src/patches/suse-2.6.27.31/patches.suse/SoN-09-mm-kmem_estimate_pages.patch

   1 From: Peter Zijlstra <a.p.zijlstra@chello.nl>
   2 Subject: mm: kmem_alloc_estimate()
   3 Patch-mainline: No
   4 References: FATE#303834
   5
   6 Provide a method to get the upper bound on the pages needed to allocate
   7 a given number of objects from a given kmem_cache.
   8
   9 This lays the foundation for a generic reserve framework as presented in
  10 a later patch in this series. This framework needs to convert object demand
  11 (kmalloc() bytes, kmem_cache_alloc() objects) to pages.
  12
  13 Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
  14 Acked-by: Neil Brown <neilb@suse.de>
  15 Acked-by: Suresh Jayaraman <sjayaraman@suse.de>
  16
  17 ---
  18  include/linux/slab.h |    4 ++
  19  mm/slab.c            |   75 +++++++++++++++++++++++++++++++++++++++++++
  20  mm/slob.c            |   67 +++++++++++++++++++++++++++++++++++++++
  21  mm/slub.c            |   87 +++++++++++++++++++++++++++++++++++++++++++++++++++
  22  4 files changed, 233 insertions(+)
  23
  24 --- a/include/linux/slab.h
  25 +++ b/include/linux/slab.h
  26 @@ -65,6 +65,8 @@ void kmem_cache_free(struct kmem_cache *
  27  unsigned int kmem_cache_size(struct kmem_cache *);
  28  const char *kmem_cache_name(struct kmem_cache *);
  29  int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr);
  30 +unsigned kmem_alloc_estimate(struct kmem_cache *cachep,
  31 +                       gfp_t flags, int objects);
  32
  33  /*
  34   * Please use this macro to create slab caches. Simply specify the
  35 @@ -100,6 +102,8 @@ void * __must_check __krealloc(const voi
  36  void * __must_check krealloc(const void *, size_t, gfp_t);
  37  void kfree(const void *);
  38  size_t ksize(const void *);
  39 +unsigned kmalloc_estimate_objs(size_t, gfp_t, int);
  40 +unsigned kmalloc_estimate_bytes(gfp_t, size_t);
  41
  42  /*
  43   * Allocator specific definitions. These are mainly used to establish optimized
  44 --- a/mm/slab.c
  45 +++ b/mm/slab.c
  46 @@ -3846,6 +3846,81 @@ const char *kmem_cache_name(struct kmem_
  47  EXPORT_SYMBOL_GPL(kmem_cache_name);
  48
  49  /*
  50 + * Calculate the upper bound of pages required to sequentially allocate
  51 + * @objects objects from @cachep.
  52 + */
  53 +unsigned kmem_alloc_estimate(struct kmem_cache *cachep,
  54 +               gfp_t flags, int objects)
  55 +{
  56 +       /*
  57 +        * (1) memory for objects,
  58 +        */
  59 +       unsigned nr_slabs = DIV_ROUND_UP(objects, cachep->num);
  60 +       unsigned nr_pages = nr_slabs << cachep->gfporder;
  61 +
  62 +       /*
  63 +        * (2) memory for each per-cpu queue (nr_cpu_ids),
  64 +        * (3) memory for each per-node alien queues (nr_cpu_ids), and
  65 +        * (4) some amount of memory for the slab management structures
  66 +        *
  67 +        * XXX: truely account these
  68 +        */
  69 +       nr_pages += 1 + ilog2(nr_pages);
  70 +
  71 +       return nr_pages;
  72 +}
  73 +
  74 +/*
  75 + * Calculate the upper bound of pages required to sequentially allocate
  76 + * @count objects of @size bytes from kmalloc given @flags.
  77 + */
  78 +unsigned kmalloc_estimate_objs(size_t size, gfp_t flags, int count)
  79 +{
  80 +       struct kmem_cache *s = kmem_find_general_cachep(size, flags);
  81 +       if (!s)
  82 +               return 0;
  83 +
  84 +       return kmem_alloc_estimate(s, flags, count);
  85 +}
  86 +EXPORT_SYMBOL_GPL(kmalloc_estimate_objs);
  87 +
  88 +/*
  89 + * Calculate the upper bound of pages requires to sequentially allocate @bytes
  90 + * from kmalloc in an unspecified number of allocations of nonuniform size.
  91 + */
  92 +unsigned kmalloc_estimate_bytes(gfp_t flags, size_t bytes)
  93 +{
  94 +       unsigned long pages;
  95 +       struct cache_sizes *csizep = malloc_sizes;
  96 +
  97 +       /*
  98 +        * multiply by two, in order to account the worst case slack space
  99 +        * due to the power-of-two allocation sizes.
 100 +        */
 101 +       pages = DIV_ROUND_UP(2 * bytes, PAGE_SIZE);
 102 +
 103 +       /*
 104 +        * add the kmem_cache overhead of each possible kmalloc cache
 105 +        */
 106 +       for (csizep = malloc_sizes; csizep->cs_cachep; csizep++) {
 107 +               struct kmem_cache *s;
 108 +
 109 +#ifdef CONFIG_ZONE_DMA
 110 +               if (unlikely(flags & __GFP_DMA))
 111 +                       s = csizep->cs_dmacachep;
 112 +               else
 113 +#endif
 114 +                       s = csizep->cs_cachep;
 115 +
 116 +               if (s)
 117 +                       pages += kmem_alloc_estimate(s, flags, 0);
 118 +       }
 119 +
 120 +       return pages;
 121 +}
 122 +EXPORT_SYMBOL_GPL(kmalloc_estimate_bytes);
 123 +
 124 +/*
 125   * This initializes kmem_list3 or resizes various caches for all nodes.
 126   */
 127  static int alloc_kmemlist(struct kmem_cache *cachep)
 128 --- a/mm/slob.c
 129 +++ b/mm/slob.c
 130 @@ -661,3 +661,70 @@ void __init kmem_cache_init(void)
 131  {
 132         slob_ready = 1;
 133  }
 134 +
 135 +static __slob_estimate(unsigned size, unsigned align, unsigned objects)
 136 +{
 137 +       unsigned nr_pages;
 138 +
 139 +       size = SLOB_UNIT * SLOB_UNITS(size + align - 1);
 140 +
 141 +       if (size <= PAGE_SIZE) {
 142 +               nr_pages = DIV_ROUND_UP(objects, PAGE_SIZE / size);
 143 +       } else {
 144 +               nr_pages = objects << get_order(size);
 145 +       }
 146 +
 147 +       return nr_pages;
 148 +}
 149 +
 150 +/*
 151 + * Calculate the upper bound of pages required to sequentially allocate
 152 + * @objects objects from @cachep.
 153 + */
 154 +unsigned kmem_alloc_estimate(struct kmem_cache *c, gfp_t flags, int objects)
 155 +{
 156 +       unsigned size = c->size;
 157 +
 158 +       if (c->flags & SLAB_DESTROY_BY_RCU)
 159 +               size += sizeof(struct slob_rcu);
 160 +
 161 +       return __slob_estimate(size, c->align, objects);
 162 +}
 163 +
 164 +/*
 165 + * Calculate the upper bound of pages required to sequentially allocate
 166 + * @count objects of @size bytes from kmalloc given @flags.
 167 + */
 168 +unsigned kmalloc_estimate_objs(size_t size, gfp_t flags, int count)
 169 +{
 170 +       unsigned align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 171 +
 172 +       return __slob_estimate(size, align, count);
 173 +}
 174 +EXPORT_SYMBOL_GPL(kmalloc_estimate_objs);
 175 +
 176 +/*
 177 + * Calculate the upper bound of pages requires to sequentially allocate @bytes
 178 + * from kmalloc in an unspecified number of allocations of nonuniform size.
 179 + */
 180 +unsigned kmalloc_estimate_bytes(gfp_t flags, size_t bytes)
 181 +{
 182 +       unsigned long pages;
 183 +
 184 +       /*
 185 +        * Multiply by two, in order to account the worst case slack space
 186 +        * due to the power-of-two allocation sizes.
 187 +        *
 188 +        * While not true for slob, it cannot do worse than that for sequential
 189 +        * allocations.
 190 +        */
 191 +       pages = DIV_ROUND_UP(2 * bytes, PAGE_SIZE);
 192 +
 193 +       /*
 194 +        * Our power of two series starts at PAGE_SIZE, so add one page.
 195 +        */
 196 +       pages++;
 197 +
 198 +       return pages;
 199 +}
 200 +EXPORT_SYMBOL_GPL(kmalloc_estimate_bytes);
 201 --- a/mm/slub.c
 202 +++ b/mm/slub.c
 203 @@ -2399,6 +2399,42 @@ const char *kmem_cache_name(struct kmem_
 204  }
 205  EXPORT_SYMBOL(kmem_cache_name);
 206
 207 +/*
 208 + * Calculate the upper bound of pages required to sequentially allocate
 209 + * @objects objects from @cachep.
 210 + *
 211 + * We should use s->min_objects because those are the least efficient.
 212 + */
 213 +unsigned kmem_alloc_estimate(struct kmem_cache *s, gfp_t flags, int objects)
 214 +{
 215 +       unsigned long pages;
 216 +       struct kmem_cache_order_objects x;
 217 +
 218 +       if (WARN_ON(!s) || WARN_ON(!oo_objects(s->min)))
 219 +               return 0;
 220 +
 221 +       x = s->min;
 222 +       pages = DIV_ROUND_UP(objects, oo_objects(x)) << oo_order(x);
 223 +
 224 +       /*
 225 +        * Account the possible additional overhead if the slab holds more that
 226 +        * one object. Use s->max_objects because that's the worst case.
 227 +        */
 228 +       x = s->oo;
 229 +       if (oo_objects(x) > 1) {
 230 +               /*
 231 +                * Account the possible additional overhead if per cpu slabs
 232 +                * are currently empty and have to be allocated. This is very
 233 +                * unlikely but a possible scenario immediately after
 234 +                * kmem_cache_shrink.
 235 +                */
 236 +               pages += num_possible_cpus() << oo_order(x);
 237 +       }
 238 +
 239 +       return pages;
 240 +}
 241 +EXPORT_SYMBOL_GPL(kmem_alloc_estimate);
 242 +
 243  static void list_slab_objects(struct kmem_cache *s, struct page *page,
 244                                                         const char *text)
 245  {
 246 @@ -2778,6 +2814,57 @@ void kfree(const void *x)
 247  EXPORT_SYMBOL(kfree);
 248
 249  /*
 250 + * Calculate the upper bound of pages required to sequentially allocate
 251 + * @count objects of @size bytes from kmalloc given @flags.
 252 + */
 253 +unsigned kmalloc_estimate_objs(size_t size, gfp_t flags, int count)
 254 +{
 255 +       struct kmem_cache *s = get_slab(size, flags);
 256 +       if (!s)
 257 +               return 0;
 258 +
 259 +       return kmem_alloc_estimate(s, flags, count);
 260 +
 261 +}
 262 +EXPORT_SYMBOL_GPL(kmalloc_estimate_objs);
 263 +
 264 +/*
 265 + * Calculate the upper bound of pages requires to sequentially allocate @bytes
 266 + * from kmalloc in an unspecified number of allocations of nonuniform size.
 267 + */
 268 +unsigned kmalloc_estimate_bytes(gfp_t flags, size_t bytes)
 269 +{
 270 +       int i;
 271 +       unsigned long pages;
 272 +
 273 +       /*
 274 +        * multiply by two, in order to account the worst case slack space
 275 +        * due to the power-of-two allocation sizes.
 276 +        */
 277 +       pages = DIV_ROUND_UP(2 * bytes, PAGE_SIZE);
 278 +
 279 +       /*
 280 +        * add the kmem_cache overhead of each possible kmalloc cache
 281 +        */
 282 +       for (i = 1; i < PAGE_SHIFT; i++) {
 283 +               struct kmem_cache *s;
 284 +
 285 +#ifdef CONFIG_ZONE_DMA
 286 +               if (unlikely(flags & SLUB_DMA))
 287 +                       s = dma_kmalloc_cache(i, flags);
 288 +               else
 289 +#endif
 290 +                       s = &kmalloc_caches[i];
 291 +
 292 +               if (s)
 293 +                       pages += kmem_alloc_estimate(s, flags, 0);
 294 +       }
 295 +
 296 +       return pages;
 297 +}
 298 +EXPORT_SYMBOL_GPL(kmalloc_estimate_bytes);
 299 +
 300 +/*
 301   * kmem_cache_shrink removes empty slabs from the partial lists and sorts
 302   * the remaining slabs by the number of items in use. The slabs with the
 303   * most items in use come first. New allocations will then fill those up