lwip/src/core/ipv4/ip_frag.c

/**
 * @file
 * This is the IPv4 packet segmentation and reassembly implementation.
 *
 */

/*
 * Copyright (c) 2001-2004 Swedish Institute of Computer Science.
 * All rights reserved. 
 * 
 * Redistribution and use in source and binary forms, with or without modification, 
 * are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice,
 *    this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 * 3. The name of the author may not be used to endorse or promote products
 *    derived from this software without specific prior written permission. 
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED 
 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 
 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT 
 * SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, 
 * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT 
 * OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 
 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 
 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING 
 * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY 
 * OF SUCH DAMAGE.
 *
 * This file is part of the lwIP TCP/IP stack.
 * 
 * Author: Jani Monoses <jani@iv.ro> 
 *         Simon Goldschmidt
 * original reassembly code by Adam Dunkels <adam@sics.se>
 * 
 */

#include "lwip/opt.h"
#include "lwip/ip.h"
#include "lwip/ip_frag.h"
#include "lwip/netif.h"
#include "lwip/snmp.h"
#include "lwip/stats.h"
#include "lwip/icmp.h"

#include <string.h>

#if IP_REASSEMBLY
/**
 * The IP reassembly code currently has the following limitations:
 * - IP header options are not supported
 * - fragments must not overlap (e.g. due to different routes),
 *   currently, overlapping or duplicate fragments are thrown away
 *   if IP_REASS_CHECK_OVERLAP=1 (the default)!
 *
 * @todo: work with IP header options
 */

/** Setting this to 0, you can turn off checking the fragments for overlapping
 * regions. The code gets a little smaller. Only use this if you know that
 * overlapping won't occur on your network! */
#ifndef IP_REASS_CHECK_OVERLAP
#define IP_REASS_CHECK_OVERLAP 1
#endif /* IP_REASS_CHECK_OVERLAP */

/** Set to 0 to prevent freeing the oldest datagram when the reassembly buffer is
 * full (IP_REASS_MAX_PBUFS pbufs are enqueued). The code gets a little smaller.
 * Datagrams will be freed by timeout only. Especially useful when MEMP_NUM_REASSDATA
 * is set to 1, so one datagram can be reassembled at a time, only. */
#ifndef IP_REASS_FREE_OLDEST
#define IP_REASS_FREE_OLDEST 1
#endif /* IP_REASS_FREE_OLDEST */

#define IP_REASS_FLAG_LASTFRAG 0x01

/** This is a helper struct which holds the starting
 * offset and the ending offset of this fragment to
 * easily chain the fragments.
 */
struct ip_reass_helper {
  struct pbuf *next_pbuf;
  u16_t start;
  u16_t end;
};

#define IP_ADDRESSES_AND_ID_MATCH(iphdrA, iphdrB)  \
  (ip_addr_cmp(&(iphdrA)->src, &(iphdrB)->src) && \
   ip_addr_cmp(&(iphdrA)->dest, &(iphdrB)->dest) && \
   IPH_ID(iphdrA) == IPH_ID(iphdrB)) ? 1 : 0

/* global variables */
static struct ip_reassdata *reassdatagrams;
static u16_t ip_reass_pbufcount;

/* function prototypes */
static void ip_reass_dequeue_datagram(struct ip_reassdata *ipr, struct ip_reassdata *prev);
static int ip_reass_free_complete_datagram(struct ip_reassdata *ipr, struct ip_reassdata *prev);

/**
 * Reassembly timer base function
 * for both NO_SYS == 0 and 1 (!).
 *
 * Should be called every 1000 msec (defined by IP_TMR_INTERVAL).
 */
void
ip_reass_tmr(void)
{
  struct ip_reassdata *r, *prev = NULL;

  r = reassdatagrams;
  while (r != NULL) {
    /* Decrement the timer. Once it reaches 0,
     * clean up the incomplete fragment assembly */
    if (r->timer > 0) {
      r->timer--;
      LWIP_DEBUGF(IP_REASS_DEBUG, ("ip_reass_tmr: timer dec %"U16_F"\n",(u16_t)r->timer));
      prev = r;
      r = r->next;
    } else {
      /* reassembly timed out */
      struct ip_reassdata *tmp;
      LWIP_DEBUGF(IP_REASS_DEBUG, ("ip_reass_tmr: timer timed out"));
      tmp = r;
      /* get the next pointer before freeing */
      r = r->next;
      /* free the helper struct and all enqueued pbufs */
      ip_reass_free_complete_datagram(tmp, prev);
     }
   }
}

/**
 * Free a datagram (struct ip_reassdata) and all its pbufs.
 * Updates the total count of enqueued pbufs (ip_reass_pbufcount),
 * SNMP counters and sends an ICMP time exceeded packet.
 *
 * @param ipr datagram to free
 * @param prev the previous datagram in the linked list
 * @return the number of pbufs freed
 */
static int
ip_reass_free_complete_datagram(struct ip_reassdata *ipr, struct ip_reassdata *prev)
{
  int pbufs_freed = 0;
  struct pbuf *p;
  struct ip_reass_helper *iprh;

  LWIP_ASSERT("prev != ipr", prev != ipr);
  if (prev != NULL) {
    LWIP_ASSERT("prev->next == ipr", prev->next == ipr);
  }

  snmp_inc_ipreasmfails();
#if LWIP_ICMP
  iprh = (struct ip_reass_helper *)ipr->p->payload;
  if (iprh->start == 0) {
    /* The first fragment was received, send ICMP time exceeded. */
    /* First, de-queue the first pbuf from r->p. */
    p = ipr->p;
    ipr->p = iprh->next_pbuf;
    /* Then, copy the original header into it. */
    SMEMCPY(p->payload, &ipr->iphdr, IP_HLEN);
    icmp_time_exceeded(p, ICMP_TE_FRAG);
    pbufs_freed += pbuf_clen(p);
    pbuf_free(p);
  }
#endif /* LWIP_ICMP */

  /* First, free all received pbufs.  The individual pbufs need to be released 
     separately as they have not yet been chained */
  p = ipr->p;
  while (p != NULL) {
    struct pbuf *pcur;
    iprh = (struct ip_reass_helper *)p->payload;
    pcur = p;
    /* get the next pointer before freeing */
    p = iprh->next_pbuf;
    pbufs_freed += pbuf_clen(pcur);
    pbuf_free(pcur);    
  }
  /* Then, unchain the struct ip_reassdata from the list and free it. */
  ip_reass_dequeue_datagram(ipr, prev);
  LWIP_ASSERT("ip_reass_pbufcount >= clen", ip_reass_pbufcount >= pbufs_freed);
  ip_reass_pbufcount -= pbufs_freed;

  return pbufs_freed;
}

#if IP_REASS_FREE_OLDEST
/**
 * Free the oldest datagram to make room for enqueueing new fragments.
 * The datagram 'fraghdr' belongs to is not freed!
 *
 * @param fraghdr IP header of the current fragment
 * @param pbufs_needed number of pbufs needed to enqueue
 *        (used for freeing other datagrams if not enough space)
 * @return the number of pbufs freed
 */
static int
ip_reass_remove_oldest_datagram(struct ip_hdr *fraghdr, int pbufs_needed)
{
  /* @todo Can't we simply remove the last datagram in the
   *       linked list behind reassdatagrams?
   */
  struct ip_reassdata *r, *oldest, *prev;
  int pbufs_freed = 0, pbufs_freed_current;
  int other_datagrams;

  /* Free datagrams until being allowed to enqueue 'pbufs_needed' pbufs,
   * but don't free the datagram that 'fraghdr' belongs to! */
  do {
    oldest = NULL;
    prev = NULL;
    other_datagrams = 0;
    r = reassdatagrams;
    while (r != NULL) {
      if (!IP_ADDRESSES_AND_ID_MATCH(&r->iphdr, fraghdr)) {
        /* Not the same datagram as fraghdr */
        other_datagrams++;
        if (oldest == NULL) {
          oldest = r;
        } else if (r->timer <= oldest->timer) {
          /* older than the previous oldest */
          oldest = r;
        }
      }
      if (r->next != NULL) {
        prev = r;
      }
      r = r->next;
    }
    if (oldest != NULL) {
      pbufs_freed_current = ip_reass_free_complete_datagram(oldest, prev);
      pbufs_freed += pbufs_freed_current;
    }
  } while ((pbufs_freed < pbufs_needed) && (other_datagrams > 1));
  return pbufs_freed;
}
#endif /* IP_REASS_FREE_OLDEST */

/**
 * Enqueues a new fragment into the fragment queue
 * @param fraghdr points to the new fragments IP hdr
 * @param clen number of pbufs needed to enqueue (used for freeing other datagrams if not enough space)
 * @return A pointer to the queue location into which the fragment was enqueued
 */
static struct ip_reassdata*
ip_reass_enqueue_new_datagram(struct ip_hdr *fraghdr, int clen)
{
  struct ip_reassdata* ipr;
  /* No matching previous fragment found, allocate a new reassdata struct */
  ipr = memp_malloc(MEMP_REASSDATA);
  if (ipr == NULL) {
#if IP_REASS_FREE_OLDEST
    if (ip_reass_remove_oldest_datagram(fraghdr, clen) >= clen) {
      ipr = memp_malloc(MEMP_REASSDATA);
    }
    if (ipr == NULL)
#endif /* IP_REASS_FREE_OLDEST */
    {
      IPFRAG_STATS_INC(ip_frag.memerr);
      LWIP_DEBUGF(IP_REASS_DEBUG,("Failed to alloc reassdata struct"));
      return NULL;
    }
  }
  memset(ipr, 0, sizeof(struct ip_reassdata));
  ipr->timer = IP_REASS_MAXAGE;

  /* enqueue the new structure to the front of the list */
  ipr->next = reassdatagrams;
  reassdatagrams = ipr;
  /* copy the ip header for later tests and input */
  /* @todo: no ip options supported? */
  SMEMCPY(&(ipr->iphdr), fraghdr, IP_HLEN);
  return ipr;
}

/**
 * Dequeues a datagram from the datagram queue. Doesn't deallocate the pbufs.
 * @param ipr points to the queue entry to dequeue
 */
static void
ip_reass_dequeue_datagram(struct ip_reassdata *ipr, struct ip_reassdata *prev)
{
  
  /* dequeue the reass struct  */
  if (reassdatagrams == ipr) {
    /* it was the first in the list */
    reassdatagrams = ipr->next;
  } else {
    /* it wasn't the first, so it must have a valid 'prev' */
    LWIP_ASSERT("sanity check linked list", prev != NULL);
    prev->next = ipr->next;
  }

  /* now we can free the ip_reass struct */
  memp_free(MEMP_REASSDATA, ipr);
}

/**
 * Chain a new pbuf into the pbuf list that composes the datagram.  The pbuf list
 * will grow over time as  new pbufs are rx.
 * Also checks that the datagram passes basic continuity checks (if the last
 * fragment was received at least once).
 * @param root_p points to the 'root' pbuf for the current datagram being assembled.
 * @param new_p points to the pbuf for the current fragment
 * @return 0 if invalid, >0 otherwise
 */
static int
ip_reass_chain_frag_into_datagram_and_validate(struct ip_reassdata *ipr, struct pbuf *new_p)
{
  struct ip_reass_helper *iprh, *iprh_tmp, *iprh_prev=NULL;
  struct pbuf *q;
  u16_t offset,len;
  struct ip_hdr *fraghdr;
  int valid = 1;

  /* Extract length and fragment offset from current fragment */
  fraghdr = (struct ip_hdr*)new_p->payload; 
  len = ntohs(IPH_LEN(fraghdr)) - IPH_HL(fraghdr) * 4;
  offset = (ntohs(IPH_OFFSET(fraghdr)) & IP_OFFMASK) * 8;

  /* overwrite the fragment's ip header from the pbuf with our helper struct,
   * and setup the embedded helper structure. */
  /* make sure the struct ip_reass_helper fits into the IP header */
  LWIP_ASSERT("sizeof(struct ip_reass_helper) <= IP_HLEN",
              sizeof(struct ip_reass_helper) <= IP_HLEN);
  iprh = (struct ip_reass_helper*)new_p->payload;
  iprh->next_pbuf = NULL;
  iprh->start = offset;
  iprh->end = offset + len;

  /* Iterate through until we either get to the end of the list (append),
   * or we find on with a larger offset (insert). */
  for (q = ipr->p; q != NULL;) {
    iprh_tmp = (struct ip_reass_helper*)q->payload;
    if (iprh->start < iprh_tmp->start) {
      /* the new pbuf should be inserted before this */
      iprh->next_pbuf = q;
      if (iprh_prev != NULL) {
        /* not the fragment with the lowest offset */
#if IP_REASS_CHECK_OVERLAP
        if ((iprh->start < iprh_prev->end) || (iprh->end > iprh_tmp->start)) {
          /* fragment overlaps with previous or following, throw away */
          goto freepbuf;
        }
#endif /* IP_REASS_CHECK_OVERLAP */
        iprh_prev->next_pbuf = new_p;
      } else {
        /* fragment with the lowest offset */
        ipr->p = new_p;
      }
      break;
    } else if(iprh->start == iprh_tmp->start) {
      /* received the same datagram twice: no need to keep the datagram */
      goto freepbuf;
#if IP_REASS_CHECK_OVERLAP
    } else if(iprh->start < iprh_tmp->end) {
      /* overlap: no need to keep the new datagram */
      goto freepbuf;
#endif /* IP_REASS_CHECK_OVERLAP */
    } else {
      /* Check if the fragments received so far have no wholes. */
      if (iprh_prev != NULL) {
        if (iprh_prev->end != iprh_tmp->start) {
          /* There is a fragment missing between the current
           * and the previous fragment */
          valid = 0;
        }
      }
    }
    q = iprh_tmp->next_pbuf;
    iprh_prev = iprh_tmp;
  }

  /* If q is NULL, then we made it to the end of the list. Determine what to do now */
  if (q == NULL) {
    if (iprh_prev != NULL) {
      /* this is (for now), the fragment with the highest offset:
       * chain it to the last fragment */
#if IP_REASS_CHECK_OVERLAP
      LWIP_ASSERT("check fragments don't overlap", iprh_prev->end <= iprh->start);
#endif /* IP_REASS_CHECK_OVERLAP */
      iprh_prev->next_pbuf = new_p;
      if (iprh_prev->end != iprh->start) {
        valid = 0;
      }
    } else {
#if IP_REASS_CHECK_OVERLAP
      LWIP_ASSERT("no previous fragment, this must be the first fragment!",
        ipr->p == NULL);
#endif /* IP_REASS_CHECK_OVERLAP */
      /* this is the first fragment we ever received for this ip datagram */
      ipr->p = new_p;
    }
  }

  /* At this point, the validation part begins: */
  /* If we already received the last fragment */
  if ((ipr->flags & IP_REASS_FLAG_LASTFRAG) != 0) {
    /* and had no wholes so far */
    if (valid) {
      /* then check if the rest of the fragments is here */
      /* Check if the queue starts with the first datagram */
      if (((struct ip_reass_helper*)ipr->p->payload)->start != 0) {
        valid = 0;
      } else {
        /* and check that there are no wholes after this datagram */
        iprh_prev = iprh;
        q = iprh->next_pbuf;
        while (q != NULL) {
          iprh = (struct ip_reass_helper*)q->payload;
          if (iprh_prev->end != iprh->start) {
            valid = 0;
            break;
          }
          iprh_prev = iprh;
          q = iprh->next_pbuf;
        }
        /* if still valid, all fragments are received
         * (because to the MF==0 already arrived */
        if (valid) {
          LWIP_ASSERT("sanity check", ipr->p != NULL);
          LWIP_ASSERT("sanity check",
            ((struct ip_reass_helper*)ipr->p->payload) != iprh);
          LWIP_ASSERT("validate_datagram:next_pbuf!=NULL",
            iprh->next_pbuf == NULL);
          LWIP_ASSERT("validate_datagram:datagram end!=datagram len",
            iprh->end == ipr->datagram_len);
        }
      }
    }
    /* If valid is 0 here, there are some fragments missing in the middle
     * (since MF == 0 has already arrived). Such datagrams simply time out if
     * no more fragments are received... */
    return valid;
  }
  /* If we come here, not all fragments were received, yet! */
  return 0; /* not yet valid! */
#if IP_REASS_CHECK_OVERLAP
freepbuf:
  ip_reass_pbufcount -= pbuf_clen(new_p);
  pbuf_free(new_p);
  return 0;
#endif /* IP_REASS_CHECK_OVERLAP */
}

/**
 * Reassembles incoming IP fragments into an IP datagram.
 *
 * @param p points to a pbuf chain of the fragment
 * @return NULL if reassembly is incomplete, ? otherwise
 */
struct pbuf *
ip_reass(struct pbuf *p)
{
  struct pbuf *r;
  struct ip_hdr *fraghdr;
  struct ip_reassdata *ipr;
  struct ip_reass_helper *iprh;
  u16_t offset, len;
  u8_t clen;
  struct ip_reassdata *ipr_prev = NULL;

  IPFRAG_STATS_INC(ip_frag.recv);
  snmp_inc_ipreasmreqds();

  fraghdr = (struct ip_hdr*)p->payload;

  if ((IPH_HL(fraghdr) * 4) != IP_HLEN) {
    LWIP_DEBUGF(IP_REASS_DEBUG,("ip_reass: IP options currently not supported!"));
    IPFRAG_STATS_INC(ip_frag.err);
    goto nullreturn;
  }

  offset = (ntohs(IPH_OFFSET(fraghdr)) & IP_OFFMASK) * 8;
  len = ntohs(IPH_LEN(fraghdr)) - IPH_HL(fraghdr) * 4;

  /* Check if we are allowed to enqueue more datagrams. */
  clen = pbuf_clen(p);
  if ((ip_reass_pbufcount + clen) > IP_REASS_MAX_PBUFS) {
#if IP_REASS_FREE_OLDEST
    if (!ip_reass_remove_oldest_datagram(fraghdr, clen) ||
        ((ip_reass_pbufcount + clen) > IP_REASS_MAX_PBUFS))
#endif /* IP_REASS_FREE_OLDEST */
    {
      /* No datagram could be freed and still too many pbufs enqueued */
      LWIP_DEBUGF(IP_REASS_DEBUG,("ip_reass: Overflow condition: pbufct=%d, clen=%d, MAX=%d",
        ip_reass_pbufcount, clen, IP_REASS_MAX_PBUFS));
      IPFRAG_STATS_INC(ip_frag.memerr);
      /* @todo: send ICMP time exceeded here? */
      /* drop this pbuf */
      goto nullreturn;
    }
  }

  /* Look for the datagram the fragment belongs to in the current datagram queue,
   * remembering the previous in the queue for later dequeueing. */
  for (ipr = reassdatagrams; ipr != NULL; ipr = ipr->next) {
    /* Check if the incoming fragment matches the one currently present
       in the reassembly buffer. If so, we proceed with copying the
       fragment into the buffer. */
    if (IP_ADDRESSES_AND_ID_MATCH(&ipr->iphdr, fraghdr)) {
      LWIP_DEBUGF(IP_REASS_DEBUG, ("ip_reass: matching previous fragment ID=%"X16_F"\n",
        ntohs(IPH_ID(fraghdr))));
      IPFRAG_STATS_INC(ip_frag.cachehit);
      break;
    }
    ipr_prev = ipr;
  }

  if (ipr == NULL) {
  /* Enqueue a new datagram into the datagram queue */
    ipr = ip_reass_enqueue_new_datagram(fraghdr, clen);
    /* Bail if unable to enqueue */
    if(ipr == NULL) {
      goto nullreturn;
    }
  } else {
    if (((ntohs(IPH_OFFSET(fraghdr)) & IP_OFFMASK) == 0) && 
      ((ntohs(IPH_OFFSET(&ipr->iphdr)) & IP_OFFMASK) != 0)) {
      /* ipr->iphdr is not the header from the first fragment, but fraghdr is
       * -> copy fraghdr into ipr->iphdr since we want to have the header
       * of the first fragment (for ICMP time exceeded and later, for copying
       * all options, if supported)*/
      SMEMCPY(&ipr->iphdr, fraghdr, IP_HLEN);
    }
  }
  /* Track the current number of pbufs current 'in-flight', in order to limit 
  the number of fragments that may be enqueued at any one time */
  ip_reass_pbufcount += clen;

  /* At this point, we have either created a new entry or pointing 
   * to an existing one */

  /* check for 'no more fragments', and update queue entry*/
  if ((ntohs(IPH_OFFSET(fraghdr)) & IP_MF) == 0) {
    ipr->flags |= IP_REASS_FLAG_LASTFRAG;
    ipr->datagram_len = offset + len;
    LWIP_DEBUGF(IP_REASS_DEBUG,
     ("ip_reass: last fragment seen, total len %"S16_F"\n",
      ipr->datagram_len));
  }
  /* find the right place to insert this pbuf */
  /* @todo: trim pbufs if fragments are overlapping */
  if (ip_reass_chain_frag_into_datagram_and_validate(ipr, p)) {
    /* the totally last fragment (flag more fragments = 0) was received at least
     * once AND all fragments are received */
    ipr->datagram_len += IP_HLEN;

    /* save the second pbuf before copying the header over the pointer */
    r = ((struct ip_reass_helper*)ipr->p->payload)->next_pbuf;

    /* copy the original ip header back to the first pbuf */
    fraghdr = (struct ip_hdr*)(ipr->p->payload);
    SMEMCPY(fraghdr, &ipr->iphdr, IP_HLEN);
    IPH_LEN_SET(fraghdr, htons(ipr->datagram_len));
    IPH_OFFSET_SET(fraghdr, 0);
    IPH_CHKSUM_SET(fraghdr, 0);
    /* @todo: do we need to set calculate the correct checksum? */
    IPH_CHKSUM_SET(fraghdr, inet_chksum(fraghdr, IP_HLEN));

    p = ipr->p;

    /* chain together the pbufs contained within the reass_data list. */
    while(r != NULL) {
      iprh = (struct ip_reass_helper*)r->payload;

      /* hide the ip header for every succeding fragment */
      pbuf_header(r, -IP_HLEN);
      pbuf_cat(p, r);
      r = iprh->next_pbuf;
    }
    /* release the sources allocate for the fragment queue entry */
    ip_reass_dequeue_datagram(ipr, ipr_prev);

    /* and adjust the number of pbufs currently queued for reassembly. */
    ip_reass_pbufcount -= pbuf_clen(p);

    /* Return the pbuf chain */
    return p;
  }
  /* the datagram is not (yet?) reassembled completely */
  LWIP_DEBUGF(IP_REASS_DEBUG,("ip_reass_pbufcount: %d out", ip_reass_pbufcount));
  return NULL;

nullreturn:
  LWIP_DEBUGF(IP_REASS_DEBUG,("nullreturn"));
  IPFRAG_STATS_INC(ip_frag.drop);
  pbuf_free(p);
  return NULL;
}
#endif /* IP_REASSEMBLY */

#if IP_FRAG
#if IP_FRAG_USES_STATIC_BUF
static u8_t buf[LWIP_MEM_ALIGN_SIZE(IP_FRAG_MAX_MTU)];
#endif /* IP_FRAG_USES_STATIC_BUF */

/**
 * Fragment an IP datagram if too large for the netif.
 *
 * Chop the datagram in MTU sized chunks and send them in order
 * by using a fixed size static memory buffer (PBUF_REF) or
 * point PBUF_REFs into p (depending on IP_FRAG_USES_STATIC_BUF).
 *
 * @param p ip packet to send
 * @param netif the netif on which to send
 * @param dest destination ip address to which to send
 *
 * @return ERR_OK if sent successfully, err_t otherwise
 */
err_t 
ip_frag(struct pbuf *p, struct netif *netif, struct ip_addr *dest)
{
  struct pbuf *rambuf;
#if IP_FRAG_USES_STATIC_BUF
  struct pbuf *header;
#else
  struct pbuf *newpbuf;
  struct ip_hdr *original_iphdr;
#endif
  struct ip_hdr *iphdr;
  u16_t nfb;
  u16_t left, cop;
  u16_t mtu = netif->mtu;
  u16_t ofo, omf;
  u16_t last;
  u16_t poff = IP_HLEN;
  u16_t tmp;
#if !IP_FRAG_USES_STATIC_BUF
  u16_t newpbuflen, left_to_copy;
#endif

  /* Get a RAM based MTU sized pbuf */
#if IP_FRAG_USES_STATIC_BUF
  /* When using a static buffer, we use a PBUF_REF, which we will
   * use to reference the packet (without link header).
   * Layer and length is irrelevant.
   */
  rambuf = pbuf_alloc(PBUF_LINK, 0, PBUF_REF);
  if (rambuf == NULL) {
    LWIP_DEBUGF(IP_REASS_DEBUG, ("ip_frag: pbuf_alloc(PBUF_LINK, 0, PBUF_REF) failed\n"));
    return ERR_MEM;
  }
  rambuf->tot_len = rambuf->len = mtu;
  rambuf->payload = LWIP_MEM_ALIGN((void *)buf);

  /* Copy the IP header in it */
  iphdr = rambuf->payload;
  SMEMCPY(iphdr, p->payload, IP_HLEN);
#else /* IP_FRAG_USES_STATIC_BUF */
  original_iphdr = p->payload;
  iphdr = original_iphdr;
#endif /* IP_FRAG_USES_STATIC_BUF */

  /* Save original offset */
  tmp = ntohs(IPH_OFFSET(iphdr));
  ofo = tmp & IP_OFFMASK;
  omf = tmp & IP_MF;

  left = p->tot_len - IP_HLEN;

  nfb = (mtu - IP_HLEN) / 8;

  while (left) {
    last = (left <= mtu - IP_HLEN);

    /* Set new offset and MF flag */
    tmp = omf | (IP_OFFMASK & (ofo));
    if (!last)
      tmp = tmp | IP_MF;

    /* Fill this fragment */
    cop = last ? left : nfb * 8;

#if IP_FRAG_USES_STATIC_BUF
    poff += pbuf_copy_partial(p, (u8_t*)iphdr + IP_HLEN, cop, poff);
#else /* IP_FRAG_USES_STATIC_BUF */
    /* When not using a static buffer, create a chain of pbufs.
     * The first will be a PBUF_RAM holding the link and IP header.
     * The rest will be PBUF_REFs mirroring the pbuf chain to be fragged,
     * but limited to the size of an mtu.
     */
    rambuf = pbuf_alloc(PBUF_LINK, IP_HLEN, PBUF_RAM);
    if (rambuf == NULL) {
      return ERR_MEM;
    }
    LWIP_ASSERT("this needs a pbuf in one piece!",
                (p->len >= (IP_HLEN)));
    SMEMCPY(rambuf->payload, original_iphdr, IP_HLEN);
    iphdr = rambuf->payload;

    /* Can just adjust p directly for needed offset. */
    p->payload = (u8_t *)p->payload + poff;
    p->len -= poff;

    left_to_copy = cop;
    while (left_to_copy) {
      newpbuflen = (left_to_copy < p->len) ? left_to_copy : p->len;
      /* Is this pbuf already empty? */
      if (!newpbuflen) {
        p = p->next;
        continue;
      }
      newpbuf = pbuf_alloc(PBUF_RAW, 0, PBUF_REF);
      if (newpbuf == NULL) {
        pbuf_free(rambuf);
        return ERR_MEM;
      }
      /* Mirror this pbuf, although we might not need all of it. */
      newpbuf->payload = p->payload;
      newpbuf->len = newpbuf->tot_len = newpbuflen;
      /* Add it to end of rambuf's chain, but using pbuf_cat, not pbuf_chain
       * so that it is removed when pbuf_dechain is later called on rambuf.
       */
      pbuf_cat(rambuf, newpbuf);
      left_to_copy -= newpbuflen;
      if (left_to_copy)
        p = p->next;
    }
    poff = newpbuflen;
#endif /* IP_FRAG_USES_STATIC_BUF */

    /* Correct header */
    IPH_OFFSET_SET(iphdr, htons(tmp));
    IPH_LEN_SET(iphdr, htons(cop + IP_HLEN));
    IPH_CHKSUM_SET(iphdr, 0);
    IPH_CHKSUM_SET(iphdr, inet_chksum(iphdr, IP_HLEN));

#if IP_FRAG_USES_STATIC_BUF
    if (last)
      pbuf_realloc(rambuf, left + IP_HLEN);

    /* This part is ugly: we alloc a RAM based pbuf for 
     * the link level header for each chunk and then 
     * free it.A PBUF_ROM style pbuf for which pbuf_header
     * worked would make things simpler.
     */
    header = pbuf_alloc(PBUF_LINK, 0, PBUF_RAM);
    if (header != NULL) {
      pbuf_chain(header, rambuf);
      netif->output(netif, header, dest);
      IPFRAG_STATS_INC(ip_frag.xmit);
      snmp_inc_ipfragcreates();
      pbuf_free(header);
    } else {
      LWIP_DEBUGF(IP_REASS_DEBUG, ("ip_frag: pbuf_alloc() for header failed\n"));
      pbuf_free(rambuf);
      return ERR_MEM;
    }
#else /* IP_FRAG_USES_STATIC_BUF */
    /* No need for separate header pbuf - we allowed room for it in rambuf
     * when allocated.
     */
    netif->output(netif, rambuf, dest);
    IPFRAG_STATS_INC(ip_frag.xmit);

    /* Unfortunately we can't reuse rambuf - the hardware may still be
     * using the buffer. Instead we free it (and the ensuing chain) and
     * recreate it next time round the loop. If we're lucky the hardware
     * will have already sent the packet, the free will really free, and
     * there will be zero memory penalty.
     */
    
    pbuf_free(rambuf);
#endif /* IP_FRAG_USES_STATIC_BUF */
    left -= cop;
    ofo += nfb;
  }
#if IP_FRAG_USES_STATIC_BUF
  pbuf_free(rambuf);
#endif /* IP_FRAG_USES_STATIC_BUF */
  snmp_inc_ipfragoks();
  return ERR_OK;
}
#endif /* IP_FRAG */