/*
* dup2 is not thread safe
*/
int dup2(int oldfd, int newfd)
{
struct fd_info *oldfdi, *newfdi;
int ret;
init_preload();
oldfdi = idm_lookup(&idm, oldfd);
if (oldfdi) {
if (oldfdi->state == fd_fork_passive)
fork_passive(oldfd);
else if (oldfdi->state == fd_fork_active)
fork_active(oldfd);
}
newfdi = idm_lookup(&idm, newfd);
if (newfdi) {
/* newfd cannot have been dup'ed directly */
if (atomic_get(&newfdi->refcnt) > 1)
return ERR(EBUSY);
close(newfd);
}
ret = real.dup2(oldfd, newfd);
if (!oldfdi || ret != newfd)
return ret;
newfdi = calloc(1, sizeof *newfdi);
if (!newfdi) {
close(newfd);
return ERR(ENOMEM);
}
pthread_mutex_lock(&mut);
idm_set(&idm, newfd, newfdi);
pthread_mutex_unlock(&mut);
newfdi->fd = oldfdi->fd;
newfdi->type = oldfdi->type;
if (oldfdi->dupfd != -1) {
newfdi->dupfd = oldfdi->dupfd;
oldfdi = idm_lookup(&idm, oldfdi->dupfd);
} else {
newfdi->dupfd = oldfd;
}
atomic_init(&newfdi->refcnt);
atomic_set(&newfdi->refcnt, 1);
atomic_inc(&oldfdi->refcnt);
return newfd;
}
static inline enum fd_type fd_gett(int index)
{
struct fd_info *fdi;
fdi = idm_lookup(&idm, index);
return fdi ? fdi->type : fd_normal;
}
static inline enum fd_fork_state fd_gets(int index)
{
struct fd_info *fdi;
fdi = idm_lookup(&idm, index);
return fdi ? fdi->state : fd_ready;
}
static inline int fd_getd(int index)
{
struct fd_info *fdi;
fdi = idm_lookup(&idm, index);
return fdi ? fdi->fd : index;
}
struct sock_conn *sock_av_lookup_addr(struct sock_ep *ep,
struct sock_av *av,
fi_addr_t addr)
{
int idx, ret;
int index = ((uint64_t)addr & av->mask);
struct sock_av_addr *av_addr;
if (index >= av->table_hdr->stored || index < 0) {
SOCK_LOG_ERROR("requested rank is larger than av table\n");
errno = EINVAL;
return NULL;
}
if (!av->cmap) {
SOCK_LOG_ERROR("EP with no AV bound\n");
errno = EINVAL;
return NULL;
}
av_addr = idm_lookup(&av->addr_idm, index);
idx = av_addr - &av->table[0];
if (!av->key[idx]) {
ret = sock_conn_map_match_or_connect(
ep, av->domain, av->cmap,
(struct sockaddr_in*)&av_addr->addr,
&av->key[idx]);
if (ret) {
SOCK_LOG_ERROR("failed to match or connect to addr %"
PRIu64 "\n", addr);
return NULL;
}
}
return sock_conn_map_lookup_key(av->cmap, av->key[idx]);
}
static int sock_av_close(struct fid *fid)
{
struct sock_av *av;
int i;
av = container_of(fid, struct sock_av, av_fid.fid);
if (atomic_get(&av->ref))
return -FI_EBUSY;
for (i=0; i<av->table_hdr->stored; i++) {
if(idm_lookup(&av->addr_idm, i))
idm_clear(&av->addr_idm , i);
}
if (!av->name)
free(av->table_hdr);
else {
shm_unlink(av->name);
free(av->name);
munmap(av->table_hdr, sizeof(struct sock_av_table_hdr) +
av->attr.count * sizeof(struct sock_av_addr));
close(av->shared_fd);
}
atomic_dec(&av->domain->ref);
free(av->key);
free(av);
return 0;
}
int close(int socket)
{
struct fd_info *fdi;
int ret;
init_preload();
fdi = idm_lookup(&idm, socket);
if (!fdi)
return real.close(socket);
if (fdi->dupfd != -1) {
ret = close(fdi->dupfd);
if (ret)
return ret;
}
if (atomic_dec(&fdi->refcnt))
return 0;
idm_clear(&idm, socket);
real.close(socket);
ret = (fdi->type == fd_rsocket) ? rclose(fdi->fd) : real.close(fdi->fd);
free(fdi);
return ret;
}
static int sock_regattr(struct fid_domain *domain, const struct fi_mr_attr *attr,
uint64_t flags, struct fid_mr **mr)
{
struct fi_eq_entry eq_entry;
struct sock_domain *dom;
struct sock_mr *_mr;
uint64_t key;
dom = container_of(domain, struct sock_domain, dom_fid);
if (!(dom->info.mode & FI_PROV_MR_ATTR) &&
((attr->requested_key > IDX_MAX_INDEX) ||
idm_lookup(&dom->mr_idm, (int) attr->requested_key)))
return -FI_ENOKEY;
_mr = calloc(1, sizeof(*_mr) + sizeof(_mr->mr_iov) * (attr->iov_count - 1));
if (!_mr)
return -FI_ENOMEM;
_mr->mr_fid.fid.fclass = FI_CLASS_MR;
_mr->mr_fid.fid.context = attr->context;
_mr->mr_fid.fid.ops = &sock_mr_fi_ops;
_mr->domain = dom;
_mr->access = attr->access;
_mr->flags = flags;
_mr->offset = (flags & FI_MR_OFFSET) ?
(uintptr_t) attr->mr_iov[0].iov_base + attr->offset :
(uintptr_t) attr->mr_iov[0].iov_base;
fastlock_acquire(&dom->lock);
key = (dom->info.mode & FI_PROV_MR_ATTR) ?
sock_get_mr_key(dom) : (uint16_t) attr->requested_key;
if (idm_set(&dom->mr_idm, key, _mr) < 0)
goto err;
_mr->mr_fid.key = key;
_mr->mr_fid.mem_desc = (void *)key;
fastlock_release(&dom->lock);
_mr->iov_count = attr->iov_count;
memcpy(&_mr->mr_iov, attr->mr_iov, sizeof(_mr->mr_iov) * attr->iov_count);
*mr = &_mr->mr_fid;
atomic_inc(&dom->ref);
if (dom->mr_eq) {
eq_entry.fid = &domain->fid;
eq_entry.context = attr->context;
return sock_eq_report_event(dom->mr_eq, FI_MR_COMPLETE,
&eq_entry, sizeof(eq_entry), 0);
}
return 0;
err:
fastlock_release(&dom->lock);
free(_mr);
return -errno;
}
static uint16_t sock_get_mr_key(struct sock_domain *dom)
{
uint16_t i;
for (i = 1; i < IDX_MAX_INDEX; i++) {
if (!idm_lookup(&dom->mr_idm, i))
return i;
}
return 0;
}
int sock_av_compare_addr(struct sock_av *av,
fi_addr_t addr1, fi_addr_t addr2)
{
int index1, index2;
struct sock_av_addr *av_addr1, *av_addr2;
index1 = ((uint64_t)addr1 & av->mask);
index2 = ((uint64_t)addr2 & av->mask);
if (index1 >= av->table_hdr->stored || index1 < 0 ||
index2 >= av->table_hdr->stored || index2 < 0) {
SOCK_LOG_ERROR("requested rank is larger than av table\n");
return -1;
}
av_addr1 = idm_lookup(&av->addr_idm, index1);
av_addr2 = idm_lookup(&av->addr_idm, index2);
return memcmp(&av_addr1->addr, &av_addr2->addr,
sizeof(struct sockaddr_in));
}
static inline enum fd_type fd_get(int index, int *fd)
{
struct fd_info *fdi;
fdi = idm_lookup(&idm, index);
if (fdi) {
*fd = fdi->fd;
return fdi->type;
} else {
*fd = index;
return fd_normal;
}
}
static inline enum fd_type fd_fork_get(int index, int *fd)
{
struct fd_info *fdi;
fdi = idm_lookup(&idm, index);
if (fdi) {
if (fdi->state == fd_fork_passive)
fork_passive(index);
else if (fdi->state == fd_fork_active)
fork_active(index);
*fd = fdi->fd;
return fdi->type;
} else {
*fd = index;
return fd_normal;
}
}
static enum fd_type fd_close(int index, int *fd)
{
struct fd_info *fdi;
enum fd_type type;
fdi = idm_lookup(&idm, index);
if (fdi) {
idm_clear(&idm, index);
*fd = fdi->fd;
type = fdi->type;
real.close(index);
free(fdi);
} else {
*fd = index;
type = fd_normal;
}
return type;
}
static int sock_av_lookup(struct fid_av *av, fi_addr_t fi_addr, void *addr,
size_t *addrlen)
{
int index;
struct sock_av *_av;
struct sock_av_addr *av_addr;
_av = container_of(av, struct sock_av, av_fid);
index = ((uint64_t)fi_addr & _av->mask);
if (index >= _av->table_hdr->stored || index < 0) {
SOCK_LOG_ERROR("requested address not inserted\n");
return -EINVAL;
}
av_addr = idm_lookup(&_av->addr_idm, index);
memcpy(addr, &av_addr->addr, MIN(*addrlen, _av->addrlen));
*addrlen = _av->addrlen;
return 0;
}
struct sock_mr *sock_mr_verify_key(struct sock_domain *domain, uint16_t key,
void *buf, size_t len, uint64_t access)
{
int i;
struct sock_mr *mr;
mr = idm_lookup(&domain->mr_idm, key);
if (!mr)
return NULL;
if (mr->flags & FI_MR_OFFSET)
buf = (char*)buf + mr->offset;
for (i = 0; i < mr->iov_count; i++) {
if ((uintptr_t)buf >= (uintptr_t)mr->mr_iov[i].iov_base &&
((uintptr_t)buf + len <= (uintptr_t) mr->mr_iov[i].iov_base +
mr->mr_iov[i].iov_len)) {
if ((access & mr->access) == access)
return mr;
}
}
SOCK_LOG_ERROR("MR check failed\n");
return NULL;
}
struct sock_conn *sock_ep_connect(struct sock_ep *ep, fi_addr_t index)
{
int conn_fd = -1, ret;
int do_retry = sock_conn_retry;
struct sock_conn *conn, *new_conn;
uint16_t idx;
struct sockaddr_in *addr;
socklen_t lon;
int valopt = 0;
struct pollfd poll_fd;
if (ep->ep_type == FI_EP_MSG) {
idx = 0;
addr = ep->dest_addr;
} else {
idx = index & ep->av->mask;
addr = (struct sockaddr_in *)&ep->av->table[idx].addr;
}
do_connect:
fastlock_acquire(&ep->cmap.lock);
conn = sock_ep_lookup_conn(ep, index, addr);
fastlock_release(&ep->cmap.lock);
if (conn != SOCK_CM_CONN_IN_PROGRESS)
return conn;
conn_fd = socket(AF_INET, SOCK_STREAM, 0);
if (conn_fd == -1) {
SOCK_LOG_ERROR("failed to create conn_fd, errno: %d\n", errno);
errno = FI_EOTHER;
return NULL;
}
ret = fd_set_nonblock(conn_fd);
if (ret) {
SOCK_LOG_ERROR("failed to set conn_fd nonblocking, errno: %d\n", errno);
errno = FI_EOTHER;
close(conn_fd);
return NULL;
}
SOCK_LOG_DBG("Connecting to: %s:%d\n", inet_ntoa(addr->sin_addr),
ntohs(addr->sin_port));
SOCK_LOG_DBG("Connecting using address:%s\n",
inet_ntoa(ep->src_addr->sin_addr));
ret = connect(conn_fd, (struct sockaddr *) addr, sizeof *addr);
if (ret < 0) {
if (errno == EINPROGRESS) {
poll_fd.fd = conn_fd;
poll_fd.events = POLLOUT;
ret = poll(&poll_fd, 1, 15 * 1000);
if (ret < 0) {
SOCK_LOG_DBG("poll failed\n");
goto retry;
}
lon = sizeof(int);
ret = getsockopt(conn_fd, SOL_SOCKET, SO_ERROR, (void*)(&valopt), &lon);
if (ret < 0) {
SOCK_LOG_DBG("getsockopt failed: %d, %d\n", ret, conn_fd);
goto retry;
}
if (valopt) {
SOCK_LOG_DBG("Error in connection() %d - %s - %d\n", valopt, strerror(valopt), conn_fd);
SOCK_LOG_DBG("Connecting to: %s:%d\n", inet_ntoa(addr->sin_addr),
ntohs(addr->sin_port));
SOCK_LOG_DBG("Connecting using address:%s\n",
inet_ntoa(ep->src_addr->sin_addr));
goto retry;
}
goto out;
} else {
SOCK_LOG_DBG("Timeout or error() - %s: %d\n", strerror(errno), conn_fd);
SOCK_LOG_DBG("Connecting to: %s:%d\n", inet_ntoa(addr->sin_addr),
ntohs(addr->sin_port));
SOCK_LOG_DBG("Connecting using address:%s\n",
inet_ntoa(ep->src_addr->sin_addr));
goto retry;
}
} else {
goto out;
}
retry:
do_retry--;
sleep(10);
if (!do_retry)
goto err;
if (conn_fd != -1) {
close(conn_fd);
conn_fd = -1;
}
SOCK_LOG_ERROR("Connect error, retrying - %s - %d\n", strerror(errno), conn_fd);
SOCK_LOG_DBG("Connecting to: %s:%d\n", inet_ntoa(addr->sin_addr),
ntohs(addr->sin_port));
SOCK_LOG_DBG("Connecting using address:%s\n",
inet_ntoa(ep->src_addr->sin_addr));
goto do_connect;
out:
fastlock_acquire(&ep->cmap.lock);
new_conn = sock_conn_map_insert(ep, addr, conn_fd, 0);
new_conn->av_index = (ep->ep_type == FI_EP_MSG) ? FI_ADDR_NOTAVAIL : (fi_addr_t) idx;
conn = idm_lookup(&ep->av_idm, index);
if (conn == SOCK_CM_CONN_IN_PROGRESS) {
idm_set(&ep->av_idm, index, new_conn);
conn = new_conn;
}
fastlock_release(&ep->cmap.lock);
return conn;
err:
close(conn_fd);
return NULL;
}
struct sock_mr * sock_mr_get_entry(struct sock_domain *domain, uint16_t key)
{
return (struct sock_mr *)idm_lookup(&domain->mr_idm, key);
}