[PATCH] Add IPIC MSI interrupt support

[Li Tony]

> -----Original Message-----
> From: Michael Ellerman [mailto:michael at ellerman.id.au] 
> Sent: 2007年12月4日 13:38
> To: Li Tony
> Cc: Li Tony; Gala Kumar; linuxppc-dev
> Subject: Re: [PATCH] Add IPIC MSI interrupt support
> 
> On Mon, 2007-12-03 at 17:07 +0800, Li Li wrote:
> > Hi Michael,
> > 
> > I emulate mpic to write this IPIC MSI routines. :)
> > 
> > 
> > > > diff --git a/arch/powerpc/platforms/83xx/mpc837x_mds.c 
> > > > b/arch/powerpc/platforms/83xx/mpc837x_mds.c
> > > > index 6048f1b..dbea34b 100644
> > > > --- a/arch/powerpc/platforms/83xx/mpc837x_mds.c
> > > > +++ b/arch/powerpc/platforms/83xx/mpc837x_mds.c
> 
> > > > +
> > > > +#define	ipic_msi_irq_to_hw(virq)	
> ((unsigned int)irq_map[virq].hwirq)
> > > 
> > > What's wrong with virq_to_hw() ?
> > > 
> > 
> > viqr_to_hw is not __inline__.
> 
> Hmm, ok. The three places you use it you also take a spin 
> lock, so I'm not sure the function call's really going to 
> kill you performance wise.
> 

I am not very sure about spin_lock influence.
But maybe somebody will change the virq_to_hw implementation.
I will take virq_to_hw instead. 

I see that the virq_to_hw is do inline in 2.6.22.
Why remove it?

> > > > +
> > > > +static void ipic_msi_compose_msg(struct ipic_msi *msi, 
> int hwirq,
> > > > +						struct 
> msi_msg *msg)
> > > > +{
> > > > +	unsigned int srs;
> > > > +	unsigned int ibs;
> > > > +
> > > > +	srs = hwirq / msi->int_per_msir;
> > > > +	ibs = hwirq - srs * msi->int_per_msir;
> > > > +
> > > > +	msg->address_lo = msi->msi_addr_lo;
> > > > +	msg->address_hi = msi->msi_addr_hi;
> > > > +	msg->data = (srs << 5) | (ibs & 0x1F);
> > > > +
> > > > +	pr_debug("%s: allocated srs: %d, ibs: %d\n",
> > > > +		__FUNCTION__, srs, ibs);
> > > > +
> > > > +}
> > > > +
> > > > +static int ipic_msi_setup_irqs(struct pci_dev *pdev, int nvec, 
> > > > +int type) {
> > > > +	struct ipic_msi *msi = ipic_msi;
> > > > +	irq_hw_number_t hwirq;
> > > > +	unsigned int virq;
> > > > +	struct msi_desc *entry;
> > > > +	struct msi_msg msg;
> > > > +
> > > > +	list_for_each_entry(entry, &pdev->msi_list, list) {
> > > > +		hwirq = ipic_msi_alloc_hwirqs(msi, 1);
> > > > +		if (hwirq < 0) {
> > > > +			pr_debug("%s: fail allocating 
> msi interrupt\n",
> > > > +					__FUNCTION__);
> > > > +			return hwirq;
> > > > +		}
> > > > +
> > > > +		/* This hwirq belongs to the irq_host 
> other than irq_host of IPIC
> > > > + 		 * So, it is independent to hwirq of IPIC */
> > > > +		virq = irq_create_mapping(msi->irqhost, hwirq);
> > > > +		if (virq == NO_IRQ) {
> > > > +			pr_debug("%s: fail mapping 
> hwirq 0x%lx\n",
> > > > +					__FUNCTION__, hwirq);
> > > > +			ipic_msi_free_hwirqs(msi, hwirq, 1);
> > > > +			return -ENOSPC;
> > > > +		}
> > > > +		set_irq_msi(virq, entry);
> > > > +		ipic_msi_compose_msg(msi, hwirq, &msg);
> > > > +		write_msi_msg(virq, &msg);
> > > > +
> > > > +		hwirq++;
> > > 
> > >                   ^^^^ this looks like my bug
> > 
> > I have a question here. Do we support more MSI interrupts 
> on ONE pci 
> > device?
> 
> I'm not sure what you mean? For MSI there is only one MSI per 
> device, but this code is used also for MSI-X which supports > 
> 1 MSI per device.
> 
> Either way we shouldn't be incrementing hwirq by hand, it's 
> reassigned at the top of the loop. I think that's left over 
> from old code that allocated nvec hwirqs in a block and then 
> created virq mappings for each one, whereas the new code 
> allocates each hwirq separately.
> 
> cheers
> 
> --
> Michael Ellerman
> OzLabs, IBM Australia Development Lab
> 
> wwweb: http://michael.ellerman.id.au
> phone: +61 2 6212 1183 (tie line 70 21183)
> 
> We do not inherit the earth from our ancestors, we borrow it 
> from our children. - S.M.A.R.T Person
>