net: IP6 defrag: use rbtrees for IPv6 defrag
authorPeter Oskolkov <posk@google.com>
Tue, 22 Jan 2019 18:02:51 +0000 (10:02 -0800)
committerDavid S. Miller <davem@davemloft.net>
Sat, 26 Jan 2019 05:37:11 +0000 (21:37 -0800)
Currently, IPv6 defragmentation code drops non-last fragments that
are smaller than 1280 bytes: see
commit 0ed4229b08c1 ("ipv6: defrag: drop non-last frags smaller than min mtu")

This behavior is not specified in IPv6 RFCs and appears to break
compatibility with some IPv6 implemenations, as reported here:
https://www.spinics.net/lists/netdev/msg543846.html

This patch re-uses common IP defragmentation queueing and reassembly
code in IPv6, removing the 1280 byte restriction.

Signed-off-by: Peter Oskolkov <posk@google.com>
Reported-by: Tom Herbert <tom@herbertland.com>
Cc: Eric Dumazet <edumazet@google.com>
Cc: Florian Westphal <fw@strlen.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
include/net/ipv6_frag.h
net/ipv6/reassembly.c

index 6ced1e6899b6ec41d137293d4957976026ba7ed5..28aa9b30aeceac9a86ee6754e4b5809be115e947 100644 (file)
@@ -82,8 +82,15 @@ ip6frag_expire_frag_queue(struct net *net, struct frag_queue *fq)
        __IP6_INC_STATS(net, __in6_dev_get(dev), IPSTATS_MIB_REASMTIMEOUT);
 
        /* Don't send error if the first segment did not arrive. */
-       head = fq->q.fragments;
-       if (!(fq->q.flags & INET_FRAG_FIRST_IN) || !head)
+       if (!(fq->q.flags & INET_FRAG_FIRST_IN))
+               goto out;
+
+       /* sk_buff::dev and sk_buff::rbnode are unionized. So we
+        * pull the head out of the tree in order to be able to
+        * deal with head->dev.
+        */
+       head = inet_frag_pull_head(&fq->q);
+       if (!head)
                goto out;
 
        head->dev = dev;
index 36a3d8dc61f56ce5abe0239da51d0a48550af5d0..24264d0a4b859b89a3e3da1dd32f78497b44c02c 100644 (file)
@@ -69,8 +69,8 @@ static u8 ip6_frag_ecn(const struct ipv6hdr *ipv6h)
 
 static struct inet_frags ip6_frags;
 
-static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *prev,
-                         struct net_device *dev);
+static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *skb,
+                         struct sk_buff *prev_tail, struct net_device *dev);
 
 static void ip6_frag_expire(struct timer_list *t)
 {
@@ -111,21 +111,26 @@ static int ip6_frag_queue(struct frag_queue *fq, struct sk_buff *skb,
                          struct frag_hdr *fhdr, int nhoff,
                          u32 *prob_offset)
 {
-       struct sk_buff *prev, *next;
-       struct net_device *dev;
-       int offset, end, fragsize;
        struct net *net = dev_net(skb_dst(skb)->dev);
+       int offset, end, fragsize;
+       struct sk_buff *prev_tail;
+       struct net_device *dev;
+       int err = -ENOENT;
        u8 ecn;
 
        if (fq->q.flags & INET_FRAG_COMPLETE)
                goto err;
 
+       err = -EINVAL;
        offset = ntohs(fhdr->frag_off) & ~0x7;
        end = offset + (ntohs(ipv6_hdr(skb)->payload_len) -
                        ((u8 *)(fhdr + 1) - (u8 *)(ipv6_hdr(skb) + 1)));
 
        if ((unsigned int)end > IPV6_MAXPLEN) {
                *prob_offset = (u8 *)&fhdr->frag_off - skb_network_header(skb);
+               /* note that if prob_offset is set, the skb is freed elsewhere,
+                * we do not free it here.
+                */
                return -1;
        }
 
@@ -170,62 +175,27 @@ static int ip6_frag_queue(struct frag_queue *fq, struct sk_buff *skb,
        if (end == offset)
                goto discard_fq;
 
+       err = -ENOMEM;
        /* Point into the IP datagram 'data' part. */
        if (!pskb_pull(skb, (u8 *) (fhdr + 1) - skb->data))
                goto discard_fq;
 
-       if (pskb_trim_rcsum(skb, end - offset))
+       err = pskb_trim_rcsum(skb, end - offset);
+       if (err)
                goto discard_fq;
 
-       /* Find out which fragments are in front and at the back of us
-        * in the chain of fragments so far.  We must know where to put
-        * this fragment, right?
-        */
-       prev = fq->q.fragments_tail;
-       if (!prev || prev->ip_defrag_offset < offset) {
-               next = NULL;
-               goto found;
-       }
-       prev = NULL;
-       for (next = fq->q.fragments; next != NULL; next = next->next) {
-               if (next->ip_defrag_offset >= offset)
-                       break;  /* bingo! */
-               prev = next;
-       }
-
-found:
-       /* RFC5722, Section 4, amended by Errata ID : 3089
-        *                          When reassembling an IPv6 datagram, if
-        *   one or more its constituent fragments is determined to be an
-        *   overlapping fragment, the entire datagram (and any constituent
-        *   fragments) MUST be silently discarded.
-        */
-
-       /* Check for overlap with preceding fragment. */
-       if (prev &&
-           (prev->ip_defrag_offset + prev->len) > offset)
-               goto discard_fq;
-
-       /* Look for overlap with succeeding segment. */
-       if (next && next->ip_defrag_offset < end)
-               goto discard_fq;
-
-       /* Note : skb->ip_defrag_offset and skb->sk share the same location */
+       /* Note : skb->rbnode and skb->dev share the same location. */
        dev = skb->dev;
-       if (dev)
-               fq->iif = dev->ifindex;
        /* Makes sure compiler wont do silly aliasing games */
        barrier();
-       skb->ip_defrag_offset = offset;
 
-       /* Insert this fragment in the chain of fragments. */
-       skb->next = next;
-       if (!next)
-               fq->q.fragments_tail = skb;
-       if (prev)
-               prev->next = skb;
-       else
-               fq->q.fragments = skb;
+       prev_tail = fq->q.fragments_tail;
+       err = inet_frag_queue_insert(&fq->q, skb, offset, end);
+       if (err)
+               goto insert_error;
+
+       if (dev)
+               fq->iif = dev->ifindex;
 
        fq->q.stamp = skb->tstamp;
        fq->q.meat += skb->len;
@@ -246,44 +216,48 @@ found:
 
        if (fq->q.flags == (INET_FRAG_FIRST_IN | INET_FRAG_LAST_IN) &&
            fq->q.meat == fq->q.len) {
-               int res;
                unsigned long orefdst = skb->_skb_refdst;
 
                skb->_skb_refdst = 0UL;
-               res = ip6_frag_reasm(fq, prev, dev);
+               err = ip6_frag_reasm(fq, skb, prev_tail, dev);
                skb->_skb_refdst = orefdst;
-               return res;
+               return err;
        }
 
        skb_dst_drop(skb);
-       return -1;
+       return -EINPROGRESS;
 
+insert_error:
+       if (err == IPFRAG_DUP) {
+               kfree_skb(skb);
+               return -EINVAL;
+       }
+       err = -EINVAL;
+       __IP6_INC_STATS(net, ip6_dst_idev(skb_dst(skb)),
+                       IPSTATS_MIB_REASM_OVERLAPS);
 discard_fq:
        inet_frag_kill(&fq->q);
-err:
        __IP6_INC_STATS(net, ip6_dst_idev(skb_dst(skb)),
                        IPSTATS_MIB_REASMFAILS);
+err:
        kfree_skb(skb);
-       return -1;
+       return err;
 }
 
 /*
  *     Check if this packet is complete.
- *     Returns NULL on failure by any reason, and pointer
- *     to current nexthdr field in reassembled frame.
  *
  *     It is called with locked fq, and caller must check that
  *     queue is eligible for reassembly i.e. it is not COMPLETE,
  *     the last and the first frames arrived and all the bits are here.
  */
-static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *prev,
-                         struct net_device *dev)
+static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *skb,
+                         struct sk_buff *prev_tail, struct net_device *dev)
 {
        struct net *net = container_of(fq->q.net, struct net, ipv6.frags);
-       struct sk_buff *fp, *head = fq->q.fragments;
-       int    payload_len, delta;
        unsigned int nhoff;
-       int sum_truesize;
+       void *reasm_data;
+       int payload_len;
        u8 ecn;
 
        inet_frag_kill(&fq->q);
@@ -292,121 +266,40 @@ static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *prev,
        if (unlikely(ecn == 0xff))
                goto out_fail;
 
-       /* Make the one we just received the head. */
-       if (prev) {
-               head = prev->next;
-               fp = skb_clone(head, GFP_ATOMIC);
-
-               if (!fp)
-                       goto out_oom;
-
-               fp->next = head->next;
-               if (!fp->next)
-                       fq->q.fragments_tail = fp;
-               prev->next = fp;
-
-               skb_morph(head, fq->q.fragments);
-               head->next = fq->q.fragments->next;
-
-               consume_skb(fq->q.fragments);
-               fq->q.fragments = head;
-       }
-
-       WARN_ON(head == NULL);
-       WARN_ON(head->ip_defrag_offset != 0);
+       reasm_data = inet_frag_reasm_prepare(&fq->q, skb, prev_tail);
+       if (!reasm_data)
+               goto out_oom;
 
-       /* Unfragmented part is taken from the first segment. */
-       payload_len = ((head->data - skb_network_header(head)) -
+       payload_len = ((skb->data - skb_network_header(skb)) -
                       sizeof(struct ipv6hdr) + fq->q.len -
                       sizeof(struct frag_hdr));
        if (payload_len > IPV6_MAXPLEN)
                goto out_oversize;
 
-       delta = - head->truesize;
-
-       /* Head of list must not be cloned. */
-       if (skb_unclone(head, GFP_ATOMIC))
-               goto out_oom;
-
-       delta += head->truesize;
-       if (delta)
-               add_frag_mem_limit(fq->q.net, delta);
-
-       /* If the first fragment is fragmented itself, we split
-        * it to two chunks: the first with data and paged part
-        * and the second, holding only fragments. */
-       if (skb_has_frag_list(head)) {
-               struct sk_buff *clone;
-               int i, plen = 0;
-
-               clone = alloc_skb(0, GFP_ATOMIC);
-               if (!clone)
-                       goto out_oom;
-               clone->next = head->next;
-               head->next = clone;
-               skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
-               skb_frag_list_init(head);
-               for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
-                       plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
-               clone->len = clone->data_len = head->data_len - plen;
-               head->data_len -= clone->len;
-               head->len -= clone->len;
-               clone->csum = 0;
-               clone->ip_summed = head->ip_summed;
-               add_frag_mem_limit(fq->q.net, clone->truesize);
-       }
-
        /* We have to remove fragment header from datagram and to relocate
         * header in order to calculate ICV correctly. */
        nhoff = fq->nhoffset;
-       skb_network_header(head)[nhoff] = skb_transport_header(head)[0];
-       memmove(head->head + sizeof(struct frag_hdr), head->head,
-               (head->data - head->head) - sizeof(struct frag_hdr));
-       if (skb_mac_header_was_set(head))
-               head->mac_header += sizeof(struct frag_hdr);
-       head->network_header += sizeof(struct frag_hdr);
-
-       skb_reset_transport_header(head);
-       skb_push(head, head->data - skb_network_header(head));
-
-       sum_truesize = head->truesize;
-       for (fp = head->next; fp;) {
-               bool headstolen;
-               int delta;
-               struct sk_buff *next = fp->next;
-
-               sum_truesize += fp->truesize;
-               if (head->ip_summed != fp->ip_summed)
-                       head->ip_summed = CHECKSUM_NONE;
-               else if (head->ip_summed == CHECKSUM_COMPLETE)
-                       head->csum = csum_add(head->csum, fp->csum);
-
-               if (skb_try_coalesce(head, fp, &headstolen, &delta)) {
-                       kfree_skb_partial(fp, headstolen);
-               } else {
-                       fp->sk = NULL;
-                       if (!skb_shinfo(head)->frag_list)
-                               skb_shinfo(head)->frag_list = fp;
-                       head->data_len += fp->len;
-                       head->len += fp->len;
-                       head->truesize += fp->truesize;
-               }
-               fp = next;
-       }
-       sub_frag_mem_limit(fq->q.net, sum_truesize);
+       skb_network_header(skb)[nhoff] = skb_transport_header(skb)[0];
+       memmove(skb->head + sizeof(struct frag_hdr), skb->head,
+               (skb->data - skb->head) - sizeof(struct frag_hdr));
+       if (skb_mac_header_was_set(skb))
+               skb->mac_header += sizeof(struct frag_hdr);
+       skb->network_header += sizeof(struct frag_hdr);
+
+       skb_reset_transport_header(skb);
+
+       inet_frag_reasm_finish(&fq->q, skb, reasm_data);
 
-       skb_mark_not_on_list(head);
-       head->dev = dev;
-       head->tstamp = fq->q.stamp;
-       ipv6_hdr(head)->payload_len = htons(payload_len);
-       ipv6_change_dsfield(ipv6_hdr(head), 0xff, ecn);
-       IP6CB(head)->nhoff = nhoff;
-       IP6CB(head)->flags |= IP6SKB_FRAGMENTED;
-       IP6CB(head)->frag_max_size = fq->q.max_size;
+       skb->dev = dev;
+       ipv6_hdr(skb)->payload_len = htons(payload_len);
+       ipv6_change_dsfield(ipv6_hdr(skb), 0xff, ecn);
+       IP6CB(skb)->nhoff = nhoff;
+       IP6CB(skb)->flags |= IP6SKB_FRAGMENTED;
+       IP6CB(skb)->frag_max_size = fq->q.max_size;
 
        /* Yes, and fold redundant checksum back. 8) */
-       skb_postpush_rcsum(head, skb_network_header(head),
-                          skb_network_header_len(head));
+       skb_postpush_rcsum(skb, skb_network_header(skb),
+                          skb_network_header_len(skb));
 
        rcu_read_lock();
        __IP6_INC_STATS(net, __in6_dev_get(dev), IPSTATS_MIB_REASMOKS);
@@ -414,6 +307,7 @@ static int ip6_frag_reasm(struct frag_queue *fq, struct sk_buff *prev,
        fq->q.fragments = NULL;
        fq->q.rb_fragments = RB_ROOT;
        fq->q.fragments_tail = NULL;
+       fq->q.last_run_head = NULL;
        return 1;
 
 out_oversize:
@@ -464,10 +358,6 @@ static int ipv6_frag_rcv(struct sk_buff *skb)
                return 1;
        }
 
-       if (skb->len - skb_network_offset(skb) < IPV6_MIN_MTU &&
-           fhdr->frag_off & htons(IP6_MF))
-               goto fail_hdr;
-
        iif = skb->dev ? skb->dev->ifindex : 0;
        fq = fq_find(net, fhdr->identification, hdr, iif);
        if (fq) {
@@ -485,6 +375,7 @@ static int ipv6_frag_rcv(struct sk_buff *skb)
                if (prob_offset) {
                        __IP6_INC_STATS(net, __in6_dev_get_safely(skb->dev),
                                        IPSTATS_MIB_INHDRERRORS);
+                       /* icmpv6_param_prob() calls kfree_skb(skb) */
                        icmpv6_param_prob(skb, ICMPV6_HDR_FIELD, prob_offset);
                }
                return ret;