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 struct flow_cache_percpu {
  39         struct flow_cache_entry **      hash_table;
  40         int                             hash_count;
  41         u32                             hash_rnd;
  42         int                             hash_rnd_recalc;
  43         struct tasklet_struct           flush_tasklet;
  44 };
  45
  46 struct flow_flush_info {
  47         struct flow_cache *             cache;
  48         atomic_t                        cpuleft;
  49         struct completion               completion;
  50 };
  51
  52 struct flow_cache {
  53         u32                             hash_shift;
  54         unsigned long                   order;
  55         struct flow_cache_percpu *      percpu;
  56         struct notifier_block           hotcpu_notifier;
  57         int                             low_watermark;
  58         int                             high_watermark;
  59         struct timer_list               rnd_timer;
  60 };
  61
  62 atomic_t flow_cache_genid = ATOMIC_INIT(0);
  63 static struct flow_cache flow_cache_global;
  64 static struct kmem_cache *flow_cachep;
  65
  66 #define flow_cache_hash_size(cache)     (1 << (cache)->hash_shift)
  67 #define FLOW_HASH_RND_PERIOD            (10 * 60 * HZ)
  68
  69 static void flow_cache_new_hashrnd(unsigned long arg)
  70 {
  71         struct flow_cache *fc = (void *) arg;
  72         int i;
  73
  74         for_each_possible_cpu(i)
  75                 per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
  76
  77         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
  78         add_timer(&fc->rnd_timer);
  79 }
  80
  81 static void flow_entry_kill(struct flow_cache *fc,
  82                             struct flow_cache_percpu *fcp,
  83                             struct flow_cache_entry *fle)
  84 {
  85         if (fle->object)
  86                 atomic_dec(fle->object_ref);
  87         kmem_cache_free(flow_cachep, fle);
  88         fcp->hash_count--;
  89 }
  90
  91 static void __flow_cache_shrink(struct flow_cache *fc,
  92                                 struct flow_cache_percpu *fcp,
  93                                 int shrink_to)
  94 {
  95         struct flow_cache_entry *fle, **flp;
  96         int i;
  97
  98         for (i = 0; i < flow_cache_hash_size(fc); i++) {
  99                 int k = 0;
 100
 101                 flp = &fcp->hash_table[i];
 102                 while ((fle = *flp) != NULL && k < shrink_to) {
 103                         k++;
 104                         flp = &fle->next;
 105                 }
 106                 while ((fle = *flp) != NULL) {
 107                         *flp = fle->next;
 108                         flow_entry_kill(fc, fcp, fle);
 109                 }
 110         }
 111 }
 112
 113 static void flow_cache_shrink(struct flow_cache *fc,
 114                               struct flow_cache_percpu *fcp)
 115 {
 116         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 117
 118         __flow_cache_shrink(fc, fcp, shrink_to);
 119 }
 120
 121 static void flow_new_hash_rnd(struct flow_cache *fc,
 122                               struct flow_cache_percpu *fcp)
 123 {
 124         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
 125         fcp->hash_rnd_recalc = 0;
 126         __flow_cache_shrink(fc, fcp, 0);
 127 }
 128
 129 static u32 flow_hash_code(struct flow_cache *fc,
 130                           struct flow_cache_percpu *fcp,
 131                           struct flowi *key)
 132 {
 133         u32 *k = (u32 *) key;
 134
 135         return (jhash2(k, (sizeof(*key) / sizeof(u32)), fcp->hash_rnd)
 136                 & (flow_cache_hash_size(fc) - 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 *fc = &flow_cache_global;
 173         struct flow_cache_percpu *fcp;
 174         struct flow_cache_entry *fle, **head;
 175         unsigned int hash;
 176
 177         local_bh_disable();
 178         fcp = per_cpu_ptr(fc->percpu, smp_processor_id());
 179
 180         fle = NULL;
 181         /* Packet really early in init?  Making flow_cache_init a
 182          * pre-smp initcall would solve this.  --RR */
 183         if (!fcp->hash_table)
 184                 goto nocache;
 185
 186         if (fcp->hash_rnd_recalc)
 187                 flow_new_hash_rnd(fc, fcp);
 188         hash = flow_hash_code(fc, fcp, key);
 189
 190         head = &fcp->hash_table[hash];
 191         for (fle = *head; fle; fle = fle->next) {
 192                 if (fle->family == family &&
 193                     fle->dir == dir &&
 194                     flow_key_compare(key, &fle->key) == 0) {
 195                         if (fle->genid == atomic_read(&flow_cache_genid)) {
 196                                 void *ret = fle->object;
 197
 198                                 if (ret)
 199                                         atomic_inc(fle->object_ref);
 200                                 local_bh_enable();
 201
 202                                 return ret;
 203                         }
 204                         break;
 205                 }
 206         }
 207
 208         if (!fle) {
 209                 if (fcp->hash_count > fc->high_watermark)
 210                         flow_cache_shrink(fc, fcp);
 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                         fcp->hash_count++;
 221                 }
 222         }
 223
 224 nocache:
 225         {
 226                 int err;
 227                 void *obj;
 228                 atomic_t *obj_ref;
 229
 230                 err = resolver(net, key, family, dir, &obj, &obj_ref);
 231
 232                 if (fle && !err) {
 233                         fle->genid = atomic_read(&flow_cache_genid);
 234
 235                         if (fle->object)
 236                                 atomic_dec(fle->object_ref);
 237
 238                         fle->object = obj;
 239                         fle->object_ref = obj_ref;
 240                         if (obj)
 241                                 atomic_inc(fle->object_ref);
 242                 }
 243                 local_bh_enable();
 244
 245                 if (err)
 246                         obj = ERR_PTR(err);
 247                 return obj;
 248         }
 249 }
 250
 251 static void flow_cache_flush_tasklet(unsigned long data)
 252 {
 253         struct flow_flush_info *info = (void *)data;
 254         struct flow_cache *fc = info->cache;
 255         struct flow_cache_percpu *fcp;
 256         int i;
 257
 258         fcp = per_cpu_ptr(fc->percpu, smp_processor_id());
 259         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 260                 struct flow_cache_entry *fle;
 261
 262                 fle = fcp->hash_table[i];
 263                 for (; fle; fle = fle->next) {
 264                         unsigned genid = atomic_read(&flow_cache_genid);
 265
 266                         if (!fle->object || fle->genid == genid)
 267                                 continue;
 268
 269                         fle->object = NULL;
 270                         atomic_dec(fle->object_ref);
 271                 }
 272         }
 273
 274         if (atomic_dec_and_test(&info->cpuleft))
 275                 complete(&info->completion);
 276 }
 277
 278 static void flow_cache_flush_per_cpu(void *data)
 279 {
 280         struct flow_flush_info *info = data;
 281         int cpu;
 282         struct tasklet_struct *tasklet;
 283
 284         cpu = smp_processor_id();
 285         tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
 286         tasklet->data = (unsigned long)info;
 287         tasklet_schedule(tasklet);
 288 }
 289
 290 void flow_cache_flush(void)
 291 {
 292         struct flow_flush_info info;
 293         static DEFINE_MUTEX(flow_flush_sem);
 294
 295         /* Don't want cpus going down or up during this. */
 296         get_online_cpus();
 297         mutex_lock(&flow_flush_sem);
 298         info.cache = &flow_cache_global;
 299         atomic_set(&info.cpuleft, num_online_cpus());
 300         init_completion(&info.completion);
 301
 302         local_bh_disable();
 303         smp_call_function(flow_cache_flush_per_cpu, &info, 0);
 304         flow_cache_flush_tasklet((unsigned long)&info);
 305         local_bh_enable();
 306
 307         wait_for_completion(&info.completion);
 308         mutex_unlock(&flow_flush_sem);
 309         put_online_cpus();
 310 }
 311
 312 static void __init flow_cache_cpu_prepare(struct flow_cache *fc,
 313                                           struct flow_cache_percpu *fcp)
 314 {
 315         fcp->hash_table = (struct flow_cache_entry **)
 316                 __get_free_pages(GFP_KERNEL|__GFP_ZERO, fc->order);
 317         if (!fcp->hash_table)
 318                 panic("NET: failed to allocate flow cache order %lu\n", fc->order);
 319
 320         fcp->hash_rnd_recalc = 1;
 321         fcp->hash_count = 0;
 322         tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 323 }
 324
 325 static int flow_cache_cpu(struct notifier_block *nfb,
 326                           unsigned long action,
 327                           void *hcpu)
 328 {
 329         struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
 330         int cpu = (unsigned long) hcpu;
 331         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 332
 333         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
 334                 __flow_cache_shrink(fc, fcp, 0);
 335         return NOTIFY_OK;
 336 }
 337
 338 static int flow_cache_init(struct flow_cache *fc)
 339 {
 340         unsigned long order;
 341         int i;
 342
 343         fc->hash_shift = 10;
 344         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 345         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 346
 347         for (order = 0;
 348              (PAGE_SIZE << order) <
 349                      (sizeof(struct flow_cache_entry *)*flow_cache_hash_size(fc));
 350              order++)
 351                 /* NOTHING */;
 352         fc->order = order;
 353         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 354
 355         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 356                     (unsigned long) fc);
 357         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 358         add_timer(&fc->rnd_timer);
 359
 360         for_each_possible_cpu(i)
 361                 flow_cache_cpu_prepare(fc, per_cpu_ptr(fc->percpu, i));
 362
 363         fc->hotcpu_notifier = (struct notifier_block){
 364                 .notifier_call = flow_cache_cpu,
 365         };
 366         register_hotcpu_notifier(&fc->hotcpu_notifier);
 367
 368         return 0;
 369 }
 370
 371 static int __init flow_cache_init_global(void)
 372 {
 373         flow_cachep = kmem_cache_create("flow_cache",
 374                                         sizeof(struct flow_cache_entry),
 375                                         0, SLAB_PANIC, NULL);
 376
 377         return flow_cache_init(&flow_cache_global);
 378 }
 379
 380 module_init(flow_cache_init_global);
 381
 382 EXPORT_SYMBOL(flow_cache_genid);
 383 EXPORT_SYMBOL(flow_cache_lookup);