static int
efab_vi_rm_mmap_mem(struct efrm_vi *virs,
unsigned long *bytes, void *opaque,
int *map_num, unsigned long *offset)
{
int queue_type;
uint32_t len;
if( virs->q[EFHW_EVQ].capacity != 0 ) {
len = efhw_iopages_size(&virs->q[EFHW_EVQ].pages);
len = CI_MIN(len, *bytes);
ci_assert_gt(len, 0);
ci_mmap_iopages(&virs->q[EFHW_EVQ].pages, 0,
len, bytes, opaque, map_num, offset);
if(*bytes == 0)
return 0;
}
for( queue_type=EFRM_VI_RM_DMA_QUEUE_COUNT-1;
queue_type>=0;
queue_type-- ) {
if( virs->q[queue_type].capacity != 0 ) {
len = efhw_iopages_size(&virs->q[queue_type].pages);
len = CI_MIN(len, *bytes);
ci_assert_gt(len, 0);
ci_mmap_iopages(&virs->q[queue_type].pages, 0,
len, bytes, opaque, map_num, offset);
if(*bytes == 0)
return 0;
}
}
return 0;
}
static int
efab_vi_rm_mmap_ctpio(struct efrm_vi *virs, unsigned long *bytes, void *opaque,
int *map_num, unsigned long *offset)
{
int rc;
int len;
int instance;
struct efhw_nic *nic;
int bar_off;
/* The CTPIO region is 12K from the start of the VI's aperture. */
const int CTPIO_OFFSET = 12 * 1024;
instance = virs->rs.rs_instance;
if( ! (virs->flags & EFHW_VI_TX_CTPIO) ) {
EFRM_ERR("%s: CTPIO is not enabled on VI instance %d\n", __FUNCTION__,
instance);
return -EINVAL;
}
/* Map the CTPIO region, which is 12K from the start of the VI's aperture. */
len = CI_MIN(*bytes, CI_PAGE_SIZE);
*bytes -= len;
nic = efrm_client_get_nic(virs->rs.rs_client);
ci_assert_ge(nic->vi_stride, CTPIO_OFFSET + len);
bar_off = (ef10_tx_dma_page_base(nic->vi_stride, instance) + CTPIO_OFFSET) &
PAGE_MASK;
rc = ci_mmap_bar(nic, bar_off, len, opaque, map_num, offset, 1);
if( rc < 0 )
EFCH_ERR("%s: ERROR: ci_mmap_bar failed rc=%d", __FUNCTION__, rc);
return rc;
}
static void ci_udp_pkt_to_zc_msg(ci_netif* ni, ci_ip_pkt_fmt* pkt,
struct onload_zc_msg* zc_msg)
{
int i, bytes_left = pkt->pf.udp.pay_len;
ci_ip_pkt_fmt* frag;
ci_ip_pkt_fmt* handle_frag;
handle_frag = frag = pkt;
i = 0;
ci_assert_nequal(zc_msg->iov, NULL);
/* Ignore first frag if zero length and there is another frag, but
* still pass the zero-length buffer as the onload_zc_handle so it
* will get freed correctly
*/
if( oo_offbuf_left(&frag->buf) == 0 && OO_PP_NOT_NULL(frag->frag_next) )
frag = PKT_CHK_NNL(ni, frag->frag_next);
do {
zc_msg->iov[i].iov_len = CI_MIN(oo_offbuf_left(&frag->buf),
bytes_left);
zc_msg->iov[i].iov_base = oo_offbuf_ptr(&frag->buf);
zc_msg->iov[i].buf = (onload_zc_handle)handle_frag;
zc_msg->iov[i].iov_flags = 0;
bytes_left -= zc_msg->iov[i].iov_len;
++i;
if( OO_PP_IS_NULL(frag->frag_next) ||
(i == CI_UDP_ZC_IOVEC_MAX) ||
(bytes_left == 0) )
break;
frag = PKT_CHK_NNL(ni, frag->frag_next);
handle_frag = frag;
} while( 1 );
zc_msg->msghdr.msg_iovlen = i;
}
static int
efab_vi_rm_mmap_pio(struct efrm_vi *virs,
unsigned long *bytes, void *opaque,
int *map_num, unsigned long *offset)
{
int rc;
int len;
int instance;
struct efhw_nic *nic;
int bar_off;
nic = efrm_client_get_nic(virs->rs.rs_client);
if( nic->devtype.arch != EFHW_ARCH_EF10 ) {
EFRM_ERR("%s: Only ef10 supports PIO."
" Expected arch=%d but got %d\n", __FUNCTION__,
EFHW_ARCH_EF10, nic->devtype.arch);
return -EINVAL;
}
instance = virs->rs.rs_instance;
/* Map the control page. */
len = CI_MIN(*bytes, CI_PAGE_SIZE);
*bytes -= len;
bar_off = (ef10_tx_dma_page_base(nic->vi_stride, instance) + 4096) &
PAGE_MASK;
rc = ci_mmap_bar(nic, bar_off, len, opaque, map_num, offset, 1);
if( rc < 0 )
EFCH_ERR("%s: ERROR: ci_mmap_bar failed rc=%d", __FUNCTION__, rc);
return rc;
}
int citp_do_init(int max_init_level)
{
int rc = 0;
int level;
int saved_errno = errno;
if( citp.init_level < max_init_level ) {
/* If threads are launched very early in program startup, then there could be
* a race here as multiple threads attempt to initialise on first access.
* The guard must be recursive, since this function might be re-entered during
* initialisation.
*/
static pthread_mutex_t mutex = PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
pthread_mutex_lock(&mutex);
_citp_do_init_inprogress++;
for (level = citp.init_level;
level < CI_MIN(max_init_level, CITP_INIT_ALL);
level++) {
rc = cipt_init_funcs[level]();
if (rc < 0)
break;
citp.init_level = level + 1;
}
--_citp_do_init_inprogress;
pthread_mutex_unlock(&mutex);
}
Log_S(log("%s: reached level %d", __FUNCTION__, citp.init_level));
if( rc == 0 )
errno = saved_errno;
return rc;
}
static int
oo_copy_pkt_to_iovec_no_adv(ci_netif* ni, const ci_ip_pkt_fmt* pkt,
ci_iovec_ptr* piov, int bytes_to_copy)
{
/* Copy data from [pkt] to [piov], following [pkt->frag_next] as
* necessary. Does not modify [pkt]. May or may not advance [piov].
* The packet must contain at least [bytes_to_copy] of data in the
* [pkt->buf]. [piov] may contain an arbitrary amount of space.
*
* Returns number of bytes copied on success, or -EFAULT otherwise.
*/
int n, pkt_left, pkt_off = 0;
int bytes_copied = 0;
while( 1 ) {
pkt_left = oo_offbuf_left(&pkt->buf) - pkt_off;
n = CI_MIN(pkt_left, CI_IOVEC_LEN(&piov->io));
n = CI_MIN(n, bytes_to_copy);
if(CI_UNLIKELY( do_copy(CI_IOVEC_BASE(&piov->io),
oo_offbuf_ptr(&pkt->buf) + pkt_off, n) != 0 ))
return -EFAULT;
bytes_copied += n;
pkt_off += n;
if( n == bytes_to_copy )
return bytes_copied;
bytes_to_copy -= n;
if( n == pkt_left ) {
/* Caller guarantees that packet contains at least [bytes_to_copy]. */
ci_assert(OO_PP_NOT_NULL(pkt->frag_next));
ci_iovec_ptr_advance(piov, n);
pkt = PKT_CHK_NNL(ni, pkt->frag_next);
pkt_off = 0;
/* We're unlikely to hit end-of-pkt-buf and end-of-iovec at the same
* time, and if we do, just go round the loop again.
*/
continue;
}
ci_assert_equal(n, CI_IOVEC_LEN(&piov->io));
if( piov->iovlen == 0 )
return bytes_copied;
piov->io = *piov->iov++;
--piov->iovlen;
}
}
static int
efab_vi_rm_mmap_io(struct efrm_vi *virs,
unsigned long *bytes, void *opaque,
int *map_num, unsigned long *offset)
{
int rc;
int len;
int instance;
int base;
unsigned vi_stride;
struct efhw_nic *nic;
nic = efrm_client_get_nic(virs->rs.rs_client);
instance = virs->rs.rs_instance;
len = CI_MIN(*bytes, CI_PAGE_SIZE);
*bytes -=len;
/* Make sure we can get away with a single page here. */
switch (nic->devtype.arch) {
case EFHW_ARCH_FALCON:
ci_assert_lt(falcon_tx_dma_page_offset(instance), CI_PAGE_SIZE);
ci_assert_lt(falcon_rx_dma_page_offset(instance), CI_PAGE_SIZE);
ci_assert_equal(falcon_tx_dma_page_base(instance),
falcon_rx_dma_page_base(instance));
base = falcon_tx_dma_page_base(instance);
break;
case EFHW_ARCH_EF10:
vi_stride = nic->vi_stride;
ci_assert_lt(ef10_tx_dma_page_offset(vi_stride, instance), CI_PAGE_SIZE);
ci_assert_lt(ef10_rx_dma_page_offset(vi_stride, instance), CI_PAGE_SIZE);
ci_assert_equal(ef10_tx_dma_page_base(vi_stride, instance),
ef10_rx_dma_page_base(vi_stride, instance));
base = ef10_tx_dma_page_base(vi_stride, instance);
break;
default:
EFCH_ERR("%s: ERROR: unknown nic type (%d)", __FUNCTION__,
nic->devtype.arch);
base = 0; /* To quiet the compiler */
BUG();
}
rc = ci_mmap_bar(nic, base, len, opaque, map_num, offset, 0);
if (rc < 0 ) {
EFCH_ERR("%s: ERROR: ci_mmap_bar failed rc=%d", __FUNCTION__, rc);
return rc;
}
return 0;
}
static void citp_get_process_name(void)
{
citp.process_name = citp.process_path;
ci_sprintf(citp.process_path, "<unknown-proc>");
{
int n;
n = readlink("/proc/self/exe", citp.process_path,
sizeof(citp.process_path));
if (n < 0)
return;
n = CI_MIN(n + 1, sizeof(citp.process_path));
citp.process_path[n - 1] = '\0';
citp.process_name = citp.process_path + n - 2;
while (citp.process_name > citp.process_path &&
citp.process_name[-1] != '/')
--citp.process_name;
}
}
int citp_do_init(int max_init_level)
{
int rc = 0;
int level;
int saved_errno = errno;
_citp_do_init_inprogress++;
for (level = citp.init_level;
level < CI_MIN(max_init_level, CITP_INIT_ALL);
level++) {
rc = cipt_init_funcs[level]();
if (rc < 0)
break;
citp.init_level = level + 1;
}
--_citp_do_init_inprogress;
Log_S(log("%s: reached level %d", __FUNCTION__, citp.init_level));
if( rc == 0 )
errno = saved_errno;
return rc;
}
/* Set the IP TOS */
void ci_udp_set_tos( ci_udp_state* us, ci_uint32 tos )
{
ci_ip_hdr_init_fixed(UDP_IP_HDR(us), IPPROTO_UDP, UDP_IP_HDR(us)->ip_ttl,
CI_MIN(tos, CI_IP_MAX_TOS));
}
static void citp_opts_getenv(citp_opts_t* opts)
{
/* ?? TODO: would like to use opts_citp_def.h here */
const char* s;
unsigned v;
opts->log_via_ioctl = 3;
/* TODO: Old name. Keeping reading 'til 2011, then purge. */
GET_ENV_OPT_HEX("EF_Log_VIA_IOCTL", log_via_ioctl);
GET_ENV_OPT_INT("EF_LOG_VIA_IOCTL", log_via_ioctl);
if( (s = getenv("EF_LOG_FILE")) && opts->log_via_ioctl == 3) {
opts->log_via_ioctl = 0;
citp_log_to_file(s);
} else if( opts->log_via_ioctl == 3 ) {
/* citp_setup_logging_early() have already detected stderr as
* tty/non-tty, so just trust it. */
if( ci_log_fn == citp_log_fn_drv )
opts->log_via_ioctl = 1;
else
opts->log_via_ioctl = 0;
}
if( opts->log_via_ioctl ) {
ci_log_options &=~ CI_LOG_PID;
citp_setup_logging_change(citp_log_fn_drv);
} else {
if( getenv("EF_LOG_TIMESTAMPS") )
ci_log_options |= CI_LOG_TIME;
citp_setup_logging_change(citp_log_fn_ul);
}
if( getenv("EF_POLL_NONBLOCK_FAST_LOOPS") &&
! getenv("EF_POLL_NONBLOCK_FAST_USEC") )
log("ERROR: EF_POLL_NONBLOCK_FAST_LOOPS is deprecated, use"
" EF_POLL_NONBLOCK_FAST_USEC instead");
if( getenv("EF_POLL_FAST_LOOPS") && ! getenv("EF_POLL_FAST_USEC") )
log("ERROR: EF_POLL_FAST_LOOPS is deprecated, use"
" EF_POLL_FAST_USEC instead");
if( (s = getenv("EF_POLL_USEC")) && atoi(s) ) {
GET_ENV_OPT_INT("EF_POLL_USEC", ul_spin_usec);
opts->ul_select_spin = 1;
opts->ul_poll_spin = 1;
#if CI_CFG_USERSPACE_EPOLL
opts->ul_epoll_spin = 1;
#endif
#if CI_CFG_UDP
opts->udp_recv_spin = 1;
opts->udp_send_spin = 1;
#endif
opts->tcp_recv_spin = 1;
opts->tcp_send_spin = 1;
opts->pkt_wait_spin = 1;
opts->sock_lock_buzz = 1;
opts->stack_lock_buzz = 1;
}
if( (s = getenv("EF_BUZZ_USEC")) && atoi(s) ) {
opts->sock_lock_buzz = 1;
opts->stack_lock_buzz = 1;
}
GET_ENV_OPT_HEX("EF_UNIX_LOG", log_level);
GET_ENV_OPT_INT("EF_PROBE", probe);
GET_ENV_OPT_INT("EF_TCP", ul_tcp);
GET_ENV_OPT_INT("EF_UDP", ul_udp);
GET_ENV_OPT_INT("EF_UL_SELECT", ul_select);
GET_ENV_OPT_INT("EF_SELECT_SPIN", ul_select_spin);
GET_ENV_OPT_INT("EF_SELECT_FAST", ul_select_fast);
GET_ENV_OPT_INT("EF_UL_POLL", ul_poll);
GET_ENV_OPT_INT("EF_POLL_SPIN", ul_poll_spin);
GET_ENV_OPT_INT("EF_POLL_FAST", ul_poll_fast);
GET_ENV_OPT_INT("EF_POLL_FAST_USEC", ul_poll_fast_usec);
GET_ENV_OPT_INT("EF_POLL_NONBLOCK_FAST_USEC", ul_poll_nonblock_fast_usec);
GET_ENV_OPT_INT("EF_SELECT_FAST_USEC", ul_select_fast_usec);
GET_ENV_OPT_INT("EF_SELECT_NONBLOCK_FAST_USEC", ul_select_nonblock_fast_usec);
#if CI_CFG_UDP
GET_ENV_OPT_INT("EF_UDP_RECV_SPIN", udp_recv_spin);
GET_ENV_OPT_INT("EF_UDP_SEND_SPIN", udp_send_spin);
#endif
GET_ENV_OPT_INT("EF_TCP_RECV_SPIN", tcp_recv_spin);
GET_ENV_OPT_INT("EF_TCP_SEND_SPIN", tcp_send_spin);
GET_ENV_OPT_INT("EF_TCP_ACCEPT_SPIN", tcp_accept_spin);
GET_ENV_OPT_INT("EF_TCP_CONNECT_SPIN",tcp_connect_spin);
GET_ENV_OPT_INT("EF_PKT_WAIT_SPIN", pkt_wait_spin);
#if CI_CFG_USERSPACE_PIPE
GET_ENV_OPT_INT("EF_PIPE_RECV_SPIN", pipe_recv_spin);
GET_ENV_OPT_INT("EF_PIPE_SEND_SPIN", pipe_send_spin);
GET_ENV_OPT_INT("EF_PIPE_SIZE", pipe_size);
#endif
GET_ENV_OPT_INT("EF_SOCK_LOCK_BUZZ", sock_lock_buzz);
GET_ENV_OPT_INT("EF_STACK_LOCK_BUZZ", stack_lock_buzz);
GET_ENV_OPT_INT("EF_SO_BUSY_POLL_SPIN", so_busy_poll_spin);
#if CI_CFG_USERSPACE_EPOLL
GET_ENV_OPT_INT("EF_UL_EPOLL", ul_epoll);
if( opts->ul_epoll == 0 && ci_cfg_opts.netif_opts.int_driven == 0 ) {
ci_log("EF_INT_DRIVEN=0 and EF_UL_EPOLL=0 are not compatible. "
"EF_INT_DRIVEN can be set to 0 implicitly, because of non-zero "
"EF_POLL_USEC. If you need both spinning and EF_UL_EPOLL=0, "
"please set EF_INT_DRIVEN=1 explicitly.");
}
GET_ENV_OPT_INT("EF_EPOLL_SPIN", ul_epoll_spin);
GET_ENV_OPT_INT("EF_EPOLL_CTL_FAST", ul_epoll_ctl_fast);
GET_ENV_OPT_INT("EF_EPOLL_CTL_HANDOFF",ul_epoll_ctl_handoff);
GET_ENV_OPT_INT("EF_EPOLL_MT_SAFE", ul_epoll_mt_safe);
#endif
GET_ENV_OPT_INT("EF_FDTABLE_SIZE", fdtable_size);
GET_ENV_OPT_INT("EF_SPIN_USEC", ul_spin_usec);
GET_ENV_OPT_INT("EF_STACK_PER_THREAD",stack_per_thread);
GET_ENV_OPT_INT("EF_DONT_ACCELERATE", dont_accelerate);
GET_ENV_OPT_INT("EF_FDTABLE_STRICT", fdtable_strict);
GET_ENV_OPT_INT("EF_FDS_MT_SAFE", fds_mt_safe);
GET_ENV_OPT_INT("EF_NO_FAIL", no_fail);
GET_ENV_OPT_INT("EF_SA_ONSTACK_INTERCEPT", sa_onstack_intercept);
GET_ENV_OPT_INT("EF_ACCEPT_INHERIT_NONBLOCK", accept_force_inherit_nonblock);
GET_ENV_OPT_INT("EF_VFORK_MODE", vfork_mode);
#if CI_CFG_USERSPACE_PIPE
GET_ENV_OPT_INT("EF_PIPE", ul_pipe);
#endif
if( (s = getenv("EF_FORK_NETIF")) && sscanf(s, "%x", &v) == 1 ) {
opts->fork_netif = CI_MIN(v, CI_UNIX_FORK_NETIF_BOTH);
}
if( (s = getenv("EF_NETIF_DTOR")) && sscanf(s, "%x", &v) == 1 ) {
opts->netif_dtor = CI_MIN(v, CITP_NETIF_DTOR_ALL);
}
if( (s = getenv("EF_SIGNALS_NOPOSTPONE")) ) {
opts->signals_no_postpone = 0;
while( sscanf(s, "%u", &v) == 1 ) {
opts->signals_no_postpone |= (1 << (v-1));
s = strchr(s, ',');
if( s == NULL )
break;
s++;
}
}
if( (s = getenv("EF_CLUSTER_NAME")) ) {
strncpy(opts->cluster_name, s, CI_CFG_CLUSTER_NAME_LEN);
opts->cluster_name[CI_CFG_CLUSTER_NAME_LEN] = '\0';
}
else {
opts->cluster_name[0] = '\0';
}
GET_ENV_OPT_INT("EF_CLUSTER_SIZE", cluster_size);
if( opts->cluster_size < 2 )
log("ERROR: cluster_size < 2 are not supported");
GET_ENV_OPT_INT("EF_CLUSTER_RESTART", cluster_restart_opt);
get_env_opt_port_list(&opts->tcp_reuseports, "EF_TCP_FORCE_REUSEPORT");
get_env_opt_port_list(&opts->udp_reuseports, "EF_UDP_FORCE_REUSEPORT");
#if CI_CFG_FD_CACHING
get_env_opt_port_list(&opts->sock_cache_ports, "EF_SOCKET_CACHE_PORTS");
#endif
}
static void citp_opts_getenv(citp_opts_t* opts)
{
/* ?? TODO: would like to use opts_citp_def.h here */
const char* s;
unsigned v;
opts->log_via_ioctl = 3;
/* TODO: Old name. Keeping reading 'til 2011, then purge. */
GET_ENV_OPT_HEX("EF_Log_VIA_IOCTL", log_via_ioctl);
GET_ENV_OPT_INT("EF_LOG_VIA_IOCTL", log_via_ioctl);
if( (s = getenv("EF_LOG_FILE")) && opts->log_via_ioctl == 3) {
opts->log_via_ioctl = 0;
citp_log_to_file(s);
} else if( opts->log_via_ioctl == 3 ) {
/* citp_setup_logging_early() have already detected stderr as
* tty/non-tty, so just trust it. */
if( ci_log_fn == citp_log_fn_drv )
opts->log_via_ioctl = 1;
else
opts->log_via_ioctl = 0;
}
if( opts->log_via_ioctl ) {
ci_log_options &=~ CI_LOG_PID;
citp_setup_logging_change(citp_log_fn_drv);
} else {
GET_ENV_OPT_INT("EF_LOG_TIMESTAMPS", log_timestamps);
if( opts->log_timestamps )
ci_log_options |= CI_LOG_TIME;
citp_setup_logging_change(citp_log_fn_ul);
}
if( getenv("EF_LOG_THREAD") )
ci_log_options |= CI_LOG_TID;
if( getenv("EF_POLL_NONBLOCK_FAST_LOOPS") &&
! getenv("EF_POLL_NONBLOCK_FAST_USEC") )
log("ERROR: EF_POLL_NONBLOCK_FAST_LOOPS is deprecated, use"
" EF_POLL_NONBLOCK_FAST_USEC instead");
if( getenv("EF_POLL_FAST_LOOPS") && ! getenv("EF_POLL_FAST_USEC") )
log("ERROR: EF_POLL_FAST_LOOPS is deprecated, use"
" EF_POLL_FAST_USEC instead");
if( (s = getenv("EF_POLL_USEC")) && atoi(s) ) {
/* Any changes to the behaviour triggered by this meta
* option must also be made to the extensions API option
* ONLOAD_SPIN_MIMIC_EF_POLL
*/
GET_ENV_OPT_INT("EF_POLL_USEC", ul_spin_usec);
GET_ENV_OPT_INT("EF_SLEEP_SPIN_USEC", sleep_spin_usec);
opts->ul_select_spin = 1;
opts->ul_poll_spin = 1;
#if CI_CFG_USERSPACE_EPOLL
opts->ul_epoll_spin = 1;
#endif
#if CI_CFG_UDP
opts->udp_recv_spin = 1;
opts->udp_send_spin = 1;
#endif
opts->tcp_recv_spin = 1;
opts->tcp_send_spin = 1;
opts->pkt_wait_spin = 1;
opts->sock_lock_buzz = 1;
opts->stack_lock_buzz = 1;
}
if( (s = getenv("EF_BUZZ_USEC")) && atoi(s) ) {
opts->sock_lock_buzz = 1;
opts->stack_lock_buzz = 1;
}
GET_ENV_OPT_HEX("EF_UNIX_LOG", log_level);
GET_ENV_OPT_INT("EF_PROBE", probe);
GET_ENV_OPT_INT("EF_TCP", ul_tcp);
GET_ENV_OPT_INT("EF_UDP", ul_udp);
GET_ENV_OPT_INT("EF_UL_SELECT", ul_select);
GET_ENV_OPT_INT("EF_SELECT_SPIN", ul_select_spin);
GET_ENV_OPT_INT("EF_SELECT_FAST", ul_select_fast);
GET_ENV_OPT_INT("EF_UL_POLL", ul_poll);
GET_ENV_OPT_INT("EF_POLL_SPIN", ul_poll_spin);
GET_ENV_OPT_INT("EF_POLL_FAST", ul_poll_fast);
GET_ENV_OPT_INT("EF_POLL_FAST_USEC", ul_poll_fast_usec);
GET_ENV_OPT_INT("EF_POLL_NONBLOCK_FAST_USEC", ul_poll_nonblock_fast_usec);
GET_ENV_OPT_INT("EF_SELECT_FAST_USEC", ul_select_fast_usec);
GET_ENV_OPT_INT("EF_SELECT_NONBLOCK_FAST_USEC", ul_select_nonblock_fast_usec);
#if CI_CFG_UDP
GET_ENV_OPT_INT("EF_UDP_RECV_SPIN", udp_recv_spin);
GET_ENV_OPT_INT("EF_UDP_SEND_SPIN", udp_send_spin);
#endif
GET_ENV_OPT_INT("EF_TCP_RECV_SPIN", tcp_recv_spin);
GET_ENV_OPT_INT("EF_TCP_SEND_SPIN", tcp_send_spin);
GET_ENV_OPT_INT("EF_TCP_ACCEPT_SPIN", tcp_accept_spin);
GET_ENV_OPT_INT("EF_TCP_CONNECT_SPIN",tcp_connect_spin);
GET_ENV_OPT_INT("EF_PKT_WAIT_SPIN", pkt_wait_spin);
#if CI_CFG_USERSPACE_PIPE
GET_ENV_OPT_INT("EF_PIPE_RECV_SPIN", pipe_recv_spin);
GET_ENV_OPT_INT("EF_PIPE_SEND_SPIN", pipe_send_spin);
GET_ENV_OPT_INT("EF_PIPE_SIZE", pipe_size);
#endif
GET_ENV_OPT_INT("EF_SOCK_LOCK_BUZZ", sock_lock_buzz);
GET_ENV_OPT_INT("EF_STACK_LOCK_BUZZ", stack_lock_buzz);
GET_ENV_OPT_INT("EF_SO_BUSY_POLL_SPIN", so_busy_poll_spin);
#if CI_CFG_USERSPACE_EPOLL
GET_ENV_OPT_INT("EF_UL_EPOLL", ul_epoll);
GET_ENV_OPT_INT("EF_EPOLL_SPIN", ul_epoll_spin);
GET_ENV_OPT_INT("EF_EPOLL_CTL_FAST", ul_epoll_ctl_fast);
GET_ENV_OPT_INT("EF_EPOLL_CTL_HANDOFF",ul_epoll_ctl_handoff);
GET_ENV_OPT_INT("EF_EPOLL_MT_SAFE", ul_epoll_mt_safe);
GET_ENV_OPT_INT("EF_WODA_SINGLE_INTERFACE", woda_single_if);
#endif
GET_ENV_OPT_INT("EF_FDTABLE_SIZE", fdtable_size);
GET_ENV_OPT_INT("EF_SPIN_USEC", ul_spin_usec);
GET_ENV_OPT_INT("EF_SLEEP_SPIN_USEC", sleep_spin_usec);
GET_ENV_OPT_INT("EF_STACK_PER_THREAD",stack_per_thread);
GET_ENV_OPT_INT("EF_DONT_ACCELERATE", dont_accelerate);
GET_ENV_OPT_INT("EF_FDTABLE_STRICT", fdtable_strict);
GET_ENV_OPT_INT("EF_FDS_MT_SAFE", fds_mt_safe);
GET_ENV_OPT_INT("EF_NO_FAIL", no_fail);
GET_ENV_OPT_INT("EF_SA_ONSTACK_INTERCEPT", sa_onstack_intercept);
GET_ENV_OPT_INT("EF_ACCEPT_INHERIT_NONBLOCK", accept_force_inherit_nonblock);
GET_ENV_OPT_INT("EF_VFORK_MODE", vfork_mode);
#if CI_CFG_USERSPACE_PIPE
GET_ENV_OPT_INT("EF_PIPE", ul_pipe);
#endif
GET_ENV_OPT_INT("EF_SYNC_CPLANE_AT_CREATE", sync_cplane);
if( (s = getenv("EF_FORK_NETIF")) && sscanf(s, "%x", &v) == 1 ) {
opts->fork_netif = CI_MIN(v, CI_UNIX_FORK_NETIF_BOTH);
}
if( (s = getenv("EF_NETIF_DTOR")) && sscanf(s, "%x", &v) == 1 ) {
opts->netif_dtor = CI_MIN(v, CITP_NETIF_DTOR_ALL);
}
if( (s = getenv("EF_SIGNALS_NOPOSTPONE")) ) {
opts->signals_no_postpone = 0;
while( sscanf(s, "%u", &v) == 1 ) {
opts->signals_no_postpone |= (1 << (v-1));
s = strchr(s, ',');
if( s == NULL )
break;
s++;
}
}
if( (s = getenv("EF_CLUSTER_NAME")) ) {
strncpy(opts->cluster_name, s, CI_CFG_CLUSTER_NAME_LEN);
opts->cluster_name[CI_CFG_CLUSTER_NAME_LEN] = '\0';
}
else {
opts->cluster_name[0] = '\0';
}
GET_ENV_OPT_INT("EF_CLUSTER_SIZE", cluster_size);
if( opts->cluster_size < 1 )
log("ERROR: cluster_size needs to be a positive number");
GET_ENV_OPT_INT("EF_CLUSTER_RESTART", cluster_restart_opt);
GET_ENV_OPT_INT("EF_CLUSTER_HOT_RESTART", cluster_hot_restart_opt);
get_env_opt_port_list(&opts->tcp_reuseports, "EF_TCP_FORCE_REUSEPORT");
get_env_opt_port_list(&opts->udp_reuseports, "EF_UDP_FORCE_REUSEPORT");
#if CI_CFG_FD_CACHING
get_env_opt_port_list(&opts->sock_cache_ports, "EF_SOCKET_CACHE_PORTS");
#endif
GET_ENV_OPT_INT("EF_ONLOAD_FD_BASE", fd_base);
}
/*
** 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;
}