void
oof_cb_sw_filter_apply(ci_netif* ni)
{
struct oof_cb_sw_filter_op* op;
ci_assert(ci_netif_is_locked(ni));
spin_lock_bh(&ni->swf_update_lock);
for( op = ni->swf_update_first; op != NULL; op = ni->swf_update_first) {
ni->swf_update_first = op->next;
if( op->next == NULL )
ni->swf_update_last = NULL;
spin_unlock_bh(&ni->swf_update_lock);
if( op->op == OOF_CB_SW_FILTER_OP_ADD ) {
ci_netif_filter_insert(ni, op->sock_id, op->laddr, op->lport,
op->raddr, op->rport, op->protocol);
}
else {
ci_netif_filter_remove(ni, op->sock_id, op->laddr, op->lport,
op->raddr, op->rport, op->protocol);
}
ci_free(op);
spin_lock_bh(&ni->swf_update_lock);
}
spin_unlock_bh(&ni->swf_update_lock);
}
/****************************************************************************
*
* close - cleanup filedescriptor and private state
*
****************************************************************************/
static int
ci_char_fop_close(struct inode *inode, struct file *filp)
{
ci_private_char_t *priv = (ci_private_char_t *) filp->private_data;
EFCH_TRACE("%s:", __FUNCTION__);
/* cleanup private state */
filp->private_data = 0;
ci_resource_table_dtor(&priv->rt);
ci_free(priv);
return 0;
}
/*! Tear down a private_t */
void
ci_resource_table_dtor( ci_resource_table_t *rt )
{
efch_resource_t *rs;
unsigned i;
ci_assert(rt);
#if CI_CFG_PRIVATE_T_DEBUG_LIST
ci_lock_lock(&priv_list_lock);
list_del(&(rt->priv_list));
ci_lock_unlock(&priv_list_lock);
#endif
for( i = 0; i < rt->resource_table_highwater; i++ ) {
rs = rt->resource_table[i];
ci_assert(rs != NULL);
efch_resource_free(rs);
CI_DEBUG(rt->resource_table[i] = NULL);
}
if( rt->resource_table != rt->resource_table_static )
ci_free(rt->resource_table);
CI_DEBUG_ZERO(rt);
}
static int
onload_alloc_file(tcp_helper_resource_t *thr, oo_sp ep_id, int flags,
int fd_type)
{
struct qstr name = { .name = "" };
#ifdef EFX_HAVE_STRUCT_PATH
struct path path;
#define my_dentry path.dentry
#else
struct dentry *dentry;
#define my_dentry dentry
#endif
struct file *file;
int fd;
struct inode *inode;
ci_private_t *priv;
struct file_operations *fops;
fops = oo_fops_by_type(fd_type);
if( fops == NULL )
return -EINVAL;
ci_assert_equal(fops->owner, THIS_MODULE);
inode = new_inode(onload_mnt->mnt_sb);
if( inode == NULL )
return -ENOMEM;
#ifdef EFX_FSTYPE_HAS_MOUNT
inode->i_ino = get_next_ino();
#endif
if( fd_type == CI_PRIV_TYPE_NETIF )
inode->i_mode = S_IRWXUGO;
if( fd_type == CI_PRIV_TYPE_TCP_EP || fd_type == CI_PRIV_TYPE_UDP_EP )
inode->i_mode =
#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,21)
/* in 2.6.18 this flag makes us "socket" and sendmsg crashes;
* see sock_from_file() */
S_IFSOCK |
#endif
S_IRWXUGO;
else
inode->i_mode = S_IFIFO | S_IRUSR | S_IWUSR;
inode->i_uid = current_fsuid();
inode->i_gid = current_fsgid();
priv = &container_of(inode, struct onload_inode, vfs_inode)->priv;
priv->thr = thr;
priv->sock_id = ep_id;
priv->fd_type = fd_type;
fd = get_unused_fd();
if( fd < 0 ) {
iput(inode);
return fd;
}
/*ci_log("[%d]%s(%d:%d) return %d priv=%p", current->pid, __func__,
thr->id, ep_id, fd, priv);*/
#ifdef EFX_FSTYPE_HAS_MOUNT
#if LINUX_VERSION_CODE <= KERNEL_VERSION(2,6,37)
path.dentry = d_alloc(onload_mnt->mnt_sb->s_root, &name);
if( path.dentry != NULL )
path.dentry->d_op = &onloadfs_dentry_operations;
#else
path.dentry = d_alloc_pseudo(onload_mnt->mnt_sb, &name);
#endif
#else /* EFX_FSTYPE_HAS_MOUNT */
#ifdef EFX_HAVE_D_DNAME
my_dentry = d_alloc(onload_mnt->mnt_sb->s_root, &name);
#else
{
char str[32];
name.len = onloadfs_name(&container_of(inode, struct onload_inode,
vfs_inode)->priv,
str, sizeof(str));
name.name = str;
name.hash = inode->i_ino;
my_dentry = d_alloc(onload_mnt->mnt_sb->s_root, &name);
}
#endif
#endif /* EFX_FSTYPE_HAS_MOUNT */
if( my_dentry == NULL ) {
put_unused_fd(fd);
iput(inode);
return -ENOMEM;
}
#if !defined(EFX_FSTYPE_HAS_MOUNT) || defined(EFX_OLD_MOUNT_PSEUDO)
my_dentry->d_op = &onloadfs_dentry_operations;
#if !defined(EFX_HAVE_STRUCT_PATH) && defined(EFX_HAVE_D_DNAME)
my_dentry->d_flags &= ~DCACHE_UNHASHED;
#endif
#endif
d_instantiate(my_dentry, inode);
#ifndef EFX_HAVE_D_DNAME
d_rehash(my_dentry);
#endif
inode->i_fop = fops;
#ifdef EFX_HAVE_STRUCT_PATH
path.mnt = mntget(onload_mnt);
file = alloc_file(&path, FMODE_READ | FMODE_WRITE, fops);
#else
file = alloc_file(onload_mnt, dentry, FMODE_READ | FMODE_WRITE, fops);
#endif
if( file == NULL) {
#ifdef EFX_HAVE_STRUCT_PATH
path_put(&path);
#else
dput(dentry);
iput(inode);
#endif
put_unused_fd(fd);
return -ENFILE;
}
priv->_filp = file;
file->f_flags = O_RDWR | (flags & O_NONBLOCK);
file->f_pos = 0;
file->private_data = priv;
if( flags & O_CLOEXEC ) {
struct files_struct *files = current->files;
struct fdtable *fdt;
spin_lock(&files->file_lock);
fdt = files_fdtable(files);
rcu_assign_pointer(fdt->fd[fd], file);
efx_set_close_on_exec(fd, fdt);
spin_unlock(&files->file_lock);
} else
fd_install(fd, file);
try_module_get(THIS_MODULE);
ci_assert_equal(file->f_op, fops);
return fd;
}
void onload_priv_free(ci_private_t *priv)
{
if( priv->_filp->f_vfsmnt != onload_mnt)
ci_free(priv);
/* inode will free the priv automatically */
}
int
oo_create_fd(tcp_helper_endpoint_t* ep, int flags, int fd_type)
{
int fd;
tcp_helper_resource_t *trs = ep->thr;
citp_waitable_obj *wo = SP_TO_WAITABLE_OBJ(&trs->netif, ep->id);
efab_thr_ref(trs);
fd = onload_alloc_file(trs, ep->id, flags, fd_type);
if( fd < 0 ) {
efab_thr_release(trs);
OO_DEBUG_ERR(ci_log("%s: onload_alloc_file failed (%d)", __FUNCTION__, fd));
return fd;
}
ci_atomic32_and(&wo-> waitable.sb_aflags,
~(CI_SB_AFLAG_ORPHAN | CI_SB_AFLAG_TCP_IN_ACCEPTQ));
return fd;
}
/*
** promote a synrecv structure to an established socket
**
** Assumes that the caller will handle a fail if we can't allocate a new
** tcp_state structure due to memory pressure or the like
*/
int ci_tcp_listenq_try_promote(ci_netif* netif, ci_tcp_socket_listen* tls,
ci_tcp_state_synrecv* tsr,
ci_ip_cached_hdrs* ipcache,
ci_tcp_state** ts_out)
{
int rc = 0;
ci_assert(netif);
ci_assert(tls);
ci_assert(tls->s.b.state == CI_TCP_LISTEN);
ci_assert(tsr);
if( (int) ci_tcp_acceptq_n(tls) < tls->acceptq_max ) {
ci_tcp_state* ts;
/* grab a tcp_state structure that will go onto the accept queue. We take
* from the cache of EPs if any are available
*/
ts = get_ts_from_cache (netif, tsr, tls);
if( !ts ) {
/* None on cache; try allocating a new ts */
ts = ci_tcp_get_state_buf(netif);
#if CI_CFG_FD_CACHING
if( ts == NULL ) {
/* We've reaped. Did this result in any being cached */
ts = get_ts_from_cache(netif, tsr, tls);
if (ts == NULL ) {
/* No -- try again to allocate. */
ts = ci_tcp_get_state_buf(netif);
}
else {
CITP_STATS_NETIF(++netif->state->stats.sockcache_hit_reap);
}
}
#endif
if( ts == NULL ) {
LOG_TV(ci_log("%s: [%d] out of socket buffers",
__FUNCTION__, NI_ID(netif)));
CITP_STATS_TCP_LISTEN(++tls->stats.n_acceptq_no_sock);
CI_SET_SO_ERROR(&tls->s, ENOMEM);
citp_waitable_wake(netif, &tls->s.b, CI_SB_FLAG_WAKE_RX);
return -ENOMEM;
}
ci_assert(ci_tcp_is_cached(ts) ||
(ts->s.b.sb_aflags & CI_SB_AFLAG_ORPHAN));
}
#ifdef ONLOAD_OFE
ts->s.ofe_code_start = tls->ofe_promote;
#endif
if( ! ci_tcp_is_cached(ts) ) {
/* Need to initialise address information for use when setting filters */
ci_tcp_set_addr_on_promote(netif, ts, tsr, tls);
/* "borrow" filter from listening socket. For loopback socket, we
* do not need filters, but we have to take a reference of the OS
* socket. */
rc = ci_tcp_ep_set_filters(netif, S_SP(ts), ts->s.cp.so_bindtodevice,
S_SP(tls));
if( rc < 0 ) {
LOG_U(ci_log("%s: Unable to set filters %d", __FUNCTION__, rc));
/* Either put this back on the list (at the head) or free it */
ci_tcp_state_free(netif, ts);
return rc;
}
}
#if CI_CFG_FD_CACHING
else {
/* Now set the s/w filter. We leave the hw filter in place for cached
* EPS. This will probably not have the correct raddr and rport, but as
* it's sharing the listening socket's filter that's not a problem. It
* will be updated if this is still around when the listener is closed.
*/
rc = ci_netif_filter_insert(netif, S_SP(ts), tsr->l_addr,
sock_lport_be16(&tls->s), tsr->r_addr,
tsr->r_port, tcp_protocol(ts));
if (rc < 0) {
/* Bung it back on the cache list */
LOG_EP(ci_log("Unable to create s/w filter!"));
ci_ni_dllist_push(netif, &tls->epcache.cache, &ts->epcache_link);
return rc;
}
/* Need to initialise address information. We do this after trying to
* insert the sw filter, so we can push the tcp state back onto the
* cache queue with as few changes as possible if we fail to add the
* sw filter.
*/
ci_tcp_set_addr_on_promote(netif, ts, tsr, tls);
LOG_EP(ci_log("Cached fd %d from cached to connected", ts->cached_on_fd));
ci_ni_dllist_push(netif, &tls->epcache_connected, &ts->epcache_link);
}
#endif
ci_assert(IS_VALID_SOCK_P(netif, S_SP(ts)));
ci_assert(ts->s.b.state == CI_TCP_CLOSED);
ts->s.domain = tls->s.domain;
cicp_ip_cache_update_from(netif, &ts->s.pkt, ipcache);
ci_pmtu_state_init(netif, &ts->s, &ts->pmtus,
CI_IP_TIMER_PMTU_DISCOVER);
ci_pmtu_set(netif, &ts->pmtus,
CI_MIN(ts->s.pkt.mtu,
tsr->tcpopts.smss + sizeof(ci_tcp_hdr)
+ sizeof(ci_ip4_hdr)));
/* If we've got SYN via local route, we can handle it */
ci_assert_equiv(ts->s.pkt.status == retrrc_localroute,
OO_SP_NOT_NULL(tsr->local_peer));
if( ts->s.pkt.status == retrrc_localroute )
ts->s.pkt.flags |= CI_IP_CACHE_IS_LOCALROUTE;
ts->amss = tsr->amss;
/* options and flags */
ts->tcpflags = 0;
ts->tcpflags |= tsr->tcpopts.flags;
ts->tcpflags |= CI_TCPT_FLAG_PASSIVE_OPENED;
ts->outgoing_hdrs_len = sizeof(ci_ip4_hdr) + sizeof(ci_tcp_hdr);
if( ts->tcpflags & CI_TCPT_FLAG_WSCL ) {
ts->snd_wscl = tsr->tcpopts.wscl_shft;
ts->rcv_wscl = tsr->rcv_wscl;
} else {
ts->snd_wscl = ts->rcv_wscl = 0u;
}
CI_IP_SOCK_STATS_VAL_TXWSCL( ts, ts->snd_wscl);
CI_IP_SOCK_STATS_VAL_RXWSCL( ts, ts->rcv_wscl);
/* Send and receive sequence numbers */
tcp_snd_una(ts) = tcp_snd_nxt(ts) = tcp_enq_nxt(ts) = tcp_snd_up(ts) =
tsr->snd_isn + 1;
ci_tcp_set_snd_max(ts, tsr->rcv_nxt, tcp_snd_una(ts), 0);
ci_tcp_rx_set_isn(ts, tsr->rcv_nxt);
tcp_rcv_up(ts) = SEQ_SUB(tcp_rcv_nxt(ts), 1);
if( ts->tcpflags & CI_TCPT_FLAG_TSO ) {
ts->incoming_tcp_hdr_len += 12;
ts->outgoing_hdrs_len += 12;
ts->tspaws = ci_tcp_time_now(netif);
ts->tsrecent = tsr->tspeer;
ts->tslastack = tsr->rcv_nxt;
}
else {
/* Must be after initialising snd_una. */
ci_tcp_clear_rtt_timing(ts);
ts->timed_ts = tsr->timest;
}
/* SACK has nothing to be done. */
/* ?? ECN */
ci_tcp_set_hdr_len(ts, (ts->outgoing_hdrs_len - sizeof(ci_ip4_hdr)));
ts->smss = tsr->tcpopts.smss;
ts->c.user_mss = tls->c.user_mss;
if (ts->c.user_mss && ts->c.user_mss < ts->smss)
ts->smss = ts->c.user_mss;
#if CI_CFG_LIMIT_SMSS
ts->smss = ci_tcp_limit_mss(ts->smss, netif, __FUNCTION__);
#endif
ci_assert(ts->smss>0);
ci_tcp_set_eff_mss(netif, ts);
ci_tcp_set_initialcwnd(netif, ts);
/* Copy socket options & related fields that should be inherited.
* Note: Windows does not inherit rcvbuf until the call to accept
* completes. The assumption here is that all options can be
* inherited at the same time (most won't have an effect until there
* is a socket available for use by the app.).
*/
ci_tcp_inherit_accept_options(netif, tls, ts, "SYN RECV (LISTENQ PROMOTE)");
/* NB. Must have already set peer (which we have). */
ci_tcp_set_established_state(netif, ts);
CITP_STATS_NETIF(++netif->state->stats.synrecv2established);
ci_assert(ts->ka_probes == 0);
ci_tcp_kalive_restart(netif, ts, ci_tcp_kalive_idle_get(ts));
ci_tcp_set_flags(ts, CI_TCP_FLAG_ACK);
/* Remove the synrecv structure from the listen queue, and free the
** buffer. */
if( tsr->tcpopts.flags & CI_TCPT_FLAG_SYNCOOKIE )
ci_free(tsr);
else {
ci_tcp_listenq_remove(netif, tls, tsr);
ci_tcp_synrecv_free(netif, tsr);
}
ci_bit_set(&ts->s.b.sb_aflags, CI_SB_AFLAG_TCP_IN_ACCEPTQ_BIT);
ci_tcp_acceptq_put(netif, tls, &ts->s.b);
LOG_TC(log(LNT_FMT "new ts=%d SYN-RECV->ESTABLISHED flags=0x%x",
LNT_PRI_ARGS(netif, tls), S_FMT(ts), ts->tcpflags);
log(LNTS_FMT RCV_WND_FMT " snd=%08x-%08x-%08x enq=%08x",
LNTS_PRI_ARGS(netif, ts), RCV_WND_ARGS(ts),
tcp_snd_una(ts),
tcp_snd_nxt(ts), ts->snd_max, tcp_enq_nxt(ts)));
citp_waitable_wake(netif, &tls->s.b, CI_SB_FLAG_WAKE_RX);
*ts_out = ts;
return 0;
}
