Re: [PATCH] [7/9] pasemi: Configure DMA controller interrupts

On Mon, Feb 05, 2007 at 10:53:49AM +1100, Benjamin Herrenschmidt wrote:

On Thu, 2007-02-01 at 22:54 -0600, Olof Johansson wrote:

quoted

plain text document attachment (pasemi-dma-irqs)
The DMA controller on PWRficient is somewhat special -- has a PCI header
so it looks like it's on the root PCI (-Express) root bus, but it uses
more than the default number of interrupts (and they are hardwired).

We need to wire up all interrupts for the DMA controller. The generic
IRQ code will only map the primary interrupt from the PCI header (128),
so add 129->211 by hand.

Or you can have them in the device-tree ... the generic code only maps
PCI IRQ index 0 but nothing prevents your fixup code to map the other
ones.

I don't see the point in having them in the device tree fixups instead
of pci fixups. I'll move them to the driver for now.

However, I do wonder what is the point of creating a mapping and nor
storing the resulting virq anywhere ?

Current usage was relying on the virqs being allocated linearly with
the rest. Yes, that's in hindsight broken.

When are those IRQs actually used ?

In the ethernet driver;

       ret = request_irq(mac->dma_pdev->irq + mac->dma_txch,
                         &pasemi_mac_tx_intr, IRQF_DISABLED,
                         mac->tx->irq_name, dev);
[...]

       ret = request_irq(mac->dma_pdev->irq + 20 + mac->dma_rxch,
                         &pasemi_mac_rx_intr, IRQF_DISABLED,
                         mac->rx->irq_name, dev);

That is the place where the mapping should happen... There is no
guarantee that the resulting virq from irq_create_mapping() will be the
same as the HW irq you passed in (it will in -most- cases with the
current code, but it's a bug to rely on that behaviour). Thus you should
store the result of irq_create_mapping() and use that. I see no
advantage in "pre-mapping" things that way.

I'll submit an incremental patch to the ethernet driver.


-Olof