Re: [PATCH] Add support for multiple MSI on x86

From: Ingo Molnar
Date: Mon Feb 14 2011 - 21:39:32 EST



* Micha Nelissen <micha@xxxxxxxxxxxxxx> wrote:

> Ingo Molnar wrote:
> >* Micha Nelissen <micha@xxxxxxxxxxxxxx> wrote:
> >>Patch is based on earlier patch from Matthew Wilcox.
> >
> >Hm, there's an awful lot of 'background', 'implementation',
> >'advantages', 'disadvantages', 'testing' description missing.
>
> Sorry, my bad, wrong assumption that this would be common knowledge.

When it comes to changes done to a kernel tree used by millions of people then
common knowledge and mundane details need to be written down, in painstaking detail.

You cannot really err when it comes to details: i've reviewed over a ten thousand
patches in my life and i've yet to see a single one where the description was too
verbose ;-)

> PCI devices interrupt the CPU using 'legacy' INTx; or PCI-e devices
> can do a special write, called an MSI (message signaled interrupt).
> The interrupt vector is chosen by the CPU, different devices use
> different vectors so that software can keep the interrupt sources
> separate (instead of using shared INTx "lines").
>
> Most PCI-e devices support what's called MSI-X, but not all do.
> Especially for FPGA based endpoints it's easier to implement only
> 'regular' MSI support. MSI-X basically involves implementing a
> lookup table that maps interrupt types in the device to interrupt
> vectors for the CPU. 'Regular' MSI (non MSI-X) only supports a
> contiguous block of interrupt vectors: a base vector with a vector
> count (which is a power of 2).
>
> The x86 code to allocate these interrupt vectors does not handle the
> <> 1 vector count case; it would return that it could only handle 1.
> So either device drivers had 1 MSI, or the device should support
> MSI-X to have multiple interrupts (and handlers) for one device.
>
> This patch adds the needed code to support multiple MSI per device.
>
> Advantages: separate interrupt handlers for separate tasks in the
> device. This allows device drivers to be better structured. Easy
> 'diagnostics' due to /proc/interrupts counting number of interrupts
> for separate functionality separately.
>
> Disadvantages: more complex code due to requirement that it is a
> contiguous block, so needs some effort to look for a free block with
> the requested count.
>
> Tested: on an Atom platform, with a Xilinx based PCI-e core in FPGA.
>
> Please review; thanks,

That was a good description, thanks Micha!

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/