lib: cpu_rmap: CPU affinity reverse-mapping
authorBen Hutchings <bhutchings@solarflare.com>
Wed, 19 Jan 2011 11:03:25 +0000 (11:03 +0000)
committerDavid S. Miller <davem@davemloft.net>
Mon, 24 Jan 2011 22:51:56 +0000 (14:51 -0800)
When initiating I/O on a multiqueue and multi-IRQ device, we may want
to select a queue for which the response will be handled on the same
or a nearby CPU.  This requires a reverse-map of IRQ affinity.  Add
library functions to support a generic reverse-mapping from CPUs to
objects with affinity and the specific case where the objects are
IRQs.

Signed-off-by: Ben Hutchings <bhutchings@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/linux/cpu_rmap.h [new file with mode: 0644]
lib/Kconfig
lib/Makefile
lib/cpu_rmap.c [new file with mode: 0644]

diff --git a/include/linux/cpu_rmap.h b/include/linux/cpu_rmap.h
new file mode 100644 (file)
index 0000000..473771a
--- /dev/null
@@ -0,0 +1,73 @@
+/*
+ * cpu_rmap.c: CPU affinity reverse-map support
+ * Copyright 2011 Solarflare Communications Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published
+ * by the Free Software Foundation, incorporated herein by reference.
+ */
+
+#include <linux/cpumask.h>
+#include <linux/gfp.h>
+#include <linux/slab.h>
+
+/**
+ * struct cpu_rmap - CPU affinity reverse-map
+ * @size: Number of objects to be reverse-mapped
+ * @used: Number of objects added
+ * @obj: Pointer to array of object pointers
+ * @near: For each CPU, the index and distance to the nearest object,
+ *      based on affinity masks
+ */
+struct cpu_rmap {
+       u16             size, used;
+       void            **obj;
+       struct {
+               u16     index;
+               u16     dist;
+       }               near[0];
+};
+#define CPU_RMAP_DIST_INF 0xffff
+
+extern struct cpu_rmap *alloc_cpu_rmap(unsigned int size, gfp_t flags);
+
+/**
+ * free_cpu_rmap - free CPU affinity reverse-map
+ * @rmap: Reverse-map allocated with alloc_cpu_rmap(), or %NULL
+ */
+static inline void free_cpu_rmap(struct cpu_rmap *rmap)
+{
+       kfree(rmap);
+}
+
+extern int cpu_rmap_add(struct cpu_rmap *rmap, void *obj);
+extern int cpu_rmap_update(struct cpu_rmap *rmap, u16 index,
+                          const struct cpumask *affinity);
+
+static inline u16 cpu_rmap_lookup_index(struct cpu_rmap *rmap, unsigned int cpu)
+{
+       return rmap->near[cpu].index;
+}
+
+static inline void *cpu_rmap_lookup_obj(struct cpu_rmap *rmap, unsigned int cpu)
+{
+       return rmap->obj[rmap->near[cpu].index];
+}
+
+#ifdef CONFIG_GENERIC_HARDIRQS
+
+/**
+ * alloc_irq_cpu_rmap - allocate CPU affinity reverse-map for IRQs
+ * @size: Number of objects to be mapped
+ *
+ * Must be called in process context.
+ */
+static inline struct cpu_rmap *alloc_irq_cpu_rmap(unsigned int size)
+{
+       return alloc_cpu_rmap(size, GFP_KERNEL);
+}
+extern void free_irq_cpu_rmap(struct cpu_rmap *rmap);
+
+extern int irq_cpu_rmap_add(struct cpu_rmap *rmap, int irq);
+
+#endif
index 0ee67e0..8334342 100644 (file)
@@ -201,6 +201,10 @@ config DISABLE_OBSOLETE_CPUMASK_FUNCTIONS
        bool "Disable obsolete cpumask functions" if DEBUG_PER_CPU_MAPS
        depends on EXPERIMENTAL && BROKEN
 
+config CPU_RMAP
+       bool
+       depends on SMP
+
 #
 # Netlink attribute parsing support is select'ed if needed
 #
index cbb774f..b73ba01 100644 (file)
@@ -110,6 +110,8 @@ obj-$(CONFIG_ATOMIC64_SELFTEST) += atomic64_test.o
 
 obj-$(CONFIG_AVERAGE) += average.o
 
+obj-$(CONFIG_CPU_RMAP) += cpu_rmap.o
+
 hostprogs-y    := gen_crc32table
 clean-files    := crc32table.h
 
diff --git a/lib/cpu_rmap.c b/lib/cpu_rmap.c
new file mode 100644 (file)
index 0000000..987acfa
--- /dev/null
@@ -0,0 +1,269 @@
+/*
+ * cpu_rmap.c: CPU affinity reverse-map support
+ * Copyright 2011 Solarflare Communications Inc.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 as published
+ * by the Free Software Foundation, incorporated herein by reference.
+ */
+
+#include <linux/cpu_rmap.h>
+#ifdef CONFIG_GENERIC_HARDIRQS
+#include <linux/interrupt.h>
+#endif
+#include <linux/module.h>
+
+/*
+ * These functions maintain a mapping from CPUs to some ordered set of
+ * objects with CPU affinities.  This can be seen as a reverse-map of
+ * CPU affinity.  However, we do not assume that the object affinities
+ * cover all CPUs in the system.  For those CPUs not directly covered
+ * by object affinities, we attempt to find a nearest object based on
+ * CPU topology.
+ */
+
+/**
+ * alloc_cpu_rmap - allocate CPU affinity reverse-map
+ * @size: Number of objects to be mapped
+ * @flags: Allocation flags e.g. %GFP_KERNEL
+ */
+struct cpu_rmap *alloc_cpu_rmap(unsigned int size, gfp_t flags)
+{
+       struct cpu_rmap *rmap;
+       unsigned int cpu;
+       size_t obj_offset;
+
+       /* This is a silly number of objects, and we use u16 indices. */
+       if (size > 0xffff)
+               return NULL;
+
+       /* Offset of object pointer array from base structure */
+       obj_offset = ALIGN(offsetof(struct cpu_rmap, near[nr_cpu_ids]),
+                          sizeof(void *));
+
+       rmap = kzalloc(obj_offset + size * sizeof(rmap->obj[0]), flags);
+       if (!rmap)
+               return NULL;
+
+       rmap->obj = (void **)((char *)rmap + obj_offset);
+
+       /* Initially assign CPUs to objects on a rota, since we have
+        * no idea where the objects are.  Use infinite distance, so
+        * any object with known distance is preferable.  Include the
+        * CPUs that are not present/online, since we definitely want
+        * any newly-hotplugged CPUs to have some object assigned.
+        */
+       for_each_possible_cpu(cpu) {
+               rmap->near[cpu].index = cpu % size;
+               rmap->near[cpu].dist = CPU_RMAP_DIST_INF;
+       }
+
+       rmap->size = size;
+       return rmap;
+}
+EXPORT_SYMBOL(alloc_cpu_rmap);
+
+/* Reevaluate nearest object for given CPU, comparing with the given
+ * neighbours at the given distance.
+ */
+static bool cpu_rmap_copy_neigh(struct cpu_rmap *rmap, unsigned int cpu,
+                               const struct cpumask *mask, u16 dist)
+{
+       int neigh;
+
+       for_each_cpu(neigh, mask) {
+               if (rmap->near[cpu].dist > dist &&
+                   rmap->near[neigh].dist <= dist) {
+                       rmap->near[cpu].index = rmap->near[neigh].index;
+                       rmap->near[cpu].dist = dist;
+                       return true;
+               }
+       }
+       return false;
+}
+
+#ifdef DEBUG
+static void debug_print_rmap(const struct cpu_rmap *rmap, const char *prefix)
+{
+       unsigned index;
+       unsigned int cpu;
+
+       pr_info("cpu_rmap %p, %s:\n", rmap, prefix);
+
+       for_each_possible_cpu(cpu) {
+               index = rmap->near[cpu].index;
+               pr_info("cpu %d -> obj %u (distance %u)\n",
+                       cpu, index, rmap->near[cpu].dist);
+       }
+}
+#else
+static inline void
+debug_print_rmap(const struct cpu_rmap *rmap, const char *prefix)
+{
+}
+#endif
+
+/**
+ * cpu_rmap_add - add object to a rmap
+ * @rmap: CPU rmap allocated with alloc_cpu_rmap()
+ * @obj: Object to add to rmap
+ *
+ * Return index of object.
+ */
+int cpu_rmap_add(struct cpu_rmap *rmap, void *obj)
+{
+       u16 index;
+
+       BUG_ON(rmap->used >= rmap->size);
+       index = rmap->used++;
+       rmap->obj[index] = obj;
+       return index;
+}
+EXPORT_SYMBOL(cpu_rmap_add);
+
+/**
+ * cpu_rmap_update - update CPU rmap following a change of object affinity
+ * @rmap: CPU rmap to update
+ * @index: Index of object whose affinity changed
+ * @affinity: New CPU affinity of object
+ */
+int cpu_rmap_update(struct cpu_rmap *rmap, u16 index,
+                   const struct cpumask *affinity)
+{
+       cpumask_var_t update_mask;
+       unsigned int cpu;
+
+       if (unlikely(!zalloc_cpumask_var(&update_mask, GFP_KERNEL)))
+               return -ENOMEM;
+
+       /* Invalidate distance for all CPUs for which this used to be
+        * the nearest object.  Mark those CPUs for update.
+        */
+       for_each_online_cpu(cpu) {
+               if (rmap->near[cpu].index == index) {
+                       rmap->near[cpu].dist = CPU_RMAP_DIST_INF;
+                       cpumask_set_cpu(cpu, update_mask);
+               }
+       }
+
+       debug_print_rmap(rmap, "after invalidating old distances");
+
+       /* Set distance to 0 for all CPUs in the new affinity mask.
+        * Mark all CPUs within their NUMA nodes for update.
+        */
+       for_each_cpu(cpu, affinity) {
+               rmap->near[cpu].index = index;
+               rmap->near[cpu].dist = 0;
+               cpumask_or(update_mask, update_mask,
+                          cpumask_of_node(cpu_to_node(cpu)));
+       }
+
+       debug_print_rmap(rmap, "after updating neighbours");
+
+       /* Update distances based on topology */
+       for_each_cpu(cpu, update_mask) {
+               if (cpu_rmap_copy_neigh(rmap, cpu,
+                                       topology_thread_cpumask(cpu), 1))
+                       continue;
+               if (cpu_rmap_copy_neigh(rmap, cpu,
+                                       topology_core_cpumask(cpu), 2))
+                       continue;
+               if (cpu_rmap_copy_neigh(rmap, cpu,
+                                       cpumask_of_node(cpu_to_node(cpu)), 3))
+                       continue;
+               /* We could continue into NUMA node distances, but for now
+                * we give up.
+                */
+       }
+
+       debug_print_rmap(rmap, "after copying neighbours");
+
+       free_cpumask_var(update_mask);
+       return 0;
+}
+EXPORT_SYMBOL(cpu_rmap_update);
+
+#ifdef CONFIG_GENERIC_HARDIRQS
+
+/* Glue between IRQ affinity notifiers and CPU rmaps */
+
+struct irq_glue {
+       struct irq_affinity_notify notify;
+       struct cpu_rmap *rmap;
+       u16 index;
+};
+
+/**
+ * free_irq_cpu_rmap - free a CPU affinity reverse-map used for IRQs
+ * @rmap: Reverse-map allocated with alloc_irq_cpu_map(), or %NULL
+ *
+ * Must be called in process context, before freeing the IRQs, and
+ * without holding any locks required by global workqueue items.
+ */
+void free_irq_cpu_rmap(struct cpu_rmap *rmap)
+{
+       struct irq_glue *glue;
+       u16 index;
+
+       if (!rmap)
+               return;
+
+       for (index = 0; index < rmap->used; index++) {
+               glue = rmap->obj[index];
+               irq_set_affinity_notifier(glue->notify.irq, NULL);
+       }
+       irq_run_affinity_notifiers();
+
+       kfree(rmap);
+}
+EXPORT_SYMBOL(free_irq_cpu_rmap);
+
+static void
+irq_cpu_rmap_notify(struct irq_affinity_notify *notify, const cpumask_t *mask)
+{
+       struct irq_glue *glue =
+               container_of(notify, struct irq_glue, notify);
+       int rc;
+
+       rc = cpu_rmap_update(glue->rmap, glue->index, mask);
+       if (rc)
+               pr_warning("irq_cpu_rmap_notify: update failed: %d\n", rc);
+}
+
+static void irq_cpu_rmap_release(struct kref *ref)
+{
+       struct irq_glue *glue =
+               container_of(ref, struct irq_glue, notify.kref);
+       kfree(glue);
+}
+
+/**
+ * irq_cpu_rmap_add - add an IRQ to a CPU affinity reverse-map
+ * @rmap: The reverse-map
+ * @irq: The IRQ number
+ *
+ * This adds an IRQ affinity notifier that will update the reverse-map
+ * automatically.
+ *
+ * Must be called in process context, after the IRQ is allocated but
+ * before it is bound with request_irq().
+ */
+int irq_cpu_rmap_add(struct cpu_rmap *rmap, int irq)
+{
+       struct irq_glue *glue = kzalloc(sizeof(*glue), GFP_KERNEL);
+       int rc;
+
+       if (!glue)
+               return -ENOMEM;
+       glue->notify.notify = irq_cpu_rmap_notify;
+       glue->notify.release = irq_cpu_rmap_release;
+       glue->rmap = rmap;
+       glue->index = cpu_rmap_add(rmap, glue);
+       rc = irq_set_affinity_notifier(irq, &glue->notify);
+       if (rc)
+               kfree(glue);
+       return rc;
+}
+EXPORT_SYMBOL(irq_cpu_rmap_add);
+
+#endif /* CONFIG_GENERIC_HARDIRQS */