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)     (1U << (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                                      unsigned int deleted,
 103                                      struct list_head *gc_list,
 104                                      struct netns_xfrm *xfrm)
 105 {
 106         if (deleted) {
 107                 atomic_add(deleted, &xfrm->flow_cache_gc_count);
 108                 fcp->hash_count -= deleted;
 109                 spin_lock_bh(&xfrm->flow_cache_gc_lock);
 110                 list_splice_tail(gc_list, &xfrm->flow_cache_gc_list);
 111                 spin_unlock_bh(&xfrm->flow_cache_gc_lock);
 112                 schedule_work(&xfrm->flow_cache_gc_work);
 113         }
 114 }
 115
 116 static void __flow_cache_shrink(struct flow_cache *fc,
 117                                 struct flow_cache_percpu *fcp,
 118                                 unsigned int shrink_to)
 119 {
 120         struct flow_cache_entry *fle;
 121         struct hlist_node *tmp;
 122         LIST_HEAD(gc_list);
 123         unsigned int deleted = 0;
 124         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 125                                                 flow_cache_global);
 126         unsigned int i;
 127
 128         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 129                 unsigned int saved = 0;
 130
 131                 hlist_for_each_entry_safe(fle, tmp,
 132                                           &fcp->hash_table[i], u.hlist) {
 133                         if (saved < shrink_to &&
 134                             flow_entry_valid(fle, xfrm)) {
 135                                 saved++;
 136                         } else {
 137                                 deleted++;
 138                                 hlist_del(&fle->u.hlist);
 139                                 list_add_tail(&fle->u.gc_list, &gc_list);
 140                         }
 141                 }
 142         }
 143
 144         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 145 }
 146
 147 static void flow_cache_shrink(struct flow_cache *fc,
 148                               struct flow_cache_percpu *fcp)
 149 {
 150         unsigned int shrink_to = fc->low_watermark / flow_cache_hash_size(fc);
 151
 152         __flow_cache_shrink(fc, fcp, shrink_to);
 153 }
 154
 155 static void flow_new_hash_rnd(struct flow_cache *fc,
 156                               struct flow_cache_percpu *fcp)
 157 {
 158         get_random_bytes(&fcp->hash_rnd, sizeof(u32));
 159         fcp->hash_rnd_recalc = 0;
 160         __flow_cache_shrink(fc, fcp, 0);
 161 }
 162
 163 static u32 flow_hash_code(struct flow_cache *fc,
 164                           struct flow_cache_percpu *fcp,
 165                           const struct flowi *key,
 166                           unsigned int keysize)
 167 {
 168         const u32 *k = (const u32 *) key;
 169         const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
 170
 171         return jhash2(k, length, fcp->hash_rnd)
 172                 & (flow_cache_hash_size(fc) - 1);
 173 }
 174
 175 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 176  * important assumptions that we can here, such as alignment.
 177  */
 178 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
 179                             unsigned int keysize)
 180 {
 181         const flow_compare_t *k1, *k1_lim, *k2;
 182
 183         k1 = (const flow_compare_t *) key1;
 184         k1_lim = k1 + keysize;
 185
 186         k2 = (const flow_compare_t *) key2;
 187
 188         do {
 189                 if (*k1++ != *k2++)
 190                         return 1;
 191         } while (k1 < k1_lim);
 192
 193         return 0;
 194 }
 195
 196 struct flow_cache_object *
 197 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
 198                   flow_resolve_t resolver, void *ctx)
 199 {
 200         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 201         struct flow_cache_percpu *fcp;
 202         struct flow_cache_entry *fle, *tfle;
 203         struct flow_cache_object *flo;
 204         unsigned int keysize;
 205         unsigned int hash;
 206
 207         local_bh_disable();
 208         fcp = this_cpu_ptr(fc->percpu);
 209
 210         fle = NULL;
 211         flo = NULL;
 212
 213         keysize = flow_key_size(family);
 214         if (!keysize)
 215                 goto nocache;
 216
 217         /* Packet really early in init?  Making flow_cache_init a
 218          * pre-smp initcall would solve this.  --RR */
 219         if (!fcp->hash_table)
 220                 goto nocache;
 221
 222         if (fcp->hash_rnd_recalc)
 223                 flow_new_hash_rnd(fc, fcp);
 224
 225         hash = flow_hash_code(fc, fcp, key, keysize);
 226         hlist_for_each_entry(tfle, &fcp->hash_table[hash], u.hlist) {
 227                 if (tfle->net == net &&
 228                     tfle->family == family &&
 229                     tfle->dir == dir &&
 230                     flow_key_compare(key, &tfle->key, keysize) == 0) {
 231                         fle = tfle;
 232                         break;
 233                 }
 234         }
 235
 236         if (unlikely(!fle)) {
 237                 if (fcp->hash_count > fc->high_watermark)
 238                         flow_cache_shrink(fc, fcp);
 239
 240                 if (atomic_read(&net->xfrm.flow_cache_gc_count) >
 241                     2 * num_online_cpus() * fc->high_watermark) {
 242                         flo = ERR_PTR(-ENOBUFS);
 243                         goto ret_object;
 244                 }
 245
 246                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 247                 if (fle) {
 248                         fle->net = net;
 249                         fle->family = family;
 250                         fle->dir = dir;
 251                         memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
 252                         fle->object = NULL;
 253                         hlist_add_head(&fle->u.hlist, &fcp->hash_table[hash]);
 254                         fcp->hash_count++;
 255                 }
 256         } else if (likely(fle->genid == atomic_read(&net->xfrm.flow_cache_genid))) {
 257                 flo = fle->object;
 258                 if (!flo)
 259                         goto ret_object;
 260                 flo = flo->ops->get(flo);
 261                 if (flo)
 262                         goto ret_object;
 263         } else if (fle->object) {
 264                 flo = fle->object;
 265                 flo->ops->delete(flo);
 266                 fle->object = NULL;
 267         }
 268
 269 nocache:
 270         flo = NULL;
 271         if (fle) {
 272                 flo = fle->object;
 273                 fle->object = NULL;
 274         }
 275         flo = resolver(net, key, family, dir, flo, ctx);
 276         if (fle) {
 277                 fle->genid = atomic_read(&net->xfrm.flow_cache_genid);
 278                 if (!IS_ERR(flo))
 279                         fle->object = flo;
 280                 else
 281                         fle->genid--;
 282         } else {
 283                 if (!IS_ERR_OR_NULL(flo))
 284                         flo->ops->delete(flo);
 285         }
 286 ret_object:
 287         local_bh_enable();
 288         return flo;
 289 }
 290 EXPORT_SYMBOL(flow_cache_lookup);
 291
 292 static void flow_cache_flush_tasklet(unsigned long data)
 293 {
 294         struct flow_flush_info *info = (void *)data;
 295         struct flow_cache *fc = info->cache;
 296         struct flow_cache_percpu *fcp;
 297         struct flow_cache_entry *fle;
 298         struct hlist_node *tmp;
 299         LIST_HEAD(gc_list);
 300         unsigned int deleted = 0;
 301         struct netns_xfrm *xfrm = container_of(fc, struct netns_xfrm,
 302                                                 flow_cache_global);
 303         unsigned int i;
 304
 305         fcp = this_cpu_ptr(fc->percpu);
 306         for (i = 0; i < flow_cache_hash_size(fc); i++) {
 307                 hlist_for_each_entry_safe(fle, tmp,
 308                                           &fcp->hash_table[i], u.hlist) {
 309                         if (flow_entry_valid(fle, xfrm))
 310                                 continue;
 311
 312                         deleted++;
 313                         hlist_del(&fle->u.hlist);
 314                         list_add_tail(&fle->u.gc_list, &gc_list);
 315                 }
 316         }
 317
 318         flow_cache_queue_garbage(fcp, deleted, &gc_list, xfrm);
 319
 320         if (atomic_dec_and_test(&info->cpuleft))
 321                 complete(&info->completion);
 322 }
 323
 324 /*
 325  * Return whether a cpu needs flushing.  Conservatively, we assume
 326  * the presence of any entries means the core may require flushing,
 327  * since the flow_cache_ops.check() function may assume it's running
 328  * on the same core as the per-cpu cache component.
 329  */
 330 static int flow_cache_percpu_empty(struct flow_cache *fc, int cpu)
 331 {
 332         struct flow_cache_percpu *fcp;
 333         unsigned int i;
 334
 335         fcp = per_cpu_ptr(fc->percpu, cpu);
 336         for (i = 0; i < flow_cache_hash_size(fc); i++)
 337                 if (!hlist_empty(&fcp->hash_table[i]))
 338                         return 0;
 339         return 1;
 340 }
 341
 342 static void flow_cache_flush_per_cpu(void *data)
 343 {
 344         struct flow_flush_info *info = data;
 345         struct tasklet_struct *tasklet;
 346
 347         tasklet = &this_cpu_ptr(info->cache->percpu)->flush_tasklet;
 348         tasklet->data = (unsigned long)info;
 349         tasklet_schedule(tasklet);
 350 }
 351
 352 void flow_cache_flush(struct net *net)
 353 {
 354         struct flow_flush_info info;
 355         cpumask_var_t mask;
 356         int i, self;
 357
 358         /* Track which cpus need flushing to avoid disturbing all cores. */
 359         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 360                 return;
 361         cpumask_clear(mask);
 362
 363         /* Don't want cpus going down or up during this. */
 364         get_online_cpus();
 365         mutex_lock(&net->xfrm.flow_flush_sem);
 366         info.cache = &net->xfrm.flow_cache_global;
 367         for_each_online_cpu(i)
 368                 if (!flow_cache_percpu_empty(info.cache, i))
 369                         cpumask_set_cpu(i, mask);
 370         atomic_set(&info.cpuleft, cpumask_weight(mask));
 371         if (atomic_read(&info.cpuleft) == 0)
 372                 goto done;
 373
 374         init_completion(&info.completion);
 375
 376         local_bh_disable();
 377         self = cpumask_test_and_clear_cpu(smp_processor_id(), mask);
 378         on_each_cpu_mask(mask, flow_cache_flush_per_cpu, &info, 0);
 379         if (self)
 380                 flow_cache_flush_tasklet((unsigned long)&info);
 381         local_bh_enable();
 382
 383         wait_for_completion(&info.completion);
 384
 385 done:
 386         mutex_unlock(&net->xfrm.flow_flush_sem);
 387         put_online_cpus();
 388         free_cpumask_var(mask);
 389 }
 390
 391 static void flow_cache_flush_task(struct work_struct *work)
 392 {
 393         struct netns_xfrm *xfrm = container_of(work, struct netns_xfrm,
 394                                                 flow_cache_flush_work);
 395         struct net *net = container_of(xfrm, struct net, xfrm);
 396
 397         flow_cache_flush(net);
 398 }
 399
 400 void flow_cache_flush_deferred(struct net *net)
 401 {
 402         schedule_work(&net->xfrm.flow_cache_flush_work);
 403 }
 404
 405 static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 406 {
 407         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 408         unsigned int sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
 409
 410         if (!fcp->hash_table) {
 411                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
 412                 if (!fcp->hash_table) {
 413                         pr_err("NET: failed to allocate flow cache sz %u\n", sz);
 414                         return -ENOMEM;
 415                 }
 416                 fcp->hash_rnd_recalc = 1;
 417                 fcp->hash_count = 0;
 418                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 419         }
 420         return 0;
 421 }
 422
 423 static int flow_cache_cpu_up_prep(unsigned int cpu, struct hlist_node *node)
 424 {
 425         struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
 426
 427         return flow_cache_cpu_prepare(fc, cpu);
 428 }
 429
 430 static int flow_cache_cpu_dead(unsigned int cpu, struct hlist_node *node)
 431 {
 432         struct flow_cache *fc = hlist_entry_safe(node, struct flow_cache, node);
 433         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 434
 435         __flow_cache_shrink(fc, fcp, 0);
 436         return 0;
 437 }
 438
 439 int flow_cache_init(struct net *net)
 440 {
 441         int i;
 442         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 443
 444         if (!flow_cachep)
 445                 flow_cachep = kmem_cache_create("flow_cache",
 446                                                 sizeof(struct flow_cache_entry),
 447                                                 0, SLAB_PANIC, NULL);
 448         spin_lock_init(&net->xfrm.flow_cache_gc_lock);
 449         INIT_LIST_HEAD(&net->xfrm.flow_cache_gc_list);
 450         INIT_WORK(&net->xfrm.flow_cache_gc_work, flow_cache_gc_task);
 451         INIT_WORK(&net->xfrm.flow_cache_flush_work, flow_cache_flush_task);
 452         mutex_init(&net->xfrm.flow_flush_sem);
 453         atomic_set(&net->xfrm.flow_cache_gc_count, 0);
 454
 455         fc->hash_shift = 10;
 456         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 457         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 458
 459         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 460         if (!fc->percpu)
 461                 return -ENOMEM;
 462
 463         if (cpuhp_state_add_instance(CPUHP_NET_FLOW_PREPARE, &fc->node))
 464                 goto err;
 465
 466         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 467                     (unsigned long) fc);
 468         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 469         add_timer(&fc->rnd_timer);
 470
 471         return 0;
 472
 473 err:
 474         for_each_possible_cpu(i) {
 475                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 476                 kfree(fcp->hash_table);
 477                 fcp->hash_table = NULL;
 478         }
 479
 480         free_percpu(fc->percpu);
 481         fc->percpu = NULL;
 482
 483         return -ENOMEM;
 484 }
 485 EXPORT_SYMBOL(flow_cache_init);
 486
 487 void flow_cache_fini(struct net *net)
 488 {
 489         int i;
 490         struct flow_cache *fc = &net->xfrm.flow_cache_global;
 491
 492         del_timer_sync(&fc->rnd_timer);
 493
 494         cpuhp_state_remove_instance_nocalls(CPUHP_NET_FLOW_PREPARE, &fc->node);
 495
 496         for_each_possible_cpu(i) {
 497                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 498                 kfree(fcp->hash_table);
 499                 fcp->hash_table = NULL;
 500         }
 501
 502         free_percpu(fc->percpu);
 503         fc->percpu = NULL;
 504 }
 505 EXPORT_SYMBOL(flow_cache_fini);
 506
 507 void __init flow_cache_hp_init(void)
 508 {
 509         int ret;
 510
 511         ret = cpuhp_setup_state_multi(CPUHP_NET_FLOW_PREPARE,
 512                                       "net/flow:prepare",
 513                                       flow_cache_cpu_up_prep,
 514                                       flow_cache_cpu_dead);
 515         WARN_ON(ret < 0);
 516 }