static inline void psock_lost_connection(FAR struct socket *psock,
FAR struct tcp_conn_s *conn)
{
FAR sq_entry_t *entry;
FAR sq_entry_t *next;
/* Do not allow any further callbacks */
psock->s_sndcb->flags = 0;
psock->s_sndcb->event = NULL;
/* Free all queued write buffers */
for (entry = sq_peek(&conn->unacked_q); entry; entry = next)
{
next = sq_next(entry);
tcp_wrbuffer_release((FAR struct tcp_wrbuffer_s *)entry);
}
for (entry = sq_peek(&conn->write_q); entry; entry = next)
{
next = sq_next(entry);
tcp_wrbuffer_release((FAR struct tcp_wrbuffer_s *)entry);
}
/* Reset write buffering variables */
sq_init(&conn->unacked_q);
sq_init(&conn->write_q);
conn->sent = 0;
}
FAR struct tcp_wrbuffer_s *tcp_wrbuffer_alloc(void)
{
FAR struct tcp_wrbuffer_s *wrb;
/* We need to allocate two things: (1) A write buffer structure and (2)
* at least one I/O buffer to start the chain.
*
* Allocate the write buffer structure first then the IOBG. In order to
* avoid deadlocks, we will need to free the IOB first, then the write
* buffer
*/
DEBUGVERIFY(net_lockedwait(&g_wrbuffer.sem));
/* Now, we are guaranteed to have a write buffer structure reserved
* for us in the free list.
*/
wrb = (FAR struct tcp_wrbuffer_s *)sq_remfirst(&g_wrbuffer.freebuffers);
DEBUGASSERT(wrb);
memset(wrb, 0, sizeof(struct tcp_wrbuffer_s));
/* Now get the first I/O buffer for the write buffer structure */
wrb->wb_iob = iob_alloc(false);
if (!wrb->wb_iob)
{
ndbg("ERROR: Failed to allocate I/O buffer\n");
tcp_wrbuffer_release(wrb);
return NULL;
}
return wrb;
}
static uint16_t psock_send_interrupt(FAR struct net_driver_s *dev,
FAR void *pvconn, FAR void *pvpriv,
uint16_t flags)
{
FAR struct tcp_conn_s *conn = (FAR struct tcp_conn_s *)pvconn;
FAR struct socket *psock = (FAR struct socket *)pvpriv;
nllvdbg("flags: %04x\n", flags);
/* If this packet contains an acknowledgement, then update the count of
* acknowledged bytes.
*/
if ((flags & TCP_ACKDATA) != 0)
{
FAR struct tcp_wrbuffer_s *wrb;
FAR sq_entry_t *entry;
FAR sq_entry_t *next;
uint32_t ackno;
ackno = tcp_getsequence(TCPBUF->ackno);
nllvdbg("ACK: ackno=%u flags=%04x\n", ackno, flags);
/* Look at every write buffer in the unacked_q. The unacked_q
* holds write buffers that have been entirely sent, but which
* have not yet been ACKed.
*/
for (entry = sq_peek(&conn->unacked_q); entry; entry = next)
{
uint32_t lastseq;
/* Check of some or all of this write buffer has been ACKed. */
next = sq_next(entry);
wrb = (FAR struct tcp_wrbuffer_s*)entry;
/* If the ACKed sequence number is greater than the start
* sequence number of the write buffer, then some or all of
* the write buffer has been ACKed.
*/
if (ackno > WRB_SEQNO(wrb))
{
/* Get the sequence number at the end of the data */
lastseq = WRB_SEQNO(wrb) + WRB_PKTLEN(wrb);
nllvdbg("ACK: wrb=%p seqno=%u lastseq=%u pktlen=%u ackno=%u\n",
wrb, WRB_SEQNO(wrb), lastseq, WRB_PKTLEN(wrb), ackno);
/* Has the entire buffer been ACKed? */
if (ackno >= lastseq)
{
nllvdbg("ACK: wrb=%p Freeing write buffer\n", wrb);
/* Yes... Remove the write buffer from ACK waiting queue */
sq_rem(entry, &conn->unacked_q);
/* And return the write buffer to the pool of free buffers */
tcp_wrbuffer_release(wrb);
}
else
{
unsigned int trimlen;
/* No, then just trim the ACKed bytes from the beginning
* of the write buffer. This will free up some I/O buffers
* that can be reused while are still sending the last
* buffers in the chain.
*/
trimlen = ackno - WRB_SEQNO(wrb);
if (trimlen > WRB_SENT(wrb))
{
/* More data has been ACKed then we have sent? */
trimlen = WRB_SENT(wrb);
}
nllvdbg("ACK: wrb=%p trim %u bytes\n", wrb, trimlen);
WRB_TRIM(wrb, trimlen);
WRB_SEQNO(wrb) = ackno;
WRB_SENT(wrb) -= trimlen;
/* Set the new sequence number for what remains */
nllvdbg("ACK: wrb=%p seqno=%u pktlen=%u\n",
wrb, WRB_SEQNO(wrb), WRB_PKTLEN(wrb));
}
}
}
/* A special case is the head of the write_q which may be partially
* sent and so can still have un-ACKed bytes that could get ACKed
* before the entire write buffer has even been sent.
*/
wrb = (FAR struct tcp_wrbuffer_s*)sq_peek(&conn->write_q);
if (wrb && WRB_SENT(wrb) > 0 && ackno > WRB_SEQNO(wrb))
{
uint32_t nacked;
/* Number of bytes that were ACKed */
nacked = ackno - WRB_SEQNO(wrb);
if (nacked > WRB_SENT(wrb))
{
/* More data has been ACKed then we have sent? ASSERT? */
nacked = WRB_SENT(wrb);
}
nllvdbg("ACK: wrb=%p seqno=%u nacked=%u sent=%u ackno=%u\n",
wrb, WRB_SEQNO(wrb), nacked, WRB_SENT(wrb), ackno);
/* Trim the ACKed bytes from the beginning of the write buffer. */
WRB_TRIM(wrb, nacked);
WRB_SEQNO(wrb) = ackno;
WRB_SENT(wrb) -= nacked;
nllvdbg("ACK: wrb=%p seqno=%u pktlen=%u sent=%u\n",
wrb, WRB_SEQNO(wrb), WRB_PKTLEN(wrb), WRB_SENT(wrb));
}
}
/* Check for a loss of connection */
else if ((flags & (TCP_CLOSE | TCP_ABORT | TCP_TIMEDOUT)) != 0)
{
nllvdbg("Lost connection: %04x\n", flags);
/* Report not connected */
net_lostconnection(psock, flags);
/* Free write buffers and terminate polling */
psock_lost_connection(psock, conn);
return flags;
}
/* Check if we are being asked to retransmit data */
else if ((flags & TCP_REXMIT) != 0)
{
FAR struct tcp_wrbuffer_s *wrb;
FAR sq_entry_t *entry;
nllvdbg("REXMIT: %04x\n", flags);
/* If there is a partially sent write buffer at the head of the
* write_q? Has anything been sent from that write buffer?
*/
wrb = (FAR struct tcp_wrbuffer_s *)sq_peek(&conn->write_q);
nllvdbg("REXMIT: wrb=%p sent=%u\n", wrb, wrb ? WRB_SENT(wrb) : 0);
if (wrb != NULL && WRB_SENT(wrb) > 0)
{
FAR struct tcp_wrbuffer_s *tmp;
uint16_t sent;
/* Yes.. Reset the number of bytes sent sent from the write buffer */
sent = WRB_SENT(wrb);
if (conn->unacked > sent)
{
conn->unacked -= sent;
}
else
{
conn->unacked = 0;
}
if (conn->sent > sent)
{
conn->sent -= sent;
}
else
{
conn->sent = 0;
}
WRB_SENT(wrb) = 0;
nllvdbg("REXMIT: wrb=%p sent=%u, conn unacked=%d sent=%d\n",
wrb, WRB_SENT(wrb), conn->unacked, conn->sent);
/* Increment the retransmit count on this write buffer. */
if (++WRB_NRTX(wrb) >= TCP_MAXRTX)
{
nlldbg("Expiring wrb=%p nrtx=%u\n", wrb, WRB_NRTX(wrb));
/* The maximum retry count as been exhausted. Remove the write
* buffer at the head of the queue.
*/
tmp = (FAR struct tcp_wrbuffer_s *)sq_remfirst(&conn->write_q);
DEBUGASSERT(tmp == wrb);
UNUSED(tmp);
/* And return the write buffer to the free list */
tcp_wrbuffer_release(wrb);
/* NOTE expired is different from un-ACKed, it is designed to
* represent the number of segments that have been sent,
* retransmitted, and un-ACKed, if expired is not zero, the
* connection will be closed.
*
* field expired can only be updated at TCP_ESTABLISHED state
*/
conn->expired++;
}
}
/* Move all segments that have been sent but not ACKed to the write
* queue again note, the un-ACKed segments are put at the head of the
* write_q so they can be resent as soon as possible.
*/
while ((entry = sq_remlast(&conn->unacked_q)) != NULL)
{
wrb = (FAR struct tcp_wrbuffer_s*)entry;
uint16_t sent;
/* Reset the number of bytes sent sent from the write buffer */
sent = WRB_SENT(wrb);
if (conn->unacked > sent)
{
conn->unacked -= sent;
}
else
{
conn->unacked = 0;
}
if (conn->sent > sent)
{
conn->sent -= sent;
}
else
{
conn->sent = 0;
}
WRB_SENT(wrb) = 0;
nllvdbg("REXMIT: wrb=%p sent=%u, conn unacked=%d sent=%d\n",
wrb, WRB_SENT(wrb), conn->unacked, conn->sent);
/* Free any write buffers that have exceed the retry count */
if (++WRB_NRTX(wrb) >= TCP_MAXRTX)
{
nlldbg("Expiring wrb=%p nrtx=%u\n", wrb, WRB_NRTX(wrb));
/* Return the write buffer to the free list */
tcp_wrbuffer_release(wrb);
/* NOTE expired is different from un-ACKed, it is designed to
* represent the number of segments that have been sent,
* retransmitted, and un-ACKed, if expired is not zero, the
* connection will be closed.
*
* field expired can only be updated at TCP_ESTABLISHED state
*/
conn->expired++;
continue;
}
else
{
/* Insert the write buffer into the write_q (in sequence
* number order). The retransmission will occur below
* when the write buffer with the lowest sequenc number
* is pulled from the write_q again.
*/
nllvdbg("REXMIT: Moving wrb=%p nrtx=%u\n", wrb, WRB_NRTX(wrb));
psock_insert_segment(wrb, &conn->write_q);
}
}
}
/* Check if the outgoing packet is available (it may have been claimed
* by a sendto interrupt serving a different thread).
*/
if (dev->d_sndlen > 0)
{
/* Another thread has beat us sending data, wait for the next poll */
return flags;
}
/* We get here if (1) not all of the data has been ACKed, (2) we have been
* asked to retransmit data, (3) the connection is still healthy, and (4)
* the outgoing packet is available for our use. In this case, we are
* now free to send more data to receiver -- UNLESS the buffer contains
* unprocessed incoming data. In that event, we will have to wait for the
* next polling cycle.
*/
if ((conn->tcpstateflags & TCP_ESTABLISHED) &&
(flags & (TCP_POLL | TCP_REXMIT)) &&
!(sq_empty(&conn->write_q)))
{
/* Check if the destination IP address is in the ARP table. If not,
* then the send won't actually make it out... it will be replaced with
* an ARP request.
*
* NOTE 1: This could be an expensive check if there are a lot of
* entries in the ARP table.
*
* NOTE 2: If we are actually harvesting IP addresses on incoming IP
* packets, then this check should not be necessary; the MAC mapping
* should already be in the ARP table in many cases.
*
* NOTE 3: If CONFIG_NET_ARP_SEND then we can be assured that the IP
* address mapping is already in the ARP table.
*/
#if defined(CONFIG_NET_ETHERNET) && !defined(CONFIG_NET_ARP_IPIN) && \
!defined(CONFIG_NET_ARP_SEND)
if (arp_find(conn->ripaddr) != NULL)
#endif
{
FAR struct tcp_wrbuffer_s *wrb;
size_t sndlen;
/* Peek at the head of the write queue (but don't remove anything
* from the write queue yet). We know from the above test that
* the write_q is not empty.
*/
wrb = (FAR struct tcp_wrbuffer_s *)sq_peek(&conn->write_q);
DEBUGASSERT(wrb);
/* Get the amount of data that we can send in the next packet.
* We will send either the remaining data in the buffer I/O
* buffer chain, or as much as will fit given the MSS and current
* window size.
*/
sndlen = WRB_PKTLEN(wrb) - WRB_SENT(wrb);
if (sndlen > tcp_mss(conn))
{
sndlen = tcp_mss(conn);
}
if (sndlen > conn->winsize)
{
sndlen = conn->winsize;
}
nllvdbg("SEND: wrb=%p pktlen=%u sent=%u sndlen=%u\n",
wrb, WRB_PKTLEN(wrb), WRB_SENT(wrb), sndlen);
/* Set the sequence number for this segment. If we are
* retransmitting, then the sequence number will already
* be set for this write buffer.
*/
if (WRB_SEQNO(wrb) == (unsigned)-1)
{
WRB_SEQNO(wrb) = conn->isn + conn->sent;
}
/* The TCP stack updates sndseq on receipt of ACK *before*
* this function is called. In that case sndseq will point
* to the next unacknowledged byte (which might have already
* been sent). We will overwrite the value of sndseq here
* before the packet is sent.
*/
tcp_setsequence(conn->sndseq, WRB_SEQNO(wrb) + WRB_SENT(wrb));
/* Then set-up to send that amount of data with the offset
* corresponding to the amount of data already sent. (this
* won't actually happen until the polling cycle completes).
*/
devif_iob_send(dev, WRB_IOB(wrb), sndlen, WRB_SENT(wrb));
/* Remember how much data we send out now so that we know
* when everything has been acknowledged. Just increment
* the amount of data sent. This will be needed in sequence
* number calculations.
*/
conn->unacked += sndlen;
conn->sent += sndlen;
nllvdbg("SEND: wrb=%p nrtx=%u unacked=%u sent=%u\n",
wrb, WRB_NRTX(wrb), conn->unacked, conn->sent);
/* Increment the count of bytes sent from this write buffer */
WRB_SENT(wrb) += sndlen;
nllvdbg("SEND: wrb=%p sent=%u pktlen=%u\n",
wrb, WRB_SENT(wrb), WRB_PKTLEN(wrb));
/* Remove the write buffer from the write queue if the
* last of the data has been sent from the buffer.
*/
DEBUGASSERT(WRB_SENT(wrb) <= WRB_PKTLEN(wrb));
if (WRB_SENT(wrb) >= WRB_PKTLEN(wrb))
{
FAR struct tcp_wrbuffer_s *tmp;
nllvdbg("SEND: wrb=%p Move to unacked_q\n", wrb);
tmp = (FAR struct tcp_wrbuffer_s *)sq_remfirst(&conn->write_q);
DEBUGASSERT(tmp == wrb);
UNUSED(tmp);
/* Put the I/O buffer chain in the un-acked queue; the
* segment is waiting for ACK again
*/
psock_insert_segment(wrb, &conn->unacked_q);
}
/* Only one data can be sent by low level driver at once,
* tell the caller stop polling the other connection.
*/
flags &= ~TCP_POLL;
}
}
/* Continue waiting */
return flags;
}
ssize_t psock_tcp_send(FAR struct socket *psock, FAR const void *buf,
size_t len)
{
FAR struct tcp_conn_s *conn;
FAR struct tcp_wrbuffer_s *wrb;
net_lock_t save;
ssize_t result = 0;
int errcode;
int ret = OK;
if (!psock || psock->s_crefs <= 0)
{
nerr("ERROR: Invalid socket\n");
errcode = EBADF;
goto errout;
}
if (psock->s_type != SOCK_STREAM || !_SS_ISCONNECTED(psock->s_flags))
{
nerr("ERROR: Not connected\n");
errcode = ENOTCONN;
goto errout;
}
/* Make sure that we have the IP address mapping */
conn = (FAR struct tcp_conn_s *)psock->s_conn;
DEBUGASSERT(conn);
#if defined(CONFIG_NET_ARP_SEND) || defined(CONFIG_NET_ICMPv6_NEIGHBOR)
#ifdef CONFIG_NET_ARP_SEND
#ifdef CONFIG_NET_ICMPv6_NEIGHBOR
if (psock->s_domain == PF_INET)
#endif
{
/* Make sure that the IP address mapping is in the ARP table */
ret = arp_send(conn->u.ipv4.raddr);
}
#endif /* CONFIG_NET_ARP_SEND */
#ifdef CONFIG_NET_ICMPv6_NEIGHBOR
#ifdef CONFIG_NET_ARP_SEND
else
#endif
{
/* Make sure that the IP address mapping is in the Neighbor Table */
ret = icmpv6_neighbor(conn->u.ipv6.raddr);
}
#endif /* CONFIG_NET_ICMPv6_NEIGHBOR */
/* Did we successfully get the address mapping? */
if (ret < 0)
{
nerr("ERROR: Not reachable\n");
errcode = ENETUNREACH;
goto errout;
}
#endif /* CONFIG_NET_ARP_SEND || CONFIG_NET_ICMPv6_NEIGHBOR */
/* Dump the incoming buffer */
BUF_DUMP("psock_tcp_send", buf, len);
/* Set the socket state to sending */
psock->s_flags = _SS_SETSTATE(psock->s_flags, _SF_SEND);
if (len > 0)
{
/* Allocate a write buffer. Careful, the network will be momentarily
* unlocked here.
*/
save = net_lock();
wrb = tcp_wrbuffer_alloc();
if (!wrb)
{
/* A buffer allocation error occurred */
nerr("ERROR: Failed to allocate write buffer\n");
errcode = ENOMEM;
goto errout_with_lock;
}
/* Allocate resources to receive a callback */
if (!psock->s_sndcb)
{
psock->s_sndcb = tcp_callback_alloc(conn);
}
/* Test if the callback has been allocated */
if (!psock->s_sndcb)
{
/* A buffer allocation error occurred */
nerr("ERROR: Failed to allocate callback\n");
errcode = ENOMEM;
goto errout_with_wrb;
}
/* Set up the callback in the connection */
psock->s_sndcb->flags = (TCP_ACKDATA | TCP_REXMIT | TCP_POLL |
TCP_DISCONN_EVENTS);
psock->s_sndcb->priv = (FAR void *)psock;
psock->s_sndcb->event = psock_send_interrupt;
/* Initialize the write buffer */
WRB_SEQNO(wrb) = (unsigned)-1;
WRB_NRTX(wrb) = 0;
result = WRB_COPYIN(wrb, (FAR uint8_t *)buf, len);
/* Dump I/O buffer chain */
WRB_DUMP("I/O buffer chain", wrb, WRB_PKTLEN(wrb), 0);
/* psock_send_interrupt() will send data in FIFO order from the
* conn->write_q
*/
sq_addlast(&wrb->wb_node, &conn->write_q);
ninfo("Queued WRB=%p pktlen=%u write_q(%p,%p)\n",
wrb, WRB_PKTLEN(wrb),
conn->write_q.head, conn->write_q.tail);
/* Notify the device driver of the availability of TX data */
send_txnotify(psock, conn);
net_unlock(save);
}
/* Set the socket state to idle */
psock->s_flags = _SS_SETSTATE(psock->s_flags, _SF_IDLE);
/* Check for errors. Errors are signalled by negative errno values
* for the send length
*/
if (result < 0)
{
errcode = result;
goto errout;
}
/* If net_lockedwait failed, then we were probably reawakened by a signal.
* In this case, net_lockedwait will have set errno appropriately.
*/
if (ret < 0)
{
errcode = -ret;
goto errout;
}
/* Return the number of bytes actually sent */
return result;
errout_with_wrb:
tcp_wrbuffer_release(wrb);
errout_with_lock:
net_unlock(save);
errout:
set_errno(errcode);
return ERROR;
}
void tcp_free(FAR struct tcp_conn_s *conn)
{
FAR struct devif_callback_s *cb;
FAR struct devif_callback_s *next;
#ifdef CONFIG_NET_TCP_WRITE_BUFFERS
FAR struct tcp_wrbuffer_s *wrbuffer;
#endif
net_lock_t flags;
/* Because g_free_tcp_connections is accessed from user level and interrupt
* level, code, it is necessary to keep interrupts disabled during this
* operation.
*/
DEBUGASSERT(conn->crefs == 0);
flags = net_lock();
/* Free remaining callbacks, actually there should be only the close callback
* left.
*/
for (cb = conn->list; cb; cb = next)
{
next = cb->flink;
tcp_callback_free(conn, cb);
}
/* TCP_ALLOCATED means that that the connection is not in the active list
* yet.
*/
if (conn->tcpstateflags != TCP_ALLOCATED)
{
/* Remove the connection from the active list */
dq_rem(&conn->node, &g_active_tcp_connections);
}
#ifdef CONFIG_NET_TCP_READAHEAD
/* Release any read-ahead buffers attached to the connection */
iob_free_queue(&conn->readahead);
#endif
#ifdef CONFIG_NET_TCP_WRITE_BUFFERS
/* Release any write buffers attached to the connection */
while ((wrbuffer = (struct tcp_wrbuffer_s *)sq_remfirst(&conn->write_q)) != NULL)
{
tcp_wrbuffer_release(wrbuffer);
}
while ((wrbuffer = (struct tcp_wrbuffer_s *)sq_remfirst(&conn->unacked_q)) != NULL)
{
tcp_wrbuffer_release(wrbuffer);
}
#endif
#ifdef CONFIG_NET_TCPBACKLOG
/* Remove any backlog attached to this connection */
if (conn->backlog)
{
tcp_backlogdestroy(conn);
}
/* If this connection is, itself, backlogged, then remove it from the
* parent connection's backlog list.
*/
if (conn->blparent)
{
tcp_backlogdelete(conn->blparent, conn);
}
#endif
/* Mark the connection available and put it into the free list */
conn->tcpstateflags = TCP_CLOSED;
dq_addlast(&conn->node, &g_free_tcp_connections);
net_unlock(flags);
}
