[PATCH] i386/x86-64: Share interrupt vectors when there is a large number of interrupt sources

Here's a patch that builds on Natalie Protasevich's IRQ compression
patch and tries to work for MPS boots as well as ACPI.  It is meant for
a 4-node IBM x460 NUMA box, which was dying because it had interrupt
pins with GSI numbers > NR_IRQS and thus overflowed irq_desc.

The problem is that this system has 270 GSIs (which are 1:1 mapped with
I/O APIC RTEs) and an 8-node box would have 540.  This is much bigger
than NR_IRQS (224 for both i386 and x86_64).  Also, there aren't enough
vectors to go around.  There are about 190 usable vectors, not counting
the reserved ones and the unused vectors at 0x20 to 0x2F.  So, my patch
attempts to compress the GSI range and share vectors by sharing IRQs.

Cc: "Protasevich, Natalie" <Natalie.Protasevich@unisys.com>

Signed-off-by: Andi Kleen <ak@suse.de>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

2 files changed, 74 insertions, 8 deletions

diff --git a/arch/x86_64/kernel/io_apic.c b/arch/x86_64/kernel/io_apic.c
index c8eee20cd519..97154ab058b4 100644
--- a/arch/x86_64/kernel/io_apic.c
+++ b/arch/x86_64/kernel/io_apic.c
@@ -57,7 +57,7 @@ int nr_ioapic_registers[MAX_IO_APICS];
  * Rough estimation of how many shared IRQs there are, can
  * be changed anytime.
  */
-#define MAX_PLUS_SHARED_IRQS NR_IRQS
+#define MAX_PLUS_SHARED_IRQS NR_IRQ_VECTORS
 #define PIN_MAP_SIZE (MAX_PLUS_SHARED_IRQS + NR_IRQS)
 
 /*
@@ -85,6 +85,7 @@ int vector_irq[NR_VECTORS] __read_mostly = { [0 ... NR_VECTORS - 1] = -1};
 	int pin;							\
 	struct irq_pin_list *entry = irq_2_pin + irq;			\
 									\
+	BUG_ON(irq >= NR_IRQS);						\
 	for (;;) {							\
 		unsigned int reg;					\
 		pin = entry->pin;					\
@@ -127,6 +128,8 @@ static void set_ioapic_affinity_irq(unsigned int irq, cpumask_t mask)
 }
 #endif
 
+static u8 gsi_2_irq[NR_IRQ_VECTORS] = { [0 ... NR_IRQ_VECTORS-1] = 0xFF };
+
 /*
  * The common case is 1:1 IRQ<->pin mappings. Sometimes there are
  * shared ISA-space IRQs, so we have to support them. We are super
@@ -137,6 +140,7 @@ static void add_pin_to_irq(unsigned int irq, int apic, int pin)
 	static int first_free_entry = NR_IRQS;
 	struct irq_pin_list *entry = irq_2_pin + irq;
 
+	BUG_ON(irq >= NR_IRQS);
 	while (entry->next)
 		entry = irq_2_pin + entry->next;
 
@@ -144,7 +148,7 @@ static void add_pin_to_irq(unsigned int irq, int apic, int pin)
 		entry->next = first_free_entry;
 		entry = irq_2_pin + entry->next;
 		if (++first_free_entry >= PIN_MAP_SIZE)
-			panic("io_apic.c: whoops");
+			panic("io_apic.c: ran out of irq_2_pin entries!");
 	}
 	entry->apic = apic;
 	entry->pin = pin;
@@ -420,6 +424,7 @@ int IO_APIC_get_PCI_irq_vector(int bus, int slot, int pin)
 				best_guess = irq;
 		}
 	}
+	BUG_ON(best_guess >= NR_IRQS);
 	return best_guess;
 }
 
@@ -610,6 +615,64 @@ static inline int irq_trigger(int idx)
 	return MPBIOS_trigger(idx);
 }
 
+static int next_irq = 16;
+
+/*
+ * gsi_irq_sharing -- Name overload!  "irq" can be either a legacy IRQ
+ * in the range 0-15, a linux IRQ in the range 0-223, or a GSI number
+ * from ACPI, which can reach 800 in large boxen.
+ *
+ * Compact the sparse GSI space into a sequential IRQ series and reuse
+ * vectors if possible.
+ */
+int gsi_irq_sharing(int gsi)
+{
+	int i, tries, vector;
+
+	BUG_ON(gsi >= NR_IRQ_VECTORS);
+
+	if (platform_legacy_irq(gsi))
+		return gsi;
+
+	if (gsi_2_irq[gsi] != 0xFF)
+		return (int)gsi_2_irq[gsi];
+
+	tries = NR_IRQS;
+  try_again:
+	vector = assign_irq_vector(gsi);
+
+	/*
+	 * Sharing vectors means sharing IRQs, so scan irq_vectors for previous
+	 * use of vector and if found, return that IRQ.  However, we never want
+	 * to share legacy IRQs, which usually have a different trigger mode
+	 * than PCI.
+	 */
+	for (i = 0; i < NR_IRQS; i++)
+		if (IO_APIC_VECTOR(i) == vector)
+			break;
+	if (platform_legacy_irq(i)) {
+		if (--tries >= 0) {
+			IO_APIC_VECTOR(i) = 0;
+			goto try_again;
+		}
+		panic("gsi_irq_sharing: didn't find an IRQ using vector 0x%02X for GSI %d", vector, gsi);
+	}
+	if (i < NR_IRQS) {
+		gsi_2_irq[gsi] = i;
+		printk(KERN_INFO "GSI %d sharing vector 0x%02X and IRQ %d\n",
+				gsi, vector, i);
+		return i;
+	}
+
+	i = next_irq++;
+	BUG_ON(i >= NR_IRQS);
+	gsi_2_irq[gsi] = i;
+	IO_APIC_VECTOR(i) = vector;
+	printk(KERN_INFO "GSI %d assigned vector 0x%02X and IRQ %d\n",
+			gsi, vector, i);
+	return i;
+}
+
 static int pin_2_irq(int idx, int apic, int pin)
 {
 	int irq, i;
@@ -639,6 +702,7 @@ static int pin_2_irq(int idx, int apic, int pin)
 			while (i < apic)
 				irq += nr_ioapic_registers[i++];
 			irq += pin;
+			irq = gsi_irq_sharing(irq);
 			break;
 		}
 		default:
@@ -648,6 +712,7 @@ static int pin_2_irq(int idx, int apic, int pin)
 			break;
 		}
 	}
+	BUG_ON(irq >= NR_IRQS);
 
 	/*
 	 * PCI IRQ command line redirection. Yes, limits are hardcoded.
@@ -663,6 +728,7 @@ static int pin_2_irq(int idx, int apic, int pin)
 			}
 		}
 	}
+	BUG_ON(irq >= NR_IRQS);
 	return irq;
 }
 
@@ -690,8 +756,8 @@ int assign_irq_vector(int irq)
 {
 	static int current_vector = FIRST_DEVICE_VECTOR, offset = 0;
 
-	BUG_ON(irq >= NR_IRQ_VECTORS);
-	if (IO_APIC_VECTOR(irq) > 0)
+	BUG_ON(irq != AUTO_ASSIGN && (unsigned)irq >= NR_IRQ_VECTORS);
+	if (irq != AUTO_ASSIGN && IO_APIC_VECTOR(irq) > 0)
 		return IO_APIC_VECTOR(irq);
 next:
 	current_vector += 8;
@@ -699,9 +765,8 @@ next:
 		goto next;
 
 	if (current_vector >= FIRST_SYSTEM_VECTOR) {
-		offset++;
-		if (!(offset%8))
-			return -ENOSPC;
+		/* If we run out of vectors on large boxen, must share them. */
+		offset = (offset + 1) % 8;
 		current_vector = FIRST_DEVICE_VECTOR + offset;
 	}
 
@@ -1917,6 +1982,7 @@ int io_apic_set_pci_routing (int ioapic, int pin, int irq, int edge_level, int a
 	entry.polarity = active_high_low;
 	entry.mask = 1;					 /* Disabled (masked) */
 
+	irq = gsi_irq_sharing(irq);
 	/*
 	 * IRQs < 16 are already in the irq_2_pin[] map
 	 */
diff --git a/arch/x86_64/kernel/mpparse.c b/arch/x86_64/kernel/mpparse.c
index f16d38d09daf..8f6958e79455 100644
--- a/arch/x86_64/kernel/mpparse.c
+++ b/arch/x86_64/kernel/mpparse.c
@@ -218,7 +218,7 @@ static void __init MP_intsrc_info (struct mpc_config_intsrc *m)
 			m->mpc_irqtype, m->mpc_irqflag & 3,
 			(m->mpc_irqflag >> 2) & 3, m->mpc_srcbus,
 			m->mpc_srcbusirq, m->mpc_dstapic, m->mpc_dstirq);
-	if (++mp_irq_entries == MAX_IRQ_SOURCES)
+	if (++mp_irq_entries >= MAX_IRQ_SOURCES)
 		panic("Max # of irq sources exceeded!!\n");
 }