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                 atomic_dec(&xfrm->flow_cache_gc_count);
  98         }
  99 }
 100
 101 static void flow_cache_queue_garbage(struct flow_cache_percpu *fcp,
 102                                      int deleted, struct list_head *gc_list,
 103                                      struct netns_xfrm *xfrm)
 104 {
 105         if (deleted) {
 106                 atomic_add(deleted, &xfrm->flow_cache_gc_count);
 107                 fcp->hash_count -= deleted;
 108                 spin_lock_bh(&xfrm->flow_cache_gc_lock);
 109                 list_splice_tail(gc_list, &xfrm->flow_cache_gc_list);
 110                 spin_unlock_bh(&xfrm->flow_cache_gc_lock);
 111                 schedule_work(&xfrm->flow_cache_gc_work);
 112         }
 113 }
 114
 115 static void __flow_cache_shrink(struct flow_cache *fc,
 116                                 struct flow_cache_percpu *fcp,
 117                                 int shrink_to)
 118 {
 119         struct flow_cache_entry *fle;
 120         struct hlist_node *tmp;
 121         LIST_HEAD(gc_list);
 122         int i, deleted = 0;
 123         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 124                                                 flow_cache_global);
 125
 126         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 127                 int saved = 0;
 128
 129                 hlist_for_each_entry_safe(fle, tmp,
 130                                           &fcp->hash_table[i], u.hlist) {
 131                         if (saved < shrink_to &&
 132                             flow_entry_valid(fle, xfrm)) {
 133                                 saved++;
 134                         } else {
 135                                 deleted++;
 136                                 hlist_del(&fle->u.hlist);
 137                                 list_add_tail(&fle->u.gc_list, &gc_list);
 138                         }
 139                 }
 140         }
 141
 142         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 143 }
 144
 145 static void flow_cache_shrink(struct flow_cache *fc,
 146                               struct flow_cache_percpu *fcp)
 147 {
 148         int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 149
 150         __flow_cache_shrink(fc, fcp, shrink_to);
 151 }
 152
 153 static void flow_new_hash_rnd(struct flow_cache *fc,
 154                               struct flow_cache_percpu *fcp)
 155 {
 156         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
 157         fcp->hash_rnd_recalc = 0;
 158         __flow_cache_shrink(fc, fcp, 0);
 159 }
 160
 161 static u32 flow_hash_code(struct flow_cache *fc,
 162                           struct flow_cache_percpu *fcp,
 163                           const struct flowi *key,
 164                           unsigned int keysize)
 165 {
 166         const u32 *k = (const u32 *) key;
 167         const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
 168
 169         return jhash2(k, length, fcp->hash_rnd)
 170                 & (flow_cache_hash_size(fc) - 1);
 171 }
 172
 173 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 174  * important assumptions that we can here, such as alignment.
 175  */
 176 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
 177                             unsigned int keysize)
 178 {
 179         const flow_compare_t *k1, *k1_lim, *k2;
 180
 181         k1 = (const flow_compare_t *) key1;
 182         k1_lim = k1 + keysize;
 183
 184         k2 = (const flow_compare_t *) key2;
 185
 186         do {
 187                 if (*k1++ != *k2++)
 188                         return 1;
 189         } while (k1 < k1_lim);
 190
 191         return 0;
 192 }
 193
 194 struct flow_cache_object *
 195 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
 196                   flow_resolve_t resolver, void *ctx)
 197 {
 198         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 199         struct flow_cache_percpu *fcp;
 200         struct flow_cache_entry *fle, *tfle;
 201         struct flow_cache_object *flo;
 202         unsigned int keysize;
 203         unsigned int hash;
 204
 205         local_bh_disable();
 206         fcp = this_cpu_ptr(fc->percpu);
 207
 208         fle = NULL;
 209         flo = NULL;
 210
 211         keysize = flow_key_size(family);
 212         if (!keysize)
 213                 goto nocache;
 214
 215         /* Packet really early in init?  Making flow_cache_init a
 216          * pre-smp initcall would solve this.  --RR */
 217         if (!fcp->hash_table)
 218                 goto nocache;
 219
 220         if (fcp->hash_rnd_recalc)
 221                 flow_new_hash_rnd(fc, fcp);
 222
 223         hash = flow_hash_code(fc, fcp, key, keysize);
 224         hlist_for_each_entry(tfle, &fcp->hash_table[hash], u.hlist) {
 225                 if (tfle->net == net &&
 226                     tfle->family == family &&
 227                     tfle->dir == dir &&
 228                     flow_key_compare(key, &tfle->key, keysize) == 0) {
 229                         fle = tfle;
 230                         break;
 231                 }
 232         }
 233
 234         if (unlikely(!fle)) {
 235                 if (fcp->hash_count > fc->high_watermark)
 236                         flow_cache_shrink(fc, fcp);
 237
 238                 if (atomic_read(&net->xfrm.flow_cache_gc_count) >
 239                     2 * num_online_cpus() * fc->high_watermark) {
 240                         flo = ERR_PTR(-ENOBUFS);
 241                         goto ret_object;
 242                 }
 243
 244                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 245                 if (fle) {
 246                         fle->net = net;
 247                         fle->family = family;
 248                         fle->dir = dir;
 249                         memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
 250                         fle->object = NULL;
 251                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
 252                         fcp->hash_count++;
 253                 }
 254         } else if (likely(fle->genid == atomic_read(&net->xfrm.flow_cache_genid))) {
 255                 flo = fle->object;
 256                 if (!flo)
 257                         goto ret_object;
 258                 flo = flo->ops->get(flo);
 259                 if (flo)
 260                         goto ret_object;
 261         } else if (fle->object) {
 262                 flo = fle->object;
 263                 flo->ops->delete(flo);
 264                 fle->object = NULL;
 265         }
 266
 267 nocache:
 268         flo = NULL;
 269         if (fle) {
 270                 flo = fle->object;
 271                 fle->object = NULL;
 272         }
 273         flo = resolver(net, key, family, dir, flo, ctx);
 274         if (fle) {
 275                 fle->genid = atomic_read(&net->xfrm.flow_cache_genid);
 276                 if (!IS_ERR(flo))
 277                         fle->object = flo;
 278                 else
 279                         fle->genid--;
 280         } else {
 281                 if (!IS_ERR_OR_NULL(flo))
 282                         flo->ops->delete(flo);
 283         }
 284 ret_object:
 285         local_bh_enable();
 286         return flo;
 287 }
 288 EXPORT_SYMBOL(flow_cache_lookup);
 289
 290 static void flow_cache_flush_tasklet(unsigned long data)
 291 {
 292         struct flow_flush_info *info = (void *)data;
 293         struct flow_cache *fc = info->cache;
 294         struct flow_cache_percpu *fcp;
 295         struct flow_cache_entry *fle;
 296         struct hlist_node *tmp;
 297         LIST_HEAD(gc_list);
 298         int i, deleted = 0;
 299         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 300                                                 flow_cache_global);
 301
 302         fcp = this_cpu_ptr(fc->percpu);
 303         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 304                 hlist_for_each_entry_safe(fle, tmp,
 305                                           &fcp->hash_table[i], u.hlist) {
 306                         if (flow_entry_valid(fle, xfrm))
 307                                 continue;
 308
 309                         deleted++;
 310                         hlist_del(&fle->u.hlist);
 311                         list_add_tail(&fle->u.gc_list, &gc_list);
 312                 }
 313         }
 314
 315         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 316
 317         if (atomic_dec_and_test(&info->cpuleft))
 318                 complete(&info->completion);
 319 }
 320
 321 /*
 322  * Return whether a cpu needs flushing.  Conservatively, we assume
 323  * the presence of any entries means the core may require flushing,
 324  * since the flow_cache_ops.check() function may assume it's running
 325  * on the same core as the per-cpu cache component.
 326  */
 327 static int flow_cache_percpu_empty(struct flow_cache *fc, int cpu)
 328 {
 329         struct flow_cache_percpu *fcp;
 330         int i;
 331
 332         fcp = per_cpu_ptr(fc->percpu, cpu);
 333         for (i = 0; i < flow_cache_hash_size(fc); i++)
 334                 if (!hlist_empty(&fcp->hash_table[i]))
 335                         return 0;
 336         return 1;
 337 }
 338
 339 static void flow_cache_flush_per_cpu(void *data)
 340 {
 341         struct flow_flush_info *info = data;
 342         struct tasklet_struct *tasklet;
 343
 344         tasklet = &this_cpu_ptr(info->cache->percpu)->flush_tasklet;
 345         tasklet->data = (unsigned long)info;
 346         tasklet_schedule(tasklet);
 347 }
 348
 349 void flow_cache_flush(struct net *net)
 350 {
 351         struct flow_flush_info info;
 352         cpumask_var_t mask;
 353         int i, self;
 354
 355         /* Track which cpus need flushing to avoid disturbing all cores. */
 356         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 357                 return;
 358         cpumask_clear(mask);
 359
 360         /* Don't want cpus going down or up during this. */
 361         get_online_cpus();
 362         mutex_lock(&net->xfrm.flow_flush_sem);
 363         info.cache = &net->xfrm.flow_cache_global;
 364         for_each_online_cpu(i)
 365                 if (!flow_cache_percpu_empty(info.cache, i))
 366                         cpumask_set_cpu(i, mask);
 367         atomic_set(&info.cpuleft, cpumask_weight(mask));
 368         if (atomic_read(&info.cpuleft) == 0)
 369                 goto done;
 370
 371         init_completion(&info.completion);
 372
 373         local_bh_disable();
 374         self = cpumask_test_and_clear_cpu(smp_processor_id(), mask);
 375         on_each_cpu_mask(mask, flow_cache_flush_per_cpu, &info, 0);
 376         if (self)
 377                 flow_cache_flush_tasklet((unsigned long)&info);
 378         local_bh_enable();
 379
 380         wait_for_completion(&info.completion);
 381
 382 done:
 383         mutex_unlock(&net->xfrm.flow_flush_sem);
 384         put_online_cpus();
 385         free_cpumask_var(mask);
 386 }
 387
 388 static void flow_cache_flush_task(struct work_struct *work)
 389 {
 390         struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
 391                                                 flow_cache_flush_work);
 392         struct net *net = container_of(xfrm, struct net, xfrm);
 393
 394         flow_cache_flush(net);
 395 }
 396
 397 void flow_cache_flush_deferred(struct net *net)
 398 {
 399         schedule_work(&net->xfrm.flow_cache_flush_work);
 400 }
 401
 402 static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 403 {
 404         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 405         size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
 406
 407         if (!fcp->hash_table) {
 408                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
 409                 if (!fcp->hash_table) {
 410                         pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
 411                         return -ENOMEM;
 412                 }
 413                 fcp->hash_rnd_recalc = 1;
 414                 fcp->hash_count = 0;
 415                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 416         }
 417         return 0;
 418 }
 419
 420 static int flow_cache_cpu_up_prep(unsigned int cpu, struct hlist_node *node)
 421 {
 422         struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
 423
 424         return flow_cache_cpu_prepare(fc, cpu);
 425 }
 426
 427 static int flow_cache_cpu_dead(unsigned int cpu, struct hlist_node *node)
 428 {
 429         struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
 430         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 431
 432         __flow_cache_shrink(fc, fcp, 0);
 433         return 0;
 434 }
 435
 436 int flow_cache_init(struct net *net)
 437 {
 438         int i;
 439         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 440
 441         if (!flow_cachep)
 442                 flow_cachep = kmem_cache_create("flow_cache",
 443                                                 sizeof(struct flow_cache_entry),
 444                                                 0, SLAB_PANIC, NULL);
 445         spin_lock_init(&net->xfrm.flow_cache_gc_lock);
 446         INIT_LIST_HEAD(&net->xfrm.flow_cache_gc_list);
 447         INIT_WORK(&net->xfrm.flow_cache_gc_work, flow_cache_gc_task);
 448         INIT_WORK(&net->xfrm.flow_cache_flush_work, flow_cache_flush_task);
 449         mutex_init(&net->xfrm.flow_flush_sem);
 450         atomic_set(&net->xfrm.flow_cache_gc_count, 0);
 451
 452         fc->hash_shift = 10;
 453         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 454         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 455
 456         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 457         if (!fc->percpu)
 458                 return -ENOMEM;
 459
 460         if (cpuhp_state_add_instance(CPUHP_NET_FLOW_PREPARE, &fc->node))
 461                 goto err;
 462
 463         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 464                     (unsigned long) fc);
 465         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 466         add_timer(&fc->rnd_timer);
 467
 468         return 0;
 469
 470 err:
 471         for_each_possible_cpu(i) {
 472                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 473                 kfree(fcp->hash_table);
 474                 fcp->hash_table = NULL;
 475         }
 476
 477         free_percpu(fc->percpu);
 478         fc->percpu = NULL;
 479
 480         return -ENOMEM;
 481 }
 482 EXPORT_SYMBOL(flow_cache_init);
 483
 484 void flow_cache_fini(struct net *net)
 485 {
 486         int i;
 487         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 488
 489         del_timer_sync(&fc->rnd_timer);
 490
 491         cpuhp_state_remove_instance_nocalls(CPUHP_NET_FLOW_PREPARE, &fc->node);
 492
 493         for_each_possible_cpu(i) {
 494                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 495                 kfree(fcp->hash_table);
 496                 fcp->hash_table = NULL;
 497         }
 498
 499         free_percpu(fc->percpu);
 500         fc->percpu = NULL;
 501 }
 502 EXPORT_SYMBOL(flow_cache_fini);
 503
 504 void __init flow_cache_hp_init(void)
 505 {
 506         int ret;
 507
 508         ret = cpuhp_setup_state_multi(CPUHP_NET_FLOW_PREPARE,
 509                                       "net/flow:prepare",
 510                                       flow_cache_cpu_up_prep,
 511                                       flow_cache_cpu_dead);
 512         WARN_ON(ret < 0);
 513 }