void
ci_netif_filter_remove(ci_netif* netif, oo_sp sock_p,
unsigned laddr, unsigned lport,
unsigned raddr, unsigned rport, unsigned protocol)
{
ci_netif_filter_table_entry* entry;
unsigned hash1, hash2, tbl_i;
ci_netif_filter_table* tbl;
int hops = 0;
unsigned first;
ci_assert(ci_netif_is_locked(netif)
#ifdef __KERNEL__
/* release_ep_tbl might be called without the stack lock.
* Do not complain about this. */
|| (netif2tcp_helper_resource(netif)->k_ref_count &
TCP_HELPER_K_RC_DEAD)
#endif
);
tbl = netif->filter_table;
hash1 = tcp_hash1(tbl, laddr, lport, raddr, rport, protocol);
hash2 = tcp_hash2(tbl, laddr, lport, raddr, rport, protocol);
first = hash1;
LOG_TC(ci_log("%s: [%d:%d] REMOVE %s %s:%u->%s:%u hash=%u:%u",
__FUNCTION__, NI_ID(netif), OO_SP_FMT(sock_p),
CI_IP_PROTOCOL_STR(protocol),
ip_addr_str(laddr), (unsigned) CI_BSWAP_BE16(lport),
ip_addr_str(raddr), (unsigned) CI_BSWAP_BE16(rport),
hash1, hash2));
tbl_i = hash1;
while( 1 ) {
entry = &tbl->table[tbl_i];
if( entry->id == OO_SP_TO_INT(sock_p) ) {
if( laddr == entry->laddr )
break;
}
else if( entry->id == EMPTY ) {
/* We allow multiple removes of the same filter -- helps avoid some
* complexity in the filter module.
*/
return;
}
tbl_i = (tbl_i + hash2) & tbl->table_size_mask;
++hops;
if( tbl_i == first ) {
LOG_E(ci_log(FN_FMT "ERROR: LOOP [%d] %s %s:%u->%s:%u",
FN_PRI_ARGS(netif), OO_SP_FMT(sock_p),
CI_IP_PROTOCOL_STR(protocol),
ip_addr_str(laddr), (unsigned) CI_BSWAP_BE16(lport),
ip_addr_str(raddr), (unsigned) CI_BSWAP_BE16(rport)));
return;
}
}
__ci_netif_filter_remove(netif, hash1, hash2, hops, tbl_i);
}
void
oof_cb_sw_filter_remove(struct oof_socket* skf, unsigned laddr, int lport,
unsigned raddr, int rport, int protocol,
int stack_locked)
{
ci_netif* ni = skf_to_ni(skf);
struct tcp_helper_resource_s *trs = netif2tcp_helper_resource(ni);
if( skf->sf_flags & OOF_SOCKET_SW_FILTER_WAS_REMOVED )
return;
/* We MAY call this function with incorrect stack_locked flag
* if OOF_SOCKET_SW_FILTER_WAS_REMOVED flag is set. */
ci_assert(!stack_locked || ci_netif_is_locked(ni));
/* We are holding a spinlock, so claim to be in driverlink context here */
if( stack_locked || efab_tcp_helper_netif_try_lock(trs, 1) ) {
ci_netif_filter_remove(ni, OO_SP_FROM_INT(ni, skf_to_ep(skf)->id),
laddr, lport, raddr, rport, protocol);
if( ! stack_locked )
efab_tcp_helper_netif_unlock(trs, 1);
}
else
oof_cb_sw_filter_postpone(skf, laddr, lport, raddr, rport, protocol,
OOF_CB_SW_FILTER_OP_REMOVE);
}
static void ci_drop_orphan(ci_netif * ni)
{
ci_irqlock_state_t lock_flags;
tcp_helper_resource_t* trs;
int dec_needed;
/* Called when connection closes AFTER the file descriptor closes
* - in kernel mode, if user mode has gone away, we call
* efab_tcp_helper_k_ref_count_dec() to decrement count
* of such connections so we can free the stack when
* they've all gone away.
*/
if( ni->flags & CI_NETIF_FLAGS_DROP_SOCK_REFS ) {
trs = netif2tcp_helper_resource(ni);
dec_needed = 0;
ci_irqlock_lock(&trs->lock, &lock_flags);
if( trs->n_ep_closing_refs > 0 ) {
--trs->n_ep_closing_refs;
dec_needed = 1;
}
ci_irqlock_unlock(&trs->lock, &lock_flags);
if( dec_needed )
efab_tcp_helper_k_ref_count_dec(trs, 0);
}
}
void citp_waitable_wake_not_in_poll(ci_netif* ni, citp_waitable* sb,
unsigned what)
{
ci_assert(what);
ci_assert((what & ~(CI_SB_FLAG_WAKE_RX|CI_SB_FLAG_WAKE_TX)) == 0u);
ci_assert(!ni->state->in_poll);
ci_wmb();
if( what & CI_SB_FLAG_WAKE_RX )
++sb->sleep_seq.rw.rx;
if( what & CI_SB_FLAG_WAKE_TX )
++sb->sleep_seq.rw.tx;
ci_mb();
#ifdef __KERNEL__
/* Normally we put an object on a ready list in ci_netif_put_on_post_poll,
* but in this case we don't go via there, so have to explicitly queue on
* the ready list here.
*/
ci_ni_dllist_remove(ni, &sb->ready_link);
ci_ni_dllist_put(ni, &ni->state->ready_lists[sb->ready_list_id],
&sb->ready_link);
if( what & sb->wake_request ) {
sb->sb_flags |= what;
citp_waitable_wakeup(ni, sb);
}
/* Wake the ready list too, if that's requested it. */
if( ni->state->ready_list_flags[sb->ready_list_id] &
CI_NI_READY_LIST_FLAG_WAKE )
efab_tcp_helper_ready_list_wakeup(netif2tcp_helper_resource(ni),
sb->ready_list_id);
#else
if( what & sb->wake_request ) {
sb->sb_flags |= what;
ci_netif_put_on_post_poll(ni, sb);
ef_eplock_holder_set_flag(&ni->state->lock, CI_EPLOCK_NETIF_NEED_WAKE);
}
else {
/* Normally we put an object on a ready list in ci_netif_put_on_post_poll,
* but in this case we don't go via there, so have to explicitly queue on
* the ready list here.
*/
ci_ni_dllist_remove(ni, &sb->ready_link);
ci_ni_dllist_put(ni, &ni->state->ready_lists[sb->ready_list_id],
&sb->ready_link);
if( ni->state->ready_list_flags[sb->ready_list_id] &
CI_NI_READY_LIST_FLAG_WAKE )
ef_eplock_holder_set_flag(&ni->state->lock, CI_EPLOCK_NETIF_NEED_WAKE);
}
#endif
}
/*! Perform system bind on the OS backing socket.
* \param ep Endpoint context
* \param fd Callers FD
* \param ip_addr_be32 Local address to which to bind
* \param port_be16 [in] requested port [out] assigned port
* \return 0 - success & [port_be16] updated
* CI_SOCKET_HANDOVER, Pass to OS, OS bound ok, (no error)
* CI_SOCKET_ERROR & errno set
*/
ci_inline int __ci_bind(ci_netif *ni, ci_sock_cmn *s,
ci_uint32 ip_addr_be32, ci_uint16* port_be16 )
{
int rc;
ci_uint16 user_port; /* Port number specified by user, not by OS.
* See bug 4015 for details */
union ci_sockaddr_u sa_u;
ci_assert(s->domain == AF_INET || s->domain == AF_INET6);
ci_assert( port_be16 );
user_port = *port_be16;
#if CI_CFG_FAKE_IPV6
ci_assert(s->domain == AF_INET || s->domain == AF_INET6);
if( s->domain == AF_INET )
ci_make_sockaddr(&sa_u.sin, s->domain, user_port, ip_addr_be32);
else
ci_make_sockaddr6(&sa_u.sin6, s->domain, user_port, ip_addr_be32);
#else
ci_assert(s->domain == AF_INET);
ci_make_sockaddr(&sa_u.sin, s->domain, user_port, ip_addr_be32);
#endif
#ifdef __ci_driver__
rc = efab_tcp_helper_bind_os_sock(netif2tcp_helper_resource(ni),
SC_SP(s),
&sa_u.sa, sizeof(sa_u), port_be16);
#else
rc = ci_tcp_helper_bind_os_sock(ni, SC_SP(s), &sa_u.sa,
sizeof(sa_u), port_be16);
#endif
/* bug1781: only do this if the earlier bind succeeded.
* check if we can handle this socket */
if( rc != 0 )
return rc;
if( user_port != 0 )
s->s_flags |= CI_SOCK_FLAG_PORT_BOUND;
if( ip_addr_be32 != INADDR_ANY )
s->s_flags |= CI_SOCK_FLAG_ADDR_BOUND;
s->s_flags &= ~CI_SOCK_FLAG_CONNECT_MUST_BIND;
#ifndef __ci_driver__
/* We do not call bind() to alien address from in-kernel code */
if( ip_addr_be32 != INADDR_ANY &&
!cicp_user_addr_is_local_efab(CICP_HANDLE(ni), &ip_addr_be32) )
s->s_flags |= CI_SOCK_FLAG_BOUND_ALIEN;
#endif
return rc;
}
int
efab_eplock_unlock_and_wake(ci_netif *ni, int in_dl_context)
{
int l = ni->state->lock.lock;
tcp_helper_resource_t *rs = netif2tcp_helper_resource(ni);
/* Allocate more packets if necessary. */
if( !in_dl_context && OO_STACK_NEEDS_MORE_PACKETS(ni) )
efab_tcp_helper_more_bufs(rs);
/* We use in_dl_context from now on, and we should remove
* CI_NETIF_FLAG_IN_DL_CONTEXT under the stack lock. */
if( in_dl_context )
ni->flags &= ~CI_NETIF_FLAG_IN_DL_CONTEXT;
again:
#ifndef NDEBUG
if( (~l & CI_EPLOCK_LOCKED) || (l & CI_EPLOCK_UNLOCKED) ) {
OO_DEBUG_ERR(ci_log("efab_eplock_unlock_and_wake: corrupt"
" (value is %x)", (unsigned) l));
OO_DEBUG_ERR(dump_stack());
return -EIO;
}
#endif
if( l & CI_EPLOCK_CALLBACK_FLAGS ) {
/* Invoke the callback while we've still got the lock. The callback
** is responsible for either
** - dropping the lock using ef_eplock_try_unlock(), and returning
** the lock value prior to unlocking, OR
** - keeping the eplock locked and returning CI_EPLOCK_LOCKED
*/
l = efab_tcp_helper_netif_lock_callback(&ni->eplock_helper, l, in_dl_context);
}
else if( ci_cas32_fail(&ni->state->lock.lock, l, CI_EPLOCK_UNLOCKED) ) {
/* Someone (probably) set a flag when we tried to unlock, so we'd
** better handle the flag(s).
*/
l = ni->state->lock.lock;
goto again;
}
if( l & CI_EPLOCK_FL_NEED_WAKE ) {
CITP_STATS_NETIF_INC(ni, lock_wakes);
wake_up_interruptible(&ni->eplock_helper.wq);
}
return 0;
}
static void
oof_cb_sw_filter_postpone(struct oof_socket* skf, unsigned laddr, int lport,
unsigned raddr, int rport, int protocol, int op_op)
{
ci_netif* ni = skf_to_ni(skf);
struct tcp_helper_resource_s *trs = netif2tcp_helper_resource(ni);
struct oof_cb_sw_filter_op* op = CI_ALLOC_OBJ(struct oof_cb_sw_filter_op);
if( op == NULL ) {
/* Linux complains about failed allocations */
return;
}
op->sock_id = OO_SP_FROM_INT(ni, skf_to_ep(skf)->id);
op->laddr = laddr;
op->raddr = raddr;
op->lport = lport;
op->rport = rport;
op->protocol = protocol;
op->op = op_op;
op->next = NULL;
spin_lock_bh(&ni->swf_update_lock);
if( ni->swf_update_last == NULL )
ni->swf_update_first = op;
else
ni->swf_update_last->next = op;
ni->swf_update_last = op;
spin_unlock_bh(&ni->swf_update_lock);
/* We are holding a spinlock, so claim to be in driverlink context here */
if( efab_tcp_helper_netif_lock_or_set_flags(trs, OO_TRUSTED_LOCK_SWF_UPDATE,
CI_EPLOCK_NETIF_SWF_UPDATE, 1) ) {
ef_eplock_holder_set_flag(&ni->state->lock, CI_EPLOCK_NETIF_SWF_UPDATE);
efab_tcp_helper_netif_unlock(trs, 1);
}
}
/* Fixme: most callers of oof_cb_sw_filter_insert do not check rc. */
int
oof_cb_sw_filter_insert(struct oof_socket* skf, unsigned laddr, int lport,
unsigned raddr, int rport, int protocol,
int stack_locked)
{
ci_netif* ni = skf_to_ni(skf);
struct tcp_helper_resource_s *trs = netif2tcp_helper_resource(ni);
int rc = 0;
ci_assert(!stack_locked || ci_netif_is_locked(ni));
/* We are holding a spinlock, so claim to be in driverlink context here */
if( stack_locked || efab_tcp_helper_netif_try_lock(trs, 1) ) {
rc = ci_netif_filter_insert(ni, OO_SP_FROM_INT(ni, skf_to_ep(skf)->id),
laddr, lport, raddr, rport, protocol);
if( ! stack_locked )
efab_tcp_helper_netif_unlock(trs, 1);
}
else
oof_cb_sw_filter_postpone(skf, laddr, lport, raddr, rport, protocol,
OOF_CB_SW_FILTER_OP_ADD);
return rc;
}
/* Locking policy:
* Enterance: priv->thr->netif is assumed to be locked.
* Exit: all stacks (the client stack and the listener's stack) are
* unlocked.
*/
int efab_tcp_loopback_connect(ci_private_t *priv, void *arg)
{
struct oo_op_loopback_connect *carg = arg;
ci_netif *alien_ni = NULL;
oo_sp tls_id;
ci_assert(ci_netif_is_locked(&priv->thr->netif));
carg->out_moved = 0;
if( !CI_PRIV_TYPE_IS_ENDPOINT(priv->fd_type) )
return -EINVAL;
if( NI_OPTS(&priv->thr->netif).tcp_client_loopback !=
CITP_TCP_LOOPBACK_TO_CONNSTACK &&
NI_OPTS(&priv->thr->netif).tcp_client_loopback !=
CITP_TCP_LOOPBACK_TO_LISTSTACK &&
NI_OPTS(&priv->thr->netif).tcp_client_loopback !=
CITP_TCP_LOOPBACK_TO_NEWSTACK) {
ci_netif_unlock(&priv->thr->netif);
return -EINVAL;
}
while( iterate_netifs_unlocked(&alien_ni) == 0 ) {
if( !efab_thr_can_access_stack(netif2tcp_helper_resource(alien_ni),
EFAB_THR_TABLE_LOOKUP_CHECK_USER) )
continue; /* no permission to look in here */
if( NI_OPTS(alien_ni).tcp_server_loopback == CITP_TCP_LOOPBACK_OFF )
continue; /* server does not accept loopback connections */
if( NI_OPTS(&priv->thr->netif).tcp_client_loopback !=
CITP_TCP_LOOPBACK_TO_LISTSTACK &&
NI_OPTS(alien_ni).tcp_server_loopback !=
CITP_TCP_LOOPBACK_ALLOW_ALIEN_IN_ACCEPTQ )
continue; /* options of the stacks to not match */
if( NI_OPTS(&priv->thr->netif).tcp_client_loopback !=
CITP_TCP_LOOPBACK_TO_LISTSTACK &&
!efab_thr_user_can_access_stack(alien_ni->uid, alien_ni->euid,
&priv->thr->netif) )
continue; /* server can't accept our socket */
tls_id = ci_tcp_connect_find_local_peer(alien_ni, carg->dst_addr,
carg->dst_port);
if( OO_SP_NOT_NULL(tls_id) ) {
int rc;
/* We are going to exit in this or other way: get ref and
* drop kref of alien_ni */
efab_thr_ref(netif2tcp_helper_resource(alien_ni));
iterate_netifs_unlocked_dropref(alien_ni);
switch( NI_OPTS(&priv->thr->netif).tcp_client_loopback ) {
case CITP_TCP_LOOPBACK_TO_CONNSTACK:
/* connect_lo_toconn unlocks priv->thr->netif */
carg->out_rc =
ci_tcp_connect_lo_toconn(&priv->thr->netif, priv->sock_id,
carg->dst_addr, alien_ni, tls_id);
efab_thr_release(netif2tcp_helper_resource(alien_ni));
return 0;
case CITP_TCP_LOOPBACK_TO_LISTSTACK:
/* Nobody should be using this socket, so trylock should succeed.
* Overwise we hand over the socket and do not accelerate this
* loopback connection. */
rc = ci_sock_trylock(&priv->thr->netif,
SP_TO_WAITABLE(&priv->thr->netif,
priv->sock_id));
if( rc == 0 ) {
ci_netif_unlock(&priv->thr->netif);
efab_thr_release(netif2tcp_helper_resource(alien_ni));
return -ECONNREFUSED;
}
/* move_to_alien changes locks - see comments near it */
rc = efab_file_move_to_alien_stack(priv, alien_ni);
if( rc != 0 ) {
/* error - everything is already unlocked */
efab_thr_release(netif2tcp_helper_resource(alien_ni));
/* if we return error, UL will hand the socket over. */
return rc;
}
/* now alien_ni is locked */
/* Connect again, using new endpoint */
carg->out_rc =
ci_tcp_connect_lo_samestack(
alien_ni,
SP_TO_TCP(alien_ni,
SP_TO_WAITABLE(&priv->thr->netif,
priv->sock_id)
->moved_to_sock_id),
tls_id);
ci_netif_unlock(alien_ni);
carg->out_moved = 1;
return 0;
case CITP_TCP_LOOPBACK_TO_NEWSTACK:
{
tcp_helper_resource_t *new_thr;
ci_resource_onload_alloc_t alloc;
/* create new stack
* todo: no hardware interfaces are necessary */
strcpy(alloc.in_version, ONLOAD_VERSION);
strcpy(alloc.in_uk_intf_ver, oo_uk_intf_ver);
alloc.in_name[0] = '\0';
alloc.in_flags = 0;
rc = tcp_helper_alloc_kernel(&alloc, &NI_OPTS(&priv->thr->netif), 0,
&new_thr);
if( rc != 0 ) {
ci_netif_unlock(&priv->thr->netif);
efab_thr_release(netif2tcp_helper_resource(alien_ni));
return -ECONNREFUSED;
}
rc = ci_sock_trylock(&priv->thr->netif,
SP_TO_WAITABLE(&priv->thr->netif,
priv->sock_id));
if( rc == 0 ) {
ci_netif_unlock(&priv->thr->netif);
efab_thr_release(netif2tcp_helper_resource(alien_ni));
efab_thr_release(new_thr);
return -ECONNREFUSED;
}
/* move connecting socket to the new stack */
rc = efab_file_move_to_alien_stack(priv, &new_thr->netif);
if( rc != 0 ) {
/* error - everything is already unlocked */
efab_thr_release(netif2tcp_helper_resource(alien_ni));
efab_thr_release(new_thr);
return -ECONNREFUSED;
}
/* now new_thr->netif is locked */
carg->out_moved = 1;
carg->out_rc = -ECONNREFUSED;
/* now connect via CITP_TCP_LOOPBACK_TO_CONNSTACK */
/* connect_lo_toconn unlocks new_thr->netif */
carg->out_rc =
ci_tcp_connect_lo_toconn(
&new_thr->netif,
SP_TO_WAITABLE(&priv->thr->netif,
priv->sock_id)->moved_to_sock_id,
carg->dst_addr, alien_ni, tls_id);
efab_thr_release(netif2tcp_helper_resource(alien_ni));
return 0;
}
}
}
else if( tls_id == OO_SP_INVALID )
break;
}
ci_netif_unlock(&priv->thr->netif);
return -ENOENT;
}
/* Move priv file to the alien_ni stack.
* Should be called with the locked priv stack and socket;
* the function returns with this stack being unlocked.
* If rc=0, it returns with alien_ni stack locked;
* otherwise, both stacks are unlocked.
* Socket is always unlocked on return. */
int efab_file_move_to_alien_stack(ci_private_t *priv, ci_netif *alien_ni)
{
tcp_helper_resource_t *old_thr = priv->thr;
tcp_helper_resource_t *new_thr = netif2tcp_helper_resource(alien_ni);
ci_sock_cmn *old_s = SP_TO_SOCK(&old_thr->netif, priv->sock_id);
ci_sock_cmn *new_s;
ci_sock_cmn *mid_s;
tcp_helper_endpoint_t *old_ep, *new_ep;
int rc, i;
int pollwait_register = 0;
#if CI_CFG_FD_CACHING
oo_p sp;
#endif
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d", __func__,
old_thr->id, priv->sock_id, new_thr->id));
/* Poll the old stack - deliver all data to our socket */
ci_netif_poll(&old_thr->netif);
/* Endpoints in epoll list should not be moved, because waitq is already
* in the epoll internal structures (bug 41152). */
if( !list_empty(&priv->_filp->f_ep_links) ) {
rc = -EBUSY;
goto fail1;
}
if( !efab_file_move_supported(&old_thr->netif, old_s) ) {
rc = -EINVAL;
goto fail1;
}
/* Lock the second stack */
i = 0;
while( ! ci_netif_trylock(alien_ni) ) {
ci_netif_unlock(&old_thr->netif);
if( i++ >= 1000 ) {
rc = -EBUSY;
goto fail1_ni_unlocked;
}
rc = ci_netif_lock(&old_thr->netif);
if( rc != 0 )
goto fail1_ni_unlocked;
}
/* Allocate a new socket in the alien_ni stack */
rc = -ENOMEM;
if( old_s->b.state == CI_TCP_STATE_UDP ) {
ci_udp_state *new_us = ci_udp_get_state_buf(alien_ni);
if( new_us == NULL )
goto fail2;
new_s = &new_us->s;
}
else {
ci_tcp_state *new_ts = ci_tcp_get_state_buf(alien_ni);
if( new_ts == NULL )
goto fail2;
new_s = &new_ts->s;
}
/* Allocate an intermediate "socket" outside of everything */
mid_s = ci_alloc(CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
if( mid_s == NULL )
goto fail3;
OO_DEBUG_TCPH(ci_log("%s: move %d:%d to %d:%d", __func__,
old_thr->id, priv->sock_id,
new_thr->id, new_s->b.bufid));
/* Copy TCP/UDP state */
memcpy(mid_s, old_s, CI_MAX(sizeof(ci_tcp_state), sizeof(ci_udp_state)));
/* do not copy old_s->b.bufid
* and other fields in stack adress space */
mid_s->b.sb_aflags |= CI_SB_AFLAG_ORPHAN;
mid_s->b.bufid = new_s->b.bufid;
mid_s->b.post_poll_link = new_s->b.post_poll_link;
mid_s->b.ready_link = new_s->b.ready_link;
mid_s->reap_link = new_s->reap_link;
if( old_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *mid_ts = SOCK_TO_TCP(mid_s);
mid_ts->timeout_q_link = new_ts->timeout_q_link;
mid_ts->tx_ready_link = new_ts->tx_ready_link;
mid_ts->rto_tid = new_ts->rto_tid;
mid_ts->delack_tid = new_ts->delack_tid;
mid_ts->zwin_tid = new_ts->zwin_tid;
mid_ts->kalive_tid = new_ts->kalive_tid;
mid_ts->cork_tid = new_ts->cork_tid;
ci_ip_queue_init(&mid_ts->recv1);
ci_ip_queue_init(&mid_ts->recv2);
ci_ip_queue_init(&mid_ts->send);
ci_ip_queue_init(&mid_ts->retrans);
mid_ts->send_prequeue = OO_PP_ID_NULL;
new_ts->retrans_ptr = OO_PP_NULL;
mid_ts->tmpl_head = OO_PP_NULL;
oo_atomic_set(&mid_ts->send_prequeue_in, 0);
*new_ts = *mid_ts;
ci_pmtu_state_init(alien_ni, &new_ts->s, &new_ts->pmtus,
CI_IP_TIMER_PMTU_DISCOVER);
#if CI_CFG_FD_CACHING
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_link, sp, "epch");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_link);
sp = TS_OFF(alien_ni, new_ts);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_state, epcache_fd_link));
ci_ni_dllist_link_init(alien_ni, &new_ts->epcache_fd_link, sp, "ecfd");
ci_ni_dllist_self_link(alien_ni, &new_ts->epcache_fd_link);
#endif
/* free temporary mid_ts storage */
CI_FREE_OBJ(mid_ts);
}
else {
ci_udp_state *mid_us = SOCK_TO_UDP(mid_s);
*SOCK_TO_UDP(new_s) = *mid_us;
CI_FREE_OBJ(mid_us);
}
/* Move the filter */
old_ep = ci_trs_ep_get(old_thr, priv->sock_id);
new_ep = ci_trs_ep_get(new_thr, new_s->b.bufid);
rc = tcp_helper_endpoint_move_filters_pre(old_ep, new_ep);
if( rc != 0 ) {
rc = -EINVAL;
goto fail3;
}
/* Allocate a new file for the new endpoint */
rc = onload_alloc_file(new_thr, new_s->b.bufid, priv->_filp->f_flags,
priv->fd_type, &old_ep->alien_ref);
if( rc != 0 )
goto fail4;
ci_assert(old_ep->alien_ref);
/* Copy F_SETOWN_EX, F_SETSIG to the new file */
#ifdef F_SETOWN_EX
rcu_read_lock();
__f_setown(old_ep->alien_ref->_filp, priv->_filp->f_owner.pid,
priv->_filp->f_owner.pid_type, 1);
rcu_read_unlock();
#endif
old_ep->alien_ref->_filp->f_owner.signum = priv->_filp->f_owner.signum;
old_ep->alien_ref->_filp->f_flags |= priv->_filp->f_flags & O_NONBLOCK;
/* Move os_socket from one ep to another */
if( tcp_helper_endpoint_set_aflags(new_ep, OO_THR_EP_AFLAG_ATTACHED) &
OO_THR_EP_AFLAG_ATTACHED ) {
fput(old_ep->alien_ref->_filp);
rc = -EBUSY;
goto fail2; /* state & filters are cleared by fput() */
}
/********* Point of no return **********/
ci_wmb();
priv->fd_type = CI_PRIV_TYPE_ALIEN_EP;
priv->_filp->f_op = &linux_tcp_helper_fops_alien;
ci_wmb();
oo_file_moved(priv);
/* Read all already-arrived packets after the filters move but before
* copying of the receive queue. */
ci_netif_poll(&old_thr->netif);
tcp_helper_endpoint_move_filters_post(old_ep, new_ep);
ci_assert( efab_file_move_supported(&old_thr->netif, old_s));
/* There's a gap between un-registering the old ep, and registering the
* the new. However, the notifications shouldn't be in use for sockets
* that are in a state that can be moved, so this shouldn't be a problem.
*/
if( old_ep->os_sock_pt.whead ) {
pollwait_register = 1;
efab_tcp_helper_os_pollwait_unregister(old_ep);
}
ci_assert_equal(new_ep->os_socket, NULL);
new_ep->os_socket = oo_file_ref_xchg(&old_ep->os_socket, NULL);
ci_assert_equal(old_ep->os_socket, NULL);
if( pollwait_register )
efab_tcp_helper_os_pollwait_register(new_ep);
ci_bit_clear(&new_s->b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
if( new_s->b.state == CI_TCP_ESTABLISHED )
CI_TCP_STATS_INC_CURR_ESTAB(alien_ni);
/* Copy recv queue */
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
ci_tcp_state *old_ts = SOCK_TO_TCP(old_s);
int i;
/* Stop timers */
ci_ip_timer_clear(&old_thr->netif, &old_ts->kalive_tid);
ci_ip_timer_clear(&old_thr->netif, &old_ts->delack_tid);
efab_ip_queue_copy(alien_ni, &new_ts->recv1,
&old_thr->netif, &old_ts->recv1);
efab_ip_queue_copy(alien_ni, &new_ts->recv2,
&old_thr->netif, &old_ts->recv2);
new_ts->recv1_extract = new_ts->recv1.head;
/* Drop reorder buffer */
ci_ip_queue_init(&new_ts->rob);
new_ts->dsack_block = OO_PP_INVALID;
new_ts->dsack_start = new_ts->dsack_end = 0;
for( i = 0; i <= CI_TCP_SACK_MAX_BLOCKS; i++ )
new_ts->last_sack[i] = OO_PP_NULL;
}
else {
/* There should not be any recv q, but drop it to be sure */
ci_udp_recv_q_init(&SOCK_TO_UDP(new_s)->recv_q);
}
/* Old stack can be unlocked */
old_s->b.sb_flags |= CI_SB_FLAG_MOVED;
ci_netif_unlock(&old_thr->netif);
ci_assert( efab_file_move_supported(alien_ni, new_s) );
/* Move done: poll for any new data. */
ci_netif_poll(alien_ni);
if( new_s->b.state & CI_TCP_STATE_TCP ) {
ci_tcp_state *new_ts = SOCK_TO_TCP(new_s);
/* Timers setup: delack, keepalive */
if( (new_ts->acks_pending & CI_TCP_ACKS_PENDING_MASK) > 0)
ci_tcp_timeout_delack(alien_ni, new_ts);
ci_tcp_kalive_reset(alien_ni, new_ts);
}
/* Old ep: we are done. */
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_BIT);
old_s->b.moved_to_stack_id = alien_ni->state->stack_id;
old_s->b.moved_to_sock_id = new_s->b.bufid;
if( ! list_empty(&priv->_filp->f_ep_links) )
ci_bit_set(&old_s->b.sb_aflags, CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL_BIT);
ci_sock_unlock(&old_thr->netif, &old_s->b);
ci_sock_unlock(alien_ni, &new_s->b);
ci_assert(ci_netif_is_locked(alien_ni));
OO_DEBUG_TCPH(ci_log("%s: -> [%d:%d] %s", __func__,
new_thr->id, new_s->b.bufid,
ci_tcp_state_str(new_s->b.state)));
return 0;
fail4:
/* We clear the filters from the new ep.
* For now, we do not need to re-insert old filters because hw filters
* are alredy here (in case of accepted socket) or not needed.
* We have not removed old sw filters yet. */
tcp_helper_endpoint_move_filters_undo(old_ep, new_ep);
fail3:
if( new_s->b.state & CI_TCP_STATE_TCP )
ci_tcp_state_free(alien_ni, SOCK_TO_TCP(new_s));
else
ci_udp_state_free(alien_ni, SOCK_TO_UDP(new_s));
fail2:
ci_netif_unlock(alien_ni);
fail1:
ci_netif_unlock(&old_thr->netif);
fail1_ni_unlocked:
ci_sock_unlock(&old_thr->netif, &old_s->b);
OO_DEBUG_TCPH(ci_log("%s: rc=%d", __func__, rc));
return rc;
}
/* c_ni is assumed to be locked on enterance and is always unlocked on
* exit. */
int ci_tcp_connect_lo_toconn(ci_netif *c_ni, oo_sp c_id, ci_uint32 dst,
ci_netif *l_ni, oo_sp l_id)
{
ci_tcp_state *ts;
ci_tcp_socket_listen *tls, *alien_tls;
citp_waitable_obj *wo;
citp_waitable *w;
int rc;
ci_assert(ci_netif_is_locked(c_ni));
ci_assert(OO_SP_NOT_NULL(c_id));
ci_assert(OO_SP_NOT_NULL(l_id));
LOG_TC(log("%s: connect %d:%d to %d:%d", __FUNCTION__,
c_ni->state->stack_id, OO_SP_TO_INT(c_id),
l_ni->state->stack_id, OO_SP_TO_INT(l_id)));
alien_tls = SP_TO_TCP_LISTEN(l_ni, l_id);
if( (int)ci_tcp_acceptq_n(alien_tls) >= alien_tls->acceptq_max ) {
ci_netif_unlock(c_ni);
return -EBUSY;
}
/* In c_ni, create shadow listening socket tls (copy l_id) */
ts = ci_tcp_get_state_buf(c_ni);
if( ts == NULL ) {
ci_netif_unlock(c_ni);
LOG_E(ci_log("%s: [%d] out of socket buffers", __FUNCTION__, NI_ID(c_ni)));
return -ENOMEM;
}
/* init common tcp fields */
ts->s.so = alien_tls->s.so;
ts->s.cp.ip_ttl = alien_tls->s.cp.ip_ttl;
S_TCP_HDR(&ts->s)->tcp_source_be16 =
S_TCP_HDR(&alien_tls->s)->tcp_source_be16;
ts->s.domain = alien_tls->s.domain;
ts->c = alien_tls->c;
ts->c.tcp_defer_accept = OO_TCP_DEFER_ACCEPT_OFF;
/* make sure nobody will ever connect to our "shadow" socket
* except us */
ci_bit_set(&ts->s.b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
ci_tcp_set_slow_state(c_ni, ts, CI_TCP_LISTEN);
tls = SOCK_TO_TCP_LISTEN(&ts->s);
/* no timer: */
tls->s.s_flags = alien_tls->s.s_flags | CI_SOCK_FLAG_BOUND_ALIEN;
tls->acceptq_max = 1;
rc = ci_tcp_listen_init(c_ni, tls);
if( rc != 0 ) {
citp_waitable_obj_free(c_ni, &tls->s.b);
return rc;
}
/* Connect c_id to tls */
ts = SP_TO_TCP(c_ni, c_id);
rc = ci_tcp_connect_lo_samestack(c_ni, ts, tls->s.b.bufid);
/* Accept as from tls */
if( !ci_tcp_acceptq_not_empty(tls) ) {
/* it is possible, for example, if ci_tcp_listenq_try_promote() failed
* because there are no endpoints */
ci_tcp_listenq_drop_all(c_ni, tls);
citp_waitable_obj_free(c_ni, &tls->s.b);
ci_netif_unlock(c_ni);
return -EBUSY;
}
w = ci_tcp_acceptq_get(c_ni, tls);
ci_assert(w);
LOG_TV(ci_log("%s: %d:%d to %d:%d shadow %d:%d accepted %d:%d",
__FUNCTION__,
c_ni->state->stack_id, OO_SP_TO_INT(c_id),
l_ni->state->stack_id, OO_SP_TO_INT(l_id),
c_ni->state->stack_id, tls->s.b.bufid,
c_ni->state->stack_id, w->bufid));
ci_assert(w->state & CI_TCP_STATE_TCP);
ci_assert(w->state != CI_TCP_LISTEN);
/* Destroy tls.
* NB: nobody could possibly connect to it, so no need to do proper
* shutdown.
*/
ci_assert_equal(ci_tcp_acceptq_n(tls), 0);
ci_tcp_listenq_drop_all(c_ni, tls);
citp_waitable_obj_free(c_ni, &tls->s.b);
ci_netif_unlock(c_ni);
/* Keep a port reference */
{
tcp_helper_endpoint_t *l_ep, *a_ep;
struct oo_file_ref* os_sock_ref;
ci_irqlock_state_t lock_flags;
l_ep = ci_trs_ep_get(netif2tcp_helper_resource(l_ni), l_id);
a_ep = ci_trs_ep_get(netif2tcp_helper_resource(c_ni), W_SP(w));
ci_irqlock_lock(&l_ep->thr->lock, &lock_flags);
os_sock_ref = l_ep->os_socket;
ci_assert_equal(a_ep->os_port_keeper, NULL);
if( os_sock_ref != NULL ) {
os_sock_ref = oo_file_ref_add(os_sock_ref);
os_sock_ref = oo_file_ref_xchg(&a_ep->os_port_keeper, os_sock_ref);
ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
if( os_sock_ref != NULL )
oo_file_ref_drop(os_sock_ref);
}
else {
ci_irqlock_unlock(&l_ep->thr->lock, &lock_flags);
goto cleanup;
}
}
/* lock l_ni: Check that l_id is the same socket it used to be */
/* create ref-sock in l_ni, put it into acc q */
if( ci_netif_lock(l_ni) != 0 )
goto cleanup;
if( alien_tls->s.b.state != CI_TCP_LISTEN ||
(alien_tls->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN) ||
S_TCP_HDR(&alien_tls->s)->tcp_source_be16 != TS_TCP(ts)->tcp_dest_be16 ||
(alien_tls->s.pkt.ip.ip_saddr_be32 != INADDR_ANY &&
alien_tls->s.pkt.ip.ip_saddr_be32 != ts->s.pkt.ip.ip_daddr_be32) ) {
ci_netif_unlock(l_ni);
goto cleanup;
}
ci_bit_mask_set(&w->sb_aflags,
CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);
wo = citp_waitable_obj_alloc(l_ni);
if( wo == NULL ) {
ci_netif_unlock(l_ni);
goto cleanup;
}
wo->waitable.state = CI_TCP_CLOSED;
wo->waitable.sb_aflags |= CI_SB_AFLAG_MOVED_AWAY;
wo->waitable.moved_to_stack_id = c_ni->state->stack_id;
wo->waitable.moved_to_sock_id = W_SP(w);
LOG_TC(log("%s: put to acceptq %d:%d referencing %d:%d", __func__,
l_ni->state->stack_id, OO_SP_TO_INT(W_SP(&wo->waitable)),
c_ni->state->stack_id, OO_SP_TO_INT(W_SP(w))));
ci_tcp_acceptq_put(l_ni, alien_tls, &wo->waitable);
citp_waitable_wake_not_in_poll(l_ni, &alien_tls->s.b, CI_SB_FLAG_WAKE_RX);
ci_netif_unlock(l_ni);
return rc;
cleanup:
ci_assert(w->sb_aflags & CI_SB_AFLAG_ORPHAN);
ci_bit_mask_clear(&w->sb_aflags,
CI_SB_AFLAG_TCP_IN_ACCEPTQ | CI_SB_AFLAG_ORPHAN);
efab_tcp_helper_close_endpoint(netif2tcp_helper_resource(c_ni), w->bufid);
/* we can not guarantee c_ni lock, so we can' call
* ci_tcp_drop(c_ni, ts). So, we return error; UL will handover
* and close ts endpoint. */
return -EBUSY;
}
static int ci_tcp_connect_ul_syn_sent(ci_netif *ni, ci_tcp_state *ts)
{
int rc = 0;
if( ts->s.b.state == CI_TCP_SYN_SENT ) {
ci_netif_poll(ni);
if( OO_SP_NOT_NULL(ts->local_peer) ) {
/* No reason to sleep. Obviously, listener have dropped our syn
* because of some reason. Go away! */
ci_tcp_drop(ni, ts, EBUSY);
RET_WITH_ERRNO(EBUSY);
}
CI_TCP_SLEEP_WHILE(ni, ts, CI_SB_FLAG_WAKE_RX,
ts->s.so.sndtimeo_msec,
ts->s.b.state == CI_TCP_SYN_SENT, &rc);
}
if( rc == -EAGAIN ) {
LOG_TC(log( LNT_FMT "timeout on sleep: %d",
LNT_PRI_ARGS(ni, ts), -rc));
if( ! (ts->tcpflags & CI_TCPT_FLAG_NONBLOCK_CONNECT) ) {
ts->tcpflags |= CI_TCPT_FLAG_NONBLOCK_CONNECT;
CI_SET_ERROR(rc, EINPROGRESS);
}
else
CI_SET_ERROR(rc, EALREADY);
return rc;
}
else if( rc == -EINTR ) {
LOG_TC(log(LNT_FMT "connect() was interrupted by a signal",
LNT_PRI_ARGS(ni, ts)));
ts->tcpflags |= CI_TCPT_FLAG_NONBLOCK_CONNECT;
CI_SET_ERROR(rc, EINTR);
return rc;
}
/*! \TODO propagate the correct error code: CONNREFUSED, NOROUTE, etc. */
if( ts->s.b.state == CI_TCP_CLOSED ) {
/* Bug 3558:
* Set OS socket state to allow/disallow next bind().
* It is Linux hack. */
#ifdef __ci_driver__
CI_TRY(efab_tcp_helper_set_tcp_close_os_sock(netif2tcp_helper_resource(ni),
S_SP(ts)));
#else
CI_TRY(ci_tcp_helper_set_tcp_close_os_sock(ni, S_SP(ts)));
#endif
/* We should re-bind socket on the next use if the port was determined by
* OS.
*/
if( ! (ts->s.s_flags & CI_SOCK_FLAG_PORT_BOUND) )
ts->s.s_flags |= CI_SOCK_FLAG_CONNECT_MUST_BIND;
/* - if SO_ERROR is set, handle it and return this value;
* - else if rx_errno is set, return it;
* - else (TCP_RX_ERRNO==0, socket is CI_SHUT_RD) return ECONNABORTED */
if( (rc = ci_tcp_connect_handle_so_error(&ts->s)) == 0)
rc = TCP_RX_ERRNO(ts) ? TCP_RX_ERRNO(ts) : ECONNABORTED;
CI_SET_ERROR(rc, rc);
if( ! (ts->s.s_flags & CI_SOCK_FLAG_ADDR_BOUND) ) {
ts->s.pkt.ip.ip_saddr_be32 = 0;
ts->s.cp.ip_laddr_be32 = 0;
}
return rc;
}
return 0;
}
int ci_tcp_listen(citp_socket* ep, ci_fd_t fd, int backlog)
{
/*
** ?? error handling on possible fails not handled robustly...
** ?? Need to check port number is valid TODO
*/
/*! \todo If not bound then we have to be listening on all interfaces.
* It's likely that we won't be coming through here as we have to
* listen on the OS socket too! */
ci_tcp_state* ts;
ci_tcp_socket_listen* tls;
ci_netif* netif = ep->netif;
ci_sock_cmn* s = ep->s;
unsigned ul_backlog = backlog;
int rc;
oo_p sp;
LOG_TC(log("%s "SK_FMT" listen backlog=%d", __FUNCTION__, SK_PRI_ARGS(ep),
backlog));
CHECK_TEP(ep);
if( NI_OPTS(netif).tcp_listen_handover )
return CI_SOCKET_HANDOVER;
if( !NI_OPTS(netif).tcp_server_loopback) {
/* We should handover if the socket is bound to alien address. */
if( s->s_flags & CI_SOCK_FLAG_BOUND_ALIEN )
return CI_SOCKET_HANDOVER;
}
if( ul_backlog < 0 )
ul_backlog = NI_OPTS(netif).max_ep_bufs;
else if( ul_backlog < NI_OPTS(netif).acceptq_min_backlog )
ul_backlog = NI_OPTS(netif).acceptq_min_backlog;
if( s->b.state == CI_TCP_LISTEN ) {
tls = SOCK_TO_TCP_LISTEN(s);
tls->acceptq_max = ul_backlog;
ci_tcp_helper_listen_os_sock(fd, ul_backlog);
return 0;
}
if( s->b.state != CI_TCP_CLOSED ) {
CI_SET_ERROR(rc, EINVAL);
return rc;
}
ts = SOCK_TO_TCP(s);
/* Bug 3376: if socket used for a previous, failed, connect then the error
* numbers will not be as expected. Only seen when not using listening
* netifs (as moving the EP to the new netif resets them).
*/
ts->s.tx_errno = EPIPE;
ts->s.rx_errno = ENOTCONN;
/* fill in address/ports and all TCP state */
if( !(ts->s.s_flags & CI_SOCK_FLAG_BOUND) ) {
ci_uint16 source_be16;
/* They haven't previously done a bind, so we need to choose
* a port. As we haven't been given a hint we let the OS choose. */
source_be16 = 0;
rc = __ci_bind(ep->netif, ep->s, ts->s.pkt.ip.ip_saddr_be32, &source_be16);
if (CI_LIKELY( rc==0 )) {
TS_TCP(ts)->tcp_source_be16 = source_be16;
ts->s.cp.lport_be16 = source_be16;
LOG_TC(log(LNT_FMT "listen: our bind returned %s:%u",
LNT_PRI_ARGS(ep->netif, ts),
ip_addr_str(ts->s.pkt.ip.ip_saddr_be32),
(unsigned) CI_BSWAP_BE16(TS_TCP(ts)->tcp_source_be16)));
} else {
LOG_U(ci_log("__ci_bind returned %d at %s:%d", CI_GET_ERROR(rc),
__FILE__, __LINE__));
return rc;
}
}
ci_sock_lock(netif, &ts->s.b);
ci_tcp_set_slow_state(netif, ts, CI_TCP_LISTEN);
tls = SOCK_TO_TCP_LISTEN(&ts->s);
tcp_raddr_be32(tls) = 0u;
tcp_rport_be16(tls) = 0u;
ci_assert_equal(tls->s.tx_errno, EPIPE);
ci_assert_equal(tls->s.rx_errno, ENOTCONN);
/* setup listen timer - do it before the first return statement,
* because __ci_tcp_listen_to_normal() will be called on error path. */
if( ~tls->s.s_flags & CI_SOCK_FLAG_BOUND_ALIEN ) {
sp = TS_OFF(netif, tls);
OO_P_ADD(sp, CI_MEMBER_OFFSET(ci_tcp_socket_listen, listenq_tid));
ci_ip_timer_init(netif, &tls->listenq_tid, sp, "lstq");
tls->listenq_tid.param1 = S_SP(tls);
tls->listenq_tid.fn = CI_IP_TIMER_TCP_LISTEN;
}
rc = ci_tcp_listen_init(netif, tls);
ci_sock_unlock(netif, &ts->s.b);
if( rc != 0 ) {
CI_SET_ERROR(rc, -rc);
goto listen_fail;
}
tls->acceptq_max = ul_backlog;
CITP_STATS_TCP_LISTEN(CI_ZERO(&tls->stats));
/* install all the filters needed for this connection
* - tcp_laddr_be32(ts) = 0 for IPADDR_ANY
*
* TODO: handle BINDTODEVICE by setting phys_port paramter to correct
* physical L5 port index
* TODO: handle REUSEADDR by setting last paramter to TRUE
*/
if( ~s->s_flags & CI_SOCK_FLAG_BOUND_ALIEN ) {
#ifdef ONLOAD_OFE
if( netif->ofe != NULL ) {
tls->s.ofe_code_start = ofe_socktbl_find(
netif->ofe, OFE_SOCKTYPE_TCP_LISTEN,
tcp_laddr_be32(tls), INADDR_ANY,
tcp_lport_be16(ts), 0);
tls->ofe_promote = ofe_socktbl_find(
netif->ofe, OFE_SOCKTYPE_TCP_PASSIVE,
tcp_laddr_be32(tls), INADDR_ANY,
tcp_lport_be16(ts), 0);
}
#endif
rc = ci_tcp_ep_set_filters(netif, S_SP(tls), tls->s.cp.so_bindtodevice,
OO_SP_NULL);
if( rc == -EFILTERSSOME ) {
if( CITP_OPTS.no_fail )
rc = 0;
else {
ci_tcp_ep_clear_filters(netif, S_SP(tls), 0);
rc = -ENOBUFS;
}
}
ci_assert_nequal(rc, -EFILTERSSOME);
VERB(ci_log("%s: set_filters returned %d", __FUNCTION__, rc));
if (rc < 0) {
CI_SET_ERROR(rc, -rc);
goto post_listen_fail;
}
}
/*
* Call of system listen() is required for listen any, local host
* communications server and multi-homed server (to accept connections
* to L5 assigned address(es), but incoming from other interfaces).
*/
#ifdef __ci_driver__
{
rc = efab_tcp_helper_listen_os_sock( netif2tcp_helper_resource(netif),
S_SP(tls), backlog);
}
#else
rc = ci_tcp_helper_listen_os_sock(fd, backlog);
#endif
if ( rc < 0 ) {
/* clear the filter we've just set */
ci_tcp_ep_clear_filters(netif, S_SP(tls), 0);
goto post_listen_fail;
}
return 0;
post_listen_fail:
ci_tcp_listenq_drop_all(netif, tls);
listen_fail:
/* revert TCP state to a non-listening socket format */
__ci_tcp_listen_to_normal(netif, tls);
/* Above function sets orphan flag but we are attached to an FD. */
ci_bit_clear(&tls->s.b.sb_aflags, CI_SB_AFLAG_ORPHAN_BIT);
#ifdef __ci_driver__
return rc;
#else
return CI_SOCKET_ERROR;
#endif
}
