net/ipv4/inet_fragment.c

   1 /*
   2  * inet fragments management
   3  *
   4  *              This program is free software; you can redistribute it and/or
   5  *              modify it under the terms of the GNU General Public License
   6  *              as published by the Free Software Foundation; either version
   7  *              2 of the License, or (at your option) any later version.
   8  *
   9  *              Authors:        Pavel Emelyanov <xemul@openvz.org>
  10  *                              Started as consolidation of ipv4/ip_fragment.c,
  11  *                              ipv6/reassembly. and ipv6 nf conntrack reassembly
  12  */
  13
  14 #include <linux/list.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/module.h>
  17 #include <linux/timer.h>
  18 #include <linux/mm.h>
  19 #include <linux/random.h>
  20 #include <linux/skbuff.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/slab.h>
  23
  24 #include <net/sock.h>
  25 #include <net/inet_frag.h>
  26 #include <net/inet_ecn.h>
  27
  28 #define INETFRAGS_EVICT_BUCKETS   128
  29 #define INETFRAGS_EVICT_MAX       512
  30
  31 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
  32  * Value : 0xff if frame should be dropped.
  33  *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
  34  */
  35 const u8 ip_frag_ecn_table[16] = {
  36         /* at least one fragment had CE, and others ECT_0 or ECT_1 */
  37         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
  38         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
  39         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
  40
  41         /* invalid combinations : drop frame */
  42         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
  43         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
  44         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
  45         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  46         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
  47         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
  48         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  49 };
  50 EXPORT_SYMBOL(ip_frag_ecn_table);
  51
  52 static unsigned int
  53 inet_frag_hashfn(const struct inet_frags *f, const struct inet_frag_queue *q)
  54 {
  55         return f->hashfn(q) & (INETFRAGS_HASHSZ - 1);
  56 }
  57
  58 static void inet_frag_secret_rebuild(unsigned long dummy)
  59 {
  60         struct inet_frags *f = (struct inet_frags *)dummy;
  61         unsigned long now = jiffies;
  62         int i;
  63
  64         /* Per bucket lock NOT needed here, due to write lock protection */
  65         write_lock(&f->lock);
  66
  67         get_random_bytes(&f->rnd, sizeof(u32));
  68         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
  69                 struct inet_frag_bucket *hb;
  70                 struct inet_frag_queue *q;
  71                 struct hlist_node *n;
  72
  73                 hb = &f->hash[i];
  74                 hlist_for_each_entry_safe(q, n, &hb->chain, list) {
  75                         unsigned int hval = inet_frag_hashfn(f, q);
  76
  77                         if (hval != i) {
  78                                 struct inet_frag_bucket *hb_dest;
  79
  80                                 hlist_del(&q->list);
  81
  82                                 /* Relink to new hash chain. */
  83                                 hb_dest = &f->hash[hval];
  84                                 hlist_add_head(&q->list, &hb_dest->chain);
  85                         }
  86                 }
  87         }
  88         write_unlock(&f->lock);
  89
  90         mod_timer(&f->secret_timer, now + f->secret_interval);
  91 }
  92
  93 static bool inet_fragq_should_evict(const struct inet_frag_queue *q)
  94 {
  95         return q->net->low_thresh == 0 ||
  96                frag_mem_limit(q->net) >= q->net->low_thresh;
  97 }
  98
  99 static unsigned int
 100 inet_evict_bucket(struct inet_frags *f, struct inet_frag_bucket *hb)
 101 {
 102         struct inet_frag_queue *fq;
 103         struct hlist_node *n;
 104         unsigned int evicted = 0;
 105         HLIST_HEAD(expired);
 106
 107 evict_again:
 108         spin_lock(&hb->chain_lock);
 109
 110         hlist_for_each_entry_safe(fq, n, &hb->chain, list) {
 111                 if (!inet_fragq_should_evict(fq))
 112                         continue;
 113
 114                 if (!del_timer(&fq->timer)) {
 115                         /* q expiring right now thus increment its refcount so
 116                          * it won't be freed under us and wait until the timer
 117                          * has finished executing then destroy it
 118                          */
 119                         atomic_inc(&fq->refcnt);
 120                         spin_unlock(&hb->chain_lock);
 121                         del_timer_sync(&fq->timer);
 122                         WARN_ON(atomic_read(&fq->refcnt) != 1);
 123                         inet_frag_put(fq, f);
 124                         goto evict_again;
 125                 }
 126
 127                 /* suppress xmit of (icmp) error packet */
 128                 fq->last_in &= ~INET_FRAG_FIRST_IN;
 129                 fq->last_in |= INET_FRAG_EVICTED;
 130                 hlist_del(&fq->list);
 131                 hlist_add_head(&fq->list, &expired);
 132                 ++evicted;
 133         }
 134
 135         spin_unlock(&hb->chain_lock);
 136
 137         hlist_for_each_entry_safe(fq, n, &expired, list)
 138                 f->frag_expire((unsigned long) fq);
 139
 140         return evicted;
 141 }
 142
 143 static void inet_frag_worker(struct work_struct *work)
 144 {
 145         unsigned int budget = INETFRAGS_EVICT_BUCKETS;
 146         unsigned int i, evicted = 0;
 147         struct inet_frags *f;
 148
 149         f = container_of(work, struct inet_frags, frags_work);
 150
 151         BUILD_BUG_ON(INETFRAGS_EVICT_BUCKETS >= INETFRAGS_HASHSZ);
 152
 153         read_lock_bh(&f->lock);
 154
 155         for (i = ACCESS_ONCE(f->next_bucket); budget; --budget) {
 156                 evicted += inet_evict_bucket(f, &f->hash[i]);
 157                 i = (i + 1) & (INETFRAGS_HASHSZ - 1);
 158                 if (evicted > INETFRAGS_EVICT_MAX)
 159                         break;
 160         }
 161
 162         f->next_bucket = i;
 163
 164         read_unlock_bh(&f->lock);
 165 }
 166
 167 static void inet_frag_schedule_worker(struct inet_frags *f)
 168 {
 169         if (unlikely(!work_pending(&f->frags_work)))
 170                 schedule_work(&f->frags_work);
 171 }
 172
 173 void inet_frags_init(struct inet_frags *f)
 174 {
 175         int i;
 176
 177         INIT_WORK(&f->frags_work, inet_frag_worker);
 178
 179         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
 180                 struct inet_frag_bucket *hb = &f->hash[i];
 181
 182                 spin_lock_init(&hb->chain_lock);
 183                 INIT_HLIST_HEAD(&hb->chain);
 184         }
 185         rwlock_init(&f->lock);
 186
 187         setup_timer(&f->secret_timer, inet_frag_secret_rebuild,
 188                         (unsigned long)f);
 189         f->secret_timer.expires = jiffies + f->secret_interval;
 190         add_timer(&f->secret_timer);
 191 }
 192 EXPORT_SYMBOL(inet_frags_init);
 193
 194 void inet_frags_init_net(struct netns_frags *nf)
 195 {
 196         init_frag_mem_limit(nf);
 197         INIT_LIST_HEAD(&nf->lru_list);
 198         spin_lock_init(&nf->lru_lock);
 199 }
 200 EXPORT_SYMBOL(inet_frags_init_net);
 201
 202 void inet_frags_fini(struct inet_frags *f)
 203 {
 204         del_timer(&f->secret_timer);
 205         cancel_work_sync(&f->frags_work);
 206 }
 207 EXPORT_SYMBOL(inet_frags_fini);
 208
 209 void inet_frags_exit_net(struct netns_frags *nf, struct inet_frags *f)
 210 {
 211         int i;
 212
 213         nf->low_thresh = 0;
 214
 215         read_lock_bh(&f->lock);
 216
 217         for (i = 0; i < INETFRAGS_HASHSZ ; i++)
 218                 inet_evict_bucket(f, &f->hash[i]);
 219
 220         read_unlock_bh(&f->lock);
 221
 222         percpu_counter_destroy(&nf->mem);
 223 }
 224 EXPORT_SYMBOL(inet_frags_exit_net);
 225
 226 static inline void fq_unlink(struct inet_frag_queue *fq, struct inet_frags *f)
 227 {
 228         struct inet_frag_bucket *hb;
 229         unsigned int hash;
 230
 231         read_lock(&f->lock);
 232         hash = inet_frag_hashfn(f, fq);
 233         hb = &f->hash[hash];
 234
 235         spin_lock(&hb->chain_lock);
 236         hlist_del(&fq->list);
 237         spin_unlock(&hb->chain_lock);
 238
 239         read_unlock(&f->lock);
 240         inet_frag_lru_del(fq);
 241 }
 242
 243 void inet_frag_kill(struct inet_frag_queue *fq, struct inet_frags *f)
 244 {
 245         if (del_timer(&fq->timer))
 246                 atomic_dec(&fq->refcnt);
 247
 248         if (!(fq->last_in & INET_FRAG_COMPLETE)) {
 249                 fq_unlink(fq, f);
 250                 atomic_dec(&fq->refcnt);
 251                 fq->last_in |= INET_FRAG_COMPLETE;
 252         }
 253 }
 254 EXPORT_SYMBOL(inet_frag_kill);
 255
 256 static inline void frag_kfree_skb(struct netns_frags *nf, struct inet_frags *f,
 257                 struct sk_buff *skb)
 258 {
 259         if (f->skb_free)
 260                 f->skb_free(skb);
 261         kfree_skb(skb);
 262 }
 263
 264 void inet_frag_destroy(struct inet_frag_queue *q, struct inet_frags *f,
 265                                         int *work)
 266 {
 267         struct sk_buff *fp;
 268         struct netns_frags *nf;
 269         unsigned int sum, sum_truesize = 0;
 270
 271         WARN_ON(!(q->last_in & INET_FRAG_COMPLETE));
 272         WARN_ON(del_timer(&q->timer) != 0);
 273
 274         /* Release all fragment data. */
 275         fp = q->fragments;
 276         nf = q->net;
 277         while (fp) {
 278                 struct sk_buff *xp = fp->next;
 279
 280                 sum_truesize += fp->truesize;
 281                 frag_kfree_skb(nf, f, fp);
 282                 fp = xp;
 283         }
 284         sum = sum_truesize + f->qsize;
 285         if (work)
 286                 *work -= sum;
 287         sub_frag_mem_limit(q, sum);
 288
 289         if (f->destructor)
 290                 f->destructor(q);
 291         kfree(q);
 292
 293 }
 294 EXPORT_SYMBOL(inet_frag_destroy);
 295
 296 static struct inet_frag_queue *inet_frag_intern(struct netns_frags *nf,
 297                 struct inet_frag_queue *qp_in, struct inet_frags *f,
 298                 void *arg)
 299 {
 300         struct inet_frag_bucket *hb;
 301         struct inet_frag_queue *qp;
 302         unsigned int hash;
 303
 304         read_lock(&f->lock); /* Protects against hash rebuild */
 305         /*
 306          * While we stayed w/o the lock other CPU could update
 307          * the rnd seed, so we need to re-calculate the hash
 308          * chain. Fortunatelly the qp_in can be used to get one.
 309          */
 310         hash = inet_frag_hashfn(f, qp_in);
 311         hb = &f->hash[hash];
 312         spin_lock(&hb->chain_lock);
 313
 314 #ifdef CONFIG_SMP
 315         /* With SMP race we have to recheck hash table, because
 316          * such entry could be created on other cpu, while we
 317          * released the hash bucket lock.
 318          */
 319         hlist_for_each_entry(qp, &hb->chain, list) {
 320                 if (qp->net == nf && f->match(qp, arg)) {
 321                         atomic_inc(&qp->refcnt);
 322                         spin_unlock(&hb->chain_lock);
 323                         read_unlock(&f->lock);
 324                         qp_in->last_in |= INET_FRAG_COMPLETE;
 325                         inet_frag_put(qp_in, f);
 326                         return qp;
 327                 }
 328         }
 329 #endif
 330         qp = qp_in;
 331         if (!mod_timer(&qp->timer, jiffies + nf->timeout))
 332                 atomic_inc(&qp->refcnt);
 333
 334         atomic_inc(&qp->refcnt);
 335         hlist_add_head(&qp->list, &hb->chain);
 336         inet_frag_lru_add(nf, qp);
 337         spin_unlock(&hb->chain_lock);
 338         read_unlock(&f->lock);
 339
 340         return qp;
 341 }
 342
 343 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
 344                 struct inet_frags *f, void *arg)
 345 {
 346         struct inet_frag_queue *q;
 347
 348         if (frag_mem_limit(nf) > nf->high_thresh) {
 349                 inet_frag_schedule_worker(f);
 350                 return NULL;
 351         }
 352
 353         q = kzalloc(f->qsize, GFP_ATOMIC);
 354         if (q == NULL)
 355                 return NULL;
 356
 357         q->net = nf;
 358         f->constructor(q, arg);
 359         add_frag_mem_limit(q, f->qsize);
 360
 361         setup_timer(&q->timer, f->frag_expire, (unsigned long)q);
 362         spin_lock_init(&q->lock);
 363         atomic_set(&q->refcnt, 1);
 364         INIT_LIST_HEAD(&q->lru_list);
 365
 366         return q;
 367 }
 368
 369 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
 370                 struct inet_frags *f, void *arg)
 371 {
 372         struct inet_frag_queue *q;
 373
 374         q = inet_frag_alloc(nf, f, arg);
 375         if (q == NULL)
 376                 return NULL;
 377
 378         return inet_frag_intern(nf, q, f, arg);
 379 }
 380
 381 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf,
 382                 struct inet_frags *f, void *key, unsigned int hash)
 383         __releases(&f->lock)
 384 {
 385         struct inet_frag_bucket *hb;
 386         struct inet_frag_queue *q;
 387         int depth = 0;
 388
 389         if (frag_mem_limit(nf) > nf->low_thresh)
 390                 inet_frag_schedule_worker(f);
 391
 392         hash &= (INETFRAGS_HASHSZ - 1);
 393         hb = &f->hash[hash];
 394
 395         spin_lock(&hb->chain_lock);
 396         hlist_for_each_entry(q, &hb->chain, list) {
 397                 if (q->net == nf && f->match(q, key)) {
 398                         atomic_inc(&q->refcnt);
 399                         spin_unlock(&hb->chain_lock);
 400                         read_unlock(&f->lock);
 401                         return q;
 402                 }
 403                 depth++;
 404         }
 405         spin_unlock(&hb->chain_lock);
 406         read_unlock(&f->lock);
 407
 408         if (depth <= INETFRAGS_MAXDEPTH)
 409                 return inet_frag_create(nf, f, key);
 410         else
 411                 return ERR_PTR(-ENOBUFS);
 412 }
 413 EXPORT_SYMBOL(inet_frag_find);
 414
 415 void inet_frag_maybe_warn_overflow(struct inet_frag_queue *q,
 416                                    const char *prefix)
 417 {
 418         static const char msg[] = "inet_frag_find: Fragment hash bucket"
 419                 " list length grew over limit " __stringify(INETFRAGS_MAXDEPTH)
 420                 ". Dropping fragment.\n";
 421
 422         if (PTR_ERR(q) == -ENOBUFS)
 423                 LIMIT_NETDEBUG(KERN_WARNING "%s%s", prefix, msg);
 424 }
 425 EXPORT_SYMBOL(inet_frag_maybe_warn_overflow);