LXR linux/mm/allocpercpu.c

   1/*
   2 * linux/mm/allocpercpu.c
   3 *
   4 * Separated from slab.c August 11, 2006 Christoph Lameter
   5 */
   6#include <linux/mm.h>
   7#include <linux/module.h>
   8
   9#ifndef cache_line_size
  10#define cache_line_size()       L1_CACHE_BYTES
  11#endif
  12
  13/**
  14 * percpu_depopulate - depopulate per-cpu data for given cpu
  15 * @__pdata: per-cpu data to depopulate
  16 * @cpu: depopulate per-cpu data for this cpu
  17 *
  18 * Depopulating per-cpu data for a cpu going offline would be a typical
  19 * use case. You need to register a cpu hotplug handler for that purpose.
  20 */
  21static void percpu_depopulate(void *__pdata, int cpu)
  22{
  23        struct percpu_data *pdata = __percpu_disguise(__pdata);
  24
  25        kfree(pdata->ptrs[cpu]);
  26        pdata->ptrs[cpu] = NULL;
  27}
  28
  29/**
  30 * percpu_depopulate_mask - depopulate per-cpu data for some cpu's
  31 * @__pdata: per-cpu data to depopulate
  32 * @mask: depopulate per-cpu data for cpu's selected through mask bits
  33 */
  34static void __percpu_depopulate_mask(void *__pdata, const cpumask_t *mask)
  35{
  36        int cpu;
  37        for_each_cpu_mask_nr(cpu, *mask)
  38                percpu_depopulate(__pdata, cpu);
  39}
  40
  41#define percpu_depopulate_mask(__pdata, mask) \
  42        __percpu_depopulate_mask((__pdata), &(mask))
  43
  44/**
  45 * percpu_populate - populate per-cpu data for given cpu
  46 * @__pdata: per-cpu data to populate further
  47 * @size: size of per-cpu object
  48 * @gfp: may sleep or not etc.
  49 * @cpu: populate per-data for this cpu
  50 *
  51 * Populating per-cpu data for a cpu coming online would be a typical
  52 * use case. You need to register a cpu hotplug handler for that purpose.
  53 * Per-cpu object is populated with zeroed buffer.
  54 */
  55static void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu)
  56{
  57        struct percpu_data *pdata = __percpu_disguise(__pdata);
  58        int node = cpu_to_node(cpu);
  59
  60        /*
  61         * We should make sure each CPU gets private memory.
  62         */
  63        size = roundup(size, cache_line_size());
  64
  65        BUG_ON(pdata->ptrs[cpu]);
  66        if (node_online(node))
  67                pdata->ptrs[cpu] = kmalloc_node(size, gfp|__GFP_ZERO, node);
  68        else
  69                pdata->ptrs[cpu] = kzalloc(size, gfp);
  70        return pdata->ptrs[cpu];
  71}
  72
  73/**
  74 * percpu_populate_mask - populate per-cpu data for more cpu's
  75 * @__pdata: per-cpu data to populate further
  76 * @size: size of per-cpu object
  77 * @gfp: may sleep or not etc.
  78 * @mask: populate per-cpu data for cpu's selected through mask bits
  79 *
  80 * Per-cpu objects are populated with zeroed buffers.
  81 */
  82static int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp,
  83                                  cpumask_t *mask)
  84{
  85        cpumask_t populated;
  86        int cpu;
  87
  88        cpus_clear(populated);
  89        for_each_cpu_mask_nr(cpu, *mask)
  90                if (unlikely(!percpu_populate(__pdata, size, gfp, cpu))) {
  91                        __percpu_depopulate_mask(__pdata, &populated);
  92                        return -ENOMEM;
  93                } else
  94                        cpu_set(cpu, populated);
  95        return 0;
  96}
  97
  98#define percpu_populate_mask(__pdata, size, gfp, mask) \
  99        __percpu_populate_mask((__pdata), (size), (gfp), &(mask))
 100
 101/**
 102 * alloc_percpu - initial setup of per-cpu data
 103 * @size: size of per-cpu object
 104 * @align: alignment
 105 *
 106 * Allocate dynamic percpu area.  Percpu objects are populated with
 107 * zeroed buffers.
 108 */
 109void *__alloc_percpu(size_t size, size_t align)
 110{
 111        /*
 112         * We allocate whole cache lines to avoid false sharing
 113         */
 114        size_t sz = roundup(nr_cpu_ids * sizeof(void *), cache_line_size());
 115        void *pdata = kzalloc(sz, GFP_KERNEL);
 116        void *__pdata = __percpu_disguise(pdata);
 117
 118        /*
 119         * Can't easily make larger alignment work with kmalloc.  WARN
 120         * on it.  Larger alignment should only be used for module
 121         * percpu sections on SMP for which this path isn't used.
 122         */
 123        WARN_ON_ONCE(align > SMP_CACHE_BYTES);
 124
 125        if (unlikely(!pdata))
 126                return NULL;
 127        if (likely(!__percpu_populate_mask(__pdata, size, GFP_KERNEL,
 128                                           &cpu_possible_map)))
 129                return __pdata;
 130        kfree(pdata);
 131        return NULL;
 132}
 133EXPORT_SYMBOL_GPL(__alloc_percpu);
 134
 135/**
 136 * free_percpu - final cleanup of per-cpu data
 137 * @__pdata: object to clean up
 138 *
 139 * We simply clean up any per-cpu object left. No need for the client to
 140 * track and specify through a bis mask which per-cpu objects are to free.
 141 */
 142void free_percpu(void *__pdata)
 143{
 144        if (unlikely(!__pdata))
 145                return;
 146        __percpu_depopulate_mask(__pdata, cpu_possible_mask);
 147        kfree(__percpu_disguise(__pdata));
 148}
 149EXPORT_SYMBOL_GPL(free_percpu);
 150