mm/percpu-km.c

   1 /*
   2  * mm/percpu-km.c - kernel memory based chunk allocation
   3  *
   4  * Copyright (C) 2010           SUSE Linux Products GmbH
   5  * Copyright (C) 2010           Tejun Heo <tj@kernel.org>
   6  *
   7  * This file is released under the GPLv2.
   8  *
   9  * Chunks are allocated as a contiguous kernel memory using gfp
  10  * allocation.  This is to be used on nommu architectures.
  11  *
  12  * To use percpu-km,
  13  *
  14  * - define CONFIG_NEED_PER_CPU_KM from the arch Kconfig.
  15  *
  16  * - CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK must not be defined.  It's
  17  *   not compatible with PER_CPU_KM.  EMBED_FIRST_CHUNK should work
  18  *   fine.
  19  *
  20  * - NUMA is not supported.  When setting up the first chunk,
  21  *   @cpu_distance_fn should be NULL or report all CPUs to be nearer
  22  *   than or at LOCAL_DISTANCE.
  23  *
  24  * - It's best if the chunk size is power of two multiple of
  25  *   PAGE_SIZE.  Because each chunk is allocated as a contiguous
  26  *   kernel memory block using alloc_pages(), memory will be wasted if
  27  *   chunk size is not aligned.  percpu-km code will whine about it.
  28  */
  29
  30 #if defined(CONFIG_SMP) && defined(CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK)
  31 #error "contiguous percpu allocation is incompatible with paged first chunk"
  32 #endif
  33
  34 #include <linux/log2.h>
  35
  36 static int pcpu_populate_chunk(struct pcpu_chunk *chunk,
  37                                int page_start, int page_end, gfp_t gfp)
  38 {
  39         return 0;
  40 }
  41
  42 static void pcpu_depopulate_chunk(struct pcpu_chunk *chunk,
  43                                   int page_start, int page_end)
  44 {
  45         /* nada */
  46 }
  47
  48 static struct pcpu_chunk *pcpu_create_chunk(gfp_t gfp)
  49 {
  50         const int nr_pages = pcpu_group_sizes[0] >> PAGE_SHIFT;
  51         struct pcpu_chunk *chunk;
  52         struct page *pages;
  53         unsigned long flags;
  54         int i;
  55
  56         chunk = pcpu_alloc_chunk(gfp);
  57         if (!chunk)
  58                 return NULL;
  59
  60         pages = alloc_pages(gfp, order_base_2(nr_pages));
  61         if (!pages) {
  62                 pcpu_free_chunk(chunk);
  63                 return NULL;
  64         }
  65
  66         for (i = 0; i < nr_pages; i++)
  67                 pcpu_set_page_chunk(nth_page(pages, i), chunk);
  68
  69         chunk->data = pages;
  70         chunk->base_addr = page_address(pages) - pcpu_group_offsets[0];
  71
  72         spin_lock_irqsave(&pcpu_lock, flags);
  73         pcpu_chunk_populated(chunk, 0, nr_pages, false);
  74         spin_unlock_irqrestore(&pcpu_lock, flags);
  75
  76         pcpu_stats_chunk_alloc();
  77         trace_percpu_create_chunk(chunk->base_addr);
  78
  79         return chunk;
  80 }
  81
  82 static void pcpu_destroy_chunk(struct pcpu_chunk *chunk)
  83 {
  84         const int nr_pages = pcpu_group_sizes[0] >> PAGE_SHIFT;
  85
  86         if (!chunk)
  87                 return;
  88
  89         pcpu_stats_chunk_dealloc();
  90         trace_percpu_destroy_chunk(chunk->base_addr);
  91
  92         if (chunk->data)
  93                 __free_pages(chunk->data, order_base_2(nr_pages));
  94         pcpu_free_chunk(chunk);
  95 }
  96
  97 static struct page *pcpu_addr_to_page(void *addr)
  98 {
  99         return virt_to_page(addr);
 100 }
 101
 102 static int __init pcpu_verify_alloc_info(const struct pcpu_alloc_info *ai)
 103 {
 104         size_t nr_pages, alloc_pages;
 105
 106         /* all units must be in a single group */
 107         if (ai->nr_groups != 1) {
 108                 pr_crit("can't handle more than one group\n");
 109                 return -EINVAL;
 110         }
 111
 112         nr_pages = (ai->groups[0].nr_units * ai->unit_size) >> PAGE_SHIFT;
 113         alloc_pages = roundup_pow_of_two(nr_pages);
 114
 115         if (alloc_pages > nr_pages)
 116                 pr_warn("wasting %zu pages per chunk\n",
 117                         alloc_pages - nr_pages);
 118
 119         return 0;
 120 }