig@pengutronix.de>, Florian Fainelli
 <florian.fainelli@broadcom.com>, Shanker Donthineni
 <sdonthineni@nvidia.com>, linux-kernel@vger.kernel.org,
 linux-serial@vger.kernel.org
Subject: Re: [RFC PATCH v2 3/4] irq: Introduce IRQ_HANDLED_MANY
In-Reply-To: <ZdghE6TNHgZ_bi19@LeoBras>
References: <20240216075948.131372-2-leobras@redhat.com>
 <20240216075948.131372-5-leobras@redhat.com> <87zfvwai62.ffs@tglx>
 <87v86kaf84.ffs@tglx> <ZdWMja3BfCZsbF_q@LeoBras> <87edd5hljz.ffs@tglx>
 <ZdghE6TNHgZ_bi19@LeoBras>
Date: Fri, 23 Feb 2024 08:33:58 +0100
Message-ID: <87cysneis9.ffs@tglx>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain

On Fri, Feb 23 2024 at 01:37, Leonardo Bras wrote:
> On Wed, Feb 21, 2024 at 04:41:20PM +0100, Thomas Gleixner wrote:
>> There is usually no concurrency at all, except for administrative
>> operations like enable/disable or affinity changes. Those administrative
>> operations are not high frequency and the resulting cache line bouncing
>> is unavoidable even without that change. But does it matter in the
>> larger picture? I don't think so.
>
> That's a fair point, but there are some use cases that use CPU Isolation on 
> top of PREEMPT_RT in order to reduce interference on a CPU running an RT 
> workload.
>
> For those cases, IIRC the handler will run on a different (housekeeping) 
> CPU when those IRQs originate on an Isolated CPU, meaning the above 
> described cacheline bouncing is expected.

No. The affinity of the interrupt and the thread are strictly coupled
and always on the same CPU except for one instance during migration, but
that's irrelevant.

Thanks,

        tglx