inet: frags: better deal with smp races
[sfrench/cifs-2.6.git] / 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 #include <linux/rhashtable.h>
24
25 #include <net/sock.h>
26 #include <net/inet_frag.h>
27 #include <net/inet_ecn.h>
28
29 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
30  * Value : 0xff if frame should be dropped.
31  *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
32  */
33 const u8 ip_frag_ecn_table[16] = {
34         /* at least one fragment had CE, and others ECT_0 or ECT_1 */
35         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
36         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
37         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
38
39         /* invalid combinations : drop frame */
40         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
41         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
42         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
43         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
44         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
45         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
46         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
47 };
48 EXPORT_SYMBOL(ip_frag_ecn_table);
49
50 int inet_frags_init(struct inet_frags *f)
51 {
52         f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
53                                             NULL);
54         if (!f->frags_cachep)
55                 return -ENOMEM;
56
57         return 0;
58 }
59 EXPORT_SYMBOL(inet_frags_init);
60
61 void inet_frags_fini(struct inet_frags *f)
62 {
63         /* We must wait that all inet_frag_destroy_rcu() have completed. */
64         rcu_barrier();
65
66         kmem_cache_destroy(f->frags_cachep);
67         f->frags_cachep = NULL;
68 }
69 EXPORT_SYMBOL(inet_frags_fini);
70
71 static void inet_frags_free_cb(void *ptr, void *arg)
72 {
73         struct inet_frag_queue *fq = ptr;
74
75         /* If we can not cancel the timer, it means this frag_queue
76          * is already disappearing, we have nothing to do.
77          * Otherwise, we own a refcount until the end of this function.
78          */
79         if (!del_timer(&fq->timer))
80                 return;
81
82         spin_lock_bh(&fq->lock);
83         if (!(fq->flags & INET_FRAG_COMPLETE)) {
84                 fq->flags |= INET_FRAG_COMPLETE;
85                 refcount_dec(&fq->refcnt);
86         }
87         spin_unlock_bh(&fq->lock);
88
89         inet_frag_put(fq);
90 }
91
92 void inet_frags_exit_net(struct netns_frags *nf)
93 {
94         nf->high_thresh = 0; /* prevent creation of new frags */
95
96         rhashtable_free_and_destroy(&nf->rhashtable, inet_frags_free_cb, NULL);
97 }
98 EXPORT_SYMBOL(inet_frags_exit_net);
99
100 void inet_frag_kill(struct inet_frag_queue *fq)
101 {
102         if (del_timer(&fq->timer))
103                 refcount_dec(&fq->refcnt);
104
105         if (!(fq->flags & INET_FRAG_COMPLETE)) {
106                 struct netns_frags *nf = fq->net;
107
108                 fq->flags |= INET_FRAG_COMPLETE;
109                 rhashtable_remove_fast(&nf->rhashtable, &fq->node, nf->f->rhash_params);
110                 refcount_dec(&fq->refcnt);
111         }
112 }
113 EXPORT_SYMBOL(inet_frag_kill);
114
115 static void inet_frag_destroy_rcu(struct rcu_head *head)
116 {
117         struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
118                                                  rcu);
119         struct inet_frags *f = q->net->f;
120
121         if (f->destructor)
122                 f->destructor(q);
123         kmem_cache_free(f->frags_cachep, q);
124 }
125
126 void inet_frag_destroy(struct inet_frag_queue *q)
127 {
128         struct sk_buff *fp;
129         struct netns_frags *nf;
130         unsigned int sum, sum_truesize = 0;
131         struct inet_frags *f;
132
133         WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
134         WARN_ON(del_timer(&q->timer) != 0);
135
136         /* Release all fragment data. */
137         fp = q->fragments;
138         nf = q->net;
139         f = nf->f;
140         if (fp) {
141                 do {
142                         struct sk_buff *xp = fp->next;
143
144                         sum_truesize += fp->truesize;
145                         kfree_skb(fp);
146                         fp = xp;
147                 } while (fp);
148         } else {
149                 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
150         }
151         sum = sum_truesize + f->qsize;
152
153         call_rcu(&q->rcu, inet_frag_destroy_rcu);
154
155         sub_frag_mem_limit(nf, sum);
156 }
157 EXPORT_SYMBOL(inet_frag_destroy);
158
159 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
160                                                struct inet_frags *f,
161                                                void *arg)
162 {
163         struct inet_frag_queue *q;
164
165         q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
166         if (!q)
167                 return NULL;
168
169         q->net = nf;
170         f->constructor(q, arg);
171         add_frag_mem_limit(nf, f->qsize);
172
173         timer_setup(&q->timer, f->frag_expire, 0);
174         spin_lock_init(&q->lock);
175         refcount_set(&q->refcnt, 3);
176
177         return q;
178 }
179
180 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
181                                                 void *arg,
182                                                 struct inet_frag_queue **prev)
183 {
184         struct inet_frags *f = nf->f;
185         struct inet_frag_queue *q;
186
187         q = inet_frag_alloc(nf, f, arg);
188         if (!q) {
189                 *prev = ERR_PTR(-ENOMEM);
190                 return NULL;
191         }
192         mod_timer(&q->timer, jiffies + nf->timeout);
193
194         *prev = rhashtable_lookup_get_insert_key(&nf->rhashtable, &q->key,
195                                                  &q->node, f->rhash_params);
196         if (*prev) {
197                 q->flags |= INET_FRAG_COMPLETE;
198                 inet_frag_kill(q);
199                 inet_frag_destroy(q);
200                 return NULL;
201         }
202         return q;
203 }
204
205 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
206 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf, void *key)
207 {
208         struct inet_frag_queue *fq = NULL, *prev;
209
210         if (!nf->high_thresh || frag_mem_limit(nf) > nf->high_thresh)
211                 return NULL;
212
213         rcu_read_lock();
214
215         prev = rhashtable_lookup(&nf->rhashtable, key, nf->f->rhash_params);
216         if (!prev)
217                 fq = inet_frag_create(nf, key, &prev);
218         if (prev && !IS_ERR(prev)) {
219                 fq = prev;
220                 if (!refcount_inc_not_zero(&fq->refcnt))
221                         fq = NULL;
222         }
223         rcu_read_unlock();
224         return fq;
225 }
226 EXPORT_SYMBOL(inet_frag_find);