linux-kernel - Re: [RFC PATCH v1] irqchip: add support for SMP irq router

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <577D0D9C.6050106@arm.com>
Date:	Wed, 6 Jul 2016 14:54:36 +0100
From:	Marc Zyngier <marc.zyngier@....com>
To:	Sebastian Frias <sf84@...oste.net>,
	Thomas Gleixner <tglx@...utronix.de>
Cc:	Jason Cooper <jason@...edaemon.net>, Mason <slash.tmp@...e.fr>,
	LKML <linux-kernel@...r.kernel.org>
Subject: Re: [RFC PATCH v1] irqchip: add support for SMP irq router

On 06/07/16 11:49, Sebastian Frias wrote:
> Hi,
> 
> On 07/06/2016 11:30 AM, Thomas Gleixner wrote:
>> On Wed, 6 Jul 2016, Marc Zyngier wrote:
>>> On 05/07/16 20:24, Thomas Gleixner wrote:
>>>> On Tue, 5 Jul 2016, Marc Zyngier wrote:
>>>>> Hardcoded? No way. You simply implement a route allocator in your
>>>>> driver, assigning them as needed. And yes, if you have more than 24
>>>>> interrupts, they get muxed.
>>>>
>>>> There is one caveat though. Under some circumstances (think RT) you want to
>>>> configure which interrupts get muxed and which not. We really should have that
>>>> option, but yes for anything which has less than 24 autorouting is the way to
>>>> go.
>>>
>>> Good point. I can see two possibilities for that:
>>>
>>> - either we describe this DT with some form of hint, indicating what are
>>> the inputs that can be muxed to a single output. Easy, but the DT guys
>>> are going to throw rocks at me for being Linux-specific.
>>
>> That's not necessarily Linux specific. The problem arises with any other OS as
>> well.
>>
>>> - or we have a way to express QoS in the irq subsystem, and a driver can
>>> request an interrupt with a "make it fast" flag. Of course, everybody
>>> and his dog are going to ask for it, and we're back to square one.
>>
>> That and the driver does not know about the particular application
>> scenario/system configuration.
>>  
>>> Do we have a way to detect which interrupt is more likely to be
>>> sensitive to muxing? My hunch is that if it is requested with
>>> IRQF_SHARED, then it is effectively muxable. Thoughts?
>>
>> That's too late. request_irq happens _after_ the interrupt is set up and the
>> routing established.
>>
> 
> What about using 3 values for the interrupt description like the GIC does?
> When connecting to the GIC we say "interrupts = <GIC_SPI 2 IRQ_TYPE_LEVEL_HIGH>;"
> If devices using this driver (the one from the RFC) requested the interrupt like:
> "interrupts = <0 38 IRQ_TYPE_LEVEL_HIGH>;"
> "interrupts = <2 27 IRQ_TYPE_LEVEL_HIGH>;"
> etc.
> with the first field being the "group", then the driver could create a domain
> for the device's IRQ (or associate it to an existing one if it has already been
> created). It would thus serve as a hint on how to create domains and how to
> share IRQs into the same line (domain).
> 
> I guess I can get such information from the .translate and .alloc callbacks
> from a newly created domain hierarchy attached to the GIC, right?

This wouldn't work. You need to instantiate the domains long before
you've parsed a single interrupt specifier, otherwise you don't know
where to allocate it from.

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...