Patchwork sparc64: fix and optimize irq distribution

login
register
mail settings
Submitter Hong H. Pham
Date June 3, 2009, 4:41 p.m.
Message ID <1244047262-15786-2-git-send-email-hong.pham@windriver.com>
Download mbox | patch
Permalink /patch/28073/
State Accepted
Delegated to: David Miller
Headers show

Comments

Hong H. Pham - June 3, 2009, 4:41 p.m.
irq_choose_cpu() should compare the affinity mask against cpu_online_map
rather than CPU_MASK_ALL, since irq_select_affinity() sets the interrupt's
affinity mask to cpu_online_map "and" CPU_MASK_ALL (which ends up being
just cpu_online_map).  The mask comparison in irq_choose_cpu() will always
fail since the two masks are not the same.  So the CPU chosen is the first CPU
in the intersection of cpu_online_map and CPU_MASK_ALL, which is always CPU0.
That means all interrupts are reassigned to CPU0...

Distributing interrupts to CPUs in a linearly increasing round robin fashion
is not optimal for the UltraSPARC T1/T2.  Also, the irq_rover in
irq_choose_cpu() causes an interrupt to be assigned to a different
processor each time the interrupt is allocated and released.  This may lead
to an unbalanced distribution over time.

A static mapping of interrupts to processors is done to optimize and balance
interrupt distribution.  For the T1/T2, interrupts are spread to different
cores first, and then to strands within a core.

The following is some benchmarks showing the effects of interrupt
distribution on a T2.  The test was done with iperf using a pair of T5220
boxes, each with a 10GBe NIU (XAUI) connected back to back.

  TCP     | Stock       Linear RR IRQ  Optimized IRQ
  Streams | 2.6.30-rc5  Distribution   Distribution
          | GBits/sec   GBits/sec      GBits/sec
  --------+-----------------------------------------
    1       0.839       0.862          0.868
    8       1.16        4.96           5.88
   16       1.15        6.40           8.04
  100       1.09        7.28           8.68

Signed-off-by: Hong H. Pham <hong.pham@windriver.com>
---
 arch/sparc/kernel/Makefile |    1 +
 arch/sparc/kernel/cpumap.c |  431 ++++++++++++++++++++++++++++++++++++++++++++
 arch/sparc/kernel/cpumap.h |   16 ++
 arch/sparc/kernel/irq_64.c |   29 +---
 arch/sparc/kernel/smp_64.c |    2 +
 5 files changed, 454 insertions(+), 25 deletions(-)
 create mode 100644 arch/sparc/kernel/cpumap.c
 create mode 100644 arch/sparc/kernel/cpumap.h
David Miller - June 4, 2009, 4:59 a.m.
From: "Hong H. Pham" <hong.pham@windriver.com>
Date: Wed,  3 Jun 2009 12:41:02 -0400

> irq_choose_cpu() should compare the affinity mask against cpu_online_map
> rather than CPU_MASK_ALL, since irq_select_affinity() sets the interrupt's
> affinity mask to cpu_online_map "and" CPU_MASK_ALL (which ends up being
> just cpu_online_map).  The mask comparison in irq_choose_cpu() will always
> fail since the two masks are not the same.  So the CPU chosen is the first CPU
> in the intersection of cpu_online_map and CPU_MASK_ALL, which is always CPU0.
> That means all interrupts are reassigned to CPU0...
> 
> Distributing interrupts to CPUs in a linearly increasing round robin fashion
> is not optimal for the UltraSPARC T1/T2.  Also, the irq_rover in
> irq_choose_cpu() causes an interrupt to be assigned to a different
> processor each time the interrupt is allocated and released.  This may lead
> to an unbalanced distribution over time.
> 
> A static mapping of interrupts to processors is done to optimize and balance
> interrupt distribution.  For the T1/T2, interrupts are spread to different
> cores first, and then to strands within a core.
> 
> The following is some benchmarks showing the effects of interrupt
> distribution on a T2.  The test was done with iperf using a pair of T5220
> boxes, each with a 10GBe NIU (XAUI) connected back to back.
> 
>   TCP     | Stock       Linear RR IRQ  Optimized IRQ
>   Streams | 2.6.30-rc5  Distribution   Distribution
>           | GBits/sec   GBits/sec      GBits/sec
>   --------+-----------------------------------------
>     1       0.839       0.862          0.868
>     8       1.16        4.96           5.88
>    16       1.15        6.40           8.04
>   100       1.09        7.28           8.68
> 
> Signed-off-by: Hong H. Pham <hong.pham@windriver.com>

Applied to sparc-next-2.6, thanks a lot!

WRT. the NIU interrupt bug, I'm still thinking about how to make
further progress on that one.  I'll post a followup soon.
--
To unsubscribe from this list: send the line "unsubscribe sparclinux" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch

diff --git a/arch/sparc/kernel/Makefile b/arch/sparc/kernel/Makefile
index 54742e5..47029c6 100644
--- a/arch/sparc/kernel/Makefile
+++ b/arch/sparc/kernel/Makefile
@@ -54,6 +54,7 @@  obj-$(CONFIG_SPARC64)   += sstate.o
 obj-$(CONFIG_SPARC64)   += mdesc.o
 obj-$(CONFIG_SPARC64)	+= pcr.o
 obj-$(CONFIG_SPARC64)	+= nmi.o
+obj-$(CONFIG_SPARC64_SMP) += cpumap.o
 
 # sparc32 do not use GENERIC_HARDIRQS but uses the generic devres implementation
 obj-$(CONFIG_SPARC32)     += devres.o
diff --git a/arch/sparc/kernel/cpumap.c b/arch/sparc/kernel/cpumap.c
new file mode 100644
index 0000000..7430ed0
--- /dev/null
+++ b/arch/sparc/kernel/cpumap.c
@@ -0,0 +1,431 @@ 
+/* cpumap.c: used for optimizing CPU assignment
+ *
+ * Copyright (C) 2009 Hong H. Pham <hong.pham@windriver.com>
+ */
+
+#include <linux/module.h>
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/cpumask.h>
+#include <linux/spinlock.h>
+#include <asm/cpudata.h>
+#include "cpumap.h"
+
+
+enum {
+	CPUINFO_LVL_ROOT = 0,
+	CPUINFO_LVL_NODE,
+	CPUINFO_LVL_CORE,
+	CPUINFO_LVL_PROC,
+	CPUINFO_LVL_MAX,
+};
+
+enum {
+	ROVER_NO_OP              = 0,
+	/* Increment rover every time level is visited */
+	ROVER_INC_ON_VISIT       = 1 << 0,
+	/* Increment parent's rover every time rover wraps around */
+	ROVER_INC_PARENT_ON_LOOP = 1 << 1,
+};
+
+struct cpuinfo_node {
+	int id;
+	int level;
+	int num_cpus;    /* Number of CPUs in this hierarchy */
+	int parent_index;
+	int child_start; /* Array index of the first child node */
+	int child_end;   /* Array index of the last child node */
+	int rover;       /* Child node iterator */
+};
+
+struct cpuinfo_level {
+	int start_index; /* Index of first node of a level in a cpuinfo tree */
+	int end_index;   /* Index of last node of a level in a cpuinfo tree */
+	int num_nodes;   /* Number of nodes in a level in a cpuinfo tree */
+};
+
+struct cpuinfo_tree {
+	int total_nodes;
+
+	/* Offsets into nodes[] for each level of the tree */
+	struct cpuinfo_level level[CPUINFO_LVL_MAX];
+	struct cpuinfo_node  nodes[0];
+};
+
+
+static struct cpuinfo_tree *cpuinfo_tree;
+
+static u16 cpu_distribution_map[NR_CPUS];
+static DEFINE_SPINLOCK(cpu_map_lock);
+
+
+/* Niagara optimized cpuinfo tree traversal. */
+static const int niagara_iterate_method[] = {
+	[CPUINFO_LVL_ROOT] = ROVER_NO_OP,
+
+	/* Strands (or virtual CPUs) within a core may not run concurrently
+	 * on the Niagara, as instruction pipeline(s) are shared.  Distribute
+	 * work to strands in different cores first for better concurrency.
+	 * Go to next NUMA node when all cores are used.
+	 */
+	[CPUINFO_LVL_NODE] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
+
+	/* Strands are grouped together by proc_id in cpuinfo_sparc, i.e.
+	 * a proc_id represents an instruction pipeline.  Distribute work to
+	 * strands in different proc_id groups if the core has multiple
+	 * instruction pipelines (e.g. the Niagara 2/2+ has two).
+	 */
+	[CPUINFO_LVL_CORE] = ROVER_INC_ON_VISIT,
+
+	/* Pick the next strand in the proc_id group. */
+	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT,
+};
+
+/* Generic cpuinfo tree traversal.  Distribute work round robin across NUMA
+ * nodes.
+ */
+static const int generic_iterate_method[] = {
+	[CPUINFO_LVL_ROOT] = ROVER_INC_ON_VISIT,
+	[CPUINFO_LVL_NODE] = ROVER_NO_OP,
+	[CPUINFO_LVL_CORE] = ROVER_INC_PARENT_ON_LOOP,
+	[CPUINFO_LVL_PROC] = ROVER_INC_ON_VISIT|ROVER_INC_PARENT_ON_LOOP,
+};
+
+
+static int cpuinfo_id(int cpu, int level)
+{
+	int id;
+
+	switch (level) {
+	case CPUINFO_LVL_ROOT:
+		id = 0;
+		break;
+	case CPUINFO_LVL_NODE:
+		id = cpu_to_node(cpu);
+		break;
+	case CPUINFO_LVL_CORE:
+		id = cpu_data(cpu).core_id;
+		break;
+	case CPUINFO_LVL_PROC:
+		id = cpu_data(cpu).proc_id;
+		break;
+	default:
+		id = -EINVAL;
+	}
+	return id;
+}
+
+/*
+ * Enumerate the CPU information in __cpu_data to determine the start index,
+ * end index, and number of nodes for each level in the cpuinfo tree.  The
+ * total number of cpuinfo nodes required to build the tree is returned.
+ */
+static int enumerate_cpuinfo_nodes(struct cpuinfo_level *tree_level)
+{
+	int prev_id[CPUINFO_LVL_MAX];
+	int i, n, num_nodes;
+
+	for (i = CPUINFO_LVL_ROOT; i < CPUINFO_LVL_MAX; i++) {
+		struct cpuinfo_level *lv = &tree_level[i];
+
+		prev_id[i] = -1;
+		lv->start_index = lv->end_index = lv->num_nodes = 0;
+	}
+
+	num_nodes = 1; /* Include the root node */
+
+	for (i = 0; i < num_possible_cpus(); i++) {
+		if (!cpu_online(i))
+			continue;
+
+		n = cpuinfo_id(i, CPUINFO_LVL_NODE);
+		if (n > prev_id[CPUINFO_LVL_NODE]) {
+			tree_level[CPUINFO_LVL_NODE].num_nodes++;
+			prev_id[CPUINFO_LVL_NODE] = n;
+			num_nodes++;
+		}
+		n = cpuinfo_id(i, CPUINFO_LVL_CORE);
+		if (n > prev_id[CPUINFO_LVL_CORE]) {
+			tree_level[CPUINFO_LVL_CORE].num_nodes++;
+			prev_id[CPUINFO_LVL_CORE] = n;
+			num_nodes++;
+		}
+		n = cpuinfo_id(i, CPUINFO_LVL_PROC);
+		if (n > prev_id[CPUINFO_LVL_PROC]) {
+			tree_level[CPUINFO_LVL_PROC].num_nodes++;
+			prev_id[CPUINFO_LVL_PROC] = n;
+			num_nodes++;
+		}
+	}
+
+	tree_level[CPUINFO_LVL_ROOT].num_nodes = 1;
+
+	n = tree_level[CPUINFO_LVL_NODE].num_nodes;
+	tree_level[CPUINFO_LVL_NODE].start_index = 1;
+	tree_level[CPUINFO_LVL_NODE].end_index   = n;
+
+	n++;
+	tree_level[CPUINFO_LVL_CORE].start_index = n;
+	n += tree_level[CPUINFO_LVL_CORE].num_nodes;
+	tree_level[CPUINFO_LVL_CORE].end_index   = n - 1;
+
+	tree_level[CPUINFO_LVL_PROC].start_index = n;
+	n += tree_level[CPUINFO_LVL_PROC].num_nodes;
+	tree_level[CPUINFO_LVL_PROC].end_index   = n - 1;
+
+	return num_nodes;
+}
+
+/* Build a tree representation of the CPU hierarchy using the per CPU
+ * information in __cpu_data.  Entries in __cpu_data[0..NR_CPUS] are
+ * assumed to be sorted in ascending order based on node, core_id, and
+ * proc_id (in order of significance).
+ */
+static struct cpuinfo_tree *build_cpuinfo_tree(void)
+{
+	struct cpuinfo_tree *new_tree;
+	struct cpuinfo_node *node;
+	struct cpuinfo_level tmp_level[CPUINFO_LVL_MAX];
+	int num_cpus[CPUINFO_LVL_MAX];
+	int level_rover[CPUINFO_LVL_MAX];
+	int prev_id[CPUINFO_LVL_MAX];
+	int n, id, cpu, prev_cpu, last_cpu, level;
+
+	n = enumerate_cpuinfo_nodes(tmp_level);
+
+	new_tree = kzalloc(sizeof(struct cpuinfo_tree) +
+	                   (sizeof(struct cpuinfo_node) * n), GFP_ATOMIC);
+	if (!new_tree)
+		return NULL;
+
+	new_tree->total_nodes = n;
+	memcpy(&new_tree->level, tmp_level, sizeof(tmp_level));
+
+	prev_cpu = cpu = first_cpu(cpu_online_map);
+
+	/* Initialize all levels in the tree with the first CPU */
+	for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT; level--) {
+		n = new_tree->level[level].start_index;
+
+		level_rover[level] = n;
+		node = &new_tree->nodes[n];
+
+		id = cpuinfo_id(cpu, level);
+		if (unlikely(id < 0)) {
+			kfree(new_tree);
+			return NULL;
+		}
+		node->id = id;
+		node->level = level;
+		node->num_cpus = 1;
+
+		node->parent_index = (level > CPUINFO_LVL_ROOT)
+		    ? new_tree->level[level - 1].start_index : -1;
+
+		node->child_start = node->child_end = node->rover =
+		    (level == CPUINFO_LVL_PROC)
+		    ? cpu : new_tree->level[level + 1].start_index;
+
+		prev_id[level] = node->id;
+		num_cpus[level] = 1;
+	}
+
+	for (last_cpu = (num_possible_cpus() - 1); last_cpu >= 0; last_cpu--) {
+		if (cpu_online(last_cpu))
+			break;
+	}
+
+	while (++cpu <= last_cpu) {
+		if (!cpu_online(cpu))
+			continue;
+
+		for (level = CPUINFO_LVL_PROC; level >= CPUINFO_LVL_ROOT;
+		     level--) {
+			id = cpuinfo_id(cpu, level);
+			if (unlikely(id < 0)) {
+				kfree(new_tree);
+				return NULL;
+			}
+
+			if ((id != prev_id[level]) || (cpu == last_cpu)) {
+				prev_id[level] = id;
+				node = &new_tree->nodes[level_rover[level]];
+				node->num_cpus = num_cpus[level];
+				num_cpus[level] = 1;
+
+				if (cpu == last_cpu)
+					node->num_cpus++;
+
+				/* Connect tree node to parent */
+				if (level == CPUINFO_LVL_ROOT)
+					node->parent_index = -1;
+				else
+					node->parent_index =
+					    level_rover[level - 1];
+
+				if (level == CPUINFO_LVL_PROC) {
+					node->child_end =
+					    (cpu == last_cpu) ? cpu : prev_cpu;
+				} else {
+					node->child_end =
+					    level_rover[level + 1] - 1;
+				}
+
+				/* Initialize the next node in the same level */
+				n = ++level_rover[level];
+				if (n <= new_tree->level[level].end_index) {
+					node = &new_tree->nodes[n];
+					node->id = id;
+					node->level = level;
+
+					/* Connect node to child */
+					node->child_start = node->child_end =
+					node->rover =
+					    (level == CPUINFO_LVL_PROC)
+					    ? cpu : level_rover[level + 1];
+				}
+			} else
+				num_cpus[level]++;
+		}
+		prev_cpu = cpu;
+	}
+
+	return new_tree;
+}
+
+static void increment_rover(struct cpuinfo_tree *t, int node_index,
+                            int root_index, const int *rover_inc_table)
+{
+	struct cpuinfo_node *node = &t->nodes[node_index];
+	int top_level, level;
+
+	top_level = t->nodes[root_index].level;
+	for (level = node->level; level >= top_level; level--) {
+		node->rover++;
+		if (node->rover <= node->child_end)
+			return;
+
+		node->rover = node->child_start;
+		/* If parent's rover does not need to be adjusted, stop here. */
+		if ((level == top_level) ||
+		    !(rover_inc_table[level] & ROVER_INC_PARENT_ON_LOOP))
+			return;
+
+		node = &t->nodes[node->parent_index];
+	}
+}
+
+static int iterate_cpu(struct cpuinfo_tree *t, unsigned int root_index)
+{
+	const int *rover_inc_table;
+	int level, new_index, index = root_index;
+
+	switch (sun4v_chip_type) {
+	case SUN4V_CHIP_NIAGARA1:
+	case SUN4V_CHIP_NIAGARA2:
+		rover_inc_table = niagara_iterate_method;
+		break;
+	default:
+		rover_inc_table = generic_iterate_method;
+	}
+
+	for (level = t->nodes[root_index].level; level < CPUINFO_LVL_MAX;
+	     level++) {
+		new_index = t->nodes[index].rover;
+		if (rover_inc_table[level] & ROVER_INC_ON_VISIT)
+			increment_rover(t, index, root_index, rover_inc_table);
+
+		index = new_index;
+	}
+	return index;
+}
+
+static void _cpu_map_rebuild(void)
+{
+	int i;
+
+	if (cpuinfo_tree) {
+		kfree(cpuinfo_tree);
+		cpuinfo_tree = NULL;
+	}
+
+	cpuinfo_tree = build_cpuinfo_tree();
+	if (!cpuinfo_tree)
+		return;
+
+	/* Build CPU distribution map that spans all online CPUs.  No need
+	 * to check if the CPU is online, as that is done when the cpuinfo
+	 * tree is being built.
+	 */
+	for (i = 0; i < cpuinfo_tree->nodes[0].num_cpus; i++)
+		cpu_distribution_map[i] = iterate_cpu(cpuinfo_tree, 0);
+}
+
+/* Fallback if the cpuinfo tree could not be built.  CPU mapping is linear
+ * round robin.
+ */
+static int simple_map_to_cpu(unsigned int index)
+{
+	int i, end, cpu_rover;
+
+	cpu_rover = 0;
+	end = index % num_online_cpus();
+	for (i = 0; i < num_possible_cpus(); i++) {
+		if (cpu_online(cpu_rover)) {
+			if (cpu_rover >= end)
+				return cpu_rover;
+
+			cpu_rover++;
+		}
+	}
+
+	/* Impossible, since num_online_cpus() <= num_possible_cpus() */
+	return first_cpu(cpu_online_map);
+}
+
+static int _map_to_cpu(unsigned int index)
+{
+	struct cpuinfo_node *root_node;
+
+	if (unlikely(!cpuinfo_tree)) {
+		_cpu_map_rebuild();
+		if (!cpuinfo_tree)
+			return simple_map_to_cpu(index);
+	}
+
+	root_node = &cpuinfo_tree->nodes[0];
+#ifdef CONFIG_HOTPLUG_CPU
+	if (unlikely(root_node->num_cpus != num_online_cpus())) {
+		_cpu_map_rebuild();
+		if (!cpuinfo_tree)
+			return simple_map_to_cpu(index);
+	}
+#endif
+	return cpu_distribution_map[index % root_node->num_cpus];
+}
+
+int map_to_cpu(unsigned int index)
+{
+	int mapped_cpu;
+	unsigned long flag;
+
+	spin_lock_irqsave(&cpu_map_lock, flag);
+	mapped_cpu = _map_to_cpu(index);
+
+#ifdef CONFIG_HOTPLUG_CPU
+	while (unlikely(!cpu_online(mapped_cpu)))
+		mapped_cpu = _map_to_cpu(index);
+#endif
+	spin_unlock_irqrestore(&cpu_map_lock, flag);
+	return mapped_cpu;
+}
+EXPORT_SYMBOL(map_to_cpu);
+
+void cpu_map_rebuild(void)
+{
+	unsigned long flag;
+
+	spin_lock_irqsave(&cpu_map_lock, flag);
+	_cpu_map_rebuild();
+	spin_unlock_irqrestore(&cpu_map_lock, flag);
+}
diff --git a/arch/sparc/kernel/cpumap.h b/arch/sparc/kernel/cpumap.h
new file mode 100644
index 0000000..e639880
--- /dev/null
+++ b/arch/sparc/kernel/cpumap.h
@@ -0,0 +1,16 @@ 
+#ifndef _CPUMAP_H
+#define _CPUMAP_H
+
+#ifdef CONFIG_SMP
+extern void cpu_map_rebuild(void);
+extern int  map_to_cpu(unsigned int index);
+#define cpu_map_init() cpu_map_rebuild()
+#else
+#define cpu_map_init() do {} while (0)
+static inline int map_to_cpu(unsigned int index)
+{
+	return raw_smp_processor_id();
+}
+#endif
+
+#endif
diff --git a/arch/sparc/kernel/irq_64.c b/arch/sparc/kernel/irq_64.c
index 5deabe9..b68386d 100644
--- a/arch/sparc/kernel/irq_64.c
+++ b/arch/sparc/kernel/irq_64.c
@@ -45,6 +45,7 @@ 
 #include <asm/cacheflush.h>
 
 #include "entry.h"
+#include "cpumap.h"
 
 #define NUM_IVECS	(IMAP_INR + 1)
 
@@ -256,35 +257,13 @@  static int irq_choose_cpu(unsigned int virt_irq)
 	int cpuid;
 
 	cpumask_copy(&mask, irq_desc[virt_irq].affinity);
-	if (cpus_equal(mask, CPU_MASK_ALL)) {
-		static int irq_rover;
-		static DEFINE_SPINLOCK(irq_rover_lock);
-		unsigned long flags;
-
-		/* Round-robin distribution... */
-	do_round_robin:
-		spin_lock_irqsave(&irq_rover_lock, flags);
-
-		while (!cpu_online(irq_rover)) {
-			if (++irq_rover >= nr_cpu_ids)
-				irq_rover = 0;
-		}
-		cpuid = irq_rover;
-		do {
-			if (++irq_rover >= nr_cpu_ids)
-				irq_rover = 0;
-		} while (!cpu_online(irq_rover));
-
-		spin_unlock_irqrestore(&irq_rover_lock, flags);
+	if (cpus_equal(mask, cpu_online_map)) {
+		cpuid = map_to_cpu(virt_irq);
 	} else {
 		cpumask_t tmp;
 
 		cpus_and(tmp, cpu_online_map, mask);
-
-		if (cpus_empty(tmp))
-			goto do_round_robin;
-
-		cpuid = first_cpu(tmp);
+		cpuid = cpus_empty(tmp) ? map_to_cpu(virt_irq) : first_cpu(tmp);
 	}
 
 	return cpuid;
diff --git a/arch/sparc/kernel/smp_64.c b/arch/sparc/kernel/smp_64.c
index f7642e5..da89ce3 100644
--- a/arch/sparc/kernel/smp_64.c
+++ b/arch/sparc/kernel/smp_64.c
@@ -1315,6 +1315,8 @@  int __cpu_disable(void)
 	cpu_clear(cpu, cpu_online_map);
 	ipi_call_unlock();
 
+	cpu_map_rebuild();
+
 	return 0;
 }