# Lustre is a trademark of Sun Microsystems, Inc.
#
-SUBDIRS = lnet klnds ulnds selftest doc utils include \
+SUBDIRS = lnet klnds selftest doc utils include \
autoconf
-DIST_SUBDIRS = lnet klnds ulnds selftest doc utils include \
+DIST_SUBDIRS = lnet klnds selftest doc utils include \
autoconf
]) # LN_CONFIG_BACKOFF
#
-# LN_CONFIG_USOCKLND
-#
-# configure support for userspace TCP/IP LND
-#
-AC_DEFUN([LN_CONFIG_USOCKLND], [
-AC_MSG_CHECKING([whether to enable usocklnd])
-AC_ARG_ENABLE([usocklnd],
- AC_HELP_STRING([--disable-usocklnd],
- [disable usocklnd]),
- [], [enable_usocklnd="yes"])
-AC_MSG_RESULT([$enable_usocklnd])
-USOCKLND=""
-AS_IF([test "x$enable_usocklnd" = xyes], [
- AC_MSG_CHECKING([whether to build usocklnd])
- AS_IF([test "$ENABLE_LIBPTHREAD" = yes], [
- AC_MSG_RESULT([yes])
- USOCKLND="usocklnd"
- ], [
- AC_MSG_RESULT([no (libpthread not present or disabled)])
- ])
-])
-AC_SUBST(USOCKLND)
-]) # LN_CONFIG_USOCKLND
-
-#
# LN_CONFIG_DLC
#
# Configure dlc if enabled
AC_SUBST(LIBWRAP)
LN_CONFIG_MAX_PAYLOAD
-LN_CONFIG_USOCKLND
LN_CONFIG_DLC
]) # LN_CONFIGURE
AM_CONDITIONAL(BUILD_RALND, test x$RALND = "xralnd")
AM_CONDITIONAL(BUILD_GNILND, test x$GNILND = "xgnilnd")
AM_CONDITIONAL(BUILD_GNILND_RCA, test x$GNILNDRCA = "xgnilndrca")
-AM_CONDITIONAL(BUILD_USOCKLND, test x$USOCKLND = "xusocklnd")
AM_CONDITIONAL(BUILD_DLC, test x$USE_DLC = "xyes")
]) # LN_CONDITIONALS
lnet/lnet/autoMakefile
lnet/selftest/Makefile
lnet/selftest/autoMakefile
-lnet/ulnds/Makefile
-lnet/ulnds/autoMakefile
-lnet/ulnds/socklnd/Makefile
lnet/utils/Makefile
lnet/utils/lnetconfig/Makefile
])
+++ /dev/null
-@BUILD_USOCKLND_TRUE@subdir-m += socklnd
-
-@INCLUDE_RULES@
-
+++ /dev/null
-#
-# GPL HEADER START
-#
-# DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
-#
-# This program is free software; you can redistribute it and/or modify
-# it under the terms of the GNU General Public License version 2 only,
-# as published by the Free Software Foundation.
-#
-# This program is distributed in the hope that it will be useful, but
-# WITHOUT ANY WARRANTY; without even the implied warranty of
-# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
-# General Public License version 2 for more details (a copy is included
-# in the LICENSE file that accompanied this code).
-#
-# You should have received a copy of the GNU General Public License
-# version 2 along with this program; If not, see
-# http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
-#
-# Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
-# CA 95054 USA or visit www.sun.com if you need additional information or
-# have any questions.
-#
-# GPL HEADER END
-#
-
-#
-# Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
-# Use is subject to license terms.
-#
-
-#
-# This file is part of Lustre, http://www.lustre.org/
-# Lustre is a trademark of Sun Microsystems, Inc.
-#
-
-SUBDIRS = socklnd
+++ /dev/null
-/Makefile.in
+++ /dev/null
-if LIBLUSTRE
-if BUILD_USOCKLND
-noinst_LIBRARIES = libsocklnd.a
-endif
-endif
-
-noinst_HEADERS = usocklnd.h
-libsocklnd_a_SOURCES = usocklnd.h usocklnd.c usocklnd_cb.c poll.c \
- handlers.c conn.c
-libsocklnd_a_CPPFLAGS = $(LLCPPFLAGS)
-libsocklnd_a_CFLAGS = $(LLCFLAGS)
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- *
- * Copyright (c) 2012, Intel Corporation.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/conn.c
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-
-#include "usocklnd.h"
-
-/* Return 1 if the conn is timed out, 0 else */
-int
-usocklnd_conn_timed_out(usock_conn_t *conn, cfs_time_t current_time)
-{
- if (conn->uc_tx_flag && /* sending is in progress */
- cfs_time_aftereq(current_time, conn->uc_tx_deadline))
- return 1;
-
- if (conn->uc_rx_flag && /* receiving is in progress */
- cfs_time_aftereq(current_time, conn->uc_rx_deadline))
- return 1;
-
- return 0;
-}
-
-void
-usocklnd_conn_kill(usock_conn_t *conn)
-{
- pthread_mutex_lock(&conn->uc_lock);
- if (conn->uc_state != UC_DEAD)
- usocklnd_conn_kill_locked(conn);
- pthread_mutex_unlock(&conn->uc_lock);
-}
-
-/* Mark the conn as DEAD and schedule its deletion */
-void
-usocklnd_conn_kill_locked(usock_conn_t *conn)
-{
- conn->uc_rx_flag = conn->uc_tx_flag = 0;
- conn->uc_state = UC_DEAD;
- usocklnd_add_killrequest(conn);
-}
-
-usock_conn_t *
-usocklnd_conn_allocate()
-{
- usock_conn_t *conn;
- usock_pollrequest_t *pr;
-
- LIBCFS_ALLOC (pr, sizeof(*pr));
- if (pr == NULL)
- return NULL;
-
- LIBCFS_ALLOC (conn, sizeof(*conn));
- if (conn == NULL) {
- LIBCFS_FREE (pr, sizeof(*pr));
- return NULL;
- }
- memset(conn, 0, sizeof(*conn));
- conn->uc_preq = pr;
-
- LIBCFS_ALLOC (conn->uc_rx_hello,
- offsetof(ksock_hello_msg_t,
- kshm_ips[LNET_MAX_INTERFACES]));
- if (conn->uc_rx_hello == NULL) {
- LIBCFS_FREE (pr, sizeof(*pr));
- LIBCFS_FREE (conn, sizeof(*conn));
- return NULL;
- }
-
- return conn;
-}
-
-void
-usocklnd_conn_free(usock_conn_t *conn)
-{
- usock_pollrequest_t *pr = conn->uc_preq;
-
- if (pr != NULL)
- LIBCFS_FREE (pr, sizeof(*pr));
-
- if (conn->uc_rx_hello != NULL)
- LIBCFS_FREE (conn->uc_rx_hello,
- offsetof(ksock_hello_msg_t,
- kshm_ips[LNET_MAX_INTERFACES]));
-
- LIBCFS_FREE (conn, sizeof(*conn));
-}
-
-void
-usocklnd_tear_peer_conn(usock_conn_t *conn)
-{
- usock_peer_t *peer = conn->uc_peer;
- int idx = usocklnd_type2idx(conn->uc_type);
- lnet_ni_t *ni;
- lnet_process_id_t id;
- int decref_flag = 0;
- int killall_flag = 0;
- void *rx_lnetmsg = NULL;
- struct list_head zombie_txs = LIST_HEAD_INIT(zombie_txs);
-
- if (peer == NULL) /* nothing to tear */
- return;
-
- pthread_mutex_lock(&peer->up_lock);
- pthread_mutex_lock(&conn->uc_lock);
-
- ni = peer->up_ni;
- id = peer->up_peerid;
-
- if (peer->up_conns[idx] == conn) {
- if (conn->uc_rx_state == UC_RX_LNET_PAYLOAD) {
- /* change state not to finalize twice */
- conn->uc_rx_state = UC_RX_KSM_HEADER;
- /* stash lnetmsg while holding locks */
- rx_lnetmsg = conn->uc_rx_lnetmsg;
- }
-
- /* we cannot finilize txs right now (bug #18844) */
- list_splice_init(&conn->uc_tx_list, &zombie_txs);
-
- peer->up_conns[idx] = NULL;
- conn->uc_peer = NULL;
- decref_flag = 1;
-
- if(conn->uc_errored && !peer->up_errored)
- peer->up_errored = killall_flag = 1;
-
- /* prevent queueing new txs to this conn */
- conn->uc_errored = 1;
- }
-
- pthread_mutex_unlock(&conn->uc_lock);
-
- if (killall_flag)
- usocklnd_del_conns_locked(peer);
-
- pthread_mutex_unlock(&peer->up_lock);
-
- if (!decref_flag)
- return;
-
- if (rx_lnetmsg != NULL)
- lnet_finalize(ni, rx_lnetmsg, -EIO);
-
- usocklnd_destroy_txlist(ni, &zombie_txs);
-
- usocklnd_conn_decref(conn);
- usocklnd_peer_decref(peer);
-
- usocklnd_check_peer_stale(ni, id);
-}
-
-/* Remove peer from hash list if all up_conns[i] is NULL &&
- * hash table is the only consumer of the peer */
-void
-usocklnd_check_peer_stale(lnet_ni_t *ni, lnet_process_id_t id)
-{
- usock_peer_t *peer;
-
- pthread_rwlock_wrlock(&usock_data.ud_peers_lock);
- peer = usocklnd_find_peer_locked(ni, id);
-
- if (peer == NULL) {
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
- return;
- }
-
- if (mt_atomic_read(&peer->up_refcount) == 2) {
- int i;
- for (i = 0; i < N_CONN_TYPES; i++)
- LASSERT (peer->up_conns[i] == NULL);
-
- list_del(&peer->up_list);
-
- if (peer->up_errored &&
- (peer->up_peerid.pid & LNET_PID_USERFLAG) == 0)
- lnet_notify (peer->up_ni, peer->up_peerid.nid, 0,
- cfs_time_seconds(peer->up_last_alive));
-
- usocklnd_peer_decref(peer);
- }
-
- usocklnd_peer_decref(peer);
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
-}
-
-/* Returns 0 on success, <0 else */
-int
-usocklnd_create_passive_conn(lnet_ni_t *ni,
- cfs_socket_t *sock, usock_conn_t **connp)
-{
- int rc;
- __u32 peer_ip;
- int peer_port;
- usock_conn_t *conn;
-
- rc = libcfs_sock_getaddr(sock, 1, &peer_ip, &peer_port);
- if (rc)
- return rc;
-
- LASSERT (peer_port >= 0); /* uc_peer_port is u16 */
-
- rc = usocklnd_set_sock_options(sock);
- if (rc)
- return rc;
-
- conn = usocklnd_conn_allocate();
- if (conn == NULL)
- return -ENOMEM;
-
- usocklnd_rx_hellomagic_state_transition(conn);
-
- conn->uc_sock = sock;
- conn->uc_peer_ip = peer_ip;
- conn->uc_peer_port = peer_port;
- conn->uc_state = UC_RECEIVING_HELLO;
- conn->uc_pt_idx = usocklnd_ip2pt_idx(peer_ip);
- conn->uc_ni = ni;
- INIT_LIST_HEAD(&conn->uc_tx_list);
- INIT_LIST_HEAD(&conn->uc_zcack_list);
- pthread_mutex_init(&conn->uc_lock, NULL);
- mt_atomic_set(&conn->uc_refcount, 1); /* 1 ref for me */
-
- *connp = conn;
- return 0;
-}
-
-/* Returns 0 on success, <0 else */
-int
-usocklnd_create_active_conn(usock_peer_t *peer, int type,
- usock_conn_t **connp)
-{
- int rc;
- cfs_socket_t *sock;
- usock_conn_t *conn;
- __u32 dst_ip = LNET_NIDADDR(peer->up_peerid.nid);
- __u16 dst_port = lnet_acceptor_port();
-
- conn = usocklnd_conn_allocate();
- if (conn == NULL)
- return -ENOMEM;
-
- conn->uc_tx_hello = usocklnd_create_cr_hello_tx(peer->up_ni, type,
- peer->up_peerid.nid);
- if (conn->uc_tx_hello == NULL) {
- usocklnd_conn_free(conn);
- return -ENOMEM;
- }
-
- if (the_lnet.ln_pid & LNET_PID_USERFLAG)
- rc = usocklnd_connect_cli_mode(&sock, dst_ip, dst_port);
- else
- rc = usocklnd_connect_srv_mode(&sock, dst_ip, dst_port);
-
- if (rc) {
- usocklnd_destroy_tx(NULL, conn->uc_tx_hello);
- usocklnd_conn_free(conn);
- return rc;
- }
-
- conn->uc_tx_deadline = cfs_time_shift(usock_tuns.ut_timeout);
- conn->uc_tx_flag = 1;
-
- conn->uc_sock = sock;
- conn->uc_peer_ip = dst_ip;
- conn->uc_peer_port = dst_port;
- conn->uc_type = type;
- conn->uc_activeflag = 1;
- conn->uc_state = UC_CONNECTING;
- conn->uc_pt_idx = usocklnd_ip2pt_idx(dst_ip);
- conn->uc_ni = NULL;
- conn->uc_peerid = peer->up_peerid;
- conn->uc_peer = peer;
-
- usocklnd_peer_addref(peer);
- INIT_LIST_HEAD(&conn->uc_tx_list);
- INIT_LIST_HEAD(&conn->uc_zcack_list);
- pthread_mutex_init(&conn->uc_lock, NULL);
- mt_atomic_set(&conn->uc_refcount, 1); /* 1 ref for me */
-
- *connp = conn;
- return 0;
-}
-
-/* Returns 0 on success, <0 else */
-int
-usocklnd_connect_srv_mode(cfs_socket_t **sockp, __u32 dst_ip, __u16 dst_port)
-{
- __u16 port;
- cfs_socket_t *sock;
- int rc;
- int fatal;
-
- for (port = LNET_ACCEPTOR_MAX_RESERVED_PORT;
- port >= LNET_ACCEPTOR_MIN_RESERVED_PORT;
- port--) {
- /* Iterate through reserved ports. */
- rc = libcfs_sock_create(&sock, &fatal, 0, port);
- if (rc) {
- if (fatal)
- return rc;
- continue;
- }
-
- rc = usocklnd_set_sock_options(sock);
- if (rc) {
- libcfs_sock_release(sock);
- return rc;
- }
-
- rc = libcfs_sock_connect(sock, dst_ip, dst_port);
- if (rc == 0) {
- *sockp = sock;
- return 0;
- }
-
- if (rc != -EADDRINUSE && rc != -EADDRNOTAVAIL) {
- libcfs_sock_release(sock);
- return rc;
- }
-
- libcfs_sock_release(sock);
- }
-
- CERROR("Can't bind to any reserved port\n");
- return rc;
-}
-
-/* Returns 0 on success, <0 else */
-int
-usocklnd_connect_cli_mode(cfs_socket_t **sockp, __u32 dst_ip, __u16 dst_port)
-{
- cfs_socket_t *sock;
- int rc;
- int fatal;
-
- rc = libcfs_sock_create(&sock, &fatal, 0, 0);
- if (rc)
- return rc;
-
- rc = usocklnd_set_sock_options(sock);
- if (rc) {
- libcfs_sock_release(sock);
- return rc;
- }
-
- rc = libcfs_sock_connect(sock, dst_ip, dst_port);
- if (rc) {
- libcfs_sock_release(sock);
- return rc;
- }
-
- *sockp = sock;
- return 0;
-}
-
-int
-usocklnd_set_sock_options(cfs_socket_t *sock)
-{
- int rc;
-
- rc = libcfs_sock_set_nagle(sock, usock_tuns.ut_socknagle);
- if (rc)
- return rc;
-
- if (usock_tuns.ut_sockbufsiz) {
- rc = libcfs_sock_set_bufsiz(sock, usock_tuns.ut_sockbufsiz);
- if (rc)
- return rc;
- }
-
- return libcfs_fcntl_nonblock(sock);
-}
-
-usock_tx_t *
-usocklnd_create_noop_tx(__u64 cookie)
-{
- usock_tx_t *tx;
-
- LIBCFS_ALLOC (tx, sizeof(usock_tx_t));
- if (tx == NULL)
- return NULL;
-
- tx->tx_size = sizeof(usock_tx_t);
- tx->tx_lnetmsg = NULL;
-
- socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_NOOP);
- tx->tx_msg.ksm_zc_cookies[1] = cookie;
-
- tx->tx_iova[0].iov_base = (void *)&tx->tx_msg;
- tx->tx_iova[0].iov_len = tx->tx_resid = tx->tx_nob =
- offsetof(ksock_msg_t, ksm_u.lnetmsg.ksnm_hdr);
- tx->tx_iov = tx->tx_iova;
- tx->tx_niov = 1;
-
- return tx;
-}
-
-usock_tx_t *
-usocklnd_create_tx(lnet_msg_t *lntmsg)
-{
- usock_tx_t *tx;
- unsigned int payload_niov = lntmsg->msg_niov;
- struct iovec *payload_iov = lntmsg->msg_iov;
- unsigned int payload_offset = lntmsg->msg_offset;
- unsigned int payload_nob = lntmsg->msg_len;
- int size = offsetof(usock_tx_t,
- tx_iova[1 + payload_niov]);
-
- LIBCFS_ALLOC (tx, size);
- if (tx == NULL)
- return NULL;
-
- tx->tx_size = size;
- tx->tx_lnetmsg = lntmsg;
-
- tx->tx_resid = tx->tx_nob = sizeof(ksock_msg_t) + payload_nob;
-
- socklnd_init_msg(&tx->tx_msg, KSOCK_MSG_LNET);
- tx->tx_msg.ksm_u.lnetmsg.ksnm_hdr = lntmsg->msg_hdr;
- tx->tx_iova[0].iov_base = (void *)&tx->tx_msg;
- tx->tx_iova[0].iov_len = sizeof(ksock_msg_t);
- tx->tx_iov = tx->tx_iova;
-
- tx->tx_niov = 1 +
- lnet_extract_iov(payload_niov, &tx->tx_iov[1],
- payload_niov, payload_iov,
- payload_offset, payload_nob);
-
- return tx;
-}
-
-void
-usocklnd_init_hello_msg(ksock_hello_msg_t *hello,
- lnet_ni_t *ni, int type, lnet_nid_t peer_nid)
-{
- usock_net_t *net = (usock_net_t *)ni->ni_data;
-
- hello->kshm_magic = LNET_PROTO_MAGIC;
- hello->kshm_version = KSOCK_PROTO_V2;
- hello->kshm_nips = 0;
- hello->kshm_ctype = type;
-
- hello->kshm_dst_incarnation = 0; /* not used */
- hello->kshm_src_incarnation = net->un_incarnation;
-
- hello->kshm_src_pid = the_lnet.ln_pid;
- hello->kshm_src_nid = ni->ni_nid;
- hello->kshm_dst_nid = peer_nid;
- hello->kshm_dst_pid = 0; /* not used */
-}
-
-usock_tx_t *
-usocklnd_create_hello_tx(lnet_ni_t *ni,
- int type, lnet_nid_t peer_nid)
-{
- usock_tx_t *tx;
- int size;
- ksock_hello_msg_t *hello;
-
- size = sizeof(usock_tx_t) + offsetof(ksock_hello_msg_t, kshm_ips);
- LIBCFS_ALLOC (tx, size);
- if (tx == NULL)
- return NULL;
-
- tx->tx_size = size;
- tx->tx_lnetmsg = NULL;
-
- hello = (ksock_hello_msg_t *)&tx->tx_iova[1];
- usocklnd_init_hello_msg(hello, ni, type, peer_nid);
-
- tx->tx_iova[0].iov_base = (void *)hello;
- tx->tx_iova[0].iov_len = tx->tx_resid = tx->tx_nob =
- offsetof(ksock_hello_msg_t, kshm_ips);
- tx->tx_iov = tx->tx_iova;
- tx->tx_niov = 1;
-
- return tx;
-}
-
-usock_tx_t *
-usocklnd_create_cr_hello_tx(lnet_ni_t *ni,
- int type, lnet_nid_t peer_nid)
-{
- usock_tx_t *tx;
- int size;
- lnet_acceptor_connreq_t *cr;
- ksock_hello_msg_t *hello;
-
- size = sizeof(usock_tx_t) +
- sizeof(lnet_acceptor_connreq_t) +
- offsetof(ksock_hello_msg_t, kshm_ips);
- LIBCFS_ALLOC (tx, size);
- if (tx == NULL)
- return NULL;
-
- tx->tx_size = size;
- tx->tx_lnetmsg = NULL;
-
- cr = (lnet_acceptor_connreq_t *)&tx->tx_iova[1];
- memset(cr, 0, sizeof(*cr));
- cr->acr_magic = LNET_PROTO_ACCEPTOR_MAGIC;
- cr->acr_version = LNET_PROTO_ACCEPTOR_VERSION;
- cr->acr_nid = peer_nid;
-
- hello = (ksock_hello_msg_t *)((char *)cr + sizeof(*cr));
- usocklnd_init_hello_msg(hello, ni, type, peer_nid);
-
- tx->tx_iova[0].iov_base = (void *)cr;
- tx->tx_iova[0].iov_len = tx->tx_resid = tx->tx_nob =
- sizeof(lnet_acceptor_connreq_t) +
- offsetof(ksock_hello_msg_t, kshm_ips);
- tx->tx_iov = tx->tx_iova;
- tx->tx_niov = 1;
-
- return tx;
-}
-
-void
-usocklnd_destroy_tx(lnet_ni_t *ni, usock_tx_t *tx)
-{
- lnet_msg_t *lnetmsg = tx->tx_lnetmsg;
- int rc = (tx->tx_resid == 0) ? 0 : -EIO;
-
- LASSERT (ni != NULL || lnetmsg == NULL);
-
- LIBCFS_FREE (tx, tx->tx_size);
-
- if (lnetmsg != NULL) /* NOOP and hello go without lnetmsg */
- lnet_finalize(ni, lnetmsg, rc);
-}
-
-void
-usocklnd_destroy_txlist(lnet_ni_t *ni, struct list_head *txlist)
-{
- usock_tx_t *tx;
-
- while (!list_empty(txlist)) {
- tx = list_entry(txlist->next, usock_tx_t, tx_list);
- list_del(&tx->tx_list);
-
- usocklnd_destroy_tx(ni, tx);
- }
-}
-
-void
-usocklnd_destroy_zcack_list(struct list_head *zcack_list)
-{
- usock_zc_ack_t *zcack;
-
- while (!list_empty(zcack_list)) {
- zcack = list_entry(zcack_list->next, usock_zc_ack_t,
- zc_list);
- list_del(&zcack->zc_list);
-
- LIBCFS_FREE (zcack, sizeof(*zcack));
- }
-}
-
-void
-usocklnd_destroy_peer(usock_peer_t *peer)
-{
- usock_net_t *net = peer->up_ni->ni_data;
- int i;
-
- for (i = 0; i < N_CONN_TYPES; i++)
- LASSERT (peer->up_conns[i] == NULL);
-
- LIBCFS_FREE (peer, sizeof (*peer));
-
- pthread_mutex_lock(&net->un_lock);
- if(--net->un_peercount == 0)
- pthread_cond_signal(&net->un_cond);
- pthread_mutex_unlock(&net->un_lock);
-}
-
-void
-usocklnd_destroy_conn(usock_conn_t *conn)
-{
- LASSERT (conn->uc_peer == NULL || conn->uc_ni == NULL);
-
- if (conn->uc_rx_state == UC_RX_LNET_PAYLOAD) {
- LASSERT (conn->uc_peer != NULL);
- lnet_finalize(conn->uc_peer->up_ni, conn->uc_rx_lnetmsg, -EIO);
- }
-
- if (!list_empty(&conn->uc_tx_list)) {
- LASSERT (conn->uc_peer != NULL);
- usocklnd_destroy_txlist(conn->uc_peer->up_ni, &conn->uc_tx_list);
- }
-
- usocklnd_destroy_zcack_list(&conn->uc_zcack_list);
-
- if (conn->uc_peer != NULL)
- usocklnd_peer_decref(conn->uc_peer);
-
- if (conn->uc_ni != NULL)
- lnet_ni_decref(conn->uc_ni);
-
- if (conn->uc_tx_hello)
- usocklnd_destroy_tx(NULL, conn->uc_tx_hello);
-
- usocklnd_conn_free(conn);
-}
-
-int
-usocklnd_get_conn_type(lnet_msg_t *lntmsg)
-{
- int nob;
-
- if (the_lnet.ln_pid & LNET_PID_USERFLAG)
- return SOCKLND_CONN_ANY;
-
- nob = sizeof(ksock_msg_t) + lntmsg->msg_len;
-
- if (nob >= usock_tuns.ut_min_bulk)
- return SOCKLND_CONN_BULK_OUT;
- else
- return SOCKLND_CONN_CONTROL;
-}
-
-int usocklnd_type2idx(int type)
-{
- switch (type) {
- case SOCKLND_CONN_ANY:
- case SOCKLND_CONN_CONTROL:
- return 0;
- case SOCKLND_CONN_BULK_IN:
- return 1;
- case SOCKLND_CONN_BULK_OUT:
- return 2;
- default:
- LBUG();
- }
-}
-
-usock_peer_t *
-usocklnd_find_peer_locked(lnet_ni_t *ni, lnet_process_id_t id)
-{
- struct list_head *peer_list = usocklnd_nid2peerlist(id.nid);
- struct list_head *tmp;
- usock_peer_t *peer;
-
- list_for_each(tmp, peer_list) {
-
- peer = list_entry(tmp, usock_peer_t, up_list);
-
- if (peer->up_ni != ni)
- continue;
-
- if (peer->up_peerid.nid != id.nid ||
- peer->up_peerid.pid != id.pid)
- continue;
-
- usocklnd_peer_addref(peer);
- return peer;
- }
- return (NULL);
-}
-
-int
-usocklnd_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
- usock_peer_t **peerp)
-{
- usock_net_t *net = ni->ni_data;
- usock_peer_t *peer;
- int i;
-
- LIBCFS_ALLOC (peer, sizeof (*peer));
- if (peer == NULL)
- return -ENOMEM;
-
- for (i = 0; i < N_CONN_TYPES; i++)
- peer->up_conns[i] = NULL;
-
- peer->up_peerid = id;
- peer->up_ni = ni;
- peer->up_incrn_is_set = 0;
- peer->up_errored = 0;
- peer->up_last_alive = 0;
- mt_atomic_set(&peer->up_refcount, 1); /* 1 ref for caller */
- pthread_mutex_init(&peer->up_lock, NULL);
-
- pthread_mutex_lock(&net->un_lock);
- net->un_peercount++;
- pthread_mutex_unlock(&net->un_lock);
-
- *peerp = peer;
- return 0;
-}
-
-/* Safely create new peer if needed. Save result in *peerp.
- * Returns 0 on success, <0 else */
-int
-usocklnd_find_or_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
- usock_peer_t **peerp)
-{
- int rc;
- usock_peer_t *peer;
- usock_peer_t *peer2;
- usock_net_t *net = ni->ni_data;
-
- pthread_rwlock_rdlock(&usock_data.ud_peers_lock);
- peer = usocklnd_find_peer_locked(ni, id);
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
-
- if (peer != NULL)
- goto find_or_create_peer_done;
-
- rc = usocklnd_create_peer(ni, id, &peer);
- if (rc)
- return rc;
-
- pthread_rwlock_wrlock(&usock_data.ud_peers_lock);
- peer2 = usocklnd_find_peer_locked(ni, id);
- if (peer2 == NULL) {
- if (net->un_shutdown) {
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
- usocklnd_peer_decref(peer); /* should destroy peer */
- CERROR("Can't create peer: network shutdown\n");
- return -ESHUTDOWN;
- }
-
- /* peer table will take 1 of my refs on peer */
- usocklnd_peer_addref(peer);
- list_add_tail(&peer->up_list,
- usocklnd_nid2peerlist(id.nid));
- } else {
- usocklnd_peer_decref(peer); /* should destroy peer */
- peer = peer2;
- }
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
-
- find_or_create_peer_done:
- *peerp = peer;
- return 0;
-}
-
-/* NB: both peer and conn locks are held */
-static int
-usocklnd_enqueue_zcack(usock_conn_t *conn, usock_zc_ack_t *zc_ack)
-{
- if (conn->uc_state == UC_READY &&
- list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list) &&
- !conn->uc_sending) {
- int rc = usocklnd_add_pollrequest(conn, POLL_TX_SET_REQUEST,
- POLLOUT);
- if (rc != 0)
- return rc;
- }
-
- list_add_tail(&zc_ack->zc_list, &conn->uc_zcack_list);
- return 0;
-}
-
-/* NB: both peer and conn locks are held
- * NB: if sending isn't in progress. the caller *MUST* send tx
- * immediately after we'll return */
-static void
-usocklnd_enqueue_tx(usock_conn_t *conn, usock_tx_t *tx,
- int *send_immediately)
-{
- if (conn->uc_state == UC_READY &&
- list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list) &&
- !conn->uc_sending) {
- conn->uc_sending = 1;
- *send_immediately = 1;
- return;
- }
-
- *send_immediately = 0;
- list_add_tail(&tx->tx_list, &conn->uc_tx_list);
-}
-
-/* Safely create new conn if needed. Save result in *connp.
- * Returns 0 on success, <0 else */
-int
-usocklnd_find_or_create_conn(usock_peer_t *peer, int type,
- usock_conn_t **connp,
- usock_tx_t *tx, usock_zc_ack_t *zc_ack,
- int *send_immediately)
-{
- usock_conn_t *conn;
- int idx;
- int rc;
- lnet_pid_t userflag = peer->up_peerid.pid & LNET_PID_USERFLAG;
-
- if (userflag)
- type = SOCKLND_CONN_ANY;
-
- idx = usocklnd_type2idx(type);
-
- pthread_mutex_lock(&peer->up_lock);
- if (peer->up_conns[idx] != NULL) {
- conn = peer->up_conns[idx];
- LASSERT(conn->uc_type == type);
- } else {
- if (userflag) {
- CERROR("Refusing to create a connection to "
- "userspace process %s\n",
- libcfs_id2str(peer->up_peerid));
- rc = -EHOSTUNREACH;
- goto find_or_create_conn_failed;
- }
-
- rc = usocklnd_create_active_conn(peer, type, &conn);
- if (rc) {
- peer->up_errored = 1;
- usocklnd_del_conns_locked(peer);
- goto find_or_create_conn_failed;
- }
-
- /* peer takes 1 of conn refcount */
- usocklnd_link_conn_to_peer(conn, peer, idx);
-
- rc = usocklnd_add_pollrequest(conn, POLL_ADD_REQUEST, POLLOUT);
- if (rc) {
- peer->up_conns[idx] = NULL;
- usocklnd_conn_decref(conn); /* should destroy conn */
- goto find_or_create_conn_failed;
- }
- usocklnd_wakeup_pollthread(conn->uc_pt_idx);
- }
-
- pthread_mutex_lock(&conn->uc_lock);
- LASSERT(conn->uc_peer == peer);
-
- LASSERT(tx == NULL || zc_ack == NULL);
- if (tx != NULL) {
- /* usocklnd_tear_peer_conn() could signal us stop queueing */
- if (conn->uc_errored) {
- rc = -EIO;
- pthread_mutex_unlock(&conn->uc_lock);
- goto find_or_create_conn_failed;
- }
-
- usocklnd_enqueue_tx(conn, tx, send_immediately);
- } else {
- rc = usocklnd_enqueue_zcack(conn, zc_ack);
- if (rc != 0) {
- usocklnd_conn_kill_locked(conn);
- pthread_mutex_unlock(&conn->uc_lock);
- goto find_or_create_conn_failed;
- }
- }
- pthread_mutex_unlock(&conn->uc_lock);
-
- usocklnd_conn_addref(conn);
- pthread_mutex_unlock(&peer->up_lock);
-
- *connp = conn;
- return 0;
-
- find_or_create_conn_failed:
- pthread_mutex_unlock(&peer->up_lock);
- return rc;
-}
-
-void
-usocklnd_link_conn_to_peer(usock_conn_t *conn, usock_peer_t *peer, int idx)
-{
- peer->up_conns[idx] = conn;
- peer->up_errored = 0; /* this new fresh conn will try
- * revitalize even stale errored peer */
-}
-
-int
-usocklnd_invert_type(int type)
-{
- switch (type)
- {
- case SOCKLND_CONN_ANY:
- case SOCKLND_CONN_CONTROL:
- return (type);
- case SOCKLND_CONN_BULK_IN:
- return SOCKLND_CONN_BULK_OUT;
- case SOCKLND_CONN_BULK_OUT:
- return SOCKLND_CONN_BULK_IN;
- default:
- return SOCKLND_CONN_NONE;
- }
-}
-
-void
-usocklnd_conn_new_state(usock_conn_t *conn, int new_state)
-{
- pthread_mutex_lock(&conn->uc_lock);
- if (conn->uc_state != UC_DEAD)
- conn->uc_state = new_state;
- pthread_mutex_unlock(&conn->uc_lock);
-}
-
-/* NB: peer is locked by caller */
-void
-usocklnd_cleanup_stale_conns(usock_peer_t *peer, __u64 incrn,
- usock_conn_t *skip_conn)
-{
- int i;
-
- if (!peer->up_incrn_is_set) {
- peer->up_incarnation = incrn;
- peer->up_incrn_is_set = 1;
- return;
- }
-
- if (peer->up_incarnation == incrn)
- return;
-
- peer->up_incarnation = incrn;
-
- for (i = 0; i < N_CONN_TYPES; i++) {
- usock_conn_t *conn = peer->up_conns[i];
-
- if (conn == NULL || conn == skip_conn)
- continue;
-
- pthread_mutex_lock(&conn->uc_lock);
- LASSERT (conn->uc_peer == peer);
- conn->uc_peer = NULL;
- peer->up_conns[i] = NULL;
- if (conn->uc_state != UC_DEAD)
- usocklnd_conn_kill_locked(conn);
- pthread_mutex_unlock(&conn->uc_lock);
-
- usocklnd_conn_decref(conn);
- usocklnd_peer_decref(peer);
- }
-}
-
-/* RX state transition to UC_RX_HELLO_MAGIC: update RX part to receive
- * MAGIC part of hello and set uc_rx_state
- */
-void
-usocklnd_rx_hellomagic_state_transition(usock_conn_t *conn)
-{
- LASSERT(conn->uc_rx_hello != NULL);
-
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_hello->kshm_magic;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- sizeof(conn->uc_rx_hello->kshm_magic);
-
- conn->uc_rx_state = UC_RX_HELLO_MAGIC;
-
- conn->uc_rx_flag = 1; /* waiting for incoming hello */
- conn->uc_rx_deadline = cfs_time_shift(usock_tuns.ut_timeout);
-}
-
-/* RX state transition to UC_RX_HELLO_VERSION: update RX part to receive
- * VERSION part of hello and set uc_rx_state
- */
-void
-usocklnd_rx_helloversion_state_transition(usock_conn_t *conn)
-{
- LASSERT(conn->uc_rx_hello != NULL);
-
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_hello->kshm_version;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- sizeof(conn->uc_rx_hello->kshm_version);
-
- conn->uc_rx_state = UC_RX_HELLO_VERSION;
-}
-
-/* RX state transition to UC_RX_HELLO_BODY: update RX part to receive
- * the rest of hello and set uc_rx_state
- */
-void
-usocklnd_rx_hellobody_state_transition(usock_conn_t *conn)
-{
- LASSERT(conn->uc_rx_hello != NULL);
-
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_hello->kshm_src_nid;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- offsetof(ksock_hello_msg_t, kshm_ips) -
- offsetof(ksock_hello_msg_t, kshm_src_nid);
-
- conn->uc_rx_state = UC_RX_HELLO_BODY;
-}
-
-/* RX state transition to UC_RX_HELLO_IPS: update RX part to receive
- * array of IPs and set uc_rx_state
- */
-void
-usocklnd_rx_helloIPs_state_transition(usock_conn_t *conn)
-{
- LASSERT(conn->uc_rx_hello != NULL);
-
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_hello->kshm_ips;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- conn->uc_rx_hello->kshm_nips *
- sizeof(conn->uc_rx_hello->kshm_ips[0]);
-
- conn->uc_rx_state = UC_RX_HELLO_IPS;
-}
-
-/* RX state transition to UC_RX_LNET_HEADER: update RX part to receive
- * LNET header and set uc_rx_state
- */
-void
-usocklnd_rx_lnethdr_state_transition(usock_conn_t *conn)
-{
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_msg.ksm_u.lnetmsg;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- sizeof(ksock_lnet_msg_t);
-
- conn->uc_rx_state = UC_RX_LNET_HEADER;
- conn->uc_rx_flag = 1;
-}
-
-/* RX state transition to UC_RX_KSM_HEADER: update RX part to receive
- * KSM header and set uc_rx_state
- */
-void
-usocklnd_rx_ksmhdr_state_transition(usock_conn_t *conn)
-{
- conn->uc_rx_niov = 1;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_iov[0].iov_base = &conn->uc_rx_msg;
- conn->uc_rx_iov[0].iov_len =
- conn->uc_rx_nob_wanted =
- conn->uc_rx_nob_left =
- offsetof(ksock_msg_t, ksm_u);
-
- conn->uc_rx_state = UC_RX_KSM_HEADER;
- conn->uc_rx_flag = 0;
-}
-
-/* RX state transition to UC_RX_SKIPPING: update RX part for
- * skipping and set uc_rx_state
- */
-void
-usocklnd_rx_skipping_state_transition(usock_conn_t *conn)
-{
- static char skip_buffer[4096];
-
- int nob;
- unsigned int niov = 0;
- int skipped = 0;
- int nob_to_skip = conn->uc_rx_nob_left;
-
- LASSERT(nob_to_skip != 0);
-
- conn->uc_rx_iov = conn->uc_rx_iova;
-
- /* Set up to skip as much as possible now. If there's more left
- * (ran out of iov entries) we'll get called again */
-
- do {
- nob = MIN (nob_to_skip, sizeof(skip_buffer));
-
- conn->uc_rx_iov[niov].iov_base = skip_buffer;
- conn->uc_rx_iov[niov].iov_len = nob;
- niov++;
- skipped += nob;
- nob_to_skip -=nob;
-
- } while (nob_to_skip != 0 && /* mustn't overflow conn's rx iov */
- niov < sizeof(conn->uc_rx_iova) / sizeof (struct iovec));
-
- conn->uc_rx_niov = niov;
- conn->uc_rx_nob_wanted = skipped;
-
- conn->uc_rx_state = UC_RX_SKIPPING;
-}
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/handlers.c
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-
-#include "usocklnd.h"
-#include <unistd.h>
-#include <sys/syscall.h>
-
-int
-usocklnd_notifier_handler(int fd)
-{
- int notification;
- return syscall(SYS_read, fd, ¬ification, sizeof(notification));
-}
-
-void
-usocklnd_exception_handler(usock_conn_t *conn)
-{
- pthread_mutex_lock(&conn->uc_lock);
-
- if (conn->uc_state == UC_CONNECTING ||
- conn->uc_state == UC_SENDING_HELLO)
- usocklnd_conn_kill_locked(conn);
-
- pthread_mutex_unlock(&conn->uc_lock);
-}
-
-int
-usocklnd_read_handler(usock_conn_t *conn)
-{
- int rc;
- int continue_reading;
- int state;
-
- read_again:
- rc = 0;
- pthread_mutex_lock(&conn->uc_lock);
- state = conn->uc_state;
-
- /* process special case: LNET calls lnd_recv() asyncronously */
- if (state == UC_READY && conn->uc_rx_state == UC_RX_PARSE) {
- /* still don't have usocklnd_recv() called */
- rc = usocklnd_add_pollrequest(conn, POLL_RX_SET_REQUEST, 0);
- if (rc == 0)
- conn->uc_rx_state = UC_RX_PARSE_WAIT;
- else
- usocklnd_conn_kill_locked(conn);
-
- pthread_mutex_unlock(&conn->uc_lock);
- return rc;
- }
-
- pthread_mutex_unlock(&conn->uc_lock);
- /* From here and below the conn cannot be changed
- * asyncronously, except:
- * 1) usocklnd_send() can work with uc_tx_list and uc_zcack_list,
- * 2) usocklnd_shutdown() can change uc_state to UC_DEAD */
-
- switch (state) {
-
- case UC_RECEIVING_HELLO:
- case UC_READY:
- if (conn->uc_rx_nob_wanted != 0) {
- /* read from conn fd as much wanted data as possible */
- rc = usocklnd_read_data(conn);
- if (rc == 0) /* partial read */
- break;
- if (rc < 0) {/* error happened or EOF */
- usocklnd_conn_kill(conn);
- break;
- }
- }
-
- /* process incoming data */
- if (state == UC_READY )
- rc = usocklnd_read_msg(conn, &continue_reading);
- else /* state == UC_RECEIVING_HELLO */
- rc = usocklnd_read_hello(conn, &continue_reading);
-
- if (rc < 0) {
- usocklnd_conn_kill(conn);
- break;
- }
-
- if (continue_reading)
- goto read_again;
-
- break;
-
- case UC_DEAD:
- break;
-
- default:
- LBUG();
- }
-
- return rc;
-}
-
-/* Switch on rx_state.
- * Return 0 on success, 1 if whole packet is read, else return <0
- * Always set cont_flag: 1 if we're ready to continue reading, else 0
- * NB: If whole packet is read, cont_flag will be set to zero to take
- * care of fairess
- */
-int
-usocklnd_read_msg(usock_conn_t *conn, int *cont_flag)
-{
- int rc = 0;
- __u64 cookie;
-
- *cont_flag = 0;
-
- /* smth. new emerged in RX part - let's process it */
- switch (conn->uc_rx_state) {
- case UC_RX_KSM_HEADER:
- if (conn->uc_flip) {
- __swab32s(&conn->uc_rx_msg.ksm_type);
- __swab32s(&conn->uc_rx_msg.ksm_csum);
- __swab64s(&conn->uc_rx_msg.ksm_zc_cookies[0]);
- __swab64s(&conn->uc_rx_msg.ksm_zc_cookies[1]);
- }
-
- /* we never send packets for wich zc-acking is required */
- if (conn->uc_rx_msg.ksm_type != KSOCK_MSG_LNET ||
- conn->uc_rx_msg.ksm_zc_cookies[1] != 0) {
- conn->uc_errored = 1;
- return -EPROTO;
- }
-
- /* zc_req will be processed later, when
- lnet payload will be received */
-
- usocklnd_rx_lnethdr_state_transition(conn);
- *cont_flag = 1;
- break;
-
- case UC_RX_LNET_HEADER:
- if (the_lnet.ln_pid & LNET_PID_USERFLAG) {
- /* replace dest_nid,pid (ksocknal sets its own) */
- conn->uc_rx_msg.ksm_u.lnetmsg.ksnm_hdr.dest_nid =
- cpu_to_le64(conn->uc_peer->up_ni->ni_nid);
- conn->uc_rx_msg.ksm_u.lnetmsg.ksnm_hdr.dest_pid =
- cpu_to_le32(the_lnet.ln_pid);
-
- } else if (conn->uc_peer->up_peerid.pid & LNET_PID_USERFLAG) {
- /* Userspace peer */
- lnet_process_id_t *id = &conn->uc_peer->up_peerid;
- lnet_hdr_t *lhdr = &conn->uc_rx_msg.ksm_u.lnetmsg.ksnm_hdr;
-
- /* Substitute process ID assigned at connection time */
- lhdr->src_pid = cpu_to_le32(id->pid);
- lhdr->src_nid = cpu_to_le64(id->nid);
- }
-
- conn->uc_rx_state = UC_RX_PARSE;
- usocklnd_conn_addref(conn); /* ++ref while parsing */
-
- rc = lnet_parse(conn->uc_peer->up_ni,
- &conn->uc_rx_msg.ksm_u.lnetmsg.ksnm_hdr,
- conn->uc_peerid.nid, conn, 0);
-
- if (rc < 0) {
- /* I just received garbage: give up on this conn */
- conn->uc_errored = 1;
- usocklnd_conn_decref(conn);
- return -EPROTO;
- }
-
- /* Race with usocklnd_recv() is possible */
- pthread_mutex_lock(&conn->uc_lock);
- LASSERT (conn->uc_rx_state == UC_RX_PARSE ||
- conn->uc_rx_state == UC_RX_LNET_PAYLOAD);
-
- /* check whether usocklnd_recv() got called */
- if (conn->uc_rx_state == UC_RX_LNET_PAYLOAD)
- *cont_flag = 1;
- pthread_mutex_unlock(&conn->uc_lock);
- break;
-
- case UC_RX_PARSE:
- LBUG(); /* it's error to be here, because this special
- * case is handled by caller */
- break;
-
- case UC_RX_PARSE_WAIT:
- LBUG(); /* it's error to be here, because the conn
- * shouldn't wait for POLLIN event in this
- * state */
- break;
-
- case UC_RX_LNET_PAYLOAD:
- /* payload all received */
-
- lnet_finalize(conn->uc_peer->up_ni, conn->uc_rx_lnetmsg, 0);
-
- cookie = conn->uc_rx_msg.ksm_zc_cookies[0];
- if (cookie != 0)
- rc = usocklnd_handle_zc_req(conn->uc_peer, cookie);
-
- if (rc != 0) {
- /* change state not to finalize twice */
- conn->uc_rx_state = UC_RX_KSM_HEADER;
- return -EPROTO;
- }
-
- /* Fall through */
-
- case UC_RX_SKIPPING:
- if (conn->uc_rx_nob_left != 0) {
- usocklnd_rx_skipping_state_transition(conn);
- *cont_flag = 1;
- } else {
- usocklnd_rx_ksmhdr_state_transition(conn);
- rc = 1; /* whole packet is read */
- }
-
- break;
-
- default:
- LBUG(); /* unknown state */
- }
-
- return rc;
-}
-
-/* Handle incoming ZC request from sender.
- * NB: it's called only from read_handler, so we're sure that
- * the conn cannot become zombie in the middle of processing */
-int
-usocklnd_handle_zc_req(usock_peer_t *peer, __u64 cookie)
-{
- usock_conn_t *conn;
- usock_zc_ack_t *zc_ack;
- int type;
- int rc;
- int dummy;
-
- LIBCFS_ALLOC (zc_ack, sizeof(*zc_ack));
- if (zc_ack == NULL)
- return -ENOMEM;
- zc_ack->zc_cookie = cookie;
-
- /* Let's assume that CONTROL is the best type for zcack,
- * but userspace clients don't use typed connections */
- if (the_lnet.ln_pid & LNET_PID_USERFLAG)
- type = SOCKLND_CONN_ANY;
- else
- type = SOCKLND_CONN_CONTROL;
-
- rc = usocklnd_find_or_create_conn(peer, type, &conn, NULL, zc_ack,
- &dummy);
- if (rc != 0) {
- LIBCFS_FREE (zc_ack, sizeof(*zc_ack));
- return rc;
- }
- usocklnd_conn_decref(conn);
-
- return 0;
-}
-
-/* Switch on rx_state.
- * Return 0 on success, else return <0
- * Always set cont_flag: 1 if we're ready to continue reading, else 0
- */
-int
-usocklnd_read_hello(usock_conn_t *conn, int *cont_flag)
-{
- int rc = 0;
- ksock_hello_msg_t *hello = conn->uc_rx_hello;
-
- *cont_flag = 0;
-
- /* smth. new emerged in hello - let's process it */
- switch (conn->uc_rx_state) {
- case UC_RX_HELLO_MAGIC:
- if (hello->kshm_magic == LNET_PROTO_MAGIC)
- conn->uc_flip = 0;
- else if (hello->kshm_magic == __swab32(LNET_PROTO_MAGIC))
- conn->uc_flip = 1;
- else
- return -EPROTO;
-
- usocklnd_rx_helloversion_state_transition(conn);
- *cont_flag = 1;
- break;
-
- case UC_RX_HELLO_VERSION:
- if ((!conn->uc_flip &&
- (hello->kshm_version != KSOCK_PROTO_V2)) ||
- (conn->uc_flip &&
- (hello->kshm_version != __swab32(KSOCK_PROTO_V2))))
- return -EPROTO;
-
- usocklnd_rx_hellobody_state_transition(conn);
- *cont_flag = 1;
- break;
-
- case UC_RX_HELLO_BODY:
- if (conn->uc_flip) {
- ksock_hello_msg_t *hello = conn->uc_rx_hello;
- __swab32s(&hello->kshm_src_pid);
- __swab64s(&hello->kshm_src_nid);
- __swab32s(&hello->kshm_dst_pid);
- __swab64s(&hello->kshm_dst_nid);
- __swab64s(&hello->kshm_src_incarnation);
- __swab64s(&hello->kshm_dst_incarnation);
- __swab32s(&hello->kshm_ctype);
- __swab32s(&hello->kshm_nips);
- }
-
- if (conn->uc_rx_hello->kshm_nips > LNET_MAX_INTERFACES) {
- CERROR("Bad nips %d from ip %u.%u.%u.%u port %d\n",
- conn->uc_rx_hello->kshm_nips,
- HIPQUAD(conn->uc_peer_ip), conn->uc_peer_port);
- return -EPROTO;
- }
-
- if (conn->uc_rx_hello->kshm_nips) {
- usocklnd_rx_helloIPs_state_transition(conn);
- *cont_flag = 1;
- break;
- }
- /* fall through */
-
- case UC_RX_HELLO_IPS:
- if (conn->uc_activeflag == 1) /* active conn */
- rc = usocklnd_activeconn_hellorecv(conn);
- else /* passive conn */
- rc = usocklnd_passiveconn_hellorecv(conn);
-
- break;
-
- default:
- LBUG(); /* unknown state */
- }
-
- return rc;
-}
-
-/* All actions that we need after receiving hello on active conn:
- * 1) Schedule removing if we're zombie
- * 2) Restart active conn if we lost the race
- * 3) Else: update RX part to receive KSM header
- */
-int
-usocklnd_activeconn_hellorecv(usock_conn_t *conn)
-{
- int rc = 0;
- ksock_hello_msg_t *hello = conn->uc_rx_hello;
- usock_peer_t *peer = conn->uc_peer;
-
- /* Active conn with peer==NULL is zombie.
- * Don't try to link it to peer because the conn
- * has already had a chance to proceed at the beginning */
- if (peer == NULL) {
- LASSERT(list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list));
-
- usocklnd_conn_kill(conn);
- return 0;
- }
-
- peer->up_last_alive = cfs_time_current();
-
- /* peer says that we lost the race */
- if (hello->kshm_ctype == SOCKLND_CONN_NONE) {
- /* Start new active conn, relink txs and zc_acks from
- * the conn to new conn, schedule removing the conn.
- * Actually, we're expecting that a passive conn will
- * make us zombie soon and take care of our txs and
- * zc_acks */
-
- struct list_head tx_list, zcack_list;
- usock_conn_t *conn2;
- int idx = usocklnd_type2idx(conn->uc_type);
-
- INIT_LIST_HEAD(&tx_list);
- INIT_LIST_HEAD(&zcack_list);
-
- /* Block usocklnd_send() to check peer->up_conns[idx]
- * and to enqueue more txs */
- pthread_mutex_lock(&peer->up_lock);
- pthread_mutex_lock(&conn->uc_lock);
-
- /* usocklnd_shutdown() could kill us */
- if (conn->uc_state == UC_DEAD) {
- pthread_mutex_unlock(&conn->uc_lock);
- pthread_mutex_unlock(&peer->up_lock);
- return 0;
- }
-
- LASSERT (peer == conn->uc_peer);
- LASSERT (peer->up_conns[idx] == conn);
-
- rc = usocklnd_create_active_conn(peer, conn->uc_type, &conn2);
- if (rc) {
- conn->uc_errored = 1;
- pthread_mutex_unlock(&conn->uc_lock);
- pthread_mutex_unlock(&peer->up_lock);
- return rc;
- }
-
- usocklnd_link_conn_to_peer(conn2, peer, idx);
- conn2->uc_peer = peer;
-
- /* unlink txs and zcack from the conn */
- list_add(&tx_list, &conn->uc_tx_list);
- list_del_init(&conn->uc_tx_list);
- list_add(&zcack_list, &conn->uc_zcack_list);
- list_del_init(&conn->uc_zcack_list);
-
- /* link they to the conn2 */
- list_add(&conn2->uc_tx_list, &tx_list);
- list_del_init(&tx_list);
- list_add(&conn2->uc_zcack_list, &zcack_list);
- list_del_init(&zcack_list);
-
- /* make conn zombie */
- conn->uc_peer = NULL;
- usocklnd_peer_decref(peer);
-
- /* schedule conn2 for processing */
- rc = usocklnd_add_pollrequest(conn2, POLL_ADD_REQUEST, POLLOUT);
- if (rc) {
- peer->up_conns[idx] = NULL;
- usocklnd_conn_decref(conn2); /* should destroy conn */
- } else {
- usocklnd_conn_kill_locked(conn);
- }
-
- pthread_mutex_unlock(&conn->uc_lock);
- pthread_mutex_unlock(&peer->up_lock);
- usocklnd_conn_decref(conn);
-
- } else { /* hello->kshm_ctype != SOCKLND_CONN_NONE */
- if (conn->uc_type != usocklnd_invert_type(hello->kshm_ctype))
- return -EPROTO;
-
- pthread_mutex_lock(&peer->up_lock);
- usocklnd_cleanup_stale_conns(peer, hello->kshm_src_incarnation,
- conn);
- pthread_mutex_unlock(&peer->up_lock);
-
- /* safely transit to UC_READY state */
- /* rc == 0 */
- pthread_mutex_lock(&conn->uc_lock);
- if (conn->uc_state != UC_DEAD) {
- usocklnd_rx_ksmhdr_state_transition(conn);
-
- /* POLLIN is already set because we just
- * received hello, but maybe we've smth. to
- * send? */
- LASSERT (conn->uc_sending == 0);
- if (!list_empty(&conn->uc_tx_list) ||
- !list_empty(&conn->uc_zcack_list)) {
-
- conn->uc_tx_deadline =
- cfs_time_shift(usock_tuns.ut_timeout);
- conn->uc_tx_flag = 1;
- rc = usocklnd_add_pollrequest(conn,
- POLL_SET_REQUEST,
- POLLIN | POLLOUT);
- }
-
- if (rc == 0)
- conn->uc_state = UC_READY;
- }
- pthread_mutex_unlock(&conn->uc_lock);
- }
-
- return rc;
-}
-
-/* All actions that we need after receiving hello on passive conn:
- * 1) Stash peer's nid, pid, incarnation and conn type
- * 2) Cope with easy case: conn[idx] is empty - just save conn there
- * 3) Resolve race:
- * a) if our nid is higher - reply with CONN_NONE and make us zombie
- * b) if peer's nid is higher - postpone race resolution till
- * READY state
- * 4) Anyhow, send reply hello
-*/
-int
-usocklnd_passiveconn_hellorecv(usock_conn_t *conn)
-{
- ksock_hello_msg_t *hello = conn->uc_rx_hello;
- int type;
- int idx;
- int rc;
- usock_peer_t *peer;
- lnet_ni_t *ni = conn->uc_ni;
- __u32 peer_ip = conn->uc_peer_ip;
- __u16 peer_port = conn->uc_peer_port;
-
- /* don't know parent peer yet and not zombie */
- LASSERT (conn->uc_peer == NULL &&
- ni != NULL);
-
- /* don't know peer's nid and incarnation yet */
- if (peer_port > LNET_ACCEPTOR_MAX_RESERVED_PORT) {
- /* do not trust liblustre clients */
- conn->uc_peerid.pid = peer_port | LNET_PID_USERFLAG;
- conn->uc_peerid.nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid),
- peer_ip);
- if (hello->kshm_ctype != SOCKLND_CONN_ANY) {
- lnet_ni_decref(ni);
- conn->uc_ni = NULL;
- CERROR("Refusing to accept connection of type=%d from "
- "userspace process %u.%u.%u.%u:%d\n", hello->kshm_ctype,
- HIPQUAD(peer_ip), peer_port);
- return -EINVAL;
- }
- } else {
- conn->uc_peerid.pid = hello->kshm_src_pid;
- conn->uc_peerid.nid = hello->kshm_src_nid;
- }
- conn->uc_type = type = usocklnd_invert_type(hello->kshm_ctype);
-
- rc = usocklnd_find_or_create_peer(ni, conn->uc_peerid, &peer);
- if (rc) {
- lnet_ni_decref(ni);
- conn->uc_ni = NULL;
- return rc;
- }
-
- peer->up_last_alive = cfs_time_current();
-
- idx = usocklnd_type2idx(conn->uc_type);
-
- /* safely check whether we're first */
- pthread_mutex_lock(&peer->up_lock);
-
- usocklnd_cleanup_stale_conns(peer, hello->kshm_src_incarnation, NULL);
-
- if (peer->up_conns[idx] == NULL) {
- peer->up_last_alive = cfs_time_current();
- conn->uc_peer = peer;
- conn->uc_ni = NULL;
- usocklnd_link_conn_to_peer(conn, peer, idx);
- usocklnd_conn_addref(conn);
- } else {
- usocklnd_peer_decref(peer);
-
- /* Resolve race in favour of higher NID */
- if (conn->uc_peerid.nid < conn->uc_ni->ni_nid) {
- /* make us zombie */
- conn->uc_ni = NULL;
- type = SOCKLND_CONN_NONE;
- }
-
- /* if conn->uc_peerid.nid > conn->uc_ni->ni_nid,
- * postpone race resolution till READY state
- * (hopefully that conn[idx] will die because of
- * incoming hello of CONN_NONE type) */
- }
- pthread_mutex_unlock(&peer->up_lock);
-
- /* allocate and initialize fake tx with hello */
- conn->uc_tx_hello = usocklnd_create_hello_tx(ni, type,
- conn->uc_peerid.nid);
- if (conn->uc_ni == NULL)
- lnet_ni_decref(ni);
-
- if (conn->uc_tx_hello == NULL)
- return -ENOMEM;
-
- /* rc == 0 */
- pthread_mutex_lock(&conn->uc_lock);
- if (conn->uc_state == UC_DEAD)
- goto passive_hellorecv_done;
-
- conn->uc_state = UC_SENDING_HELLO;
- conn->uc_tx_deadline = cfs_time_shift(usock_tuns.ut_timeout);
- conn->uc_tx_flag = 1;
- rc = usocklnd_add_pollrequest(conn, POLL_SET_REQUEST, POLLOUT);
-
- passive_hellorecv_done:
- pthread_mutex_unlock(&conn->uc_lock);
- return rc;
-}
-
-int
-usocklnd_write_handler(usock_conn_t *conn)
-{
- usock_tx_t *tx;
- int ret;
- int rc = 0;
- int state;
- usock_peer_t *peer;
- lnet_ni_t *ni;
-
- pthread_mutex_lock(&conn->uc_lock); /* like membar */
- state = conn->uc_state;
- pthread_mutex_unlock(&conn->uc_lock);
-
- switch (state) {
- case UC_CONNECTING:
- /* hello_tx has already been initialized
- * in usocklnd_create_active_conn() */
- usocklnd_conn_new_state(conn, UC_SENDING_HELLO);
- /* fall through */
-
- case UC_SENDING_HELLO:
- rc = usocklnd_send_tx(conn, conn->uc_tx_hello);
- if (rc <= 0) /* error or partial send or connection closed */
- break;
-
- /* tx with hello was sent successfully */
- usocklnd_destroy_tx(NULL, conn->uc_tx_hello);
- conn->uc_tx_hello = NULL;
-
- if (conn->uc_activeflag == 1) /* active conn */
- rc = usocklnd_activeconn_hellosent(conn);
- else /* passive conn */
- rc = usocklnd_passiveconn_hellosent(conn);
-
- break;
-
- case UC_READY:
- pthread_mutex_lock(&conn->uc_lock);
-
- peer = conn->uc_peer;
- LASSERT (peer != NULL);
- ni = peer->up_ni;
-
- if (list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list)) {
- LASSERT(usock_tuns.ut_fair_limit > 1);
- pthread_mutex_unlock(&conn->uc_lock);
- return 0;
- }
-
- tx = usocklnd_try_piggyback(&conn->uc_tx_list,
- &conn->uc_zcack_list);
- if (tx != NULL)
- conn->uc_sending = 1;
- else
- rc = -ENOMEM;
-
- pthread_mutex_unlock(&conn->uc_lock);
-
- if (rc)
- break;
-
- rc = usocklnd_send_tx(conn, tx);
- if (rc == 0) { /* partial send or connection closed */
- pthread_mutex_lock(&conn->uc_lock);
- list_add(&tx->tx_list, &conn->uc_tx_list);
- conn->uc_sending = 0;
- pthread_mutex_unlock(&conn->uc_lock);
- break;
- }
- if (rc < 0) { /* real error */
- usocklnd_destroy_tx(ni, tx);
- break;
- }
-
- /* rc == 1: tx was sent completely */
- usocklnd_destroy_tx(ni, tx);
-
- pthread_mutex_lock(&conn->uc_lock);
- conn->uc_sending = 0;
- if (conn->uc_state != UC_DEAD &&
- list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list)) {
- conn->uc_tx_flag = 0;
- ret = usocklnd_add_pollrequest(conn,
- POLL_TX_SET_REQUEST, 0);
- if (ret)
- rc = ret;
- }
- pthread_mutex_unlock(&conn->uc_lock);
-
- break;
-
- case UC_DEAD:
- break;
-
- default:
- LBUG();
- }
-
- if (rc < 0)
- usocklnd_conn_kill(conn);
-
- return rc;
-}
-
-/* Return the first tx from tx_list with piggybacked zc_ack
- * from zcack_list when possible. If tx_list is empty, return
- * brand new noop tx for zc_ack from zcack_list. Return NULL
- * if an error happened */
-usock_tx_t *
-usocklnd_try_piggyback(struct list_head *tx_list_p,
- struct list_head *zcack_list_p)
-{
- usock_tx_t *tx;
- usock_zc_ack_t *zc_ack;
-
- /* assign tx and zc_ack */
- if (list_empty(tx_list_p))
- tx = NULL;
- else {
- tx = list_entry(tx_list_p->next, usock_tx_t, tx_list);
- list_del(&tx->tx_list);
-
- /* already piggybacked or partially send */
- if (tx->tx_msg.ksm_zc_cookies[1] != 0 ||
- tx->tx_resid != tx->tx_nob)
- return tx;
- }
-
- if (list_empty(zcack_list_p)) {
- /* nothing to piggyback */
- return tx;
- } else {
- zc_ack = list_entry(zcack_list_p->next,
- usock_zc_ack_t, zc_list);
- list_del(&zc_ack->zc_list);
- }
-
- if (tx != NULL)
- /* piggyback the zc-ack cookie */
- tx->tx_msg.ksm_zc_cookies[1] = zc_ack->zc_cookie;
- else
- /* cannot piggyback, need noop */
- tx = usocklnd_create_noop_tx(zc_ack->zc_cookie);
-
- LIBCFS_FREE (zc_ack, sizeof(*zc_ack));
- return tx;
-}
-
-/* All actions that we need after sending hello on active conn:
- * 1) update RX iov to receive hello
- * 2) state transition to UC_RECEIVING_HELLO
- * 3) notify poll_thread that we're waiting for incoming hello */
-int
-usocklnd_activeconn_hellosent(usock_conn_t *conn)
-{
- int rc = 0;
-
- pthread_mutex_lock(&conn->uc_lock);
-
- if (conn->uc_state != UC_DEAD) {
- usocklnd_rx_hellomagic_state_transition(conn);
- conn->uc_state = UC_RECEIVING_HELLO;
- conn->uc_tx_flag = 0;
- rc = usocklnd_add_pollrequest(conn, POLL_SET_REQUEST, POLLIN);
- }
-
- pthread_mutex_unlock(&conn->uc_lock);
-
- return rc;
-}
-
-/* All actions that we need after sending hello on passive conn:
- * 1) Cope with 1st easy case: conn is already linked to a peer
- * 2) Cope with 2nd easy case: remove zombie conn
- * 3) Resolve race:
- * a) find the peer
- * b) link the conn to the peer if conn[idx] is empty
- * c) if the conn[idx] isn't empty and is in READY state,
- * remove the conn as duplicated
- * d) if the conn[idx] isn't empty and isn't in READY state,
- * override conn[idx] with the conn
- */
-int
-usocklnd_passiveconn_hellosent(usock_conn_t *conn)
-{
- usock_conn_t *conn2;
- usock_peer_t *peer;
- struct list_head tx_list;
- struct list_head zcack_list;
- int idx;
- int rc = 0;
-
- /* almost nothing to do if conn is already linked to peer hash table */
- if (conn->uc_peer != NULL)
- goto passive_hellosent_done;
-
- /* conn->uc_peer == NULL, so the conn isn't accessible via
- * peer hash list, so nobody can touch the conn but us */
-
- if (conn->uc_ni == NULL) /* remove zombie conn */
- goto passive_hellosent_connkill;
-
- /* all code below is race resolution, because normally
- * passive conn is linked to peer just after receiving hello */
- INIT_LIST_HEAD(&tx_list);
- INIT_LIST_HEAD(&zcack_list);
-
- /* conn is passive and isn't linked to any peer,
- so its tx and zc_ack lists have to be empty */
- LASSERT(list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list) &&
- conn->uc_sending == 0);
-
- rc = usocklnd_find_or_create_peer(conn->uc_ni, conn->uc_peerid, &peer);
- if (rc)
- return rc;
-
- idx = usocklnd_type2idx(conn->uc_type);
-
- /* try to link conn to peer */
- pthread_mutex_lock(&peer->up_lock);
- if (peer->up_conns[idx] == NULL) {
- usocklnd_link_conn_to_peer(conn, peer, idx);
- usocklnd_conn_addref(conn);
- conn->uc_peer = peer;
- usocklnd_peer_addref(peer);
- } else {
- conn2 = peer->up_conns[idx];
- pthread_mutex_lock(&conn2->uc_lock);
-
- if (conn2->uc_state == UC_READY) {
- /* conn2 is in READY state, so conn is "duplicated" */
- pthread_mutex_unlock(&conn2->uc_lock);
- pthread_mutex_unlock(&peer->up_lock);
- usocklnd_peer_decref(peer);
- goto passive_hellosent_connkill;
- }
-
- /* uc_state != UC_READY => switch conn and conn2 */
- /* Relink txs and zc_acks from conn2 to conn.
- * We're sure that nobody but us can access to conn,
- * nevertheless we use mutex (if we're wrong yet,
- * deadlock is easy to see that corrupted list */
- list_add(&tx_list, &conn2->uc_tx_list);
- list_del_init(&conn2->uc_tx_list);
- list_add(&zcack_list, &conn2->uc_zcack_list);
- list_del_init(&conn2->uc_zcack_list);
-
- pthread_mutex_lock(&conn->uc_lock);
- list_add_tail(&conn->uc_tx_list, &tx_list);
- list_del_init(&tx_list);
- list_add_tail(&conn->uc_zcack_list, &zcack_list);
- list_del_init(&zcack_list);
- conn->uc_peer = peer;
- pthread_mutex_unlock(&conn->uc_lock);
-
- conn2->uc_peer = NULL; /* make conn2 zombie */
- pthread_mutex_unlock(&conn2->uc_lock);
- usocklnd_conn_decref(conn2);
-
- usocklnd_link_conn_to_peer(conn, peer, idx);
- usocklnd_conn_addref(conn);
- conn->uc_peer = peer;
- }
-
- lnet_ni_decref(conn->uc_ni);
- conn->uc_ni = NULL;
- pthread_mutex_unlock(&peer->up_lock);
- usocklnd_peer_decref(peer);
-
- passive_hellosent_done:
- /* safely transit to UC_READY state */
- /* rc == 0 */
- pthread_mutex_lock(&conn->uc_lock);
- if (conn->uc_state != UC_DEAD) {
- usocklnd_rx_ksmhdr_state_transition(conn);
-
- /* we're ready to recive incoming packets and maybe
- already have smth. to transmit */
- LASSERT (conn->uc_sending == 0);
- if (list_empty(&conn->uc_tx_list) &&
- list_empty(&conn->uc_zcack_list)) {
- conn->uc_tx_flag = 0;
- rc = usocklnd_add_pollrequest(conn, POLL_SET_REQUEST,
- POLLIN);
- } else {
- conn->uc_tx_deadline =
- cfs_time_shift(usock_tuns.ut_timeout);
- conn->uc_tx_flag = 1;
- rc = usocklnd_add_pollrequest(conn, POLL_SET_REQUEST,
- POLLIN | POLLOUT);
- }
-
- if (rc == 0)
- conn->uc_state = UC_READY;
- }
- pthread_mutex_unlock(&conn->uc_lock);
- return rc;
-
- passive_hellosent_connkill:
- usocklnd_conn_kill(conn);
- return 0;
-}
-
-/* Send as much tx data as possible.
- * Returns 0 or 1 on succsess, <0 if fatal error.
- * 0 means partial send or non-fatal error, 1 - complete.
- * Rely on libcfs_sock_writev() for differentiating fatal and
- * non-fatal errors. An error should be considered as non-fatal if:
- * 1) it still makes sense to continue reading &&
- * 2) anyway, poll() will set up POLLHUP|POLLERR flags */
-int
-usocklnd_send_tx(usock_conn_t *conn, usock_tx_t *tx)
-{
- struct iovec *iov;
- int nob;
- cfs_time_t t;
-
- LASSERT (tx->tx_resid != 0);
-
- do {
- usock_peer_t *peer = conn->uc_peer;
-
- LASSERT (tx->tx_niov > 0);
-
- nob = libcfs_sock_writev(conn->uc_sock,
- tx->tx_iov, tx->tx_niov);
- if (nob < 0)
- conn->uc_errored = 1;
- if (nob <= 0) /* write queue is flow-controlled or error */
- return nob;
-
- LASSERT (nob <= tx->tx_resid);
- tx->tx_resid -= nob;
- t = cfs_time_current();
- conn->uc_tx_deadline = cfs_time_add(t, cfs_time_seconds(usock_tuns.ut_timeout));
-
- if(peer != NULL)
- peer->up_last_alive = t;
-
- /* "consume" iov */
- iov = tx->tx_iov;
- do {
- LASSERT (tx->tx_niov > 0);
-
- if (nob < iov->iov_len) {
- iov->iov_base = (void *)(((unsigned long)(iov->iov_base)) + nob);
- iov->iov_len -= nob;
- break;
- }
-
- nob -= iov->iov_len;
- tx->tx_iov = ++iov;
- tx->tx_niov--;
- } while (nob != 0);
-
- } while (tx->tx_resid != 0);
-
- return 1; /* send complete */
-}
-
-/* Read from wire as much data as possible.
- * Returns 0 or 1 on succsess, <0 if error or EOF.
- * 0 means partial read, 1 - complete */
-int
-usocklnd_read_data(usock_conn_t *conn)
-{
- struct iovec *iov;
- int nob;
- cfs_time_t t;
-
- LASSERT (conn->uc_rx_nob_wanted != 0);
-
- do {
- usock_peer_t *peer = conn->uc_peer;
-
- LASSERT (conn->uc_rx_niov > 0);
-
- nob = libcfs_sock_readv(conn->uc_sock,
- conn->uc_rx_iov, conn->uc_rx_niov);
- if (nob <= 0) {/* read nothing or error */
- if (nob < 0)
- conn->uc_errored = 1;
- return nob;
- }
-
- LASSERT (nob <= conn->uc_rx_nob_wanted);
- conn->uc_rx_nob_wanted -= nob;
- conn->uc_rx_nob_left -= nob;
- t = cfs_time_current();
- conn->uc_rx_deadline = cfs_time_add(t, cfs_time_seconds(usock_tuns.ut_timeout));
-
- if(peer != NULL)
- peer->up_last_alive = t;
-
- /* "consume" iov */
- iov = conn->uc_rx_iov;
- do {
- LASSERT (conn->uc_rx_niov > 0);
-
- if (nob < iov->iov_len) {
- iov->iov_base = (void *)(((unsigned long)(iov->iov_base)) + nob);
- iov->iov_len -= nob;
- break;
- }
-
- nob -= iov->iov_len;
- conn->uc_rx_iov = ++iov;
- conn->uc_rx_niov--;
- } while (nob != 0);
-
- } while (conn->uc_rx_nob_wanted != 0);
-
- return 1; /* read complete */
-}
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- *
- * Copyright (c) 2012, Intel Corporation.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/poll.c
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-
-#include "usocklnd.h"
-#include <unistd.h>
-#include <sys/syscall.h>
-
-void
-usocklnd_process_stale_list(usock_pollthread_t *pt_data)
-{
- while (!list_empty(&pt_data->upt_stale_list)) {
- usock_conn_t *conn;
- conn = list_entry(pt_data->upt_stale_list.next,
- usock_conn_t, uc_stale_list);
-
- list_del(&conn->uc_stale_list);
-
- usocklnd_tear_peer_conn(conn);
- usocklnd_conn_decref(conn); /* -1 for idx2conn[idx] or pr */
- }
-}
-
-int
-usocklnd_poll_thread(void *arg)
-{
- int rc = 0;
- usock_pollthread_t *pt_data = (usock_pollthread_t *)arg;
- cfs_time_t current_time;
- cfs_time_t planned_time;
- int idx;
- int idx_start;
- int idx_finish;
- int chunk;
- int saved_nfds;
- int extra;
- int times;
-
- /* mask signals to avoid SIGPIPE, etc */
- sigset_t sigs;
- sigfillset (&sigs);
- pthread_sigmask (SIG_SETMASK, &sigs, 0);
-
- LASSERT(pt_data != NULL);
-
- planned_time = cfs_time_shift(usock_tuns.ut_poll_timeout);
- chunk = usocklnd_calculate_chunk_size(pt_data->upt_nfds);
- saved_nfds = pt_data->upt_nfds;
- idx_start = 1;
-
- /* Main loop */
- while (usock_data.ud_shutdown == 0) {
- rc = 0;
-
- /* Process all enqueued poll requests */
- pthread_mutex_lock(&pt_data->upt_pollrequests_lock);
- while (!list_empty(&pt_data->upt_pollrequests)) {
- usock_pollrequest_t *pr;
- pr = list_entry(pt_data->upt_pollrequests.next,
- usock_pollrequest_t, upr_list);
-
- list_del(&pr->upr_list);
- rc = usocklnd_process_pollrequest(pr, pt_data);
- if (rc)
- break;
- }
- pthread_mutex_unlock(&pt_data->upt_pollrequests_lock);
-
- if (rc)
- break;
-
- /* Delete conns orphaned due to POLL_DEL_REQUESTs */
- usocklnd_process_stale_list(pt_data);
-
- /* Actual polling for events */
- rc = poll(pt_data->upt_pollfd,
- pt_data->upt_nfds,
- usock_tuns.ut_poll_timeout * 1000);
-
- if (rc < 0 && errno != EINTR) {
- CERROR("Cannot poll(2): errno=%d\n", errno);
- break;
- }
-
- if (rc > 0)
- usocklnd_execute_handlers(pt_data);
-
- current_time = cfs_time_current();
-
- if (pt_data->upt_nfds < 2 ||
- cfs_time_before(current_time, planned_time))
- continue;
-
- /* catch up growing pollfd[] */
- if (pt_data->upt_nfds > saved_nfds) {
- extra = pt_data->upt_nfds - saved_nfds;
- saved_nfds = pt_data->upt_nfds;
- } else {
- extra = 0;
- }
-
- times = cfs_duration_sec(cfs_time_sub(current_time, planned_time)) + 1;
- idx_finish = MIN(idx_start + chunk*times + extra, pt_data->upt_nfds);
-
- for (idx = idx_start; idx < idx_finish; idx++) {
- usock_conn_t *conn = pt_data->upt_idx2conn[idx];
- pthread_mutex_lock(&conn->uc_lock);
- if (usocklnd_conn_timed_out(conn, current_time) &&
- conn->uc_state != UC_DEAD) {
- conn->uc_errored = 1;
- usocklnd_conn_kill_locked(conn);
- }
- pthread_mutex_unlock(&conn->uc_lock);
- }
-
- if (idx_finish == pt_data->upt_nfds) {
- chunk = usocklnd_calculate_chunk_size(pt_data->upt_nfds);
- saved_nfds = pt_data->upt_nfds;
- idx_start = 1;
- }
- else {
- idx_start = idx_finish;
- }
-
- planned_time = cfs_time_add(current_time,
- cfs_time_seconds(usock_tuns.ut_poll_timeout));
- }
-
- /* All conns should be deleted by POLL_DEL_REQUESTs while shutdown */
- LASSERT (rc != 0 || pt_data->upt_nfds == 1);
-
- if (rc) {
- pthread_mutex_lock(&pt_data->upt_pollrequests_lock);
-
- /* Block new poll requests to be enqueued */
- pt_data->upt_errno = rc;
-
- while (!list_empty(&pt_data->upt_pollrequests)) {
- usock_pollrequest_t *pr;
- pr = list_entry(pt_data->upt_pollrequests.next,
- usock_pollrequest_t, upr_list);
-
- list_del(&pr->upr_list);
-
- if (pr->upr_type == POLL_ADD_REQUEST) {
- libcfs_sock_release(pr->upr_conn->uc_sock);
- list_add_tail(&pr->upr_conn->uc_stale_list,
- &pt_data->upt_stale_list);
- } else {
- usocklnd_conn_decref(pr->upr_conn);
- }
-
- LIBCFS_FREE (pr, sizeof(*pr));
- }
- pthread_mutex_unlock(&pt_data->upt_pollrequests_lock);
-
- usocklnd_process_stale_list(pt_data);
-
- for (idx = 1; idx < pt_data->upt_nfds; idx++) {
- usock_conn_t *conn = pt_data->upt_idx2conn[idx];
- LASSERT(conn != NULL);
- libcfs_sock_release(conn->uc_sock);
- usocklnd_tear_peer_conn(conn);
- usocklnd_conn_decref(conn);
- }
- }
-
- /* unblock usocklnd_shutdown() */
- complete(&pt_data->upt_completion);
-
- return 0;
-}
-
-/* Returns 0 on success, <0 else */
-int
-usocklnd_add_pollrequest(usock_conn_t *conn, int type, short value)
-{
- int pt_idx = conn->uc_pt_idx;
- usock_pollthread_t *pt = &usock_data.ud_pollthreads[pt_idx];
- usock_pollrequest_t *pr;
-
- LIBCFS_ALLOC(pr, sizeof(*pr));
- if (pr == NULL) {
- CERROR ("Cannot allocate poll request\n");
- return -ENOMEM;
- }
-
- pr->upr_conn = conn;
- pr->upr_type = type;
- pr->upr_value = value;
-
- usocklnd_conn_addref(conn); /* +1 for poll request */
-
- pthread_mutex_lock(&pt->upt_pollrequests_lock);
-
- if (pt->upt_errno) { /* very rare case: errored poll thread */
- int rc = pt->upt_errno;
- pthread_mutex_unlock(&pt->upt_pollrequests_lock);
- usocklnd_conn_decref(conn);
- LIBCFS_FREE(pr, sizeof(*pr));
- return rc;
- }
-
- list_add_tail(&pr->upr_list, &pt->upt_pollrequests);
- pthread_mutex_unlock(&pt->upt_pollrequests_lock);
- return 0;
-}
-
-void
-usocklnd_add_killrequest(usock_conn_t *conn)
-{
- int pt_idx = conn->uc_pt_idx;
- usock_pollthread_t *pt = &usock_data.ud_pollthreads[pt_idx];
- usock_pollrequest_t *pr = conn->uc_preq;
-
- /* Use preallocated poll request because there is no good
- * workaround for ENOMEM error while killing connection */
- if (pr) {
- pr->upr_conn = conn;
- pr->upr_type = POLL_DEL_REQUEST;
- pr->upr_value = 0;
-
- usocklnd_conn_addref(conn); /* +1 for poll request */
-
- pthread_mutex_lock(&pt->upt_pollrequests_lock);
-
- if (pt->upt_errno) { /* very rare case: errored poll thread */
- pthread_mutex_unlock(&pt->upt_pollrequests_lock);
- usocklnd_conn_decref(conn);
- return; /* conn will be killed in poll thread anyway */
- }
-
- list_add_tail(&pr->upr_list, &pt->upt_pollrequests);
- pthread_mutex_unlock(&pt->upt_pollrequests_lock);
-
- conn->uc_preq = NULL;
- }
-}
-
-/* Process poll request. Update poll data.
- * Returns 0 on success, <0 else */
-int
-usocklnd_process_pollrequest(usock_pollrequest_t *pr,
- usock_pollthread_t *pt_data)
-{
- int type = pr->upr_type;
- short value = pr->upr_value;
- usock_conn_t *conn = pr->upr_conn;
- int idx = 0;
- struct pollfd *pollfd = pt_data->upt_pollfd;
- int *fd2idx = pt_data->upt_fd2idx;
- usock_conn_t **idx2conn = pt_data->upt_idx2conn;
- int *skip = pt_data->upt_skip;
-
- LASSERT(conn != NULL);
- LASSERT(conn->uc_sock != NULL);
- LASSERT(type == POLL_ADD_REQUEST ||
- LIBCFS_SOCK2FD(conn->uc_sock) < pt_data->upt_nfd2idx);
-
- if (type != POLL_ADD_REQUEST) {
- idx = fd2idx[LIBCFS_SOCK2FD(conn->uc_sock)];
- if (idx > 0 && idx < pt_data->upt_nfds) { /* hot path */
- LASSERT(pollfd[idx].fd ==
- LIBCFS_SOCK2FD(conn->uc_sock));
- } else { /* unlikely */
- CWARN("Very unlikely event happend: trying to"
- " handle poll request of type %d but idx=%d"
- " is out of range [1 ... %d]. Is shutdown"
- " in progress (%d)?\n",
- type, idx, pt_data->upt_nfds - 1,
- usock_data.ud_shutdown);
-
- LIBCFS_FREE (pr, sizeof(*pr));
- usocklnd_conn_decref(conn);
- return 0;
- }
- }
-
- LIBCFS_FREE (pr, sizeof(*pr));
-
- switch (type) {
- case POLL_ADD_REQUEST:
- if (pt_data->upt_nfds >= pt_data->upt_npollfd) {
- /* resize pollfd[], idx2conn[] and skip[] */
- struct pollfd *new_pollfd;
- int new_npollfd = pt_data->upt_npollfd * 2;
- usock_conn_t **new_idx2conn;
- int *new_skip;
-
- new_pollfd = LIBCFS_REALLOC(pollfd, new_npollfd *
- sizeof(struct pollfd));
- if (new_pollfd == NULL)
- goto process_pollrequest_enomem;
- pt_data->upt_pollfd = pollfd = new_pollfd;
-
- new_idx2conn = LIBCFS_REALLOC(idx2conn, new_npollfd *
- sizeof(usock_conn_t *));
- if (new_idx2conn == NULL)
- goto process_pollrequest_enomem;
- pt_data->upt_idx2conn = idx2conn = new_idx2conn;
-
- new_skip = LIBCFS_REALLOC(skip, new_npollfd *
- sizeof(int));
- if (new_skip == NULL)
- goto process_pollrequest_enomem;
- pt_data->upt_skip = new_skip;
-
- pt_data->upt_npollfd = new_npollfd;
- }
-
- if (LIBCFS_SOCK2FD(conn->uc_sock) >= pt_data->upt_nfd2idx) {
- /* resize fd2idx[] */
- int *new_fd2idx;
- int new_nfd2idx = pt_data->upt_nfd2idx * 2;
-
- while (new_nfd2idx <= LIBCFS_SOCK2FD(conn->uc_sock))
- new_nfd2idx *= 2;
-
- new_fd2idx = LIBCFS_REALLOC(fd2idx, new_nfd2idx *
- sizeof(int));
- if (new_fd2idx == NULL)
- goto process_pollrequest_enomem;
-
- pt_data->upt_fd2idx = fd2idx = new_fd2idx;
- memset(fd2idx + pt_data->upt_nfd2idx, 0,
- (new_nfd2idx - pt_data->upt_nfd2idx)
- * sizeof(int));
- pt_data->upt_nfd2idx = new_nfd2idx;
- }
-
- LASSERT(fd2idx[LIBCFS_SOCK2FD(conn->uc_sock)] == 0);
-
- idx = pt_data->upt_nfds++;
- idx2conn[idx] = conn;
- fd2idx[LIBCFS_SOCK2FD(conn->uc_sock)] = idx;
-
- pollfd[idx].fd = LIBCFS_SOCK2FD(conn->uc_sock);
- pollfd[idx].events = value;
- pollfd[idx].revents = 0;
- break;
- case POLL_DEL_REQUEST:
- fd2idx[LIBCFS_SOCK2FD(conn->uc_sock)] = 0; /* invalidate this
- * entry */
- --pt_data->upt_nfds;
- if (idx != pt_data->upt_nfds) {
- /* shift last entry into released position */
- memcpy(&pollfd[idx], &pollfd[pt_data->upt_nfds],
- sizeof(struct pollfd));
- idx2conn[idx] = idx2conn[pt_data->upt_nfds];
- fd2idx[pollfd[idx].fd] = idx;
- }
-
- libcfs_sock_release(conn->uc_sock);
- list_add_tail(&conn->uc_stale_list,
- &pt_data->upt_stale_list);
- break;
- case POLL_RX_SET_REQUEST:
- pollfd[idx].events = (pollfd[idx].events & ~POLLIN) | value;
- break;
- case POLL_TX_SET_REQUEST:
- pollfd[idx].events = (pollfd[idx].events & ~POLLOUT) | value;
- break;
- case POLL_SET_REQUEST:
- pollfd[idx].events = value;
- break;
- default:
- LBUG(); /* unknown type */
- }
-
- /* In the case of POLL_ADD_REQUEST, idx2conn[idx] takes the
- * reference that poll request possesses */
- if (type != POLL_ADD_REQUEST)
- usocklnd_conn_decref(conn);
-
- return 0;
-
- process_pollrequest_enomem:
- usocklnd_conn_decref(conn);
- return -ENOMEM;
-}
-
-/* Loop on poll data executing handlers repeatedly until
- * fair_limit is reached or all entries are exhausted */
-void
-usocklnd_execute_handlers(usock_pollthread_t *pt_data)
-{
- struct pollfd *pollfd = pt_data->upt_pollfd;
- int nfds = pt_data->upt_nfds;
- usock_conn_t **idx2conn = pt_data->upt_idx2conn;
- int *skip = pt_data->upt_skip;
- int j;
-
- if (pollfd[0].revents & POLLIN)
- while (usocklnd_notifier_handler(pollfd[0].fd) > 0)
- ;
-
- skip[0] = 1; /* always skip notifier fd */
-
- for (j = 0; j < usock_tuns.ut_fair_limit; j++) {
- int prev = 0;
- int i = skip[0];
-
- if (i >= nfds) /* nothing ready */
- break;
-
- do {
- usock_conn_t *conn = idx2conn[i];
- int next;
-
- if (j == 0) /* first pass... */
- next = skip[i] = i+1; /* set skip chain */
- else /* later passes... */
- next = skip[i]; /* skip unready pollfds */
-
- /* kill connection if it's closed by peer and
- * there is no data pending for reading */
- if ((pollfd[i].revents & POLLERR) != 0 ||
- (pollfd[i].revents & POLLHUP) != 0) {
- if ((pollfd[i].events & POLLIN) != 0 &&
- (pollfd[i].revents & POLLIN) == 0)
- usocklnd_conn_kill(conn);
- else
- usocklnd_exception_handler(conn);
- }
-
- if ((pollfd[i].revents & POLLIN) != 0 &&
- usocklnd_read_handler(conn) <= 0)
- pollfd[i].revents &= ~POLLIN;
-
- if ((pollfd[i].revents & POLLOUT) != 0 &&
- usocklnd_write_handler(conn) <= 0)
- pollfd[i].revents &= ~POLLOUT;
-
- if ((pollfd[i].revents & (POLLIN | POLLOUT)) == 0)
- skip[prev] = next; /* skip this entry next pass */
- else
- prev = i;
-
- i = next;
- } while (i < nfds);
- }
-}
-
-int
-usocklnd_calculate_chunk_size(int num)
-{
- const int n = 4;
- const int p = usock_tuns.ut_poll_timeout;
- int chunk = num;
-
- /* chunk should be big enough to detect a timeout on any
- * connection within (n+1)/n times the timeout interval
- * if we checks every 'p' seconds 'chunk' conns */
-
- if (usock_tuns.ut_timeout > n * p)
- chunk = (chunk * n * p) / usock_tuns.ut_timeout;
-
- if (chunk == 0)
- chunk = 1;
-
- return chunk;
-}
-
-void
-usocklnd_wakeup_pollthread(int i)
-{
- usock_pollthread_t *pt = &usock_data.ud_pollthreads[i];
- int notification = 0;
- int rc;
-
- rc = syscall(SYS_write, LIBCFS_SOCK2FD(pt->upt_notifier[0]),
- ¬ification, sizeof(notification));
-
- if (rc != sizeof(notification))
- CERROR("Very unlikely event happend: "
- "cannot write to notifier fd (rc=%d; errno=%d)\n",
- rc, errno);
-}
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- *
- * Copyright (c) 2012, Intel Corporation.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/usocklnd.c
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-
-#include "usocklnd.h"
-#include <sys/time.h>
-
-lnd_t the_tcplnd = {
- .lnd_type = SOCKLND,
- .lnd_startup = usocklnd_startup,
- .lnd_shutdown = usocklnd_shutdown,
- .lnd_send = usocklnd_send,
- .lnd_recv = usocklnd_recv,
- .lnd_accept = usocklnd_accept,
-};
-
-usock_data_t usock_data;
-usock_tunables_t usock_tuns = {
- .ut_timeout = 50,
- .ut_poll_timeout = 1,
- .ut_fair_limit = 1,
- .ut_npollthreads = 0,
- .ut_min_bulk = 1<<10,
- .ut_txcredits = 256,
- .ut_peertxcredits = 8,
- .ut_socknagle = 0,
- .ut_sockbufsiz = 0,
-};
-
-#define MAX_REASONABLE_TIMEOUT 36000 /* 10 hours */
-#define MAX_REASONABLE_NPT 1000
-
-int
-usocklnd_validate_tunables()
-{
- if (usock_tuns.ut_timeout <= 0 ||
- usock_tuns.ut_timeout > MAX_REASONABLE_TIMEOUT) {
- CERROR("USOCK_TIMEOUT: %d is out of reasonable limits\n",
- usock_tuns.ut_timeout);
- return -1;
- }
-
- if (usock_tuns.ut_poll_timeout <= 0 ||
- usock_tuns.ut_poll_timeout > MAX_REASONABLE_TIMEOUT) {
- CERROR("USOCK_POLL_TIMEOUT: %d is out of reasonable limits\n",
- usock_tuns.ut_poll_timeout);
- return -1;
- }
-
- if (usock_tuns.ut_fair_limit <= 0) {
- CERROR("Invalid USOCK_FAIR_LIMIT: %d (should be >0)\n",
- usock_tuns.ut_fair_limit);
- return -1;
- }
-
- if (usock_tuns.ut_npollthreads < 0 ||
- usock_tuns.ut_npollthreads > MAX_REASONABLE_NPT) {
- CERROR("USOCK_NPOLLTHREADS: %d is out of reasonable limits\n",
- usock_tuns.ut_npollthreads);
- return -1;
- }
-
- if (usock_tuns.ut_txcredits <= 0) {
- CERROR("USOCK_TXCREDITS: %d should be positive\n",
- usock_tuns.ut_txcredits);
- return -1;
- }
-
- if (usock_tuns.ut_peertxcredits <= 0) {
- CERROR("USOCK_PEERTXCREDITS: %d should be positive\n",
- usock_tuns.ut_peertxcredits);
- return -1;
- }
-
- if (usock_tuns.ut_peertxcredits > usock_tuns.ut_txcredits) {
- CERROR("USOCK_PEERTXCREDITS: %d should not be greater"
- " than USOCK_TXCREDITS: %d\n",
- usock_tuns.ut_peertxcredits, usock_tuns.ut_txcredits);
- return -1;
- }
-
- if (usock_tuns.ut_socknagle != 0 &&
- usock_tuns.ut_socknagle != 1) {
- CERROR("USOCK_SOCKNAGLE: %d should be 0 or 1\n",
- usock_tuns.ut_socknagle);
- return -1;
- }
-
- if (usock_tuns.ut_sockbufsiz < 0) {
- CERROR("USOCK_SOCKBUFSIZ: %d should be 0 or positive\n",
- usock_tuns.ut_sockbufsiz);
- return -1;
- }
-
- return 0;
-}
-
-void
-usocklnd_release_poll_states(int n)
-{
- int i;
-
- for (i = 0; i < n; i++) {
- usock_pollthread_t *pt = &usock_data.ud_pollthreads[i];
-
- libcfs_sock_release(pt->upt_notifier[0]);
- libcfs_sock_release(pt->upt_notifier[1]);
-
- pthread_mutex_destroy(&pt->upt_pollrequests_lock);
- fini_completion(&pt->upt_completion);
-
- LIBCFS_FREE (pt->upt_pollfd,
- sizeof(struct pollfd) * pt->upt_npollfd);
- LIBCFS_FREE (pt->upt_idx2conn,
- sizeof(usock_conn_t *) * pt->upt_npollfd);
- LIBCFS_FREE (pt->upt_fd2idx,
- sizeof(int) * pt->upt_nfd2idx);
- }
-}
-
-int
-usocklnd_update_tunables()
-{
- int rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_timeout,
- "USOCK_TIMEOUT");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_poll_timeout,
- "USOCK_POLL_TIMEOUT");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_npollthreads,
- "USOCK_NPOLLTHREADS");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_fair_limit,
- "USOCK_FAIR_LIMIT");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_min_bulk,
- "USOCK_MIN_BULK");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_txcredits,
- "USOCK_TXCREDITS");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_peertxcredits,
- "USOCK_PEERTXCREDITS");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_socknagle,
- "USOCK_SOCKNAGLE");
- if (rc)
- return rc;
-
- rc = lnet_parse_int_tunable(&usock_tuns.ut_sockbufsiz,
- "USOCK_SOCKBUFSIZ");
- if (rc)
- return rc;
-
- if (usocklnd_validate_tunables())
- return -EINVAL;
-
- if (usock_tuns.ut_npollthreads == 0) {
- usock_tuns.ut_npollthreads = cfs_online_cpus();
-
- if (usock_tuns.ut_npollthreads <= 0) {
- CERROR("Cannot find out the number of online CPUs\n");
- return -EINVAL;
- }
- }
-
- return 0;
-}
-
-
-int
-usocklnd_base_startup()
-{
- usock_pollthread_t *pt;
- int i;
- int rc;
-
- rc = usocklnd_update_tunables();
- if (rc)
- return rc;
-
- usock_data.ud_npollthreads = usock_tuns.ut_npollthreads;
-
- LIBCFS_ALLOC (usock_data.ud_pollthreads,
- usock_data.ud_npollthreads *
- sizeof(usock_pollthread_t));
- if (usock_data.ud_pollthreads == NULL)
- return -ENOMEM;
-
- /* Initialize poll thread state structures */
- for (i = 0; i < usock_data.ud_npollthreads; i++) {
-
- pt = &usock_data.ud_pollthreads[i];
-
- rc = -ENOMEM;
-
- LIBCFS_ALLOC (pt->upt_pollfd,
- sizeof(struct pollfd) * UPT_START_SIZ);
- if (pt->upt_pollfd == NULL)
- goto base_startup_failed_0;
-
- LIBCFS_ALLOC (pt->upt_idx2conn,
- sizeof(usock_conn_t *) * UPT_START_SIZ);
- if (pt->upt_idx2conn == NULL)
- goto base_startup_failed_1;
-
- LIBCFS_ALLOC (pt->upt_fd2idx,
- sizeof(int) * UPT_START_SIZ);
- if (pt->upt_fd2idx == NULL)
- goto base_startup_failed_2;
-
- memset(pt->upt_fd2idx, 0,
- sizeof(int) * UPT_START_SIZ);
-
- LIBCFS_ALLOC (pt->upt_skip,
- sizeof(int) * UPT_START_SIZ);
- if (pt->upt_skip == NULL)
- goto base_startup_failed_3;
-
- pt->upt_npollfd = pt->upt_nfd2idx = UPT_START_SIZ;
-
- rc = libcfs_socketpair(pt->upt_notifier);
- if (rc != 0)
- goto base_startup_failed_4;
-
- pt->upt_pollfd[0].fd = LIBCFS_SOCK2FD(pt->upt_notifier[1]);
- pt->upt_pollfd[0].events = POLLIN;
- pt->upt_pollfd[0].revents = 0;
-
- pt->upt_nfds = 1;
- pt->upt_idx2conn[0] = NULL;
-
- pt->upt_errno = 0;
- INIT_LIST_HEAD(&pt->upt_pollrequests);
- INIT_LIST_HEAD(&pt->upt_stale_list);
- pthread_mutex_init(&pt->upt_pollrequests_lock, NULL);
- init_completion(&pt->upt_completion);
- }
-
- /* Initialize peer hash list */
- for (i = 0; i < UD_PEER_HASH_SIZE; i++)
- INIT_LIST_HEAD(&usock_data.ud_peers[i]);
-
- pthread_rwlock_init(&usock_data.ud_peers_lock, NULL);
-
- /* Spawn poll threads */
- for (i = 0; i < usock_data.ud_npollthreads; i++) {
- struct task_struct *task;
-
- task = kthread_run(usocklnd_poll_thread,
- &usock_data.ud_pollthreads[i], "");
- if (IS_ERR(task)) {
- usocklnd_base_shutdown(i);
- return PTR_ERR(task);
- }
- }
-
- usock_data.ud_state = UD_STATE_INITIALIZED;
-
- return 0;
-
- base_startup_failed_4:
- LIBCFS_FREE (pt->upt_skip, sizeof(int) * UPT_START_SIZ);
- base_startup_failed_3:
- LIBCFS_FREE (pt->upt_fd2idx, sizeof(int) * UPT_START_SIZ);
- base_startup_failed_2:
- LIBCFS_FREE (pt->upt_idx2conn, sizeof(usock_conn_t *) * UPT_START_SIZ);
- base_startup_failed_1:
- LIBCFS_FREE (pt->upt_pollfd, sizeof(struct pollfd) * UPT_START_SIZ);
- base_startup_failed_0:
- LASSERT(rc != 0);
- usocklnd_release_poll_states(i);
- LIBCFS_FREE (usock_data.ud_pollthreads,
- usock_data.ud_npollthreads *
- sizeof(usock_pollthread_t));
- return rc;
-}
-
-void
-usocklnd_base_shutdown(int n)
-{
- int i;
-
- usock_data.ud_shutdown = 1;
- for (i = 0; i < n; i++) {
- usock_pollthread_t *pt = &usock_data.ud_pollthreads[i];
- usocklnd_wakeup_pollthread(i);
- wait_for_completion(&pt->upt_completion);
- }
-
- pthread_rwlock_destroy(&usock_data.ud_peers_lock);
-
- usocklnd_release_poll_states(usock_data.ud_npollthreads);
-
- LIBCFS_FREE (usock_data.ud_pollthreads,
- usock_data.ud_npollthreads *
- sizeof(usock_pollthread_t));
-
- usock_data.ud_state = UD_STATE_INIT_NOTHING;
-}
-
-__u64
-usocklnd_new_incarnation()
-{
- struct timeval tv;
- int rc = gettimeofday(&tv, NULL);
- LASSERT (rc == 0);
- return (((__u64)tv.tv_sec) * 1000000) + tv.tv_usec;
-}
-
-static int
-usocklnd_assign_ni_nid(lnet_ni_t *ni)
-{
- int rc;
- int up;
- __u32 ipaddr;
-
- /* Find correct IP-address and update ni_nid with it.
- * Two cases are supported:
- * 1) no explicit interfaces are defined. NID will be assigned to
- * first non-lo interface that is up;
- * 2) exactly one explicit interface is defined. For example,
- * LNET_NETWORKS='tcp(eth0)' */
-
- if (ni->ni_interfaces[0] == NULL) {
- char **names;
- int i, n;
-
- n = libcfs_ipif_enumerate(&names);
- if (n <= 0) {
- CERROR("Can't enumerate interfaces: %d\n", n);
- return -1;
- }
-
- for (i = 0; i < n; i++) {
-
- if (!strcmp(names[i], "lo")) /* skip the loopback IF */
- continue;
-
- rc = libcfs_ipif_query(names[i], &up, &ipaddr);
- if (rc != 0) {
- CWARN("Can't get interface %s info: %d\n",
- names[i], rc);
- continue;
- }
-
- if (!up) {
- CWARN("Ignoring interface %s (down)\n",
- names[i]);
- continue;
- }
-
- break; /* one address is quite enough */
- }
-
- libcfs_ipif_free_enumeration(names, n);
-
- if (i >= n) {
- CERROR("Can't find any usable interfaces\n");
- return -1;
- }
-
- CDEBUG(D_NET, "No explicit interfaces defined. "
- "%u.%u.%u.%u used\n", HIPQUAD(ipaddr));
- } else {
- if (ni->ni_interfaces[1] != NULL) {
- CERROR("only one explicit interface is allowed\n");
- return -1;
- }
-
- rc = libcfs_ipif_query(ni->ni_interfaces[0], &up, &ipaddr);
- if (rc != 0) {
- CERROR("Can't get interface %s info: %d\n",
- ni->ni_interfaces[0], rc);
- return -1;
- }
-
- if (!up) {
- CERROR("Explicit interface defined: %s but is down\n",
- ni->ni_interfaces[0]);
- return -1;
- }
-
- CDEBUG(D_NET, "Explicit interface defined: %s. "
- "%u.%u.%u.%u used\n",
- ni->ni_interfaces[0], HIPQUAD(ipaddr));
-
- }
-
- ni->ni_nid = LNET_MKNID(LNET_NIDNET(ni->ni_nid), ipaddr);
-
- return 0;
-}
-
-int
-usocklnd_startup(lnet_ni_t *ni)
-{
- int rc;
- usock_net_t *net;
-
- if (usock_data.ud_state == UD_STATE_INIT_NOTHING) {
- rc = usocklnd_base_startup();
- if (rc != 0)
- return rc;
- }
-
- LIBCFS_ALLOC(net, sizeof(*net));
- if (net == NULL)
- goto startup_failed_0;
-
- memset(net, 0, sizeof(*net));
- net->un_incarnation = usocklnd_new_incarnation();
- pthread_mutex_init(&net->un_lock, NULL);
- pthread_cond_init(&net->un_cond, NULL);
-
- ni->ni_data = net;
-
- rc = usocklnd_assign_ni_nid(ni);
- if (rc != 0)
- goto startup_failed_1;
-
- LASSERT (ni->ni_lnd == &the_tcplnd);
-
- ni->ni_maxtxcredits = usock_tuns.ut_txcredits;
- ni->ni_peertxcredits = usock_tuns.ut_peertxcredits;
-
- usock_data.ud_nets_count++;
- return 0;
-
- startup_failed_1:
- pthread_mutex_destroy(&net->un_lock);
- pthread_cond_destroy(&net->un_cond);
- LIBCFS_FREE(net, sizeof(*net));
- startup_failed_0:
- if (usock_data.ud_nets_count == 0)
- usocklnd_base_shutdown(usock_data.ud_npollthreads);
-
- return -ENETDOWN;
-}
-
-void
-usocklnd_shutdown(lnet_ni_t *ni)
-{
- usock_net_t *net = ni->ni_data;
-
- net->un_shutdown = 1;
-
- usocklnd_del_all_peers(ni);
-
- /* Wait for all peer state to clean up */
- pthread_mutex_lock(&net->un_lock);
- while (net->un_peercount != 0)
- pthread_cond_wait(&net->un_cond, &net->un_lock);
- pthread_mutex_unlock(&net->un_lock);
-
- /* Release usock_net_t structure */
- pthread_mutex_destroy(&net->un_lock);
- pthread_cond_destroy(&net->un_cond);
- LIBCFS_FREE(net, sizeof(*net));
-
- usock_data.ud_nets_count--;
- if (usock_data.ud_nets_count == 0)
- usocklnd_base_shutdown(usock_data.ud_npollthreads);
-}
-
-void
-usocklnd_del_all_peers(lnet_ni_t *ni)
-{
- struct list_head *ptmp;
- struct list_head *pnxt;
- usock_peer_t *peer;
- int i;
-
- pthread_rwlock_wrlock(&usock_data.ud_peers_lock);
-
- for (i = 0; i < UD_PEER_HASH_SIZE; i++) {
- list_for_each_safe(ptmp, pnxt, &usock_data.ud_peers[i]) {
- peer = list_entry(ptmp, usock_peer_t, up_list);
-
- if (peer->up_ni != ni)
- continue;
-
- usocklnd_del_peer_and_conns(peer);
- }
- }
-
- pthread_rwlock_unlock(&usock_data.ud_peers_lock);
-
- /* wakeup all threads */
- for (i = 0; i < usock_data.ud_npollthreads; i++)
- usocklnd_wakeup_pollthread(i);
-}
-
-void
-usocklnd_del_peer_and_conns(usock_peer_t *peer)
-{
- /* peer cannot disappear because it's still in hash list */
-
- pthread_mutex_lock(&peer->up_lock);
- /* content of conn[] array cannot change now */
- usocklnd_del_conns_locked(peer);
- pthread_mutex_unlock(&peer->up_lock);
-
- /* peer hash list is still protected by the caller */
- list_del(&peer->up_list);
-
- usocklnd_peer_decref(peer); /* peer isn't in hash list anymore */
-}
-
-void
-usocklnd_del_conns_locked(usock_peer_t *peer)
-{
- int i;
-
- for (i=0; i < N_CONN_TYPES; i++) {
- usock_conn_t *conn = peer->up_conns[i];
- if (conn != NULL)
- usocklnd_conn_kill(conn);
- }
-}
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- *
- * Copyright (c) 2012, Intel Corporation.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/usocklnd.h
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-#ifndef _GNU_SOURCE
-#define _GNU_SOURCE
-#endif
-#include <pthread.h>
-#include <poll.h>
-#include <lnet/lib-lnet.h>
-#include <lnet/socklnd.h>
-
-typedef struct {
- struct list_head tx_list; /* neccessary to form tx list */
- lnet_msg_t *tx_lnetmsg; /* lnet message for lnet_finalize() */
- ksock_msg_t tx_msg; /* buffer for wire header of ksock msg */
- int tx_resid; /* # of residual bytes */
- int tx_nob; /* # of packet bytes */
- int tx_size; /* size of this descriptor */
- struct iovec *tx_iov; /* points to tx_iova[i] */
- int tx_niov; /* # of packet iovec frags */
- struct iovec tx_iova[1]; /* iov for header */
-} usock_tx_t;
-
-struct usock_peer_s;
-
-typedef struct {
- cfs_socket_t *uc_sock; /* socket */
- int uc_type; /* conn type */
- int uc_activeflag; /* active side of connection? */
- int uc_flip; /* is peer other endian? */
- int uc_state; /* connection state */
- struct usock_peer_s *uc_peer; /* owning peer */
- lnet_process_id_t uc_peerid; /* id of remote peer */
- int uc_pt_idx; /* index in ud_pollthreads[] of
- * owning poll thread */
- lnet_ni_t *uc_ni; /* parent NI while accepting */
- struct usock_preq_s *uc_preq; /* preallocated request */
- __u32 uc_peer_ip; /* IP address of the peer */
- __u16 uc_peer_port; /* port of the peer */
- struct list_head uc_stale_list; /* orphaned connections */
-
- /* Receive state */
- int uc_rx_state; /* message or hello state */
- ksock_hello_msg_t *uc_rx_hello; /* hello buffer */
- struct iovec *uc_rx_iov; /* points to uc_rx_iova[i] */
- struct iovec uc_rx_iova[LNET_MAX_IOV]; /* message frags */
- int uc_rx_niov; /* # frags */
- int uc_rx_nob_left; /* # bytes to next hdr/body */
- int uc_rx_nob_wanted; /* # of bytes actually wanted */
- void *uc_rx_lnetmsg; /* LNET message being received */
- cfs_time_t uc_rx_deadline; /* when to time out */
- int uc_rx_flag; /* deadline valid? */
- ksock_msg_t uc_rx_msg; /* message buffer */
-
- /* Send state */
- struct list_head uc_tx_list; /* pending txs */
- struct list_head uc_zcack_list; /* pending zc_acks */
- cfs_time_t uc_tx_deadline; /* when to time out */
- int uc_tx_flag; /* deadline valid? */
- int uc_sending; /* send op is in progress */
- usock_tx_t *uc_tx_hello; /* fake tx with hello */
-
- mt_atomic_t uc_refcount; /* # of users */
- pthread_mutex_t uc_lock; /* serialize */
- int uc_errored; /* a flag for lnet_notify() */
-} usock_conn_t;
-
-/* Allowable conn states are: */
-#define UC_CONNECTING 1
-#define UC_SENDING_HELLO 2
-#define UC_RECEIVING_HELLO 3
-#define UC_READY 4
-#define UC_DEAD 5
-
-/* Allowable RX states are: */
-#define UC_RX_HELLO_MAGIC 1
-#define UC_RX_HELLO_VERSION 2
-#define UC_RX_HELLO_BODY 3
-#define UC_RX_HELLO_IPS 4
-#define UC_RX_KSM_HEADER 5
-#define UC_RX_LNET_HEADER 6
-#define UC_RX_PARSE 7
-#define UC_RX_PARSE_WAIT 8
-#define UC_RX_LNET_PAYLOAD 9
-#define UC_RX_SKIPPING 10
-
-#define N_CONN_TYPES 3 /* CONTROL, BULK_IN and BULK_OUT */
-
-typedef struct usock_peer_s {
- /* neccessary to form peer list */
- struct list_head up_list;
- lnet_process_id_t up_peerid; /* id of remote peer */
- usock_conn_t *up_conns[N_CONN_TYPES]; /* conns that connect us
- * us with the peer */
- lnet_ni_t *up_ni; /* pointer to parent NI */
- __u64 up_incarnation; /* peer's incarnation */
- int up_incrn_is_set;/* 0 if peer's incarnation
- * hasn't been set so far */
- mt_atomic_t up_refcount; /* # of users */
- pthread_mutex_t up_lock; /* serialize */
- int up_errored; /* a flag for lnet_notify() */
- cfs_time_t up_last_alive; /* when the peer was last alive */
-} usock_peer_t;
-
-typedef struct {
- cfs_socket_t *upt_notifier[2]; /* notifier sockets: 1st for
- * writing, 2nd for reading */
- struct pollfd *upt_pollfd; /* poll fds */
- int upt_nfds; /* active poll fds */
- int upt_npollfd; /* allocated poll fds */
- usock_conn_t **upt_idx2conn; /* conns corresponding to
- * upt_pollfd[idx] */
- int *upt_skip; /* skip chain */
- int *upt_fd2idx; /* index into upt_pollfd[]
- * by fd */
- int upt_nfd2idx; /* # of allocated elements
- * of upt_fd2idx[] */
- struct list_head upt_stale_list; /* list of orphaned conns */
- struct list_head upt_pollrequests; /* list of poll requests */
- pthread_mutex_t upt_pollrequests_lock; /* serialize */
- int upt_errno; /* non-zero if errored */
- struct completion upt_completion; /* wait/signal facility for
- * syncronizing shutdown */
-} usock_pollthread_t;
-
-/* Number of elements in upt_pollfd[], upt_idx2conn[] and upt_fd2idx[]
- * at initialization time. Will be resized on demand */
-#define UPT_START_SIZ 32
-
-/* # peer lists */
-#define UD_PEER_HASH_SIZE 101
-
-typedef struct {
- int ud_state; /* initialization state */
- int ud_npollthreads; /* # of poll threads */
- usock_pollthread_t *ud_pollthreads; /* their state */
- int ud_shutdown; /* shutdown flag */
- int ud_nets_count; /* # of instances */
- struct list_head ud_peers[UD_PEER_HASH_SIZE]; /* peer hash table */
- pthread_rwlock_t ud_peers_lock; /* serialize */
-} usock_data_t;
-
-extern usock_data_t usock_data;
-
-/* ud_state allowed values */
-#define UD_STATE_INIT_NOTHING 0
-#define UD_STATE_INITIALIZED 1
-
-typedef struct {
- int un_peercount; /* # of peers */
- int un_shutdown; /* shutdown flag */
- __u64 un_incarnation; /* my epoch */
- pthread_cond_t un_cond; /* condvar to wait for notifications */
- pthread_mutex_t un_lock; /* a lock to protect un_cond */
-} usock_net_t;
-
-typedef struct {
- int ut_poll_timeout; /* the third arg for poll(2) (seconds) */
- int ut_timeout; /* "stuck" socket timeout (seconds) */
- int ut_npollthreads; /* number of poll thread to spawn */
- int ut_fair_limit; /* how many packets can we receive or transmit
- * without calling poll(2) */
- int ut_min_bulk; /* smallest "large" message */
- int ut_txcredits; /* # concurrent sends */
- int ut_peertxcredits; /* # concurrent sends to 1 peer */
- int ut_socknagle; /* Is Nagle alg on ? */
- int ut_sockbufsiz; /* size of socket buffers */
-} usock_tunables_t;
-
-extern usock_tunables_t usock_tuns;
-
-typedef struct usock_preq_s {
- int upr_type; /* type of requested action */
- short upr_value; /* bitmask of POLLIN and POLLOUT bits */
- usock_conn_t * upr_conn; /* a conn for the sake of which
- * action will be performed */
- struct list_head upr_list; /* neccessary to form list */
-} usock_pollrequest_t;
-
-/* Allowable poll request types are: */
-#define POLL_ADD_REQUEST 1
-#define POLL_DEL_REQUEST 2
-#define POLL_RX_SET_REQUEST 3
-#define POLL_TX_SET_REQUEST 4
-#define POLL_SET_REQUEST 5
-
-typedef struct {
- struct list_head zc_list; /* neccessary to form zc_ack list */
- __u64 zc_cookie; /* zero-copy cookie */
-} usock_zc_ack_t;
-
-static inline void
-usocklnd_conn_addref(usock_conn_t *conn)
-{
- LASSERT(mt_atomic_read(&conn->uc_refcount) > 0);
- mt_atomic_inc(&conn->uc_refcount);
-}
-
-void usocklnd_destroy_conn(usock_conn_t *conn);
-
-static inline void
-usocklnd_conn_decref(usock_conn_t *conn)
-{
- LASSERT(mt_atomic_read(&conn->uc_refcount) > 0);
- if (mt_atomic_dec_and_test(&conn->uc_refcount))
- usocklnd_destroy_conn(conn);
-}
-
-static inline void
-usocklnd_peer_addref(usock_peer_t *peer)
-{
- LASSERT(mt_atomic_read(&peer->up_refcount) > 0);
- mt_atomic_inc(&peer->up_refcount);
-}
-
-void usocklnd_destroy_peer(usock_peer_t *peer);
-
-static inline void
-usocklnd_peer_decref(usock_peer_t *peer)
-{
- LASSERT(mt_atomic_read(&peer->up_refcount) > 0);
- if (mt_atomic_dec_and_test(&peer->up_refcount))
- usocklnd_destroy_peer(peer);
-}
-
-static inline int
-usocklnd_ip2pt_idx(__u32 ip) {
- return ip % usock_data.ud_npollthreads;
-}
-
-static inline struct list_head *
-usocklnd_nid2peerlist(lnet_nid_t nid)
-{
- unsigned int hash = ((unsigned int)nid) % UD_PEER_HASH_SIZE;
-
- return &usock_data.ud_peers[hash];
-}
-
-int usocklnd_startup(lnet_ni_t *ni);
-void usocklnd_shutdown(lnet_ni_t *ni);
-int usocklnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg);
-int usocklnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
- unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
- unsigned int offset, unsigned int mlen, unsigned int rlen);
-int usocklnd_accept(lnet_ni_t *ni, cfs_socket_t *sock);
-
-int usocklnd_poll_thread(void *arg);
-int usocklnd_add_pollrequest(usock_conn_t *conn, int type, short value);
-void usocklnd_add_killrequest(usock_conn_t *conn);
-int usocklnd_process_pollrequest(usock_pollrequest_t *pr,
- usock_pollthread_t *pt_data);
-void usocklnd_execute_handlers(usock_pollthread_t *pt_data);
-int usocklnd_calculate_chunk_size(int num);
-void usocklnd_wakeup_pollthread(int i);
-
-int usocklnd_notifier_handler(int fd);
-void usocklnd_exception_handler(usock_conn_t *conn);
-int usocklnd_read_handler(usock_conn_t *conn);
-int usocklnd_read_msg(usock_conn_t *conn, int *cont_flag);
-int usocklnd_handle_zc_req(usock_peer_t *peer, __u64 cookie);
-int usocklnd_read_hello(usock_conn_t *conn, int *cont_flag);
-int usocklnd_activeconn_hellorecv(usock_conn_t *conn);
-int usocklnd_passiveconn_hellorecv(usock_conn_t *conn);
-int usocklnd_write_handler(usock_conn_t *conn);
-usock_tx_t *usocklnd_try_piggyback(struct list_head *tx_list_p,
- struct list_head *zcack_list_p);
-int usocklnd_activeconn_hellosent(usock_conn_t *conn);
-int usocklnd_passiveconn_hellosent(usock_conn_t *conn);
-int usocklnd_send_tx(usock_conn_t *conn, usock_tx_t *tx);
-int usocklnd_read_data(usock_conn_t *conn);
-
-void usocklnd_release_poll_states(int n);
-int usocklnd_base_startup();
-void usocklnd_base_shutdown(int n);
-__u64 usocklnd_new_incarnation();
-void usocklnd_del_all_peers(lnet_ni_t *ni);
-void usocklnd_del_peer_and_conns(usock_peer_t *peer);
-void usocklnd_del_conns_locked(usock_peer_t *peer);
-
-int usocklnd_conn_timed_out(usock_conn_t *conn, cfs_time_t current_time);
-void usocklnd_conn_kill(usock_conn_t *conn);
-void usocklnd_conn_kill_locked(usock_conn_t *conn);
-usock_conn_t *usocklnd_conn_allocate();
-void usocklnd_conn_free(usock_conn_t *conn);
-void usocklnd_tear_peer_conn(usock_conn_t *conn);
-void usocklnd_check_peer_stale(lnet_ni_t *ni, lnet_process_id_t id);
-int usocklnd_create_passive_conn(lnet_ni_t *ni,
- cfs_socket_t *sock, usock_conn_t **connp);
-int usocklnd_create_active_conn(usock_peer_t *peer, int type,
- usock_conn_t **connp);
-int usocklnd_connect_srv_mode(cfs_socket_t **sockp,
- __u32 dst_ip, __u16 dst_port);
-int usocklnd_connect_cli_mode(cfs_socket_t **sockp,
- __u32 dst_ip, __u16 dst_port);
-int usocklnd_set_sock_options(cfs_socket_t *sock);
-usock_tx_t *usocklnd_create_noop_tx(__u64 cookie);
-usock_tx_t *usocklnd_create_tx(lnet_msg_t *lntmsg);
-void usocklnd_init_hello_msg(ksock_hello_msg_t *hello,
- lnet_ni_t *ni, int type, lnet_nid_t peer_nid);
-usock_tx_t *usocklnd_create_hello_tx(lnet_ni_t *ni,
- int type, lnet_nid_t peer_nid);
-usock_tx_t *usocklnd_create_cr_hello_tx(lnet_ni_t *ni,
- int type, lnet_nid_t peer_nid);
-void usocklnd_destroy_tx(lnet_ni_t *ni, usock_tx_t *tx);
-void usocklnd_destroy_txlist(lnet_ni_t *ni, struct list_head *txlist);
-void usocklnd_destroy_zcack_list(struct list_head *zcack_list);
-void usocklnd_destroy_peer (usock_peer_t *peer);
-int usocklnd_get_conn_type(lnet_msg_t *lntmsg);
-int usocklnd_type2idx(int type);
-usock_peer_t *usocklnd_find_peer_locked(lnet_ni_t *ni, lnet_process_id_t id);
-int usocklnd_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
- usock_peer_t **peerp);
-int usocklnd_find_or_create_peer(lnet_ni_t *ni, lnet_process_id_t id,
- usock_peer_t **peerp);
-int usocklnd_find_or_create_conn(usock_peer_t *peer, int type,
- usock_conn_t **connp,
- usock_tx_t *tx, usock_zc_ack_t *zc_ack,
- int *send_immediately_flag);
-void usocklnd_link_conn_to_peer(usock_conn_t *conn, usock_peer_t *peer, int idx);
-int usocklnd_invert_type(int type);
-void usocklnd_conn_new_state(usock_conn_t *conn, int new_state);
-void usocklnd_cleanup_stale_conns(usock_peer_t *peer, __u64 incrn,
- usock_conn_t *skip_conn);
-
-void usocklnd_rx_hellomagic_state_transition(usock_conn_t *conn);
-void usocklnd_rx_helloversion_state_transition(usock_conn_t *conn);
-void usocklnd_rx_hellobody_state_transition(usock_conn_t *conn);
-void usocklnd_rx_helloIPs_state_transition(usock_conn_t *conn);
-void usocklnd_rx_lnethdr_state_transition(usock_conn_t *conn);
-void usocklnd_rx_ksmhdr_state_transition(usock_conn_t *conn);
-void usocklnd_rx_skipping_state_transition(usock_conn_t *conn);
+++ /dev/null
-/*
- * GPL HEADER START
- *
- * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
- *
- * This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License version 2 only,
- * as published by the Free Software Foundation.
- *
- * This program is distributed in the hope that it will be useful, but
- * WITHOUT ANY WARRANTY; without even the implied warranty of
- * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
- * General Public License version 2 for more details (a copy is included
- * in the LICENSE file that accompanied this code).
- *
- * You should have received a copy of the GNU General Public License
- * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
- *
- * GPL HEADER END
- */
-/*
- * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
- * Use is subject to license terms.
- */
-/*
- * This file is part of Lustre, http://www.lustre.org/
- * Lustre is a trademark of Sun Microsystems, Inc.
- *
- * lnet/ulnds/socklnd/usocklnd_cb.c
- *
- * Author: Maxim Patlasov <maxim@clusterfs.com>
- */
-
-#include "usocklnd.h"
-
-static int
-usocklnd_send_tx_immediately(usock_conn_t *conn, usock_tx_t *tx)
-{
- int rc;
- int rc2;
- int partial_send = 0;
- usock_peer_t *peer = conn->uc_peer;
-
- LASSERT (peer != NULL);
-
- /* usocklnd_enqueue_tx() turned it on for us */
- LASSERT(conn->uc_sending);
-
- //counter_imm_start++;
- rc = usocklnd_send_tx(conn, tx);
- if (rc == 0) { /* partial send or connection closed */
- pthread_mutex_lock(&conn->uc_lock);
- list_add(&tx->tx_list, &conn->uc_tx_list);
- conn->uc_sending = 0;
- pthread_mutex_unlock(&conn->uc_lock);
- partial_send = 1;
- } else {
- usocklnd_destroy_tx(peer->up_ni, tx);
- /* NB: lnetmsg was finalized, so we *must* return 0 */
-
- if (rc < 0) { /* real error */
- usocklnd_conn_kill(conn);
- return 0;
- }
-
- /* rc == 1: tx was sent completely */
- rc = 0; /* let's say to caller 'Ok' */
- //counter_imm_complete++;
- }
-
- pthread_mutex_lock(&conn->uc_lock);
- conn->uc_sending = 0;
-
- /* schedule write handler */
- if (partial_send ||
- (conn->uc_state == UC_READY &&
- (!list_empty(&conn->uc_tx_list) ||
- !list_empty(&conn->uc_zcack_list)))) {
- conn->uc_tx_deadline =
- cfs_time_shift(usock_tuns.ut_timeout);
- conn->uc_tx_flag = 1;
- rc2 = usocklnd_add_pollrequest(conn, POLL_TX_SET_REQUEST, POLLOUT);
- if (rc2 != 0)
- usocklnd_conn_kill_locked(conn);
- else
- usocklnd_wakeup_pollthread(conn->uc_pt_idx);
- }
-
- pthread_mutex_unlock(&conn->uc_lock);
-
- return rc;
-}
-
-int
-usocklnd_send(lnet_ni_t *ni, void *private, lnet_msg_t *lntmsg)
-{
- usock_tx_t *tx;
- lnet_process_id_t target = lntmsg->msg_target;
- usock_peer_t *peer;
- int type;
- int rc;
- usock_conn_t *conn;
- int send_immediately;
-
- tx = usocklnd_create_tx(lntmsg);
- if (tx == NULL)
- return -ENOMEM;
-
- rc = usocklnd_find_or_create_peer(ni, target, &peer);
- if (rc) {
- LIBCFS_FREE (tx, tx->tx_size);
- return rc;
- }
- /* peer cannot disappear now because its refcount was incremented */
-
- type = usocklnd_get_conn_type(lntmsg);
- rc = usocklnd_find_or_create_conn(peer, type, &conn, tx, NULL,
- &send_immediately);
- if (rc != 0) {
- usocklnd_peer_decref(peer);
- usocklnd_check_peer_stale(ni, target);
- LIBCFS_FREE (tx, tx->tx_size);
- return rc;
- }
- /* conn cannot disappear now because its refcount was incremented */
-
- if (send_immediately)
- rc = usocklnd_send_tx_immediately(conn, tx);
-
- usocklnd_conn_decref(conn);
- usocklnd_peer_decref(peer);
- return rc;
-}
-
-int
-usocklnd_recv(lnet_ni_t *ni, void *private, lnet_msg_t *msg, int delayed,
- unsigned int niov, struct iovec *iov, lnet_kiov_t *kiov,
- unsigned int offset, unsigned int mlen, unsigned int rlen)
-{
- int rc = 0;
- usock_conn_t *conn = (usock_conn_t *)private;
-
- /* I don't think that we'll win much concurrency moving lock()
- * call below lnet_extract_iov() */
- pthread_mutex_lock(&conn->uc_lock);
-
- conn->uc_rx_lnetmsg = msg;
- conn->uc_rx_nob_wanted = mlen;
- conn->uc_rx_nob_left = rlen;
- conn->uc_rx_iov = conn->uc_rx_iova;
- conn->uc_rx_niov =
- lnet_extract_iov(LNET_MAX_IOV, conn->uc_rx_iov,
- niov, iov, offset, mlen);
-
- /* the gap between lnet_parse() and usocklnd_recv() happened? */
- if (conn->uc_rx_state == UC_RX_PARSE_WAIT) {
- conn->uc_rx_flag = 1; /* waiting for incoming lnet payload */
- conn->uc_rx_deadline =
- cfs_time_shift(usock_tuns.ut_timeout);
- rc = usocklnd_add_pollrequest(conn, POLL_RX_SET_REQUEST, POLLIN);
- if (rc != 0) {
- usocklnd_conn_kill_locked(conn);
- goto recv_out;
- }
- usocklnd_wakeup_pollthread(conn->uc_pt_idx);
- }
-
- conn->uc_rx_state = UC_RX_LNET_PAYLOAD;
- recv_out:
- pthread_mutex_unlock(&conn->uc_lock);
- usocklnd_conn_decref(conn);
- return rc;
-}
-
-int
-usocklnd_accept(lnet_ni_t *ni, cfs_socket_t *sock)
-{
- int rc;
- usock_conn_t *conn;
-
- rc = usocklnd_create_passive_conn(ni, sock, &conn);
- if (rc)
- return rc;
- LASSERT(conn != NULL);
-
- /* disable shutdown event temporarily */
- lnet_ni_addref(ni);
-
- rc = usocklnd_add_pollrequest(conn, POLL_ADD_REQUEST, POLLIN);
- if (rc == 0)
- usocklnd_wakeup_pollthread(conn->uc_pt_idx);
-
- /* NB: conn reference counter was incremented while adding
- * poll request if rc == 0 */
-
- usocklnd_conn_decref(conn); /* should destroy conn if rc != 0 */
- return rc;
-}
$(LIBCFSUTIL) $(LIBCFS) $(LIBREADLINE) $(LIBEFENCE)
lnetctl_DEPENDENCIES = $(top_builddir)/lnet/utils/lnetconfig/liblnetconfig.so
endif # UTILS
-
-LND_LIBS =
-if BUILD_USOCKLND
-LND_LIBS += $(top_builddir)/lnet/ulnds/socklnd/libsocklnd.a
-endif
-
-EXTRA_DIST = genlib.sh
+++ /dev/null
-#!/bin/bash
-#set -xv
-set -e
-
-CWD=`pwd`
-
-LIBS=$1
-LND_LIBS=$2
-PTHREAD_LIBS=$3
-
-# do cleanup at first
-rm -f liblst.so
-
-ALL_OBJS=
-
-build_obj_list() {
- _objs=`$AR -t $1/$2 | grep -v SYMDEF | grep -v SORTED`
- for _lib in $_objs; do
- ALL_OBJS=$ALL_OBJS"$1/$_lib ";
- done;
-}
-
-# lnet components libs
-build_obj_list ../../libcfs/libcfs libcfs.a
-build_obj_list ../../libcfs/libcfs libcfsutil.a
-if $(echo "$LND_LIBS" | grep "socklnd" >/dev/null) ; then
- build_obj_list ../../lnet/ulnds/socklnd libsocklnd.a
-fi
-build_obj_list ../../lnet/lnet liblnet.a
-build_obj_list ../../lnet/selftest libselftest.a
-
-# create static lib lustre
-rm -f $CWD/liblst.a
-$AR -cru $CWD/liblst.a $ALL_OBJS
-$RANLIB $CWD/liblst.a
git://git.whamcloud.com / fs / lustre-release.git / commitdiff