void citp_waitable_print(citp_waitable* w)
{
/* Output socket using netstat style output:
* TCP 2 0 0.0.0.0:12865 0.0.0.0:0 LISTEN
* UDP 0 0 172.16.129.131:57521 0.0.0.0:0 UDP
*/
if( CI_TCP_STATE_IS_SOCKET(w->state) ) {
ci_sock_cmn* s = CI_CONTAINER(ci_sock_cmn, b, w);
citp_waitable_obj* wo = CI_CONTAINER(citp_waitable_obj, waitable, w);
int tq = 0;
int rq = 0;
if( (w->state & CI_TCP_STATE_TCP) &&
!(w->state & CI_TCP_STATE_NOT_CONNECTED) ) {
tq = ci_tcp_sendq_n_pkts(&wo->tcp);
rq = wo->tcp.recv1.num + wo->tcp.recv2.num;
}
else if( w->state == CI_TCP_STATE_UDP ) {
tq = wo->udp.tx_count + oo_atomic_read(&wo->udp.tx_async_q_level);
rq = ci_udp_recv_q_pkts(&wo->udp.recv_q);
}
log("%s %d %d "OOF_IP4PORT" "OOF_IP4PORT" %s",
citp_waitable_type_str(w), rq, tq,
OOFA_IP4PORT(sock_laddr_be32(s), sock_lport_be16(s)),
OOFA_IP4PORT(sock_raddr_be32(s), sock_rport_be16(s)),
ci_tcp_state_str(w->state));
}
}
void __citp_fdinfo_ref_count_zero(citp_fdinfo* fdi, int fdt_locked)
{
Log_V(log("%s: fd=%d on_rcz=%d", __FUNCTION__, fdi->fd,
fdi->on_ref_count_zero));
citp_fdinfo_assert_valid(fdi);
ci_assert(oo_atomic_read(&fdi->ref_count) == 0);
ci_assert_ge(fdi->fd, 0);
ci_assert_lt(fdi->fd, citp_fdtable.inited_count);
ci_assert_nequal(fdi_to_fdip(fdi), citp_fdtable.table[fdi->fd].fdip);
switch( fdi->on_ref_count_zero ) {
case FDI_ON_RCZ_CLOSE:
#if CI_CFG_FD_CACHING
if( citp_fdinfo_get_ops(fdi)->cache(fdi) == 1 ) {
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_LOCK();
fdtable_swap(fdi->fd, fdip_closing, fdip_unknown,
fdt_locked | fdtable_strict());
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked | fdtable_strict());
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_UNLOCK();
citp_fdinfo_free(fdi);
break;
}
else
#endif
{
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_LOCK();
ci_tcp_helper_close_no_trampoline(fdi->fd);
/* The swap must occur after the close, otherwise another thread could
* cause a probe of the old endpoint info, which is about be freed.
*/
fdtable_swap(fdi->fd, fdip_closing, fdip_unknown,
fdt_locked | fdtable_strict());
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked | fdtable_strict());
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_UNLOCK();
citp_fdinfo_free(fdi);
break;
}
case FDI_ON_RCZ_DUP2:
dup2_complete(fdi, fdi_to_fdip(fdi), fdt_locked);
break;
case FDI_ON_RCZ_HANDOVER:
citp_fdinfo_do_handover(fdi, fdt_locked);
break;
case FDI_ON_RCZ_MOVED:
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked);
citp_fdinfo_free(fdi);
break;
default:
CI_DEBUG(ci_log("%s: fd=%d on_ref_count_zero=%d", __FUNCTION__,
fdi->fd, fdi->on_ref_count_zero));
ci_assert(0);
}
}
void citp_fdtable_insert(citp_fdinfo* fdi, unsigned fd, int fdt_locked)
{
ci_assert(fdi);
ci_assert(fdi->protocol);
ci_assert(citp_fdtable.inited_count > fd);
ci_assert_ge(oo_atomic_read(&fdi->ref_count), 1);
fdi->fd = fd;
CI_DEBUG(fdi->on_ref_count_zero = FDI_ON_RCZ_NONE);
fdi->is_special = 0;
citp_fdtable_busy_clear(fd, fdi_to_fdip(fdi), fdt_locked);
}
/* This function does some simple tests to ensure that the fdtable makes sense.
* There are many more tests we could do; feel free to add them at your
* leisure!
*/
void
citp_fdtable_assert_valid(void)
{
int i;
if( ! citp_fdtable.table ) return;
CITP_FDTABLE_LOCK_RD();
for( i = 0; i < citp_fdtable.inited_count; i++ ) {
citp_fdinfo_p fdip = citp_fdtable.table[i].fdip;
if( fdip_is_normal(fdip) ) {
citp_fdinfo * fdi = fdip_to_fdi(fdip);
ci_assert(fdi);
ci_assert(fdi->protocol);
if( ( fdi->protocol->type == CITP_TCP_SOCKET ||
fdi->protocol->type == CITP_UDP_SOCKET )
&& fdi_to_socket(fdi)->s )
ci_assert(! (fdi_to_socket(fdi)->s->b.sb_aflags & CI_SB_AFLAG_ORPHAN));
if (!fdi->is_special) {
/* Ensure the "back pointer" makes sense */
ci_assert (fdi->fd == i);
/* Ensure that the reference count is in a vaguely sensible range */
ci_assert ((oo_atomic_read (&fdi->ref_count) > 0) &&
(oo_atomic_read (&fdi->ref_count) < 10000));
/* 10,000 threads is a bit mad, warn if more than 20 */
if (oo_atomic_read (&fdi->ref_count) > 20) {
Log_U (log ("Warning: fd %d's ref-count suspiciously large (%d)\n",
i, oo_atomic_read (&fdi->ref_count)));
}
}
}
}
CITP_FDTABLE_UNLOCK_RD();
}
/* 2004/08/16 stg: added [fdt_locked] to allow the fd table to be
* locked before this function. [fdt_locked] = 0 for legacy operation
*/
void __citp_epinfo_ref_count_zero(citp_epinfo* epinfo,citp_fdinfo* last_fdinfo,
int fdt_locked)
{
Log_V(log(LPF "ref_count_zero(%p, %d)", epinfo, last_fdinfo->fd));
ci_assert(epinfo);
ci_assert(oo_atomic_read(&epinfo->ref_count) == 0);
ci_assert(epinfo->protocol);
ci_assert(last_fdinfo);
ci_assert(last_fdinfo->ep == epinfo);
epinfo->protocol->ops.dtor(epinfo, last_fdinfo, fdt_locked);
}
static int efab_file_move_supported_tcp(ci_netif *ni, ci_tcp_state *ts)
{
#if CI_CFG_FD_CACHING
/* Don't support moving cached sockets for now */
if( ci_tcp_is_cached(ts) ||
!ci_ni_dllist_is_self_linked(ni, &ts->epcache_link) )
return false;
#endif
/* TCP closed: supported */
if( ts->s.b.state == CI_TCP_CLOSED )
return true;
/* everything except TCP connected is not supported */
if( !(ts->s.b.state & CI_TCP_STATE_TCP_CONN) )
return false;
if( ts->local_peer != OO_SP_NULL )
return false;
if( !(ts->tcpflags & CI_TCPT_FLAG_PASSIVE_OPENED) )
return false;
/* send queue is not supported
* NB: retrans_ptr is uninitialised when retrans was not used yet,
* so do not check for !OO_PP_IS_NULL(ts->retrans_ptr) */
if( !ci_ip_queue_is_empty(&ts->send) ||
ts->send_prequeue != OO_PP_ID_NULL ||
oo_atomic_read(&ts->send_prequeue_in) != 0 ||
!ci_ip_queue_is_empty(&ts->retrans) ||
ci_ip_timer_pending(ni, &ts->rto_tid) ||
ci_ip_timer_pending(ni, &ts->zwin_tid) ||
#if CI_CFG_TAIL_DROP_PROBE
ci_ip_timer_pending(ni, &ts->taildrop_tid) ||
#endif
ci_ip_timer_pending(ni, &ts->cork_tid) )
return false;
/* Sockets with allocated templates are not supported */
if( OO_PP_NOT_NULL(ts->tmpl_head) )
return false;
return true;
}
citp_fdinfo*
citp_fdtable_lookup_fast(citp_lib_context_t* ctx, unsigned fd)
{
/* Note that if we haven't yet initialised this module, then
** [inited_count] will be zero, and the following test will fail. So the
** test for initialisation is done further down...
**
** This is highly performance critial. DO NOT add any code between here
** and the first [return] statement.
*/
citp_fdinfo* fdi;
/* Try to avoid entering lib. */
ctx->thread = NULL;
if( fd < citp_fdtable.inited_count ) {
volatile citp_fdinfo_p* p_fdip = &citp_fdtable.table[fd].fdip;
citp_fdinfo_p fdip;
again:
fdip = *p_fdip;
if( fdip_is_normal(fdip) ) {
citp_enter_lib_if(ctx);
if( citp_fdtable_is_mt_safe() ) {
/* No need to use atomic ops or add a ref to the fdi when MT-safe.
* The definition of "fds_mt_safe" is that the app does not change
* the meaning of a file descriptor in one thread when it is being
* used in another thread.
*/
fdi = fdip_to_fdi(fdip);
if( ! citp_fdinfo_is_consistent(fdi) )
fdi = citp_reprobe_moved(fdi, CI_TRUE, CI_FALSE);
return fdi;
}
else {
/* Swap in the busy marker. */
if( fdip_cas_succeed(p_fdip, fdip, fdip_busy) ) {
fdi = fdip_to_fdi(fdip);
ci_assert(fdi);
ci_assert_gt(oo_atomic_read(&fdi->ref_count), 0);
ci_assert(fdip_is_closing(fdip) || fdip_is_reserved(fdip) ||
fdi->fd == fd);
/* Bump the reference count. */
citp_fdinfo_ref(fdi);
if( ! citp_fdinfo_is_consistent(fdi) )
fdi = citp_reprobe_moved(fdi, CI_FALSE, CI_TRUE);
else {
/* Swap the busy marker out again. */
citp_fdtable_busy_clear(fd, fdip, 0);
}
return fdi;
}
goto again;
}
}
/* Not normal! */
if( fdip_is_passthru(fdip) )
return NULL;
citp_enter_lib_if(ctx);
if( fdip_is_busy(fdip) ) {
citp_fdtable_busy_wait(fd, 0);
goto again;
}
ci_assert(fdip_is_unknown(fdip));
goto probe;
}
if( citp.init_level < CITP_INIT_FDTABLE ) {
if( _citp_do_init_inprogress == 0 )
CI_TRY(citp_do_init(CITP_INIT_ALL));
else
CI_TRY(citp_do_init(CITP_INIT_FDTABLE)); /* get what we need */
}
if( fd >= citp_fdtable.size )
return NULL;
probe:
citp_enter_lib_if(ctx);
fdi = citp_fdtable_probe(fd);
if( fdi && citp_fdtable_is_mt_safe() )
citp_fdinfo_release_ref(fdi, 0);
return fdi;
}
citp_fdinfo *
citp_fdtable_lookup(unsigned fd)
{
/* Note that if we haven't yet initialised this module, then
** [inited_count] will be zero, and the following test will fail. So the
** test for initialisation is done further down...
**
** This is highly performance critial. DO NOT add any code between here
** and the first [return] statement.
*/
citp_fdinfo* fdi;
/* In some cases, we'll lock fdtable. Assert that it is possible: */
ci_assert(oo_per_thread_get()->sig.inside_lib);
if( fd < citp_fdtable.inited_count ) {
volatile citp_fdinfo_p* p_fdip = &citp_fdtable.table[fd].fdip;
citp_fdinfo_p fdip;
again:
/* Swap in the busy marker. */
fdip = *p_fdip;
if( fdip_is_normal(fdip) ) {
if( citp_fdtable_not_mt_safe() ) {
if( fdip_cas_succeed(p_fdip, fdip, fdip_busy) ) {
fdi = fdip_to_fdi(fdip);
ci_assert(fdi);
ci_assert_gt(oo_atomic_read(&fdi->ref_count), 0);
ci_assert(fdip_is_closing(fdip) || fdip_is_reserved(fdip) ||
fdi->fd == fd);
/* Bump the reference count. */
citp_fdinfo_ref(fdi);
if( ! citp_fdinfo_is_consistent(fdi) ) {
/* Something is wrong. Re-probe. */
fdi = citp_reprobe_moved(fdi, CI_FALSE, CI_TRUE);
}
else {
/* Swap the busy marker out again. */
citp_fdtable_busy_clear(fd, fdip, 0);
}
return fdi;
}
goto again;
}
else {
/* No need to use atomic ops when single-threaded. The definition
* of "fds_mt_safe" is that the app does not change the meaning of
* a file descriptor in one thread when it is being used in another
* thread. In that case I'm hoping this should be safe, but at
* time of writing I'm really not confident. (FIXME).
*/
fdi = fdip_to_fdi(fdip);
if( ci_is_multithreaded() )
citp_fdinfo_ref(fdi);
else
++fdi->ref_count.n;
if( ! citp_fdinfo_is_consistent(fdi) )
fdi = citp_reprobe_moved(fdi, CI_FALSE, CI_FALSE);
return fdi;
}
}
/* Not normal! */
if( fdip_is_passthru(fdip) ) return NULL;
if( fdip_is_busy(fdip) ) {
citp_fdtable_busy_wait(fd, 0);
goto again;
}
ci_assert(fdip_is_unknown(fdip));
goto probe;
}
if (citp.init_level < CITP_INIT_FDTABLE) {
if (_citp_do_init_inprogress == 0)
CI_TRY(citp_do_init(CITP_INIT_ALL));
else
CI_TRY(citp_do_init(CITP_INIT_FDTABLE)); /* get what we need */
}
if( fd >= citp_fdtable.size ) return NULL;
probe:
fdi = citp_fdtable_probe(fd);
return fdi;
}
int citp_ep_dup3(unsigned fromfd, unsigned tofd, int flags)
{
volatile citp_fdinfo_p* p_tofdip;
citp_fdinfo_p tofdip;
unsigned max;
Log_V(log("%s(%d, %d)", __FUNCTION__, fromfd, tofd));
/* Must be checked by callers. */
ci_assert(fromfd != tofd);
/* Hack: if [tofd] is the fd we're using for logging, we'd better choose
** a different one!
*/
if( tofd == citp.log_fd ) citp_log_change_fd();
ci_assert(citp.init_level >= CITP_INIT_FDTABLE);
max = CI_MAX(fromfd, tofd);
if( max >= citp_fdtable.inited_count ) {
ci_assert(max < citp_fdtable.size);
CITP_FDTABLE_LOCK();
__citp_fdtable_extend(max);
CITP_FDTABLE_UNLOCK();
}
/* Bug1151: Concurrent threads doing dup2(x,y) and dup2(y,x) can deadlock
** against one another. So we take out a fat lock to prevent concurrent
** dup2()s.
*/
/* Lock tofd. We need to interlock against select and poll etc, so we
** also grab the exclusive lock. Also grab the bug1151 lock.
*/
pthread_mutex_lock(&citp_dup_lock);
CITP_FDTABLE_LOCK();
p_tofdip = &citp_fdtable.table[tofd].fdip;
lock_tofdip_again:
tofdip = *p_tofdip;
if( fdip_is_busy(tofdip) )
tofdip = citp_fdtable_busy_wait(tofd, 1);
if( fdip_is_closing(tofdip) )
tofdip = citp_fdtable_closing_wait(tofd, 1);
if( fdip_is_reserved(tofdip) ) {
/* ?? FIXME: we can't cope with this at the moment */
CITP_FDTABLE_UNLOCK();
Log_U(log("%s(%d, %d): target is reserved", __FUNCTION__, fromfd, tofd));
errno = EBUSY;
tofd = -1;
goto out;
}
if( fdip_cas_fail(p_tofdip, tofdip, fdip_busy) )
goto lock_tofdip_again;
CITP_FDTABLE_UNLOCK();
ci_assert(fdip_is_normal(tofdip) | fdip_is_passthru(tofdip) |
fdip_is_unknown(tofdip));
if( fdip_is_normal(tofdip) ) {
/* We're duping onto a user-level socket. */
citp_fdinfo* tofdi = fdip_to_fdi(tofdip);
if( tofdi->epoll_fd >= 0 ) {
citp_fdinfo* epoll_fdi = citp_epoll_fdi_from_member(tofdi, 0);
if( epoll_fdi ) {
if( epoll_fdi->protocol->type == CITP_EPOLL_FD )
citp_epoll_on_close(epoll_fdi, tofdi, 0);
citp_fdinfo_release_ref(epoll_fdi, 0);
}
}
ci_assert_equal(tofdi->on_ref_count_zero, FDI_ON_RCZ_NONE);
tofdi->on_ref_count_zero = FDI_ON_RCZ_DUP2;
tofdi->on_rcz.dup3_args.fd = fromfd;
tofdi->on_rcz.dup3_args.flags = flags;
citp_fdinfo_release_ref(tofdi, 0);
{
int i = 0;
/* We need to free this fdi. If someone is using it right now,
* we are in trouble. So, we spin for a while and interrupt the
* user. See bug 28123. */
while( tofdi->on_ref_count_zero != FDI_ON_RCZ_DONE ) {
if( ci_is_multithreaded() && i % 10000 == 9999 ) {
pthread_t pth = tofdi->thread_id;
if( pth != pthread_self() && pth != PTHREAD_NULL ) {
pthread_kill(pth, SIGONLOAD);
sleep(1);
}
}
ci_spinloop_pause();
i++;
}
ci_rmb();
}
if( tofdi->on_rcz.dup2_result < 0 ) {
errno = -tofdi->on_rcz.dup2_result;
/* Need to re-insert [tofdi] into the table. */
ci_assert_equal(oo_atomic_read(&tofdi->ref_count), 0);
oo_atomic_set(&tofdi->ref_count, 1);
CI_DEBUG(tofdi->on_ref_count_zero = FDI_ON_RCZ_NONE);
citp_fdtable_busy_clear(tofd, tofdip, 0);
tofd = -1;
}
else {
ci_assert(tofdi->on_rcz.dup2_result == tofd);
citp_fdinfo_get_ops(tofdi)->dtor(tofdi, 0);
citp_fdinfo_free(tofdi);
}
goto out;
}
ci_assert(fdip_is_passthru(tofdip) | fdip_is_unknown(tofdip));
{ /* We're dupping onto an O/S descriptor, or it may be closed. Create a
** dummy [citp_fdinfo], just so we can share code with the case above.
*/
citp_fdinfo fdi;
fdi.fd = tofd;
fdi.on_rcz.dup3_args.fd = fromfd;
fdi.on_rcz.dup3_args.flags = flags;
dup2_complete(&fdi, tofdip, 0);
if( fdi.on_rcz.dup2_result < 0 ) {
errno = -fdi.on_rcz.dup2_result;
citp_fdtable_busy_clear(tofd, tofdip, 0);
tofd = -1;
}
else
ci_assert(fdi.on_rcz.dup2_result == tofd);
}
out:
pthread_mutex_unlock(&citp_dup_lock);
return tofd;
}
void ci_udp_state_dump(ci_netif* ni, ci_udp_state* us, const char* pf,
oo_dump_log_fn_t logger, void* log_arg)
{
ci_udp_socket_stats uss = us->stats;
unsigned rx_added = us->recv_q.pkts_added;
unsigned rx_os = uss.n_rx_os + uss.n_rx_os_slow;
unsigned rx_total = rx_added + uss.n_rx_mem_drop + uss.n_rx_overflow + rx_os;
unsigned n_tx_onload = uss.n_tx_onload_uc + uss.n_tx_onload_c;
unsigned tx_total = n_tx_onload + uss.n_tx_os;
ci_ip_cached_hdrs* ipcache;
(void) rx_total; /* unused on 32-bit builds in kernel */
(void) tx_total;
#if CI_CFG_TIMESTAMPING
if( us->s.timestamping_flags & ONLOAD_SOF_TIMESTAMPING_TX_HARDWARE )
ci_udp_recvq_dump(ni, &us->timestamp_q, pf, " TX timestamping queue:",
logger, log_arg);
#endif
/* General. */
logger(log_arg, "%s udpflags: "CI_UDP_STATE_FLAGS_FMT, pf,
CI_UDP_STATE_FLAGS_PRI_ARG(us));
/* Receive path. */
ci_udp_recvq_dump(ni, &us->recv_q, pf, " rcv:", logger, log_arg);
logger(log_arg,
"%s rcv: oflow_drop=%u(%u%%) mem_drop=%u eagain=%u pktinfo=%u "
"q_max_pkts=%u", pf, uss.n_rx_overflow,
percent(uss.n_rx_overflow, rx_total),
uss.n_rx_mem_drop, uss.n_rx_eagain, uss.n_rx_pktinfo,
uss.max_recvq_pkts);
logger(log_arg, "%s rcv: os=%u(%u%%) os_slow=%u os_error=%u", pf,
rx_os, percent(rx_os, rx_total), uss.n_rx_os_slow, uss.n_rx_os_error);
/* Send path. */
logger(log_arg, "%s snd: q=%u+%u ul=%u os=%u(%u%%)", pf,
us->tx_count, oo_atomic_read(&us->tx_async_q_level),
n_tx_onload, uss.n_tx_os, percent(uss.n_tx_os, tx_total));
logger(log_arg,
"%s snd: LOCK cp=%u(%u%%) pkt=%u(%u%%) snd=%u(%u%%) poll=%u(%u%%) "
"defer=%u(%u%%)", pf,
uss.n_tx_lock_cp, percent(uss.n_tx_lock_cp, n_tx_onload),
uss.n_tx_lock_pkt, percent(uss.n_tx_lock_pkt, n_tx_onload),
uss.n_tx_lock_snd, percent(uss.n_tx_lock_snd, n_tx_onload),
uss.n_tx_lock_poll, percent(uss.n_tx_lock_poll, n_tx_onload),
uss.n_tx_lock_defer, percent(uss.n_tx_lock_defer, n_tx_onload));
logger(log_arg, "%s snd: MCAST if=%d src="OOF_IP4" ttl=%d", pf,
us->s.cp.ip_multicast_if,
OOFA_IP4(us->s.cp.ip_multicast_if_laddr_be32),
(int) us->s.cp.ip_mcast_ttl);
/* State relating to unconnected sends. */
ipcache = &us->ephemeral_pkt;
logger(log_arg,
"%s snd: TO n=%u match=%u(%u%%) "
"lookup=%u+%u(%u%%) "OOF_IPCACHE_STATE,
pf, uss.n_tx_onload_uc,
uss.n_tx_cp_match, percent(uss.n_tx_cp_match, uss.n_tx_onload_uc),
uss.n_tx_cp_uc_lookup, uss.n_tx_cp_a_lookup,
percent(uss.n_tx_cp_uc_lookup + uss.n_tx_cp_a_lookup,
uss.n_tx_onload_uc),
OOFA_IPCACHE_STATE(ni, ipcache));
logger(log_arg, "%s snd: TO "OOF_IPCACHE_DETAIL, pf,
OOFA_IPCACHE_DETAIL(ipcache));
logger(log_arg, "%s snd: TO "OOF_IP4PORT" => "OOF_IP4PORT, pf,
OOFA_IP4PORT(ipcache->ip_saddr.ip4, udp_lport_be16(us)),
OOFA_IP4PORT(ipcache->ip.ip_daddr_be32, ipcache->dport_be16));
/* State relating to connected sends. */
ipcache = &us->s.pkt;
logger(log_arg, "%s snd: CON n=%d lookup=%d "OOF_IPCACHE_STATE, pf,
uss.n_tx_onload_c, uss.n_tx_cp_c_lookup,
OOFA_IPCACHE_STATE(ni,ipcache));
logger(log_arg, "%s snd: CON "OOF_IPCACHE_DETAIL, pf,
OOFA_IPCACHE_DETAIL(ipcache));
logger(log_arg, "%s snd: eagain=%d spin=%d block=%d", pf,
uss.n_tx_eagain, uss.n_tx_spin, uss.n_tx_block);
logger(log_arg, "%s snd: poll_avoids_full=%d fragments=%d confirm=%d", pf,
uss.n_tx_poll_avoids_full, uss.n_tx_fragments, uss.n_tx_msg_confirm);
logger(log_arg,
"%s snd: os_slow=%d os_late=%d unconnect_late=%d nomac=%u(%u%%)", pf,
uss.n_tx_os_slow, uss.n_tx_os_late, uss.n_tx_unconnect_late,
uss.n_tx_cp_no_mac, percent(uss.n_tx_cp_no_mac, tx_total));
}
static int ci_udp_ioctl_locked(ci_netif* ni, ci_udp_state* us,
ci_fd_t fd, int request, void* arg)
{
int rc;
switch( request ) {
case FIONREAD: /* synonym of SIOCINQ */
if( ! CI_IOCTL_ARG_OK(int, arg) )
return -EFAULT;
rc = 1;
if( rc ) {
/* Return the size of the datagram at the head of the receive queue.
*
* Careful: extract side of receive queue is owned by sock lock,
* which we don't have. However, freeing of bufs is owned by netif
* lock, which we do have. So we're safe so long as we only read
* [extract] once.
*/
oo_pkt_p extract = us->recv_q.extract;
if( OO_PP_NOT_NULL(extract) ) {
ci_ip_pkt_fmt* pkt = PKT_CHK(ni, extract);
if( (pkt->rx_flags & CI_PKT_RX_FLAG_RECV_Q_CONSUMED) &&
OO_PP_NOT_NULL(pkt->udp_rx_next) )
pkt = PKT_CHK(ni, pkt->udp_rx_next);
if( !(pkt->rx_flags & CI_PKT_RX_FLAG_RECV_Q_CONSUMED) ) {
*(int*) arg = pkt->pf.udp.pay_len;
return 0;
}
}
}
/* Nothing in userlevel receive queue: So take the value returned by
* the O/S socket.
*/
if( !(us->s.os_sock_status & OO_OS_STATUS_RX) ) {
*(int*)arg = 0;
return 0;
}
goto sys_ioctl;
case TIOCOUTQ: /* synonym of SIOCOUTQ */
if( ! CI_IOCTL_ARG_OK(int, arg) )
return -EFAULT;
*(int*)arg = us->tx_count + oo_atomic_read(&us->tx_async_q_level);
return 0;
case SIOCGSTAMP:
#if defined( __linux__) && defined(__KERNEL__)
/* The following code assumes the width of the timespec and timeval fields */
# error "Need to consider 32-on-64 bit setting of timeval arg"
#endif
if( ! (us->udpflags & CI_UDPF_LAST_RECV_ON) )
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
return ci_udp_ioctl_siocgstamp(ni, us, arg, 1);
case SIOCGSTAMPNS:
if( ! (us->udpflags & CI_UDPF_LAST_RECV_ON) )
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
return ci_udp_ioctl_siocgstamp(ni, us, arg, 0);
}
return ci_udp_ioctl_slow(ni, us, fd, request, arg);
sys_ioctl:
return oo_os_sock_ioctl(ni, us->s.b.bufid, request, arg, NULL);
}
static int
efab_tcp_helper_get_info(ci_private_t *unused, void *arg)
{
ci_netif_info_t *info = arg;
int index, rc=0;
tcp_helper_resource_t* thr = NULL;
ci_netif* ni = NULL;
int flags = EFAB_THR_TABLE_LOOKUP_CHECK_USER | EFAB_THR_TABLE_LOOKUP_NO_WARN;
#if CI_CFG_EFAB_EPLOCK_RECORD_CONTENTIONS
int j;
eplock_resource_t* eplock_rs;
#endif
info->ni_exists = 0;
info->ni_no_perms_exists = 0;
if( info->ni_orphan ) {
flags |= EFAB_THR_TABLE_LOOKUP_NO_UL;
info->ni_orphan = 0;
}
rc = efab_thr_table_lookup(NULL, info->ni_index, flags, &thr);
if( rc == 0 ) {
info->ni_exists = 1;
info->ni_orphan = (thr->k_ref_count & TCP_HELPER_K_RC_NO_USERLAND);
ni = &thr->netif;
info->mmap_bytes = thr->mem_mmap_bytes;
info->k_ref_count = thr->k_ref_count;
info->rs_ref_count = oo_atomic_read(&thr->ref_count);
memcpy(info->ni_name, ni->state->name, sizeof(ni->state->name));
} else if( rc == -EACCES ) {
info->ni_no_perms_id = info->ni_index;
if( efab_thr_get_inaccessible_stack_info(info->ni_index,
&info->ni_no_perms_uid,
&info->ni_no_perms_euid,
&info->ni_no_perms_share_with,
info->ni_no_perms_name) == 0 )
info->ni_no_perms_exists = 1;
}
/* sub-ops that do not need the netif to exist */
if( info->ni_subop == CI_DBG_NETIF_INFO_GET_NEXT_NETIF ) {
tcp_helper_resource_t* next_thr;
info->u.ni_next_ni.index = -1;
for( index = info->ni_index + 1;
index < 10000 /* FIXME: magic! */;
++index ) {
rc = efab_thr_table_lookup(NULL, index, flags, &next_thr);
if( rc == 0 ) {
if( next_thr->k_ref_count & TCP_HELPER_K_RC_NO_USERLAND )
efab_tcp_helper_k_ref_count_dec(next_thr, 1);
else
efab_thr_release(next_thr);
info->u.ni_next_ni.index = index;
break;
}
if( rc == -EACCES ) {
info->u.ni_next_ni.index = index;
break;
}
}
rc = 0;
}
else if( info->ni_subop == CI_DBG_NETIF_INFO_NOOP ) {
rc = 0;
}
if (!info->ni_exists)
return 0;
/* sub-ops that need the netif to exist */
switch (info->ni_subop)
{
case CI_DBG_NETIF_INFO_GET_ENDPOINT_STATE:
index = info->u.ni_endpoint.index;
info->u.ni_endpoint.max = thr->netif.ep_tbl_n;
if ((index < 0) || (index >= (int)thr->netif.ep_tbl_n)) {
info->u.ni_endpoint.state = CI_TCP_STATE_FREE;
}
else {
citp_waitable_obj* wo = ID_TO_WAITABLE_OBJ(ni, index);
info->u.ni_endpoint.state = wo->waitable.state;
if( wo->waitable.state == CI_TCP_STATE_UDP ) {
ci_udp_state* us = &wo->udp;
info->u.ni_endpoint.udpstate = us->udpflags;
info->u.ni_endpoint.rx_pkt_ul = us->recv_q.pkts_delivered;
info->u.ni_endpoint.rx_pkt_kn = us->stats.n_rx_os;
}
else if( wo->waitable.state & CI_TCP_STATE_TCP_CONN ) {
ci_tcp_state* ts = &wo->tcp;
info->u.ni_endpoint.tx_pkts_max = ts->so_sndbuf_pkts;
info->u.ni_endpoint.tx_pkts_num = ts->send.num;
}
if( wo->waitable.state & CI_TCP_STATE_SOCKET ) {
ci_sock_cmn* s = &wo->sock;
info->u.ni_endpoint.protocol = (int) sock_protocol(s);
info->u.ni_endpoint.laddr = sock_laddr_be32(s);
info->u.ni_endpoint.lport = (int) sock_lport_be16(s);
info->u.ni_endpoint.raddr = sock_raddr_be32(s);
info->u.ni_endpoint.rport = (int) sock_rport_be16(s);
}
}
break;
case CI_DBG_NETIF_INFO_GET_NEXT_NETIF:
/* If the current netif is found, we need to succeed */
break;
case CI_DBG_NETIF_INFO_NOOP:
/* Always succeeds, rc already set */
break;
default:
rc = -EINVAL;
break;
}
if( thr ) {
/* Lookup needs a matching efab_thr_release() in case of ordinary
* stack but just a ref_count_dec in case of orphan
*/
if( thr->k_ref_count & TCP_HELPER_K_RC_NO_USERLAND )
efab_tcp_helper_k_ref_count_dec(thr, 1);
else
efab_thr_release(thr);
}
return rc;
}
