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
  28 struct flow_cache_entry {
  29         union {
  30                 struct hlist_node       hlist;
  31                 struct list_head        gc_list;
  32         } u;
  33         struct net                      *net;
  34         u16                             family;
  35         u8                              dir;
  36         u32                             genid;
  37         struct flowi                    key;
  38         struct flow_cache_object        *object;
  39 };
  40
  41 struct flow_cache_percpu {
  42         struct hlist_head               *hash_table;
  43         int                             hash_count;
  44         u32                             hash_rnd;
  45         int                             hash_rnd_recalc;
  46         struct tasklet_struct           flush_tasklet;
  47 };
  48
  49 struct flow_flush_info {
  50         struct flow_cache               *cache;
  51         atomic_t                        cpuleft;
  52         struct completion               completion;
  53 };
  54
  55 struct flow_cache {
  56         u32                             hash_shift;
  57         struct flow_cache_percpu __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 __read_mostly;
  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 *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, 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                           const struct flowi *key,
 177                           size_t keysize)
 178 {
 179         const u32 *k = (const u32 *) key;
 180         const u32 length = keysize * sizeof(flow_compare_t) / sizeof(u32);
 181
 182         return jhash2(k, length, fcp->hash_rnd)
 183                 & (flow_cache_hash_size(fc) - 1);
 184 }
 185
 186 /* I hear what you're saying, use memcmp.  But memcmp cannot make
 187  * important assumptions that we can here, such as alignment.
 188  */
 189 static int flow_key_compare(const struct flowi *key1, const struct flowi *key2,
 190                             size_t keysize)
 191 {
 192         const flow_compare_t *k1, *k1_lim, *k2;
 193
 194         k1 = (const flow_compare_t *) key1;
 195         k1_lim = k1 + keysize;
 196
 197         k2 = (const flow_compare_t *) key2;
 198
 199         do {
 200                 if (*k1++ != *k2++)
 201                         return 1;
 202         } while (k1 < k1_lim);
 203
 204         return 0;
 205 }
 206
 207 struct flow_cache_object *
 208 flow_cache_lookup(struct net *net, const struct flowi *key, u16 family, u8 dir,
 209                   flow_resolve_t resolver, void *ctx)
 210 {
 211         struct flow_cache *fc = &flow_cache_global;
 212         struct flow_cache_percpu *fcp;
 213         struct flow_cache_entry *fle, *tfle;
 214         struct flow_cache_object *flo;
 215         size_t keysize;
 216         unsigned int hash;
 217
 218         local_bh_disable();
 219         fcp = this_cpu_ptr(fc->percpu);
 220
 221         fle = NULL;
 222         flo = NULL;
 223
 224         keysize = flow_key_size(family);
 225         if (!keysize)
 226                 goto nocache;
 227
 228         /* Packet really early in init?  Making flow_cache_init a
 229          * pre-smp initcall would solve this.  --RR */
 230         if (!fcp->hash_table)
 231                 goto nocache;
 232
 233         if (fcp->hash_rnd_recalc)
 234                 flow_new_hash_rnd(fc, fcp);
 235
 236         hash = flow_hash_code(fc, fcp, key, keysize);
 237         hlist_for_each_entry(tfle, &fcp->hash_table[hash], u.hlist) {
 238                 if (tfle->net == net &&
 239                     tfle->family == family &&
 240                     tfle->dir == dir &&
 241                     flow_key_compare(key, &tfle->key, keysize) == 0) {
 242                         fle = tfle;
 243                         break;
 244                 }
 245         }
 246
 247         if (unlikely(!fle)) {
 248                 if (fcp->hash_count > fc->high_watermark)
 249                         flow_cache_shrink(fc, fcp);
 250
 251                 fle = kmem_cache_alloc(flow_cachep, GFP_ATOMIC);
 252                 if (fle) {
 253                         fle->net = net;
 254                         fle->family = family;
 255                         fle->dir = dir;
 256                         memcpy(&fle->key, key, keysize * sizeof(flow_compare_t));
 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 (!IS_ERR_OR_NULL(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 *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, 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 /*
 327  * Return whether a cpu needs flushing.  Conservatively, we assume
 328  * the presence of any entries means the core may require flushing,
 329  * since the flow_cache_ops.check() function may assume it's running
 330  * on the same core as the per-cpu cache component.
 331  */
 332 static int flow_cache_percpu_empty(struct flow_cache *fc, int cpu)
 333 {
 334         struct flow_cache_percpu *fcp;
 335         int i;
 336
 337         fcp = per_cpu_ptr(fc->percpu, cpu);
 338         for (i = 0; i < flow_cache_hash_size(fc); i++)
 339                 if (!hlist_empty(&fcp->hash_table[i]))
 340                         return 0;
 341         return 1;
 342 }
 343
 344 static void flow_cache_flush_per_cpu(void *data)
 345 {
 346         struct flow_flush_info *info = data;
 347         struct tasklet_struct *tasklet;
 348
 349         tasklet = &this_cpu_ptr(info->cache->percpu)->flush_tasklet;
 350         tasklet->data = (unsigned long)info;
 351         tasklet_schedule(tasklet);
 352 }
 353
 354 void flow_cache_flush(void)
 355 {
 356         struct flow_flush_info info;
 357         static DEFINE_MUTEX(flow_flush_sem);
 358         cpumask_var_t mask;
 359         int i, self;
 360
 361         /* Track which cpus need flushing to avoid disturbing all cores. */
 362         if (!alloc_cpumask_var(&mask, GFP_KERNEL))
 363                 return;
 364         cpumask_clear(mask);
 365
 366         /* Don't want cpus going down or up during this. */
 367         get_online_cpus();
 368         mutex_lock(&flow_flush_sem);
 369         info.cache = &flow_cache_global;
 370         for_each_online_cpu(i)
 371                 if (!flow_cache_percpu_empty(info.cache, i))
 372                         cpumask_set_cpu(i, mask);
 373         atomic_set(&info.cpuleft, cpumask_weight(mask));
 374         if (atomic_read(&info.cpuleft) == 0)
 375                 goto done;
 376
 377         init_completion(&info.completion);
 378
 379         local_bh_disable();
 380         self = cpumask_test_and_clear_cpu(smp_processor_id(), mask);
 381         on_each_cpu_mask(mask, flow_cache_flush_per_cpu, &info, 0);
 382         if (self)
 383                 flow_cache_flush_tasklet((unsigned long)&info);
 384         local_bh_enable();
 385
 386         wait_for_completion(&info.completion);
 387
 388 done:
 389         mutex_unlock(&flow_flush_sem);
 390         put_online_cpus();
 391         free_cpumask_var(mask);
 392 }
 393
 394 static void flow_cache_flush_task(struct work_struct *work)
 395 {
 396         flow_cache_flush();
 397 }
 398
 399 static DECLARE_WORK(flow_cache_flush_work, flow_cache_flush_task);
 400
 401 void flow_cache_flush_deferred(void)
 402 {
 403         schedule_work(&flow_cache_flush_work);
 404 }
 405
 406 static int flow_cache_cpu_prepare(struct flow_cache *fc, int cpu)
 407 {
 408         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 409         size_t sz = sizeof(struct hlist_head) * flow_cache_hash_size(fc);
 410
 411         if (!fcp->hash_table) {
 412                 fcp->hash_table = kzalloc_node(sz, GFP_KERNEL, cpu_to_node(cpu));
 413                 if (!fcp->hash_table) {
 414                         pr_err("NET: failed to allocate flow cache sz %zu\n", sz);
 415                         return -ENOMEM;
 416                 }
 417                 fcp->hash_rnd_recalc = 1;
 418                 fcp->hash_count = 0;
 419                 tasklet_init(&fcp->flush_tasklet, flow_cache_flush_tasklet, 0);
 420         }
 421         return 0;
 422 }
 423
 424 static int flow_cache_cpu(struct notifier_block *nfb,
 425                           unsigned long action,
 426                           void *hcpu)
 427 {
 428         struct flow_cache *fc = container_of(nfb, struct flow_cache, hotcpu_notifier);
 429         int res, cpu = (unsigned long) hcpu;
 430         struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, cpu);
 431
 432         switch (action) {
 433         case CPU_UP_PREPARE:
 434         case CPU_UP_PREPARE_FROZEN:
 435                 res = flow_cache_cpu_prepare(fc, cpu);
 436                 if (res)
 437                         return notifier_from_errno(res);
 438                 break;
 439         case CPU_DEAD:
 440         case CPU_DEAD_FROZEN:
 441                 __flow_cache_shrink(fc, fcp, 0);
 442                 break;
 443         }
 444         return NOTIFY_OK;
 445 }
 446
 447 static int __init flow_cache_init(struct flow_cache *fc)
 448 {
 449         int i;
 450
 451         fc->hash_shift = 10;
 452         fc->low_watermark = 2 * flow_cache_hash_size(fc);
 453         fc->high_watermark = 4 * flow_cache_hash_size(fc);
 454
 455         fc->percpu = alloc_percpu(struct flow_cache_percpu);
 456         if (!fc->percpu)
 457                 return -ENOMEM;
 458
 459         for_each_online_cpu(i) {
 460                 if (flow_cache_cpu_prepare(fc, i))
 461                         goto err;
 462         }
 463         fc->hotcpu_notifier = (struct notifier_block){
 464                 .notifier_call = flow_cache_cpu,
 465         };
 466         register_hotcpu_notifier(&fc->hotcpu_notifier);
 467
 468         setup_timer(&fc->rnd_timer, flow_cache_new_hashrnd,
 469                     (unsigned long) fc);
 470         fc->rnd_timer.expires = jiffies + FLOW_HASH_RND_PERIOD;
 471         add_timer(&fc->rnd_timer);
 472
 473         return 0;
 474
 475 err:
 476         for_each_possible_cpu(i) {
 477                 struct flow_cache_percpu *fcp = per_cpu_ptr(fc->percpu, i);
 478                 kfree(fcp->hash_table);
 479                 fcp->hash_table = NULL;
 480         }
 481
 482         free_percpu(fc->percpu);
 483         fc->percpu = NULL;
 484
 485         return -ENOMEM;
 486 }
 487
 488 static int __init flow_cache_init_global(void)
 489 {
 490         flow_cachep = kmem_cache_create("flow_cache",
 491                                         sizeof(struct flow_cache_entry),
 492                                         0, SLAB_PANIC, NULL);
 493
 494         return flow_cache_init(&flow_cache_global);
 495 }
 496
 497 module_init(flow_cache_init_global);