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 <linux/atomic.h>
  26 #include <linux/security.h>
  27 #include <net/net_namespace.h>
  28
  29 struct flow_cache_entry {
  30         union {
  31                 struct hlist_node       hlist;
  32                 struct list_head        gc_list;
  33         } u;
  34         struct net                      *net;
  35         u16                             family;
  36         u8                              dir;
  37         u32                             genid;
  38         struct flowi                    key;
  39         struct flow_cache_object        *object;
  40 };
  41
  42 struct flow_flush_info {
  43         struct flow_cache               *cache;
  44         atomic_t                        cpuleft;
  45         struct completion               completion;
  46 };
  47
  48 static struct kmem_cache *flow_cachep __read_mostly;
  49
  50 #define flow_cache_hash_size(cache)     (1 << (cache)->hash_shift)
  51 #define FLOW_HASH_RND_PERIOD            (10 * 60 * HZ)
  52
  53 static void flow_cache_new_hashrnd(unsigned long arg)
  54 {
  55         struct flow_cache *fc = (void *) arg;
  56         int i;
  57
  58         for_each_possible_cpu(i)
  59                 per_cpu_ptr(fc->percpu, i)->hash_rnd_recalc = 1;
  60
  61         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
  62         add_timer(&fc->rnd_timer);
  63 }
  64
  65 static int flow_entry_valid(struct flow_cache_entry *fle,
  66                                 struct netns_xfrm *xfrm)
  67 {
  68         if (atomic_read(&xfrm->flow_cache_genid) != fle->genid)
  69                 return 0;
  70         if (fle->object && !fle->object->ops->check(fle->object))
  71                 return 0;
  72         return 1;
  73 }
  74
  75 static void flow_entry_kill(struct flow_cache_entry *fle,
  76                                 struct netns_xfrm *xfrm)
  77 {
  78         if (fle->object)
  79                 fle->object->ops->delete(fle->object);
  80         kmem_cache_free(flow_cachep, fle);
  81 }
  82
  83 static void flow_cache_gc_task(struct work_struct *work)
  84 {
  85         struct list_head gc_list;
  86         struct flow_cache_entry *fce, *n;
  87         struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
  88                                                 flow_cache_gc_work);
  89
  90         INIT_LIST_HEAD(&gc_list);
  91         spin_lock_bh(&xfrm->flow_cache_gc_lock);
  92         list_splice_tail_init(&xfrm->flow_cache_gc_list, &gc_list);
  93         spin_unlock_bh(&xfrm->flow_cache_gc_lock);
  94
  95         list_for_each_entry_safe(fce, n, &gc_list, u.gc_list)
  96                 flow_entry_kill(fce, xfrm);
  97 }
  98
  99 static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
 100                                      int deleted, struct list_head *gc_list,
 101                                      struct netns_xfrm *xfrm)
 102 {
 103         if (deleted) {
 104                 fcp->hash_count -= deleted;
 105                 spin_lock_bh(&xfrm->flow_cache_gc_lock);
 106                 list_splice_tail(gc_list, &xfrm->flow_cache_gc_list);
 107                 spin_unlock_bh(&xfrm->flow_cache_gc_lock);
 108                 schedule_work(&xfrm->flow_cache_gc_work);
 109         }
 110 }
 111
 112 static void __flow_cache_shrink(struct flow_cache *fc,
 113                                 struct flow_cache_percpu *fcp,
 114                                 int shrink_to)
 115 {
 116         struct flow_cache_entry *fle;
 117         struct hlist_node *tmp;
 118         LIST_HEAD(gc_list);
 119         int i, deleted = 0;
 120         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 121                                                 flow_cache_global);
 122
 123         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 124                 int saved = 0;
 125
 126                 hlist_for_each_entry_safe(fle, tmp,
 127                                           &fcp->hash_table[i], u.hlist) {
 128                         if (saved < shrink_to &&
 129                             flow_entry_valid(fle, xfrm)) {
 130                                 saved++;
 131                         } else {
 132                                 deleted++;
 133                                 hlist_del(&fle->u.hlist);
 134                                 list_add_tail(&fle->u.gc_list, &gc_list);
 135                         }
 136                 }
 137         }
 138
 139         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 140 }
 141
 142 static void flow_cache_shrink(struct flow_cache *fc,
 143                               struct flow_cache_percpu *fcp)
 144 {
 145         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 146
 147         __flow_cache_shrink(fc, fcp, shrink_to);
 148 }
 149
 150 static void flow_new_hash_rnd(struct flow_cache *fc,
 151                               struct flow_cache_percpu *fcp)
 152 {
 153         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
 154         fcp->hash_rnd_recalc = 0;
 155         __flow_cache_shrink(fc, fcp, 0);
 156 }
 157
 158 static u32 flow_hash_code(struct flow_cache *fc,
 159                           struct flow_cache_percpu *fcp,
 160                           const struct flowi *key,
 161                           size_t keysize)
 162 {
 163         const u32 *k = (const u32 *) key;
 164         const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
 165
 166         return jhash2(k, length, fcp->hash_rnd)
 167                 & (flow_cache_hash_size(fc) - 1);
 168 }
 169
 170 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 171  * important assumptions that we can here, such as alignment.
 172  */
 173 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
 174                             size_t keysize)
 175 {
 176         const flow_compare_t *k1, *k1_lim, *k2;
 177
 178         k1 = (const flow_compare_t *) key1;
 179         k1_lim = k1 + keysize;
 180
 181         k2 = (const flow_compare_t *) key2;
 182
 183         do {
 184                 if (*k1++ != *k2++)
 185                         return 1;
 186         } while (k1 < k1_lim);
 187
 188         return 0;
 189 }
 190
 191 struct flow_cache_object *
 192 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
 193                   flow_resolve_t resolver, void *ctx)
 194 {
 195         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 196         struct flow_cache_percpu *fcp;
 197         struct flow_cache_entry *fle, *tfle;
 198         struct flow_cache_object *flo;
 199         size_t keysize;
 200         unsigned int hash;
 201
 202         local_bh_disable();
 203         fcp = this_cpu_ptr(fc->percpu);
 204
 205         fle = NULL;
 206         flo = NULL;
 207
 208         keysize = flow_key_size(family);
 209         if (!keysize)
 210                 goto nocache;
 211
 212         /* Packet really early in init?  Making flow_cache_init a
 213          * pre-smp initcall would solve this.  --RR */
 214         if (!fcp->hash_table)
 215                 goto nocache;
 216
 217         if (fcp->hash_rnd_recalc)
 218                 flow_new_hash_rnd(fc, fcp);
 219
 220         hash = flow_hash_code(fc, fcp, key, keysize);
 221         hlist_for_each_entry(tfle, &fcp->hash_table[hash], u.hlist) {
 222                 if (tfle->net == net &&
 223                     tfle->family == family &&
 224                     tfle->dir == dir &&
 225                     flow_key_compare(key, &tfle->key, keysize) == 0) {
 226                         fle = tfle;
 227                         break;
 228                 }
 229         }
 230
 231         if (unlikely(!fle)) {
 232                 if (fcp->hash_count > fc->high_watermark)
 233                         flow_cache_shrink(fc, fcp);
 234
 235                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 236                 if (fle) {
 237                         fle->net = net;
 238                         fle->family = family;
 239                         fle->dir = dir;
 240                         memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
 241                         fle->object = NULL;
 242                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
 243                         fcp->hash_count++;
 244                 }
 245         } else if (likely(fle->genid == atomic_read(&net->xfrm.flow_cache_genid))) {
 246                 flo = fle->object;
 247                 if (!flo)
 248                         goto ret_object;
 249                 flo = flo->ops->get(flo);
 250                 if (flo)
 251                         goto ret_object;
 252         } else if (fle->object) {
 253                 flo = fle->object;
 254                 flo->ops->delete(flo);
 255                 fle->object = NULL;
 256         }
 257
 258 nocache:
 259         flo = NULL;
 260         if (fle) {
 261                 flo = fle->object;
 262                 fle->object = NULL;
 263         }
 264         flo = resolver(net, key, family, dir, flo, ctx);
 265         if (fle) {
 266                 fle->genid = atomic_read(&net->xfrm.flow_cache_genid);
 267                 if (!IS_ERR(flo))
 268                         fle->object = flo;
 269                 else
 270                         fle->genid--;
 271         } else {
 272                 if (!IS_ERR_OR_NULL(flo))
 273                         flo->ops->delete(flo);
 274         }
 275 ret_object:
 276         local_bh_enable();
 277         return flo;
 278 }
 279 EXPORT_SYMBOL(flow_cache_lookup);
 280
 281 static void flow_cache_flush_tasklet(unsigned long data)
 282 {
 283         struct flow_flush_info *info = (void *)data;
 284         struct flow_cache *fc = info->cache;
 285         struct flow_cache_percpu *fcp;
 286         struct flow_cache_entry *fle;
 287         struct hlist_node *tmp;
 288         LIST_HEAD(gc_list);
 289         int i, deleted = 0;
 290         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 291                                                 flow_cache_global);
 292
 293         fcp = this_cpu_ptr(fc->percpu);
 294         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 295                 hlist_for_each_entry_safe(fle, tmp,
 296                                           &fcp->hash_table[i], u.hlist) {
 297                         if (flow_entry_valid(fle, xfrm))
 298                                 continue;
 299
 300                         deleted++;
 301                         hlist_del(&fle->u.hlist);
 302                         list_add_tail(&fle->u.gc_list, &gc_list);
 303                 }
 304         }
 305
 306         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 307
 308         if (atomic_dec_and_test(&info->cpuleft))
 309                 complete(&info->completion);
 310 }
 311
 312 /*
 313  * Return whether a cpu needs flushing.  Conservatively, we assume
 314  * the presence of any entries means the core may require flushing,
 315  * since the flow_cache_ops.check() function may assume it's running
 316  * on the same core as the per-cpu cache component.
 317  */
 318 static int flow_cache_percpu_empty(struct flow_cache *fc, int cpu)
 319 {
 320         struct flow_cache_percpu *fcp;
 321         int i;
 322
 323         fcp = per_cpu_ptr(fc->percpu, cpu);
 324         for (i = 0; i < flow_cache_hash_size(fc); i++)
 325                 if (!hlist_empty(&fcp->hash_table[i]))
 326                         return 0;
 327         return 1;
 328 }
 329
 330 static void flow_cache_flush_per_cpu(void *data)
 331 {
 332         struct flow_flush_info *info = data;
 333         struct tasklet_struct *tasklet;
 334
 335         tasklet = &this_cpu_ptr(info->cache->percpu)->flush_tasklet;
 336         tasklet->data = (unsigned long)info;
 337         tasklet_schedule(tasklet);
 338 }
 339
 340 void flow_cache_flush(struct net *net)
 341 {
 342         struct flow_flush_info info;
 343         cpumask_var_t mask;
 344         int i, self;
 345
 346         /* Track which cpus need flushing to avoid disturbing all cores. */
 347         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 348                 return;
 349         cpumask_clear(mask);
 350
 351         /* Don't want cpus going down or up during this. */
 352         get_online_cpus();
 353         mutex_lock(&net->xfrm.flow_flush_sem);
 354         info.cache = &net->xfrm.flow_cache_global;
 355         for_each_online_cpu(i)
 356                 if (!flow_cache_percpu_empty(info.cache, i))
 357                         cpumask_set_cpu(i, mask);
 358         atomic_set(&info.cpuleft, cpumask_weight(mask));
 359         if (atomic_read(&info.cpuleft) == 0)
 360                 goto done;
 361
 362         init_completion(&info.completion);
 363
 364         local_bh_disable();
 365         self = cpumask_test_and_clear_cpu(smp_processor_id(), mask);
 366         on_each_cpu_mask(mask, flow_cache_flush_per_cpu, &info, 0);
 367         if (self)
 368                 flow_cache_flush_tasklet((unsigned long)&info);
 369         local_bh_enable();
 370
 371         wait_for_completion(&info.completion);
 372
 373 done:
 374         mutex_unlock(&net->xfrm.flow_flush_sem);
 375         put_online_cpus();
 376         free_cpumask_var(mask);
 377 }
 378
 379 static void flow_cache_flush_task(struct work_struct *work)
 380 {
 381         struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
 382                                                 flow_cache_flush_work);
 383         struct net *net = container_of(xfrm, struct net, xfrm);
 384
 385         flow_cache_flush(net);
 386 }
 387
 388 void flow_cache_flush_deferred(struct net *net)
 389 {
 390         schedule_work(&net->xfrm.flow_cache_flush_work);
 391 }
 392
 393 static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 394 {
 395         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 396         size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
 397
 398         if (!fcp->hash_table) {
 399                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
 400                 if (!fcp->hash_table) {
 401                         pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
 402                         return -ENOMEM;
 403                 }
 404                 fcp->hash_rnd_recalc = 1;
 405                 fcp->hash_count = 0;
 406                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 407         }
 408         return 0;
 409 }
 410
 411 static int flow_cache_cpu(struct notifier_block *nfb,
 412                           unsigned long action,
 413                           void *hcpu)
 414 {
 415         struct flow_cache *fc = container_of(nfb, struct flow_cache,
 416                                                 hotcpu_notifier);
 417         int res, cpu = (unsigned long) hcpu;
 418         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 419
 420         switch (action) {
 421         case CPU_UP_PREPARE:
 422         case CPU_UP_PREPARE_FROZEN:
 423                 res = flow_cache_cpu_prepare(fc, cpu);
 424                 if (res)
 425                         return notifier_from_errno(res);
 426                 break;
 427         case CPU_DEAD:
 428         case CPU_DEAD_FROZEN:
 429                 __flow_cache_shrink(fc, fcp, 0);
 430                 break;
 431         }
 432         return NOTIFY_OK;
 433 }
 434
 435 int flow_cache_init(struct net *net)
 436 {
 437         int i;
 438         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 439
 440         if (!flow_cachep)
 441                 flow_cachep = kmem_cache_create("flow_cache",
 442                                                 sizeof(struct flow_cache_entry),
 443                                                 0, SLAB_PANIC, NULL);
 444         spin_lock_init(&net->xfrm.flow_cache_gc_lock);
 445         INIT_LIST_HEAD(&net->xfrm.flow_cache_gc_list);
 446         INIT_WORK(&net->xfrm.flow_cache_gc_work, flow_cache_gc_task);
 447         INIT_WORK(&net->xfrm.flow_cache_flush_work, flow_cache_flush_task);
 448         mutex_init(&net->xfrm.flow_flush_sem);
 449
 450         fc->hash_shift = 10;
 451         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 452         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 453
 454         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 455         if (!fc->percpu)
 456                 return -ENOMEM;
 457
 458         cpu_notifier_register_begin();
 459
 460         for_each_online_cpu(i) {
 461                 if (flow_cache_cpu_prepare(fc, i))
 462                         goto err;
 463         }
 464         fc->hotcpu_notifier = (struct notifier_block){
 465                 .notifier_call = flow_cache_cpu,
 466         };
 467         __register_hotcpu_notifier(&fc->hotcpu_notifier);
 468
 469         cpu_notifier_register_done();
 470
 471         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 472                     (unsigned long) fc);
 473         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 474         add_timer(&fc->rnd_timer);
 475
 476         return 0;
 477
 478 err:
 479         for_each_possible_cpu(i) {
 480                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 481                 kfree(fcp->hash_table);
 482                 fcp->hash_table = NULL;
 483         }
 484
 485         cpu_notifier_register_done();
 486
 487         free_percpu(fc->percpu);
 488         fc->percpu = NULL;
 489
 490         return -ENOMEM;
 491 }
 492 EXPORT_SYMBOL(flow_cache_init);
 493
 494 void flow_cache_fini(struct net *net)
 495 {
 496         int i;
 497         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 498
 499         del_timer_sync(&fc->rnd_timer);
 500         unregister_hotcpu_notifier(&fc->hotcpu_notifier);
 501
 502         for_each_possible_cpu(i) {
 503                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 504                 kfree(fcp->hash_table);
 505                 fcp->hash_table = NULL;
 506         }
 507
 508         free_percpu(fc->percpu);
 509         fc->percpu = NULL;
 510 }
 511 EXPORT_SYMBOL(flow_cache_fini);