kernel/2.6.29.6-aldebaran-rt/net/core/flow.c

   1 /* flow.c: Generic flow cache.
   2  *
   3  * Copyright (C) 2003 Alexey N. Kuznetsov (kuznet@ms2.inr.ac.ru)
   4  * Copyright (C) 2003 David S. Miller (davem@redhat.com)
   5  */
   6
   7 #include <linux/kernel.h>
   8 #include <linux/module.h>
   9 #include <linux/list.h>
  10 #include <linux/jhash.h>
  11 #include <linux/interrupt.h>
  12 #include <linux/mm.h>
  13 #include <linux/random.h>
  14 #include <linux/init.h>
  15 #include <linux/slab.h>
  16 #include <linux/smp.h>
  17 #include <linux/completion.h>
  18 #include <linux/percpu.h>
  19 #include <linux/bitops.h>
  20 #include <linux/notifier.h>
  21 #include <linux/cpu.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/mutex.h>
  24 #include <net/flow.h>
  25 #include <asm/atomic.h>
  26 #include <linux/security.h>
  27
  28 struct flow_cache_entry {
  29         struct flow_cache_entry *next;
  30         u16                     family;
  31         u8                      dir;
  32         u32                     genid;
  33         struct flowi            key;
  34         void                    *object;
  35         atomic_t                *object_ref;
  36 };
  37
  38 atomic_t flow_cache_genid = ATOMIC_INIT(0);
  39
  40 static u32 flow_hash_shift;
  41 #define flow_hash_size  (1 << flow_hash_shift)
  42
  43 static DEFINE_PER_CPU_LOCKED(struct flow_cache_entry **, flow_tables);
  44
  45 #define flow_table(cpu) (per_cpu_var_locked(flow_tables, cpu))
  46
  47 static struct kmem_cache *flow_cachep __read_mostly;
  48
  49 static int flow_lwm, flow_hwm;
  50
  51 struct flow_percpu_info {
  52         int hash_rnd_recalc;
  53         u32 hash_rnd;
  54         int count;
  55 };
  56 static DEFINE_PER_CPU(struct flow_percpu_info, flow_hash_info) = { 0 };
  57
  58 #define flow_hash_rnd_recalc(cpu) \
  59         (per_cpu(flow_hash_info, cpu).hash_rnd_recalc)
  60 #define flow_hash_rnd(cpu) \
  61         (per_cpu(flow_hash_info, cpu).hash_rnd)
  62 #define flow_count(cpu) \
  63         (per_cpu(flow_hash_info, cpu).count)
  64
  65 static struct timer_list flow_hash_rnd_timer;
  66
  67 #define FLOW_HASH_RND_PERIOD    (10 * 60 * HZ)
  68
  69 struct flow_flush_info {
  70         atomic_t cpuleft;
  71         struct completion completion;
  72 };
  73 static DEFINE_PER_CPU(struct tasklet_struct, flow_flush_tasklets) = { NULL };
  74
  75 #define flow_flush_tasklet(cpu) (&per_cpu(flow_flush_tasklets, cpu))
  76
  77 static void flow_cache_new_hashrnd(unsigned long arg)
  78 {
  79         int i;
  80
  81         for_each_possible_cpu(i)
  82                 flow_hash_rnd_recalc(i) = 1;
  83
  84         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
  85         add_timer(&flow_hash_rnd_timer);
  86 }
  87
  88 static void flow_entry_kill(int cpu, struct flow_cache_entry *fle)
  89 {
  90         if (fle->object)
  91                 atomic_dec(fle->object_ref);
  92         kmem_cache_free(flow_cachep, fle);
  93         flow_count(cpu)--;
  94 }
  95
  96 static void __flow_cache_shrink(int cpu, int shrink_to)
  97 {
  98         struct flow_cache_entry *fle, **flp;
  99         int i;
 100
 101         for (i = 0; i < flow_hash_size; i++) {
 102                 int k = 0;
 103
 104                 flp = &flow_table(cpu)[i];
 105                 while ((fle = *flp) != NULL && k < shrink_to) {
 106                         k++;
 107                         flp = &fle->next;
 108                 }
 109                 while ((fle = *flp) != NULL) {
 110                         *flp = fle->next;
 111                         flow_entry_kill(cpu, fle);
 112                 }
 113         }
 114 }
 115
 116 static void flow_cache_shrink(int cpu)
 117 {
 118         int shrink_to = flow_lwm / flow_hash_size;
 119
 120         __flow_cache_shrink(cpu, shrink_to);
 121 }
 122
 123 static void flow_new_hash_rnd(int cpu)
 124 {
 125         get_random_bytes(&flow_hash_rnd(cpu), sizeof(u32));
 126         flow_hash_rnd_recalc(cpu) = 0;
 127
 128         __flow_cache_shrink(cpu, 0);
 129 }
 130
 131 static u32 flow_hash_code(struct flowi *key, int cpu)
 132 {
 133         u32 *k = (u32 *) key;
 134
 135         return (jhash2(k, (sizeof(*key) / sizeof(u32)), flow_hash_rnd(cpu)) &
 136                 (flow_hash_size - 1));
 137 }
 138
 139 #if (BITS_PER_LONG == 64)
 140 typedef u64 flow_compare_t;
 141 #else
 142 typedef u32 flow_compare_t;
 143 #endif
 144
 145 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 146  * important assumptions that we can here, such as alignment and
 147  * constant size.
 148  */
 149 static int flow_key_compare(struct flowi *key1, struct flowi *key2)
 150 {
 151         flow_compare_t *k1, *k1_lim, *k2;
 152         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
 153
 154         BUILD_BUG_ON(sizeof(struct flowi) % sizeof(flow_compare_t));
 155
 156         k1 = (flow_compare_t *) key1;
 157         k1_lim = k1 + n_elem;
 158
 159         k2 = (flow_compare_t *) key2;
 160
 161         do {
 162                 if (*k1++ != *k2++)
 163                         return 1;
 164         } while (k1 < k1_lim);
 165
 166         return 0;
 167 }
 168
 169 void *flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
 170                         flow_resolve_t resolver)
 171 {
 172         struct flow_cache_entry **table, *fle, **head = NULL /* shut up GCC */;
 173         unsigned int hash;
 174         int cpu;
 175
 176         local_bh_disable();
 177         table = get_cpu_var_locked(flow_tables, &cpu);
 178
 179         fle = NULL;
 180         /* Packet really early in init?  Making flow_cache_init a
 181          * pre-smp initcall would solve this.  --RR */
 182         if (!table)
 183                 goto nocache;
 184
 185         if (flow_hash_rnd_recalc(cpu))
 186                 flow_new_hash_rnd(cpu);
 187         hash = flow_hash_code(key, cpu);
 188
 189         head = &table[hash];
 190         for (fle = *head; fle; fle = fle->next) {
 191                 if (fle->family == family &&
 192                     fle->dir == dir &&
 193                     flow_key_compare(key, &fle->key) == 0) {
 194                         if (fle->genid == atomic_read(&flow_cache_genid)) {
 195                                 void *ret = fle->object;
 196
 197                                 if (ret)
 198                                         atomic_inc(fle->object_ref);
 199                                 put_cpu_var_locked(flow_tables, cpu);
 200                                 local_bh_enable();
 201
 202                                 return ret;
 203                         }
 204                         break;
 205                 }
 206         }
 207
 208         if (!fle) {
 209                 if (flow_count(cpu) > flow_hwm)
 210                         flow_cache_shrink(cpu);
 211
 212                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 213                 if (fle) {
 214                         fle->next = *head;
 215                         *head = fle;
 216                         fle->family = family;
 217                         fle->dir = dir;
 218                         memcpy(&fle->key, key, sizeof(*key));
 219                         fle->object = NULL;
 220                         flow_count(cpu)++;
 221                 }
 222         }
 223
 224 nocache:
 225         put_cpu_var_locked(flow_tables, cpu);
 226
 227         {
 228                 int err;
 229                 void *obj;
 230                 atomic_t *obj_ref;
 231
 232                 err = resolver(net, key, family, dir, &obj, &obj_ref);
 233
 234                 if (fle && !err) {
 235                         fle->genid = atomic_read(&flow_cache_genid);
 236
 237                         if (fle->object)
 238                                 atomic_dec(fle->object_ref);
 239
 240                         fle->object = obj;
 241                         fle->object_ref = obj_ref;
 242                         if (obj)
 243                                 atomic_inc(fle->object_ref);
 244                 }
 245                 local_bh_enable();
 246
 247                 if (err)
 248                         obj = ERR_PTR(err);
 249                 return obj;
 250         }
 251 }
 252
 253 static void flow_cache_flush_tasklet(unsigned long data)
 254 {
 255         struct flow_flush_info *info = (void *)data;
 256         struct flow_cache_entry **table;
 257         int i;
 258         int cpu;
 259
 260         table = get_cpu_var_locked(flow_tables, &cpu);
 261         for (i = 0; i < flow_hash_size; i++) {
 262                 struct flow_cache_entry *fle;
 263
 264                 fle = table[i];
 265                 for (; fle; fle = fle->next) {
 266                         unsigned genid = atomic_read(&flow_cache_genid);
 267
 268                         if (!fle->object || fle->genid == genid)
 269                                 continue;
 270
 271                         fle->object = NULL;
 272                         atomic_dec(fle->object_ref);
 273                 }
 274         }
 275         put_cpu_var_locked(flow_tables, cpu);
 276
 277         if (atomic_dec_and_test(&info->cpuleft))
 278                 complete(&info->completion);
 279 }
 280
 281 static void flow_cache_flush_per_cpu(void *) __attribute__((__unused__));
 282 static void flow_cache_flush_per_cpu(void *data)
 283 {
 284         struct flow_flush_info *info = data;
 285         int cpu;
 286         struct tasklet_struct *tasklet;
 287
 288         cpu = smp_processor_id();
 289
 290         tasklet = flow_flush_tasklet(cpu);
 291         tasklet->data = (unsigned long)info;
 292         tasklet_schedule(tasklet);
 293 }
 294
 295 void flow_cache_flush(void)
 296 {
 297         struct flow_flush_info info;
 298         static DEFINE_MUTEX(flow_flush_sem);
 299
 300         /* Don't want cpus going down or up during this. */
 301         get_online_cpus();
 302         mutex_lock(&flow_flush_sem);
 303         atomic_set(&info.cpuleft, num_online_cpus());
 304         init_completion(&info.completion);
 305
 306         local_bh_disable();
 307         smp_call_function(flow_cache_flush_per_cpu, &info, 0);
 308         flow_cache_flush_tasklet((unsigned long)&info);
 309         local_bh_enable();
 310
 311         wait_for_completion(&info.completion);
 312         mutex_unlock(&flow_flush_sem);
 313         put_online_cpus();
 314 }
 315
 316 static void __init flow_cache_cpu_prepare(int cpu)
 317 {
 318         struct tasklet_struct *tasklet;
 319         unsigned long order;
 320
 321         for (order = 0;
 322              (PAGE_SIZE << order) <
 323                      (sizeof(struct flow_cache_entry *)*flow_hash_size);
 324              order++)
 325                 /* NOTHING */;
 326
 327         flow_table(cpu) = (struct flow_cache_entry **)
 328                 __get_free_pages(GFP_KERNEL|__GFP_ZERO, order);
 329         if (!flow_table(cpu))
 330                 panic("NET: failed to allocate flow cache order %lu\n", order);
 331
 332         flow_hash_rnd_recalc(cpu) = 1;
 333         flow_count(cpu) = 0;
 334
 335         tasklet = flow_flush_tasklet(cpu);
 336         tasklet_init(tasklet, flow_cache_flush_tasklet, 0);
 337 }
 338
 339 static int flow_cache_cpu(struct notifier_block *nfb,
 340                           unsigned long action,
 341                           void *hcpu)
 342 {
 343         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
 344                 __flow_cache_shrink((unsigned long)hcpu, 0);
 345         return NOTIFY_OK;
 346 }
 347
 348 static int __init flow_cache_init(void)
 349 {
 350         int i;
 351
 352         flow_cachep = kmem_cache_create("flow_cache",
 353                                         sizeof(struct flow_cache_entry),
 354                                         0, SLAB_PANIC,
 355                                         NULL);
 356         flow_hash_shift = 10;
 357         flow_lwm = 2 * flow_hash_size;
 358         flow_hwm = 4 * flow_hash_size;
 359
 360         setup_timer(&flow_hash_rnd_timer, flow_cache_new_hashrnd, 0);
 361         flow_hash_rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 362         add_timer(&flow_hash_rnd_timer);
 363
 364         for_each_possible_cpu(i)
 365                 flow_cache_cpu_prepare(i);
 366
 367         hotcpu_notifier(flow_cache_cpu, 0);
 368         return 0;
 369 }
 370
 371 module_init(flow_cache_init);
 372
 373 EXPORT_SYMBOL(flow_cache_genid);
 374 EXPORT_SYMBOL(flow_cache_lookup);