Re: INTC issue

From: Magnus Damm <magnus.damm@gmail.com>
To: linux-sh@vger.kernel.org
Subject: Re: INTC issue
Date: Tue, 10 Feb 2009 08:13:14 +0000	[thread overview]
Message-ID: <aec7e5c30902100013y45bffa85g3553de9064d56905@mail.gmail.com> (raw)
In-Reply-To: <498AD5E4.8050809@renesas.com>

Hi Shimoda-san,

Thanks for bringing up this issue and thanks for the patch!

On Thu, Feb 5, 2009 at 9:04 PM, Yoshihiro Shimoda
<shimoda.yoshihiro@renesas.com> wrote:
> Hi,
>
> I found INTC issue. For example, an interrupt is prohibited when I handle it as follows.
>
>  1) Driver A call request_irq of DMAC ch0
>  2) Driver B call request_irq of DMAC ch1.
>  3) Driver A call free_irq of DMAC ch0.
>  4) Driver B use DMAC ch1, but DMAC ch1 interrupt disabled.
>
> The interrupt mask register of SH7785 is INT2MSKR.
> The bit 11 of INT2MSKR is DMAC channels 0 to 5 interrupts mask bit.
>
> I think that it is necessary to count the number of times admitted
> an interrupt to solve this problem. I think that intc have a counter
> every interrupt bit for a simple implementation method, but there
> is much waste.
>
> Any ideas?

Hm, counter may work half-ok, but I think we end up with broken logic
anyhow. Counters would result in that enabling and disabling of
interrupts sometimes result in enable/disable depending on the count
value. For interrupt sources belonging to a single driver we can
probably live with that, but it may become very strange if we share
enable bit between multiple drivers. A driver most likely expects that
interrupt disable really disables interrupts. =)

I would prefer to map all vectors that share a enable/disable bit to a
single interrupt number instead. This is somewhat similar to interrupt
sharing. The common irq code handles interrupt count for us. Look at
desc->depth inside linux/kernel/irq/*.

As an example, right now in the sh7785 dma case you have 14 vectors
divided into 7 DMAC0 sources and 7 DMAC1 sources. A total of 14 dma
interrupts in linux. I would prefer to map them to 2 linux interrupts
instead, one for DMAC0 and one for DMAC1. So all DMAC0 vectors map to
the DMAC0 linux interrupt and same for DMAC1. Each linux interrupt can
be enabled and disabled as a regular interrupt source.

Would that work for you? I hope you can determine interrupt source by
checking status bits in the dma hardware block. Or maybe you can't and
need separate interrupt handlers?

Cheers,

/ magnus