sparse irq_desc[] array: core kernel and x86 changes

Impact: new feature

Problem on distro kernels: irq_desc[NR_IRQS] takes megabytes of RAM with
NR_CPUS set to large values. The goal is to be able to scale up to much
larger NR_IRQS value without impacting the (important) common case.

To solve this, we generalize irq_desc[NR_IRQS] to an (optional) array of
irq_desc pointers.

When CONFIG_SPARSE_IRQ=y is used, we use kzalloc_node to get irq_desc,
this also makes the IRQ descriptors NUMA-local (to the site that calls
request_irq()).

This gets rid of the irq_cfg[] static array on x86 as well: irq_cfg now
uses desc->chip_data for x86 to store irq_cfg.

Signed-off-by: Yinghai Lu <yinghai@kernel.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>

kernel/irq/autoprobe.c

@@ -40,6 +40,9 @@ unsigned long probe_irq_on(void)
 	 * flush such a longstanding irq before considering it as spurious.
 	 */
 	for_each_irq_desc_reverse(i, desc) {
+		if (!desc)
+			continue;
+
 		spin_lock_irq(&desc->lock);
 		if (!desc->action && !(desc->status & IRQ_NOPROBE)) {
 			/*
@@ -68,6 +71,9 @@ unsigned long probe_irq_on(void)
 	 * happened in the previous stage, it may have masked itself)
 	 */
 	for_each_irq_desc_reverse(i, desc) {
+		if (!desc)
+			continue;
+
 		spin_lock_irq(&desc->lock);
 		if (!desc->action && !(desc->status & IRQ_NOPROBE)) {
 			desc->status |= IRQ_AUTODETECT | IRQ_WAITING;
@@ -86,6 +92,9 @@ unsigned long probe_irq_on(void)
 	 * Now filter out any obviously spurious interrupts
 	 */
 	for_each_irq_desc(i, desc) {
+		if (!desc)
+			continue;
+
 		spin_lock_irq(&desc->lock);
 		status = desc->status;
 
@@ -124,6 +133,9 @@ unsigned int probe_irq_mask(unsigned long val)
 	int i;
 
 	for_each_irq_desc(i, desc) {
+		if (!desc)
+			continue;
+
 		spin_lock_irq(&desc->lock);
 		status = desc->status;
 
@@ -166,6 +178,9 @@ int probe_irq_off(unsigned long val)
 	unsigned int status;
 
 	for_each_irq_desc(i, desc) {
+		if (!desc)
+			continue;
+
 		spin_lock_irq(&desc->lock);
 		status = desc->status;
 

kernel/irq/chip.c

@@ -24,9 +24,10 @@
  */
 void dynamic_irq_init(unsigned int irq)
 {
-	struct irq_desc *desc = irq_to_desc(irq);
+	struct irq_desc *desc;
 	unsigned long flags;
 
+	desc = irq_to_desc(irq);
 	if (!desc) {
 		WARN(1, KERN_ERR "Trying to initialize invalid IRQ%d\n", irq);
 		return;

kernel/irq/handle.c

@@ -15,9 +15,16 @@
 #include <linux/random.h>
 #include <linux/interrupt.h>
 #include <linux/kernel_stat.h>
+#include <linux/rculist.h>
+#include <linux/hash.h>
 
 #include "internals.h"
 
+/*
+ * lockdep: we want to handle all irq_desc locks as a single lock-class:
+ */
+static struct lock_class_key irq_desc_lock_class;
+
 /**
  * handle_bad_irq - handle spurious and unhandled irqs
  * @irq:       the interrupt number
@@ -49,6 +56,155 @@ void handle_bad_irq(unsigned int irq, struct irq_desc *desc)
 int nr_irqs = NR_IRQS;
 EXPORT_SYMBOL_GPL(nr_irqs);
 
+void __init __attribute__((weak)) arch_early_irq_init(void)
+{
+}
+
+#ifdef CONFIG_SPARSE_IRQ
+static struct irq_desc irq_desc_init = {
+	.irq	    = -1,
+	.status	    = IRQ_DISABLED,
+	.chip	    = &no_irq_chip,
+	.handle_irq = handle_bad_irq,
+	.depth      = 1,
+	.lock       = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
+#ifdef CONFIG_SMP
+	.affinity   = CPU_MASK_ALL
+#endif
+};
+
+static void init_kstat_irqs(struct irq_desc *desc, int cpu, int nr)
+{
+	unsigned long bytes;
+	char *ptr;
+	int node;
+
+	/* Compute how many bytes we need per irq and allocate them */
+	bytes = nr * sizeof(unsigned int);
+
+	node = cpu_to_node(cpu);
+	ptr = kzalloc_node(bytes, GFP_ATOMIC, node);
+	printk(KERN_DEBUG "  alloc kstat_irqs on cpu %d node %d\n", cpu, node);
+
+	if (ptr)
+		desc->kstat_irqs = (unsigned int *)ptr;
+}
+
+void __attribute__((weak)) arch_init_chip_data(struct irq_desc *desc, int cpu)
+{
+}
+
+static void init_one_irq_desc(int irq, struct irq_desc *desc, int cpu)
+{
+	memcpy(desc, &irq_desc_init, sizeof(struct irq_desc));
+	desc->irq = irq;
+#ifdef CONFIG_SMP
+	desc->cpu = cpu;
+#endif
+	lockdep_set_class(&desc->lock, &irq_desc_lock_class);
+	init_kstat_irqs(desc, cpu, nr_cpu_ids);
+	if (!desc->kstat_irqs) {
+		printk(KERN_ERR "can not alloc kstat_irqs\n");
+		BUG_ON(1);
+	}
+	arch_init_chip_data(desc, cpu);
+}
+
+/*
+ * Protect the sparse_irqs:
+ */
+static DEFINE_SPINLOCK(sparse_irq_lock);
+
+struct irq_desc *irq_desc_ptrs[NR_IRQS] __read_mostly;
+
+static struct irq_desc irq_desc_legacy[16] __cacheline_aligned_in_smp = {
+	[0 ... 15] = {
+		.irq	    = -1,
+		.status	    = IRQ_DISABLED,
+		.chip	    = &no_irq_chip,
+		.handle_irq = handle_bad_irq,
+		.depth	    = 1,
+		.lock	    = __SPIN_LOCK_UNLOCKED(irq_desc_init.lock),
+#ifdef CONFIG_SMP
+		.affinity   = CPU_MASK_ALL
+#endif
+	}
+};
+
+/* FIXME: use bootmem alloc ...*/
+static unsigned int kstat_irqs_legacy[16][NR_CPUS];
+
+void __init early_irq_init(void)
+{
+	struct irq_desc *desc;
+	int legacy_count;
+	int i;
+
+	desc = irq_desc_legacy;
+	legacy_count = ARRAY_SIZE(irq_desc_legacy);
+
+	for (i = 0; i < legacy_count; i++) {
+		desc[i].irq = i;
+		desc[i].kstat_irqs = kstat_irqs_legacy[i];
+
+		irq_desc_ptrs[i] = desc + i;
+	}
+
+	for (i = legacy_count; i < NR_IRQS; i++)
+		irq_desc_ptrs[i] = NULL;
+
+	arch_early_irq_init();
+}
+
+struct irq_desc *irq_to_desc(unsigned int irq)
+{
+	return (irq < NR_IRQS) ? irq_desc_ptrs[irq] : NULL;
+}
+
+struct irq_desc *irq_to_desc_alloc_cpu(unsigned int irq, int cpu)
+{
+	struct irq_desc *desc;
+	unsigned long flags;
+	int node;
+
+	if (irq >= NR_IRQS) {
+		printk(KERN_WARNING "irq >= NR_IRQS in irq_to_desc_alloc: %d %d\n",
+				irq, NR_IRQS);
+		WARN_ON(1);
+		return NULL;
+	}
+
+	desc = irq_desc_ptrs[irq];
+	if (desc)
+		return desc;
+
+	spin_lock_irqsave(&sparse_irq_lock, flags);
+
+	/* We have to check it to avoid races with another CPU */
+	desc = irq_desc_ptrs[irq];
+	if (desc)
+		goto out_unlock;
+
+	node = cpu_to_node(cpu);
+	desc = kzalloc_node(sizeof(*desc), GFP_ATOMIC, node);
+	printk(KERN_DEBUG "  alloc irq_desc for %d on cpu %d node %d\n",
+		 irq, cpu, node);
+	if (!desc) {
+		printk(KERN_ERR "can not alloc irq_desc\n");
+		BUG_ON(1);
+	}
+	init_one_irq_desc(irq, desc, cpu);
+
+	irq_desc_ptrs[irq] = desc;
+
+out_unlock:
+	spin_unlock_irqrestore(&sparse_irq_lock, flags);
+
+	return desc;
+}
+
+#else
+
 struct irq_desc irq_desc[NR_IRQS] __cacheline_aligned_in_smp = {
 	[0 ... NR_IRQS-1] = {
 		.status = IRQ_DISABLED,
@@ -62,6 +218,8 @@ struct irq_desc irq_desc[NR_IRQS] __cacheline_aligned_in_smp = {
 	}
 };
 
+#endif
+
 /*
  * What should we do if we get a hw irq event on an illegal vector?
  * Each architecture has to answer this themself.
@@ -261,17 +419,28 @@ out:
 
 
 #ifdef CONFIG_TRACE_IRQFLAGS
-/*
- * lockdep: we want to handle all irq_desc locks as a single lock-class:
- */
-static struct lock_class_key irq_desc_lock_class;
-
 void early_init_irq_lock_class(void)
 {
+#ifndef CONFIG_SPARSE_IRQ
 	struct irq_desc *desc;
 	int i;
 
-	for_each_irq_desc(i, desc)
+	for_each_irq_desc(i, desc) {
+		if (!desc)
+			continue;
+
 		lockdep_set_class(&desc->lock, &irq_desc_lock_class);
+	}
+#endif
 }
 #endif
+
+#ifdef CONFIG_SPARSE_IRQ
+unsigned int kstat_irqs_cpu(unsigned int irq, int cpu)
+{
+	struct irq_desc *desc = irq_to_desc(irq);
+	return desc->kstat_irqs[cpu];
+}
+#endif
+EXPORT_SYMBOL(kstat_irqs_cpu);
+

kernel/irq/proc.c

@@ -243,7 +243,11 @@ void init_irq_proc(void)
 	/*
 	 * Create entries for all existing IRQs.
 	 */
-	for_each_irq_desc(irq, desc)
+	for_each_irq_desc(irq, desc) {
+		if (!desc)
+			continue;
+
 		register_irq_proc(irq, desc);
+	}
 }
 

kernel/irq/spurious.c

@@ -91,6 +91,9 @@ static int misrouted_irq(int irq)
 	int i, ok = 0;
 
 	for_each_irq_desc(i, desc) {
+		if (!desc)
+			continue;
+
 		if (!i)
 			 continue;
 
@@ -112,6 +115,8 @@ static void poll_spurious_irqs(unsigned long dummy)
 	for_each_irq_desc(i, desc) {
 		unsigned int status;
 
+		if (!desc)
+			continue;
 		if (!i)
 			 continue;