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 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
  29  * Value : 0xff if frame should be dropped.
  30  *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
  31  */
  32 const u8 ip_frag_ecn_table[16] = {
  33         /* at least one fragment had CE, and others ECT_0 or ECT_1 */
  34         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
  35         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
  36         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
  37
  38         /* invalid combinations : drop frame */
  39         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
  40         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
  41         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
  42         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  43         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
  44         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
  45         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  46 };
  47 EXPORT_SYMBOL(ip_frag_ecn_table);
  48
  49 static unsigned int
  50 inet_frag_hashfn(const struct inet_frags *f, const struct inet_frag_queue *q)
  51 {
  52         return f->hashfn(q) & (INETFRAGS_HASHSZ - 1);
  53 }
  54
  55 static void inet_frag_secret_rebuild(unsigned long dummy)
  56 {
  57         struct inet_frags *f = (struct inet_frags *)dummy;
  58         unsigned long now = jiffies;
  59         int i;
  60
  61         /* Per bucket lock NOT needed here, due to write lock protection */
  62         write_lock(&f->lock);
  63
  64         get_random_bytes(&f->rnd, sizeof(u32));
  65         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
  66                 struct inet_frag_bucket *hb;
  67                 struct inet_frag_queue *q;
  68                 struct hlist_node *n;
  69
  70                 hb = &f->hash[i];
  71                 hlist_for_each_entry_safe(q, n, &hb->chain, list) {
  72                         unsigned int hval = inet_frag_hashfn(f, q);
  73
  74                         if (hval != i) {
  75                                 struct inet_frag_bucket *hb_dest;
  76
  77                                 hlist_del(&q->list);
  78
  79                                 /* Relink to new hash chain. */
  80                                 hb_dest = &f->hash[hval];
  81                                 hlist_add_head(&q->list, &hb_dest->chain);
  82                         }
  83                 }
  84         }
  85         write_unlock(&f->lock);
  86
  87         mod_timer(&f->secret_timer, now + f->secret_interval);
  88 }
  89
  90 void inet_frags_init(struct inet_frags *f)
  91 {
  92         int i;
  93
  94         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
  95                 struct inet_frag_bucket *hb = &f->hash[i];
  96
  97                 spin_lock_init(&hb->chain_lock);
  98                 INIT_HLIST_HEAD(&hb->chain);
  99         }
 100         rwlock_init(&f->lock);
 101
 102         setup_timer(&f->secret_timer, inet_frag_secret_rebuild,
 103                         (unsigned long)f);
 104         f->secret_timer.expires = jiffies + f->secret_interval;
 105         add_timer(&f->secret_timer);
 106 }
 107 EXPORT_SYMBOL(inet_frags_init);
 108
 109 void inet_frags_init_net(struct netns_frags *nf)
 110 {
 111         nf->nqueues = 0;
 112         init_frag_mem_limit(nf);
 113         INIT_LIST_HEAD(&nf->lru_list);
 114         spin_lock_init(&nf->lru_lock);
 115 }
 116 EXPORT_SYMBOL(inet_frags_init_net);
 117
 118 void inet_frags_fini(struct inet_frags *f)
 119 {
 120         del_timer(&f->secret_timer);
 121 }
 122 EXPORT_SYMBOL(inet_frags_fini);
 123
 124 void inet_frags_exit_net(struct netns_frags *nf, struct inet_frags *f)
 125 {
 126         nf->low_thresh = 0;
 127
 128         local_bh_disable();
 129         inet_frag_evictor(nf, f, true);
 130         local_bh_enable();
 131
 132         percpu_counter_destroy(&nf->mem);
 133 }
 134 EXPORT_SYMBOL(inet_frags_exit_net);
 135
 136 static inline void fq_unlink(struct inet_frag_queue *fq, struct inet_frags *f)
 137 {
 138         struct inet_frag_bucket *hb;
 139         unsigned int hash;
 140
 141         read_lock(&f->lock);
 142         hash = inet_frag_hashfn(f, fq);
 143         hb = &f->hash[hash];
 144
 145         spin_lock(&hb->chain_lock);
 146         hlist_del(&fq->list);
 147         spin_unlock(&hb->chain_lock);
 148
 149         read_unlock(&f->lock);
 150         inet_frag_lru_del(fq);
 151 }
 152
 153 void inet_frag_kill(struct inet_frag_queue *fq, struct inet_frags *f)
 154 {
 155         if (del_timer(&fq->timer))
 156                 atomic_dec(&fq->refcnt);
 157
 158         if (!(fq->last_in & INET_FRAG_COMPLETE)) {
 159                 fq_unlink(fq, f);
 160                 atomic_dec(&fq->refcnt);
 161                 fq->last_in |= INET_FRAG_COMPLETE;
 162         }
 163 }
 164 EXPORT_SYMBOL(inet_frag_kill);
 165
 166 static inline void frag_kfree_skb(struct netns_frags *nf, struct inet_frags *f,
 167                 struct sk_buff *skb)
 168 {
 169         if (f->skb_free)
 170                 f->skb_free(skb);
 171         kfree_skb(skb);
 172 }
 173
 174 void inet_frag_destroy(struct inet_frag_queue *q, struct inet_frags *f,
 175                                         int *work)
 176 {
 177         struct sk_buff *fp;
 178         struct netns_frags *nf;
 179         unsigned int sum, sum_truesize = 0;
 180
 181         WARN_ON(!(q->last_in & INET_FRAG_COMPLETE));
 182         WARN_ON(del_timer(&q->timer) != 0);
 183
 184         /* Release all fragment data. */
 185         fp = q->fragments;
 186         nf = q->net;
 187         while (fp) {
 188                 struct sk_buff *xp = fp->next;
 189
 190                 sum_truesize += fp->truesize;
 191                 frag_kfree_skb(nf, f, fp);
 192                 fp = xp;
 193         }
 194         sum = sum_truesize + f->qsize;
 195         if (work)
 196                 *work -= sum;
 197         sub_frag_mem_limit(q, sum);
 198
 199         if (f->destructor)
 200                 f->destructor(q);
 201         kfree(q);
 202
 203 }
 204 EXPORT_SYMBOL(inet_frag_destroy);
 205
 206 int inet_frag_evictor(struct netns_frags *nf, struct inet_frags *f, bool force)
 207 {
 208         struct inet_frag_queue *q;
 209         int work, evicted = 0;
 210
 211         if (!force) {
 212                 if (frag_mem_limit(nf) <= nf->high_thresh)
 213                         return 0;
 214         }
 215
 216         work = frag_mem_limit(nf) - nf->low_thresh;
 217         while (work > 0 || force) {
 218                 spin_lock(&nf->lru_lock);
 219
 220                 if (list_empty(&nf->lru_list)) {
 221                         spin_unlock(&nf->lru_lock);
 222                         break;
 223                 }
 224
 225                 q = list_first_entry(&nf->lru_list,
 226                                 struct inet_frag_queue, lru_list);
 227                 atomic_inc(&q->refcnt);
 228                 /* Remove q from list to avoid several CPUs grabbing it */
 229                 list_del_init(&q->lru_list);
 230
 231                 spin_unlock(&nf->lru_lock);
 232
 233                 spin_lock(&q->lock);
 234                 if (!(q->last_in & INET_FRAG_COMPLETE))
 235                         inet_frag_kill(q, f);
 236                 spin_unlock(&q->lock);
 237
 238                 if (atomic_dec_and_test(&q->refcnt))
 239                         inet_frag_destroy(q, f, &work);
 240                 evicted++;
 241         }
 242
 243         return evicted;
 244 }
 245 EXPORT_SYMBOL(inet_frag_evictor);
 246
 247 static struct inet_frag_queue *inet_frag_intern(struct netns_frags *nf,
 248                 struct inet_frag_queue *qp_in, struct inet_frags *f,
 249                 void *arg)
 250 {
 251         struct inet_frag_bucket *hb;
 252         struct inet_frag_queue *qp;
 253         unsigned int hash;
 254
 255         read_lock(&f->lock); /* Protects against hash rebuild */
 256         /*
 257          * While we stayed w/o the lock other CPU could update
 258          * the rnd seed, so we need to re-calculate the hash
 259          * chain. Fortunatelly the qp_in can be used to get one.
 260          */
 261         hash = inet_frag_hashfn(f, qp_in);
 262         hb = &f->hash[hash];
 263         spin_lock(&hb->chain_lock);
 264
 265 #ifdef CONFIG_SMP
 266         /* With SMP race we have to recheck hash table, because
 267          * such entry could be created on other cpu, while we
 268          * released the hash bucket lock.
 269          */
 270         hlist_for_each_entry(qp, &hb->chain, list) {
 271                 if (qp->net == nf && f->match(qp, arg)) {
 272                         atomic_inc(&qp->refcnt);
 273                         spin_unlock(&hb->chain_lock);
 274                         read_unlock(&f->lock);
 275                         qp_in->last_in |= INET_FRAG_COMPLETE;
 276                         inet_frag_put(qp_in, f);
 277                         return qp;
 278                 }
 279         }
 280 #endif
 281         qp = qp_in;
 282         if (!mod_timer(&qp->timer, jiffies + nf->timeout))
 283                 atomic_inc(&qp->refcnt);
 284
 285         atomic_inc(&qp->refcnt);
 286         hlist_add_head(&qp->list, &hb->chain);
 287         inet_frag_lru_add(nf, qp);
 288         spin_unlock(&hb->chain_lock);
 289         read_unlock(&f->lock);
 290
 291         return qp;
 292 }
 293
 294 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
 295                 struct inet_frags *f, void *arg)
 296 {
 297         struct inet_frag_queue *q;
 298
 299         q = kzalloc(f->qsize, GFP_ATOMIC);
 300         if (q == NULL)
 301                 return NULL;
 302
 303         q->net = nf;
 304         f->constructor(q, arg);
 305         add_frag_mem_limit(q, f->qsize);
 306
 307         setup_timer(&q->timer, f->frag_expire, (unsigned long)q);
 308         spin_lock_init(&q->lock);
 309         atomic_set(&q->refcnt, 1);
 310         INIT_LIST_HEAD(&q->lru_list);
 311
 312         return q;
 313 }
 314
 315 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
 316                 struct inet_frags *f, void *arg)
 317 {
 318         struct inet_frag_queue *q;
 319
 320         q = inet_frag_alloc(nf, f, arg);
 321         if (q == NULL)
 322                 return NULL;
 323
 324         return inet_frag_intern(nf, q, f, arg);
 325 }
 326
 327 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf,
 328                 struct inet_frags *f, void *key, unsigned int hash)
 329         __releases(&f->lock)
 330 {
 331         struct inet_frag_bucket *hb;
 332         struct inet_frag_queue *q;
 333         int depth = 0;
 334
 335         hash &= (INETFRAGS_HASHSZ - 1);
 336         hb = &f->hash[hash];
 337
 338         spin_lock(&hb->chain_lock);
 339         hlist_for_each_entry(q, &hb->chain, list) {
 340                 if (q->net == nf && f->match(q, key)) {
 341                         atomic_inc(&q->refcnt);
 342                         spin_unlock(&hb->chain_lock);
 343                         read_unlock(&f->lock);
 344                         return q;
 345                 }
 346                 depth++;
 347         }
 348         spin_unlock(&hb->chain_lock);
 349         read_unlock(&f->lock);
 350
 351         if (depth <= INETFRAGS_MAXDEPTH)
 352                 return inet_frag_create(nf, f, key);
 353         else
 354                 return ERR_PTR(-ENOBUFS);
 355 }
 356 EXPORT_SYMBOL(inet_frag_find);
 357
 358 void inet_frag_maybe_warn_overflow(struct inet_frag_queue *q,
 359                                    const char *prefix)
 360 {
 361         static const char msg[] = "inet_frag_find: Fragment hash bucket"
 362                 " list length grew over limit " __stringify(INETFRAGS_MAXDEPTH)
 363                 ". Dropping fragment.\n";
 364
 365         if (PTR_ERR(q) == -ENOBUFS)
 366                 LIMIT_NETDEBUG(KERN_WARNING "%s%s", prefix, msg);
 367 }
 368 EXPORT_SYMBOL(inet_frag_maybe_warn_overflow);