static void citp_fdinfo_do_handover(citp_fdinfo* fdi, int fdt_locked)
{
int rc;
citp_fdinfo* epoll_fdi = NULL;
int os_fd = fdi->fd;
#ifndef NDEBUG
/* Yuk: does for UDP too. */
volatile citp_fdinfo_p* p_fdip;
p_fdip = &citp_fdtable.table[fdi->fd].fdip;
ci_assert(fdip_is_busy(*p_fdip));
#endif
Log_V(ci_log("%s: fd=%d nonb_switch=%d", __FUNCTION__, fdi->fd,
fdi->on_rcz.handover_nonb_switch));
if( fdi->epoll_fd >= 0 ) {
epoll_fdi = citp_epoll_fdi_from_member(fdi, fdt_locked);
if( epoll_fdi->protocol->type == CITP_EPOLLB_FD )
citp_epollb_on_handover(epoll_fdi, fdi);
}
rc = fdtable_fd_move(fdi->fd, OO_IOC_TCP_HANDOVER);
if( rc == -EBUSY && fdi->epoll_fd >= 0 ) {
ci_assert(fdi_to_sock_fdi(fdi)->sock.s->b.sb_aflags &
CI_SB_AFLAG_MOVED_AWAY);
/* If this is our epoll, we can do full handover: we manually add os
* fd into the epoll set.
* Fixme: ensure we are not in _other_ epoll sets */
ci_bit_clear(&fdi_to_sock_fdi(fdi)->sock.s->b.sb_aflags,
CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL_BIT);
rc = fdtable_fd_move(fdi->fd, OO_IOC_FILE_MOVED);
}
if( rc != 0 ) {
citp_fdinfo* new_fdi;
if( ! fdt_locked ) CITP_FDTABLE_LOCK();
new_fdi = citp_fdtable_probe_locked(fdi->fd, CI_TRUE, CI_TRUE);
citp_fdinfo_release_ref(new_fdi, 1);
if( ! fdt_locked ) CITP_FDTABLE_UNLOCK();
ci_assert_equal(citp_fdinfo_get_type(new_fdi), CITP_PASSTHROUGH_FD);
os_fd = fdi_to_alien_fdi(new_fdi)->os_socket;
}
if( fdi->on_rcz.handover_nonb_switch >= 0 ) {
int on_off = !! fdi->on_rcz.handover_nonb_switch;
int rc = ci_sys_ioctl(os_fd, FIONBIO, &on_off);
if( rc < 0 )
Log_E(ci_log("%s: ioctl failed on_off=%d", __FUNCTION__, on_off));
}
if( rc != 0 )
goto exit;
citp_fdtable_busy_clear(fdi->fd, fdip_passthru, fdt_locked);
exit:
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked);
if( epoll_fdi != NULL && epoll_fdi->protocol->type == CITP_EPOLL_FD )
citp_epoll_on_handover(epoll_fdi, fdi, fdt_locked);
if( epoll_fdi != NULL )
citp_fdinfo_release_ref(epoll_fdi, fdt_locked);
citp_fdinfo_free(fdi);
}
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);
}
/* Looks up the user-level 'FD info' for a given file descriptor.
** Returns pointer to the user-level 'FD info' for a given file
** descriptor, or NULL if the FD is not user-level.
** NOTE: The reference count of the 'FD info' is incremented, the
** caller should ensure the reference is dropped when no
** longer needed by calling citp_fdinfo_release_ref().
*/
citp_fdinfo* citp_fdtable_lookup_noprobe(unsigned fd)
{
/* Need to be initialised before we can try and grab the lock at the
** moment. TODO: make this more efficient by using a trylock to grab the
** fdtable lock, and on fail see if we need to initialise it.
*/
if( CI_UNLIKELY(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 */
return NULL;
}
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( fdip_cas_succeed(p_fdip, fdip, fdip_busy) ) {
/* Bump the reference count. */
citp_fdinfo* fdi = fdip_to_fdi(fdip);
citp_fdinfo_ref(fdi);
/* Swap the busy marker out again. */
citp_fdtable_busy_clear(fd, fdip, 0);
return fdi;
}
goto again;
}
/* Not normal! */
else if( fdip_is_busy(fdip) ) {
citp_fdtable_busy_wait(fd, 0);
goto again;
}
}
return NULL;
}
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;
}
/* Find out what sort of thing [fd] is, and if it is a user-level socket
* then map in the user-level state.
*/
static citp_fdinfo * citp_fdtable_probe_locked(unsigned fd, int print_banner,
int fdip_is_already_busy)
{
citp_fdinfo* fdi = NULL;
struct stat64 st;
ci_ep_info_t info;
if( ! fdip_is_already_busy ) {
volatile citp_fdinfo_p* p_fdip;
citp_fdinfo_p fdip;
/* ?? We're repeating some effort already expended in lookup() here, but
** this keeps it cleaner. May optimise down the line when I understand
** what other code needs to call this.
*/
p_fdip = &citp_fdtable.table[fd].fdip;
again:
fdip = *p_fdip;
if( fdip_is_busy(fdip) ) fdip = citp_fdtable_busy_wait(fd, 1);
if( ! fdip_is_unknown(fdip) && ! fdip_is_normal(fdip) ) goto exit;
if( fdip_cas_fail(p_fdip, fdip, fdip_busy) ) goto again;
if( fdip_is_normal(fdip) ) {
fdi = fdip_to_fdi(fdip);
citp_fdinfo_ref(fdi);
citp_fdtable_busy_clear(fd, fdip, 1);
goto exit;
}
}
if( ci_sys_fstat64(fd, &st) != 0 ) {
/* fstat() failed. Must be a bad (closed) file descriptor, so
** leave this entry as unknown. Return citp_the_closed_fd to avoid the
** caller passing through to an fd that is created asynchronously.
*/
citp_fdtable_busy_clear(fd, fdip_unknown, 1);
fdi = &citp_the_closed_fd;
citp_fdinfo_ref(fdi);
goto exit;
}
/* oo_get_st_rdev() and oo_onloadfs_dev_t() open-and-close fd, so
* fdtable should be locked if strict mode requested. */
if( fdtable_strict() ) { CITP_FDTABLE_ASSERT_LOCKED(1); }
if( st.st_dev == oo_onloadfs_dev_t() ) {
/* Retrieve user-level endpoint info */
if( oo_ep_info(fd, &info) < 0 ) {
Log_V(log("%s: fd=%d type=%d unknown", __FUNCTION__,fd,info.fd_type));
citp_fdtable_busy_clear(fd, fdip_passthru, 1);
goto exit;
}
switch( info.fd_type ) {
case CI_PRIV_TYPE_TCP_EP:
case CI_PRIV_TYPE_UDP_EP:
case CI_PRIV_TYPE_PASSTHROUGH_EP:
case CI_PRIV_TYPE_ALIEN_EP:
#if CI_CFG_USERSPACE_PIPE
case CI_PRIV_TYPE_PIPE_READER:
case CI_PRIV_TYPE_PIPE_WRITER:
#endif
{
citp_fdinfo_p fdip;
Log_V(log("%s: fd=%d %s restore", __FUNCTION__, fd,
info.fd_type == CI_PRIV_TYPE_TCP_EP ? "TCP":
#if CI_CFG_USERSPACE_PIPE
info.fd_type != CI_PRIV_TYPE_UDP_EP ? "PIPE" :
#endif
"UDP"));
fdip = citp_fdtable_probe_restore(fd, &info, print_banner);
if( fdip_is_normal(fdip) )
fdi = fdip_to_fdi(fdip);
else
citp_fdtable_busy_clear(fd, fdip, 1);
goto exit;
}
case CI_PRIV_TYPE_NETIF:
/* This should never happen, because netif fds are close-on-exec.
** But let's leave this code here just in case my reasoning is bad.
*/
Log_U(log("%s: fd=%d NETIF reserved", __FUNCTION__, fd));
citp_fdtable_busy_clear(fd, fdip_reserved, 1);
fdi = &citp_the_reserved_fd;
citp_fdinfo_ref(fdi);
goto exit;
case CI_PRIV_TYPE_NONE:
/* This happens if a thread gets at an onload driver fd that has just
* been created, but not yet specialised. On Linux I think this
* means it will shortly be a new netif internal fd. (fds associated
* with sockets and pipes are never unspecialised).
*/
Log_V(log("%s: fd=%d TYPE_NONE", __FUNCTION__, fd));
citp_fdtable_busy_clear(fd, fdip_passthru, 1);
goto exit;
default:
CI_TEST(0);
break;
}
}
else if( ci_major(st.st_rdev) == ci_major(oo_get_st_rdev(OO_EPOLL_DEV)) ) {
citp_epollb_fdi *epi = CI_ALLOC_OBJ(citp_epollb_fdi);
if( ! epi ) {
Log_E(log("%s: out of memory (epoll_fdi)", __FUNCTION__));
citp_fdtable_busy_clear(fd, fdip_passthru, 1);
goto exit;
}
oo_epollb_ctor(epi);
fdi = &epi->fdinfo;
citp_fdinfo_init(fdi, &citp_epollb_protocol_impl);
citp_fdinfo_ref(fdi);
citp_fdtable_insert(fdi, fd, 1);
goto exit;
}
#ifndef NDEBUG
/* /dev/onload may be netif only; they are closed on fork or exec */
if( ci_major(st.st_rdev) == ci_major(oo_get_st_rdev(OO_STACK_DEV)) )
Log_U(log("%s: %d is /dev/onload", __FUNCTION__, fd));
#endif
/* Not one of ours, so pass-through. */
Log_V(log("%s: fd=%u non-efab", __FUNCTION__, fd));
citp_fdtable_busy_clear(fd, fdip_passthru, 1);
exit:
return fdi;
}
static citp_fdinfo_p
citp_fdtable_probe_restore(int fd, ci_ep_info_t * info, int print_banner)
{
citp_protocol_impl* proto = 0;
citp_fdinfo* fdi = 0;
ci_netif* ni;
int rc;
int c_sock_fdi = 1;
/* Must be holding the FD table writer lock */
CITP_FDTABLE_ASSERT_LOCKED(1);
ci_assert_nequal(info->resource_id, CI_ID_POOL_ID_NONE);
/* Will need to review this function if the following assert fires */
switch( info->fd_type ) {
case CI_PRIV_TYPE_TCP_EP: proto = &citp_tcp_protocol_impl; break;
case CI_PRIV_TYPE_UDP_EP: proto = &citp_udp_protocol_impl; break;
case CI_PRIV_TYPE_PASSTHROUGH_EP:
proto = &citp_passthrough_protocol_impl;
c_sock_fdi = 0;
break;
case CI_PRIV_TYPE_ALIEN_EP:
proto = NULL;
c_sock_fdi = 0;
break;
#if CI_CFG_USERSPACE_PIPE
case CI_PRIV_TYPE_PIPE_READER:
proto = &citp_pipe_read_protocol_impl;
c_sock_fdi = 0;
break;
case CI_PRIV_TYPE_PIPE_WRITER:
proto = &citp_pipe_write_protocol_impl;
c_sock_fdi = 0;
break;
#endif
default: ci_assert(0);
}
/* Attempt to find the user-level netif for this endpoint */
ni = citp_find_ul_netif(info->resource_id, 1);
if( ! ni ) {
ef_driver_handle netif_fd;
/* Not found, rebuild/restore the netif for this endpoint */
rc = citp_netif_recreate_probed(fd, &netif_fd, &ni);
if ( rc < 0 ) {
Log_E(log("%s: citp_netif_recreate_probed failed! (%d)",
__FUNCTION__, rc));
goto fail;
}
if( print_banner ) {
ci_log("Importing "ONLOAD_PRODUCT" "ONLOAD_VERSION" "ONLOAD_COPYRIGHT
" [%s]", ni->state->pretty_name);
}
}
else
citp_netif_add_ref(ni);
/* There is a race condition where the fd can have been created, but it has
* not yet been initialised, as we can't put a busy marker in the right place
* in the fdtable until we know what the fd is. In this case we don't want
* to probe this new info, so return the closed fd.
*/
if( SP_TO_WAITABLE(ni, info->sock_id)->sb_aflags & CI_SB_AFLAG_NOT_READY ) {
citp_fdtable_busy_clear(fd, fdip_unknown, 1);
fdi = &citp_the_closed_fd;
citp_fdinfo_ref(fdi);
return fdi_to_fdip(fdi);
}
if (c_sock_fdi) {
citp_sock_fdi* sock_fdi;
sock_fdi = CI_ALLOC_OBJ(citp_sock_fdi);
if( ! sock_fdi ) {
Log_E(log("%s: out of memory (sock_fdi)", __FUNCTION__));
goto fail;
}
fdi = &sock_fdi->fdinfo;
sock_fdi->sock.s = SP_TO_SOCK_CMN(ni, info->sock_id);
sock_fdi->sock.netif = ni;
}
else if( info->fd_type == CI_PRIV_TYPE_PASSTHROUGH_EP ) {
citp_waitable* w = SP_TO_WAITABLE(ni, info->sock_id);
citp_alien_fdi* alien_fdi;
if( ~w->sb_aflags & CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL &&
fdtable_fd_move(fd, OO_IOC_FILE_MOVED) == 0 ) {
citp_netif_release_ref(ni, 1);
return fdip_passthru;
}
alien_fdi = CI_ALLOC_OBJ(citp_alien_fdi);
if( ! alien_fdi ) {
Log_E(log("%s: out of memory (alien_fdi)", __FUNCTION__));
goto fail;
}
fdi = &alien_fdi->fdinfo;
alien_fdi->netif = ni;
alien_fdi->ep = SP_TO_WAITABLE(ni, info->sock_id);
citp_passthrough_init(alien_fdi);
}
else if( info->fd_type == CI_PRIV_TYPE_ALIEN_EP ) {
citp_waitable* w = SP_TO_WAITABLE(ni, info->sock_id);
citp_sock_fdi* sock_fdi;
ci_netif* alien_ni;
sock_fdi = CI_ALLOC_OBJ(citp_sock_fdi);
if( ! sock_fdi ) {
Log_E(log("%s: out of memory (alien sock_fdi)", __FUNCTION__));
goto fail;
}
fdi = &sock_fdi->fdinfo;
rc = citp_netif_by_id(w->moved_to_stack_id, &alien_ni, 1);
if( rc != 0 ) {
goto fail;
}
sock_fdi->sock.s = SP_TO_SOCK_CMN(alien_ni, w->moved_to_sock_id);
sock_fdi->sock.netif = alien_ni;
citp_netif_release_ref(ni, 1);
/* Replace the file under this fd if possible */
if( ~w->sb_aflags & CI_SB_AFLAG_MOVED_AWAY_IN_EPOLL )
fdtable_fd_move(fd, OO_IOC_FILE_MOVED);
if( sock_fdi->sock.s->b.state & CI_TCP_STATE_TCP )
proto = &citp_tcp_protocol_impl;
else if( sock_fdi->sock.s->b.state == CI_TCP_STATE_UDP )
proto = &citp_udp_protocol_impl;
else {
CI_TEST(0);
}
}
#if CI_CFG_USERSPACE_PIPE
else {
citp_pipe_fdi* pipe_fdi;
pipe_fdi = CI_ALLOC_OBJ(citp_pipe_fdi);
if( ! pipe_fdi ) {
Log_E(log("%s: out of memory (pipe_fdi)", __FUNCTION__));
goto fail;
}
fdi = &pipe_fdi->fdinfo;
pipe_fdi->pipe = SP_TO_PIPE(ni, info->sock_id);
pipe_fdi->ni = ni;
}
#endif
citp_fdinfo_init(fdi, proto);
/* We're returning a reference to the caller. */
citp_fdinfo_ref(fdi);
citp_fdtable_insert(fdi, fd, 1);
return fdi_to_fdip(fdi);
fail:
if( ni ) citp_netif_release_ref(ni, 1);
return fdip_unknown;
}
/* Re-probe fdinfo after endpoint was moved to another stack.
* The function assumes that fdinfo was obtained via citp_fdtable_lookup()
* or from citp_fdtable_lookup_fast(). */
citp_fdinfo* citp_reprobe_moved(citp_fdinfo* fdinfo, int from_fast_lookup,
int fdip_is_already_busy)
{
int fd = fdinfo->fd;
citp_fdinfo* new_fdinfo = NULL;
CITP_FDTABLE_LOCK();
if( ! fdip_is_already_busy ) {
volatile citp_fdinfo_p* p_fdip;
citp_fdinfo_p fdip;
p_fdip = &citp_fdtable.table[fd].fdip;
again:
fdip = *p_fdip;
if( fdip_is_busy(fdip) ) fdip = citp_fdtable_busy_wait(fd, 1);
ci_assert( fdip_is_normal(fdip) || fdip_is_passthru(fdip) );
if( fdip_cas_fail(p_fdip, fdip, fdip_busy) ) goto again;
/* Possibly, a parrallel thread have already called
* citp_reprobe_moved() for us. */
if( fdip_is_passthru(fdip) ) {
citp_fdtable_busy_clear(fd, fdip, 1);
if( new_fdinfo != NULL )
citp_fdinfo_ref(new_fdinfo);
goto done;
}
ci_assert( fdip_is_normal(fdip) );
new_fdinfo = fdip_to_fdi(fdip);
if( new_fdinfo != fdinfo) {
citp_fdtable_busy_clear(fd, fdip, 1);
if( new_fdinfo != NULL )
citp_fdinfo_ref(new_fdinfo);
goto done;
}
}
else
ci_assert(fdip_is_busy(citp_fdtable.table[fd].fdip));
/* re-probe new fd */
new_fdinfo = citp_fdtable_probe_locked(fd, CI_TRUE, CI_TRUE);
if( fdinfo->epoll_fd >= 0 ) {
citp_fdinfo* epoll_fdi = citp_epoll_fdi_from_member(fdinfo, 1);
if( epoll_fdi->protocol->type == CITP_EPOLL_FD )
citp_epoll_on_move(epoll_fdi, fdinfo, new_fdinfo, 1);
else
citp_epollb_on_handover(epoll_fdi, fdinfo);
citp_fdinfo_release_ref(epoll_fdi, 1);
}
/* Drop refcount from fdtable */
fdinfo->on_ref_count_zero = FDI_ON_RCZ_MOVED;
citp_fdinfo_release_ref(fdinfo, 1);
done:
/* One refcount from the caller */
if( from_fast_lookup )
citp_fdinfo_release_ref_fast(fdinfo);
else
citp_fdinfo_release_ref(fdinfo, 1);
CITP_FDTABLE_UNLOCK();
if( new_fdinfo == NULL )
return NULL;
if( from_fast_lookup ) {
citp_fdinfo_ref_fast(new_fdinfo);
citp_fdinfo_release_ref(new_fdinfo, 0);
}
return new_fdinfo;
}
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;
}
static void dup2_complete(citp_fdinfo* prev_tofdi,
citp_fdinfo_p prev_tofdip, int fdt_locked)
{
volatile citp_fdinfo_p *p_fromfdip;
unsigned fromfd = prev_tofdi->on_rcz.dup3_args.fd;
unsigned tofd = prev_tofdi->fd;
citp_fdinfo_p fromfdip;
int rc;
#if CI_LIBC_HAS_dup3 || !defined(NDEBUG)
int flags = prev_tofdi->on_rcz.dup3_args.flags;
#endif
#ifndef NDEBUG
volatile citp_fdinfo_p* p_tofdip;
p_tofdip = &citp_fdtable.table[tofd].fdip;
ci_assert(fdip_is_busy(*p_tofdip));
#endif
citp_fdinfo* fromfdi;
p_fromfdip = &citp_fdtable.table[fromfd].fdip;
lock_fromfdip_again:
fromfdip = *p_fromfdip;
if( fdip_is_busy(fromfdip) )
fromfdip = citp_fdtable_busy_wait(fromfd, fdt_locked);
if( fdip_is_closing(fromfdip) | fdip_is_reserved(fromfdip) ) {
prev_tofdi->on_rcz.dup2_result = -EBADF;
ci_wmb();
prev_tofdi->on_ref_count_zero = FDI_ON_RCZ_DONE;
return;
}
#if CI_CFG_FD_CACHING
/* Need to check in case this sucker's cached */
if( fdip_is_unknown(fromfdip) ) {
if( !fdt_locked ) CITP_FDTABLE_LOCK();
fromfdi = citp_fdtable_probe_locked(fromfd, CI_FALSE, CI_FALSE);
if( !fdt_locked ) CITP_FDTABLE_UNLOCK();
if( fromfdi == &citp_the_closed_fd ) {
prev_tofdi->on_rcz.dup2_result = -EBADF;
ci_wmb();
prev_tofdi->on_ref_count_zero = FDI_ON_RCZ_DONE;
citp_fdinfo_release_ref(fromfdi, CI_TRUE);
return;
}
if( fromfdi )
citp_fdinfo_release_ref(fromfdi, CI_TRUE);
}
#endif
if( fdip_cas_fail(p_fromfdip, fromfdip, fdip_busy) )
goto lock_fromfdip_again;
oo_rwlock_lock_write(&citp_dup2_lock);
#if CI_LIBC_HAS_dup3
rc = ci_sys_dup3(fromfd, tofd, flags);
#else
ci_assert_equal(flags, 0);
rc = ci_sys_dup2(fromfd, tofd);
#endif
oo_rwlock_unlock_write(&citp_dup2_lock);
if( rc < 0 ) {
citp_fdtable_busy_clear(fromfd, fromfdip, fdt_locked);
prev_tofdi->on_rcz.dup2_result = -errno;
ci_wmb();
prev_tofdi->on_ref_count_zero = FDI_ON_RCZ_DONE;
return;
}
ci_assert(fdip_is_normal(fromfdip) | fdip_is_passthru(fromfdip) |
fdip_is_unknown(fromfdip));
if( fdip_is_normal(fromfdip) &&
(((fromfdi = fdip_to_fdi(fromfdip))->protocol->type) == CITP_EPOLL_FD) ) {
citp_fdinfo* newfdi = citp_fdinfo_get_ops(fromfdi)->dup(fromfdi);
if( newfdi ) {
citp_fdinfo_init(newfdi, fdip_to_fdi(fromfdip)->protocol);
citp_fdtable_insert(newfdi, tofd, fdt_locked);
}
else {
/* Out of memory. Can't probe epoll1 fd later on, so fail. */
citp_fdtable_busy_clear(fromfd, fromfdip, fdt_locked);
prev_tofdi->on_rcz.dup2_result = -ENOMEM;
ci_wmb();
prev_tofdi->on_ref_count_zero = FDI_ON_RCZ_DONE;
return;
}
}
else {
/* Mark newfd as unknown. When used, it'll get probed.
*
* We are not just being lazy here: Setting to unknown rather than
* installing a proper fdi (when oldfd is accelerated) is essential to
* vfork()+dup2()+exec() working properly. Reason is that child and
* parent share address space, so child is modifying the parent's
* fdtable. Setting an entry to unknown is safe.
*/
citp_fdtable_busy_clear(tofd, fdip_unknown, fdt_locked);
#if CI_CFG_FD_CACHING
/* Multiple refs to this now, don't allow it to be cached. */
if( fdip_is_normal(fromfdip) )
fdip_to_fdi(fromfdip)->can_cache = 0;
#endif
}
citp_fdtable_busy_clear(fromfd, fromfdip, fdt_locked);
prev_tofdi->on_rcz.dup2_result = tofd;
ci_wmb();
prev_tofdi->on_ref_count_zero = FDI_ON_RCZ_DONE;
}
/*
** Why do these live here? Because they need to hack into the low-level
** dirty nastiness of the fdtable.
*/
int citp_ep_dup(unsigned oldfd, int (*syscall)(int oldfd, long arg),
long arg)
{
/* This implements dup(oldfd) and fcntl(oldfd, F_DUPFD, arg). */
volatile citp_fdinfo_p* p_oldfdip;
citp_fdinfo_p oldfdip;
citp_fdinfo* newfdi = 0;
citp_fdinfo* oldfdi;
int newfd;
Log_V(log("%s(%d)", __FUNCTION__, oldfd));
if(CI_UNLIKELY( citp.init_level < CITP_INIT_FDTABLE ||
oo_per_thread_get()->in_vfork_child ))
/* Lib not initialised, so no U/L state, and therefore system dup()
** will do just fine. */
return syscall(oldfd, arg);
if( oldfd >= citp_fdtable.inited_count ) {
/* NB. We can't just pass through in this case because we need to worry
** about other threads racing with us. So we need to be able to lock
** this fd while we do the dup. */
ci_assert(oldfd < citp_fdtable.size);
CITP_FDTABLE_LOCK();
__citp_fdtable_extend(oldfd);
CITP_FDTABLE_UNLOCK();
}
p_oldfdip = &citp_fdtable.table[oldfd].fdip;
again:
oldfdip = *p_oldfdip;
if( fdip_is_busy(oldfdip) )
oldfdip = citp_fdtable_busy_wait(oldfd, 0);
if( fdip_is_closing(oldfdip) | fdip_is_reserved(oldfdip) ) {
errno = EBADF;
return -1;
}
#if CI_CFG_FD_CACHING
/* Need to check in case this sucker's cached */
if( fdip_is_unknown(oldfdip) ) {
CITP_FDTABLE_LOCK();
oldfdi = citp_fdtable_probe_locked(oldfd, CI_FALSE, CI_FALSE);
CITP_FDTABLE_UNLOCK();
if( oldfdi == &citp_the_closed_fd ) {
citp_fdinfo_release_ref(oldfdi, CI_TRUE);
errno = EBADF;
return -1;
}
if( oldfdi )
citp_fdinfo_release_ref(oldfdi, CI_TRUE);
}
#endif
if( fdip_cas_fail(p_oldfdip, oldfdip, fdip_busy) )
goto again;
#if CI_CFG_FD_CACHING
/* May end up with multiple refs to this, don't allow it to be cached. */
if( fdip_is_normal(oldfdip) )
fdip_to_fdi(oldfdip)->can_cache = 0;
#endif
if( fdip_is_normal(oldfdip) &&
(((oldfdi = fdip_to_fdi(oldfdip))->protocol->type) == CITP_EPOLL_FD) ) {
newfdi = citp_fdinfo_get_ops(oldfdi)->dup(oldfdi);
if( ! newfdi ) {
citp_fdtable_busy_clear(oldfd, oldfdip, 0);
errno = ENOMEM;
return -1;
}
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
newfd = syscall(oldfd, arg);
if( newfd >= 0 )
citp_fdtable_new_fd_set(newfd, fdip_busy, fdtable_strict());
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
if( newfd >= 0 ) {
citp_fdtable_insert(newfdi, newfd, 0);
newfdi = 0;
}
}
else {
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
newfd = syscall(oldfd, arg);
if( newfd >= 0 && newfd < citp_fdtable.inited_count ) {
/* Mark newfd as unknown. When used, it'll get probed.
*
* We are not just being lazy here: Setting to unknown rather than
* installing a proper fdi (when oldfd is accelerated) is essential to
* vfork()+dup()+exec() working properly. Reason is that child and
* parent share address space, so child is modifying the parent's
* fdtable. Setting an entry to unknown is safe.
*/
citp_fdtable_new_fd_set(newfd, fdip_unknown, fdtable_strict());
}
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
}
citp_fdtable_busy_clear(oldfd, oldfdip, 0);
if( newfdi ) citp_fdinfo_free(newfdi);
return newfd;
}
int citp_epoll_create(int size, int flags)
{
citp_fdinfo *fdi;
citp_epoll_fdi *epi;
struct citp_epoll_fd* ep;
int fd;
if( (epi = CI_ALLOC_OBJ(citp_epoll_fdi)) == NULL )
goto fail0;
if( (ep = CI_ALLOC_OBJ(struct citp_epoll_fd)) == NULL )
goto fail1;
fdi = &epi->fdinfo;
citp_fdinfo_init(fdi, &citp_epoll_protocol_impl);
/* Create the epoll fd. */
CITP_FDTABLE_LOCK();
if( (fd = ci_sys_epoll_create_compat(size, flags, 0)) < 0 )
goto fail2;
citp_fdtable_new_fd_set(fd, fdip_busy, TRUE);
/* Init epfd_os */
#ifdef O_CLOEXEC
ep->epfd_os = ci_sys_open(OO_EPOLL_DEV, O_RDWR | O_CLOEXEC);
#else
ep->epfd_os = ci_sys_open(OO_EPOLL_DEV, O_RDWR);
if( ep->epfd_os >= 0 )
ci_sys_fcntl(ep->epfd_os, F_SETFD, FD_CLOEXEC);
#endif
if( ep->epfd_os < 0 ) {
Log_E(ci_log("%s: ERROR: failed to open(%s) errno=%d",
__FUNCTION__, OO_EPOLL_DEV, errno));
goto fail3;
}
__citp_fdtable_reserve(ep->epfd_os, 1);
ep->shared = mmap(NULL, sizeof(*ep->shared), PROT_READ, MAP_SHARED,
ep->epfd_os, 0);
if( ep->shared == MAP_FAILED ) {
Log_E(ci_log("%s: ERROR: failed to mmap shared segment errno=%d",
__FUNCTION__, errno));
goto fail4;
}
__citp_fdtable_reserve(ep->shared->epfd, 1);
CITP_FDTABLE_UNLOCK();
epi->epoll = ep;
ep->size = size;
oo_wqlock_init(&ep->lock);
ep->not_mt_safe = ! CITP_OPTS.ul_epoll_mt_safe;
ci_dllist_init(&ep->oo_sockets);
ep->oo_sockets_n = 0;
ci_dllist_init(&ep->dead_sockets);
oo_atomic_set(&ep->refcount, 1);
ep->epfd_syncs_needed = 0;
ep->blocking = 0;
citp_fdtable_insert(fdi, fd, 0);
Log_POLL(ci_log("%s: fd=%d driver_fd=%d epfd=%d", __FUNCTION__,
fd, ep->epfd_os, (int) ep->shared->epfd));
return fd;
fail4:
__citp_fdtable_reserve(ep->epfd_os, 0);
ci_sys_close(ep->epfd_os);
fail3:
ci_sys_close(fd);
citp_fdtable_busy_clear(fd, fdip_unknown, 1);
fail2:
CITP_FDTABLE_UNLOCK();
CI_FREE_OBJ(ep);
fail1:
CI_FREE_OBJ(epi);
fail0:
return -1;
}