Re: [PATCH v3 22/25] irq_domain/x86: Convert x86 (embedded) to use common irq_domain

On Thu, Feb 23, 2012 at 10:22:15PM +0100, Sebastian Andrzej Siewior wrote:

On 02/23/2012 08:56 PM, Grant Likely wrote:

quoted

On Wed, Feb 1, 2012 at 11:06 AM, Grant Likely[off-list ref]  wrote:

quoted

On Wed, Feb 1, 2012 at 7:17 AM, Sebastian Andrzej Siewior
[off-list ref]  wrote:

quoted

* Grant Likely | 2012-01-30 12:58:42 [-0700]:

quoted

Ugh.  This isn't easy.  The legacy mapping really needs all the

Feel free to merge this patch. I don't have the time to look at this now
so I take a look at the ioapic later.

There's no rush here.  I can leave it as-is with IRQ_DOMAIN turned off
for x86 for now.

Turns out I have to enable IRQ_DOMAIN for x86 because the TI TWL4030
driver needs it.  I do need to apply this patch.  Until something
better can be implemented, can I change ioapic_add_ofnode() so that it
allocates all irq_descs immediately.  It's not ideal, but every other
approach I've looked at results in nasty hacks.

Looking at the ioapic code, it appears to handle preallocated
irq_descs gracefully.

Please merge your initial patch as-it.

Okay, will do.

quoted

Does adding this loop help (apologies if it is whitespace damaged, I
cut&paste it):

diff --git a/arch/x86/kernel/devicetree.c b/arch/x86/kernel/devicetree.c
index 3ae2ced..89c1310 100644
--- a/arch/x86/kernel/devicetree.c
+++ b/arch/x86/kernel/devicetree.c

@@ -345,7 +345,7 @@ const struct irq_domain_ops ioapic_irq_domain_ops = {
 static void __init ioapic_add_ofnode(struct device_node *np)
 {
 	struct resource r;
-	int i, ret;
+	int i, j, ret;

 	ret = of_address_to_resource(np, 0,&r);
 	if (ret) {

@@ -361,6 +361,14 @@ static void __init ioapic_add_ofnode(struct

device_node *np)

 			gsi_cfg = mp_ioapic_gsi_routing(i);

+			/*
+			 * Preallocate irq_descs so that the legacy mapping
+			 * works, but don't set them up.
+			 * io_apic_setup_irq_pin_once() will finish the set up.
+			 */
+			for (j = 0; j<  32; j++)

It is not 32. If I remember correctly the first ioapic had 24 pins so
did the second. This is ioapic specifc.

Is this better (bigger diffstat, but mostly due to renaming a local varable):

---

diff --git a/arch/x86/kernel/devicetree.c b/arch/x86/kernel/devicetree.c
index 3ae2ced..4028dc6 100644
--- a/arch/x86/kernel/devicetree.c
+++ b/arch/x86/kernel/devicetree.c

@@ -345,7 +345,7 @@ const struct irq_domain_ops ioapic_irq_domain_ops = {
 static void __init ioapic_add_ofnode(struct device_node *np)
 {
 	struct resource r;
-	int i, ret;
+	int i, irq, ret;
 
 	ret = of_address_to_resource(np, 0, &r);
 	if (ret) {

@@ -357,13 +357,22 @@ static void __init ioapic_add_ofnode(struct device_node *np)
 	for (i = 0; i < nr_ioapics; i++) {
 		if (r.start == mpc_ioapic_addr(i)) {
 			struct irq_domain *id;
-			struct mp_ioapic_gsi *gsi_cfg;
-
-			gsi_cfg = mp_ioapic_gsi_routing(i);
-
-			id = irq_domain_add_legacy(np, 32, gsi_cfg->gsi_base, 0,
-						   &ioapic_irq_domain_ops,
-						   (void*)i);
+			struct mp_ioapic_gsi *cfg;
+
+			cfg = mp_ioapic_gsi_routing(i);
+
+			/*
+			 * Preallocate irq_descs so that the legacy mapping
+			 * works, but don't set them up.
+			 * io_apic_setup_irq_pin_once() will finish the set up.
+			 */
+			for (irq = cfg->gsi_base; irq <= cfg->gsi_end; irq++)
+				irq_alloc_desc_at(irq, cpu_to_node(0));
+
+			id = irq_domain_add_legacy(np,
+					cfg->gsi_end - cfg->gsi_base + 1,
+					cfg->gsi_base, 0,
+					&ioapic_irq_domain_ops, (void*)i);
 			BUG_ON(!id);
 			return;
 		}