On 2014/9/17 2:31, Thomas Gleixner wrote:

On Thu, 11 Sep 2014, Jiang Liu wrote:

quoted

 #ifdef CONFIG_HPET_TIMER
+#define	HPET_DOMAIN_REMAPPED		0x80000000
+
+static inline int hpet_dev_id(struct irq_domain *domain)
+{
+	return (int)((long)domain->host_data & ~HPET_DOMAIN_REMAPPED);
+}
+
+static inline bool hpet_remapped(struct irq_domain *domain)
+{
+	return (bool)((long)domain->host_data & HPET_DOMAIN_REMAPPED);
+}

It's kinda odd to have this encoded in domain->host_data.

Hi Thomas,
	I have thought about add a "domain_flags" field to struct
irq_domain to host the remapping flag. But remapping flag is not a
common flag for all architectures, so I adopted the dirty solution
to hide the remapped flag in x86 arch specific code.
How about adding domain_flags field and define IRQ_DOMAIN_FLAG_ARCH1?

quoted

 static int hpet_msi_set_affinity(struct irq_data *data,
 				 const struct cpumask *mask, bool force)
 {
+	struct irq_data *parent = data->parent_data;
 	struct irq_cfg *cfg = irqd_cfg(data);
 	struct msi_msg msg;
-	unsigned int dest;
 	int ret;
 
-	ret = apic_set_affinity(data, mask, &dest);
-	if (ret)
-		return ret;
-
-	hpet_msi_read(data->handler_data, &msg);
-
-	msg.data &= ~MSI_DATA_VECTOR_MASK;
-	msg.data |= MSI_DATA_VECTOR(cfg->vector);
-	msg.address_lo &= ~MSI_ADDR_DEST_ID_MASK;
-	msg.address_lo |= MSI_ADDR_DEST_ID(dest);
-
-	hpet_msi_write(data->handler_data, &msg);
+	ret = parent->chip->irq_set_affinity(parent, mask, force);
+	/* No need to rewrite HPET registers if interrupt is remapped */
+	if (ret >= 0 && !hpet_remapped(data->domain)) {

So we really should use irq_data->chip_data for this, i.e. storing

struct hpet_msi {
       struct msi_msg msg;
       bool remapped;
       /* whatever you need here */
};

OK, we add this flag for HPET, MSI and IOAPIC.

quoted

+		hpet_msi_read(data->handler_data, &msg);
+		msg.data &= ~MSI_DATA_VECTOR_MASK;
+		msg.data |= MSI_DATA_VECTOR(cfg->vector);
+		msg.address_lo &= ~MSI_ADDR_DEST_ID_MASK;
+		msg.address_lo |= MSI_ADDR_DEST_ID(cfg->dest_apicid);

We need the same thing for MSI so this should be a helper function

   msi_update_msg(struct msi_msg *msg, struct irq_cfg *cfg)

Good suggestion, will do it in next version
Regards!
Gerry

quoted

+		hpet_msi_write(data->handler_data, &msg);
+	}
 
-	return IRQ_SET_MASK_OK_NOCOPY;
+	return ret;
 }

 
Thanks,

	tglx