Date: Mon, 25 Sep 2017 23:06:39 +0200 (CEST)
From: Thomas Gleixner <tglx@linutronix.de>
To: Paul Burton <paul.burton@imgtec.com>
cc: Ralf Baechle <ralf@linux-mips.org>, dianders@chromium.org,
        James Hogan <james.hogan@imgtec.com>,
        Brian Norris <briannorris@chromium.org>,
        Jason Cooper <jason@lakedaemon.net>, jeffy.chen@rock-chips.com,
        Marc Zyngier <marc.zyngier@arm.com>, linux-kernel@vger.kernel.org,
        linux-mips@linux-mips.org, tfiga@chromium.org
Subject: Re: [RFC PATCH v1 2/9] genirq: Support shared per_cpu_devid
 interrupts
In-Reply-To: <20170907232542.20589-3-paul.burton@imgtec.com>
Message-ID: <alpine.DEB.2.20.1709252259270.2418@nanos>
References: <1682867.tATABVWsV9@np-p-burton> <20170907232542.20589-1-paul.burton@imgtec.com> <20170907232542.20589-3-paul.burton@imgtec.com>
User-Agent: Alpine 2.20 (DEB 67 2015-01-07)
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1927
Lines: 43

On Thu, 7 Sep 2017, Paul Burton wrote:
> Up until now per_cpu_devid interrupts have not supported sharing. On
> MIPS we have some percpu interrupts which are shared in many systems -
> a single CPU interrupt line may be used to indicate a timer interrupt,
> performance counter interrupt or fast debug channel interrupt. We have
> up until now supported this with a series of hacks, wherein drivers call
> each other's interrupt handlers & our MIPS GIC irqchip driver includes a
> hack which configures the interrupt(s) for all CPUs. In order to allow
> this mess to be cleaned up, this patch introduces support for shared
> per_cpu_devid interrupts.
> 
> The major portion of this is supporting per_cpu_devid interrupts in
> __handle_irq_event_percpu() and then making use of this, via
> handle_irq_event_percpu(), from handler_percpu_devif_irq() to invoke the
> handler for all actions associated with the shared interrupt. This does
> have a few side effects worth noting:
> 
>  - per_cpu_devid interrupts will now add to the entropy pool via
>    add_interrupt_randomness(), where they previously did not.
> 
>  - per_cpu_devid interrupts will record timings when IRQS_TIMINGS is
>    set, via record_irq_time(), where they previously did not.
> 
>  - per_cpu_devid interrupts will handle an IRQ_WAKE_THREAD return from
>    their handlers to wake a thread, where they previously did not.

That's broken because it lacks the magic synchronization which is described
in the comment in __irq_wake_thread().

> I'm not aware of any reason the above should be bad side effects, so
> sharing __handle_irq_event_percpu() for per_cpu_devid interrupts seems
> like a positive.

Now you are :)

The other side effect of this is the extra overhead. You add an extra
conditional into the main interrupt handling function
__handle_irq_event_percpu() and the extra loop and hoops overhead for
handle_percpu_devid_irq().

Thanks,

	tglx