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         union {
  30                 struct hlist_node       hlist;
  31                 struct list_head        gc_list;
  32         } u;
  33         u16                             family;
  34         u8                              dir;
  35         u32                             genid;
  36         struct flowi                    key;
  37         struct flow_cache_object        *object;
  38 };
  39
  40 struct flow_cache_percpu {
  41         struct hlist_head               *hash_table;
  42         int                             hash_count;
  43         u32                             hash_rnd;
  44         int                             hash_rnd_recalc;
  45         struct tasklet_struct           flush_tasklet;
  46 };
  47
  48 struct flow_flush_info {
  49         struct flow_cache               *cache;
  50         atomic_t                        cpuleft;
  51         struct completion               completion;
  52 };
  53
  54 struct flow_cache {
  55         u32                             hash_shift;
  56         unsigned long                   order;
  57         struct flow_cache_percpu        *percpu;
  58         struct notifier_block           hotcpu_notifier;
  59         int                             low_watermark;
  60         int                             high_watermark;
  61         struct timer_list               rnd_timer;
  62 };
  63
  64 atomic_t flow_cache_genid = ATOMIC_INIT(0);
  65 EXPORT_SYMBOL(flow_cache_genid);
  66 static struct flow_cache flow_cache_global;
  67 static struct kmem_cache *flow_cachep;
  68
  69 static DEFINE_SPINLOCK(flow_cache_gc_lock);
  70 static LIST_HEAD(flow_cache_gc_list);
  71
  72 #define flow_cache_hash_size(cache)     (1 << (cache)->hash_shift)
  73 #define FLOW_HASH_RND_PERIOD            (10 * 60 * HZ)
  74
  75 static void flow_cache_new_hashrnd(unsigned long arg)
  76 {
  77         struct flow_cache *fc = (void *) arg;
  78         int i;
  79
  80         for_each_possible_cpu(i)
  81                 per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
  82
  83         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
  84         add_timer(&fc->rnd_timer);
  85 }
  86
  87 static int flow_entry_valid(struct flow_cache_entry *fle)
  88 {
  89         if (atomic_read(&flow_cache_genid) != fle->genid)
  90                 return 0;
  91         if (fle->object && !fle->object->ops->check(fle->object))
  92                 return 0;
  93         return 1;
  94 }
  95
  96 static void flow_entry_kill(struct flow_cache_entry *fle)
  97 {
  98         if (fle->object)
  99                 fle->object->ops->delete(fle->object);
 100         kmem_cache_free(flow_cachep, fle);
 101 }
 102
 103 static void flow_cache_gc_task(struct work_struct *work)
 104 {
 105         struct list_head gc_list;
 106         struct flow_cache_entry *fce, *n;
 107
 108         INIT_LIST_HEAD(&gc_list);
 109         spin_lock_bh(&flow_cache_gc_lock);
 110         list_splice_tail_init(&flow_cache_gc_list, &gc_list);
 111         spin_unlock_bh(&flow_cache_gc_lock);
 112
 113         list_for_each_entry_safe(fce, n, &gc_list, u.gc_list)
 114                 flow_entry_kill(fce);
 115 }
 116 static DECLARE_WORK(flow_cache_gc_work, flow_cache_gc_task);
 117
 118 static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
 119                                      int deleted, struct list_head *gc_list)
 120 {
 121         if (deleted) {
 122                 fcp->hash_count -= deleted;
 123                 spin_lock_bh(&flow_cache_gc_lock);
 124                 list_splice_tail(gc_list, &flow_cache_gc_list);
 125                 spin_unlock_bh(&flow_cache_gc_lock);
 126                 schedule_work(&flow_cache_gc_work);
 127         }
 128 }
 129
 130 static void __flow_cache_shrink(struct flow_cache *fc,
 131                                 struct flow_cache_percpu *fcp,
 132                                 int shrink_to)
 133 {
 134         struct flow_cache_entry *fle;
 135         struct hlist_node *entry, *tmp;
 136         LIST_HEAD(gc_list);
 137         int i, deleted = 0;
 138
 139         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 140                 int saved = 0;
 141
 142                 hlist_for_each_entry_safe(fle, entry, tmp,
 143                                           &fcp->hash_table[i], u.hlist) {
 144                         if (saved < shrink_to &&
 145                             flow_entry_valid(fle)) {
 146                                 saved++;
 147                         } else {
 148                                 deleted++;
 149                                 hlist_del(&fle->u.hlist);
 150                                 list_add_tail(&fle->u.gc_list, &gc_list);
 151                         }
 152                 }
 153         }
 154
 155         flow_cache_queue_garbage(fcp, deleted, &gc_list);
 156 }
 157
 158 static void flow_cache_shrink(struct flow_cache *fc,
 159                               struct flow_cache_percpu *fcp)
 160 {
 161         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 162
 163         __flow_cache_shrink(fc, fcp, shrink_to);
 164 }
 165
 166 static void flow_new_hash_rnd(struct flow_cache *fc,
 167                               struct flow_cache_percpu *fcp)
 168 {
 169         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
 170         fcp->hash_rnd_recalc = 0;
 171         __flow_cache_shrink(fc, fcp, 0);
 172 }
 173
 174 static u32 flow_hash_code(struct flow_cache *fc,
 175                           struct flow_cache_percpu *fcp,
 176                           struct flowi *key)
 177 {
 178         u32 *k = (u32 *) key;
 179
 180         return (jhash2(k, (sizeof(*key) / sizeof(u32)), fcp->hash_rnd)
 181                 & (flow_cache_hash_size(fc) - 1));
 182 }
 183
 184 #if (BITS_PER_LONG == 64)
 185 typedef u64 flow_compare_t;
 186 #else
 187 typedef u32 flow_compare_t;
 188 #endif
 189
 190 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 191  * important assumptions that we can here, such as alignment and
 192  * constant size.
 193  */
 194 static int flow_key_compare(struct flowi *key1, struct flowi *key2)
 195 {
 196         flow_compare_t *k1, *k1_lim, *k2;
 197         const int n_elem = sizeof(struct flowi) / sizeof(flow_compare_t);
 198
 199         BUILD_BUG_ON(sizeof(struct flowi) % sizeof(flow_compare_t));
 200
 201         k1 = (flow_compare_t *) key1;
 202         k1_lim = k1 + n_elem;
 203
 204         k2 = (flow_compare_t *) key2;
 205
 206         do {
 207                 if (*k1++ != *k2++)
 208                         return 1;
 209         } while (k1 < k1_lim);
 210
 211         return 0;
 212 }
 213
 214 struct flow_cache_object *
 215 flow_cache_lookup(struct net *net, struct flowi *key, u16 family, u8 dir,
 216                   flow_resolve_t resolver, void *ctx)
 217 {
 218         struct flow_cache *fc = &flow_cache_global;
 219         struct flow_cache_percpu *fcp;
 220         struct flow_cache_entry *fle, *tfle;
 221         struct hlist_node *entry;
 222         struct flow_cache_object *flo;
 223         unsigned int hash;
 224
 225         local_bh_disable();
 226         fcp = this_cpu_ptr(fc->percpu);
 227
 228         fle = NULL;
 229         flo = NULL;
 230         /* Packet really early in init?  Making flow_cache_init a
 231          * pre-smp initcall would solve this.  --RR */
 232         if (!fcp->hash_table)
 233                 goto nocache;
 234
 235         if (fcp->hash_rnd_recalc)
 236                 flow_new_hash_rnd(fc, fcp);
 237
 238         hash = flow_hash_code(fc, fcp, key);
 239         hlist_for_each_entry(tfle, entry, &fcp->hash_table[hash], u.hlist) {
 240                 if (tfle->family == family &&
 241                     tfle->dir == dir &&
 242                     flow_key_compare(key, &tfle->key) == 0) {
 243                         fle = tfle;
 244                         break;
 245                 }
 246         }
 247
 248         if (unlikely(!fle)) {
 249                 if (fcp->hash_count > fc->high_watermark)
 250                         flow_cache_shrink(fc, fcp);
 251
 252                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 253                 if (fle) {
 254                         fle->family = family;
 255                         fle->dir = dir;
 256                         memcpy(&fle->key, key, sizeof(*key));
 257                         fle->object = NULL;
 258                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
 259                         fcp->hash_count++;
 260                 }
 261         } else if (likely(fle->genid == atomic_read(&flow_cache_genid))) {
 262                 flo = fle->object;
 263                 if (!flo)
 264                         goto ret_object;
 265                 flo = flo->ops->get(flo);
 266                 if (flo)
 267                         goto ret_object;
 268         } else if (fle->object) {
 269                 flo = fle->object;
 270                 flo->ops->delete(flo);
 271                 fle->object = NULL;
 272         }
 273
 274 nocache:
 275         flo = NULL;
 276         if (fle) {
 277                 flo = fle->object;
 278                 fle->object = NULL;
 279         }
 280         flo = resolver(net, key, family, dir, flo, ctx);
 281         if (fle) {
 282                 fle->genid = atomic_read(&flow_cache_genid);
 283                 if (!IS_ERR(flo))
 284                         fle->object = flo;
 285                 else
 286                         fle->genid--;
 287         } else {
 288                 if (flo && !IS_ERR(flo))
 289                         flo->ops->delete(flo);
 290         }
 291 ret_object:
 292         local_bh_enable();
 293         return flo;
 294 }
 295 EXPORT_SYMBOL(flow_cache_lookup);
 296
 297 static void flow_cache_flush_tasklet(unsigned long data)
 298 {
 299         struct flow_flush_info *info = (void *)data;
 300         struct flow_cache *fc = info->cache;
 301         struct flow_cache_percpu *fcp;
 302         struct flow_cache_entry *fle;
 303         struct hlist_node *entry, *tmp;
 304         LIST_HEAD(gc_list);
 305         int i, deleted = 0;
 306
 307         fcp = this_cpu_ptr(fc->percpu);
 308         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 309                 hlist_for_each_entry_safe(fle, entry, tmp,
 310                                           &fcp->hash_table[i], u.hlist) {
 311                         if (flow_entry_valid(fle))
 312                                 continue;
 313
 314                         deleted++;
 315                         hlist_del(&fle->u.hlist);
 316                         list_add_tail(&fle->u.gc_list, &gc_list);
 317                 }
 318         }
 319
 320         flow_cache_queue_garbage(fcp, deleted, &gc_list);
 321
 322         if (atomic_dec_and_test(&info->cpuleft))
 323                 complete(&info->completion);
 324 }
 325
 326 static void flow_cache_flush_per_cpu(void *data)
 327 {
 328         struct flow_flush_info *info = data;
 329         int cpu;
 330         struct tasklet_struct *tasklet;
 331
 332         cpu = smp_processor_id();
 333         tasklet = &per_cpu_ptr(info->cache->percpu, cpu)->flush_tasklet;
 334         tasklet->data = (unsigned long)info;
 335         tasklet_schedule(tasklet);
 336 }
 337
 338 void flow_cache_flush(void)
 339 {
 340         struct flow_flush_info info;
 341         static DEFINE_MUTEX(flow_flush_sem);
 342
 343         /* Don't want cpus going down or up during this. */
 344         get_online_cpus();
 345         mutex_lock(&flow_flush_sem);
 346         info.cache = &flow_cache_global;
 347         atomic_set(&info.cpuleft, num_online_cpus());
 348         init_completion(&info.completion);
 349
 350         local_bh_disable();
 351         smp_call_function(flow_cache_flush_per_cpu, &info, 0);
 352         flow_cache_flush_tasklet((unsigned long)&info);
 353         local_bh_enable();
 354
 355         wait_for_completion(&info.completion);
 356         mutex_unlock(&flow_flush_sem);
 357         put_online_cpus();
 358 }
 359
 360 static void __init flow_cache_cpu_prepare(struct flow_cache *fc,
 361                                           struct flow_cache_percpu *fcp)
 362 {
 363         fcp->hash_table = (struct hlist_head *)
 364                 __get_free_pages(GFP_KERNEL|__GFP_ZERO, fc->order);
 365         if (!fcp->hash_table)
 366                 panic("NET: failed to allocate flow cache order %lu\n", fc->order);
 367
 368         fcp->hash_rnd_recalc = 1;
 369         fcp->hash_count = 0;
 370         tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 371 }
 372
 373 static int flow_cache_cpu(struct notifier_block *nfb,
 374                           unsigned long action,
 375                           void *hcpu)
 376 {
 377         struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
 378         int cpu = (unsigned long) hcpu;
 379         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 380
 381         if (action == CPU_DEAD || action == CPU_DEAD_FROZEN)
 382                 __flow_cache_shrink(fc, fcp, 0);
 383         return NOTIFY_OK;
 384 }
 385
 386 static int flow_cache_init(struct flow_cache *fc)
 387 {
 388         unsigned long order;
 389         int i;
 390
 391         fc->hash_shift = 10;
 392         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 393         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 394
 395         for (order = 0;
 396              (PAGE_SIZE << order) <
 397                      (sizeof(struct hlist_head)*flow_cache_hash_size(fc));
 398              order++)
 399                 /* NOTHING */;
 400         fc->order = order;
 401         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 402
 403         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 404                     (unsigned long) fc);
 405         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 406         add_timer(&fc->rnd_timer);
 407
 408         for_each_possible_cpu(i)
 409                 flow_cache_cpu_prepare(fc, per_cpu_ptr(fc->percpu, i));
 410
 411         fc->hotcpu_notifier = (struct notifier_block){
 412                 .notifier_call = flow_cache_cpu,
 413         };
 414         register_hotcpu_notifier(&fc->hotcpu_notifier);
 415
 416         return 0;
 417 }
 418
 419 static int __init flow_cache_init_global(void)
 420 {
 421         flow_cachep = kmem_cache_create("flow_cache",
 422                                         sizeof(struct flow_cache_entry),
 423                                         0, SLAB_PANIC, NULL);
 424
 425         return flow_cache_init(&flow_cache_global);
 426 }
 427
 428 module_init(flow_cache_init_global);