diff options
Diffstat (limited to 'rpc/rpc-transport')
| -rw-r--r-- | rpc/rpc-transport/Makefile.am | 2 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/Makefile.am | 1 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/Makefile.am | 24 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/name.c | 715 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/name.h | 35 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/rdma.c | 4904 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/rdma.h | 384 | ||||
| -rw-r--r-- | rpc/rpc-transport/rdma/src/rpc-trans-rdma-messages.h | 66 | ||||
| -rw-r--r-- | rpc/rpc-transport/socket/src/name.c | 114 | ||||
| -rw-r--r-- | rpc/rpc-transport/socket/src/name.h | 2 | ||||
| -rw-r--r-- | rpc/rpc-transport/socket/src/socket-mem-types.h | 3 | ||||
| -rw-r--r-- | rpc/rpc-transport/socket/src/socket.c | 1134 | ||||
| -rw-r--r-- | rpc/rpc-transport/socket/src/socket.h | 108 |
13 files changed, 734 insertions, 6758 deletions
diff --git a/rpc/rpc-transport/Makefile.am b/rpc/rpc-transport/Makefile.am index 221fd640514..7dd9f026cfc 100644 --- a/rpc/rpc-transport/Makefile.am +++ b/rpc/rpc-transport/Makefile.am @@ -1 +1 @@ -SUBDIRS = socket $(RDMA_SUBDIR) +SUBDIRS = socket diff --git a/rpc/rpc-transport/rdma/Makefile.am b/rpc/rpc-transport/rdma/Makefile.am deleted file mode 100644 index f963effea22..00000000000 --- a/rpc/rpc-transport/rdma/Makefile.am +++ /dev/null @@ -1 +0,0 @@ -SUBDIRS = src
\ No newline at end of file diff --git a/rpc/rpc-transport/rdma/src/Makefile.am b/rpc/rpc-transport/rdma/src/Makefile.am deleted file mode 100644 index 40b5a19d3d1..00000000000 --- a/rpc/rpc-transport/rdma/src/Makefile.am +++ /dev/null @@ -1,24 +0,0 @@ -# TODO : need to change transportdir - -transport_LTLIBRARIES = rdma.la -transportdir = $(libdir)/glusterfs/$(PACKAGE_VERSION)/rpc-transport - -rdma_la_LDFLAGS = -module -avoid-version -nostartfiles - -rdma_la_SOURCES = rdma.c name.c -rdma_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la \ - $(top_builddir)/rpc/xdr/src/libgfxdr.la \ - $(top_builddir)/rpc/rpc-lib/src/libgfrpc.la \ - -libverbs -lrdmacm - -noinst_HEADERS = rdma.h name.h rpc-trans-rdma-messages.h - -AM_CPPFLAGS = $(GF_CPPFLAGS) \ - -I$(top_srcdir)/libglusterfs/src \ - -I$(top_srcdir)/rpc/rpc-lib/src/ \ - -I$(top_srcdir)/rpc/xdr/src \ - -I$(top_builddir)/rpc/xdr/src - -AM_CFLAGS = -Wall $(GF_CFLAGS) - -CLEANFILES = *~ diff --git a/rpc/rpc-transport/rdma/src/name.c b/rpc/rpc-transport/rdma/src/name.c deleted file mode 100644 index ea960cba4eb..00000000000 --- a/rpc/rpc-transport/rdma/src/name.c +++ /dev/null @@ -1,715 +0,0 @@ -/* - Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> - This file is part of GlusterFS. - - This file is licensed to you under your choice of the GNU Lesser - General Public License, version 3 or any later version (LGPLv3 or - later), or the GNU General Public License, version 2 (GPLv2), in all - cases as published by the Free Software Foundation. -*/ - -#include <sys/types.h> -#include <sys/socket.h> -#include <errno.h> -#include <netdb.h> -#include <string.h> -#include <rdma/rdma_cma.h> - -#ifndef AF_INET_SDP -#define AF_INET_SDP 27 -#endif - -#include "rpc-transport.h" -#include "rdma.h" -#include "common-utils.h" -#include "rpc-lib-messages.h" -#include "rpc-trans-rdma-messages.h" - -int32_t -gf_resolve_ip6(const char *hostname, uint16_t port, int family, void **dnscache, - struct addrinfo **addr_info); - -static void -_assign_port(struct sockaddr *sockaddr, uint16_t port) -{ - switch (sockaddr->sa_family) { - case AF_INET6: - ((struct sockaddr_in6 *)sockaddr)->sin6_port = htons(port); - break; - - case AF_INET_SDP: - case AF_INET: - ((struct sockaddr_in *)sockaddr)->sin_port = htons(port); - break; - } -} - -static int32_t -af_inet_bind_to_port_lt_ceiling(struct rdma_cm_id *cm_id, - struct sockaddr *sockaddr, - socklen_t sockaddr_len, uint32_t ceiling) -{ -#if GF_DISABLE_PRIVPORT_TRACKING - _assign_port(sockaddr, 0); - return rdma_bind_addr(cm_id, sockaddr); -#else - int32_t ret = -1; - uint16_t port = ceiling - 1; - unsigned char ports[GF_PORT_ARRAY_SIZE] = { - 0, - }; - int i = 0; - -loop: - ret = gf_process_reserved_ports(ports, ceiling); - - while (port) { - if (port == GF_CLIENT_PORT_CEILING) { - ret = -1; - break; - } - - /* ignore the reserved ports */ - if (BIT_VALUE(ports, port)) { - port--; - continue; - } - - _assign_port(sockaddr, port); - - ret = rdma_bind_addr(cm_id, sockaddr); - - if (ret == 0) - break; - - if (ret == -1 && errno == EACCES) - break; - - port--; - } - - /* In case if all the secure ports are exhausted, we are no more - * binding to secure ports, hence instead of getting a random - * port, lets define the range to restrict it from getting from - * ports reserved for bricks i.e from range of 49152 - 65535 - * which further may lead to port clash */ - if (!port) { - ceiling = port = GF_CLNT_INSECURE_PORT_CEILING; - for (i = 0; i <= ceiling; i++) - BIT_CLEAR(ports, i); - goto loop; - } - - return ret; -#endif /* GF_DISABLE_PRIVPORT_TRACKING */ -} - -#if 0 -static int32_t -af_unix_client_bind (rpc_transport_t *this, struct sockaddr *sockaddr, - socklen_t sockaddr_len, struct rdma_cm_id *cm_id) -{ - data_t *path_data = NULL; - struct sockaddr_un *addr = NULL; - int32_t ret = -1; - - path_data = dict_get (this->options, - "transport.rdma.bind-path"); - if (path_data) { - char *path = data_to_str (path_data); - if (!path || strlen (path) > UNIX_PATH_MAX) { - gf_msg_debug (this->name, 0, - "transport.rdma.bind-path not specified " - "for unix socket, letting connect to " - "assign default value"); - goto err; - } - - addr = (struct sockaddr_un *) sockaddr; - strcpy (addr->sun_path, path); - ret = bind (sock, (struct sockaddr *)addr, sockaddr_len); - if (ret == -1) { - gf_msg (this->name, GF_LOG_ERROR, errno, - TRANS_MSG_SOCKET_BIND_ERROR, - "cannot bind to unix-domain socket %d ", - sock); - goto err; - } - } - -err: - return ret; -} -#endif - -static int32_t -client_fill_address_family(rpc_transport_t *this, struct sockaddr *sockaddr) -{ - data_t *address_family_data = NULL; - - address_family_data = dict_get(this->options, "transport.address-family"); - if (!address_family_data) { - data_t *remote_host_data = NULL, *connect_path_data = NULL; - remote_host_data = dict_get(this->options, "remote-host"); - connect_path_data = dict_get(this->options, - "transport.rdma.connect-path"); - - if (!(remote_host_data || connect_path_data) || - (remote_host_data && connect_path_data)) { - gf_msg(this->name, GF_LOG_ERROR, 0, - TRANS_MSG_ADDR_FAMILY_NOT_SPECIFIED, - "address-family not specified and not able to " - "determine the same from other options " - "(remote-host:%s and connect-path:%s)", - data_to_str(remote_host_data), - data_to_str(connect_path_data)); - return -1; - } - - if (remote_host_data) { - gf_msg_debug(this->name, 0, - "address-family not " - "specified, guessing it to be " - "inet/inet6"); - sockaddr->sa_family = AF_UNSPEC; - } else { - gf_msg_debug(this->name, 0, - "address-family not " - "specified, guessing it to be unix"); - sockaddr->sa_family = AF_UNIX; - } - - } else { - char *address_family = data_to_str(address_family_data); - if (!strcasecmp(address_family, "unix")) { - sockaddr->sa_family = AF_UNIX; - } else if (!strcasecmp(address_family, "inet")) { - sockaddr->sa_family = AF_INET; - } else if (!strcasecmp(address_family, "inet6")) { - sockaddr->sa_family = AF_INET6; - } else if (!strcasecmp(address_family, "inet-sdp")) { - sockaddr->sa_family = AF_INET_SDP; - } else { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_UNKNOWN_ADDR_FAMILY, - "unknown address-family (%s) specified", address_family); - sockaddr->sa_family = AF_UNSPEC; - return -1; - } - } - - return 0; -} - -static int32_t -af_inet_client_get_remote_sockaddr(rpc_transport_t *this, - struct sockaddr *sockaddr, - socklen_t *sockaddr_len, int16_t remote_port) -{ - dict_t *options = this->options; - data_t *remote_host_data = NULL; - data_t *remote_port_data = NULL; - char *remote_host = NULL; - struct addrinfo *addr_info = NULL; - int32_t ret = 0; - - remote_host_data = dict_get(options, "remote-host"); - if (remote_host_data == NULL) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_REMOTE_HOST_ERROR, - "option remote-host " - "missing in volume %s", - this->name); - ret = -1; - goto err; - } - - remote_host = data_to_str(remote_host_data); - if (remote_host == NULL) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_REMOTE_HOST_ERROR, - "option remote-host " - "has data NULL in volume %s", - this->name); - ret = -1; - goto err; - } - - if (remote_port == 0) { - remote_port_data = dict_get(options, "remote-port"); - if (remote_port_data == NULL) { - gf_msg_debug(this->name, 0, - "option remote-port " - "missing in volume %s. Defaulting to %d", - this->name, GF_DEFAULT_RDMA_LISTEN_PORT); - - remote_port = GF_DEFAULT_RDMA_LISTEN_PORT; - } else { - remote_port = data_to_uint16(remote_port_data); - } - } - - if (remote_port == -1) { - gf_msg(this->name, GF_LOG_ERROR, EINVAL, RDMA_MSG_INVALID_ENTRY, - "option remote-port has " - "invalid port in volume %s", - this->name); - ret = -1; - goto err; - } - - /* TODO: gf_resolve is a blocking call. kick in some - non blocking dns techniques */ - ret = gf_resolve_ip6(remote_host, remote_port, sockaddr->sa_family, - &this->dnscache, &addr_info); - if (ret == -1) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_DNS_RESOL_FAILED, - "DNS resolution failed on host %s", remote_host); - goto err; - } - - memcpy(sockaddr, addr_info->ai_addr, addr_info->ai_addrlen); - *sockaddr_len = addr_info->ai_addrlen; - -err: - return ret; -} - -static int32_t -af_unix_client_get_remote_sockaddr(rpc_transport_t *this, - struct sockaddr *sockaddr, - socklen_t *sockaddr_len) -{ - struct sockaddr_un *sockaddr_un = NULL; - char *connect_path = NULL; - data_t *connect_path_data = NULL; - int32_t ret = 0; - - connect_path_data = dict_get(this->options, "transport.rdma.connect-path"); - if (!connect_path_data) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_CONNECT_PATH_ERROR, - "option " - "transport.rdma.connect-path not specified for " - "address-family unix"); - ret = -1; - goto err; - } - - connect_path = data_to_str(connect_path_data); - if (!connect_path) { - gf_msg(this->name, GF_LOG_ERROR, EINVAL, RDMA_MSG_INVALID_ENTRY, - "connect-path is null-string"); - ret = -1; - goto err; - } - - if (strlen(connect_path) > UNIX_PATH_MAX) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_CONNECT_PATH_ERROR, - "connect-path value length %" GF_PRI_SIZET - " > " - "%d octets", - strlen(connect_path), UNIX_PATH_MAX); - ret = -1; - goto err; - } - - gf_msg_debug(this->name, 0, "using connect-path %s", connect_path); - sockaddr_un = (struct sockaddr_un *)sockaddr; - strcpy(sockaddr_un->sun_path, connect_path); - *sockaddr_len = sizeof(struct sockaddr_un); - -err: - return ret; -} - -static int32_t -af_unix_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, - socklen_t *addr_len) -{ - data_t *listen_path_data = NULL; - char *listen_path = NULL; - int32_t ret = 0; - struct sockaddr_un *sunaddr = (struct sockaddr_un *)addr; - - listen_path_data = dict_get(this->options, "transport.rdma.listen-path"); - if (!listen_path_data) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_LISTEN_PATH_ERROR, - "missing option listen-path"); - ret = -1; - goto err; - } - - listen_path = data_to_str(listen_path_data); - -#ifndef UNIX_PATH_MAX -#define UNIX_PATH_MAX 108 -#endif - - if (strlen(listen_path) > UNIX_PATH_MAX) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_LISTEN_PATH_ERROR, - "option listen-path has " - "value length %" GF_PRI_SIZET " > %d", - strlen(listen_path), UNIX_PATH_MAX); - ret = -1; - goto err; - } - - sunaddr->sun_family = AF_UNIX; - strcpy(sunaddr->sun_path, listen_path); - *addr_len = sizeof(struct sockaddr_un); - -err: - return ret; -} - -static int32_t -af_inet_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, - socklen_t *addr_len) -{ - struct addrinfo hints, *res = 0; - data_t *listen_port_data = NULL, *listen_host_data = NULL; - uint16_t listen_port = -1; - char service[NI_MAXSERV], *listen_host = NULL; - dict_t *options = NULL; - int32_t ret = 0; - - options = this->options; - - listen_port_data = dict_get(options, "transport.rdma.listen-port"); - listen_host_data = dict_get(options, "transport.rdma.bind-address"); - - if (listen_port_data) { - listen_port = data_to_uint16(listen_port_data); - } else { - listen_port = GF_DEFAULT_RDMA_LISTEN_PORT; - - if (addr->sa_family == AF_INET6) { - struct sockaddr_in6 *in = (struct sockaddr_in6 *)addr; - in->sin6_addr = in6addr_any; - in->sin6_port = htons(listen_port); - *addr_len = sizeof(struct sockaddr_in6); - goto out; - } else if (addr->sa_family == AF_INET) { - struct sockaddr_in *in = (struct sockaddr_in *)addr; - in->sin_addr.s_addr = htonl(INADDR_ANY); - in->sin_port = htons(listen_port); - *addr_len = sizeof(struct sockaddr_in); - goto out; - } - } - - if (listen_port == (uint16_t)-1) - listen_port = GF_DEFAULT_RDMA_LISTEN_PORT; - - if (listen_host_data) { - listen_host = data_to_str(listen_host_data); - } - - sprintf(service, "%d", listen_port); - - memset(&hints, 0, sizeof(hints)); - hints.ai_family = addr->sa_family; - hints.ai_socktype = SOCK_STREAM; - hints.ai_flags = AI_ADDRCONFIG | AI_PASSIVE; - - ret = getaddrinfo(listen_host, service, &hints, &res); - if (ret != 0) { - gf_msg(this->name, GF_LOG_ERROR, ret, TRANS_MSG_GET_ADDR_INFO_FAILED, - "getaddrinfo failed for host %s, service %s", listen_host, - service); - ret = -1; - goto out; - } - - memcpy(addr, res->ai_addr, res->ai_addrlen); - *addr_len = res->ai_addrlen; - - freeaddrinfo(res); - -out: - return ret; -} - -int32_t -gf_rdma_client_bind(rpc_transport_t *this, struct sockaddr *sockaddr, - socklen_t *sockaddr_len, struct rdma_cm_id *cm_id) -{ - int ret = 0; - - *sockaddr_len = sizeof(struct sockaddr_in6); - switch (sockaddr->sa_family) { - case AF_INET_SDP: - case AF_INET: - *sockaddr_len = sizeof(struct sockaddr_in); - /* Fall through */ - case AF_INET6: - if (!this->bind_insecure) { - ret = af_inet_bind_to_port_lt_ceiling( - cm_id, sockaddr, *sockaddr_len, GF_CLIENT_PORT_CEILING); - if (ret == -1) { - gf_msg(this->name, GF_LOG_WARNING, errno, - RDMA_MSG_PORT_BIND_FAILED, - "cannot bind rdma_cm_id to port " - "less than %d", - GF_CLIENT_PORT_CEILING); - } - } else { - ret = af_inet_bind_to_port_lt_ceiling( - cm_id, sockaddr, *sockaddr_len, GF_IANA_PRIV_PORTS_START); - if (ret == -1) { - gf_msg(this->name, GF_LOG_WARNING, errno, - RDMA_MSG_PORT_BIND_FAILED, - "cannot bind rdma_cm_id to port " - "less than %d", - GF_IANA_PRIV_PORTS_START); - } - } - break; - - case AF_UNIX: - *sockaddr_len = sizeof(struct sockaddr_un); -#if 0 - ret = af_unix_client_bind (this, (struct sockaddr *)sockaddr, - *sockaddr_len, sock); -#endif - break; - - default: - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_UNKNOWN_ADDR_FAMILY, - "unknown address family %d", sockaddr->sa_family); - ret = -1; - break; - } - - return ret; -} - -int32_t -gf_rdma_client_get_remote_sockaddr(rpc_transport_t *this, - struct sockaddr *sockaddr, - socklen_t *sockaddr_len, int16_t remote_port) -{ - int32_t ret = 0; - char is_inet_sdp = 0; - - ret = client_fill_address_family(this, sockaddr); - if (ret) { - ret = -1; - goto err; - } - - switch (sockaddr->sa_family) { - case AF_INET_SDP: - sockaddr->sa_family = AF_INET; - is_inet_sdp = 1; - /* Fall through */ - case AF_INET: - case AF_INET6: - case AF_UNSPEC: - ret = af_inet_client_get_remote_sockaddr(this, sockaddr, - sockaddr_len, remote_port); - - if (is_inet_sdp) { - sockaddr->sa_family = AF_INET_SDP; - } - - break; - - case AF_UNIX: - ret = af_unix_client_get_remote_sockaddr(this, sockaddr, - sockaddr_len); - break; - - default: - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_UNKNOWN_ADDR_FAMILY, - "unknown address-family %d", sockaddr->sa_family); - ret = -1; - } - -err: - return ret; -} - -int32_t -gf_rdma_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, - socklen_t *addr_len) -{ - data_t *address_family_data = NULL; - int32_t ret = 0; - char is_inet_sdp = 0; - - address_family_data = dict_get(this->options, "transport.address-family"); - if (address_family_data) { - char *address_family = NULL; - address_family = data_to_str(address_family_data); - - if (!strcasecmp(address_family, "inet")) { - addr->sa_family = AF_INET; - } else if (!strcasecmp(address_family, "inet6")) { - addr->sa_family = AF_INET6; - } else if (!strcasecmp(address_family, "inet-sdp")) { - addr->sa_family = AF_INET_SDP; - } else if (!strcasecmp(address_family, "unix")) { - addr->sa_family = AF_UNIX; - } else { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_UNKNOWN_ADDR_FAMILY, - "unknown address" - " family (%s) specified", - address_family); - addr->sa_family = AF_UNSPEC; - ret = -1; - goto err; - } - } else { - gf_msg_debug(this->name, 0, - "option address-family not " - "specified, defaulting to inet"); - addr->sa_family = AF_INET; - } - - switch (addr->sa_family) { - case AF_INET_SDP: - is_inet_sdp = 1; - addr->sa_family = AF_INET; - /* Fall through */ - case AF_INET: - case AF_INET6: - case AF_UNSPEC: - ret = af_inet_server_get_local_sockaddr(this, addr, addr_len); - if (is_inet_sdp && !ret) { - addr->sa_family = AF_INET_SDP; - } - break; - - case AF_UNIX: - ret = af_unix_server_get_local_sockaddr(this, addr, addr_len); - break; - } - -err: - return ret; -} - -int32_t -fill_inet6_inet_identifiers(rpc_transport_t *this, - struct sockaddr_storage *addr, int32_t addr_len, - char *identifier) -{ - int32_t ret = 0, tmpaddr_len = 0; - char service[NI_MAXSERV], host[NI_MAXHOST]; - union gf_sock_union sock_union; - - memset(&sock_union, 0, sizeof(sock_union)); - sock_union.storage = *addr; - tmpaddr_len = addr_len; - - if (sock_union.sa.sa_family == AF_INET6) { - int32_t one_to_four, four_to_eight, twelve_to_sixteen; - int16_t eight_to_ten, ten_to_twelve; - - one_to_four = four_to_eight = twelve_to_sixteen = 0; - eight_to_ten = ten_to_twelve = 0; - - one_to_four = sock_union.sin6.sin6_addr.s6_addr32[0]; - four_to_eight = sock_union.sin6.sin6_addr.s6_addr32[1]; -#ifdef GF_SOLARIS_HOST_OS - eight_to_ten = S6_ADDR16(sock_union.sin6.sin6_addr)[4]; -#else - eight_to_ten = sock_union.sin6.sin6_addr.s6_addr16[4]; -#endif - -#ifdef GF_SOLARIS_HOST_OS - ten_to_twelve = S6_ADDR16(sock_union.sin6.sin6_addr)[5]; -#else - ten_to_twelve = sock_union.sin6.sin6_addr.s6_addr16[5]; -#endif - twelve_to_sixteen = sock_union.sin6.sin6_addr.s6_addr32[3]; - - /* ipv4 mapped ipv6 address has - bits 0-80: 0 - bits 80-96: 0xffff - bits 96-128: ipv4 address - */ - - if (one_to_four == 0 && four_to_eight == 0 && eight_to_ten == 0 && - ten_to_twelve == -1) { - struct sockaddr_in *in_ptr = &sock_union.sin; - memset(&sock_union, 0, sizeof(sock_union)); - - in_ptr->sin_family = AF_INET; - in_ptr->sin_port = ((struct sockaddr_in6 *)addr)->sin6_port; - in_ptr->sin_addr.s_addr = twelve_to_sixteen; - tmpaddr_len = sizeof(*in_ptr); - } - } - - ret = getnameinfo(&sock_union.sa, tmpaddr_len, host, sizeof(host), service, - sizeof(service), NI_NUMERICHOST | NI_NUMERICSERV); - if (ret != 0) { - gf_msg(this->name, GF_LOG_ERROR, ret, TRANS_MSG_GET_NAME_INFO_FAILED, - "getnameinfo failed"); - } - - sprintf(identifier, "%s:%s", host, service); - - return ret; -} - -int32_t -gf_rdma_get_transport_identifiers(rpc_transport_t *this) -{ - int32_t ret = 0; - char is_inet_sdp = 0; - - switch (((struct sockaddr *)&this->myinfo.sockaddr)->sa_family) { - case AF_INET_SDP: - is_inet_sdp = 1; - ((struct sockaddr *)&this->peerinfo.sockaddr) - ->sa_family = ((struct sockaddr *)&this->myinfo.sockaddr) - ->sa_family = AF_INET; - /* Fall through */ - case AF_INET: - case AF_INET6: { - ret = fill_inet6_inet_identifiers(this, &this->myinfo.sockaddr, - this->myinfo.sockaddr_len, - this->myinfo.identifier); - if (ret == -1) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_INET_ERROR, - "can't fill inet/inet6 identifier for server"); - goto err; - } - - ret = fill_inet6_inet_identifiers(this, &this->peerinfo.sockaddr, - this->peerinfo.sockaddr_len, - this->peerinfo.identifier); - if (ret == -1) { - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_INET_ERROR, - "can't fill inet/inet6 identifier for client"); - goto err; - } - - if (is_inet_sdp) { - ((struct sockaddr *)&this->peerinfo.sockaddr) - ->sa_family = ((struct sockaddr *)&this->myinfo.sockaddr) - ->sa_family = AF_INET_SDP; - } - } break; - - case AF_UNIX: { - struct sockaddr_un *sunaddr = NULL; - - sunaddr = (struct sockaddr_un *)&this->myinfo.sockaddr; - strcpy(this->myinfo.identifier, sunaddr->sun_path); - - sunaddr = (struct sockaddr_un *)&this->peerinfo.sockaddr; - strcpy(this->peerinfo.identifier, sunaddr->sun_path); - } break; - - default: - gf_msg(this->name, GF_LOG_ERROR, 0, TRANS_MSG_UNKNOWN_ADDR_FAMILY, - "unknown address family (%d)", - ((struct sockaddr *)&this->myinfo.sockaddr)->sa_family); - ret = -1; - break; - } - -err: - return ret; -} diff --git a/rpc/rpc-transport/rdma/src/name.h b/rpc/rpc-transport/rdma/src/name.h deleted file mode 100644 index 6c782b52db5..00000000000 --- a/rpc/rpc-transport/rdma/src/name.h +++ /dev/null @@ -1,35 +0,0 @@ -/* - Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> - This file is part of GlusterFS. - - This file is licensed to you under your choice of the GNU Lesser - General Public License, version 3 or any later version (LGPLv3 or - later), or the GNU General Public License, version 2 (GPLv2), in all - cases as published by the Free Software Foundation. -*/ - -#ifndef _IB_VERBS_NAME_H -#define _IB_VERBS_NAME_H - -#include <rdma/rdma_cma.h> - -#include "compat.h" - -int32_t -gf_rdma_client_bind(rpc_transport_t *this, struct sockaddr *sockaddr, - socklen_t *sockaddr_len, struct rdma_cm_id *cm_id); - -int32_t -gf_rdma_client_get_remote_sockaddr(rpc_transport_t *this, - struct sockaddr *sockaddr, - socklen_t *sockaddr_len, - int16_t remote_port); - -int32_t -gf_rdma_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, - socklen_t *addr_len); - -int32_t -gf_rdma_get_transport_identifiers(rpc_transport_t *this); - -#endif /* _IB_VERBS_NAME_H */ diff --git a/rpc/rpc-transport/rdma/src/rdma.c b/rpc/rpc-transport/rdma/src/rdma.c deleted file mode 100644 index 92c9ce95929..00000000000 --- a/rpc/rpc-transport/rdma/src/rdma.c +++ /dev/null @@ -1,4904 +0,0 @@ -/* - Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> - This file is part of GlusterFS. - - This file is licensed to you under your choice of the GNU Lesser - General Public License, version 3 or any later version (LGPLv3 or - later), or the GNU General Public License, version 2 (GPLv2), in all - cases as published by the Free Software Foundation. -*/ - -#include "dict.h" -#include "glusterfs.h" -#include "iobuf.h" -#include "logging.h" -#include "rdma.h" -#include "name.h" -#include "byte-order.h" -#include "xlator.h" -#include "xdr-rpc.h" -#include "rpc-lib-messages.h" -#include "rpc-trans-rdma-messages.h" -#include <signal.h> - -#define GF_RDMA_LOG_NAME "rpc-transport/rdma" - -static int32_t -__gf_rdma_ioq_churn(gf_rdma_peer_t *peer); - -gf_rdma_post_t * -gf_rdma_post_ref(gf_rdma_post_t *post); - -int -gf_rdma_post_unref(gf_rdma_post_t *post); - -static void * -gf_rdma_send_completion_proc(void *data); - -static void * -gf_rdma_recv_completion_proc(void *data); - -void * -gf_rdma_async_event_thread(void *context); - -static int32_t -gf_rdma_create_qp(rpc_transport_t *this); - -static int32_t -__gf_rdma_teardown(rpc_transport_t *this); - -static int32_t -gf_rdma_teardown(rpc_transport_t *this); - -static int32_t -gf_rdma_disconnect(rpc_transport_t *this, gf_boolean_t wait); - -static void -gf_rdma_cm_handle_disconnect(rpc_transport_t *this); - -static int -gf_rdma_cm_handle_connect_init(struct rdma_cm_event *event); - -static void -gf_rdma_put_post(gf_rdma_queue_t *queue, gf_rdma_post_t *post) -{ - post->ctx.is_request = 0; - - pthread_mutex_lock(&queue->lock); - { - if (post->prev) { - queue->active_count--; - post->prev->next = post->next; - } - - if (post->next) { - post->next->prev = post->prev; - } - - post->prev = &queue->passive_posts; - post->next = post->prev->next; - post->prev->next = post; - post->next->prev = post; - queue->passive_count++; - } - pthread_mutex_unlock(&queue->lock); -} - -static gf_rdma_post_t * -gf_rdma_new_post(rpc_transport_t *this, gf_rdma_device_t *device, int32_t len, - gf_rdma_post_type_t type) -{ - gf_rdma_post_t *post = NULL; - int ret = -1; - - post = (gf_rdma_post_t *)GF_CALLOC(1, sizeof(*post), - gf_common_mt_rdma_post_t); - if (post == NULL) { - goto out; - } - - pthread_mutex_init(&post->lock, NULL); - - post->buf_size = len; - - post->buf = valloc(len); - if (!post->buf) { - gf_msg_nomem(GF_RDMA_LOG_NAME, GF_LOG_ERROR, len); - goto out; - } - - post->mr = ibv_reg_mr(device->pd, post->buf, post->buf_size, - IBV_ACCESS_LOCAL_WRITE); - if (!post->mr) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_MR_ALOC_FAILED, - "memory registration failed"); - goto out; - } - - post->device = device; - post->type = type; - - ret = 0; -out: - if (ret != 0 && post) { - free(post->buf); - - GF_FREE(post); - post = NULL; - } - - return post; -} - -static gf_rdma_post_t * -gf_rdma_get_post(gf_rdma_queue_t *queue) -{ - gf_rdma_post_t *post = NULL; - - pthread_mutex_lock(&queue->lock); - { - post = queue->passive_posts.next; - if (post == &queue->passive_posts) - post = NULL; - - if (post) { - if (post->prev) - post->prev->next = post->next; - if (post->next) - post->next->prev = post->prev; - post->prev = &queue->active_posts; - post->next = post->prev->next; - post->prev->next = post; - post->next->prev = post; - post->reused++; - queue->active_count++; - } - } - pthread_mutex_unlock(&queue->lock); - - return post; -} - -void -gf_rdma_destroy_post(gf_rdma_post_t *post) -{ - ibv_dereg_mr(post->mr); - free(post->buf); - GF_FREE(post); -} - -static int32_t -__gf_rdma_quota_get(gf_rdma_peer_t *peer) -{ - int32_t ret = -1; - gf_rdma_private_t *priv = NULL; - - priv = peer->trans->private; - - if (priv->connected && peer->quota > 0) { - ret = peer->quota--; - } - - return ret; -} - -static void -__gf_rdma_ioq_entry_free(gf_rdma_ioq_t *entry) -{ - list_del_init(&entry->list); - - if (entry->iobref) { - iobref_unref(entry->iobref); - entry->iobref = NULL; - } - - if (entry->msg.request.rsp_iobref) { - iobref_unref(entry->msg.request.rsp_iobref); - entry->msg.request.rsp_iobref = NULL; - } - - mem_put(entry); -} - -static void -__gf_rdma_ioq_flush(gf_rdma_peer_t *peer) -{ - gf_rdma_ioq_t *entry = NULL, *dummy = NULL; - - list_for_each_entry_safe(entry, dummy, &peer->ioq, list) - { - __gf_rdma_ioq_entry_free(entry); - } -} - -static int32_t -__gf_rdma_disconnect(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - - priv = this->private; - - if (priv->connected) { - rdma_disconnect(priv->peer.cm_id); - } - - return 0; -} - -static void -gf_rdma_queue_init(gf_rdma_queue_t *queue) -{ - pthread_mutex_init(&queue->lock, NULL); - - queue->active_posts.next = &queue->active_posts; - queue->active_posts.prev = &queue->active_posts; - queue->passive_posts.next = &queue->passive_posts; - queue->passive_posts.prev = &queue->passive_posts; -} - -static void -__gf_rdma_destroy_queue(gf_rdma_post_t *post) -{ - gf_rdma_post_t *tmp = NULL; - - while (post->next != post) { - tmp = post->next; - - post->next = post->next->next; - post->next->prev = post; - - gf_rdma_destroy_post(tmp); - } -} - -static void -gf_rdma_destroy_queue(gf_rdma_queue_t *queue) -{ - if (queue == NULL) { - goto out; - } - - pthread_mutex_lock(&queue->lock); - { - if (queue->passive_count > 0) { - __gf_rdma_destroy_queue(&queue->passive_posts); - queue->passive_count = 0; - } - - if (queue->active_count > 0) { - __gf_rdma_destroy_queue(&queue->active_posts); - queue->active_count = 0; - } - } - pthread_mutex_unlock(&queue->lock); - -out: - return; -} - -static void -gf_rdma_destroy_posts(rpc_transport_t *this) -{ - gf_rdma_device_t *device = NULL; - gf_rdma_private_t *priv = NULL; - - if (this == NULL) { - goto out; - } - - priv = this->private; - device = priv->device; - - gf_rdma_destroy_queue(&device->sendq); - gf_rdma_destroy_queue(&device->recvq); - -out: - return; -} - -static int32_t -__gf_rdma_create_posts(rpc_transport_t *this, int32_t count, int32_t size, - gf_rdma_queue_t *q, gf_rdma_post_type_t type) -{ - int32_t i = 0; - int32_t ret = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - - priv = this->private; - device = priv->device; - - for (i = 0; i < count; i++) { - gf_rdma_post_t *post = NULL; - - post = gf_rdma_new_post(this, device, size + 2048, type); - if (!post) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_POST_CREATION_FAILED, - "post creation failed"); - ret = -1; - break; - } - - gf_rdma_put_post(q, post); - } - return ret; -} - -static int32_t -gf_rdma_post_recv(struct ibv_srq *srq, gf_rdma_post_t *post) -{ - struct ibv_sge list = {.addr = (unsigned long)post->buf, - .length = post->buf_size, - .lkey = post->mr->lkey}; - - struct ibv_recv_wr wr = - { - .wr_id = (unsigned long)post, - .sg_list = &list, - .num_sge = 1, - }, - *bad_wr; - - gf_rdma_post_ref(post); - - return ibv_post_srq_recv(srq, &wr, &bad_wr); -} - -static void -gf_rdma_deregister_iobuf_pool(gf_rdma_device_t *device) -{ - gf_rdma_arena_mr *arena_mr = NULL; - gf_rdma_arena_mr *tmp = NULL; - - while (device) { - pthread_mutex_lock(&device->all_mr_lock); - { - if (!list_empty(&device->all_mr)) { - list_for_each_entry_safe(arena_mr, tmp, &device->all_mr, list) - { - if (ibv_dereg_mr(arena_mr->mr)) { - gf_msg("rdma", GF_LOG_WARNING, 0, - RDMA_MSG_DEREGISTER_ARENA_FAILED, - "deallocation of memory region " - "failed"); - pthread_mutex_unlock(&device->all_mr_lock); - return; - } - list_del(&arena_mr->list); - GF_FREE(arena_mr); - } - } - } - pthread_mutex_unlock(&device->all_mr_lock); - - device = device->next; - } -} - -int -gf_rdma_deregister_arena(struct list_head **mr_list, - struct iobuf_arena *iobuf_arena) -{ - gf_rdma_arena_mr *tmp = NULL; - gf_rdma_arena_mr *dummy = NULL; - gf_rdma_device_t *device = NULL; - int count = 0, i = 0; - - count = iobuf_arena->iobuf_pool->rdma_device_count; - for (i = 0; i < count; i++) { - device = iobuf_arena->iobuf_pool->device[i]; - pthread_mutex_lock(&device->all_mr_lock); - { - list_for_each_entry_safe(tmp, dummy, mr_list[i], list) - { - if (tmp->iobuf_arena == iobuf_arena) { - if (ibv_dereg_mr(tmp->mr)) { - gf_msg("rdma", GF_LOG_WARNING, 0, - RDMA_MSG_DEREGISTER_ARENA_FAILED, - "deallocation of memory region " - "failed"); - pthread_mutex_unlock(&device->all_mr_lock); - return -1; - } - list_del(&tmp->list); - GF_FREE(tmp); - break; - } - } - } - pthread_mutex_unlock(&device->all_mr_lock); - } - - return 0; -} - -int -gf_rdma_register_arena(void **arg1, void *arg2) -{ - struct ibv_mr *mr = NULL; - gf_rdma_arena_mr *new = NULL; - struct iobuf_pool *iobuf_pool = NULL; - gf_rdma_device_t **device = (gf_rdma_device_t **)arg1; - struct iobuf_arena *iobuf_arena = arg2; - int count = 0, i = 0; - - iobuf_pool = iobuf_arena->iobuf_pool; - count = iobuf_pool->rdma_device_count; - for (i = 0; i < count; i++) { - new = GF_CALLOC(1, sizeof(gf_rdma_arena_mr), - gf_common_mt_rdma_arena_mr); - if (new == NULL) { - gf_msg("rdma", GF_LOG_INFO, ENOMEM, RDMA_MSG_MR_ALOC_FAILED, - "Out of " - "memory: registering pre allocated buffer " - "with rdma device failed."); - return -1; - } - INIT_LIST_HEAD(&new->list); - new->iobuf_arena = iobuf_arena; - - mr = ibv_reg_mr(device[i]->pd, iobuf_arena->mem_base, - iobuf_arena->arena_size, - IBV_ACCESS_REMOTE_READ | IBV_ACCESS_LOCAL_WRITE | - IBV_ACCESS_REMOTE_WRITE); - if (!mr) - gf_msg("rdma", GF_LOG_WARNING, 0, RDMA_MSG_MR_ALOC_FAILED, - "allocation of mr " - "failed"); - - new->mr = mr; - pthread_mutex_lock(&device[i]->all_mr_lock); - { - list_add(&new->list, &device[i]->all_mr); - } - pthread_mutex_unlock(&device[i]->all_mr_lock); - new = NULL; - } - - return 0; -} - -static void -gf_rdma_register_iobuf_pool(gf_rdma_device_t *device, - struct iobuf_pool *iobuf_pool) -{ - struct iobuf_arena *tmp = NULL; - struct iobuf_arena *dummy = NULL; - struct ibv_mr *mr = NULL; - gf_rdma_arena_mr *new = NULL; - - if (!list_empty(&iobuf_pool->all_arenas)) { - list_for_each_entry_safe(tmp, dummy, &iobuf_pool->all_arenas, all_list) - { - new = GF_CALLOC(1, sizeof(gf_rdma_arena_mr), - gf_common_mt_rdma_arena_mr); - if (new == NULL) { - gf_msg("rdma", GF_LOG_INFO, ENOMEM, RDMA_MSG_MR_ALOC_FAILED, - "Out of " - "memory: registering pre allocated " - "buffer with rdma device failed."); - return; - } - INIT_LIST_HEAD(&new->list); - new->iobuf_arena = tmp; - - mr = ibv_reg_mr(device->pd, tmp->mem_base, tmp->arena_size, - IBV_ACCESS_REMOTE_READ | IBV_ACCESS_LOCAL_WRITE | - IBV_ACCESS_REMOTE_WRITE); - if (!mr) { - gf_msg("rdma", GF_LOG_WARNING, 0, RDMA_MSG_MR_ALOC_FAILED, - "failed" - " to pre register buffers with rdma " - "devices."); - } - new->mr = mr; - pthread_mutex_lock(&device->all_mr_lock); - { - list_add(&new->list, &device->all_mr); - } - pthread_mutex_unlock(&device->all_mr_lock); - - new = NULL; - } - } - - return; -} - -static void -gf_rdma_register_iobuf_pool_with_device(gf_rdma_device_t *device, - struct iobuf_pool *iobuf_pool) -{ - while (device) { - gf_rdma_register_iobuf_pool(device, iobuf_pool); - device = device->next; - } -} - -static struct ibv_mr * -gf_rdma_get_pre_registred_mr(rpc_transport_t *this, void *ptr, int size) -{ - gf_rdma_arena_mr *tmp = NULL; - gf_rdma_arena_mr *dummy = NULL; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - - priv = this->private; - device = priv->device; - - pthread_mutex_lock(&device->all_mr_lock); - { - if (!list_empty(&device->all_mr)) { - list_for_each_entry_safe(tmp, dummy, &device->all_mr, list) - { - if (tmp->iobuf_arena->mem_base <= ptr && - ptr < tmp->iobuf_arena->mem_base + - tmp->iobuf_arena->arena_size) { - pthread_mutex_unlock(&device->all_mr_lock); - return tmp->mr; - } - } - } - } - pthread_mutex_unlock(&device->all_mr_lock); - - return NULL; -} - -static int32_t -gf_rdma_create_posts(rpc_transport_t *this) -{ - int32_t i = 0, ret = 0; - gf_rdma_post_t *post = NULL; - gf_rdma_private_t *priv = NULL; - gf_rdma_options_t *options = NULL; - gf_rdma_device_t *device = NULL; - - priv = this->private; - options = &priv->options; - device = priv->device; - - ret = __gf_rdma_create_posts(this, options->send_count, options->send_size, - &device->sendq, GF_RDMA_SEND_POST); - if (!ret) - ret = __gf_rdma_create_posts(this, options->recv_count, - options->recv_size, &device->recvq, - GF_RDMA_RECV_POST); - - if (!ret) { - for (i = 0; i < options->recv_count; i++) { - post = gf_rdma_get_post(&device->recvq); - if (gf_rdma_post_recv(device->srq, post) != 0) { - ret = -1; - break; - } - } - } - - if (ret) - gf_rdma_destroy_posts(this); - - return ret; -} - -static void -gf_rdma_destroy_cq(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - - priv = this->private; - device = priv->device; - - if (device->recv_cq) - ibv_destroy_cq(device->recv_cq); - device->recv_cq = NULL; - - if (device->send_cq) - ibv_destroy_cq(device->send_cq); - device->send_cq = NULL; - - return; -} - -static int32_t -gf_rdma_create_cq(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_options_t *options = NULL; - gf_rdma_device_t *device = NULL; - uint64_t send_cqe = 0; - int32_t ret = 0; - struct ibv_device_attr device_attr = { - {0}, - }; - - priv = this->private; - options = &priv->options; - device = priv->device; - - device->recv_cq = ibv_create_cq(priv->device->context, - options->recv_count * 2, device, - device->recv_chan, 0); - if (!device->recv_cq) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_CQ_CREATION_FAILED, - "creation of CQ for " - "device %s failed", - device->device_name); - ret = -1; - goto out; - } else if (ibv_req_notify_cq(device->recv_cq, 0)) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_REQ_NOTIFY_CQ_REVQ_FAILED, - "ibv_req_notify_" - "cq on recv CQ of device %s failed", - device->device_name); - ret = -1; - goto out; - } - - do { - ret = ibv_query_device(priv->device->context, &device_attr); - if (ret != 0) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_QUERY_DEVICE_FAILED, - "ibv_query_" - "device on %s returned %d (%s)", - priv->device->device_name, ret, - (ret > 0) ? strerror(ret) : ""); - ret = -1; - goto out; - } - - send_cqe = (uint64_t)options->send_count * 128; - send_cqe = (send_cqe > device_attr.max_cqe) ? device_attr.max_cqe - : send_cqe; - - /* TODO: make send_cq size dynamically adaptive */ - device->send_cq = ibv_create_cq(priv->device->context, send_cqe, device, - device->send_chan, 0); - if (!device->send_cq) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_CQ_CREATION_FAILED, - "creation of send_cq " - "for device %s failed", - device->device_name); - ret = -1; - goto out; - } - - if (ibv_req_notify_cq(device->send_cq, 0)) { - gf_msg(this->name, GF_LOG_ERROR, 0, - RDMA_MSG_REQ_NOTIFY_CQ_SENDQ_FAILED, - "ibv_req_notify_cq on send_cq for device %s" - " failed", - device->device_name); - ret = -1; - goto out; - } - } while (0); - -out: - if (ret != 0) - gf_rdma_destroy_cq(this); - - return ret; -} - -static gf_rdma_device_t * -gf_rdma_get_device(rpc_transport_t *this, struct ibv_context *ibctx, - char *device_name) -{ - glusterfs_ctx_t *ctx = NULL; - gf_rdma_private_t *priv = NULL; - gf_rdma_options_t *options = NULL; - int32_t ret = 0; - int32_t i = 0; - gf_rdma_device_t *trav = NULL, *device = NULL; - gf_rdma_ctx_t *rdma_ctx = NULL; - struct iobuf_pool *iobuf_pool = NULL; - - priv = this->private; - options = &priv->options; - ctx = this->ctx; - rdma_ctx = ctx->ib; - iobuf_pool = ctx->iobuf_pool; - - trav = rdma_ctx->device; - - while (trav) { - if (!strcmp(trav->device_name, device_name)) - break; - trav = trav->next; - } - - if (!trav) { - trav = GF_CALLOC(1, sizeof(*trav), gf_common_mt_rdma_device_t); - if (trav == NULL) { - goto out; - } - priv->device = trav; - trav->context = ibctx; - - trav->next = rdma_ctx->device; - rdma_ctx->device = trav; - - iobuf_pool->device[iobuf_pool->rdma_device_count] = trav; - iobuf_pool->mr_list[iobuf_pool->rdma_device_count++] = &trav->all_mr; - trav->request_ctx_pool = mem_pool_new(gf_rdma_request_context_t, - GF_RDMA_POOL_SIZE); - if (trav->request_ctx_pool == NULL) { - goto out; - } - - trav->ioq_pool = mem_pool_new(gf_rdma_ioq_t, GF_RDMA_POOL_SIZE); - if (trav->ioq_pool == NULL) { - goto out; - } - - trav->reply_info_pool = mem_pool_new(gf_rdma_reply_info_t, - GF_RDMA_POOL_SIZE); - if (trav->reply_info_pool == NULL) { - goto out; - } - - trav->device_name = gf_strdup(device_name); - - trav->send_chan = ibv_create_comp_channel(trav->context); - if (!trav->send_chan) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_SEND_COMP_CHAN_FAILED, - "could not " - "create send completion channel for " - "device (%s)", - device_name); - goto out; - } - - trav->recv_chan = ibv_create_comp_channel(trav->context); - if (!trav->recv_chan) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_RECV_COMP_CHAN_FAILED, - "could not " - "create recv completion channel for " - "device (%s)", - device_name); - - /* TODO: cleanup current mess */ - goto out; - } - - if (gf_rdma_create_cq(this) < 0) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_CQ_CREATION_FAILED, - "could not create CQ for device (%s)", device_name); - goto out; - } - - /* protection domain */ - trav->pd = ibv_alloc_pd(trav->context); - - if (!trav->pd) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_ALOC_PROT_DOM_FAILED, - "could not " - "allocate protection domain for device (%s)", - device_name); - goto out; - } - - struct ibv_srq_init_attr attr = {.attr = {.max_wr = options->recv_count, - .max_sge = 1, - .srq_limit = 10}}; - trav->srq = ibv_create_srq(trav->pd, &attr); - - if (!trav->srq) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_CRE_SRQ_FAILED, - "could not create SRQ" - " for device (%s)", - device_name); - goto out; - } - - /* queue init */ - gf_rdma_queue_init(&trav->sendq); - gf_rdma_queue_init(&trav->recvq); - - INIT_LIST_HEAD(&trav->all_mr); - pthread_mutex_init(&trav->all_mr_lock, NULL); - gf_rdma_register_iobuf_pool(trav, iobuf_pool); - - if (gf_rdma_create_posts(this) < 0) { - gf_msg(this->name, GF_LOG_ERROR, 0, RDMA_MSG_ALOC_POST_FAILED, - "could not allocate" - "posts for device (%s)", - device_name); - goto out; - } - - /* completion threads */ - ret = gf_thread_create(&trav->send_thread, NULL, - gf_rdma_send_completion_proc, trav->send_chan, - "rdmascom"); - if (ret) { - gf_msg(this->name, GF_LOG_ERROR, 0, - RDMA_MSG_SEND_COMP_THREAD_FAILED, - "could not create send completion thread for " - "device (%s)", - device_name); - goto out; - } - - ret = gf_thread_create(&trav->recv_thread, NULL, - gf_rdma_recv_completion_proc, trav->recv_chan, - "rdmarcom"); - if (ret) { - gf_msg(this->name, GF_LOG_ERROR, 0, - RDMA_MSG_RECV_COMP_THREAD_FAILED, - "could not create recv completion thread " - "for device (%s)", - device_name); - return NULL; - } - - ret = gf_thread_create(&trav->async_event_thread, NULL, - gf_rdma_async_event_thread, ibctx, "rdmaAsyn"); - if (ret) { - gf_msg(this->name, GF_LOG_ERROR, 0, - RDMA_MSG_ASYNC_EVENT_THEAD_FAILED, - "could not create async_event_thread"); - return NULL; - } - - /* qpreg */ - pthread_mutex_init(&trav->qpreg.lock, NULL); - for (i = 0; i < 42; i++) { - trav->qpreg.ents[i].next = &trav->qpreg.ents[i]; - trav->qpreg.ents[i].prev = &trav->qpreg.ents[i]; - } - } - - device = trav; - trav = NULL; -out: - - if (trav != NULL) { - rdma_ctx->device = trav->next; - gf_rdma_destroy_posts(this); - mem_pool_destroy(trav->ioq_pool); - mem_pool_destroy(trav->request_ctx_pool); - mem_pool_destroy(trav->reply_info_pool); - if (trav->pd != NULL) { - ibv_dealloc_pd(trav->pd); - } - gf_rdma_destroy_cq(this); - ibv_destroy_comp_channel(trav->recv_chan); - ibv_destroy_comp_channel(trav->send_chan); - GF_FREE((char *)trav->device_name); - GF_FREE(trav); - } - - return device; -} - -static rpc_transport_t * -gf_rdma_transport_new(rpc_transport_t *listener, struct rdma_cm_id *cm_id) -{ - gf_rdma_private_t *listener_priv = NULL, *priv = NULL; - rpc_transport_t *this = NULL, *new = NULL; - gf_rdma_options_t *options = NULL; - char *device_name = NULL; - - listener_priv = listener->private; - - this = GF_CALLOC(1, sizeof(rpc_transport_t), gf_common_mt_rpc_transport_t); - if (this == NULL) { - goto out; - } - - this->listener = listener; - - priv = GF_CALLOC(1, sizeof(gf_rdma_private_t), gf_common_mt_rdma_private_t); - if (priv == NULL) { - goto out; - } - - this->private = priv; - priv->options = listener_priv->options; - - priv->listener = listener; - priv->entity = GF_RDMA_SERVER; - - options = &priv->options; - - this->ops = listener->ops; - this->init = listener->init; - this->fini = listener->fini; - this->ctx = listener->ctx; - this->name = gf_strdup(listener->name); - this->notify = listener->notify; - this->mydata = listener->mydata; - this->xl = listener->xl; - - this->myinfo.sockaddr_len = sizeof(cm_id->route.addr.src_addr); - memcpy(&this->myinfo.sockaddr, &cm_id->route.addr.src_addr, - this->myinfo.sockaddr_len); - - this->peerinfo.sockaddr_len = sizeof(cm_id->route.addr.dst_addr); - memcpy(&this->peerinfo.sockaddr, &cm_id->route.addr.dst_addr, - this->peerinfo.sockaddr_len); - - priv->peer.trans = this; - gf_rdma_get_transport_identifiers(this); - - device_name = (char *)ibv_get_device_name(cm_id->verbs->device); - if (device_name == NULL) { - gf_msg(listener->name, GF_LOG_WARNING, 0, - RDMA_MSG_GET_DEVICE_NAME_FAILED, - "cannot get device " - "name (peer:%s me:%s)", - this->peerinfo.identifier, this->myinfo.identifier); - goto out; - } - - priv->device = gf_rdma_get_device(this, cm_id->verbs, device_name); - if (priv->device == NULL) { - gf_msg(listener->name, GF_LOG_WARNING, 0, RDMA_MSG_GET_IB_DEVICE_FAILED, - "cannot get infiniband" - " device %s (peer:%s me:%s)", - device_name, this->peerinfo.identifier, this->myinfo.identifier); - goto out; - } - - priv->peer.send_count = options->send_count; - priv->peer.recv_count = options->recv_count; - priv->peer.send_size = options->send_size; - priv->peer.recv_size = options->recv_size; - priv->peer.cm_id = cm_id; - INIT_LIST_HEAD(&priv->peer.ioq); - - pthread_mutex_init(&priv->write_mutex, NULL); - pthread_mutex_init(&priv->recv_mutex, NULL); - - cm_id->context = this; - - new = rpc_transport_ref(this); - this = NULL; -out: - if (this != NULL) { - if (this->private != NULL) { - GF_FREE(this->private); - } - - if (this->name != NULL) { - GF_FREE(this->name); - } - - GF_FREE(this); - } - - return new; -} - -static int -gf_rdma_cm_handle_connect_request(struct rdma_cm_event *event) -{ - int ret = -1; - rpc_transport_t *this = NULL, *listener = NULL; - struct rdma_cm_id *child_cm_id = NULL, *listener_cm_id = NULL; - struct rdma_conn_param conn_param = { - 0, - }; - gf_rdma_private_t *priv = NULL; - gf_rdma_options_t *options = NULL; - - child_cm_id = event->id; - listener_cm_id = event->listen_id; - - listener = listener_cm_id->context; - priv = listener->private; - options = &priv->options; - - this = gf_rdma_transport_new(listener, child_cm_id); - if (this == NULL) { - gf_msg(listener->name, GF_LOG_WARNING, 0, - RDMA_MSG_CREAT_INC_TRANS_FAILED, - "could not create " - "a transport for incoming connection" - " (me.name:%s me.identifier:%s)", - listener->name, listener->myinfo.identifier); - rdma_destroy_id(child_cm_id); - goto out; - } - - gf_msg_trace(listener->name, 0, - "got a connect request (me:%s peer:" - "%s)", - listener->myinfo.identifier, this->peerinfo.identifier); - - ret = gf_rdma_create_qp(this); - if (ret < 0) { - gf_msg(listener->name, GF_LOG_WARNING, 0, RDMA_MSG_CREAT_QP_FAILED, - "could not create QP " - "(peer:%s me:%s)", - this->peerinfo.identifier, this->myinfo.identifier); - gf_rdma_cm_handle_disconnect(this); - goto out; - } - - conn_param.responder_resources = 1; - conn_param.initiator_depth = 1; - conn_param.retry_count = options->attr_retry_cnt; - conn_param.rnr_retry_count = options->attr_rnr_retry; - - ret = rdma_accept(child_cm_id, &conn_param); - if (ret < 0) { - gf_msg(listener->name, GF_LOG_WARNING, errno, RDMA_MSG_ACCEPT_FAILED, - "rdma_accept failed peer:%s " - "me:%s", - this->peerinfo.identifier, this->myinfo.identifier); - gf_rdma_cm_handle_disconnect(this); - goto out; - } - gf_rdma_cm_handle_connect_init(event); - ret = 0; - -out: - return ret; -} - -static int -gf_rdma_cm_handle_route_resolved(struct rdma_cm_event *event) -{ - struct rdma_conn_param conn_param = { - 0, - }; - int ret = 0; - rpc_transport_t *this = NULL; - gf_rdma_private_t *priv = NULL; - gf_rdma_peer_t *peer = NULL; - gf_rdma_options_t *options = NULL; - - if (event == NULL) { - goto out; - } - - this = event->id->context; - - priv = this->private; - peer = &priv->peer; - options = &priv->options; - - ret = gf_rdma_create_qp(this); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_CREAT_QP_FAILED, - "could not create QP " - "(peer:%s me:%s)", - this->peerinfo.identifier, this->myinfo.identifier); - gf_rdma_cm_handle_disconnect(this); - goto out; - } - - memset(&conn_param, 0, sizeof conn_param); - conn_param.responder_resources = 1; - conn_param.initiator_depth = 1; - conn_param.retry_count = options->attr_retry_cnt; - conn_param.rnr_retry_count = options->attr_rnr_retry; - - ret = rdma_connect(peer->cm_id, &conn_param); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_CONNECT_FAILED, - "rdma_connect failed"); - gf_rdma_cm_handle_disconnect(this); - goto out; - } - - gf_msg_trace(this->name, 0, "route resolved (me:%s peer:%s)", - this->myinfo.identifier, this->peerinfo.identifier); - - ret = 0; -out: - return ret; -} - -static int -gf_rdma_cm_handle_addr_resolved(struct rdma_cm_event *event) -{ - rpc_transport_t *this = NULL; - gf_rdma_peer_t *peer = NULL; - gf_rdma_private_t *priv = NULL; - int ret = 0; - - this = event->id->context; - - priv = this->private; - peer = &priv->peer; - - GF_ASSERT(peer->cm_id == event->id); - - this->myinfo.sockaddr_len = sizeof(peer->cm_id->route.addr.src_addr); - memcpy(&this->myinfo.sockaddr, &peer->cm_id->route.addr.src_addr, - this->myinfo.sockaddr_len); - - this->peerinfo.sockaddr_len = sizeof(peer->cm_id->route.addr.dst_addr); - memcpy(&this->peerinfo.sockaddr, &peer->cm_id->route.addr.dst_addr, - this->peerinfo.sockaddr_len); - - gf_rdma_get_transport_identifiers(this); - - ret = rdma_resolve_route(peer->cm_id, 2000); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_ROUTE_RESOLVE_FAILED, - "rdma_resolve_route " - "failed (me:%s peer:%s)", - this->myinfo.identifier, this->peerinfo.identifier); - gf_rdma_cm_handle_disconnect(this); - return ret; - } - - gf_msg_trace(this->name, 0, "Address resolved (me:%s peer:%s)", - this->myinfo.identifier, this->peerinfo.identifier); - - return ret; -} - -static void -gf_rdma_cm_handle_disconnect(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - char need_unref = 0; - - priv = this->private; - gf_msg_debug(this->name, 0, "peer disconnected, cleaning up"); - - pthread_mutex_lock(&priv->write_mutex); - { - if (priv->peer.cm_id != NULL) { - need_unref = 1; - priv->connected = 0; - } - - __gf_rdma_teardown(this); - } - pthread_mutex_unlock(&priv->write_mutex); - - rpc_transport_notify(this, RPC_TRANSPORT_DISCONNECT, this); - - if (need_unref) - rpc_transport_unref(this); -} - -static int -gf_rdma_cm_handle_connect_init(struct rdma_cm_event *event) -{ - rpc_transport_t *this = NULL; - gf_rdma_private_t *priv = NULL; - struct rdma_cm_id *cm_id = NULL; - int ret = 0; - - cm_id = event->id; - this = cm_id->context; - priv = this->private; - - if (priv->connected == 1) { - gf_msg_trace(this->name, 0, - "received event " - "RDMA_CM_EVENT_ESTABLISHED (me:%s peer:%s)", - this->myinfo.identifier, this->peerinfo.identifier); - return ret; - } - - priv->connected = 1; - - pthread_mutex_lock(&priv->write_mutex); - { - priv->peer.quota = 1; - priv->peer.quota_set = 0; - } - pthread_mutex_unlock(&priv->write_mutex); - - if (priv->entity == GF_RDMA_CLIENT) { - gf_msg_trace(this->name, 0, - "received event " - "RDMA_CM_EVENT_ESTABLISHED (me:%s peer:%s)", - this->myinfo.identifier, this->peerinfo.identifier); - ret = rpc_transport_notify(this, RPC_TRANSPORT_CONNECT, this); - - } else if (priv->entity == GF_RDMA_SERVER) { - ret = rpc_transport_notify(priv->listener, RPC_TRANSPORT_ACCEPT, this); - } - - if (ret < 0) { - gf_rdma_disconnect(this, _gf_false); - } - - return ret; -} - -static int -gf_rdma_cm_handle_event_error(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - - priv = this->private; - - if (priv->entity != GF_RDMA_SERVER_LISTENER) { - gf_rdma_cm_handle_disconnect(this); - } - - return 0; -} - -static int -gf_rdma_cm_handle_device_removal(struct rdma_cm_event *event) -{ - return 0; -} - -static void * -gf_rdma_cm_event_handler(void *data) -{ - struct rdma_cm_event *event = NULL; - int ret = 0; - rpc_transport_t *this = NULL; - struct rdma_event_channel *event_channel = NULL; - - event_channel = data; - - while (1) { - ret = rdma_get_cm_event(event_channel, &event); - if (ret != 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_CM_EVENT_FAILED, "rdma_cm_get_event failed"); - break; - } - - switch (event->event) { - case RDMA_CM_EVENT_ADDR_RESOLVED: - gf_rdma_cm_handle_addr_resolved(event); - break; - - case RDMA_CM_EVENT_ROUTE_RESOLVED: - gf_rdma_cm_handle_route_resolved(event); - break; - - case RDMA_CM_EVENT_CONNECT_REQUEST: - gf_rdma_cm_handle_connect_request(event); - break; - - case RDMA_CM_EVENT_ESTABLISHED: - gf_rdma_cm_handle_connect_init(event); - break; - - case RDMA_CM_EVENT_ADDR_ERROR: - case RDMA_CM_EVENT_ROUTE_ERROR: - case RDMA_CM_EVENT_CONNECT_ERROR: - case RDMA_CM_EVENT_UNREACHABLE: - case RDMA_CM_EVENT_REJECTED: - this = event->id->context; - - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_CM_EVENT_FAILED, - "cma event %s, " - "error %d (me:%s peer:%s)\n", - rdma_event_str(event->event), event->status, - this->myinfo.identifier, this->peerinfo.identifier); - - rdma_ack_cm_event(event); - event = NULL; - - gf_rdma_cm_handle_event_error(this); - continue; - - case RDMA_CM_EVENT_DISCONNECTED: - this = event->id->context; - - gf_msg_debug(this->name, 0, - "received disconnect " - "(me:%s peer:%s)\n", - this->myinfo.identifier, - this->peerinfo.identifier); - - rdma_ack_cm_event(event); - event = NULL; - - gf_rdma_cm_handle_disconnect(this); - continue; - - case RDMA_CM_EVENT_DEVICE_REMOVAL: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CM_EVENT_FAILED, - "device " - "removed"); - gf_rdma_cm_handle_device_removal(event); - break; - - default: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CM_EVENT_FAILED, - "unhandled event: %s, ignoring", - rdma_event_str(event->event)); - break; - } - - rdma_ack_cm_event(event); - } - - return NULL; -} - -static int32_t -gf_rdma_post_send(struct ibv_qp *qp, gf_rdma_post_t *post, int32_t len) -{ - struct ibv_sge list = {.addr = (unsigned long)post->buf, - .length = len, - .lkey = post->mr->lkey}; - - struct ibv_send_wr wr = - { - .wr_id = (unsigned long)post, - .sg_list = &list, - .num_sge = 1, - .opcode = IBV_WR_SEND, - .send_flags = IBV_SEND_SIGNALED, - }, - *bad_wr; - - if (!qp) - return EINVAL; - - return ibv_post_send(qp, &wr, &bad_wr); -} - -int -__gf_rdma_encode_error(gf_rdma_peer_t *peer, gf_rdma_reply_info_t *reply_info, - struct iovec *rpchdr, gf_rdma_header_t *hdr, - gf_rdma_errcode_t err) -{ - struct rpc_msg *rpc_msg = NULL; - - if (reply_info != NULL) { - hdr->rm_xid = hton32(reply_info->rm_xid); - } else { - rpc_msg = rpchdr[0].iov_base; /* assume rpchdr contains - * only one vector. - * (which is true) - */ - hdr->rm_xid = rpc_msg->rm_xid; - } - - hdr->rm_vers = hton32(GF_RDMA_VERSION); - hdr->rm_credit = hton32(peer->send_count); - hdr->rm_type = hton32(GF_RDMA_ERROR); - hdr->rm_body.rm_error.rm_type = hton32(err); - if (err == ERR_VERS) { - hdr->rm_body.rm_error.rm_version.gf_rdma_vers_low = hton32( - GF_RDMA_VERSION); - hdr->rm_body.rm_error.rm_version.gf_rdma_vers_high = hton32( - GF_RDMA_VERSION); - } - - return sizeof(*hdr); -} - -int32_t -__gf_rdma_send_error(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post, gf_rdma_reply_info_t *reply_info, - gf_rdma_errcode_t err) -{ - int32_t ret = -1, len = 0; - - len = __gf_rdma_encode_error(peer, reply_info, entry->rpchdr, - (gf_rdma_header_t *)post->buf, err); - if (len == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, 0, RDMA_MSG_ENCODE_ERROR, - "encode error returned -1"); - goto out; - } - - gf_rdma_post_ref(post); - - ret = gf_rdma_post_send(peer->qp, post, len); - if (!ret) { - ret = len; - } else { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_POST_SEND_FAILED, - "gf_rdma_post_send (to %s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - gf_rdma_post_unref(post); - __gf_rdma_disconnect(peer->trans); - ret = -1; - } - -out: - return ret; -} - -int32_t -__gf_rdma_create_read_chunks_from_vector(gf_rdma_peer_t *peer, - gf_rdma_read_chunk_t **readch_ptr, - int32_t *pos, struct iovec *vector, - int count, - gf_rdma_request_context_t *request_ctx) -{ - int i = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - struct ibv_mr *mr = NULL; - gf_rdma_read_chunk_t *readch = NULL; - int32_t ret = -1; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, readch_ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, *readch_ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, vector, out); - - priv = peer->trans->private; - device = priv->device; - readch = *readch_ptr; - - for (i = 0; i < count; i++) { - readch->rc_discrim = hton32(1); - readch->rc_position = hton32(*pos); - - mr = gf_rdma_get_pre_registred_mr( - peer->trans, (void *)vector[i].iov_base, vector[i].iov_len); - if (!mr) { - mr = ibv_reg_mr(device->pd, vector[i].iov_base, vector[i].iov_len, - IBV_ACCESS_REMOTE_READ); - } - if (!mr) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_MR_ALOC_FAILED, - "memory registration failed (peer:%s)", - peer->trans->peerinfo.identifier); - goto out; - } - - request_ctx->mr[request_ctx->mr_count++] = mr; - - readch->rc_target.rs_handle = hton32(mr->rkey); - readch->rc_target.rs_length = hton32(vector[i].iov_len); - readch->rc_target.rs_offset = hton64( - (uint64_t)(unsigned long)vector[i].iov_base); - - *pos = *pos + vector[i].iov_len; - readch++; - } - - *readch_ptr = readch; - - ret = 0; -out: - return ret; -} - -int32_t -__gf_rdma_create_read_chunks(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_chunktype_t type, uint32_t **ptr, - gf_rdma_request_context_t *request_ctx) -{ - int32_t ret = -1; - int pos = 0; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, entry, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, *ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, request_ctx, out); - - request_ctx->iobref = iobref_ref(entry->iobref); - - if (type == gf_rdma_areadch) { - pos = 0; - ret = __gf_rdma_create_read_chunks_from_vector( - peer, (gf_rdma_read_chunk_t **)ptr, &pos, entry->rpchdr, - entry->rpchdr_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_READ_CHUNK_VECTOR_FAILED, - "cannot create read chunks from vector " - "entry->rpchdr"); - goto out; - } - - ret = __gf_rdma_create_read_chunks_from_vector( - peer, (gf_rdma_read_chunk_t **)ptr, &pos, entry->proghdr, - entry->proghdr_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_READ_CHUNK_VECTOR_FAILED, - "cannot create read chunks from vector " - "entry->proghdr"); - } - - if (entry->prog_payload_count != 0) { - ret = __gf_rdma_create_read_chunks_from_vector( - peer, (gf_rdma_read_chunk_t **)ptr, &pos, entry->prog_payload, - entry->prog_payload_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_READ_CHUNK_VECTOR_FAILED, - "cannot create read chunks from vector" - " entry->prog_payload"); - } - } - } else { - pos = iov_length(entry->rpchdr, entry->rpchdr_count); - ret = __gf_rdma_create_read_chunks_from_vector( - peer, (gf_rdma_read_chunk_t **)ptr, &pos, entry->prog_payload, - entry->prog_payload_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_READ_CHUNK_VECTOR_FAILED, - "cannot create read chunks from vector " - "entry->prog_payload"); - } - } - - /* terminate read-chunk list*/ - **ptr = 0; - *ptr = *ptr + 1; -out: - return ret; -} - -int32_t -__gf_rdma_create_write_chunks_from_vector( - gf_rdma_peer_t *peer, gf_rdma_write_chunk_t **writech_ptr, - struct iovec *vector, int count, gf_rdma_request_context_t *request_ctx) -{ - int i = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - struct ibv_mr *mr = NULL; - gf_rdma_write_chunk_t *writech = NULL; - int32_t ret = -1; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, writech_ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, *writech_ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, vector, out); - - writech = *writech_ptr; - - priv = peer->trans->private; - device = priv->device; - - for (i = 0; i < count; i++) { - mr = gf_rdma_get_pre_registred_mr( - peer->trans, (void *)vector[i].iov_base, vector[i].iov_len); - if (!mr) { - mr = ibv_reg_mr(device->pd, vector[i].iov_base, vector[i].iov_len, - IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_LOCAL_WRITE); - } - - if (!mr) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_MR_ALOC_FAILED, - "memory " - "registration failed (peer:%s)", - peer->trans->peerinfo.identifier); - goto out; - } - - request_ctx->mr[request_ctx->mr_count++] = mr; - - writech->wc_target.rs_handle = hton32(mr->rkey); - writech->wc_target.rs_length = hton32(vector[i].iov_len); - writech->wc_target.rs_offset = hton64( - ((uint64_t)(unsigned long)vector[i].iov_base)); - - writech++; - } - - *writech_ptr = writech; - - ret = 0; -out: - return ret; -} - -int32_t -__gf_rdma_create_write_chunks(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_chunktype_t chunk_type, uint32_t **ptr, - gf_rdma_request_context_t *request_ctx) -{ - int32_t ret = -1; - gf_rdma_write_array_t *warray = NULL; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, *ptr, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, entry, out); - - if ((chunk_type == gf_rdma_replych) && - ((entry->msg.request.rsphdr_count != 1) || - (entry->msg.request.rsphdr_vec[0].iov_base == NULL))) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_BUFFER_ERROR, - (entry->msg.request.rsphdr_count == 1) - ? "chunktype specified as reply chunk but the vector " - "specifying the buffer to be used for holding reply" - " header is not correct" - : "chunktype specified as reply chunk, but more than one " - "buffer provided for holding reply"); - goto out; - } - - /* - if ((chunk_type == gf_rdma_writech) - && ((entry->msg.request.rsphdr_count == 0) - || (entry->msg.request.rsphdr_vec[0].iov_base == NULL))) { - gf_msg_debug (GF_RDMA_LOG_NAME, 0, - "vector specifying buffer to hold the program's reply " - "header should also be provided when buffers are " - "provided for holding the program's payload in reply"); - goto out; - } - */ - - if (chunk_type == gf_rdma_writech) { - warray = (gf_rdma_write_array_t *)*ptr; - warray->wc_discrim = hton32(1); - warray->wc_nchunks = hton32(entry->msg.request.rsp_payload_count); - - *ptr = (uint32_t *)&warray->wc_array[0]; - - ret = __gf_rdma_create_write_chunks_from_vector( - peer, (gf_rdma_write_chunk_t **)ptr, entry->msg.request.rsp_payload, - entry->msg.request.rsp_payload_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_WRITE_CHUNK_VECTOR_FAILED, - "cannot create write chunks from vector " - "entry->rpc_payload"); - goto out; - } - - /* terminate write chunklist */ - **ptr = 0; - *ptr = *ptr + 1; - - /* no reply chunklist */ - **ptr = 0; - *ptr = *ptr + 1; - } else { - /* no write chunklist */ - **ptr = 0; - *ptr = *ptr + 1; - - warray = (gf_rdma_write_array_t *)*ptr; - warray->wc_discrim = hton32(1); - warray->wc_nchunks = hton32(entry->msg.request.rsphdr_count); - - *ptr = (uint32_t *)&warray->wc_array[0]; - - ret = __gf_rdma_create_write_chunks_from_vector( - peer, (gf_rdma_write_chunk_t **)ptr, entry->msg.request.rsphdr_vec, - entry->msg.request.rsphdr_count, request_ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_WRITE_CHUNK_VECTOR_FAILED, - "cannot create write chunks from vector " - "entry->rpchdr"); - goto out; - } - - /* terminate reply chunklist */ - **ptr = 0; - *ptr = *ptr + 1; - } - -out: - return ret; -} - -static void -__gf_rdma_deregister_mr(gf_rdma_device_t *device, struct ibv_mr **mr, int count) -{ - gf_rdma_arena_mr *tmp = NULL; - gf_rdma_arena_mr *dummy = NULL; - int i = 0; - int found = 0; - - if (mr == NULL) { - goto out; - } - - for (i = 0; i < count; i++) { - found = 0; - pthread_mutex_lock(&device->all_mr_lock); - { - if (!list_empty(&device->all_mr)) { - list_for_each_entry_safe(tmp, dummy, &device->all_mr, list) - { - if (tmp->mr == mr[i]) { - found = 1; - break; - } - } - } - } - pthread_mutex_unlock(&device->all_mr_lock); - if (!found) - ibv_dereg_mr(mr[i]); - } - -out: - return; -} - -static int32_t -__gf_rdma_quota_put(gf_rdma_peer_t *peer) -{ - int32_t ret = 0; - - peer->quota++; - ret = peer->quota; - - if (!list_empty(&peer->ioq)) { - ret = __gf_rdma_ioq_churn(peer); - } - - return ret; -} - -static int32_t -gf_rdma_quota_put(gf_rdma_peer_t *peer) -{ - int32_t ret = 0; - gf_rdma_private_t *priv = NULL; - - priv = peer->trans->private; - pthread_mutex_lock(&priv->write_mutex); - { - ret = __gf_rdma_quota_put(peer); - } - pthread_mutex_unlock(&priv->write_mutex); - - return ret; -} - -/* to be called with priv->mutex held */ -void -__gf_rdma_request_context_destroy(gf_rdma_request_context_t *context) -{ - gf_rdma_peer_t *peer = NULL; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - int32_t ret = 0; - - if (context == NULL) { - goto out; - } - - peer = context->peer; - - priv = peer->trans->private; - device = priv->device; - __gf_rdma_deregister_mr(device, context->mr, context->mr_count); - - if (priv->connected) { - ret = __gf_rdma_quota_put(peer); - if (ret < 0) { - gf_msg_debug("rdma", 0, "failed to send message"); - mem_put(context); - __gf_rdma_disconnect(peer->trans); - goto out; - } - } - - if (context->iobref != NULL) { - iobref_unref(context->iobref); - context->iobref = NULL; - } - - if (context->rsp_iobref != NULL) { - iobref_unref(context->rsp_iobref); - context->rsp_iobref = NULL; - } - - mem_put(context); - -out: - return; -} - -void -gf_rdma_post_context_destroy(gf_rdma_device_t *device, - gf_rdma_post_context_t *ctx) -{ - if (ctx == NULL) { - goto out; - } - - __gf_rdma_deregister_mr(device, ctx->mr, ctx->mr_count); - - if (ctx->iobref != NULL) { - iobref_unref(ctx->iobref); - } - - if (ctx->hdr_iobuf != NULL) { - iobuf_unref(ctx->hdr_iobuf); - } - - memset(ctx, 0, sizeof(*ctx)); -out: - return; -} - -int -gf_rdma_post_unref(gf_rdma_post_t *post) -{ - int refcount = -1; - - if (post == NULL) { - goto out; - } - - pthread_mutex_lock(&post->lock); - { - refcount = --post->refcount; - } - pthread_mutex_unlock(&post->lock); - - if (refcount == 0) { - gf_rdma_post_context_destroy(post->device, &post->ctx); - if (post->type == GF_RDMA_SEND_POST) { - gf_rdma_put_post(&post->device->sendq, post); - } else { - gf_rdma_post_recv(post->device->srq, post); - } - } -out: - return refcount; -} - -int -gf_rdma_post_get_refcount(gf_rdma_post_t *post) -{ - int refcount = -1; - - if (post == NULL) { - goto out; - } - - pthread_mutex_lock(&post->lock); - { - refcount = post->refcount; - } - pthread_mutex_unlock(&post->lock); - -out: - return refcount; -} - -gf_rdma_post_t * -gf_rdma_post_ref(gf_rdma_post_t *post) -{ - if (post == NULL) { - goto out; - } - - pthread_mutex_lock(&post->lock); - { - post->refcount++; - } - pthread_mutex_unlock(&post->lock); - -out: - return post; -} - -int32_t -__gf_rdma_ioq_churn_request(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post) -{ - gf_rdma_chunktype_t rtype = gf_rdma_noch; - gf_rdma_chunktype_t wtype = gf_rdma_noch; - uint64_t send_size = 0; - gf_rdma_header_t *hdr = NULL; - struct rpc_msg *rpc_msg = NULL; - uint32_t *chunkptr = NULL; - char *buf = NULL; - int32_t ret = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - int chunk_count = 0; - gf_rdma_request_context_t *request_ctx = NULL; - uint32_t prog_payload_length = 0, len = 0; - struct rpc_req *rpc_req = NULL; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, entry, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, post, out); - - if ((entry->msg.request.rsphdr_count != 0) && - (entry->msg.request.rsp_payload_count != 0)) { - ret = -1; - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_WRITE_REPLY_CHUNCK_CONFLICT, - "both write-chunklist and reply-chunk cannot be " - "present"); - goto out; - } - - post->ctx.is_request = 1; - priv = peer->trans->private; - device = priv->device; - - hdr = (gf_rdma_header_t *)post->buf; - - send_size = iov_length(entry->rpchdr, entry->rpchdr_count) + - iov_length(entry->proghdr, entry->proghdr_count) + - GLUSTERFS_RDMA_MAX_HEADER_SIZE; - - if (entry->prog_payload_count != 0) { - prog_payload_length = iov_length(entry->prog_payload, - entry->prog_payload_count); - } - - if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - rtype = gf_rdma_areadch; - } else if ((send_size + prog_payload_length) < - GLUSTERFS_RDMA_INLINE_THRESHOLD) { - rtype = gf_rdma_noch; - } else if (entry->prog_payload_count != 0) { - rtype = gf_rdma_readch; - } - - if (entry->msg.request.rsphdr_count != 0) { - wtype = gf_rdma_replych; - } else if (entry->msg.request.rsp_payload_count != 0) { - wtype = gf_rdma_writech; - } - - if (rtype == gf_rdma_readch) { - chunk_count += entry->prog_payload_count; - } else if (rtype == gf_rdma_areadch) { - chunk_count += entry->rpchdr_count; - chunk_count += entry->proghdr_count; - } - - if (wtype == gf_rdma_writech) { - chunk_count += entry->msg.request.rsp_payload_count; - } else if (wtype == gf_rdma_replych) { - chunk_count += entry->msg.request.rsphdr_count; - } - - if (chunk_count > GF_RDMA_MAX_SEGMENTS) { - ret = -1; - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CHUNK_COUNT_GREAT_MAX_SEGMENTS, - "chunk count(%d) exceeding maximum allowed RDMA " - "segment count(%d)", - chunk_count, GF_RDMA_MAX_SEGMENTS); - goto out; - } - - if ((wtype != gf_rdma_noch) || (rtype != gf_rdma_noch)) { - request_ctx = mem_get(device->request_ctx_pool); - if (request_ctx == NULL) { - ret = -1; - goto out; - } - - memset(request_ctx, 0, sizeof(*request_ctx)); - - request_ctx->pool = device->request_ctx_pool; - request_ctx->peer = peer; - - entry->msg.request.rpc_req->conn_private = request_ctx; - - if (entry->msg.request.rsp_iobref != NULL) { - request_ctx->rsp_iobref = iobref_ref(entry->msg.request.rsp_iobref); - } - } - - rpc_msg = (struct rpc_msg *)entry->rpchdr[0].iov_base; - - hdr->rm_xid = rpc_msg->rm_xid; /* no need of hton32(rpc_msg->rm_xid), - * since rpc_msg->rm_xid is already - * hton32ed value of actual xid - */ - hdr->rm_vers = hton32(GF_RDMA_VERSION); - hdr->rm_credit = hton32(peer->send_count); - if (rtype == gf_rdma_areadch) { - hdr->rm_type = hton32(GF_RDMA_NOMSG); - } else { - hdr->rm_type = hton32(GF_RDMA_MSG); - } - - chunkptr = &hdr->rm_body.rm_chunks[0]; - if (rtype != gf_rdma_noch) { - ret = __gf_rdma_create_read_chunks(peer, entry, rtype, &chunkptr, - request_ctx); - if (ret != 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CREATE_READ_CHUNK_FAILED, - "creation of read chunks failed"); - goto out; - } - } else { - *chunkptr++ = 0; /* no read chunks */ - } - - if (wtype != gf_rdma_noch) { - ret = __gf_rdma_create_write_chunks(peer, entry, wtype, &chunkptr, - request_ctx); - if (ret != 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CREATE_WRITE_REPLAY_FAILED, - "creation of write/reply chunk failed"); - goto out; - } - } else { - *chunkptr++ = 0; /* no write chunks */ - *chunkptr++ = 0; /* no reply chunk */ - } - - buf = (char *)chunkptr; - - if (rtype != gf_rdma_areadch) { - iov_unload(buf, entry->rpchdr, entry->rpchdr_count); - buf += iov_length(entry->rpchdr, entry->rpchdr_count); - - iov_unload(buf, entry->proghdr, entry->proghdr_count); - buf += iov_length(entry->proghdr, entry->proghdr_count); - - if (rtype != gf_rdma_readch) { - iov_unload(buf, entry->prog_payload, entry->prog_payload_count); - buf += iov_length(entry->prog_payload, entry->prog_payload_count); - } - } - - len = buf - post->buf; - - gf_rdma_post_ref(post); - - ret = gf_rdma_post_send(peer->qp, post, len); - if (!ret) { - ret = len; - } else { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_POST_SEND_FAILED, - "gf_rdma_post_send (to %s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - gf_rdma_post_unref(post); - __gf_rdma_disconnect(peer->trans); - ret = -1; - } - -out: - if (ret == -1) { - rpc_req = entry->msg.request.rpc_req; - - if (request_ctx != NULL) { - __gf_rdma_request_context_destroy(rpc_req->conn_private); - } - - rpc_req->conn_private = NULL; - } - - return ret; -} - -static void -__gf_rdma_fill_reply_header(gf_rdma_header_t *header, struct iovec *rpchdr, - gf_rdma_reply_info_t *reply_info, int credits) -{ - struct rpc_msg *rpc_msg = NULL; - - if (reply_info != NULL) { - header->rm_xid = hton32(reply_info->rm_xid); - } else { - rpc_msg = rpchdr[0].iov_base; /* assume rpchdr contains - * only one vector. - * (which is true) - */ - header->rm_xid = rpc_msg->rm_xid; - } - - header->rm_type = hton32(GF_RDMA_MSG); - header->rm_vers = hton32(GF_RDMA_VERSION); - header->rm_credit = hton32(credits); - - header->rm_body.rm_chunks[0] = 0; /* no read chunks */ - header->rm_body.rm_chunks[1] = 0; /* no write chunks */ - header->rm_body.rm_chunks[2] = 0; /* no reply chunks */ - - return; -} - -int32_t -__gf_rdma_send_reply_inline(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post, - gf_rdma_reply_info_t *reply_info) -{ - gf_rdma_header_t *header = NULL; - int32_t send_size = 0, ret = 0; - char *buf = NULL; - - send_size = iov_length(entry->rpchdr, entry->rpchdr_count) + - iov_length(entry->proghdr, entry->proghdr_count) + - iov_length(entry->prog_payload, entry->prog_payload_count) + - sizeof(gf_rdma_header_t); /* - * remember, no chunklists in the - * reply - */ - - if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - ret = __gf_rdma_send_error(peer, entry, post, reply_info, ERR_CHUNK); - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_SEND_SIZE_GREAT_INLINE_THRESHOLD, - "msg size (%d) is greater than maximum size " - "of msg that can be sent inlined (%d)", - send_size, GLUSTERFS_RDMA_INLINE_THRESHOLD); - goto out; - } - - header = (gf_rdma_header_t *)post->buf; - - __gf_rdma_fill_reply_header(header, entry->rpchdr, reply_info, - peer->send_count); - - buf = (char *)&header->rm_body.rm_chunks[3]; - - if (entry->rpchdr_count != 0) { - iov_unload(buf, entry->rpchdr, entry->rpchdr_count); - buf += iov_length(entry->rpchdr, entry->rpchdr_count); - } - - if (entry->proghdr_count != 0) { - iov_unload(buf, entry->proghdr, entry->proghdr_count); - buf += iov_length(entry->proghdr, entry->proghdr_count); - } - - if (entry->prog_payload_count != 0) { - iov_unload(buf, entry->prog_payload, entry->prog_payload_count); - buf += iov_length(entry->prog_payload, entry->prog_payload_count); - } - - gf_rdma_post_ref(post); - - ret = gf_rdma_post_send(peer->qp, post, (buf - post->buf)); - if (!ret) { - ret = send_size; - } else { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_POST_SEND_FAILED, - "posting send (to %s) " - "failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - gf_rdma_post_unref(post); - __gf_rdma_disconnect(peer->trans); - ret = -1; - } - -out: - return ret; -} - -int32_t -__gf_rdma_reply_encode_write_chunks(gf_rdma_peer_t *peer, uint32_t payload_size, - gf_rdma_post_t *post, - gf_rdma_reply_info_t *reply_info, - uint32_t **ptr) -{ - uint32_t chunk_size = 0; - int32_t ret = -1; - gf_rdma_write_array_t *target_array = NULL; - int i = 0; - - target_array = (gf_rdma_write_array_t *)*ptr; - - for (i = 0; i < reply_info->wc_array->wc_nchunks; i++) { - chunk_size += reply_info->wc_array->wc_array[i].wc_target.rs_length; - } - - if (chunk_size < payload_size) { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "length of payload (%d) is " - "exceeding the total write chunk length (%d)", - payload_size, chunk_size); - goto out; - } - - target_array->wc_discrim = hton32(1); - for (i = 0; (i < reply_info->wc_array->wc_nchunks) && (payload_size != 0); - i++) { - target_array->wc_array[i].wc_target.rs_offset = hton64( - reply_info->wc_array->wc_array[i].wc_target.rs_offset); - - target_array->wc_array[i].wc_target.rs_length = hton32( - min(payload_size, - reply_info->wc_array->wc_array[i].wc_target.rs_length)); - } - - target_array->wc_nchunks = hton32(i); - target_array->wc_array[i].wc_target.rs_handle = 0; /* terminate - chunklist */ - - ret = 0; - - *ptr = &target_array->wc_array[i].wc_target.rs_length; -out: - return ret; -} - -static int32_t -__gf_rdma_register_local_mr_for_rdma(gf_rdma_peer_t *peer, struct iovec *vector, - int count, gf_rdma_post_context_t *ctx) -{ - int i = 0; - int32_t ret = -1; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, ctx, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, vector, out); - - priv = peer->trans->private; - device = priv->device; - - for (i = 0; i < count; i++) { - /* what if the memory is registered more than once? - * Assume that a single write buffer is passed to afr, which - * then passes it to its children. If more than one children - * happen to use rdma, then the buffer is registered more than - * once. - * Ib-verbs specification says that multiple registrations of - * same memory location is allowed. Refer to 10.6.3.8 of - * Infiniband Architecture Specification Volume 1 - * (Release 1.2.1) - */ - ctx->mr[ctx->mr_count] = gf_rdma_get_pre_registred_mr( - peer->trans, (void *)vector[i].iov_base, vector[i].iov_len); - - if (!ctx->mr[ctx->mr_count]) { - ctx->mr[ctx->mr_count] = ibv_reg_mr(device->pd, vector[i].iov_base, - vector[i].iov_len, - IBV_ACCESS_LOCAL_WRITE); - } - if (ctx->mr[ctx->mr_count] == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_MR_ALOC_FAILED, - "registering memory for IBV_ACCESS_LOCAL_WRITE" - " failed"); - goto out; - } - - ctx->mr_count++; - } - - ret = 0; -out: - return ret; -} - -/* 1. assumes xfer_len of data is pointed by vector(s) starting from vec[*idx] - * 2. modifies vec - */ -int32_t -__gf_rdma_write(gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *vec, - uint32_t xfer_len, int *idx, gf_rdma_write_chunk_t *writech) -{ - int size = 0, num_sge = 0, i = 0; - int32_t ret = -1; - struct ibv_sge *sg_list = NULL; - struct ibv_send_wr wr = - { - .opcode = IBV_WR_RDMA_WRITE, - .send_flags = IBV_SEND_SIGNALED, - }, - *bad_wr; - - if ((peer == NULL) || (writech == NULL) || (idx == NULL) || - (post == NULL) || (vec == NULL) || (xfer_len == 0)) { - goto out; - } - - for (i = *idx; size < xfer_len; i++) { - size += vec[i].iov_len; - } - - num_sge = i - *idx; - - sg_list = GF_CALLOC(num_sge, sizeof(struct ibv_sge), gf_common_mt_sge); - if (sg_list == NULL) { - ret = -1; - goto out; - } - - for ((i = *idx), (num_sge = 0); (xfer_len != 0); i++, num_sge++) { - size = min(xfer_len, vec[i].iov_len); - - sg_list[num_sge].addr = (unsigned long)vec[i].iov_base; - sg_list[num_sge].length = size; - sg_list[num_sge].lkey = post->ctx.mr[i]->lkey; - - xfer_len -= size; - } - - *idx = i; - - if (size < vec[i - 1].iov_len) { - vec[i - 1].iov_base += size; - vec[i - 1].iov_len -= size; - *idx = i - 1; - } - - wr.sg_list = sg_list; - wr.num_sge = num_sge; - wr.wr_id = (unsigned long)gf_rdma_post_ref(post); - wr.wr.rdma.rkey = writech->wc_target.rs_handle; - wr.wr.rdma.remote_addr = writech->wc_target.rs_offset; - - ret = ibv_post_send(peer->qp, &wr, &bad_wr); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_WRITE_CLIENT_ERROR, - "rdma write to " - "client (%s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - ret = -1; - } - - GF_FREE(sg_list); -out: - return ret; -} - -int32_t -__gf_rdma_do_gf_rdma_write(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - struct iovec *vector, int count, - struct iobref *iobref, - gf_rdma_reply_info_t *reply_info) -{ - int i = 0, payload_idx = 0; - uint32_t payload_size = 0, xfer_len = 0; - int32_t ret = -1; - - if (count != 0) { - payload_size = iov_length(vector, count); - } - - if (payload_size == 0) { - ret = 0; - goto out; - } - - ret = __gf_rdma_register_local_mr_for_rdma(peer, vector, count, &post->ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_MR_ALOC_FAILED, - "registering memory region for rdma failed"); - goto out; - } - - post->ctx.iobref = iobref_ref(iobref); - - for (i = 0; (i < reply_info->wc_array->wc_nchunks) && (payload_size != 0); - i++) { - xfer_len = min(payload_size, - reply_info->wc_array->wc_array[i].wc_target.rs_length); - - ret = __gf_rdma_write(peer, post, vector, xfer_len, &payload_idx, - &reply_info->wc_array->wc_array[i]); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_WRITE_CLIENT_ERROR, - "rdma write to " - "client (%s) failed", - peer->trans->peerinfo.identifier); - goto out; - } - - payload_size -= xfer_len; - } - - ret = 0; -out: - - return ret; -} - -int32_t -__gf_rdma_send_reply_type_nomsg(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post, - gf_rdma_reply_info_t *reply_info) -{ - gf_rdma_header_t *header = NULL; - char *buf = NULL; - uint32_t payload_size = 0; - int count = 0, i = 0; - int32_t ret = 0; - struct iovec vector[MAX_IOVEC]; - - header = (gf_rdma_header_t *)post->buf; - - __gf_rdma_fill_reply_header(header, entry->rpchdr, reply_info, - peer->send_count); - - header->rm_type = hton32(GF_RDMA_NOMSG); - - payload_size = iov_length(entry->rpchdr, entry->rpchdr_count) + - iov_length(entry->proghdr, entry->proghdr_count); - - /* encode reply chunklist */ - buf = (char *)&header->rm_body.rm_chunks[2]; - ret = __gf_rdma_reply_encode_write_chunks(peer, payload_size, post, - reply_info, (uint32_t **)&buf); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_ENCODE_ERROR, - "encoding write chunks failed"); - ret = __gf_rdma_send_error(peer, entry, post, reply_info, ERR_CHUNK); - goto out; - } - - gf_rdma_post_ref(post); - - for (i = 0; i < entry->rpchdr_count; i++) { - vector[count++] = entry->rpchdr[i]; - } - - for (i = 0; i < entry->proghdr_count; i++) { - vector[count++] = entry->proghdr[i]; - } - - ret = __gf_rdma_do_gf_rdma_write(peer, post, vector, count, entry->iobref, - reply_info); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_WRITE_PEER_FAILED, - "rdma write to peer " - "(%s) failed", - peer->trans->peerinfo.identifier); - gf_rdma_post_unref(post); - goto out; - } - - ret = gf_rdma_post_send(peer->qp, post, (buf - post->buf)); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_POST_SEND_FAILED, - "posting a send request " - "to client (%s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - ret = -1; - gf_rdma_post_unref(post); - } else { - ret = payload_size; - } - -out: - return ret; -} - -int32_t -__gf_rdma_send_reply_type_msg(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post, - gf_rdma_reply_info_t *reply_info) -{ - gf_rdma_header_t *header = NULL; - int32_t send_size = 0, ret = 0; - char *ptr = NULL; - uint32_t payload_size = 0; - - send_size = iov_length(entry->rpchdr, entry->rpchdr_count) + - iov_length(entry->proghdr, entry->proghdr_count) + - GLUSTERFS_RDMA_MAX_HEADER_SIZE; - - if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_SEND_SIZE_GREAT_INLINE_THRESHOLD, - "client has provided only write chunks, but the " - "combined size of rpc and program header (%d) is " - "exceeding the size of msg that can be sent using " - "RDMA send (%d)", - send_size, GLUSTERFS_RDMA_INLINE_THRESHOLD); - - ret = __gf_rdma_send_error(peer, entry, post, reply_info, ERR_CHUNK); - goto out; - } - - header = (gf_rdma_header_t *)post->buf; - - __gf_rdma_fill_reply_header(header, entry->rpchdr, reply_info, - peer->send_count); - - payload_size = iov_length(entry->prog_payload, entry->prog_payload_count); - ptr = (char *)&header->rm_body.rm_chunks[1]; - - ret = __gf_rdma_reply_encode_write_chunks(peer, payload_size, post, - reply_info, (uint32_t **)&ptr); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_ENCODE_ERROR, - "encoding write chunks failed"); - ret = __gf_rdma_send_error(peer, entry, post, reply_info, ERR_CHUNK); - goto out; - } - - *(uint32_t *)ptr = 0; /* terminate reply chunklist */ - ptr += sizeof(uint32_t); - - gf_rdma_post_ref(post); - - ret = __gf_rdma_do_gf_rdma_write(peer, post, entry->prog_payload, - entry->prog_payload_count, entry->iobref, - reply_info); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_WRITE_PEER_FAILED, - "rdma write to peer " - "(%s) failed", - peer->trans->peerinfo.identifier); - gf_rdma_post_unref(post); - goto out; - } - - iov_unload(ptr, entry->rpchdr, entry->rpchdr_count); - ptr += iov_length(entry->rpchdr, entry->rpchdr_count); - - iov_unload(ptr, entry->proghdr, entry->proghdr_count); - ptr += iov_length(entry->proghdr, entry->proghdr_count); - - ret = gf_rdma_post_send(peer->qp, post, (ptr - post->buf)); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_SEND_CLIENT_ERROR, - "rdma send to client (%s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - gf_rdma_post_unref(post); - ret = -1; - } else { - ret = send_size + payload_size; - } - -out: - return ret; -} - -void -gf_rdma_reply_info_destroy(gf_rdma_reply_info_t *reply_info) -{ - if (reply_info == NULL) { - goto out; - } - - if (reply_info->wc_array != NULL) { - GF_FREE(reply_info->wc_array); - reply_info->wc_array = NULL; - } - - mem_put(reply_info); -out: - return; -} - -gf_rdma_reply_info_t * -gf_rdma_reply_info_alloc(gf_rdma_peer_t *peer) -{ - gf_rdma_reply_info_t *reply_info = NULL; - gf_rdma_private_t *priv = NULL; - - priv = peer->trans->private; - - reply_info = mem_get(priv->device->reply_info_pool); - if (reply_info == NULL) { - goto out; - } - - memset(reply_info, 0, sizeof(*reply_info)); - reply_info->pool = priv->device->reply_info_pool; - -out: - return reply_info; -} - -int32_t -__gf_rdma_ioq_churn_reply(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, - gf_rdma_post_t *post) -{ - gf_rdma_reply_info_t *reply_info = NULL; - int32_t ret = -1; - gf_rdma_chunktype_t type = gf_rdma_noch; - - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, entry, out); - GF_VALIDATE_OR_GOTO(GF_RDMA_LOG_NAME, post, out); - - reply_info = entry->msg.reply_info; - if (reply_info != NULL) { - type = reply_info->type; - } - - switch (type) { - case gf_rdma_noch: - ret = __gf_rdma_send_reply_inline(peer, entry, post, reply_info); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_SEND_REPLY_FAILED, - "failed to send reply to peer (%s) as an " - "inlined rdma msg", - peer->trans->peerinfo.identifier); - } - break; - - case gf_rdma_replych: - ret = __gf_rdma_send_reply_type_nomsg(peer, entry, post, - reply_info); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_SEND_REPLY_FAILED, - "failed to send reply to peer (%s) as " - "RDMA_NOMSG", - peer->trans->peerinfo.identifier); - } - break; - - case gf_rdma_writech: - ret = __gf_rdma_send_reply_type_msg(peer, entry, post, reply_info); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_SEND_REPLY_FAILED, - "failed to send reply with write chunks " - "to peer (%s)", - peer->trans->peerinfo.identifier); - } - break; - - default: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_INVALID_CHUNK_TYPE, - "invalid chunktype (%d) specified for sending reply " - " (peer:%s)", - type, peer->trans->peerinfo.identifier); - break; - } - - if (reply_info != NULL) { - gf_rdma_reply_info_destroy(reply_info); - } -out: - return ret; -} - -int32_t -__gf_rdma_ioq_churn_entry(gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry) -{ - int32_t ret = 0, quota = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - gf_rdma_options_t *options = NULL; - gf_rdma_post_t *post = NULL; - - priv = peer->trans->private; - options = &priv->options; - device = priv->device; - - quota = __gf_rdma_quota_get(peer); - if (quota > 0) { - post = gf_rdma_get_post(&device->sendq); - if (post == NULL) { - post = gf_rdma_new_post(peer->trans, device, - (options->send_size + 2048), - GF_RDMA_SEND_POST); - } - - if (post == NULL) { - ret = -1; - gf_msg_callingfn(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_POST_SEND_FAILED, - "not able to get a post to send msg"); - goto out; - } - - if (entry->is_request) { - ret = __gf_rdma_ioq_churn_request(peer, entry, post); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PROC_IOQ_ENTRY_FAILED, - "failed to process request ioq entry " - "to peer(%s)", - peer->trans->peerinfo.identifier); - } - } else { - ret = __gf_rdma_ioq_churn_reply(peer, entry, post); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PROC_IOQ_ENTRY_FAILED, - "failed to process reply ioq entry " - "to peer (%s)", - peer->trans->peerinfo.identifier); - } - } - - if (ret != 0) { - __gf_rdma_ioq_entry_free(entry); - } - } else { - ret = 0; - } - -out: - return ret; -} - -static int32_t -__gf_rdma_ioq_churn(gf_rdma_peer_t *peer) -{ - gf_rdma_ioq_t *entry = NULL; - int32_t ret = 0; - - while (!list_empty(&peer->ioq)) { - /* pick next entry */ - entry = peer->ioq_next; - - ret = __gf_rdma_ioq_churn_entry(peer, entry); - - if (ret <= 0) - break; - } - - /* - list_for_each_entry_safe (entry, dummy, &peer->ioq, list) { - ret = __gf_rdma_ioq_churn_entry (peer, entry); - if (ret <= 0) { - break; - } - } - */ - - return ret; -} - -static int32_t -gf_rdma_writev(rpc_transport_t *this, gf_rdma_ioq_t *entry) -{ - int32_t ret = 0, need_append = 1; - gf_rdma_private_t *priv = NULL; - gf_rdma_peer_t *peer = NULL; - - priv = this->private; - pthread_mutex_lock(&priv->write_mutex); - { - if (!priv->connected) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_PEER_DISCONNECTED, - "rdma is not connected to peer (%s)", - this->peerinfo.identifier); - ret = -1; - goto unlock; - } - - peer = &priv->peer; - if (list_empty(&peer->ioq)) { - ret = __gf_rdma_ioq_churn_entry(peer, entry); - if (ret != 0) { - need_append = 0; - - if (ret < 0) { - gf_msg(this->name, GF_LOG_WARNING, 0, - RDMA_MSG_PROC_IOQ_ENTRY_FAILED, - "processing ioq entry destined" - " to (%s) failed", - this->peerinfo.identifier); - } - } - } - - if (need_append) { - list_add_tail(&entry->list, &peer->ioq); - } - } -unlock: - pthread_mutex_unlock(&priv->write_mutex); - return ret; -} - -gf_rdma_ioq_t * -gf_rdma_ioq_new(rpc_transport_t *this, rpc_transport_data_t *data) -{ - gf_rdma_ioq_t *entry = NULL; - int count = 0, i = 0; - rpc_transport_msg_t *msg = NULL; - gf_rdma_private_t *priv = NULL; - - if ((data == NULL) || (this == NULL)) { - goto out; - } - - priv = this->private; - - entry = mem_get(priv->device->ioq_pool); - if (entry == NULL) { - goto out; - } - memset(entry, 0, sizeof(*entry)); - entry->pool = priv->device->ioq_pool; - - if (data->is_request) { - msg = &data->data.req.msg; - if (data->data.req.rsp.rsphdr_count != 0) { - for (i = 0; i < data->data.req.rsp.rsphdr_count; i++) { - entry->msg.request.rsphdr_vec[i] = data->data.req.rsp.rsphdr[i]; - } - - entry->msg.request.rsphdr_count = data->data.req.rsp.rsphdr_count; - } - - if (data->data.req.rsp.rsp_payload_count != 0) { - for (i = 0; i < data->data.req.rsp.rsp_payload_count; i++) { - entry->msg.request.rsp_payload[i] = data->data.req.rsp - .rsp_payload[i]; - } - - entry->msg.request.rsp_payload_count = data->data.req.rsp - .rsp_payload_count; - } - - entry->msg.request.rpc_req = data->data.req.rpc_req; - - if (data->data.req.rsp.rsp_iobref != NULL) { - entry->msg.request.rsp_iobref = iobref_ref( - data->data.req.rsp.rsp_iobref); - } - } else { - msg = &data->data.reply.msg; - entry->msg.reply_info = data->data.reply.private; - } - - entry->is_request = data->is_request; - - count = msg->rpchdrcount + msg->proghdrcount + msg->progpayloadcount; - - GF_ASSERT(count <= MAX_IOVEC); - - if (msg->rpchdr != NULL) { - memcpy(&entry->rpchdr[0], msg->rpchdr, - sizeof(struct iovec) * msg->rpchdrcount); - entry->rpchdr_count = msg->rpchdrcount; - } - - if (msg->proghdr != NULL) { - memcpy(&entry->proghdr[0], msg->proghdr, - sizeof(struct iovec) * msg->proghdrcount); - entry->proghdr_count = msg->proghdrcount; - } - - if (msg->progpayload != NULL) { - memcpy(&entry->prog_payload[0], msg->progpayload, - sizeof(struct iovec) * msg->progpayloadcount); - entry->prog_payload_count = msg->progpayloadcount; - } - - if (msg->iobref != NULL) { - entry->iobref = iobref_ref(msg->iobref); - } - - INIT_LIST_HEAD(&entry->list); - -out: - return entry; -} - -int32_t -gf_rdma_submit_request(rpc_transport_t *this, rpc_transport_req_t *req) -{ - int32_t ret = 0; - gf_rdma_ioq_t *entry = NULL; - rpc_transport_data_t data = { - 0, - }; - gf_rdma_private_t *priv = NULL; - gf_rdma_peer_t *peer = NULL; - - if (req == NULL) { - goto out; - } - - priv = this->private; - if (priv == NULL) { - ret = -1; - goto out; - } - - peer = &priv->peer; - data.is_request = 1; - data.data.req = *req; - /* - * when fist message is received on a transport, quota variable will - * initiaize and quota_set will set to one. In gluster code client - * process with respect to transport is the one who sends the first - * message. Before settng quota_set variable if a submit request is - * came on server, then the message should not send. - */ - - if (priv->entity == GF_RDMA_SERVER && peer->quota_set == 0) { - ret = 0; - goto out; - } - - entry = gf_rdma_ioq_new(this, &data); - if (entry == NULL) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_NEW_IOQ_ENTRY_FAILED, - "getting a new ioq entry failed (peer:%s)", - this->peerinfo.identifier); - goto out; - } - - ret = gf_rdma_writev(this, entry); - - if (ret > 0) { - ret = 0; - } else if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_WRITE_PEER_FAILED, - "sending request to peer (%s) failed", - this->peerinfo.identifier); - rpc_transport_disconnect(this, _gf_false); - } - -out: - return ret; -} - -int32_t -gf_rdma_submit_reply(rpc_transport_t *this, rpc_transport_reply_t *reply) -{ - int32_t ret = 0; - gf_rdma_ioq_t *entry = NULL; - rpc_transport_data_t data = { - 0, - }; - - if (reply == NULL) { - goto out; - } - - data.data.reply = *reply; - - entry = gf_rdma_ioq_new(this, &data); - if (entry == NULL) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_NEW_IOQ_ENTRY_FAILED, - "getting a new ioq entry failed (peer:%s)", - this->peerinfo.identifier); - goto out; - } - - ret = gf_rdma_writev(this, entry); - if (ret > 0) { - ret = 0; - } else if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_WRITE_PEER_FAILED, - "sending request to peer (%s) failed", - this->peerinfo.identifier); - rpc_transport_disconnect(this, _gf_false); - } - -out: - return ret; -} - -static int -gf_rdma_register_peer(gf_rdma_device_t *device, int32_t qp_num, - gf_rdma_peer_t *peer) -{ - struct _qpent *ent = NULL; - gf_rdma_qpreg_t *qpreg = NULL; - int32_t hash = 0; - int ret = -1; - - qpreg = &device->qpreg; - hash = qp_num % 42; - - pthread_mutex_lock(&qpreg->lock); - { - ent = qpreg->ents[hash].next; - while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) { - ent = ent->next; - } - - if (ent->qp_num == qp_num) { - ret = 0; - goto unlock; - } - - ent = (struct _qpent *)GF_CALLOC(1, sizeof(*ent), gf_common_mt_qpent); - if (ent == NULL) { - goto unlock; - } - - /* TODO: ref reg->peer */ - ent->peer = peer; - ent->next = &qpreg->ents[hash]; - ent->prev = ent->next->prev; - ent->next->prev = ent; - ent->prev->next = ent; - ent->qp_num = qp_num; - qpreg->count++; - ret = 0; - } -unlock: - pthread_mutex_unlock(&qpreg->lock); - - return ret; -} - -static void -gf_rdma_unregister_peer(gf_rdma_device_t *device, int32_t qp_num) -{ - struct _qpent *ent = NULL; - gf_rdma_qpreg_t *qpreg = NULL; - int32_t hash = 0; - - qpreg = &device->qpreg; - hash = qp_num % 42; - - pthread_mutex_lock(&qpreg->lock); - { - ent = qpreg->ents[hash].next; - while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) - ent = ent->next; - if (ent->qp_num != qp_num) { - pthread_mutex_unlock(&qpreg->lock); - return; - } - ent->prev->next = ent->next; - ent->next->prev = ent->prev; - /* TODO: unref reg->peer */ - GF_FREE(ent); - qpreg->count--; - } - pthread_mutex_unlock(&qpreg->lock); -} - -static gf_rdma_peer_t * -__gf_rdma_lookup_peer(gf_rdma_device_t *device, int32_t qp_num) -{ - struct _qpent *ent = NULL; - gf_rdma_peer_t *peer = NULL; - gf_rdma_qpreg_t *qpreg = NULL; - int32_t hash = 0; - - qpreg = &device->qpreg; - hash = qp_num % 42; - ent = qpreg->ents[hash].next; - while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) - ent = ent->next; - - if (ent != &qpreg->ents[hash]) { - peer = ent->peer; - } - - return peer; -} - -static void -__gf_rdma_destroy_qp(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - - priv = this->private; - if (priv->peer.qp) { - gf_rdma_unregister_peer(priv->device, priv->peer.qp->qp_num); - rdma_destroy_qp(priv->peer.cm_id); - } - priv->peer.qp = NULL; - - return; -} - -static int32_t -gf_rdma_create_qp(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_device_t *device = NULL; - int32_t ret = 0; - gf_rdma_peer_t *peer = NULL; - char *device_name = NULL; - - priv = this->private; - - peer = &priv->peer; - - device_name = (char *)ibv_get_device_name(peer->cm_id->verbs->device); - if (device_name == NULL) { - ret = -1; - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_GET_DEVICE_NAME_FAILED, - "cannot get " - "device_name"); - goto out; - } - - device = gf_rdma_get_device(this, peer->cm_id->verbs, device_name); - if (device == NULL) { - ret = -1; - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_GET_DEVICE_FAILED, - "cannot get device for " - "device %s", - device_name); - goto out; - } - - if (priv->device == NULL) { - priv->device = device; - } - - struct ibv_qp_init_attr init_attr = { - .send_cq = device->send_cq, - .recv_cq = device->recv_cq, - .srq = device->srq, - .cap = {.max_send_wr = peer->send_count, - .max_recv_wr = peer->recv_count, - .max_send_sge = 2, - .max_recv_sge = 1}, - .qp_type = IBV_QPT_RC}; - - ret = rdma_create_qp(peer->cm_id, device->pd, &init_attr); - if (ret != 0) { - gf_msg(peer->trans->name, GF_LOG_CRITICAL, errno, - RDMA_MSG_CREAT_QP_FAILED, "%s: could not create QP", this->name); - ret = -1; - goto out; - } - - peer->qp = peer->cm_id->qp; - - ret = gf_rdma_register_peer(device, peer->qp->qp_num, peer); - -out: - if (ret == -1) - __gf_rdma_destroy_qp(this); - - return ret; -} - -static int32_t -__gf_rdma_teardown(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_peer_t *peer = NULL; - - priv = this->private; - peer = &priv->peer; - - if (peer->cm_id && peer->cm_id->qp != NULL) { - __gf_rdma_destroy_qp(this); - } - - if (!list_empty(&priv->peer.ioq)) { - __gf_rdma_ioq_flush(peer); - } - - if (peer->cm_id != NULL) { - rdma_destroy_id(peer->cm_id); - peer->cm_id = NULL; - } - - /* TODO: decrement cq size */ - return 0; -} - -static int32_t -gf_rdma_teardown(rpc_transport_t *this) -{ - int32_t ret = 0; - gf_rdma_private_t *priv = NULL; - - if (this == NULL) { - goto out; - } - - priv = this->private; - - pthread_mutex_lock(&priv->write_mutex); - { - ret = __gf_rdma_teardown(this); - } - pthread_mutex_unlock(&priv->write_mutex); - -out: - return ret; -} - -/* - * allocates new memory to hold write-chunklist. New memory is needed since - * write-chunklist will be used while sending reply and the post holding initial - * write-chunklist sent from client will be put back to srq before a pollin - * event is sent to upper layers. - */ -int32_t -gf_rdma_get_write_chunklist(char **ptr, gf_rdma_write_array_t **write_ary) -{ - gf_rdma_write_array_t *from = NULL, *to = NULL; - int32_t ret = -1, size = 0, i = 0; - - from = (gf_rdma_write_array_t *)*ptr; - if (from->wc_discrim == 0) { - ret = 0; - goto out; - } - - from->wc_nchunks = ntoh32(from->wc_nchunks); - - size = sizeof(*from) + (sizeof(gf_rdma_write_chunk_t) * from->wc_nchunks); - - to = GF_CALLOC(1, size, gf_common_mt_char); - if (to == NULL) { - ret = -1; - goto out; - } - - to->wc_discrim = ntoh32(from->wc_discrim); - to->wc_nchunks = from->wc_nchunks; - - for (i = 0; i < to->wc_nchunks; i++) { - to->wc_array[i].wc_target.rs_handle = ntoh32( - from->wc_array[i].wc_target.rs_handle); - to->wc_array[i].wc_target.rs_length = ntoh32( - from->wc_array[i].wc_target.rs_length); - to->wc_array[i].wc_target.rs_offset = ntoh64( - from->wc_array[i].wc_target.rs_offset); - } - - *write_ary = to; - ret = 0; - *ptr = (char *)&from->wc_array[i].wc_target.rs_handle; -out: - return ret; -} - -/* - * does not allocate new memory to hold read-chunklist. New memory is not - * needed, since post is not put back to srq till we've completed all the - * rdma-reads and hence readchunk-list can point to memory held by post. - */ -int32_t -gf_rdma_get_read_chunklist(char **ptr, gf_rdma_read_chunk_t **readch) -{ - int32_t ret = -1; - gf_rdma_read_chunk_t *chunk = NULL; - int i = 0; - - chunk = (gf_rdma_read_chunk_t *)*ptr; - if (chunk[0].rc_discrim == 0) { - ret = 0; - goto out; - } - - for (i = 0; chunk[i].rc_discrim != 0; i++) { - chunk[i].rc_discrim = ntoh32(chunk[i].rc_discrim); - chunk[i].rc_position = ntoh32(chunk[i].rc_position); - chunk[i].rc_target.rs_handle = ntoh32(chunk[i].rc_target.rs_handle); - chunk[i].rc_target.rs_length = ntoh32(chunk[i].rc_target.rs_length); - chunk[i].rc_target.rs_offset = ntoh64(chunk[i].rc_target.rs_offset); - } - - *readch = &chunk[0]; - ret = 0; - *ptr = (char *)&chunk[i].rc_discrim; -out: - return ret; -} - -static int32_t -gf_rdma_decode_error_msg(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - size_t bytes_in_post) -{ - gf_rdma_header_t *header = NULL; - struct iobuf *iobuf = NULL; - struct iobref *iobref = NULL; - int32_t ret = -1; - struct rpc_msg rpc_msg = { - 0, - }; - - header = (gf_rdma_header_t *)post->buf; - header->rm_body.rm_error.rm_type = ntoh32(header->rm_body.rm_error.rm_type); - if (header->rm_body.rm_error.rm_type == ERR_VERS) { - header->rm_body.rm_error.rm_version.gf_rdma_vers_low = ntoh32( - header->rm_body.rm_error.rm_version.gf_rdma_vers_low); - header->rm_body.rm_error.rm_version.gf_rdma_vers_high = ntoh32( - header->rm_body.rm_error.rm_version.gf_rdma_vers_high); - } - - rpc_msg.rm_xid = header->rm_xid; - rpc_msg.rm_direction = REPLY; - rpc_msg.rm_reply.rp_stat = MSG_DENIED; - - iobuf = iobuf_get2(peer->trans->ctx->iobuf_pool, bytes_in_post); - if (iobuf == NULL) { - ret = -1; - goto out; - } - - post->ctx.iobref = iobref = iobref_new(); - if (iobref == NULL) { - ret = -1; - goto out; - } - - iobref_add(iobref, iobuf); - iobuf_unref(iobuf); - - ret = rpc_reply_to_xdr(&rpc_msg, iobuf_ptr(iobuf), iobuf_pagesize(iobuf), - &post->ctx.vector[0]); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_RPC_REPLY_CREATE_FAILED, - "Failed to create " - "RPC reply"); - goto out; - } - - post->ctx.count = 1; - - iobuf = NULL; - iobref = NULL; - -out: - if (ret == -1) { - if (iobuf != NULL) { - iobuf_unref(iobuf); - } - - if (iobref != NULL) { - iobref_unref(iobref); - } - } - - return 0; -} - -int32_t -gf_rdma_decode_msg(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - gf_rdma_read_chunk_t **readch, size_t bytes_in_post) -{ - int32_t ret = -1; - gf_rdma_header_t *header = NULL; - gf_rdma_reply_info_t *reply_info = NULL; - char *ptr = NULL; - gf_rdma_write_array_t *write_ary = NULL; - size_t header_len = 0; - - header = (gf_rdma_header_t *)post->buf; - - ptr = (char *)&header->rm_body.rm_chunks[0]; - - ret = gf_rdma_get_read_chunklist(&ptr, readch); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_GET_READ_CHUNK_FAILED, - "cannot get read " - "chunklist from msg"); - goto out; - } - - /* skip terminator of read-chunklist */ - ptr = ptr + sizeof(uint32_t); - - ret = gf_rdma_get_write_chunklist(&ptr, &write_ary); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_GET_WRITE_CHUNK_FAILED, - "cannot get write " - "chunklist from msg"); - goto out; - } - - /* skip terminator of write-chunklist */ - ptr = ptr + sizeof(uint32_t); - - if (write_ary != NULL) { - reply_info = gf_rdma_reply_info_alloc(peer); - if (reply_info == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_REPLY_INFO_ALLOC_FAILED, "reply_info_alloc failed"); - ret = -1; - goto out; - } - - reply_info->type = gf_rdma_writech; - reply_info->wc_array = write_ary; - reply_info->rm_xid = header->rm_xid; - } else { - ret = gf_rdma_get_write_chunklist(&ptr, &write_ary); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_CHUNKLIST_ERROR, - "cannot get reply " - "chunklist from msg"); - goto out; - } - - if (write_ary != NULL) { - reply_info = gf_rdma_reply_info_alloc(peer); - if (reply_info == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_REPLY_INFO_ALLOC_FAILED, - "reply_info_alloc_failed"); - ret = -1; - goto out; - } - - reply_info->type = gf_rdma_replych; - reply_info->wc_array = write_ary; - reply_info->rm_xid = header->rm_xid; - } - } - - /* skip terminator of reply chunk */ - ptr = ptr + sizeof(uint32_t); - if (header->rm_type != GF_RDMA_NOMSG) { - header_len = (long)ptr - (long)post->buf; - post->ctx.vector[0].iov_len = (bytes_in_post - header_len); - - post->ctx.hdr_iobuf = iobuf_get2(peer->trans->ctx->iobuf_pool, - (bytes_in_post - header_len)); - if (post->ctx.hdr_iobuf == NULL) { - ret = -1; - goto out; - } - - post->ctx.vector[0].iov_base = iobuf_ptr(post->ctx.hdr_iobuf); - memcpy(post->ctx.vector[0].iov_base, ptr, post->ctx.vector[0].iov_len); - post->ctx.count = 1; - } - - post->ctx.reply_info = reply_info; -out: - if (ret == -1) { - if (*readch != NULL) { - GF_FREE(*readch); - *readch = NULL; - } - - GF_FREE(write_ary); - } - - return ret; -} - -/* Assumes only one of either write-chunklist or a reply chunk is present */ -int32_t -gf_rdma_decode_header(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - gf_rdma_read_chunk_t **readch, size_t bytes_in_post) -{ - int32_t ret = -1; - gf_rdma_header_t *header = NULL; - - header = (gf_rdma_header_t *)post->buf; - - header->rm_xid = ntoh32(header->rm_xid); - header->rm_vers = ntoh32(header->rm_vers); - header->rm_credit = ntoh32(header->rm_credit); - header->rm_type = ntoh32(header->rm_type); - - switch (header->rm_type) { - case GF_RDMA_MSG: - case GF_RDMA_NOMSG: - ret = gf_rdma_decode_msg(peer, post, readch, bytes_in_post); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_ENCODE_ERROR, - "cannot decode msg of " - "type (%d)", - header->rm_type); - } - - break; - - case GF_RDMA_MSGP: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_INVALID_ENTRY, - "rdma msg of msg-type " - "GF_RDMA_MSGP should not have been received"); - ret = -1; - break; - - case GF_RDMA_DONE: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_INVALID_ENTRY, - "rdma msg of msg-type " - "GF_RDMA_DONE should not have been received"); - ret = -1; - break; - - case GF_RDMA_ERROR: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_RDMA_ERROR_RECEIVED, - "received a msg of type" - " RDMA_ERROR"); - ret = gf_rdma_decode_error_msg(peer, post, bytes_in_post); - break; - - default: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_INVALID_ENTRY, - "unknown rdma msg-type (%d)", header->rm_type); - } - - return ret; -} - -int32_t -gf_rdma_do_reads(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - gf_rdma_read_chunk_t *readch) -{ - int32_t ret = -1, i = 0, count = 0; - size_t size = 0; - char *ptr = NULL; - struct iobuf *iobuf = NULL; - gf_rdma_private_t *priv = NULL; - struct ibv_sge *list = NULL; - struct ibv_send_wr *wr = NULL, *bad_wr = NULL; - int total_ref = 0; - priv = peer->trans->private; - - for (i = 0; readch[i].rc_discrim != 0; i++) { - size += readch[i].rc_target.rs_length; - } - - if (i == 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_INVALID_CHUNK_TYPE, - "message type specified " - "as rdma-read but there are no rdma read-chunks " - "present"); - goto out; - } - - post->ctx.gf_rdma_reads = i; - i = 0; - iobuf = iobuf_get2(peer->trans->ctx->iobuf_pool, size); - if (iobuf == NULL) { - goto out; - } - - if (post->ctx.iobref == NULL) { - post->ctx.iobref = iobref_new(); - if (post->ctx.iobref == NULL) { - iobuf_unref(iobuf); - iobuf = NULL; - goto out; - } - } - - iobref_add(post->ctx.iobref, iobuf); - iobuf_unref(iobuf); - - ptr = iobuf_ptr(iobuf); - iobuf = NULL; - - pthread_mutex_lock(&priv->write_mutex); - { - if (!priv->connected) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PEER_DISCONNECTED, - "transport not " - "connected to peer (%s), not doing rdma reads", - peer->trans->peerinfo.identifier); - goto unlock; - } - - list = GF_CALLOC(post->ctx.gf_rdma_reads, sizeof(struct ibv_sge), - gf_common_mt_sge); - - if (list == NULL) { - errno = ENOMEM; - ret = -1; - goto unlock; - } - wr = GF_CALLOC(post->ctx.gf_rdma_reads, sizeof(struct ibv_send_wr), - gf_common_mt_wr); - if (wr == NULL) { - errno = ENOMEM; - ret = -1; - goto unlock; - } - for (i = 0; readch[i].rc_discrim != 0; i++) { - count = post->ctx.count++; - post->ctx.vector[count].iov_base = ptr; - post->ctx.vector[count].iov_len = readch[i].rc_target.rs_length; - - ret = __gf_rdma_register_local_mr_for_rdma( - peer, &post->ctx.vector[count], 1, &post->ctx); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_MR_ALOC_FAILED, - "registering local memory" - " for rdma read failed"); - goto unlock; - } - - list[i].addr = (unsigned long)post->ctx.vector[count].iov_base; - list[i].length = post->ctx.vector[count].iov_len; - list[i].lkey = post->ctx.mr[post->ctx.mr_count - 1]->lkey; - - wr[i].wr_id = (unsigned long)gf_rdma_post_ref(post); - wr[i].sg_list = &list[i]; - wr[i].next = &wr[i + 1]; - wr[i].num_sge = 1; - wr[i].opcode = IBV_WR_RDMA_READ; - wr[i].send_flags = IBV_SEND_SIGNALED; - wr[i].wr.rdma.remote_addr = readch[i].rc_target.rs_offset; - wr[i].wr.rdma.rkey = readch[i].rc_target.rs_handle; - - ptr += readch[i].rc_target.rs_length; - total_ref++; - } - wr[i - 1].next = NULL; - ret = ibv_post_send(peer->qp, wr, &bad_wr); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_READ_CLIENT_ERROR, - "rdma read from " - "client (%s) failed with ret = %d (%s)", - peer->trans->peerinfo.identifier, ret, - (ret > 0) ? strerror(ret) : ""); - - if (!bad_wr) { - ret = -1; - goto unlock; - } - - for (i = 0; i < post->ctx.gf_rdma_reads; i++) { - if (&wr[i] != bad_wr) - total_ref--; - else - break; - } - - ret = -1; - } - } -unlock: - pthread_mutex_unlock(&priv->write_mutex); -out: - if (list) - GF_FREE(list); - if (wr) - GF_FREE(wr); - - if (ret == -1) { - while (total_ref-- > 0) - gf_rdma_post_unref(post); - } - - return ret; -} - -int32_t -gf_rdma_pollin_notify(gf_rdma_peer_t *peer, gf_rdma_post_t *post) -{ - int32_t ret = -1; - enum msg_type msg_type = 0; - struct rpc_req *rpc_req = NULL; - gf_rdma_request_context_t *request_context = NULL; - rpc_request_info_t request_info = { - 0, - }; - gf_rdma_private_t *priv = NULL; - uint32_t *ptr = NULL; - rpc_transport_pollin_t *pollin = NULL; - - if ((peer == NULL) || (post == NULL)) { - goto out; - } - - if (post->ctx.iobref == NULL) { - post->ctx.iobref = iobref_new(); - if (post->ctx.iobref == NULL) { - goto out; - } - - /* handling the case where both hdr and payload of - * GF_FOP_READ_CBK were received in a single iobuf - * because of server sending entire msg as inline without - * doing rdma writes. - */ - if (post->ctx.hdr_iobuf) - iobref_add(post->ctx.iobref, post->ctx.hdr_iobuf); - } - - pollin = rpc_transport_pollin_alloc(peer->trans, post->ctx.vector, - post->ctx.count, post->ctx.hdr_iobuf, - post->ctx.iobref, post->ctx.reply_info); - if (pollin == NULL) { - goto out; - } - - ptr = (uint32_t *)pollin->vector[0].iov_base; - - request_info.xid = ntoh32(*ptr); - msg_type = ntoh32(*(ptr + 1)); - - if (msg_type == REPLY) { - ret = rpc_transport_notify(peer->trans, RPC_TRANSPORT_MAP_XID_REQUEST, - &request_info); - if (ret == -1) { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "cannot get request" - "information from rpc layer"); - goto out; - } - - rpc_req = request_info.rpc_req; - if (rpc_req == NULL) { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "rpc request " - "structure not found"); - ret = -1; - goto out; - } - - request_context = rpc_req->conn_private; - rpc_req->conn_private = NULL; - - priv = peer->trans->private; - if (request_context != NULL) { - pthread_mutex_lock(&priv->write_mutex); - { - __gf_rdma_request_context_destroy(request_context); - } - pthread_mutex_unlock(&priv->write_mutex); - } else { - gf_rdma_quota_put(peer); - } - - pollin->is_reply = 1; - } - - ret = rpc_transport_notify(peer->trans, RPC_TRANSPORT_MSG_RECEIVED, pollin); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, TRANS_MSG_TRANSPORT_ERROR, - "transport_notify failed"); - } - -out: - if (pollin != NULL) { - pollin->private = NULL; - rpc_transport_pollin_destroy(pollin); - } - - return ret; -} - -int32_t -gf_rdma_recv_reply(gf_rdma_peer_t *peer, gf_rdma_post_t *post) -{ - int32_t ret = -1; - gf_rdma_header_t *header = NULL; - gf_rdma_reply_info_t *reply_info = NULL; - gf_rdma_write_array_t *wc_array = NULL; - int i = 0; - uint32_t *ptr = NULL; - gf_rdma_request_context_t *ctx = NULL; - rpc_request_info_t request_info = { - 0, - }; - struct rpc_req *rpc_req = NULL; - - header = (gf_rdma_header_t *)post->buf; - reply_info = post->ctx.reply_info; - - /* no write chunklist, just notify upper layers */ - if (reply_info == NULL) { - ret = 0; - goto out; - } - - wc_array = reply_info->wc_array; - - if (header->rm_type == GF_RDMA_NOMSG) { - post->ctx.vector[0].iov_base = (void *)(long)wc_array->wc_array[0] - .wc_target.rs_offset; - post->ctx.vector[0].iov_len = wc_array->wc_array[0].wc_target.rs_length; - - post->ctx.count = 1; - } else { - for (i = 0; i < wc_array->wc_nchunks; i++) { - post->ctx.vector[i + 1].iov_base = - (void *)(long)wc_array->wc_array[i].wc_target.rs_offset; - post->ctx.vector[i + 1].iov_len = wc_array->wc_array[i] - .wc_target.rs_length; - } - - post->ctx.count += wc_array->wc_nchunks; - } - - ptr = (uint32_t *)post->ctx.vector[0].iov_base; - request_info.xid = ntoh32(*ptr); - - ret = rpc_transport_notify(peer->trans, RPC_TRANSPORT_MAP_XID_REQUEST, - &request_info); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, TRANS_MSG_TRANSPORT_ERROR, - "cannot get request " - "information (peer:%s) from rpc layer", - peer->trans->peerinfo.identifier); - goto out; - } - - rpc_req = request_info.rpc_req; - if (rpc_req == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_RPC_ST_ERROR, - "rpc request structure not " - "found"); - ret = -1; - goto out; - } - - ctx = rpc_req->conn_private; - if ((post->ctx.iobref == NULL) && ctx->rsp_iobref) { - post->ctx.iobref = iobref_ref(ctx->rsp_iobref); - } - - ret = 0; - - gf_rdma_reply_info_destroy(reply_info); - -out: - if (ret == 0) { - ret = gf_rdma_pollin_notify(peer, post); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_POLL_IN_NOTIFY_FAILED, "pollin notify failed"); - } - } - - return ret; -} - -static int32_t -gf_rdma_recv_request(gf_rdma_peer_t *peer, gf_rdma_post_t *post, - gf_rdma_read_chunk_t *readch) -{ - int32_t ret = -1; - - if (readch != NULL) { - ret = gf_rdma_do_reads(peer, post, readch); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PEER_READ_FAILED, "rdma read from peer (%s) failed", - peer->trans->peerinfo.identifier); - } - } else { - ret = gf_rdma_pollin_notify(peer, post); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_POLL_IN_NOTIFY_FAILED, - "pollin notification failed"); - } - } - - return ret; -} - -void -gf_rdma_process_recv(gf_rdma_peer_t *peer, struct ibv_wc *wc) -{ - gf_rdma_post_t *post = NULL; - gf_rdma_read_chunk_t *readch = NULL; - int ret = -1; - uint32_t *ptr = NULL; - enum msg_type msg_type = 0; - gf_rdma_header_t *header = NULL; - gf_rdma_private_t *priv = NULL; - - post = (gf_rdma_post_t *)(long)wc->wr_id; - if (post == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_POST_MISSING, - "no post found in successful " - "work completion element"); - goto out; - } - - ret = gf_rdma_decode_header(peer, post, &readch, wc->byte_len); - if (ret == -1) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_HEADER_DECODE_FAILED, - "decoding of header " - "failed"); - goto out; - } - - header = (gf_rdma_header_t *)post->buf; - - priv = peer->trans->private; - - pthread_mutex_lock(&priv->write_mutex); - { - if (!priv->peer.quota_set) { - priv->peer.quota_set = 1; - - /* Initially peer.quota is set to 1 as per RFC 5666. We - * have to account for the quota used while sending - * first msg (which may or may not be returned to pool - * at this point) while deriving peer.quota from - * header->rm_credit. Hence the arithmetic below, - * instead of directly setting it to header->rm_credit. - */ - priv->peer.quota = header->rm_credit - (1 - priv->peer.quota); - } - } - pthread_mutex_unlock(&priv->write_mutex); - - switch (header->rm_type) { - case GF_RDMA_MSG: - ptr = (uint32_t *)post->ctx.vector[0].iov_base; - msg_type = ntoh32(*(ptr + 1)); - break; - - case GF_RDMA_NOMSG: - if (readch != NULL) { - msg_type = CALL; - } else { - msg_type = REPLY; - } - break; - - case GF_RDMA_ERROR: - if (header->rm_body.rm_error.rm_type == ERR_CHUNK) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_RDMA_ERROR_RECEIVED, - "peer (%s), couldn't encode or decode the msg " - "properly or write chunks were not provided " - "for replies that were bigger than " - "RDMA_INLINE_THRESHOLD (%d)", - peer->trans->peerinfo.identifier, - GLUSTERFS_RDMA_INLINE_THRESHOLD); - ret = gf_rdma_pollin_notify(peer, post); - if (ret == -1) { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "pollin " - "notification failed"); - } - goto out; - } else { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, 0, - TRANS_MSG_TRANSPORT_ERROR, - "an error has " - "happened while transmission of msg, " - "disconnecting the transport"); - ret = -1; - goto out; - } - - default: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_INVALID_ENTRY, - "invalid rdma msg-type (%d)", header->rm_type); - goto out; - } - - if (msg_type == CALL) { - ret = gf_rdma_recv_request(peer, post, readch); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PEER_REQ_FAILED, - "receiving a request" - " from peer (%s) failed", - peer->trans->peerinfo.identifier); - } - } else { - ret = gf_rdma_recv_reply(peer, post); - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_PEER_REP_FAILED, - "receiving a reply " - "from peer (%s) failed", - peer->trans->peerinfo.identifier); - } - } - -out: - if (ret == -1) { - rpc_transport_disconnect(peer->trans, _gf_false); - } - - return; -} - -void * -gf_rdma_async_event_thread(void *context) -{ - struct ibv_async_event event; - int ret; - - while (1) { - do { - ret = ibv_get_async_event((struct ibv_context *)context, &event); - - if (ret && errno != EINTR) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_EVENT_ERROR, - "Error getting " - "event"); - } - } while (ret && errno == EINTR); - - switch (event.event_type) { - case IBV_EVENT_SRQ_LIMIT_REACHED: - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_EVENT_SRQ_LIMIT_REACHED, - "received " - "srq_limit reached"); - break; - - default: - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "event (%d) " - "received", - event.event_type); - break; - } - - ibv_ack_async_event(&event); - } - - return 0; -} - -static void * -gf_rdma_recv_completion_proc(void *data) -{ - struct ibv_comp_channel *chan = NULL; - gf_rdma_device_t *device = NULL; - ; - gf_rdma_post_t *post = NULL; - gf_rdma_peer_t *peer = NULL; - struct ibv_cq *event_cq = NULL; - struct ibv_wc wc[10] = { - {0}, - }; - void *event_ctx = NULL; - int32_t ret = 0; - int32_t num_wr = 0, index = 0; - uint8_t failed = 0; - - chan = data; - - while (1) { - failed = 0; - ret = ibv_get_cq_event(chan, &event_cq, &event_ctx); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_GET_CQ_FAILED, - "ibv_get_cq_event failed, terminating recv " - "thread %d (%d)", - ret, errno); - continue; - } - - device = event_ctx; - - ret = ibv_req_notify_cq(event_cq, 0); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_REQ_NOTIFY_CQ_FAILED, - "ibv_req_notify_cq on %s failed, terminating " - "recv thread: %d (%d)", - device->device_name, ret, errno); - continue; - } - - device = (gf_rdma_device_t *)event_ctx; - - while (!failed && (num_wr = ibv_poll_cq(event_cq, 10, wc)) > 0) { - for (index = 0; index < num_wr && !failed; index++) { - post = (gf_rdma_post_t *)(long)wc[index].wr_id; - - pthread_mutex_lock(&device->qpreg.lock); - { - peer = __gf_rdma_lookup_peer(device, wc[index].qp_num); - - /* - * keep a refcount on transport so that it - * does not get freed because of some error - * indicated by wc.status till we are done - * with usage of peer and thereby that of - * trans. - */ - if (peer != NULL) { - rpc_transport_ref(peer->trans); - } - } - pthread_mutex_unlock(&device->qpreg.lock); - - if (wc[index].status != IBV_WC_SUCCESS) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, 0, - RDMA_MSG_RECV_ERROR, - "recv work " - "request on `%s' returned error (%d)", - device->device_name, wc[index].status); - failed = 1; - if (peer) { - ibv_ack_cq_events(event_cq, num_wr); - rpc_transport_disconnect(peer->trans, _gf_false); - rpc_transport_unref(peer->trans); - } - - if (post) { - gf_rdma_post_unref(post); - } - - continue; - } - - if (peer) { - gf_rdma_process_recv(peer, &wc[index]); - rpc_transport_unref(peer->trans); - } else { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "could not lookup peer " - "for qp_num: %d", - wc[index].qp_num); - } - - gf_rdma_post_unref(post); - } - } - - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_POLL_CQ_ERROR, - "ibv_poll_cq on `%s' returned error " - "(ret = %d, errno = %d)", - device->device_name, ret, errno); - continue; - } - if (!failed) - ibv_ack_cq_events(event_cq, num_wr); - } - - return NULL; -} - -void -gf_rdma_handle_failed_send_completion(gf_rdma_peer_t *peer, struct ibv_wc *wc) -{ - gf_rdma_post_t *post = NULL; - gf_rdma_device_t *device = NULL; - gf_rdma_private_t *priv = NULL; - - if (peer != NULL) { - priv = peer->trans->private; - if (priv != NULL) { - device = priv->device; - } - } - - post = (gf_rdma_post_t *)(long)wc->wr_id; - - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, RDMA_MSG_RDMA_HANDLE_FAILED, - "send work request on `%s' returned error " - "wc.status = %d, wc.vendor_err = %d, post->buf = %p, " - "wc.byte_len = %d, post->reused = %d", - (device != NULL) ? device->device_name : NULL, wc->status, - wc->vendor_err, post->buf, wc->byte_len, post->reused); - - if (wc->status == IBV_WC_RETRY_EXC_ERR) { - gf_msg("rdma", GF_LOG_ERROR, 0, TRANS_MSG_TIMEOUT_EXCEEDED, - "connection between client and server not working. " - "check by running 'ibv_srq_pingpong'. also make sure " - "subnet manager is running (eg: 'opensm'), or check " - "if rdma port is valid (or active) by running " - "'ibv_devinfo'. contact Gluster Support Team if the " - "problem persists."); - } - - if (peer) { - rpc_transport_disconnect(peer->trans, _gf_false); - } - - return; -} - -void -gf_rdma_handle_successful_send_completion(gf_rdma_peer_t *peer, - struct ibv_wc *wc) -{ - gf_rdma_post_t *post = NULL; - int reads = 0, ret = 0; - gf_rdma_header_t *header = NULL; - - if (wc->opcode != IBV_WC_RDMA_READ) { - goto out; - } - - post = (gf_rdma_post_t *)(long)wc->wr_id; - - pthread_mutex_lock(&post->lock); - { - reads = --post->ctx.gf_rdma_reads; - } - pthread_mutex_unlock(&post->lock); - - if (reads != 0) { - /* if it is not the last rdma read, we've got nothing to do */ - goto out; - } - - header = (gf_rdma_header_t *)post->buf; - - if (header->rm_type == GF_RDMA_NOMSG) { - post->ctx.count = 1; - post->ctx.vector[0].iov_len += post->ctx.vector[1].iov_len; - } - /* - * if reads performed as vectored, then all the buffers are actually - * contiguous memory, so that we can use it as single vector, instead - * of multiple. - */ - while (post->ctx.count > 2) { - post->ctx.vector[1].iov_len += post->ctx.vector[post->ctx.count - 1] - .iov_len; - post->ctx.count--; - } - - ret = gf_rdma_pollin_notify(peer, post); - if ((ret == -1) && (peer != NULL)) { - rpc_transport_disconnect(peer->trans, _gf_false); - } - -out: - return; -} - -static void * -gf_rdma_send_completion_proc(void *data) -{ - struct ibv_comp_channel *chan = NULL; - gf_rdma_post_t *post = NULL; - gf_rdma_peer_t *peer = NULL; - struct ibv_cq *event_cq = NULL; - void *event_ctx = NULL; - gf_rdma_device_t *device = NULL; - struct ibv_wc wc[10] = { - {0}, - }; - char is_request = 0; - int32_t ret = 0, quota_ret = 0, num_wr = 0; - int32_t index = 0, failed = 0; - chan = data; - while (1) { - failed = 0; - ret = ibv_get_cq_event(chan, &event_cq, &event_ctx); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_GET_CQ_FAILED, - "ibv_get_cq_event on failed, terminating " - "send thread: %d (%d)", - ret, errno); - continue; - } - - device = event_ctx; - - ret = ibv_req_notify_cq(event_cq, 0); - if (ret) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_REQ_NOTIFY_CQ_FAILED, - "ibv_req_notify_cq on %s failed, terminating " - "send thread: %d (%d)", - device->device_name, ret, errno); - continue; - } - - while (!failed && (num_wr = ibv_poll_cq(event_cq, 10, wc)) > 0) { - for (index = 0; index < num_wr && !failed; index++) { - post = (gf_rdma_post_t *)(long)wc[index].wr_id; - - pthread_mutex_lock(&device->qpreg.lock); - { - peer = __gf_rdma_lookup_peer(device, wc[index].qp_num); - - /* - * keep a refcount on transport so that it - * does not get freed because of some error - * indicated by wc.status, till we are done - * with usage of peer and thereby that of trans. - */ - if (peer != NULL) { - rpc_transport_ref(peer->trans); - } - } - pthread_mutex_unlock(&device->qpreg.lock); - - if (wc[index].status != IBV_WC_SUCCESS) { - ibv_ack_cq_events(event_cq, num_wr); - failed = 1; - gf_rdma_handle_failed_send_completion(peer, &wc[index]); - } else { - gf_rdma_handle_successful_send_completion(peer, &wc[index]); - } - - if (post) { - is_request = post->ctx.is_request; - - ret = gf_rdma_post_unref(post); - if ((ret == 0) && (wc[index].status == IBV_WC_SUCCESS) && - !is_request && (post->type == GF_RDMA_SEND_POST) && - (peer != NULL)) { - /* An GF_RDMA_RECV_POST can end up in - * gf_rdma_send_completion_proc for - * rdma-reads, and we do not take - * quota for getting an GF_RDMA_RECV_POST. - */ - - /* - * if it is request, quota is returned - * after reply has come. - */ - quota_ret = gf_rdma_quota_put(peer); - if (quota_ret < 0) { - gf_msg_debug("rdma", 0, - "failed to send " - "message"); - } - } - } - - if (peer) { - rpc_transport_unref(peer->trans); - } else { - gf_msg_debug(GF_RDMA_LOG_NAME, 0, - "could not lookup peer for qp_num: %d", - wc[index].qp_num); - } - } - } - - if (ret < 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_ERROR, errno, - RDMA_MSG_IBV_POLL_CQ_ERROR, - "ibv_poll_cq on `%s' returned error (ret = %d," - " errno = %d)", - device->device_name, ret, errno); - continue; - } - if (!failed) - ibv_ack_cq_events(event_cq, num_wr); - } - - return NULL; -} - -static void -gf_rdma_options_init(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_options_t *options = NULL; - int32_t mtu = 0; - data_t *temp = NULL; - - /* TODO: validate arguments from options below */ - - priv = this->private; - options = &priv->options; - options->send_size = - GLUSTERFS_RDMA_INLINE_THRESHOLD; /*this->ctx->page_size * 4; 512 KB*/ - options->recv_size = - GLUSTERFS_RDMA_INLINE_THRESHOLD; /*this->ctx->page_size * 4; 512 KB*/ - options->send_count = 4096; - options->recv_count = 4096; - options->attr_timeout = GF_RDMA_TIMEOUT; - options->attr_retry_cnt = GF_RDMA_RETRY_CNT; - options->attr_rnr_retry = GF_RDMA_RNR_RETRY; - - temp = dict_get(this->options, "transport.listen-backlog"); - if (temp) - options->backlog = data_to_uint32(temp); - else - options->backlog = GLUSTERFS_SOCKET_LISTEN_BACKLOG; - - temp = dict_get(this->options, "transport.rdma.work-request-send-count"); - if (temp) - options->send_count = data_to_int32(temp); - - temp = dict_get(this->options, "transport.rdma.work-request-recv-count"); - if (temp) - options->recv_count = data_to_int32(temp); - - temp = dict_get(this->options, "transport.rdma.attr-timeout"); - - if (temp) - options->attr_timeout = data_to_uint8(temp); - - temp = dict_get(this->options, "transport.rdma.attr-retry-cnt"); - - if (temp) - options->attr_retry_cnt = data_to_uint8(temp); - - temp = dict_get(this->options, "transport.rdma.attr-rnr-retry"); - - if (temp) - options->attr_rnr_retry = data_to_uint8(temp); - - options->port = 1; - temp = dict_get(this->options, "transport.rdma.port"); - if (temp) - options->port = data_to_uint64(temp); - - options->mtu = mtu = IBV_MTU_2048; - temp = dict_get(this->options, "transport.rdma.mtu"); - if (temp) - mtu = data_to_int32(temp); - switch (mtu) { - case 256: - options->mtu = IBV_MTU_256; - break; - - case 512: - options->mtu = IBV_MTU_512; - break; - - case 1024: - options->mtu = IBV_MTU_1024; - break; - - case 2048: - options->mtu = IBV_MTU_2048; - break; - - case 4096: - options->mtu = IBV_MTU_4096; - break; - default: - if (temp) - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, 0, - RDMA_MSG_UNRECG_MTU_VALUE, - "%s: unrecognized " - "MTU value '%s', defaulting to '2048'", - this->name, data_to_str(temp)); - else - gf_msg_trace(GF_RDMA_LOG_NAME, 0, - "%s: defaulting " - "MTU to '2048'", - this->name); - options->mtu = IBV_MTU_2048; - break; - } - - temp = dict_get(this->options, "transport.rdma.device-name"); - if (temp) - options->device_name = gf_strdup(temp->data); - - return; -} - -gf_rdma_ctx_t * -__gf_rdma_ctx_create(void) -{ - gf_rdma_ctx_t *rdma_ctx = NULL; - int ret = -1; - - rdma_ctx = GF_CALLOC(1, sizeof(*rdma_ctx), gf_common_mt_char); - if (rdma_ctx == NULL) { - goto out; - } - pthread_mutex_init(&rdma_ctx->lock, NULL); - rdma_ctx->rdma_cm_event_channel = rdma_create_event_channel(); - if (rdma_ctx->rdma_cm_event_channel == NULL) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, errno, - RDMA_MSG_CM_EVENT_FAILED, - "rdma_cm event channel " - "creation failed"); - goto out; - } - - ret = gf_thread_create(&rdma_ctx->rdma_cm_thread, NULL, - gf_rdma_cm_event_handler, - rdma_ctx->rdma_cm_event_channel, "rdmaehan"); - if (ret != 0) { - gf_msg(GF_RDMA_LOG_NAME, GF_LOG_WARNING, ret, RDMA_MSG_CM_EVENT_FAILED, - "creation of thread to " - "handle rdma-cm events failed"); - goto out; - } - -out: - if (ret < 0 && rdma_ctx) { - if (rdma_ctx->rdma_cm_event_channel != NULL) { - rdma_destroy_event_channel(rdma_ctx->rdma_cm_event_channel); - } - - GF_FREE(rdma_ctx); - rdma_ctx = NULL; - } - - return rdma_ctx; -} - -static int32_t -gf_rdma_init(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - int32_t ret = 0; - glusterfs_ctx_t *ctx = NULL; - gf_rdma_options_t *options = NULL; - - ctx = this->ctx; - - priv = this->private; - - ibv_fork_init(); - gf_rdma_options_init(this); - - options = &priv->options; - priv->peer.send_count = options->send_count; - priv->peer.recv_count = options->recv_count; - priv->peer.send_size = options->send_size; - priv->peer.recv_size = options->recv_size; - priv->backlog = options->backlog; - - priv->peer.trans = this; - INIT_LIST_HEAD(&priv->peer.ioq); - - pthread_mutex_init(&priv->write_mutex, NULL); - pthread_mutex_init(&priv->recv_mutex, NULL); - pthread_cond_init(&priv->recv_cond, NULL); - - LOCK(&ctx->lock); - { - if (ctx->ib == NULL) { - ctx->ib = __gf_rdma_ctx_create(); - if (ctx->ib == NULL) { - ret = -1; - } - } - } - UNLOCK(&ctx->lock); - - return ret; -} - -static int32_t -gf_rdma_disconnect(rpc_transport_t *this, gf_boolean_t wait) -{ - gf_rdma_private_t *priv = NULL; - int32_t ret = 0; - - priv = this->private; - gf_msg_callingfn(this->name, GF_LOG_DEBUG, 0, RDMA_MSG_PEER_DISCONNECTED, - "disconnect called (peer:%s)", this->peerinfo.identifier); - - pthread_mutex_lock(&priv->write_mutex); - { - ret = __gf_rdma_disconnect(this); - } - pthread_mutex_unlock(&priv->write_mutex); - - return ret; -} - -static int32_t -gf_rdma_connect(struct rpc_transport *this, int port) -{ - gf_rdma_private_t *priv = NULL; - int32_t ret = 0; - union gf_sock_union sock_union = { - { - 0, - }, - }; - socklen_t sockaddr_len = 0; - gf_rdma_peer_t *peer = NULL; - gf_rdma_ctx_t *rdma_ctx = NULL; - gf_boolean_t connected = _gf_false; - - priv = this->private; - - peer = &priv->peer; - - rpc_transport_ref(this); - - ret = gf_rdma_client_get_remote_sockaddr(this, &sock_union.sa, - &sockaddr_len, port); - if (ret != 0) { - gf_msg_debug(this->name, 0, - "cannot get remote address to " - "connect"); - goto out; - } - - rdma_ctx = this->ctx->ib; - - pthread_mutex_lock(&priv->write_mutex); - { - if (peer->cm_id != NULL) { - ret = -1; - errno = EINPROGRESS; - connected = _gf_true; - goto unlock; - } - - priv->entity = GF_RDMA_CLIENT; - - ret = rdma_create_id(rdma_ctx->rdma_cm_event_channel, &peer->cm_id, - this, RDMA_PS_TCP); - if (ret != 0) { - gf_msg(this->name, GF_LOG_ERROR, errno, RDMA_MSG_CM_EVENT_FAILED, - "creation of " - "rdma_cm_id failed"); - ret = -errno; - goto unlock; - } - - memcpy(&this->peerinfo.sockaddr, &sock_union.storage, sockaddr_len); - this->peerinfo.sockaddr_len = sockaddr_len; - - if (port > 0) - sock_union.sin.sin_port = htons(port); - - ((struct sockaddr *)&this->myinfo.sockaddr)->sa_family = - ((struct sockaddr *)&this->peerinfo.sockaddr)->sa_family; - - ret = gf_rdma_client_bind(this, - (struct sockaddr *)&this->myinfo.sockaddr, - &this->myinfo.sockaddr_len, peer->cm_id); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, - RDMA_MSG_CLIENT_BIND_FAILED, "client bind failed"); - goto unlock; - } - - ret = rdma_resolve_addr(peer->cm_id, NULL, &sock_union.sa, 2000); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, - RDMA_MSG_RDMA_RESOLVE_ADDR_FAILED, - "rdma_resolve_addr failed"); - goto unlock; - } - - priv->connected = 0; - } -unlock: - pthread_mutex_unlock(&priv->write_mutex); - -out: - if (ret != 0) { - if (!connected) { - gf_rdma_teardown(this); - } - - rpc_transport_unref(this); - } - - return ret; -} - -static int32_t -gf_rdma_listen(rpc_transport_t *this) -{ - union gf_sock_union sock_union = { - { - 0, - }, - }; - socklen_t sockaddr_len = 0; - gf_rdma_private_t *priv = NULL; - gf_rdma_peer_t *peer = NULL; - int ret = 0; - gf_rdma_ctx_t *rdma_ctx = NULL; - char service[NI_MAXSERV], host[NI_MAXHOST]; - int optval = 2; - - priv = this->private; - peer = &priv->peer; - - priv->entity = GF_RDMA_SERVER_LISTENER; - - rdma_ctx = this->ctx->ib; - - ret = gf_rdma_server_get_local_sockaddr(this, &sock_union.sa, - &sockaddr_len); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_NW_ADDR_UNKNOWN, - "cannot find network address of server to bind to"); - goto err; - } - - ret = rdma_create_id(rdma_ctx->rdma_cm_event_channel, &peer->cm_id, this, - RDMA_PS_TCP); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_CM_EVENT_FAILED, - "creation of rdma_cm_id " - "failed"); - goto err; - } - - memcpy(&this->myinfo.sockaddr, &sock_union.storage, sockaddr_len); - this->myinfo.sockaddr_len = sockaddr_len; - - ret = getnameinfo((struct sockaddr *)&this->myinfo.sockaddr, - this->myinfo.sockaddr_len, host, sizeof(host), service, - sizeof(service), NI_NUMERICHOST); - if (ret != 0) { - gf_msg(this->name, GF_LOG_ERROR, ret, TRANS_MSG_GET_NAME_INFO_FAILED, - "getnameinfo failed"); - goto err; - } - - sprintf(this->myinfo.identifier, "%s:%s", host, service); - - ret = rdma_set_option(peer->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR, - (void *)&optval, sizeof(optval)); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_OPTION_SET_FAILED, - "rdma option set failed"); - goto err; - } - - ret = rdma_bind_addr(peer->cm_id, &sock_union.sa); - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, - RDMA_MSG_RDMA_BIND_ADDR_FAILED, "rdma_bind_addr failed"); - goto err; - } - - ret = rdma_listen(peer->cm_id, priv->backlog); - - if (ret != 0) { - gf_msg(this->name, GF_LOG_WARNING, errno, RDMA_MSG_LISTEN_FAILED, - "rdma_listen failed"); - goto err; - } - - rpc_transport_ref(this); - - ret = 0; -err: - if (ret < 0) { - if (peer->cm_id != NULL) { - rdma_destroy_id(peer->cm_id); - peer->cm_id = NULL; - } - } - - return ret; -} - -struct rpc_transport_ops tops = { - .submit_request = gf_rdma_submit_request, - .submit_reply = gf_rdma_submit_reply, - .connect = gf_rdma_connect, - .disconnect = gf_rdma_disconnect, - .listen = gf_rdma_listen, -}; - -int32_t -init(rpc_transport_t *this) -{ - gf_rdma_private_t *priv = NULL; - gf_rdma_ctx_t *rdma_ctx = NULL; - struct iobuf_pool *iobuf_pool = NULL; - - priv = GF_CALLOC(1, sizeof(*priv), gf_common_mt_rdma_private_t); - if (!priv) - return -1; - - this->private = priv; - - if (gf_rdma_init(this)) { - gf_msg(this->name, GF_LOG_WARNING, 0, RDMA_MSG_INIT_IB_DEVICE_FAILED, - "Failed to initialize IB Device"); - this->private = NULL; - GF_FREE(priv); - return -1; - } - rdma_ctx = this->ctx->ib; - if (!rdma_ctx) - return -1; - - pthread_mutex_lock(&rdma_ctx->lock); - { - if (this->dl_handle && (++(rdma_ctx->dlcount)) == 1) { - iobuf_pool = this->ctx->iobuf_pool; - iobuf_pool->rdma_registration = gf_rdma_register_arena; - iobuf_pool->rdma_deregistration = gf_rdma_deregister_arena; - gf_rdma_register_iobuf_pool_with_device(rdma_ctx->device, - iobuf_pool); - } - } - pthread_mutex_unlock(&rdma_ctx->lock); - - return 0; -} - -int -reconfigure(rpc_transport_t *this, dict_t *options) -{ - gf_rdma_private_t *priv = NULL; - uint32_t backlog = 0; - int ret = -1; - - GF_VALIDATE_OR_GOTO("rdma", this, out); - GF_VALIDATE_OR_GOTO("rdma", this->private, out); - - priv = this->private; - - if (dict_get_uint32(options, "transport.listen-backlog", &backlog) == 0) { - priv->backlog = backlog; - gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.listen-backlog=%d", - priv->backlog); - } - ret = 0; -out: - return ret; -} -void -fini(struct rpc_transport *this) -{ - /* TODO: verify this function does graceful finish */ - gf_rdma_private_t *priv = NULL; - struct iobuf_pool *iobuf_pool = NULL; - gf_rdma_ctx_t *rdma_ctx = NULL; - - priv = this->private; - - this->private = NULL; - - if (priv) { - pthread_mutex_destroy(&priv->recv_mutex); - pthread_mutex_destroy(&priv->write_mutex); - - gf_msg_trace(this->name, 0, "called fini on transport: %p", this); - GF_FREE(priv); - } - - rdma_ctx = this->ctx->ib; - if (!rdma_ctx) - return; - - pthread_mutex_lock(&rdma_ctx->lock); - { - if (this->dl_handle && (--(rdma_ctx->dlcount)) == 0) { - iobuf_pool = this->ctx->iobuf_pool; - gf_rdma_deregister_iobuf_pool(rdma_ctx->device); - iobuf_pool->rdma_registration = NULL; - iobuf_pool->rdma_deregistration = NULL; - } - } - pthread_mutex_unlock(&rdma_ctx->lock); - - return; -} - -/* TODO: expand each option */ -struct volume_options options[] = { - {.key = {"transport.rdma.port", "rdma-port"}, - .type = GF_OPTION_TYPE_INT, - .min = 1, - .max = 4, - .description = "check the option by 'ibv_devinfo'"}, - { - .key = {"transport.rdma.mtu", "rdma-mtu"}, - .type = GF_OPTION_TYPE_INT, - }, - {.key = {"transport.rdma.device-name", "rdma-device-name"}, - .type = GF_OPTION_TYPE_ANY, - .description = "check by 'ibv_devinfo'"}, - { - .key = {"transport.rdma.work-request-send-count", - "rdma-work-request-send-count"}, - .type = GF_OPTION_TYPE_INT, - }, - { - .key = {"transport.rdma.work-request-recv-count", - "rdma-work-request-recv-count"}, - .type = GF_OPTION_TYPE_INT, - }, - {.key = {"remote-port", "transport.remote-port", - "transport.rdma.remote-port"}, - .type = GF_OPTION_TYPE_INT}, - {.key = {"transport.rdma.attr-timeout", "rdma-attr-timeout"}, - .type = GF_OPTION_TYPE_INT}, - {.key = {"transport.rdma.attr-retry-cnt", "rdma-attr-retry-cnt"}, - .type = GF_OPTION_TYPE_INT}, - {.key = {"transport.rdma.attr-rnr-retry", "rdma-attr-rnr-retry"}, - .type = GF_OPTION_TYPE_INT}, - {.key = {"transport.rdma.listen-port", "listen-port"}, - .type = GF_OPTION_TYPE_INT}, - {.key = {"transport.rdma.connect-path", "connect-path"}, - .type = GF_OPTION_TYPE_ANY}, - {.key = {"transport.rdma.bind-path", "bind-path"}, - .type = GF_OPTION_TYPE_ANY}, - {.key = {"transport.rdma.listen-path", "listen-path"}, - .type = GF_OPTION_TYPE_ANY}, - {.key = {"transport.address-family", "address-family"}, - .value = {"inet", "inet6", "inet/inet6", "inet6/inet", "unix", "inet-sdp"}, - .type = GF_OPTION_TYPE_STR}, - {.key = {"transport.socket.lowlat"}, .type = GF_OPTION_TYPE_BOOL}, - {.key = {NULL}}}; diff --git a/rpc/rpc-transport/rdma/src/rdma.h b/rpc/rpc-transport/rdma/src/rdma.h deleted file mode 100644 index 84c1397054b..00000000000 --- a/rpc/rpc-transport/rdma/src/rdma.h +++ /dev/null @@ -1,384 +0,0 @@ -/* - Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> - This file is part of GlusterFS. - - This file is licensed to you under your choice of the GNU Lesser - General Public License, version 3 or any later version (LGPLv3 or - later), or the GNU General Public License, version 2 (GPLv2), in all - cases as published by the Free Software Foundation. -*/ - -#ifndef _XPORT_RDMA_H -#define _XPORT_RDMA_H - -#ifndef MAX_IOVEC -#define MAX_IOVEC 16 -#endif /* MAX_IOVEC */ - -#include "rpc-clnt.h" -#include "rpc-transport.h" -#include "xlator.h" -#include "gf-event.h" -#include <stdio.h> -#include <list.h> -#include <arpa/inet.h> -#include <infiniband/verbs.h> -#include <rdma/rdma_cma.h> - -/* FIXME: give appropriate values to these macros */ -#define GF_DEFAULT_RDMA_LISTEN_PORT (GF_DEFAULT_BASE_PORT + 1) - -/* If you are changing GF_RDMA_MAX_SEGMENTS, please make sure to update - * GLUSTERFS_GF_RDMA_MAX_HEADER_SIZE defined in glusterfs.h . - */ -#define GF_RDMA_MAX_SEGMENTS 8 - -#define GF_RDMA_VERSION 1 -#define GF_RDMA_POOL_SIZE 512 - -/* Additional attributes */ -#define GF_RDMA_TIMEOUT 14 -#define GF_RDMA_RETRY_CNT 7 -#define GF_RDMA_RNR_RETRY 7 - -typedef enum gf_rdma_errcode { ERR_VERS = 1, ERR_CHUNK = 2 } gf_rdma_errcode_t; - -struct gf_rdma_err_vers { - uint32_t gf_rdma_vers_low; /* Version range supported by peer */ - uint32_t gf_rdma_vers_high; -} __attribute__((packed)); -typedef struct gf_rdma_err_vers gf_rdma_err_vers_t; - -typedef enum gf_rdma_proc { - GF_RDMA_MSG = 0, /* An RPC call or reply msg */ - GF_RDMA_NOMSG = 1, /* An RPC call or reply msg - separate body */ - GF_RDMA_MSGP = 2, /* An RPC call or reply msg with padding */ - GF_RDMA_DONE = 3, /* Client signals reply completion */ - GF_RDMA_ERROR = 4 /* An RPC RDMA encoding error */ -} gf_rdma_proc_t; - -typedef enum gf_rdma_chunktype { - gf_rdma_noch = 0, /* no chunk */ - gf_rdma_readch, /* some argument through rdma read */ - gf_rdma_areadch, /* entire request through rdma read */ - gf_rdma_writech, /* some result through rdma write */ - gf_rdma_replych /* entire reply through rdma write */ -} gf_rdma_chunktype_t; - -/* If you are modifying __gf_rdma_header, please make sure to change - * GLUSTERFS_GF_RDMA_MAX_HEADER_SIZE defined in glusterfs.h to reflect your - * changes - */ -struct __gf_rdma_header { - uint32_t rm_xid; /* Mirrors the RPC header xid */ - uint32_t rm_vers; /* Version of this protocol */ - uint32_t rm_credit; /* Buffers requested/granted */ - uint32_t rm_type; /* Type of message (enum gf_rdma_proc) */ - union { - struct { /* no chunks */ - uint32_t rm_empty[3]; /* 3 empty chunk lists */ - } __attribute__((packed)) rm_nochunks; - - struct { /* no chunks and padded */ - uint32_t rm_align; /* Padding alignment */ - uint32_t rm_thresh; /* Padding threshold */ - uint32_t rm_pempty[3]; /* 3 empty chunk lists */ - } __attribute__((packed)) rm_padded; - - struct { - uint32_t rm_type; - gf_rdma_err_vers_t rm_version; - } __attribute__((packed)) rm_error; - - uint32_t rm_chunks[0]; /* read, write and reply chunks */ - } __attribute__((packed)) rm_body; -} __attribute__((packed)); -typedef struct __gf_rdma_header gf_rdma_header_t; - -/* If you are modifying __gf_rdma_segment or __gf_rdma_read_chunk, please make - * sure to change GLUSTERFS_GF_RDMA_MAX_HEADER_SIZE defined in glusterfs.h to - * reflect your changes. - */ -struct __gf_rdma_segment { - uint32_t rs_handle; /* Registered memory handle */ - uint32_t rs_length; /* Length of the chunk in bytes */ - uint64_t rs_offset; /* Chunk virtual address or offset */ -} __attribute__((packed)); -typedef struct __gf_rdma_segment gf_rdma_segment_t; - -/* read chunk(s), encoded as a linked list. */ -struct __gf_rdma_read_chunk { - uint32_t rc_discrim; /* 1 indicates presence */ - uint32_t rc_position; /* Position in XDR stream */ - gf_rdma_segment_t rc_target; -} __attribute__((packed)); -typedef struct __gf_rdma_read_chunk gf_rdma_read_chunk_t; - -/* write chunk, and reply chunk. */ -struct __gf_rdma_write_chunk { - gf_rdma_segment_t wc_target; -} __attribute__((packed)); -typedef struct __gf_rdma_write_chunk gf_rdma_write_chunk_t; - -/* write chunk(s), encoded as a counted array. */ -struct __gf_rdma_write_array { - uint32_t wc_discrim; /* 1 indicates presence */ - uint32_t wc_nchunks; /* Array count */ - struct __gf_rdma_write_chunk wc_array[0]; -} __attribute__((packed)); -typedef struct __gf_rdma_write_array gf_rdma_write_array_t; - -/* options per transport end point */ -struct __gf_rdma_options { - int32_t port; - char *device_name; - enum ibv_mtu mtu; - int32_t send_count; - int32_t recv_count; - uint64_t recv_size; - uint64_t send_size; - uint8_t attr_timeout; - uint8_t attr_retry_cnt; - uint8_t attr_rnr_retry; - uint32_t backlog; -}; -typedef struct __gf_rdma_options gf_rdma_options_t; - -struct __gf_rdma_reply_info { - uint32_t rm_xid; /* xid in network endian */ - gf_rdma_chunktype_t type; /* - * can be either gf_rdma_replych - * or gf_rdma_writech. - */ - gf_rdma_write_array_t *wc_array; - struct mem_pool *pool; -}; -typedef struct __gf_rdma_reply_info gf_rdma_reply_info_t; - -struct __gf_rdma_ioq { - union { - struct list_head list; - struct { - struct __gf_rdma_ioq *next; - struct __gf_rdma_ioq *prev; - }; - }; - - char is_request; - struct iovec rpchdr[MAX_IOVEC]; - int rpchdr_count; - struct iovec proghdr[MAX_IOVEC]; - int proghdr_count; - struct iovec prog_payload[MAX_IOVEC]; - int prog_payload_count; - - struct iobref *iobref; - - union { - struct __gf_rdma_ioq_request { - /* used to build reply_chunk for GF_RDMA_NOMSG type msgs */ - struct iovec rsphdr_vec[MAX_IOVEC]; - int rsphdr_count; - - /* - * used to build write_array during operations like - * read. - */ - struct iovec rsp_payload[MAX_IOVEC]; - int rsp_payload_count; - - struct rpc_req *rpc_req; /* FIXME: hack! hack! should be - * cleaned up later - */ - struct iobref *rsp_iobref; - } request; - - gf_rdma_reply_info_t *reply_info; - } msg; - - struct mem_pool *pool; -}; -typedef struct __gf_rdma_ioq gf_rdma_ioq_t; - -typedef enum __gf_rdma_send_post_type { - GF_RDMA_SEND_POST_NO_CHUNKLIST, /* post which is sent using rdma-send - * and the msg carries no - * chunklists. - */ - GF_RDMA_SEND_POST_READ_CHUNKLIST, /* post which is sent using rdma-send - * and the msg carries only read - * chunklist. - */ - GF_RDMA_SEND_POST_WRITE_CHUNKLIST, /* post which is sent using - * rdma-send and the msg carries - * only write chunklist. - */ - GF_RDMA_SEND_POST_READ_WRITE_CHUNKLIST, /* post which is sent using - * rdma-send and the msg - * carries both read and - * write chunklists. - */ - GF_RDMA_SEND_POST_GF_RDMA_READ, /* RDMA read */ - GF_RDMA_SEND_POST_GF_RDMA_WRITE, /* RDMA write */ -} gf_rdma_send_post_type_t; - -/* represents one communication peer, two per transport_t */ -struct __gf_rdma_peer { - rpc_transport_t *trans; - struct rdma_cm_id *cm_id; - struct ibv_qp *qp; - pthread_t rdma_event_thread; - char quota_set; - - int32_t recv_count; - int32_t send_count; - int32_t recv_size; - int32_t send_size; - - int32_t quota; - union { - struct list_head ioq; - struct { - gf_rdma_ioq_t *ioq_next; - gf_rdma_ioq_t *ioq_prev; - }; - }; - - /* QP attributes, needed to connect with remote QP */ - int32_t local_lid; - int32_t local_psn; - int32_t local_qpn; - int32_t remote_lid; - int32_t remote_psn; - int32_t remote_qpn; -}; -typedef struct __gf_rdma_peer gf_rdma_peer_t; - -struct __gf_rdma_post_context { - struct ibv_mr *mr[GF_RDMA_MAX_SEGMENTS]; - int mr_count; - struct iovec vector[MAX_IOVEC]; - int count; - struct iobref *iobref; - struct iobuf *hdr_iobuf; - char is_request; - int gf_rdma_reads; - gf_rdma_reply_info_t *reply_info; -}; -typedef struct __gf_rdma_post_context gf_rdma_post_context_t; - -typedef enum { GF_RDMA_SEND_POST, GF_RDMA_RECV_POST } gf_rdma_post_type_t; - -struct __gf_rdma_post { - struct __gf_rdma_post *next, *prev; - struct ibv_mr *mr; - char *buf; - int32_t buf_size; - char aux; - int32_t reused; - struct __gf_rdma_device *device; - gf_rdma_post_type_t type; - gf_rdma_post_context_t ctx; - int refcount; - pthread_mutex_t lock; -}; -typedef struct __gf_rdma_post gf_rdma_post_t; - -struct __gf_rdma_queue { - gf_rdma_post_t active_posts, passive_posts; - int32_t active_count, passive_count; - pthread_mutex_t lock; -}; -typedef struct __gf_rdma_queue gf_rdma_queue_t; - -struct __gf_rdma_qpreg { - pthread_mutex_t lock; - int32_t count; - struct _qpent { - struct _qpent *next, *prev; - int32_t qp_num; - gf_rdma_peer_t *peer; - } ents[42]; -}; -typedef struct __gf_rdma_qpreg gf_rdma_qpreg_t; - -/* context per device, stored in global glusterfs_ctx_t->ib */ -struct __gf_rdma_device { - struct __gf_rdma_device *next; - const char *device_name; - struct ibv_context *context; - int32_t port; - struct ibv_pd *pd; - struct ibv_srq *srq; - gf_rdma_qpreg_t qpreg; - struct ibv_comp_channel *send_chan, *recv_chan; - struct ibv_cq *send_cq, *recv_cq; - gf_rdma_queue_t sendq, recvq; - pthread_t send_thread, recv_thread, async_event_thread; - struct mem_pool *request_ctx_pool; - struct mem_pool *ioq_pool; - struct mem_pool *reply_info_pool; - struct list_head all_mr; - pthread_mutex_t all_mr_lock; -}; -typedef struct __gf_rdma_device gf_rdma_device_t; - -struct __gf_rdma_arena_mr { - struct list_head list; - struct iobuf_arena *iobuf_arena; - struct ibv_mr *mr; -}; - -typedef struct __gf_rdma_arena_mr gf_rdma_arena_mr; -struct __gf_rdma_ctx { - gf_rdma_device_t *device; - struct rdma_event_channel *rdma_cm_event_channel; - pthread_t rdma_cm_thread; - pthread_mutex_t lock; - int32_t dlcount; -}; -typedef struct __gf_rdma_ctx gf_rdma_ctx_t; - -struct __gf_rdma_request_context { - struct ibv_mr *mr[GF_RDMA_MAX_SEGMENTS]; - int mr_count; - struct mem_pool *pool; - gf_rdma_peer_t *peer; - struct iobref *iobref; - struct iobref *rsp_iobref; -}; -typedef struct __gf_rdma_request_context gf_rdma_request_context_t; - -typedef enum { - GF_RDMA_SERVER_LISTENER, - GF_RDMA_SERVER, - GF_RDMA_CLIENT, -} gf_rdma_transport_entity_t; - -struct __gf_rdma_private { - int32_t idx; - unsigned char connected; - in_addr_t addr; - unsigned short port; - - /* IB Verbs Driver specific variables, pointers */ - gf_rdma_peer_t peer; - struct __gf_rdma_device *device; - gf_rdma_options_t options; - - /* Used by trans->op->receive */ - char *data_ptr; - int32_t data_offset; - int32_t data_len; - - /* Mutex */ - pthread_mutex_t write_mutex; - rpc_transport_t *listener; - pthread_mutex_t recv_mutex; - pthread_cond_t recv_cond; - gf_rdma_transport_entity_t entity; - uint32_t backlog; -}; -typedef struct __gf_rdma_private gf_rdma_private_t; - -#endif /* _XPORT_GF_RDMA_H */ diff --git a/rpc/rpc-transport/rdma/src/rpc-trans-rdma-messages.h b/rpc/rpc-transport/rdma/src/rpc-trans-rdma-messages.h deleted file mode 100644 index 6e9ea9a2925..00000000000 --- a/rpc/rpc-transport/rdma/src/rpc-trans-rdma-messages.h +++ /dev/null @@ -1,66 +0,0 @@ -/* - Copyright (c) 2015 Red Hat, Inc. <http://www.redhat.com> - This file is part of GlusterFS. - - This file is licensed to you under your choice of the GNU Lesser - General Public License, version 3 or any later version (LGPLv3 or - later), or the GNU General Public License, version 2 (GPLv2), in all - cases as published by the Free Software Foundation. -*/ - -#ifndef _RPC_TRANS_RDMA_MESSAGES_H_ -#define _RPC_TRANS_RDMA_MESSAGES_H_ - -#include "glfs-message-id.h" - -/* To add new message IDs, append new identifiers at the end of the list. - * - * Never remove a message ID. If it's not used anymore, you can rename it or - * leave it as it is, but not delete it. This is to prevent reutilization of - * IDs by other messages. - * - * The component name must match one of the entries defined in - * glfs-message-id.h. - */ - -GLFS_MSGID( - RPC_TRANS_RDMA, RDMA_MSG_PORT_BIND_FAILED, RDMA_MSG_POST_CREATION_FAILED, - RDMA_MSG_DEREGISTER_ARENA_FAILED, RDMA_MSG_MR_ALOC_FAILED, - RDMA_MSG_PREREG_BUFFER_FAILED, RDMA_MSG_CQ_CREATION_FAILED, - RDMA_MSG_REQ_NOTIFY_CQ_REVQ_FAILED, RDMA_MSG_QUERY_DEVICE_FAILED, - RDMA_MSG_REQ_NOTIFY_CQ_SENDQ_FAILED, RDMA_MSG_SEND_COMP_CHAN_FAILED, - RDMA_MSG_RECV_COMP_CHAN_FAILED, RDMA_MSG_ALOC_PROT_DOM_FAILED, - RDMA_MSG_CRE_SRQ_FAILED, RDMA_MSG_ALOC_POST_FAILED, - RDMA_MSG_SEND_COMP_THREAD_FAILED, RDMA_MSG_RECV_COMP_THREAD_FAILED, - RDMA_MSG_ASYNC_EVENT_THEAD_FAILED, RDMA_MSG_GET_DEVICE_NAME_FAILED, - RDMA_MSG_GET_IB_DEVICE_FAILED, RDMA_MSG_CREAT_INC_TRANS_FAILED, - RDMA_MSG_CREAT_QP_FAILED, RDMA_MSG_ACCEPT_FAILED, RDMA_MSG_CONNECT_FAILED, - RDMA_MSG_ROUTE_RESOLVE_FAILED, RDMA_MSG_GET_DEVICE_FAILED, - RDMA_MSG_PEER_DISCONNECTED, RDMA_MSG_ENCODE_ERROR, - RDMA_MSG_POST_SEND_FAILED, RDMA_MSG_READ_CHUNK_VECTOR_FAILED, - RDMA_MSG_WRITE_CHUNK_VECTOR_FAILED, RDMA_MSG_WRITE_REPLY_CHUNCK_CONFLICT, - RDMA_MSG_CHUNK_COUNT_GREAT_MAX_SEGMENTS, RDMA_MSG_CREATE_READ_CHUNK_FAILED, - RDMA_MSG_CREATE_WRITE_REPLAY_FAILED, - RDMA_MSG_SEND_SIZE_GREAT_INLINE_THRESHOLD, - RDMA_MSG_REG_ACCESS_LOCAL_WRITE_FAILED, RDMA_MSG_WRITE_PEER_FAILED, - RDMA_MSG_SEND_REPLY_FAILED, RDMA_MSG_INVALID_CHUNK_TYPE, - RDMA_MSG_PROC_IOQ_ENTRY_FAILED, RDMA_MSG_NEW_IOQ_ENTRY_FAILED, - RDMA_MSG_RPC_REPLY_CREATE_FAILED, RDMA_MSG_GET_READ_CHUNK_FAILED, - RDMA_MSG_GET_WRITE_CHUNK_FAILED, RDMA_MSG_REPLY_INFO_ALLOC_FAILED, - RDMA_MSG_RDMA_ERROR_RECEIVED, RDMA_MSG_GET_REQ_INFO_RPC_FAILED, - RDMA_MSG_POLL_IN_NOTIFY_FAILED, RDMA_MSG_HEADER_DECODE_FAILED, - RDMA_MSG_EVENT_SRQ_LIMIT_REACHED, RDMA_MSG_UNRECG_MQ_VALUE, - RDMA_MSG_BUFFER_ERROR, RDMA_MSG_OPTION_SET_FAILED, RDMA_MSG_LISTEN_FAILED, - RDMA_MSG_INIT_IB_DEVICE_FAILED, RDMA_MSG_WRITE_CLIENT_ERROR, - RDMA_MSG_CHUNKLIST_ERROR, RDMA_MSG_INVALID_ENTRY, - RDMA_MSG_READ_CLIENT_ERROR, RDMA_MSG_RPC_ST_ERROR, - RDMA_MSG_PEER_READ_FAILED, RDMA_MSG_POST_MISSING, RDMA_MSG_PEER_REQ_FAILED, - RDMA_MSG_PEER_REP_FAILED, RDMA_MSG_EVENT_ERROR, RDMA_MSG_IBV_GET_CQ_FAILED, - RDMA_MSG_IBV_REQ_NOTIFY_CQ_FAILED, RDMA_MSG_RECV_ERROR, - RDMA_MSG_IBV_POLL_CQ_ERROR, RDMA_MSG_RDMA_HANDLE_FAILED, - RDMA_MSG_CM_EVENT_FAILED, RDMA_MSG_CLIENT_BIND_FAILED, - RDMA_MSG_RDMA_RESOLVE_ADDR_FAILED, RDMA_MSG_NW_ADDR_UNKNOWN, - RDMA_MSG_RDMA_BIND_ADDR_FAILED, RDMA_MSG_SEND_CLIENT_ERROR, - RDMA_MSG_UNRECG_MTU_VALUE); - -#endif /* !_RPC_TRANS_RDMA_MESSAGES_H_ */ diff --git a/rpc/rpc-transport/socket/src/name.c b/rpc/rpc-transport/socket/src/name.c index 3194a7cf369..9286bbb236d 100644 --- a/rpc/rpc-transport/socket/src/name.c +++ b/rpc/rpc-transport/socket/src/name.c @@ -21,7 +21,7 @@ #include "rpc-transport.h" #include "socket.h" -#include "common-utils.h" +#include <glusterfs/common-utils.h> static void _assign_port(struct sockaddr *sockaddr, uint16_t port) @@ -105,10 +105,10 @@ af_unix_client_bind(rpc_transport_t *this, struct sockaddr *sockaddr, struct sockaddr_un *addr = NULL; int32_t ret = 0; - path_data = dict_get(this->options, "transport.socket.bind-path"); + path_data = dict_get_sizen(this->options, "transport.socket.bind-path"); if (path_data) { char *path = data_to_str(path_data); - if (!path || strlen(path) > UNIX_PATH_MAX) { + if (!path || path_data->len > 108) { /* 108 = addr->sun_path length */ gf_log(this->name, GF_LOG_TRACE, "bind-path not specified for unix socket, " "letting connect to assign default value"); @@ -134,7 +134,7 @@ err: return ret; } -int32_t +static int32_t client_fill_address_family(rpc_transport_t *this, sa_family_t *sa_family) { data_t *address_family_data = NULL; @@ -145,12 +145,13 @@ client_fill_address_family(rpc_transport_t *this, sa_family_t *sa_family) goto out; } - address_family_data = dict_get(this->options, "transport.address-family"); + address_family_data = dict_get_sizen(this->options, + "transport.address-family"); if (!address_family_data) { data_t *remote_host_data = NULL, *connect_path_data = NULL; - remote_host_data = dict_get(this->options, "remote-host"); - connect_path_data = dict_get(this->options, - "transport.socket.connect-path"); + remote_host_data = dict_get_sizen(this->options, "remote-host"); + connect_path_data = dict_get_sizen(this->options, + "transport.socket.connect-path"); if (!(remote_host_data || connect_path_data) || (remote_host_data && connect_path_data)) { @@ -179,7 +180,7 @@ client_fill_address_family(rpc_transport_t *this, sa_family_t *sa_family) } } else { - char *address_family = data_to_str(address_family_data); + const char *address_family = data_to_str(address_family_data); if (!strcasecmp(address_family, "unix")) { *sa_family = AF_UNIX; } else if (!strcasecmp(address_family, "inet")) { @@ -211,11 +212,12 @@ af_inet_client_get_remote_sockaddr(rpc_transport_t *this, data_t *remote_host_data = NULL; data_t *remote_port_data = NULL; char *remote_host = NULL; - uint16_t remote_port = 0; + uint16_t remote_port = GF_DEFAULT_SOCKET_LISTEN_PORT; struct addrinfo *addr_info = NULL; int32_t ret = 0; + struct in6_addr serveraddr; - remote_host_data = dict_get(options, "remote-host"); + remote_host_data = dict_get_sizen(options, "remote-host"); if (remote_host_data == NULL) { gf_log(this->name, GF_LOG_ERROR, "option remote-host missing in volume %s", this->name); @@ -231,22 +233,27 @@ af_inet_client_get_remote_sockaddr(rpc_transport_t *this, goto err; } - remote_port_data = dict_get(options, "remote-port"); + remote_port_data = dict_get_sizen(options, "remote-port"); if (remote_port_data == NULL) { gf_log(this->name, GF_LOG_TRACE, "option remote-port missing in volume %s. Defaulting to %d", this->name, GF_DEFAULT_SOCKET_LISTEN_PORT); - - remote_port = GF_DEFAULT_SOCKET_LISTEN_PORT; } else { remote_port = data_to_uint16(remote_port_data); + if (remote_port == (uint16_t)-1) { + gf_log(this->name, GF_LOG_ERROR, + "option remote-port has invalid port in volume %s", + this->name); + ret = -1; + goto err; + } } - if (remote_port == (uint16_t)-1) { - gf_log(this->name, GF_LOG_ERROR, - "option remote-port has invalid port in volume %s", this->name); - ret = -1; - goto err; + /* Need to update transport-address family if address-family is not provided + to command-line arguments + */ + if (inet_pton(AF_INET6, remote_host, &serveraddr)) { + sockaddr->sa_family = AF_INET6; } /* TODO: gf_resolve is a blocking call. kick in some @@ -274,31 +281,29 @@ af_unix_client_get_remote_sockaddr(rpc_transport_t *this, struct sockaddr_un *sockaddr_un = NULL; char *connect_path = NULL; data_t *connect_path_data = NULL; - int32_t ret = 0; + int32_t ret = -1; - connect_path_data = dict_get(this->options, - "transport.socket.connect-path"); + connect_path_data = dict_get_sizen(this->options, + "transport.socket.connect-path"); if (!connect_path_data) { gf_log(this->name, GF_LOG_ERROR, "option transport.unix.connect-path not specified for " "address-family unix"); - ret = -1; goto err; } - connect_path = data_to_str(connect_path_data); - if (!connect_path) { + /* 108 = sockaddr_un->sun_path length */ + if ((connect_path_data->len + 1) > 108) { gf_log(this->name, GF_LOG_ERROR, - "transport.unix.connect-path is null-string"); - ret = -1; + "connect-path value length %d > %d octets", + connect_path_data->len + 1, UNIX_PATH_MAX); goto err; } - if ((strlen(connect_path) + 1) > UNIX_PATH_MAX) { + connect_path = data_to_str(connect_path_data); + if (!connect_path) { gf_log(this->name, GF_LOG_ERROR, - "connect-path value length %" GF_PRI_SIZET " > %d octets", - strlen(connect_path), UNIX_PATH_MAX); - ret = -1; + "transport.unix.connect-path is null-string"); goto err; } @@ -307,6 +312,7 @@ af_unix_client_get_remote_sockaddr(rpc_transport_t *this, strcpy(sockaddr_un->sun_path, connect_path); *sockaddr_len = sizeof(struct sockaddr_un); + ret = 0; err: return ret; } @@ -320,7 +326,8 @@ af_unix_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, int32_t ret = 0; struct sockaddr_un *sunaddr = (struct sockaddr_un *)addr; - listen_path_data = dict_get(this->options, "transport.socket.listen-path"); + listen_path_data = dict_get_sizen(this->options, + "transport.socket.listen-path"); if (!listen_path_data) { gf_log(this->name, GF_LOG_ERROR, "missing option transport.socket.listen-path"); @@ -334,7 +341,7 @@ af_unix_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, #define UNIX_PATH_MAX 108 #endif - if ((strlen(listen_path) + 1) > UNIX_PATH_MAX) { + if ((listen_path_data->len + 1) > UNIX_PATH_MAX) { gf_log(this->name, GF_LOG_ERROR, "option transport.unix.listen-path has value length " "%" GF_PRI_SIZET " > %d", @@ -357,23 +364,24 @@ af_inet_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, { struct addrinfo hints, *res = 0, *rp = NULL; data_t *listen_port_data = NULL, *listen_host_data = NULL; - uint16_t listen_port = -1; + uint16_t listen_port = 0; char service[NI_MAXSERV], *listen_host = NULL; dict_t *options = NULL; int32_t ret = 0; - options = this->options; + /* initializes addr_len */ + *addr_len = 0; - listen_port_data = dict_get(options, "transport.socket.listen-port"); - listen_host_data = dict_get(options, "transport.socket.bind-address"); + options = this->options; + listen_port_data = dict_get_sizen(options, "transport.socket.listen-port"); if (listen_port_data) { listen_port = data_to_uint16(listen_port_data); - } - - if (listen_port == (uint16_t)-1) + } else { listen_port = GF_DEFAULT_SOCKET_LISTEN_PORT; + } + listen_host_data = dict_get_sizen(options, "transport.socket.bind-address"); if (listen_host_data) { listen_host = data_to_str(listen_host_data); } else { @@ -418,8 +426,12 @@ af_inet_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, } if (!(*addr_len)) { - memcpy(addr, res->ai_addr, res->ai_addrlen); - *addr_len = res->ai_addrlen; + if (res && res->ai_addr) { + memcpy(addr, res->ai_addr, res->ai_addrlen); + *addr_len = res->ai_addrlen; + } else { + ret = -1; + } } freeaddrinfo(res); @@ -522,7 +534,10 @@ socket_client_get_remote_sockaddr(rpc_transport_t *this, ret = -1; } - if (*sa_family == AF_UNSPEC) { + /* Address-family is updated based on remote_host in + af_inet_client_get_remote_sockaddr + */ + if (*sa_family != sockaddr->sa_family) { *sa_family = sockaddr->sa_family; } @@ -530,23 +545,24 @@ err: return ret; } -int32_t +static int32_t server_fill_address_family(rpc_transport_t *this, sa_family_t *sa_family) { data_t *address_family_data = NULL; int32_t ret = -1; #ifdef IPV6_DEFAULT - char *addr_family = "inet6"; + const char *addr_family = "inet6"; sa_family_t default_family = AF_INET6; #else - char *addr_family = "inet"; + const char *addr_family = "inet"; sa_family_t default_family = AF_INET; #endif GF_VALIDATE_OR_GOTO("socket", sa_family, out); - address_family_data = dict_get(this->options, "transport.address-family"); + address_family_data = dict_get_sizen(this->options, + "transport.address-family"); if (address_family_data) { char *address_family = NULL; address_family = data_to_str(address_family_data); @@ -598,7 +614,7 @@ socket_server_get_local_sockaddr(rpc_transport_t *this, struct sockaddr *addr, switch (addr->sa_family) { case AF_INET_SDP: addr->sa_family = AF_INET; - /* Fall through */ + /* Fall through */ case AF_INET: case AF_INET6: case AF_UNSPEC: @@ -618,7 +634,7 @@ err: return ret; } -int32_t +static int32_t fill_inet6_inet_identifiers(rpc_transport_t *this, struct sockaddr_storage *addr, int32_t addr_len, char *identifier) diff --git a/rpc/rpc-transport/socket/src/name.h b/rpc/rpc-transport/socket/src/name.h index 85f8bb7040b..080c7588f5a 100644 --- a/rpc/rpc-transport/socket/src/name.h +++ b/rpc/rpc-transport/socket/src/name.h @@ -11,7 +11,7 @@ #ifndef _SOCKET_NAME_H #define _SOCKET_NAME_H -#include "compat.h" +#include <glusterfs/compat.h> int32_t client_bind(rpc_transport_t *this, struct sockaddr *sockaddr, diff --git a/rpc/rpc-transport/socket/src/socket-mem-types.h b/rpc/rpc-transport/socket/src/socket-mem-types.h index 81635b72320..241ce67f670 100644 --- a/rpc/rpc-transport/socket/src/socket-mem-types.h +++ b/rpc/rpc-transport/socket/src/socket-mem-types.h @@ -11,12 +11,11 @@ #ifndef __SOCKET_MEM_TYPES_H__ #define __SOCKET_MEM_TYPES_H__ -#include "mem-types.h" +#include <glusterfs/mem-types.h> typedef enum gf_sock_mem_types_ { gf_sock_connect_error_state_t = gf_common_mt_end + 1, gf_sock_mt_lock_array, - gf_sock_mt_tid_wrap, gf_sock_mt_end } gf_sock_mem_types_t; diff --git a/rpc/rpc-transport/socket/src/socket.c b/rpc/rpc-transport/socket/src/socket.c index b3f8b7fc580..ed8b473be23 100644 --- a/rpc/rpc-transport/socket/src/socket.c +++ b/rpc/rpc-transport/socket/src/socket.c @@ -10,22 +10,16 @@ #include "socket.h" #include "name.h" -#include "dict.h" -#include "rpc-transport.h" -#include "logging.h" -#include "xlator.h" -#include "syscall.h" -#include "byte-order.h" -#include "common-utils.h" -#include "compat-errno.h" +#include <glusterfs/dict.h> +#include <glusterfs/syscall.h> +#include <glusterfs/byte-order.h> +#include <glusterfs/compat-errno.h> #include "socket-mem-types.h" -#include "timer.h" /* ugly #includes below */ #include "protocol-common.h" #include "glusterfs3-xdr.h" #include "glusterfs4-xdr.h" -#include "xdr-nfs3.h" #include "rpcsvc.h" /* for TCP_USER_TIMEOUT */ @@ -35,7 +29,6 @@ #include <netinet/tcp.h> #endif -#include <fcntl.h> #include <errno.h> #include <rpc/xdr.h> #include <sys/ioctl.h> @@ -139,7 +132,7 @@ ssl_setup_connection_params(rpc_transport_t *this); \ gf_log(this->name, GF_LOG_TRACE, \ "partial read on non-blocking socket"); \ - \ + ret = 0; \ break; \ } \ } @@ -173,7 +166,7 @@ ssl_setup_connection_params(rpc_transport_t *this); \ ret = __socket_readv(this, in->pending_vector, 1, &in->pending_vector, \ &in->pending_count, &bytes_read); \ - if (ret == -1) \ + if (ret < 0) \ break; \ __socket_proto_update_priv_after_read(priv, ret, bytes_read); \ } @@ -198,7 +191,7 @@ socket_dump_info(struct sockaddr *sa, int is_server, int is_ssl, int sock, 0, }; char *addr = NULL; - char *peer_type = NULL; + const char *peer_type = NULL; int af = sa->sa_family; int so_error = -1; socklen_t slen = sizeof(so_error); @@ -223,7 +216,7 @@ socket_dump_info(struct sockaddr *sa, int is_server, int is_ssl, int sock, else peer_type = "client"; - getsockopt(sock, SOL_SOCKET, SO_ERROR, &so_error, &slen); + (void)getsockopt(sock, SOL_SOCKET, SO_ERROR, &so_error, &slen); gf_log(log_domain, GF_LOG_TRACE, "$$$ %s: %s (af:%d,sock:%d) %s %s (errno:%d:%s)", peer_type, @@ -253,7 +246,6 @@ ssl_do(rpc_transport_t *this, void *buf, size_t len, SSL_trinary_func *func) int r = (-1); socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO(this->name, this->private, out); priv = this->private; if (buf) { @@ -269,11 +261,9 @@ ssl_do(rpc_transport_t *this, void *buf, size_t len, SSL_trinary_func *func) } r = func(priv->ssl_ssl, buf, len); } else { - /* - * We actually need these functions to get to - * priv->connected == 1. - */ - r = ((SSL_unary_func *)func)(priv->ssl_ssl); + /* This should be treated as error */ + gf_log(this->name, GF_LOG_ERROR, "buffer is empty %s", __func__); + goto out; } switch (SSL_get_error(priv->ssl_ssl, r)) { case SSL_ERROR_NONE: @@ -308,14 +298,69 @@ out: #define ssl_write_one(t, b, l) \ ssl_do((t), (b), (l), (SSL_trinary_func *)SSL_write) -int -ssl_setup_connection_prefix(rpc_transport_t *this) +/* set crl verify flags only for server */ +/* see man X509_VERIFY_PARAM_SET_FLAGS(3) + * X509_V_FLAG_CRL_CHECK enables CRL checking for the certificate chain + * leaf certificate. An error occurs if a suitable CRL cannot be found. + * Since we're never going to revoke a gluster node cert, we better disable + * CRL check for server certs to avoid getting error and failed connection + * attempts. + */ +static void +ssl_clear_crl_verify_flags(SSL_CTX *ssl_ctx) +{ +#ifdef X509_V_FLAG_CRL_CHECK_ALL +#ifdef HAVE_SSL_CTX_GET0_PARAM + X509_VERIFY_PARAM *vpm; + + vpm = SSL_CTX_get0_param(ssl_ctx); + if (vpm) { + X509_VERIFY_PARAM_clear_flags( + vpm, (X509_V_FLAG_CRL_CHECK | X509_V_FLAG_CRL_CHECK_ALL)); + } +#else + /* CRL verify flag need not be cleared for rhel6 kind of clients */ +#endif +#else + gf_log(this->name, GF_LOG_ERROR, "OpenSSL version does not support CRL"); +#endif + return; +} + +/* set crl verify flags only for server */ +static void +ssl_set_crl_verify_flags(SSL_CTX *ssl_ctx) +{ +#ifdef X509_V_FLAG_CRL_CHECK_ALL +#ifdef HAVE_SSL_CTX_GET0_PARAM + X509_VERIFY_PARAM *vpm; + + vpm = SSL_CTX_get0_param(ssl_ctx); + if (vpm) { + unsigned long flags; + + flags = X509_VERIFY_PARAM_get_flags(vpm); + flags |= (X509_V_FLAG_CRL_CHECK | X509_V_FLAG_CRL_CHECK_ALL); + X509_VERIFY_PARAM_set_flags(vpm, flags); + } +#else + X509_STORE *x509store; + + x509store = SSL_CTX_get_cert_store(ssl_ctx); + X509_STORE_set_flags(x509store, + X509_V_FLAG_CRL_CHECK | X509_V_FLAG_CRL_CHECK_ALL); +#endif +#else + gf_log(this->name, GF_LOG_ERROR, "OpenSSL version does not support CRL"); +#endif +} + +static int +ssl_setup_connection_prefix(rpc_transport_t *this, gf_boolean_t server) { int ret = -1; socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO(this->name, this->private, done); - priv = this->private; if (ssl_setup_connection_params(this) < 0) { @@ -332,6 +377,9 @@ ssl_setup_connection_prefix(rpc_transport_t *this) priv->ssl_accepted = _gf_false; priv->ssl_context_created = _gf_false; + if (!server && priv->crl_path) + ssl_clear_crl_verify_flags(priv->ssl_ctx); + priv->ssl_ssl = SSL_new(priv->ssl_ctx); if (!priv->ssl_ssl) { gf_log(this->name, GF_LOG_ERROR, "SSL_new failed"); @@ -364,7 +412,6 @@ ssl_setup_connection_postfix(rpc_transport_t *this) char peer_CN[256] = ""; socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO(this->name, this->private, done); priv = this->private; /* Make sure _SSL verification_ succeeded, yielding an identity. */ @@ -386,6 +433,7 @@ ssl_setup_connection_postfix(rpc_transport_t *this) gf_log(this->name, GF_LOG_DEBUG, "SSL verification succeeded (client: %s) (server: %s)", this->peerinfo.identifier, this->myinfo.identifier); + X509_free(peer); return gf_strdup(peer_CN); /* Error paths. */ @@ -397,11 +445,10 @@ ssl_error: SSL_free(priv->ssl_ssl); priv->ssl_ssl = NULL; -done: return NULL; } -int +static int ssl_complete_connection(rpc_transport_t *this) { int ret = -1; /* 1 : implies go back to epoll_wait() @@ -640,7 +687,6 @@ __socket_rwv(rpc_transport_t *this, struct iovec *vector, int count, int opcount = 0; int moved = 0; - GF_VALIDATE_OR_GOTO("socket", this, out); GF_VALIDATE_OR_GOTO("socket", this->private, out); priv = this->private; @@ -679,24 +725,26 @@ __socket_rwv(rpc_transport_t *this, struct iovec *vector, int count, ret = sys_writev(sock, opvector, IOV_MIN(opcount)); } - if (ret == 0 || (ret == -1 && errno == EAGAIN)) { + if ((ret == 0) || ((ret < 0) && (errno == EAGAIN))) { /* done for now */ break; - } - this->total_bytes_write += ret; + } else if (ret > 0) + this->total_bytes_write += ret; } else { ret = __socket_cached_read(this, opvector, opcount); if (ret == 0) { - gf_log(this->name, GF_LOG_DEBUG, "EOF on socket (errno:%d:%s)", - errno, strerror(errno)); + gf_log(this->name, GF_LOG_DEBUG, + "EOF on socket %d (errno:%d:%s); returning ENODATA", + sock, errno, strerror(errno)); + errno = ENODATA; ret = -1; } - if (ret == -1 && errno == EAGAIN) { + if ((ret < 0) && (errno == EAGAIN)) { /* done for now */ break; - } - this->total_bytes_read += ret; + } else if (ret > 0) + this->total_bytes_read += ret; } if (ret == 0) { @@ -708,7 +756,7 @@ __socket_rwv(rpc_transport_t *this, struct iovec *vector, int count, errno = ENOTCONN; break; } - if (ret == -1) { + if (ret < 0) { if (errno == EINTR) continue; @@ -771,24 +819,16 @@ static int __socket_readv(rpc_transport_t *this, struct iovec *vector, int count, struct iovec **pending_vector, int *pending_count, size_t *bytes) { - int ret = -1; - - ret = __socket_rwv(this, vector, count, pending_vector, pending_count, - bytes, 0); - - return ret; + return __socket_rwv(this, vector, count, pending_vector, pending_count, + bytes, 0); } static int __socket_writev(rpc_transport_t *this, struct iovec *vector, int count, struct iovec **pending_vector, int *pending_count) { - int ret = -1; - - ret = __socket_rwv(this, vector, count, pending_vector, pending_count, NULL, - 1); - - return ret; + return __socket_rwv(this, vector, count, pending_vector, pending_count, + NULL, 1); } static int @@ -804,6 +844,9 @@ __socket_shutdown(rpc_transport_t *this) why it failed to shutdown in normal cases */ gf_log(this->name, GF_LOG_DEBUG, "shutdown() returned %d. %s", ret, strerror(errno)); + } else { + GF_LOG_OCCASIONALLY(priv->shutdown_log_ctr, this->name, GF_LOG_INFO, + "intentional socket shutdown(%d)", priv->sock); } return ret; @@ -812,20 +855,14 @@ __socket_shutdown(rpc_transport_t *this) static int __socket_teardown_connection(rpc_transport_t *this) { - int ret = -1; socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; if (priv->use_ssl) ssl_teardown_connection(priv); - ret = __socket_shutdown(this); -out: - return ret; + return __socket_shutdown(this); } static int @@ -834,15 +871,12 @@ __socket_disconnect(rpc_transport_t *this) int ret = -1; socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; gf_log(this->name, GF_LOG_TRACE, "disconnecting %p, sock=%d", this, priv->sock); - if (priv->sock != -1) { + if (priv->sock >= 0) { gf_log_callingfn(this->name, GF_LOG_TRACE, "tearing down socket connection"); ret = __socket_teardown_connection(this); @@ -853,7 +887,6 @@ __socket_disconnect(rpc_transport_t *this) } } -out: return ret; } @@ -861,19 +894,22 @@ static int __socket_server_bind(rpc_transport_t *this) { socket_private_t *priv = NULL; + glusterfs_ctx_t *ctx = NULL; + cmd_args_t *cmd_args = NULL; + struct sockaddr_storage unix_addr = {0}; int ret = -1; int opt = 1; int reuse_check_sock = -1; - struct sockaddr_storage unix_addr = {0}; - - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); + uint16_t sin_port = 0; + int retries = 0; priv = this->private; + ctx = this->ctx; + cmd_args = &ctx->cmd_args; ret = setsockopt(priv->sock, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof(opt)); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_ERROR, "setsockopt() for SO_REUSEADDR failed (%s)", strerror(errno)); } @@ -882,27 +918,72 @@ __socket_server_bind(rpc_transport_t *this) if (AF_UNIX == SA(&this->myinfo.sockaddr)->sa_family) { memcpy(&unix_addr, SA(&this->myinfo.sockaddr), this->myinfo.sockaddr_len); - reuse_check_sock = socket(AF_UNIX, SOCK_STREAM, 0); + reuse_check_sock = sys_socket(AF_UNIX, SOCK_STREAM, 0); if (reuse_check_sock >= 0) { ret = connect(reuse_check_sock, SA(&unix_addr), this->myinfo.sockaddr_len); - if ((ret == -1) && (ECONNREFUSED == errno)) { + if ((ret != 0) && (ECONNREFUSED == errno)) { sys_unlink(((struct sockaddr_un *)&unix_addr)->sun_path); } + gf_log(this->name, GF_LOG_INFO, + "closing (AF_UNIX) reuse check socket %d", reuse_check_sock); sys_close(reuse_check_sock); } } - ret = bind(priv->sock, (struct sockaddr *)&this->myinfo.sockaddr, - this->myinfo.sockaddr_len); - - if (ret == -1) { - gf_log(this->name, GF_LOG_ERROR, "binding to %s failed: %s", - this->myinfo.identifier, strerror(errno)); - if (errno == EADDRINUSE) { - gf_log(this->name, GF_LOG_ERROR, "Port is already in use"); + if (AF_UNIX != SA(&this->myinfo.sockaddr)->sa_family) { + sin_port = (int)ntohs( + ((struct sockaddr_in *)&this->myinfo.sockaddr)->sin_port); + if (!sin_port) { + sin_port = GF_DEFAULT_SOCKET_LISTEN_PORT; + ((struct sockaddr_in *)&this->myinfo.sockaddr)->sin_port = htons( + sin_port); + } + retries = 10; + while (retries) { + ret = bind(priv->sock, (struct sockaddr *)&this->myinfo.sockaddr, + this->myinfo.sockaddr_len); + if (ret != 0) { + gf_log(this->name, GF_LOG_ERROR, "binding to %s failed: %s", + this->myinfo.identifier, strerror(errno)); + if (errno == EADDRINUSE) { + gf_log(this->name, GF_LOG_ERROR, "Port is already in use"); + sleep(1); + retries--; + } else { + break; + } + } else { + break; + } + } + } else { + ret = bind(priv->sock, (struct sockaddr *)&this->myinfo.sockaddr, + this->myinfo.sockaddr_len); - ret = -EADDRINUSE; + if (ret != 0) { + gf_log(this->name, GF_LOG_ERROR, "binding to %s failed: %s", + this->myinfo.identifier, strerror(errno)); + if (errno == EADDRINUSE) { + gf_log(this->name, GF_LOG_ERROR, "Port is already in use"); + } + } + } + if (AF_UNIX != SA(&this->myinfo.sockaddr)->sa_family) { + if (getsockname(priv->sock, SA(&this->myinfo.sockaddr), + &this->myinfo.sockaddr_len) != 0) { + gf_log(this->name, GF_LOG_WARNING, + "getsockname on (%d) failed (%s)", priv->sock, + strerror(errno)); + ret = -1; + goto out; + } + if (!cmd_args->brick_port) { + cmd_args->brick_port = (int)ntohs( + ((struct sockaddr_in *)&this->myinfo.sockaddr)->sin_port); + gf_log(this->name, GF_LOG_INFO, + "process started listening on port (%d)", + cmd_args->brick_port); } } @@ -918,7 +999,7 @@ __socket_nonblock(int fd) flags = fcntl(fd, F_GETFL); - if (flags != -1) + if (flags >= 0) ret = fcntl(fd, F_SETFL, flags | O_NONBLOCK); return ret; @@ -948,7 +1029,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, #endif ret = setsockopt(fd, SOL_SOCKET, SO_KEEPALIVE, &on, sizeof(on)); - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set keep alive option on socket %d", fd); goto err; @@ -965,7 +1046,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPALIVE, &keepaliveintvl, sizeof(keepaliveintvl)); #endif - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set keep alive interval on socket %d", fd); goto err; @@ -976,7 +1057,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPIDLE, &keepaliveidle, sizeof(keepaliveidle)); - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set keep idle %d on socket %d, %s", keepaliveidle, fd, strerror(errno)); @@ -984,7 +1065,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, } ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPINTVL, &keepaliveintvl, sizeof(keepaliveintvl)); - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set keep interval %d on socket %d, %s", keepaliveintvl, fd, strerror(errno)); @@ -996,7 +1077,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, goto done; ret = setsockopt(fd, IPPROTO_TCP, TCP_USER_TIMEOUT, &timeout_ms, sizeof(timeout_ms)); - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set " "TCP_USER_TIMEOUT %d on socket %d, %s", @@ -1007,7 +1088,7 @@ __socket_keepalive(int fd, int family, int keepaliveintvl, int keepaliveidle, #if defined(TCP_KEEPCNT) ret = setsockopt(fd, IPPROTO_TCP, TCP_KEEPCNT, &keepalivecnt, sizeof(keepalivecnt)); - if (ret == -1) { + if (ret != 0) { gf_log("socket", GF_LOG_WARNING, "failed to set " "TCP_KEEPCNT %d on socket %d, %s", @@ -1049,9 +1130,6 @@ __socket_reset(rpc_transport_t *this) { socket_private_t *priv = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; /* TODO: use mem-pool on incoming data */ @@ -1070,8 +1148,16 @@ __socket_reset(rpc_transport_t *this) memset(&priv->incoming, 0, sizeof(priv->incoming)); - event_unregister_close(this->ctx->event_pool, priv->sock, priv->idx); - + gf_event_unregister_close(this->ctx->event_pool, priv->sock, priv->idx); + if (priv->use_ssl && priv->ssl_ssl) { + SSL_clear(priv->ssl_ssl); + SSL_free(priv->ssl_ssl); + priv->ssl_ssl = NULL; + } + if (priv->ssl_ctx) { + SSL_CTX_free(priv->ssl_ctx); + priv->ssl_ctx = NULL; + } priv->sock = -1; priv->idx = -1; priv->connected = -1; @@ -1091,8 +1177,6 @@ __socket_reset(rpc_transport_t *this) GF_FREE(priv->ssl_ca_list); priv->ssl_ca_list = NULL; } -out: - return; } static void @@ -1122,8 +1206,6 @@ __socket_ioq_new(rpc_transport_t *this, rpc_transport_msg_t *msg) int count = 0; uint32_t size = 0; - GF_VALIDATE_OR_GOTO("socket", this, out); - /* TODO: use mem-pool */ entry = GF_CALLOC(1, sizeof(*entry), gf_common_mt_ioq); if (!entry) @@ -1178,7 +1260,6 @@ __socket_ioq_new(rpc_transport_t *this, rpc_transport_msg_t *msg) INIT_LIST_HEAD(&entry->list); -out: return entry; } @@ -1199,27 +1280,18 @@ out: } static void -__socket_ioq_flush(rpc_transport_t *this) +__socket_ioq_flush(socket_private_t *priv) { - socket_private_t *priv = NULL; struct ioq *entry = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - - priv = this->private; - while (!list_empty(&priv->ioq)) { entry = priv->ioq_next; __socket_ioq_entry_free(entry); } - -out: - return; } static int -__socket_ioq_churn_entry(rpc_transport_t *this, struct ioq *entry, int direct) +__socket_ioq_churn_entry(rpc_transport_t *this, struct ioq *entry) { int ret = -1; @@ -1242,16 +1314,13 @@ __socket_ioq_churn(rpc_transport_t *this) int ret = 0; struct ioq *entry = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; while (!list_empty(&priv->ioq)) { /* pick next entry */ entry = priv->ioq_next; - ret = __socket_ioq_churn_entry(this, entry, 0); + ret = __socket_ioq_churn_entry(this, entry); if (ret != 0) break; @@ -1259,11 +1328,10 @@ __socket_ioq_churn(rpc_transport_t *this) if (list_empty(&priv->ioq)) { /* all pending writes done, not interested in POLLOUT */ - priv->idx = event_select_on(this->ctx->event_pool, priv->sock, - priv->idx, -1, 0); + priv->idx = gf_event_select_on(this->ctx->event_pool, priv->sock, + priv->idx, -1, 0); } -out: return ret; } @@ -1273,22 +1341,17 @@ socket_event_poll_err(rpc_transport_t *this, int gen, int idx) socket_private_t *priv = NULL; gf_boolean_t socket_closed = _gf_false; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { - if ((priv->gen == gen) && (priv->idx == idx) && (priv->sock != -1)) { - __socket_ioq_flush(this); + if ((priv->gen == gen) && (priv->idx == idx) && (priv->sock >= 0)) { + __socket_ioq_flush(priv); __socket_reset(this); socket_closed = _gf_true; } } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); if (socket_closed) { pthread_mutex_lock(&priv->notify.lock); @@ -1301,7 +1364,6 @@ socket_event_poll_err(rpc_transport_t *this, int gen, int idx) rpc_transport_notify(this, RPC_TRANSPORT_DISCONNECT, this); } -out: return socket_closed; } @@ -1311,9 +1373,6 @@ socket_event_poll_out(rpc_transport_t *this) socket_private_t *priv = NULL; int ret = -1; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; pthread_mutex_lock(&priv->out_lock); @@ -1321,7 +1380,7 @@ socket_event_poll_out(rpc_transport_t *this) if (priv->connected == 1) { ret = __socket_ioq_churn(this); - if (ret == -1) { + if (ret < 0) { gf_log(this->name, GF_LOG_TRACE, "__socket_ioq_churn returned -1; " "disconnecting socket"); @@ -1332,9 +1391,11 @@ socket_event_poll_out(rpc_transport_t *this) pthread_mutex_unlock(&priv->out_lock); if (ret == 0) - ret = rpc_transport_notify(this, RPC_TRANSPORT_MSG_SENT, NULL); + rpc_transport_notify(this, RPC_TRANSPORT_MSG_SENT, NULL); + + if (ret > 0) + ret = 0; -out: return ret; } @@ -1377,7 +1438,7 @@ __socket_read_simple_msg(rpc_transport_t *this) &bytes_read); } - if (ret == -1) { + if (ret < 0) { gf_log(this->name, GF_LOG_WARNING, "reading from socket failed. Error (%s), " "peer (%s)", @@ -1391,6 +1452,7 @@ __socket_read_simple_msg(rpc_transport_t *this) if (ret > 0) { gf_log(this->name, GF_LOG_TRACE, "partial read on non-blocking socket."); + ret = 0; break; } @@ -1403,12 +1465,6 @@ out: return ret; } -static int -__socket_read_simple_request(rpc_transport_t *this) -{ - return __socket_read_simple_msg(this); -} - #define rpc_cred_addr(buf) (buf + RPC_MSGTYPE_SIZE + RPC_CALL_BODY_SIZE - 4) #define rpc_verf_addr(fragcurrent) (fragcurrent - 4) @@ -1435,9 +1491,6 @@ __socket_read_vectored_request(rpc_transport_t *this, struct gf_sock_incoming_frag *frag = NULL; sp_rpcfrag_request_state_t *request = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; /* used to reduce the indirection */ @@ -1574,8 +1627,8 @@ __socket_read_vectored_request(rpc_transport_t *this, remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; - if ((ret == -1) || ((ret == 0) && (remaining_size == 0) && - RPC_LASTFRAG(in->fraghdr))) { + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + RPC_LASTFRAG(in->fraghdr))) { request->vector_state = SP_STATE_VECTORED_REQUEST_INIT; in->payload_vector.iov_len = ((unsigned long)frag->fragcurrent - (unsigned long) @@ -1584,7 +1637,6 @@ __socket_read_vectored_request(rpc_transport_t *this, break; } -out: return ret; } @@ -1601,9 +1653,6 @@ __socket_read_request(rpc_transport_t *this) struct gf_sock_incoming_frag *frag = NULL; sp_rpcfrag_request_state_t *request = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; /* used to reduce the indirection */ @@ -1647,20 +1696,19 @@ __socket_read_request(rpc_transport_t *this) if (vector_sizer) { ret = __socket_read_vectored_request(this, vector_sizer); } else { - ret = __socket_read_simple_request(this); + ret = __socket_read_simple_msg(this); } remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; - if ((ret == -1) || ((ret == 0) && (remaining_size == 0) && - (RPC_LASTFRAG(in->fraghdr)))) { + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + (RPC_LASTFRAG(in->fraghdr)))) { request->header_state = SP_STATE_REQUEST_HEADER_INIT; } break; } -out: return ret; } @@ -1678,9 +1726,7 @@ __socket_read_accepted_successful_reply(rpc_transport_t *this) XDR xdr; struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); + uint32_t remaining_size = 0; priv = this->private; @@ -1718,9 +1764,9 @@ __socket_read_accepted_successful_reply(rpc_transport_t *this) free(read_rsp.xdata.xdata_val); - /* need to round off to proper roof (%4), as XDR packing pads + /* need to round off to proper gf_roof (%4), as XDR packing pads the end of opaque object with '0' */ - size = roof(read_rsp.xdata.xdata_len, 4); + size = gf_roof(read_rsp.xdata.xdata_len, 4); if (!size) { frag->call_body.reply @@ -1781,7 +1827,9 @@ __socket_read_accepted_successful_reply(rpc_transport_t *this) case SP_STATE_READ_PROC_HEADER: /* now read the entire remaining msg into new iobuf */ ret = __socket_read_simple_msg(this); - if ((ret == -1) || ((ret == 0) && RPC_LASTFRAG(in->fraghdr))) { + remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + RPC_LASTFRAG(in->fraghdr))) { frag->call_body.reply.accepted_success_state = SP_STATE_ACCEPTED_SUCCESS_REPLY_INIT; } @@ -1807,9 +1855,7 @@ __socket_read_accepted_successful_reply_v2(rpc_transport_t *this) XDR xdr; struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); + uint32_t remaining_size = 0; priv = this->private; @@ -1848,9 +1894,9 @@ __socket_read_accepted_successful_reply_v2(rpc_transport_t *this) free(read_rsp.xdata.pairs.pairs_val); - /* need to round off to proper roof (%4), as XDR packing pads + /* need to round off to proper gf_roof (%4), as XDR packing pads the end of opaque object with '0' */ - size = roof(read_rsp.xdata.xdr_size, 4); + size = gf_roof(read_rsp.xdata.xdr_size, 4); if (!size) { frag->call_body.reply @@ -1911,7 +1957,9 @@ __socket_read_accepted_successful_reply_v2(rpc_transport_t *this) case SP_STATE_READ_PROC_HEADER: /* now read the entire remaining msg into new iobuf */ ret = __socket_read_simple_msg(this); - if ((ret == -1) || ((ret == 0) && RPC_LASTFRAG(in->fraghdr))) { + remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + RPC_LASTFRAG(in->fraghdr))) { frag->call_body.reply.accepted_success_state = SP_STATE_ACCEPTED_SUCCESS_REPLY_INIT; } @@ -1937,9 +1985,6 @@ __socket_read_accepted_reply(rpc_transport_t *this) struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; /* used to reduce the indirection */ in = &priv->incoming; @@ -2010,8 +2055,8 @@ __socket_read_accepted_reply(rpc_transport_t *this) remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; - if ((ret == -1) || ((ret == 0) && (remaining_size == 0) && - (RPC_LASTFRAG(in->fraghdr)))) { + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + (RPC_LASTFRAG(in->fraghdr)))) { frag->call_body.reply .accepted_state = SP_STATE_ACCEPTED_REPLY_INIT; } @@ -2019,7 +2064,6 @@ __socket_read_accepted_reply(rpc_transport_t *this) break; } -out: return ret; } @@ -2041,9 +2085,6 @@ __socket_read_vectored_reply(rpc_transport_t *this) struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; in = &priv->incoming; frag = &in->frag; @@ -2076,8 +2117,8 @@ __socket_read_vectored_reply(rpc_transport_t *this) remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; - if ((ret == -1) || ((ret == 0) && (remaining_size == 0) && - (RPC_LASTFRAG(in->fraghdr)))) { + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + (RPC_LASTFRAG(in->fraghdr)))) { frag->call_body.reply .status_state = SP_STATE_VECTORED_REPLY_STATUS_INIT; in->payload_vector.iov_len = (unsigned long)frag->fragcurrent - @@ -2087,7 +2128,6 @@ __socket_read_vectored_reply(rpc_transport_t *this) break; } -out: return ret; } @@ -2110,9 +2150,6 @@ __socket_read_reply(rpc_transport_t *this) struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; in = &priv->incoming; frag = &in->frag; @@ -2138,17 +2175,13 @@ __socket_read_reply(rpc_transport_t *this) * and priv->lock, since we are doing an upcall here. */ frag->state = SP_STATE_NOTIFYING_XID; - pthread_mutex_unlock(&priv->in_lock); - { - ret = rpc_transport_notify(this, RPC_TRANSPORT_MAP_XID_REQUEST, - in->request_info); - } - pthread_mutex_lock(&priv->in_lock); + ret = rpc_transport_notify(this, RPC_TRANSPORT_MAP_XID_REQUEST, + in->request_info); /* Transition back to externally visible state. */ frag->state = SP_STATE_READ_MSGTYPE; - if (ret == -1) { + if (ret < 0) { gf_log(this->name, GF_LOG_WARNING, "notify for event MAP_XID failed for %s", this->peerinfo.identifier); @@ -2182,9 +2215,6 @@ __socket_read_frag(rpc_transport_t *this) struct gf_sock_incoming *in = NULL; struct gf_sock_incoming_frag *frag = NULL; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; /* used to reduce the indirection */ in = &priv->incoming; @@ -2226,8 +2256,8 @@ __socket_read_frag(rpc_transport_t *this) remaining_size = RPC_FRAGSIZE(in->fraghdr) - frag->bytes_read; - if ((ret == -1) || ((ret == 0) && (remaining_size == 0) && - (RPC_LASTFRAG(in->fraghdr)))) { + if ((ret < 0) || ((ret == 0) && (remaining_size == 0) && + (RPC_LASTFRAG(in->fraghdr)))) { /* frag->state = SP_STATE_NADA; */ frag->state = SP_STATE_RPCFRAG_INIT; } @@ -2242,7 +2272,6 @@ __socket_read_frag(rpc_transport_t *this) break; } -out: return ret; } @@ -2311,13 +2340,14 @@ __socket_proto_state_machine(rpc_transport_t *this, ret = __socket_readv(this, in->pending_vector, 1, &in->pending_vector, &in->pending_count, NULL); - if (ret == -1) + if (ret < 0) goto out; if (ret > 0) { gf_log(this->name, GF_LOG_TRACE, "partial " "fragment header read"); + ret = 0; goto out; } @@ -2332,7 +2362,7 @@ __socket_proto_state_machine(rpc_transport_t *this, in->total_bytes_read += RPC_FRAGSIZE(in->fraghdr); if (in->total_bytes_read >= GF_UNIT_GB) { - ret = -ENOMEM; + ret = -1; goto out; } @@ -2340,7 +2370,7 @@ __socket_proto_state_machine(rpc_transport_t *this, this->ctx->iobuf_pool, (in->total_bytes_read + sizeof(in->fraghdr))); if (!iobuf) { - ret = -ENOMEM; + ret = -1; goto out; } @@ -2367,7 +2397,7 @@ __socket_proto_state_machine(rpc_transport_t *this, case SP_STATE_READING_FRAG: ret = __socket_read_frag(this); - if ((ret == -1) || + if ((ret < 0) || (frag->bytes_read != RPC_FRAGSIZE(in->fraghdr))) { goto out; } @@ -2448,10 +2478,6 @@ __socket_proto_state_machine(rpc_transport_t *this, } out: - if ((ret == -1) && (errno == EAGAIN)) { - ret = 0; - } - return ret; } @@ -2459,22 +2485,34 @@ static int socket_proto_state_machine(rpc_transport_t *this, rpc_transport_pollin_t **pollin) { - socket_private_t *priv = NULL; - int ret = 0; + return __socket_proto_state_machine(this, pollin); +} - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); +static void +socket_event_poll_in_async(xlator_t *xl, gf_async_t *async) +{ + rpc_transport_pollin_t *pollin; + rpc_transport_t *this; + socket_private_t *priv; + pollin = caa_container_of(async, rpc_transport_pollin_t, async); + this = pollin->trans; priv = this->private; - pthread_mutex_lock(&priv->in_lock); + rpc_transport_notify(this, RPC_TRANSPORT_MSG_RECEIVED, pollin); + + rpc_transport_unref(this); + + rpc_transport_pollin_destroy(pollin); + + pthread_mutex_lock(&priv->notify.lock); { - ret = __socket_proto_state_machine(this, pollin); - } - pthread_mutex_unlock(&priv->in_lock); + --priv->notify.in_progress; -out: - return ret; + if (!priv->notify.in_progress) + pthread_cond_signal(&priv->notify.cond); + } + pthread_mutex_unlock(&priv->notify.lock); } static int @@ -2497,22 +2535,12 @@ socket_event_poll_in(rpc_transport_t *this, gf_boolean_t notify_handled) pthread_mutex_unlock(&priv->notify.lock); } - if (notify_handled && (ret != -1)) - event_handled(ctx->event_pool, priv->sock, priv->idx, priv->gen); + if (notify_handled && (ret >= 0)) + gf_event_handled(ctx->event_pool, priv->sock, priv->idx, priv->gen); if (pollin) { - ret = rpc_transport_notify(this, RPC_TRANSPORT_MSG_RECEIVED, pollin); - - rpc_transport_pollin_destroy(pollin); - - pthread_mutex_lock(&priv->notify.lock); - { - --priv->notify.in_progress; - - if (!priv->notify.in_progress) - pthread_cond_signal(&priv->notify.cond); - } - pthread_mutex_unlock(&priv->notify.lock); + rpc_transport_ref(this); + gf_async(&pollin->async, THIS, socket_event_poll_in_async); } return ret; @@ -2526,12 +2554,8 @@ socket_connect_finish(rpc_transport_t *this) rpc_transport_event_t event = 0; char notify_rpc = 0; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { if (priv->connected != 0) @@ -2541,10 +2565,10 @@ socket_connect_finish(rpc_transport_t *this) ret = __socket_connect_finish(priv->sock); - if (ret == -1 && errno == EINPROGRESS) + if ((ret < 0) && (errno == EINPROGRESS)) ret = 1; - if (ret == -1 && errno != EINPROGRESS) { + if ((ret < 0) && (errno != EINPROGRESS)) { if (!priv->connect_finish_log) { gf_log(this->name, GF_LOG_ERROR, "connection to %s failed (%s); " @@ -2563,7 +2587,7 @@ socket_connect_finish(rpc_transport_t *this) ret = getsockname(priv->sock, SA(&this->myinfo.sockaddr), &this->myinfo.sockaddr_len); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_WARNING, "getsockname on (%d) failed (%s) - " "disconnecting socket", @@ -2580,12 +2604,11 @@ socket_connect_finish(rpc_transport_t *this) } unlock: pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); if (notify_rpc) { rpc_transport_notify(this, event, this); } -out: + return ret; } @@ -2594,12 +2617,8 @@ socket_disconnect(rpc_transport_t *this, gf_boolean_t wait); /* socket_is_connected() is for use only in socket_event_handler() */ static inline gf_boolean_t -socket_is_connected(rpc_transport_t *this) +socket_is_connected(socket_private_t *priv) { - socket_private_t *priv = NULL; - - priv = this->private; - if (priv->use_ssl) { return priv->is_server ? priv->ssl_accepted : priv->ssl_connected; } else { @@ -2624,10 +2643,10 @@ ssl_rearm_event_fd(rpc_transport_t *this) fd = priv->sock; if (priv->ssl_error_required == SSL_ERROR_WANT_READ) - event_select_on(ctx->event_pool, fd, idx, 1, -1); + gf_event_select_on(ctx->event_pool, fd, idx, 1, -1); if (priv->ssl_error_required == SSL_ERROR_WANT_WRITE) - event_select_on(ctx->event_pool, fd, idx, -1, 1); - event_handled(ctx->event_pool, fd, idx, gen); + gf_event_select_on(ctx->event_pool, fd, idx, -1, 1); + gf_event_handled(ctx->event_pool, fd, idx, gen); } static int @@ -2648,7 +2667,7 @@ ssl_handle_server_connection_attempt(rpc_transport_t *this) fd = priv->sock; if (!priv->ssl_context_created) { - ret = ssl_setup_connection_prefix(this); + ret = ssl_setup_connection_prefix(this, _gf_true); if (ret < 0) { gf_log(this->name, GF_LOG_TRACE, "> ssl_setup_connection_prefix() failed!"); @@ -2661,8 +2680,8 @@ ssl_handle_server_connection_attempt(rpc_transport_t *this) ret = ssl_complete_connection(this); if (ret == 0) { /* nothing to do */ - event_select_on(ctx->event_pool, fd, idx, 1, 0); - event_handled(ctx->event_pool, fd, idx, gen); + gf_event_select_on(ctx->event_pool, fd, idx, 1, 0); + gf_event_handled(ctx->event_pool, fd, idx, gen); ret = 1; } else { if (errno == EAGAIN) { @@ -2696,13 +2715,13 @@ ssl_handle_client_connection_attempt(rpc_transport_t *this) /* SSL client */ if (priv->connect_failed) { gf_log(this->name, GF_LOG_TRACE, ">>> disconnecting SSL socket"); - ret = socket_disconnect(this, _gf_false); + (void)socket_disconnect(this, _gf_false); /* Force ret to be -1, as we are officially done with this socket */ ret = -1; } else { if (!priv->ssl_context_created) { - ret = ssl_setup_connection_prefix(this); + ret = ssl_setup_connection_prefix(this, _gf_false); if (ret < 0) { gf_log(this->name, GF_LOG_TRACE, "> ssl_setup_connection_prefix() " @@ -2716,7 +2735,7 @@ ssl_handle_client_connection_attempt(rpc_transport_t *this) ret = ssl_complete_connection(this); if (ret == 0) { ret = socket_connect_finish(this); - event_select_on(ctx->event_pool, fd, idx, 1, 0); + gf_event_select_on(ctx->event_pool, fd, idx, 1, 0); gf_log(this->name, GF_LOG_TRACE, ">>> completed client connect"); } else { if (errno == EAGAIN) { @@ -2726,7 +2745,7 @@ ssl_handle_client_connection_attempt(rpc_transport_t *this) ret = 1; } else { /* this is a connection failure */ - ret = socket_connect_finish(this); + (void)socket_connect_finish(this); gf_log(this->name, GF_LOG_TRACE, "ssl_complete_connection " "returned error"); @@ -2785,7 +2804,7 @@ socket_handle_client_connection_attempt(rpc_transport_t *this) * return 1 */ ret = 1; - event_handled(ctx->event_pool, fd, idx, gen); + gf_event_handled(ctx->event_pool, fd, idx, gen); } } return ret; @@ -2821,7 +2840,7 @@ socket_complete_connection(rpc_transport_t *this) * socket_server_event_handler() */ priv->accepted = _gf_true; - event_handled(ctx->event_pool, fd, idx, gen); + gf_event_handled(ctx->event_pool, fd, idx, gen); ret = 1; } else { ret = socket_handle_client_connection_attempt(this); @@ -2831,9 +2850,9 @@ socket_complete_connection(rpc_transport_t *this) } /* reads rpc_requests during pollin */ -static int +static void socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, - int poll_out, int poll_err) + int poll_out, int poll_err, char event_thread_died) { rpc_transport_t *this = NULL; socket_private_t *priv = NULL; @@ -2843,6 +2862,18 @@ socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, this = data; + if (event_thread_died) { + /* to avoid duplicate notifications, notify only for listener sockets */ + return; + } + + /* At this point we are sure no other thread is using the transport because + * we cannot receive more events until we call gf_event_handled(). However + * this function may call gf_event_handled() in some cases. When this is + * done, the transport may be destroyed at any moment if another thread + * handled an error event. To prevent that we take a reference here. */ + rpc_transport_ref(this); + GF_VALIDATE_OR_GOTO("socket", this, out); GF_VALIDATE_OR_GOTO("socket", this->private, out); GF_VALIDATE_OR_GOTO("socket", this->xl, out); @@ -2851,21 +2882,19 @@ socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, priv = this->private; ctx = this->ctx; - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { priv->idx = idx; priv->gen = gen; } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); gf_log(this->name, GF_LOG_TRACE, "%s (sock:%d) in:%d, out:%d, err:%d", (priv->is_server ? "server" : "client"), priv->sock, poll_in, poll_out, poll_err); if (!poll_err) { - if (!socket_is_connected(this)) { + if (!socket_is_connected(priv)) { gf_log(this->name, GF_LOG_TRACE, "%s (sock:%d) socket is not connected, " "completing connection", @@ -2881,7 +2910,7 @@ socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, if (ret > 0) { gf_log(this->name, GF_LOG_TRACE, "(sock:%d) returning to wait on socket", priv->sock); - return 0; + goto out; } } else { char *sock_type = (priv->is_server ? "Server" : "Client"); @@ -2921,6 +2950,13 @@ socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, socket_dump_info(sa, priv->is_server, priv->use_ssl, priv->sock, this->name, "disconnecting from"); + /* Dump the SSL error stack to clear any errors that may otherwise + * resurface in the future. + */ + if (priv->use_ssl && priv->ssl_ssl) { + ssl_dump_error_stack(this->name); + } + /* Logging has happened already in earlier cases */ gf_log("transport", ((ret >= 0) ? GF_LOG_INFO : GF_LOG_DEBUG), "EPOLLERR - disconnecting (sock:%d) (%s)", priv->sock, @@ -2932,16 +2968,16 @@ socket_event_handler(int fd, int idx, int gen, void *data, int poll_in, rpc_transport_unref(this); } else if (!notify_handled) { - event_handled(ctx->event_pool, fd, idx, gen); + gf_event_handled(ctx->event_pool, fd, idx, gen); } out: - return ret; + rpc_transport_unref(this); } -static int +static void socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, - int poll_out, int poll_err) + int poll_out, int poll_err, char event_thread_died) { rpc_transport_t *this = NULL; socket_private_t *priv = NULL; @@ -2959,11 +2995,18 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, GF_VALIDATE_OR_GOTO("socket", this, out); GF_VALIDATE_OR_GOTO("socket", this->private, out); GF_VALIDATE_OR_GOTO("socket", this->xl, out); + GF_VALIDATE_OR_GOTO("socket", this->ctx, out); THIS = this->xl; priv = this->private; ctx = this->ctx; + if (event_thread_died) { + rpc_transport_notify(this, RPC_TRANSPORT_EVENT_THREAD_DIED, + (void *)(unsigned long)gen); + return; + } + /* NOTE: * We have done away with the critical section in this function. since * there's little that it helps with. There's no other code that @@ -2971,48 +3014,66 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, * thread context while we are using it here. */ priv->idx = idx; + priv->gen = gen; + + if (poll_err) { + socket_event_poll_err(this, gen, idx); + goto out; + } if (poll_in) { - new_sock = accept(priv->sock, SA(&new_sockaddr), &addrlen); + int aflags = 0; + + if (!priv->bio) + aflags = O_NONBLOCK; - if (ctx) - event_handled(ctx->event_pool, fd, idx, gen); + new_sock = sys_accept(priv->sock, SA(&new_sockaddr), &addrlen, aflags); - if (new_sock == -1) { + gf_event_handled(ctx->event_pool, fd, idx, gen); + + if (new_sock < 0) { gf_log(this->name, GF_LOG_WARNING, "accept on %d failed (%s)", priv->sock, strerror(errno)); goto out; } - if (priv->nodelay && (new_sockaddr.ss_family != AF_UNIX)) { - ret = __socket_nodelay(new_sock); - if (ret == -1) { - gf_log(this->name, GF_LOG_WARNING, - "setsockopt() failed for " - "NODELAY (%s)", - strerror(errno)); + if (new_sockaddr.ss_family != AF_UNIX) { + if (priv->nodelay) { + ret = __socket_nodelay(new_sock); + if (ret != 0) { + gf_log(this->name, GF_LOG_WARNING, + "setsockopt() failed for " + "NODELAY (%s)", + strerror(errno)); + } } - } - if (priv->keepalive && new_sockaddr.ss_family != AF_UNIX) { - ret = __socket_keepalive(new_sock, new_sockaddr.ss_family, - priv->keepaliveintvl, priv->keepaliveidle, - priv->keepalivecnt, priv->timeout); - if (ret == -1) - gf_log(this->name, GF_LOG_WARNING, - "Failed to set keep-alive: %s", strerror(errno)); + if (priv->keepalive) { + ret = __socket_keepalive( + new_sock, new_sockaddr.ss_family, priv->keepaliveintvl, + priv->keepaliveidle, priv->keepalivecnt, priv->timeout); + if (ret != 0) + gf_log(this->name, GF_LOG_WARNING, + "Failed to set keep-alive: %s", strerror(errno)); + } } new_trans = GF_CALLOC(1, sizeof(*new_trans), gf_common_mt_rpc_trans_t); if (!new_trans) { sys_close(new_sock); + gf_log(this->name, GF_LOG_WARNING, + "transport object allocation failure; " + "closed newly accepted socket %d", + new_sock); goto out; } ret = pthread_mutex_init(&new_trans->lock, NULL); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_WARNING, - "pthread_mutex_init() failed: %s", strerror(errno)); + "pthread_mutex_init() failed: %s; closing newly accepted " + "socket %d", + strerror(errno), new_sock); sys_close(new_sock); GF_FREE(new_trans); goto out; @@ -3028,8 +3089,10 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, ret = getsockname(new_sock, SA(&new_trans->myinfo.sockaddr), &new_trans->myinfo.sockaddr_len); - if (ret == -1) { - gf_log(this->name, GF_LOG_WARNING, "getsockname on %d failed (%s)", + if (ret != 0) { + gf_log(this->name, GF_LOG_WARNING, + "getsockname on socket %d " + "failed (errno:%s); closing newly accepted socket", new_sock, strerror(errno)); sys_close(new_sock); GF_FREE(new_trans->name); @@ -3041,8 +3104,35 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, gf_log(this->name, GF_LOG_TRACE, "XXX server:%s, client:%s", new_trans->myinfo.identifier, new_trans->peerinfo.identifier); + /* Make options available to local socket_init() to create new + * SSL_CTX per transport. A separate SSL_CTX per transport is + * required to avoid setting crl checking options for client + * connections. The verification options eventually get copied + * to the SSL object. Unfortunately, there's no way to identify + * whether socket_init() is being called after a client-side + * connect() or a server-side accept(). Although, we could pass + * a flag from the transport init() to the socket_init() and + * from this place, this doesn't identify the case where the + * server-side transport loading is done for the first time. + * Also, SSL doesn't apply for UNIX sockets. + */ + if (new_sockaddr.ss_family != AF_UNIX) + new_trans->options = dict_ref(this->options); + new_trans->ctx = this->ctx; + ret = socket_init(new_trans); + + /* reset options to NULL to avoid double free */ + if (new_sockaddr.ss_family != AF_UNIX) { + dict_unref(new_trans->options); + new_trans->options = NULL; + } + if (ret != 0) { + gf_log(this->name, GF_LOG_WARNING, + "initialization of new_trans " + "failed; closing newly accepted socket %d", + new_sock); sys_close(new_sock); GF_FREE(new_trans->name); GF_FREE(new_trans); @@ -3056,6 +3146,7 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, new_trans->mydata = this->mydata; new_trans->notify = this->notify; new_trans->listener = this; + new_trans->notify_poller_death = this->poller_death_accept; new_priv = new_trans->private; if (new_sockaddr.ss_family == AF_UNIX) { @@ -3078,25 +3169,9 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, new_priv->sock = new_sock; new_priv->ssl_enabled = priv->ssl_enabled; - new_priv->ssl_ctx = priv->ssl_ctx; new_priv->connected = 1; new_priv->is_server = _gf_true; - /* set O_NONBLOCK for plain text as well as ssl connections */ - if (!priv->bio) { - gf_log(this->name, GF_LOG_TRACE, "### use non-blocking IO"); - ret = __socket_nonblock(new_sock); - - if (ret == -1) { - gf_log(this->name, GF_LOG_WARNING, "NBIO on %d failed (%s)", - new_sock, strerror(errno)); - - sys_close(new_sock); - GF_FREE(new_trans->name); - GF_FREE(new_trans); - goto out; - } - } /* * This is the first ref on the newly accepted * transport. @@ -3120,10 +3195,10 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, */ ret = rpc_transport_notify(this, RPC_TRANSPORT_ACCEPT, new_trans); - if (ret != -1) { - new_priv->idx = event_register(ctx->event_pool, new_sock, - socket_event_handler, new_trans, - 1, 0); + if (ret >= 0) { + new_priv->idx = gf_event_register( + ctx->event_pool, new_sock, socket_event_handler, new_trans, + 1, 0, new_trans->notify_poller_death); if (new_priv->idx == -1) { ret = -1; gf_log(this->name, GF_LOG_ERROR, @@ -3158,7 +3233,8 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, rpc_transport_unref(new_trans); } - if (ret == -1) { + if (ret < 0) { + gf_log(this->name, GF_LOG_WARNING, "closing newly accepted socket"); sys_close(new_sock); /* this unref is to actually cause the destruction of * the new_trans since we've failed at everything so far @@ -3167,7 +3243,7 @@ socket_server_event_handler(int fd, int idx, int gen, void *data, int poll_in, } } out: - return ret; + return; } static int @@ -3176,20 +3252,14 @@ socket_disconnect(rpc_transport_t *this, gf_boolean_t wait) socket_private_t *priv = NULL; int ret = -1; - GF_VALIDATE_OR_GOTO("socket", this, out); - GF_VALIDATE_OR_GOTO("socket", this->private, out); - priv = this->private; - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { ret = __socket_disconnect(this); } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); -out: return ret; } @@ -3288,10 +3358,9 @@ socket_connect(rpc_transport_t *this, int port) goto err; } - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { - if (priv->sock != -1) { + if (priv->sock >= 0) { gf_log_callingfn(this->name, GF_LOG_TRACE, "connect () called on transport " "already connected"); @@ -3305,7 +3374,7 @@ socket_connect(rpc_transport_t *this, int port) ret = socket_client_get_remote_sockaddr(this, &sock_union.sa, &sockaddr_len, &sa_family); - if (ret == -1) { + if (ret < 0) { /* logged inside client_get_remote_sockaddr */ goto unlock; } @@ -3323,8 +3392,8 @@ socket_connect(rpc_transport_t *this, int port) memcpy(&this->peerinfo.sockaddr, &sock_union.storage, sockaddr_len); this->peerinfo.sockaddr_len = sockaddr_len; - priv->sock = socket(sa_family, SOCK_STREAM, 0); - if (priv->sock == -1) { + priv->sock = sys_socket(sa_family, SOCK_STREAM, 0); + if (priv->sock < 0) { gf_log(this->name, GF_LOG_ERROR, "socket creation failed (%s)", strerror(errno)); ret = -1; @@ -3336,7 +3405,7 @@ socket_connect(rpc_transport_t *this, int port) */ if (priv->windowsize != 0) { if (setsockopt(priv->sock, SOL_SOCKET, SO_RCVBUF, &priv->windowsize, - sizeof(priv->windowsize)) < 0) { + sizeof(priv->windowsize)) != 0) { gf_log(this->name, GF_LOG_ERROR, "setting receive window " "size failed: %d: %d: %s", @@ -3344,7 +3413,7 @@ socket_connect(rpc_transport_t *this, int port) } if (setsockopt(priv->sock, SOL_SOCKET, SO_SNDBUF, &priv->windowsize, - sizeof(priv->windowsize)) < 0) { + sizeof(priv->windowsize)) != 0) { gf_log(this->name, GF_LOG_ERROR, "setting send window size " "failed: %d: %d: %s", @@ -3366,30 +3435,32 @@ socket_connect(rpc_transport_t *this, int port) } #endif - if (priv->nodelay && (sa_family != AF_UNIX)) { - ret = __socket_nodelay(priv->sock); - - if (ret == -1) { - gf_log(this->name, GF_LOG_ERROR, "NODELAY on %d failed (%s)", - priv->sock, strerror(errno)); + if (sa_family != AF_UNIX) { + if (priv->nodelay) { + ret = __socket_nodelay(priv->sock); + if (ret != 0) { + gf_log(this->name, GF_LOG_ERROR, + "NODELAY on %d failed (%s)", priv->sock, + strerror(errno)); + } } - } - if (priv->keepalive && sa_family != AF_UNIX) { - ret = __socket_keepalive(priv->sock, sa_family, - priv->keepaliveintvl, priv->keepaliveidle, - priv->keepalivecnt, priv->timeout); - if (ret == -1) - gf_log(this->name, GF_LOG_ERROR, "Failed to set keep-alive: %s", - strerror(errno)); + if (priv->keepalive) { + ret = __socket_keepalive( + priv->sock, sa_family, priv->keepaliveintvl, + priv->keepaliveidle, priv->keepalivecnt, priv->timeout); + if (ret != 0) + gf_log(this->name, GF_LOG_ERROR, + "Failed to set keep-alive: %s", strerror(errno)); + } } SA(&this->myinfo.sockaddr)->sa_family = SA(&this->peerinfo.sockaddr) ->sa_family; /* If a source addr is explicitly specified, use it */ - ret = dict_get_str(this->options, "transport.socket.source-addr", - &local_addr); + ret = dict_get_str_sizen(this->options, "transport.socket.source-addr", + &local_addr); if (!ret && SA(&this->myinfo.sockaddr)->sa_family == AF_INET) { addr = (struct sockaddr_in *)(&this->myinfo.sockaddr); ret = inet_pton(AF_INET, local_addr, &(addr->sin_addr.s_addr)); @@ -3401,7 +3472,7 @@ socket_connect(rpc_transport_t *this, int port) ret = client_bind(this, SA(&this->myinfo.sockaddr), &this->myinfo.sockaddr_len, priv->sock); - if (ret == -1) { + if (ret < 0) { gf_log(this->name, GF_LOG_WARNING, "client bind failed: %s", strerror(errno)); goto handler; @@ -3410,7 +3481,7 @@ socket_connect(rpc_transport_t *this, int port) /* make socket non-blocking for all types of sockets */ if (!priv->bio) { ret = __socket_nonblock(priv->sock); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_ERROR, "NBIO on %d failed (%s)", priv->sock, strerror(errno)); goto handler; @@ -3437,7 +3508,7 @@ socket_connect(rpc_transport_t *this, int port) connect_attempted = _gf_true; - if (ret == -1 && errno == ENOENT && ign_enoent) { + if ((ret != 0) && (errno == ENOENT) && ign_enoent) { gf_log(this->name, GF_LOG_WARNING, "Ignore failed connection attempt on %s, (%s) ", this->peerinfo.identifier, strerror(errno)); @@ -3455,7 +3526,7 @@ socket_connect(rpc_transport_t *this, int port) goto handler; } - if (ret == -1 && ((errno != EINPROGRESS) && (errno != ENOENT))) { + if ((ret != 0) && (errno != EINPROGRESS) && (errno != ENOENT)) { /* For unix path based sockets, the socket path is * cryptic (md5sum of path) and may not be useful for * the user in debugging so log it in DEBUG @@ -3489,8 +3560,9 @@ socket_connect(rpc_transport_t *this, int port) will result in EPOLLERR, so cleanup is done in socket_event_handler or socket_poller */ shutdown(priv->sock, SHUT_RDWR); - gf_log(this->name, GF_LOG_TRACE, - "@@@ client shutdown(%d, SHUT_RDWR)", priv->sock); + ret = 0; + gf_log(this->name, GF_LOG_INFO, + "intentional client shutdown(%d, SHUT_RDWR)", priv->sock); } priv->connected = 0; @@ -3499,10 +3571,14 @@ socket_connect(rpc_transport_t *this, int port) refd = _gf_true; this->listener = this; - priv->idx = event_register(ctx->event_pool, priv->sock, - socket_event_handler, this, 1, 1); + priv->idx = gf_event_register(ctx->event_pool, priv->sock, + socket_event_handler, this, 1, 1, + this->notify_poller_death); if (priv->idx == -1) { - gf_log("", GF_LOG_WARNING, "failed to register the event"); + gf_log("", GF_LOG_WARNING, + "failed to register the event; " + "closing socket %d", + priv->sock); sys_close(priv->sock); priv->sock = -1; ret = -1; @@ -3512,11 +3588,10 @@ socket_connect(rpc_transport_t *this, int port) sock = priv->sock; } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); err: - /* if sock != -1, then cleanup is done from the event handler */ - if (ret == -1 && sock == -1) { + /* if sock >= 0, then cleanup is done from the event handler */ + if ((ret < 0) && (sock < 0)) { /* Cleaup requires to send notification to upper layer which intern holds the big_lock. There can be dead-lock situation if big_lock is already held by the current thread. @@ -3536,6 +3611,8 @@ err: GF_FREE(arg); GF_ASSERT(0); } + + ret = 0; } return ret; @@ -3562,29 +3639,26 @@ socket_listen(rpc_transport_t *this) myinfo = &this->myinfo; ctx = this->ctx; - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { sock = priv->sock; } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); - if (sock != -1) { + if (sock >= 0) { gf_log_callingfn(this->name, GF_LOG_DEBUG, "already listening"); return ret; } ret = socket_server_get_local_sockaddr(this, SA(&sockaddr), &sockaddr_len, &sa_family); - if (ret == -1) { + if (ret < 0) { return ret; } - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { - if (priv->sock != -1) { + if (priv->sock >= 0) { gf_log(this->name, GF_LOG_DEBUG, "already listening"); goto unlock; } @@ -3592,9 +3666,9 @@ socket_listen(rpc_transport_t *this) memcpy(&myinfo->sockaddr, &sockaddr, sockaddr_len); myinfo->sockaddr_len = sockaddr_len; - priv->sock = socket(sa_family, SOCK_STREAM, 0); + priv->sock = sys_socket(sa_family, SOCK_STREAM, 0); - if (priv->sock == -1) { + if (priv->sock < 0) { gf_log(this->name, GF_LOG_ERROR, "socket creation failed (%s)", strerror(errno)); goto unlock; @@ -3605,7 +3679,7 @@ socket_listen(rpc_transport_t *this) */ if (priv->windowsize != 0) { if (setsockopt(priv->sock, SOL_SOCKET, SO_RCVBUF, &priv->windowsize, - sizeof(priv->windowsize)) < 0) { + sizeof(priv->windowsize)) != 0) { gf_log(this->name, GF_LOG_ERROR, "setting receive window size " "failed: %d: %d: %s", @@ -3613,7 +3687,7 @@ socket_listen(rpc_transport_t *this) } if (setsockopt(priv->sock, SOL_SOCKET, SO_SNDBUF, &priv->windowsize, - sizeof(priv->windowsize)) < 0) { + sizeof(priv->windowsize)) != 0) { gf_log(this->name, GF_LOG_ERROR, "setting send window size failed:" " %d: %d: %s", @@ -3623,7 +3697,7 @@ socket_listen(rpc_transport_t *this) if (priv->nodelay && (sa_family != AF_UNIX)) { ret = __socket_nodelay(priv->sock); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_ERROR, "setsockopt() failed for NODELAY (%s)", strerror(errno)); } @@ -3632,8 +3706,10 @@ socket_listen(rpc_transport_t *this) if (!priv->bio) { ret = __socket_nonblock(priv->sock); - if (ret == -1) { - gf_log(this->name, GF_LOG_ERROR, "NBIO on %d failed (%s)", + if (ret != 0) { + gf_log(this->name, GF_LOG_ERROR, + "NBIO on socket %d failed " + "(errno:%s); closing socket", priv->sock, strerror(errno)); sys_close(priv->sock); priv->sock = -1; @@ -3641,10 +3717,15 @@ socket_listen(rpc_transport_t *this) } } + /* coverity[SLEEP] */ ret = __socket_server_bind(this); - if ((ret == -EADDRINUSE) || (ret == -1)) { + if (ret < 0) { /* logged inside __socket_server_bind() */ + gf_log(this->name, GF_LOG_ERROR, + "__socket_server_bind failed;" + "closing socket %d", + priv->sock); sys_close(priv->sock); priv->sock = -1; goto unlock; @@ -3655,10 +3736,11 @@ socket_listen(rpc_transport_t *this) ret = listen(priv->sock, priv->backlog); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_ERROR, - "could not set socket %d to listen mode (%s)", priv->sock, - strerror(errno)); + "could not set socket %d to listen mode (errno:%s); " + "closing socket", + priv->sock, strerror(errno)); sys_close(priv->sock); priv->sock = -1; goto unlock; @@ -3666,12 +3748,15 @@ socket_listen(rpc_transport_t *this) rpc_transport_ref(this); - priv->idx = event_register(ctx->event_pool, priv->sock, - socket_server_event_handler, this, 1, 0); + priv->idx = gf_event_register(ctx->event_pool, priv->sock, + socket_server_event_handler, this, 1, 0, + this->notify_poller_death); if (priv->idx == -1) { gf_log(this->name, GF_LOG_WARNING, - "could not register socket %d with events", priv->sock); + "could not register socket %d with events; " + "closing socket", + priv->sock); ret = -1; sys_close(priv->sock); priv->sock = -1; @@ -3680,7 +3765,6 @@ socket_listen(rpc_transport_t *this) } unlock: pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); out: return ret; @@ -3719,7 +3803,7 @@ socket_submit_outgoing_msg(rpc_transport_t *this, rpc_transport_msg_t *msg) goto unlock; if (list_empty(&priv->ioq)) { - ret = __socket_ioq_churn_entry(this, entry, 1); + ret = __socket_ioq_churn_entry(this, entry); if (ret == 0) { need_append = 0; @@ -3735,8 +3819,8 @@ socket_submit_outgoing_msg(rpc_transport_t *this, rpc_transport_msg_t *msg) } if (need_poll_out) { /* first entry to wait. continue writing on POLLOUT */ - priv->idx = event_select_on(ctx->event_pool, priv->sock, priv->idx, - -1, 1); + priv->idx = gf_event_select_on(ctx->event_pool, priv->sock, + priv->idx, -1, 1); } } unlock: @@ -3845,7 +3929,6 @@ socket_throttle(rpc_transport_t *this, gf_boolean_t onoff) will never read() any more data until throttling is turned off. */ - pthread_mutex_lock(&priv->in_lock); pthread_mutex_lock(&priv->out_lock); { /* Throttling is useless on a disconnected transport. In fact, @@ -3854,11 +3937,10 @@ socket_throttle(rpc_transport_t *this, gf_boolean_t onoff) * registered fd mapping. */ if (priv->connected == 1) - priv->idx = event_select_on(this->ctx->event_pool, priv->sock, - priv->idx, (int)!onoff, -1); + priv->idx = gf_event_select_on(this->ctx->event_pool, priv->sock, + priv->idx, (int)!onoff, -1); } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); return 0; } @@ -3881,31 +3963,23 @@ reconfigure(rpc_transport_t *this, dict_t *options) socket_private_t *priv = NULL; gf_boolean_t tmp_bool = _gf_false; char *optstr = NULL; - int ret = 0; + int ret = -1; uint32_t backlog = 0; uint64_t windowsize = 0; - uint32_t timeout = 0; - int keepaliveidle = GF_KEEPALIVE_TIME; - int keepaliveintvl = GF_KEEPALIVE_INTERVAL; - int keepalivecnt = GF_KEEPALIVE_COUNT; + data_t *data; GF_VALIDATE_OR_GOTO("socket", this, out); GF_VALIDATE_OR_GOTO("socket", this->private, out); - if (!this || !this->private) { - ret = -1; - goto out; - } - priv = this->private; - if (dict_get_str(options, "transport.socket.keepalive", &optstr) == 0) { - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + if (dict_get_str_sizen(options, "transport.socket.keepalive", &optstr) == + 0) { + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_ERROR, "'transport.socket.keepalive' takes only " "boolean options, not taking any action"); priv->keepalive = 1; - ret = -1; goto out; } gf_log(this->name, GF_LOG_DEBUG, @@ -3915,48 +3989,43 @@ reconfigure(rpc_transport_t *this, dict_t *options) } else priv->keepalive = 1; - if (dict_get_int32(options, "transport.tcp-user-timeout", - &(priv->timeout)) != 0) - priv->timeout = timeout; + if (dict_get_int32_sizen(options, "transport.tcp-user-timeout", + &(priv->timeout)) != 0) + priv->timeout = GF_NETWORK_TIMEOUT; gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.tcp-user-timeout=%d", - priv->timeout); + "Reconfigured transport.tcp-user-timeout=%d", priv->timeout); if (dict_get_uint32(options, "transport.listen-backlog", &backlog) == 0) { priv->backlog = backlog; gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.listen-backlog=%d", - priv->backlog); + "Reconfigured transport.listen-backlog=%d", priv->backlog); } - if (dict_get_int32(options, "transport.socket.keepalive-time", - &(priv->keepaliveidle)) != 0) - priv->keepaliveidle = keepaliveidle; - gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.socket.keepalive-time=%d", - priv->keepaliveidle); + if (priv->keepalive) { + if (dict_get_int32_sizen(options, "transport.socket.keepalive-time", + &(priv->keepaliveidle)) != 0) + priv->keepaliveidle = GF_KEEPALIVE_TIME; + gf_log(this->name, GF_LOG_DEBUG, + "Reconfigured transport.socket.keepalive-time=%d", + priv->keepaliveidle); - if (dict_get_int32(options, "transport.socket.keepalive-interval", - &(priv->keepaliveintvl)) != 0) - priv->keepaliveintvl = keepaliveintvl; - gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.socket.keepalive-interval=%d", - priv->keepaliveintvl); + if (dict_get_int32_sizen(options, "transport.socket.keepalive-interval", + &(priv->keepaliveintvl)) != 0) + priv->keepaliveintvl = GF_KEEPALIVE_INTERVAL; + gf_log(this->name, GF_LOG_DEBUG, + "Reconfigured transport.socket.keepalive-interval=%d", + priv->keepaliveintvl); - if (dict_get_int32(options, "transport.socket.keepalive-count", - &(priv->keepalivecnt)) != 0) - priv->keepalivecnt = keepalivecnt; - gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.socket.keepalive-count=%d", - priv->keepalivecnt); + if (dict_get_int32_sizen(options, "transport.socket.keepalive-count", + &(priv->keepalivecnt)) != 0) + priv->keepalivecnt = GF_KEEPALIVE_COUNT; + gf_log(this->name, GF_LOG_DEBUG, + "Reconfigured transport.socket.keepalive-count=%d", + priv->keepalivecnt); + } optstr = NULL; - if (dict_get_str(options, "tcp-window-size", &optstr) == 0) { + if (dict_get_str_sizen(options, "tcp-window-size", &optstr) == 0) { if (gf_string2uint64(optstr, &windowsize) != 0) { gf_log(this->name, GF_LOG_ERROR, "invalid number format: %s", optstr); @@ -3966,10 +4035,11 @@ reconfigure(rpc_transport_t *this, dict_t *options) priv->windowsize = (int)windowsize; - if (dict_get(options, "non-blocking-io")) { - optstr = data_to_str(dict_get(options, "non-blocking-io")); + data = dict_get_sizen(options, "non-blocking-io"); + if (data) { + optstr = data_to_str(data); - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_ERROR, "'non-blocking-io' takes only boolean options," " not taking any action"); @@ -3984,7 +4054,7 @@ reconfigure(rpc_transport_t *this, dict_t *options) if (!priv->bio) { ret = __socket_nonblock(priv->sock); - if (ret == -1) { + if (ret != 0) { gf_log(this->name, GF_LOG_WARNING, "NBIO on %d failed (%s)", priv->sock, strerror(errno)); goto out; @@ -4098,6 +4168,34 @@ static void __attribute__((destructor)) fini_openssl_mt(void) ERR_free_strings(); } +/* The function returns 0 if AES bit is enabled on the CPU */ +static int +ssl_check_aes_bit(void) +{ + FILE *fp = fopen("/proc/cpuinfo", "r"); + int ret = 1; + size_t len = 0; + char *line = NULL; + char *match = NULL; + + GF_ASSERT(fp != NULL); + + while (getline(&line, &len, fp) > 0) { + if (!strncmp(line, "flags", 5)) { + match = strstr(line, " aes"); + if ((match != NULL) && ((match[4] == ' ') || (match[4] == 0))) { + ret = 0; + break; + } + } + } + + free(line); + fclose(fp); + + return ret; +} + static int ssl_setup_connection_params(rpc_transport_t *this) { @@ -4108,7 +4206,7 @@ ssl_setup_connection_params(rpc_transport_t *this) char *cipher_list = DEFAULT_CIPHER_LIST; char *dh_param = DEFAULT_DH_PARAM; char *ec_curve = DEFAULT_EC_CURVE; - char *crl_path = NULL; + gf_boolean_t dh_flag = _gf_false; priv = this->private; @@ -4117,8 +4215,16 @@ ssl_setup_connection_params(rpc_transport_t *this) return 0; } + if (!priv->ssl_enabled && !priv->mgmt_ssl) { + return 0; + } + + if (!ssl_check_aes_bit()) { + cipher_list = "AES128:" DEFAULT_CIPHER_LIST; + } + priv->ssl_own_cert = DEFAULT_CERT_PATH; - if (dict_get_str(this->options, SSL_OWN_CERT_OPT, &optstr) == 0) { + if (dict_get_str_sizen(this->options, SSL_OWN_CERT_OPT, &optstr) == 0) { if (!priv->ssl_enabled) { gf_log(this->name, GF_LOG_WARNING, "%s specified without %s (ignored)", SSL_OWN_CERT_OPT, @@ -4129,7 +4235,7 @@ ssl_setup_connection_params(rpc_transport_t *this) priv->ssl_own_cert = gf_strdup(priv->ssl_own_cert); priv->ssl_private_key = DEFAULT_KEY_PATH; - if (dict_get_str(this->options, SSL_PRIVATE_KEY_OPT, &optstr) == 0) { + if (dict_get_str_sizen(this->options, SSL_PRIVATE_KEY_OPT, &optstr) == 0) { if (!priv->ssl_enabled) { gf_log(this->name, GF_LOG_WARNING, "%s specified without %s (ignored)", SSL_PRIVATE_KEY_OPT, @@ -4140,7 +4246,7 @@ ssl_setup_connection_params(rpc_transport_t *this) priv->ssl_private_key = gf_strdup(priv->ssl_private_key); priv->ssl_ca_list = DEFAULT_CA_PATH; - if (dict_get_str(this->options, SSL_CA_LIST_OPT, &optstr) == 0) { + if (dict_get_str_sizen(this->options, SSL_CA_LIST_OPT, &optstr) == 0) { if (!priv->ssl_enabled) { gf_log(this->name, GF_LOG_WARNING, "%s specified without %s (ignored)", SSL_CA_LIST_OPT, @@ -4150,42 +4256,41 @@ ssl_setup_connection_params(rpc_transport_t *this) } priv->ssl_ca_list = gf_strdup(priv->ssl_ca_list); - if (dict_get_str(this->options, SSL_CRL_PATH_OPT, &optstr) == 0) { + optstr = NULL; + if (dict_get_str_sizen(this->options, SSL_CRL_PATH_OPT, &optstr) == 0) { if (!priv->ssl_enabled) { gf_log(this->name, GF_LOG_WARNING, "%s specified without %s (ignored)", SSL_CRL_PATH_OPT, SSL_ENABLED_OPT); } if (strcasecmp(optstr, "NULL") == 0) - crl_path = NULL; + priv->crl_path = NULL; else - crl_path = optstr; + priv->crl_path = gf_strdup(optstr); } - gf_log(this->name, priv->ssl_enabled ? GF_LOG_INFO : GF_LOG_DEBUG, - "SSL support on the I/O path is %s", - priv->ssl_enabled ? "ENABLED" : "NOT enabled"); - gf_log(this->name, priv->mgmt_ssl ? GF_LOG_INFO : GF_LOG_DEBUG, - "SSL support for glusterd is %s", - priv->mgmt_ssl ? "ENABLED" : "NOT enabled"); - if (!priv->mgmt_ssl) { - if (!dict_get_int32(this->options, SSL_CERT_DEPTH_OPT, &cert_depth)) { - gf_log(this->name, GF_LOG_INFO, "using certificate depth %d", - cert_depth); + if (!dict_get_int32_sizen(this->options, SSL_CERT_DEPTH_OPT, + &cert_depth)) { } } else { cert_depth = this->ctx->ssl_cert_depth; - gf_log(this->name, GF_LOG_INFO, "using certificate depth %d", - cert_depth); } - if (!dict_get_str(this->options, SSL_CIPHER_LIST_OPT, &cipher_list)) { + gf_log(this->name, priv->ssl_enabled ? GF_LOG_INFO : GF_LOG_DEBUG, + "SSL support for MGMT is %s IO path is %s certificate depth is %d " + "for peer %s", + (priv->mgmt_ssl ? "ENABLED" : "NOT enabled"), + (priv->ssl_enabled ? "ENABLED" : "NOT enabled"), cert_depth, + this->peerinfo.identifier); + + if (!dict_get_str_sizen(this->options, SSL_CIPHER_LIST_OPT, &cipher_list)) { gf_log(this->name, GF_LOG_INFO, "using cipher list %s", cipher_list); } - if (!dict_get_str(this->options, SSL_DH_PARAM_OPT, &dh_param)) { + if (!dict_get_str_sizen(this->options, SSL_DH_PARAM_OPT, &dh_param)) { + dh_flag = _gf_true; gf_log(this->name, GF_LOG_INFO, "using DH parameters %s", dh_param); } - if (!dict_get_str(this->options, SSL_EC_CURVE_OPT, &ec_curve)) { + if (!dict_get_str_sizen(this->options, SSL_EC_CURVE_OPT, &ec_curve)) { gf_log(this->name, GF_LOG_INFO, "using EC curve %s", ec_curve); } @@ -4217,12 +4322,15 @@ ssl_setup_connection_params(rpc_transport_t *this) #ifdef SSL_OP_NO_COMPRESSION SSL_CTX_set_options(priv->ssl_ctx, SSL_OP_NO_COMPRESSION); #endif - - if ((bio = BIO_new_file(dh_param, "r")) == NULL) { - gf_log(this->name, GF_LOG_INFO, - "failed to open %s, " - "DH ciphers are disabled", - dh_param); + /* Upload file to bio wrapper only if dh param is configured + */ + if (dh_flag) { + if ((bio = BIO_new_file(dh_param, "r")) == NULL) { + gf_log(this->name, GF_LOG_ERROR, + "failed to open %s, " + "DH ciphers are disabled", + dh_param); + } } if (bio != NULL) { @@ -4301,25 +4409,15 @@ ssl_setup_connection_params(rpc_transport_t *this) } if (!SSL_CTX_load_verify_locations(priv->ssl_ctx, priv->ssl_ca_list, - crl_path)) { + priv->crl_path)) { gf_log(this->name, GF_LOG_ERROR, "could not load CA list"); goto err; } SSL_CTX_set_verify_depth(priv->ssl_ctx, cert_depth); - if (crl_path) { -#ifdef X509_V_FLAG_CRL_CHECK_ALL - X509_STORE *x509store; - - x509store = SSL_CTX_get_cert_store(priv->ssl_ctx); - X509_STORE_set_flags( - x509store, X509_V_FLAG_CRL_CHECK | X509_V_FLAG_CRL_CHECK_ALL); -#else - gf_log(this->name, GF_LOG_ERROR, - "OpenSSL version does not support CRL"); -#endif - } + if (priv->crl_path) + ssl_set_crl_verify_flags(priv->ssl_ctx); priv->ssl_session_id = session_id++; SSL_CTX_set_session_id_context(priv->ssl_ctx, @@ -4349,25 +4447,19 @@ socket_init(rpc_transport_t *this) gf_boolean_t tmp_bool = 0; uint64_t windowsize = GF_DEFAULT_SOCKET_WINDOW_SIZE; char *optstr = NULL; - uint32_t timeout = 0; - int keepaliveidle = GF_KEEPALIVE_TIME; - int keepaliveintvl = GF_KEEPALIVE_INTERVAL; - int keepalivecnt = GF_KEEPALIVE_COUNT; - uint32_t backlog = 0; + data_t *data; if (this->private) { gf_log_callingfn(this->name, GF_LOG_ERROR, "double init attempted"); return -1; } - priv = GF_MALLOC(sizeof(*priv), gf_common_mt_socket_private_t); + priv = GF_CALLOC(1, sizeof(*priv), gf_common_mt_socket_private_t); if (!priv) { return -1; } - memset(priv, 0, sizeof(*priv)); this->private = priv; - pthread_mutex_init(&priv->in_lock, NULL); pthread_mutex_init(&priv->out_lock, NULL); pthread_mutex_init(&priv->cond_lock, NULL); pthread_cond_init(&priv->cond, NULL); @@ -4390,10 +4482,11 @@ socket_init(rpc_transport_t *this) if (!this->options) goto out; - if (dict_get(this->options, "non-blocking-io")) { - optstr = data_to_str(dict_get(this->options, "non-blocking-io")); + data = dict_get_sizen(this->options, "non-blocking-io"); + if (data) { + optstr = data_to_str(data); - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_ERROR, "'non-blocking-io' takes only boolean options," " not taking any action"); @@ -4409,11 +4502,11 @@ socket_init(rpc_transport_t *this) optstr = NULL; /* By default, we enable NODELAY */ - if (dict_get(this->options, "transport.socket.nodelay")) { - optstr = data_to_str( - dict_get(this->options, "transport.socket.nodelay")); + data = dict_get_sizen(this->options, "transport.socket.nodelay"); + if (data) { + optstr = data_to_str(data); - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_ERROR, "'transport.socket.nodelay' takes only " "boolean options, not taking any action"); @@ -4426,7 +4519,7 @@ socket_init(rpc_transport_t *this) } optstr = NULL; - if (dict_get_str(this->options, "tcp-window-size", &optstr) == 0) { + if (dict_get_str_sizen(this->options, "tcp-window-size", &optstr) == 0) { if (gf_string2uint64(optstr, &windowsize) != 0) { gf_log(this->name, GF_LOG_ERROR, "invalid number format: %s", optstr); @@ -4442,9 +4535,9 @@ socket_init(rpc_transport_t *this) priv->keepaliveintvl = GF_KEEPALIVE_INTERVAL; priv->keepaliveidle = GF_KEEPALIVE_TIME; priv->keepalivecnt = GF_KEEPALIVE_COUNT; - if (dict_get_str(this->options, "transport.socket.keepalive", &optstr) == - 0) { - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + if (dict_get_str_sizen(this->options, "transport.socket.keepalive", + &optstr) == 0) { + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_ERROR, "'transport.socket.keepalive' takes only " "boolean options, not taking any action"); @@ -4455,46 +4548,46 @@ socket_init(rpc_transport_t *this) priv->keepalive = 0; } - if (dict_get_int32(this->options, "transport.tcp-user-timeout", - &(priv->timeout)) != 0) - priv->timeout = timeout; - gf_log(this->name, GF_LOG_DEBUG, - "Configued " - "transport.tcp-user-timeout=%d", + if (dict_get_int32_sizen(this->options, "transport.tcp-user-timeout", + &(priv->timeout)) != 0) + priv->timeout = GF_NETWORK_TIMEOUT; + gf_log(this->name, GF_LOG_DEBUG, "Configured transport.tcp-user-timeout=%d", priv->timeout); - if (dict_get_int32(this->options, "transport.socket.keepalive-time", - &(priv->keepaliveidle)) != 0) { - priv->keepaliveidle = keepaliveidle; - } + if (priv->keepalive) { + if (dict_get_int32_sizen(this->options, + "transport.socket.keepalive-time", + &(priv->keepaliveidle)) != 0) { + priv->keepaliveidle = GF_KEEPALIVE_TIME; + } - if (dict_get_int32(this->options, "transport.socket.keepalive-interval", - &(priv->keepaliveintvl)) != 0) { - priv->keepaliveintvl = keepaliveintvl; - } + if (dict_get_int32_sizen(this->options, + "transport.socket.keepalive-interval", + &(priv->keepaliveintvl)) != 0) { + priv->keepaliveintvl = GF_KEEPALIVE_INTERVAL; + } - if (dict_get_int32(this->options, "transport.socket.keepalive-count", - &(priv->keepalivecnt)) != 0) - priv->keepalivecnt = keepalivecnt; - gf_log(this->name, GF_LOG_DEBUG, - "Reconfigued " - "transport.keepalivecnt=%d", - keepalivecnt); + if (dict_get_int32_sizen(this->options, + "transport.socket.keepalive-count", + &(priv->keepalivecnt)) != 0) + priv->keepalivecnt = GF_KEEPALIVE_COUNT; + gf_log(this->name, GF_LOG_DEBUG, + "Reconfigured transport.keepalivecnt=%d", priv->keepalivecnt); + } - if (dict_get_uint32(this->options, "transport.listen-backlog", &backlog) != - 0) { - backlog = GLUSTERFS_SOCKET_LISTEN_BACKLOG; + if (dict_get_uint32(this->options, "transport.listen-backlog", + &(priv->backlog)) != 0) { + priv->backlog = GLUSTERFS_SOCKET_LISTEN_BACKLOG; } - priv->backlog = backlog; optstr = NULL; /* Check if socket read failures are to be logged */ priv->read_fail_log = 1; - if (dict_get(this->options, "transport.socket.read-fail-log")) { - optstr = data_to_str( - dict_get(this->options, "transport.socket.read-fail-log")); - if (gf_string2boolean(optstr, &tmp_bool) == -1) { + data = dict_get_sizen(this->options, "transport.socket.read-fail-log"); + if (data) { + optstr = data_to_str(data); + if (gf_string2boolean(optstr, &tmp_bool) != 0) { gf_log(this->name, GF_LOG_WARNING, "'transport.socket.read-fail-log' takes only " "boolean options; logging socket read fails"); @@ -4506,7 +4599,7 @@ socket_init(rpc_transport_t *this) priv->windowsize = (int)windowsize; priv->ssl_enabled = _gf_false; - if (dict_get_str(this->options, SSL_ENABLED_OPT, &optstr) == 0) { + if (dict_get_str_sizen(this->options, SSL_ENABLED_OPT, &optstr) == 0) { if (gf_string2boolean(optstr, &priv->ssl_enabled) != 0) { gf_log(this->name, GF_LOG_ERROR, "invalid value given for ssl-enabled boolean"); @@ -4531,22 +4624,34 @@ fini(rpc_transport_t *this) priv = this->private; if (priv) { - if (priv->sock != -1) { - pthread_mutex_lock(&priv->in_lock); + if (priv->sock >= 0) { pthread_mutex_lock(&priv->out_lock); { - __socket_ioq_flush(this); + __socket_ioq_flush(priv); __socket_reset(this); } pthread_mutex_unlock(&priv->out_lock); - pthread_mutex_unlock(&priv->in_lock); } gf_log(this->name, GF_LOG_TRACE, "transport %p destroyed", this); - pthread_mutex_destroy(&priv->in_lock); pthread_mutex_destroy(&priv->out_lock); pthread_mutex_destroy(&priv->cond_lock); pthread_cond_destroy(&priv->cond); + + GF_ASSERT(priv->notify.in_progress == 0); + pthread_mutex_destroy(&priv->notify.lock); + pthread_cond_destroy(&priv->notify.cond); + + if (priv->use_ssl && priv->ssl_ssl) { + SSL_clear(priv->ssl_ssl); + SSL_free(priv->ssl_ssl); + priv->ssl_ssl = NULL; + } + if (priv->ssl_ctx) { + SSL_CTX_free(priv->ssl_ctx); + priv->ssl_ctx = NULL; + } + if (priv->ssl_private_key) { GF_FREE(priv->ssl_private_key); } @@ -4571,7 +4676,7 @@ init(rpc_transport_t *this) ret = socket_init(this); - if (ret == -1) { + if (ret < 0) { gf_log(this->name, GF_LOG_DEBUG, "socket_init() failed"); } @@ -4617,11 +4722,10 @@ struct volume_options options[] = { {.key = {"transport.tcp-user-timeout"}, .type = GF_OPTION_TYPE_INT, .op_version = {GD_OP_VERSION_3_10_2}, - .default_value = "0"}, + .default_value = TOSTRING(GF_NETWORK_TIMEOUT)}, {.key = {"transport.socket.nodelay"}, .type = GF_OPTION_TYPE_BOOL, .default_value = "1"}, - {.key = {"transport.socket.lowlat"}, .type = GF_OPTION_TYPE_BOOL}, {.key = {"transport.socket.keepalive"}, .type = GF_OPTION_TYPE_BOOL, .op_version = {1}, diff --git a/rpc/rpc-transport/socket/src/socket.h b/rpc/rpc-transport/socket/src/socket.h index 27db0ed427b..8a2eda70605 100644 --- a/rpc/rpc-transport/socket/src/socket.h +++ b/rpc/rpc-transport/socket/src/socket.h @@ -14,6 +14,7 @@ #include <openssl/ssl.h> #include <openssl/err.h> #include <openssl/x509v3.h> +#include <openssl/x509_vfy.h> #ifdef HAVE_OPENSSL_DH_H #include <openssl/dh.h> #endif @@ -21,13 +22,7 @@ #include <openssl/ecdh.h> #endif -#include "gf-event.h" #include "rpc-transport.h" -#include "logging.h" -#include "dict.h" -#include "mem-pool.h" -#include "globals.h" -#include "refcount.h" #ifndef MAX_IOVEC #define MAX_IOVEC 16 @@ -108,12 +103,13 @@ struct ioq { }; }; - uint32_t fraghdr; struct iovec vector[MAX_IOVEC]; - int count; struct iovec *pending_vector; + int count; int pending_count; struct iobref *iobref; + uint32_t fraghdr; + char _pad[4]; }; typedef struct { @@ -169,49 +165,30 @@ struct gf_sock_incoming_frag { #define GF_SOCKET_RA_MAX 1024 struct gf_sock_incoming { - sp_rpcrecord_state_t record_state; - struct gf_sock_incoming_frag frag; char *proghdr_base_addr; struct iobuf *iobuf; size_t iobuf_size; + struct gf_sock_incoming_frag frag; struct iovec vector[2]; - int count; struct iovec payload_vector; struct iobref *iobref; rpc_request_info_t *request_info; struct iovec *pending_vector; + int count; int pending_count; - uint32_t fraghdr; - char complete_record; - msg_type_t msg_type; size_t total_bytes_read; size_t ra_read; size_t ra_max; size_t ra_served; char *ra_buf; + uint32_t fraghdr; + msg_type_t msg_type; + sp_rpcrecord_state_t record_state; + char _pad[4]; }; -typedef enum { - OT_IDLE, /* Uninitialized or termination complete. */ - OT_SPAWNING, /* Past pthread_create but not in thread yet. */ - OT_RUNNING, /* Poller thread running normally. */ - OT_CALLBACK, /* Poller thread in the middle of a callback. */ - OT_PLEASE_DIE, /* Poller termination requested. */ -} ot_state_t; - typedef struct { - int32_t sock; - int32_t idx; - int32_t gen; - /* -1 = not connected. 0 = in progress. 1 = connected */ - char connected; - /* 1 = connect failed for reasons other than EINPROGRESS/ENOENT - see socket_connect for details */ - char connect_failed; - char bio; - char connect_finish_log; - char submit_log; union { struct list_head ioq; struct { @@ -219,37 +196,60 @@ typedef struct { struct ioq *ioq_prev; }; }; - struct gf_sock_incoming incoming; - pthread_mutex_t in_lock; pthread_mutex_t out_lock; pthread_mutex_t cond_lock; pthread_cond_t cond; int windowsize; - char lowlat; - char nodelay; int keepalive; int keepaliveidle; int keepaliveintvl; int keepalivecnt; int timeout; + int log_ctr; + int shutdown_log_ctr; + /* ssl_error_required is used only during the SSL connection setup + * phase. + * It holds the error code returned by SSL_get_error() and is used to + * arm the epoll event set for the required event for the specific fd. + */ + int ssl_error_required; + int ssl_session_id; + + GF_REF_DECL; /* refcount to keep track of socket_poller + threads */ + struct { + pthread_mutex_t lock; + pthread_cond_t cond; + uint64_t in_progress; + } notify; + int32_t sock; + int32_t idx; + int32_t gen; uint32_t backlog; - gf_boolean_t read_fail_log; - gf_boolean_t ssl_enabled; /* outbound I/O */ - gf_boolean_t mgmt_ssl; /* outbound mgmt */ - mgmt_ssl_t srvr_ssl; - gf_boolean_t use_ssl; SSL_METHOD *ssl_meth; SSL_CTX *ssl_ctx; - int ssl_session_id; BIO *ssl_sbio; SSL *ssl_ssl; char *ssl_own_cert; char *ssl_private_key; char *ssl_ca_list; - pthread_t thread; - int pipe[2]; + char *crl_path; + struct gf_sock_incoming incoming; + mgmt_ssl_t srvr_ssl; + /* -1 = not connected. 0 = in progress. 1 = connected */ + char connected; + /* 1 = connect failed for reasons other than EINPROGRESS/ENOENT + see socket_connect for details */ + char connect_failed; + char bio; + char connect_finish_log; + char submit_log; + char nodelay; + gf_boolean_t read_fail_log; + gf_boolean_t ssl_enabled; /* outbound I/O */ + gf_boolean_t mgmt_ssl; /* outbound mgmt */ gf_boolean_t is_server; - int log_ctr; + gf_boolean_t use_ssl; gf_boolean_t ssl_accepted; /* To indicate SSL_accept() */ gf_boolean_t ssl_connected; /* or SSL_connect() has been * been completed on this socket. @@ -270,21 +270,7 @@ typedef struct { * socket_event_handler() for * newly accepted socket */ - - /* ssl_error_required is used only during the SSL connection setup - * phase. - * It holds the error code returned by SSL_get_error() and is used to - * arm the epoll event set for the required event for the specific fd. - */ - int ssl_error_required; - - GF_REF_DECL; /* refcount to keep track of socket_poller - threads */ - struct { - pthread_mutex_t lock; - pthread_cond_t cond; - uint64_t in_progress; - } notify; + char _pad[4]; } socket_private_t; #endif |