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
  23 #include <net/inet_frag.h>
  24
  25 static void inet_frag_secret_rebuild(unsigned long dummy)
  26 {
  27         struct inet_frags *f = (struct inet_frags *)dummy;
  28         unsigned long now = jiffies;
  29         int i;
  30
  31         write_lock(&f->lock);
  32         get_random_bytes(&f->rnd, sizeof(u32));
  33         for (i = 0; i < INETFRAGS_HASHSZ; i++) {
  34                 struct inet_frag_queue *q;
  35                 struct hlist_node *p, *n;
  36
  37                 hlist_for_each_entry_safe(q, p, n, &f->hash[i], list) {
  38                         unsigned int hval = f->hashfn(q);
  39
  40                         if (hval != i) {
  41                                 hlist_del(&q->list);
  42
  43                                 /* Relink to new hash chain. */
  44                                 hlist_add_head(&q->list, &f->hash[hval]);
  45                         }
  46                 }
  47         }
  48         write_unlock(&f->lock);
  49
  50         mod_timer(&f->secret_timer, now + f->ctl->secret_interval);
  51 }
  52
  53 void inet_frags_init(struct inet_frags *f)
  54 {
  55         int i;
  56
  57         for (i = 0; i < INETFRAGS_HASHSZ; i++)
  58                 INIT_HLIST_HEAD(&f->hash[i]);
  59
  60         INIT_LIST_HEAD(&f->lru_list);
  61         rwlock_init(&f->lock);
  62
  63         f->rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
  64                                    (jiffies ^ (jiffies >> 6)));
  65
  66         f->nqueues = 0;
  67         atomic_set(&f->mem, 0);
  68
  69         init_timer(&f->secret_timer);
  70         f->secret_timer.function = inet_frag_secret_rebuild;
  71         f->secret_timer.data = (unsigned long)f;
  72         f->secret_timer.expires = jiffies + f->ctl->secret_interval;
  73         add_timer(&f->secret_timer);
  74 }
  75 EXPORT_SYMBOL(inet_frags_init);
  76
  77 void inet_frags_fini(struct inet_frags *f)
  78 {
  79         del_timer(&f->secret_timer);
  80 }
  81 EXPORT_SYMBOL(inet_frags_fini);
  82
  83 static inline void fq_unlink(struct inet_frag_queue *fq, struct inet_frags *f)
  84 {
  85         write_lock(&f->lock);
  86         hlist_del(&fq->list);
  87         list_del(&fq->lru_list);
  88         f->nqueues--;
  89         write_unlock(&f->lock);
  90 }
  91
  92 void inet_frag_kill(struct inet_frag_queue *fq, struct inet_frags *f)
  93 {
  94         if (del_timer(&fq->timer))
  95                 atomic_dec(&fq->refcnt);
  96
  97         if (!(fq->last_in & COMPLETE)) {
  98                 fq_unlink(fq, f);
  99                 atomic_dec(&fq->refcnt);
 100                 fq->last_in |= COMPLETE;
 101         }
 102 }
 103
 104 EXPORT_SYMBOL(inet_frag_kill);
 105
 106 static inline void frag_kfree_skb(struct inet_frags *f, struct sk_buff *skb,
 107                                                 int *work)
 108 {
 109         if (work)
 110                 *work -= skb->truesize;
 111
 112         atomic_sub(skb->truesize, &f->mem);
 113         if (f->skb_free)
 114                 f->skb_free(skb);
 115         kfree_skb(skb);
 116 }
 117
 118 void inet_frag_destroy(struct inet_frag_queue *q, struct inet_frags *f,
 119                                         int *work)
 120 {
 121         struct sk_buff *fp;
 122
 123         BUG_TRAP(q->last_in & COMPLETE);
 124         BUG_TRAP(del_timer(&q->timer) == 0);
 125
 126         /* Release all fragment data. */
 127         fp = q->fragments;
 128         while (fp) {
 129                 struct sk_buff *xp = fp->next;
 130
 131                 frag_kfree_skb(f, fp, work);
 132                 fp = xp;
 133         }
 134
 135         if (work)
 136                 *work -= f->qsize;
 137         atomic_sub(f->qsize, &f->mem);
 138
 139         f->destructor(q);
 140
 141 }
 142 EXPORT_SYMBOL(inet_frag_destroy);
 143
 144 int inet_frag_evictor(struct inet_frags *f)
 145 {
 146         struct inet_frag_queue *q;
 147         int work, evicted = 0;
 148
 149         work = atomic_read(&f->mem) - f->ctl->low_thresh;
 150         while (work > 0) {
 151                 read_lock(&f->lock);
 152                 if (list_empty(&f->lru_list)) {
 153                         read_unlock(&f->lock);
 154                         break;
 155                 }
 156
 157                 q = list_first_entry(&f->lru_list,
 158                                 struct inet_frag_queue, lru_list);
 159                 atomic_inc(&q->refcnt);
 160                 read_unlock(&f->lock);
 161
 162                 spin_lock(&q->lock);
 163                 if (!(q->last_in & COMPLETE))
 164                         inet_frag_kill(q, f);
 165                 spin_unlock(&q->lock);
 166
 167                 if (atomic_dec_and_test(&q->refcnt))
 168                         inet_frag_destroy(q, f, &work);
 169                 evicted++;
 170         }
 171
 172         return evicted;
 173 }
 174 EXPORT_SYMBOL(inet_frag_evictor);
 175
 176 struct inet_frag_queue *inet_frag_intern(struct inet_frag_queue *qp_in,
 177                 struct inet_frags *f, unsigned int hash)
 178 {
 179         struct inet_frag_queue *qp;
 180 #ifdef CONFIG_SMP
 181         struct hlist_node *n;
 182 #endif
 183
 184         write_lock(&f->lock);
 185 #ifdef CONFIG_SMP
 186         /* With SMP race we have to recheck hash table, because
 187          * such entry could be created on other cpu, while we
 188          * promoted read lock to write lock.
 189          */
 190         hlist_for_each_entry(qp, n, &f->hash[hash], list) {
 191                 if (f->equal(qp, qp_in)) {
 192                         atomic_inc(&qp->refcnt);
 193                         write_unlock(&f->lock);
 194                         qp_in->last_in |= COMPLETE;
 195                         inet_frag_put(qp_in, f);
 196                         return qp;
 197                 }
 198         }
 199 #endif
 200         qp = qp_in;
 201         if (!mod_timer(&qp->timer, jiffies + f->ctl->timeout))
 202                 atomic_inc(&qp->refcnt);
 203
 204         atomic_inc(&qp->refcnt);
 205         hlist_add_head(&qp->list, &f->hash[hash]);
 206         list_add_tail(&qp->lru_list, &f->lru_list);
 207         f->nqueues++;
 208         write_unlock(&f->lock);
 209         return qp;
 210 }
 211 EXPORT_SYMBOL(inet_frag_intern);