void __citp_fdinfo_ref_count_zero(citp_fdinfo* fdi, int fdt_locked)
{
Log_V(log("%s: fd=%d on_rcz=%d", __FUNCTION__, fdi->fd,
fdi->on_ref_count_zero));
citp_fdinfo_assert_valid(fdi);
ci_assert(oo_atomic_read(&fdi->ref_count) == 0);
ci_assert_ge(fdi->fd, 0);
ci_assert_lt(fdi->fd, citp_fdtable.inited_count);
ci_assert_nequal(fdi_to_fdip(fdi), citp_fdtable.table[fdi->fd].fdip);
switch( fdi->on_ref_count_zero ) {
case FDI_ON_RCZ_CLOSE:
#if CI_CFG_FD_CACHING
if( citp_fdinfo_get_ops(fdi)->cache(fdi) == 1 ) {
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_LOCK();
fdtable_swap(fdi->fd, fdip_closing, fdip_unknown,
fdt_locked | fdtable_strict());
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked | fdtable_strict());
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_UNLOCK();
citp_fdinfo_free(fdi);
break;
}
else
#endif
{
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_LOCK();
ci_tcp_helper_close_no_trampoline(fdi->fd);
/* The swap must occur after the close, otherwise another thread could
* cause a probe of the old endpoint info, which is about be freed.
*/
fdtable_swap(fdi->fd, fdip_closing, fdip_unknown,
fdt_locked | fdtable_strict());
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked | fdtable_strict());
if( ! fdt_locked && fdtable_strict() ) CITP_FDTABLE_UNLOCK();
citp_fdinfo_free(fdi);
break;
}
case FDI_ON_RCZ_DUP2:
dup2_complete(fdi, fdi_to_fdip(fdi), fdt_locked);
break;
case FDI_ON_RCZ_HANDOVER:
citp_fdinfo_do_handover(fdi, fdt_locked);
break;
case FDI_ON_RCZ_MOVED:
citp_fdinfo_get_ops(fdi)->dtor(fdi, fdt_locked);
citp_fdinfo_free(fdi);
break;
default:
CI_DEBUG(ci_log("%s: fd=%d on_ref_count_zero=%d", __FUNCTION__,
fdi->fd, fdi->on_ref_count_zero));
ci_assert(0);
}
}
static citp_fdinfo_p citp_fdtable_closing_wait(unsigned fd, int fdt_locked)
{
/* We're currently spinning in this case. Not ideal, but implementing
** blocking here is slightly tricky. (Can be done, but I want proof that
** it's needed first!)
*/
volatile citp_fdinfo_p* p_fdip = &citp_fdtable.table[fd].fdip;
citp_fdinfo_p fdip;
Log_V(ci_log("%s: fd=%u", __FUNCTION__, fd));
again:
fdip = *p_fdip;
if( fdip_is_busy(fdip) ) fdip = citp_fdtable_busy_wait(fd, fdt_locked);
if( fdip_is_closing(fdip) ) {
if( fdt_locked ) {
/* Need to drop the lock to avoid deadlock with the other thread
** trying to closing this fd! */
CITP_FDTABLE_UNLOCK();
CITP_FDTABLE_LOCK();
}
goto again;
}
return fdip;
}
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);
}
static void citp_passthrough_dtor(citp_fdinfo* fdi, int fdt_locked)
{
citp_alien_fdi* epi = fdi_to_alien_fdi(fdi);
CITP_FDTABLE_LOCK();
ci_tcp_helper_close_no_trampoline(epi->os_socket);
__citp_fdtable_reserve(epi->os_socket, 0);
CITP_FDTABLE_UNLOCK();
citp_netif_release_ref(fdi_to_alien_fdi(fdi)->netif, fdt_locked);
}
void citp_fdinfo_handover(citp_fdinfo* fdi, int nonb_switch)
{
/* Please see comments in internal.h. */
volatile citp_fdinfo_p* p_fdip;
citp_fdinfo_p fdip;
unsigned fd = fdi->fd;
/* We're about to free some user-level state, so we need to interlock
** against select and poll.
*/
CITP_FDTABLE_LOCK();
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 == fdi_to_fdip(fdi) ) {
if( fdip_cas_fail(p_fdip, fdip, fdip_busy) )
goto again;
}
else {
/* [fd] must have changed meaning under our feet. It must be closing,
** so do nothing except drop the ref passed in.
*/
ci_assert(fdip_is_closing(fdip));
ci_assert_nequal(fdi->on_ref_count_zero, FDI_ON_RCZ_NONE);
}
if( fdip == fdi_to_fdip(fdi) ) {
ci_assert_equal(fdi->on_ref_count_zero, FDI_ON_RCZ_NONE);
fdi->on_ref_count_zero = FDI_ON_RCZ_HANDOVER;
fdi->on_rcz.handover_nonb_switch = nonb_switch;
/* Drop the fdtable ref. When the ref count goes to zero, the handover
** will be done. We return without waiting, because the caller
** shouldn't do anything more with this socket anyway.
*/
citp_fdinfo_release_ref(fdi, 1);
}
/* Drop the ref passed in. */
citp_fdinfo_release_ref(fdi, 1);
CITP_FDTABLE_UNLOCK();
}
citp_fdinfo_p
citp_fdtable_new_fd_set(unsigned fd, citp_fdinfo_p new_fdip, int fdt_locked)
{
volatile citp_fdinfo_p* p_fdip;
citp_fdinfo_p prev;
if( fd >= citp_fdtable.inited_count ) {
ci_assert_lt(fd, citp_fdtable.size);
if( ! fdt_locked ) CITP_FDTABLE_LOCK();
__citp_fdtable_extend(fd);
if( ! fdt_locked ) CITP_FDTABLE_UNLOCK();
}
p_fdip = &citp_fdtable.table[fd].fdip;
do {
prev = *p_fdip;
/* Busy? Perhaps just closed, but not yet marked unknown. Or perhaps it
** is being probed. */
if( fdip_is_busy(prev) )
prev = citp_fdtable_busy_wait(fd, fdt_locked);
/* There is a close in progress, so we wait until it is resolved. */
if( fdip_is_closing(prev) )
prev = citp_fdtable_closing_wait(fd, fdt_locked);
/* Reserved? Perhaps it was a netif fd that has just been closed. So it
** should be about to be unreserved. */
} while (fdip_is_reserved(prev) || fdip_cas_fail(p_fdip, prev, new_fdip) );
if( fdip_is_normal(prev) ) {
/* We can get here is close-trampolining fails. So for release
** builds we accept that the user-level state got out-of-sync, and
** leak [fdi] since it seems like a suitably cautious thing to do.
*/
ci_log("%s: ERROR: Orphaned entry in user-level fd-table", __FUNCTION__);
}
else
/* We (at time of writing) only register a trampoline handler when we
** create a netif, so we can miss the closing of pass-through
** descriptors.
*/
ci_assert(fdip_is_unknown(prev) || fdip_is_passthru(prev));
return prev;
}
static void citp_epoll_dtor(citp_fdinfo* fdi, int fdt_locked)
{
struct citp_epoll_fd* ep = fdi_to_epoll(fdi);
if (!oo_atomic_dec_and_test(&ep->refcount))
return;
CITP_FDTABLE_LOCK();
ci_tcp_helper_close_no_trampoline(ep->shared->epfd);
__citp_fdtable_reserve(ep->shared->epfd, 0);
munmap(ep->shared, sizeof(*ep->shared));
ci_tcp_helper_close_no_trampoline(ep->epfd_os);
__citp_fdtable_reserve(ep->epfd_os, 0);
CITP_FDTABLE_UNLOCK();
CI_FREE_OBJ(ep);
}
static citp_fdinfo *
citp_fdtable_probe(unsigned fd)
{
citp_fdinfo* fdi;
int saved_errno;
ci_assert(fd < citp_fdtable.size);
if( ! CITP_OPTS.probe || oo_per_thread_get()->in_vfork_child )
return NULL;
saved_errno = errno;
CITP_FDTABLE_LOCK();
__citp_fdtable_extend(fd);
fdi = citp_fdtable_probe_locked(fd, CI_FALSE, CI_FALSE);
CITP_FDTABLE_UNLOCK();
errno = saved_errno;
return fdi;
}
void citp_log_change_fd(void)
{
int newfd, prev;
/* We need to change logging fd, probably because someone wants to do a
** dup2() onto it.
**
** No need to set 'close-on-exec' (FD_CLOEXEC) again for the newfd as
** it will be copied by the dup().
*/
CITP_FDTABLE_LOCK();
prev = citp.log_fd;
newfd = oo_fcntl_dupfd_cloexec(prev, 3);
if( newfd >= 0 ) {
__citp_fdtable_reserve(newfd, 1);
citp.log_fd = newfd;
}
Log_S(log("%s: old=%d new=%d", __FUNCTION__, prev, newfd));
__citp_fdtable_reserve(prev, 0);
ci_sys_close(prev);
CITP_FDTABLE_UNLOCK();
}
/* 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_close(unsigned fd)
{
volatile citp_fdinfo_p* p_fdip;
citp_fdinfo_p fdip;
int rc, got_lock;
citp_fdinfo* fdi;
/* Do not touch shared fdtable when in vfork child. */
if( oo_per_thread_get()->in_vfork_child )
return ci_tcp_helper_close_no_trampoline(fd);
/* Interlock against other closes, against the fdtable being extended,
** and against select and poll.
*/
CITP_FDTABLE_LOCK();
got_lock = 1;
__citp_fdtable_extend(fd);
if( fd >= citp_fdtable.inited_count ) {
rc = ci_sys_close(fd);
goto done;
}
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_closing(fdip) | fdip_is_reserved(fdip) ) {
/* Concurrent close or attempt to close reserved. */
Log_V(ci_log("%s: fd=%d closing=%d reserved=%d", __FUNCTION__, fd,
fdip_is_closing(fdip), fdip_is_reserved(fdip)));
errno = EBADF;
rc = -1;
goto done;
}
#if CI_CFG_FD_CACHING
/* Need to check in case this sucker's cached */
if( fdip_is_unknown(fdip) ) {
fdi = citp_fdtable_probe_locked(fd, CI_FALSE, CI_FALSE);
if( fdi == &citp_the_closed_fd ) {
citp_fdinfo_release_ref(fdi, CI_TRUE);
errno = EBADF;
rc = -1;
goto done;
}
if( fdi )
citp_fdinfo_release_ref(fdi, CI_TRUE);
}
#endif
ci_assert(fdip_is_normal(fdip) | fdip_is_passthru(fdip) |
fdip_is_unknown(fdip));
/* Swap in the "closed" pseudo-fdinfo. This lets any other thread know
** that we're in the middle of closing this fd.
*/
if( fdip_cas_fail(p_fdip, fdip, fdip_closing) )
goto again;
if( fdip_is_normal(fdip) ) {
fdi = fdip_to_fdi(fdip);
CITP_FDTABLE_UNLOCK();
got_lock = 0;
if( fdi->is_special ) {
Log_V(ci_log("%s: fd=%d is_special, returning EBADF", __FUNCTION__, fd));
errno = EBADF;
rc = -1;
fdtable_swap(fd, fdip_closing, fdip, 0);
goto done;
}
Log_V(ci_log("%s: fd=%d u/l socket", __FUNCTION__, fd));
ci_assert_equal(fdi->fd, fd);
ci_assert_equal(fdi->on_ref_count_zero, FDI_ON_RCZ_NONE);
fdi->on_ref_count_zero = FDI_ON_RCZ_CLOSE;
if( fdi->epoll_fd >= 0 ) {
citp_fdinfo* epoll_fdi = citp_epoll_fdi_from_member(fdi, 0);
if( epoll_fdi ) {
if( epoll_fdi->protocol->type == CITP_EPOLL_FD )
citp_epoll_on_close(epoll_fdi, fdi, 0);
citp_fdinfo_release_ref(epoll_fdi, 0);
}
}
citp_fdinfo_release_ref(fdi, 0);
rc = 0;
}
else {
ci_assert(fdip_is_passthru(fdip) ||
fdip_is_unknown(fdip));
if( ! fdtable_strict() ) {
CITP_FDTABLE_UNLOCK();
got_lock = 0;
}
Log_V(ci_log("%s: fd=%d passthru=%d unknown=%d", __FUNCTION__, fd,
fdip_is_passthru(fdip), fdip_is_unknown(fdip)));
fdtable_swap(fd, fdip_closing, fdip_unknown, fdtable_strict());
rc = ci_tcp_helper_close_no_trampoline(fd);
}
done:
if( got_lock ) CITP_FDTABLE_UNLOCK();
FDTABLE_ASSERT_VALID();
return rc;
}
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;
}
/* we don't register protocol impl */
int citp_pipe_create(int fds[2], int flags)
{
citp_pipe_fdi* epi_read;
citp_pipe_fdi* epi_write;
struct oo_pipe* p = NULL; /* make compiler happy */
ci_netif* ni;
int rc = -1;
ef_driver_handle fd = -1;
Log_V(log(LPF "pipe()"));
/* citp_netif_exists() does not need citp_ul_lock here */
if( CITP_OPTS.ul_pipe == CI_UNIX_PIPE_ACCELERATE_IF_NETIF &&
! citp_netif_exists() ) {
return CITP_NOT_HANDLED;
}
rc = citp_netif_alloc_and_init(&fd, &ni);
if( rc != 0 ) {
if( rc == CI_SOCKET_HANDOVER ) {
/* This implies EF_DONT_ACCELERATE is set, so we handover
* regardless of CITP_OPTS.no_fail */
return CITP_NOT_HANDLED;
}
/* may be lib mismatch - errno will be ELIBACC */
goto fail1;
}
rc = -1;
CI_MAGIC_CHECK(ni, NETIF_MAGIC);
/* add another reference as we have 2 fdis */
citp_netif_add_ref(ni);
epi_read = citp_pipe_epi_alloc(ni, O_RDONLY);
if( epi_read == NULL )
goto fail2;
epi_write = citp_pipe_epi_alloc(ni, O_WRONLY);
if( epi_write == NULL )
goto fail3;
/* oo_pipe init code */
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
rc = oo_pipe_ctor(ni, &p, fds, flags);
if( rc < 0 )
goto fail4;
citp_fdtable_new_fd_set(fds[0], fdip_busy, fdtable_strict());
citp_fdtable_new_fd_set(fds[1], fdip_busy, fdtable_strict());
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
LOG_PIPE("%s: pipe=%p id=%d", __FUNCTION__, p, p->b.bufid);
/* as pipe is created it should be attached to the end-points */
epi_read->pipe = p;
epi_write->pipe = p;
/* We're ready. Unleash us onto the world! */
ci_assert(epi_read->pipe->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_assert(epi_write->pipe->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi_read->pipe->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi_read->pipe->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
citp_fdtable_insert(&epi_read->fdinfo, fds[0], 0);
citp_fdtable_insert(&epi_write->fdinfo, fds[1], 0);
CI_MAGIC_CHECK(ni, NETIF_MAGIC);
return 0;
fail4:
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
fail3:
CI_FREE_OBJ(epi_write);
fail2:
CI_FREE_OBJ(epi_read);
citp_netif_release_ref(ni, 0);
citp_netif_release_ref(ni, 0);
fail1:
if( CITP_OPTS.no_fail && errno != ELIBACC ) {
Log_U(ci_log("%s: failed (errno:%d) - PASSING TO OS", __FUNCTION__, errno));
return CITP_NOT_HANDLED;
}
return rc;
}
/*
** 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;
}
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;
}
static int citp_udp_socket(int domain, int type, int protocol)
{
citp_fdinfo* fdi;
citp_sock_fdi* epi;
ef_driver_handle fd;
int rc;
ci_netif* ni;
Log_V(log(LPF "socket(%d, %d, %d)", domain, type, protocol));
epi = CI_ALLOC_OBJ(citp_sock_fdi);
if( ! epi ) {
Log_U(ci_log(LPF "socket: failed to allocate epi"));
errno = ENOMEM;
goto fail1;
}
fdi = &epi->fdinfo;
citp_fdinfo_init(fdi, &citp_udp_protocol_impl);
rc = citp_netif_alloc_and_init(&fd, &ni);
if( rc != 0 ) {
if( rc == CI_SOCKET_HANDOVER ) {
/* This implies EF_DONT_ACCELERATE is set, so we handover
* regardless of CITP_OPTS.no_fail */
CI_FREE_OBJ(epi);
return rc;
}
goto fail2;
}
/* Protect the fdtable entry until we're done initialising. */
if( fdtable_strict() ) CITP_FDTABLE_LOCK();
if((fd = ci_udp_ep_ctor(&epi->sock, ni, domain, type)) < 0) {
/*! ?? \TODO unpick the ci_udp_ep_ctor according to how failed */
Log_U(ci_log(LPF "socket: udp_ep_ctor failed"));
errno = -fd;
goto fail3;
}
citp_fdtable_new_fd_set(fd, fdip_busy, fdtable_strict());
if( fdtable_strict() ) CITP_FDTABLE_UNLOCK();
CI_DEBUG(epi->sock.s->pid = getpid());
/* We're ready. Unleash us onto the world! */
ci_assert(epi->sock.s->b.sb_aflags & CI_SB_AFLAG_NOT_READY);
ci_atomic32_and(&epi->sock.s->b.sb_aflags, ~CI_SB_AFLAG_NOT_READY);
citp_fdtable_insert(fdi, fd, 0);
Log_VSS(log(LPF "socket(%d, %d, %d) = "EF_FMT, domain, type, protocol,
EF_PRI_ARGS(epi,fd)));
return fd;
fail3:
if( CITP_OPTS.no_fail && errno != ELIBACC )
CITP_STATS_NETIF(++ni->state->stats.udp_handover_socket);
citp_netif_release_ref(ni, 0);
fail2:
CI_FREE_OBJ(epi);
fail1:
/* BUG1408: Graceful failure. We'll only fail outright if there's a
* driver/library mismatch */
if( CITP_OPTS.no_fail && errno != ELIBACC ) {
Log_U(ci_log("%s: failed (errno:%d) - PASSING TO OS", __FUNCTION__, errno));
return CI_SOCKET_HANDOVER;
}
return -1;
}
