summaryrefslogtreecommitdiffstats
path: root/rpc/rpc-transport
diff options
context:
space:
mode:
Diffstat (limited to 'rpc/rpc-transport')
-rw-r--r--rpc/rpc-transport/rdma/src/Makefile.am10
-rw-r--r--rpc/rpc-transport/rdma/src/name.c80
-rw-r--r--rpc/rpc-transport/rdma/src/name.h9
-rw-r--r--rpc/rpc-transport/rdma/src/rdma.c3145
-rw-r--r--rpc/rpc-transport/rdma/src/rdma.h105
-rw-r--r--rpc/rpc-transport/socket/src/Makefile.am10
-rw-r--r--rpc/rpc-transport/socket/src/name.c56
-rw-r--r--rpc/rpc-transport/socket/src/socket.c1879
-rw-r--r--rpc/rpc-transport/socket/src/socket.h109
9 files changed, 3005 insertions, 2398 deletions
diff --git a/rpc/rpc-transport/rdma/src/Makefile.am b/rpc/rpc-transport/rdma/src/Makefile.am
index b4b940bca..2bf7cf238 100644
--- a/rpc/rpc-transport/rdma/src/Makefile.am
+++ b/rpc/rpc-transport/rdma/src/Makefile.am
@@ -3,18 +3,20 @@
transport_LTLIBRARIES = rdma.la
transportdir = $(libdir)/glusterfs/$(PACKAGE_VERSION)/rpc-transport
-rdma_la_LDFLAGS = -module -avoidversion
+rdma_la_LDFLAGS = -module -avoid-version
rdma_la_SOURCES = rdma.c name.c
rdma_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la \
- -libverbs
+ -libverbs -lrdmacm
noinst_HEADERS = rdma.h name.h
-I$(top_srcdir)/libglusterfs/src -I$(top_srcdir)/rpc/rpc-lib/src/ \
-I$(top_srcdir)/xlators/protocol/lib/src/ -shared -nostartfiles $(GF_CFLAGS)
-AM_CFLAGS = -fPIC -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -Wall -D$(GF_HOST_OS) \
+AM_CPPFLAGS = $(GF_CPPFLAGS) \
-I$(top_srcdir)/libglusterfs/src -I$(top_srcdir)/rpc/rpc-lib/src/ \
- -I$(top_srcdir)/rpc/xdr/src -shared -nostartfiles $(GF_CFLAGS)
+ -I$(top_srcdir)/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
index e5f7ba2ca..c57428ad6 100644
--- a/rpc/rpc-transport/rdma/src/name.c
+++ b/rpc/rpc-transport/rdma/src/name.c
@@ -13,12 +13,7 @@
#include <errno.h>
#include <netdb.h>
#include <string.h>
-
-#ifdef CLIENT_PORT_CEILING
-#undef CLIENT_PORT_CEILING
-#endif
-
-#define CLIENT_PORT_CEILING 1024
+#include <rdma/rdma_cma.h>
#ifndef AF_INET_SDP
#define AF_INET_SDP 27
@@ -26,6 +21,8 @@
#include "rpc-transport.h"
#include "rdma.h"
+#include "common-utils.h"
+
int32_t
gf_resolve_ip6 (const char *hostname,
@@ -35,28 +32,43 @@ gf_resolve_ip6 (const char *hostname,
struct addrinfo **addr_info);
static int32_t
-af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
+af_inet_bind_to_port_lt_ceiling (struct rdma_cm_id *cm_id,
+ struct sockaddr *sockaddr,
socklen_t sockaddr_len, int ceiling)
{
- int32_t ret = -1;
- /* struct sockaddr_in sin = {0, }; */
- uint16_t port = ceiling - 1;
+ int32_t ret = -1;
+ uint16_t port = ceiling - 1;
+ // by default assume none of the ports are blocked and all are available
+ gf_boolean_t ports[1024] = {_gf_false,};
+ int i = 0;
+
+ ret = gf_process_reserved_ports (ports);
+ if (ret != 0) {
+ for (i = 0; i < 1024; i++)
+ ports[i] = _gf_false;
+ }
while (port)
{
switch (sockaddr->sa_family)
{
case AF_INET6:
- ((struct sockaddr_in6 *)sockaddr)->sin6_port = htons (port);
+ ((struct sockaddr_in6 *)sockaddr)->sin6_port
+ = htons (port);
break;
case AF_INET_SDP:
case AF_INET:
- ((struct sockaddr_in *)sockaddr)->sin_port = htons (port);
+ ((struct sockaddr_in *)sockaddr)->sin_port
+ = htons (port);
break;
}
-
- ret = bind (fd, sockaddr, sockaddr_len);
+ // ignore the reserved ports
+ if (ports[port] == _gf_true) {
+ port--;
+ continue;
+ }
+ ret = rdma_bind_addr (cm_id, sockaddr);
if (ret == 0)
break;
@@ -70,11 +82,10 @@ af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
return ret;
}
+#if 0
static int32_t
-af_unix_client_bind (rpc_transport_t *this,
- struct sockaddr *sockaddr,
- socklen_t sockaddr_len,
- int sock)
+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;
@@ -106,6 +117,7 @@ af_unix_client_bind (rpc_transport_t *this,
err:
return ret;
}
+#endif
static int32_t
client_fill_address_family (rpc_transport_t *this, struct sockaddr *sockaddr)
@@ -153,13 +165,11 @@ client_fill_address_family (rpc_transport_t *this, struct sockaddr *sockaddr)
sockaddr->sa_family = AF_INET6;
} else if (!strcasecmp (address_family, "inet-sdp")) {
sockaddr->sa_family = AF_INET_SDP;
- } else if (!strcasecmp (address_family, "inet/inet6")
- || !strcasecmp (address_family, "inet6/inet")) {
- sockaddr->sa_family = AF_UNSPEC;
} else {
gf_log (this->name, GF_LOG_ERROR,
"unknown address-family (%s) specified",
address_family);
+ sockaddr->sa_family = AF_UNSPEC;
return -1;
}
}
@@ -354,6 +364,8 @@ af_inet_server_get_local_sockaddr (rpc_transport_t *this,
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;
@@ -404,10 +416,8 @@ out:
}
int32_t
-gf_rdma_client_bind (rpc_transport_t *this,
- struct sockaddr *sockaddr,
- socklen_t *sockaddr_len,
- int sock)
+gf_rdma_client_bind (rpc_transport_t *this, struct sockaddr *sockaddr,
+ socklen_t *sockaddr_len, struct rdma_cm_id *cm_id)
{
int ret = 0;
@@ -419,22 +429,24 @@ gf_rdma_client_bind (rpc_transport_t *this,
*sockaddr_len = sizeof (struct sockaddr_in);
case AF_INET6:
- ret = af_inet_bind_to_port_lt_ceiling (sock, sockaddr,
+ ret = af_inet_bind_to_port_lt_ceiling (cm_id, sockaddr,
*sockaddr_len,
- CLIENT_PORT_CEILING);
+ GF_CLIENT_PORT_CEILING);
if (ret == -1) {
gf_log (this->name, GF_LOG_WARNING,
- "cannot bind inet socket (%d) to port "
- "less than %d (%s)",
- sock, CLIENT_PORT_CEILING, strerror (errno));
+ "cannot bind rdma_cm_id to port "
+ "less than %d (%s)", GF_CLIENT_PORT_CEILING,
+ strerror (errno));
ret = 0;
}
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:
@@ -521,21 +533,19 @@ gf_rdma_server_get_local_sockaddr (rpc_transport_t *this,
addr->sa_family = AF_INET_SDP;
} else if (!strcasecmp (address_family, "unix")) {
addr->sa_family = AF_UNIX;
- } else if (!strcasecmp (address_family, "inet/inet6")
- || !strcasecmp (address_family, "inet6/inet")) {
- addr->sa_family = AF_UNSPEC;
} else {
gf_log (this->name, GF_LOG_ERROR,
"unknown address family (%s) specified",
address_family);
+ addr->sa_family = AF_UNSPEC;
ret = -1;
goto err;
}
} else {
gf_log (this->name, GF_LOG_DEBUG,
"option address-family not specified, defaulting "
- "to inet/inet6");
- addr->sa_family = AF_UNSPEC;
+ "to inet");
+ addr->sa_family = AF_INET;
}
switch (addr->sa_family)
diff --git a/rpc/rpc-transport/rdma/src/name.h b/rpc/rpc-transport/rdma/src/name.h
index 114ed1661..742fc5fc3 100644
--- a/rpc/rpc-transport/rdma/src/name.h
+++ b/rpc/rpc-transport/rdma/src/name.h
@@ -11,16 +11,13 @@
#ifndef _IB_VERBS_NAME_H
#define _IB_VERBS_NAME_H
-#include <sys/socket.h>
-#include <sys/un.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,
- int sock);
+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,
diff --git a/rpc/rpc-transport/rdma/src/rdma.c b/rpc/rpc-transport/rdma/src/rdma.c
index 8e9472fc6..6e6099a98 100644
--- a/rpc/rpc-transport/rdma/src/rdma.c
+++ b/rpc/rpc-transport/rdma/src/rdma.c
@@ -8,7 +8,6 @@
cases as published by the Free Software Foundation.
*/
-
#ifndef _CONFIG_H
#define _CONFIG_H
#include "config.h"
@@ -21,6 +20,7 @@
#include "name.h"
#include "byte-order.h"
#include "xlator.h"
+#include "xdr-rpc.h"
#include <signal.h>
#define GF_RDMA_LOG_NAME "rpc-transport/rdma"
@@ -34,99 +34,29 @@ gf_rdma_post_ref (gf_rdma_post_t *post);
int
gf_rdma_post_unref (gf_rdma_post_t *post);
-int32_t
-gf_resolve_ip6 (const char *hostname,
- uint16_t port,
- int family,
- void **dnscache,
- struct addrinfo **addr_info);
-
-static uint16_t
-gf_rdma_get_local_lid (struct ibv_context *context,
- int32_t port)
-{
- struct ibv_port_attr attr;
+static void *
+gf_rdma_send_completion_proc (void *data);
- if (ibv_query_port (context, port, &attr))
- return 0;
+static void *
+gf_rdma_recv_completion_proc (void *data);
- return attr.lid;
-}
-
-static const char *
-get_port_state_str(enum ibv_port_state pstate)
-{
- switch (pstate) {
- case IBV_PORT_DOWN: return "PORT_DOWN";
- case IBV_PORT_INIT: return "PORT_INIT";
- case IBV_PORT_ARMED: return "PORT_ARMED";
- case IBV_PORT_ACTIVE: return "PORT_ACTIVE";
- case IBV_PORT_ACTIVE_DEFER: return "PORT_ACTIVE_DEFER";
- default: return "invalid state";
- }
-}
+void *
+gf_rdma_async_event_thread (void *context);
static int32_t
-ib_check_active_port (struct ibv_context *ctx, uint8_t port)
-{
- struct ibv_port_attr port_attr = {0, };
- int32_t ret = 0;
- const char *state_str = NULL;
-
- if (!ctx) {
- gf_log_callingfn (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "Error in supplied context");
- return -1;
- }
-
- ret = ibv_query_port (ctx, port, &port_attr);
-
- if (ret) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "Failed to query port %u properties", port);
- return -1;
- }
-
- state_str = get_port_state_str (port_attr.state);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_TRACE,
- "Infiniband PORT: (%u) STATE: (%s)",
- port, state_str);
-
- if (port_attr.state == IBV_PORT_ACTIVE)
- return 0;
-
- return -1;
-}
+gf_rdma_create_qp (rpc_transport_t *this);
static int32_t
-ib_get_active_port (struct ibv_context *ib_ctx)
-{
- struct ibv_device_attr ib_device_attr = {{0, }, };
- int32_t ret = -1;
- uint8_t ib_port = 0;
+__gf_rdma_teardown (rpc_transport_t *this);
- if (!ib_ctx) {
- gf_log_callingfn (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "Error in supplied context");
- return -1;
- }
- if (ibv_query_device (ib_ctx, &ib_device_attr)) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "Failed to query device properties");
- return -1;
- }
+static int32_t
+gf_rdma_teardown (rpc_transport_t *this);
- for (ib_port = 1; ib_port <= ib_device_attr.phys_port_cnt; ++ib_port) {
- ret = ib_check_active_port (ib_ctx, ib_port);
- if (ret == 0)
- return ib_port;
+static int32_t
+gf_rdma_disconnect (rpc_transport_t *this);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_TRACE,
- "Port:(%u) not active", ib_port);
- continue;
- }
- return ret;
-}
+static void
+gf_rdma_cm_handle_disconnect (rpc_transport_t *this);
static void
@@ -156,7 +86,7 @@ gf_rdma_put_post (gf_rdma_queue_t *queue, gf_rdma_post_t *post)
static gf_rdma_post_t *
-gf_rdma_new_post (gf_rdma_device_t *device, int32_t len,
+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;
@@ -183,8 +113,9 @@ gf_rdma_new_post (gf_rdma_device_t *device, int32_t len,
post->buf_size,
IBV_ACCESS_LOCAL_WRITE);
if (!post->mr) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "memory registration failed");
+ gf_log (this->name, GF_LOG_WARNING,
+ "memory registration failed (%s)",
+ strerror (errno));
goto out;
}
@@ -194,9 +125,7 @@ gf_rdma_new_post (gf_rdma_device_t *device, int32_t len,
ret = 0;
out:
if (ret != 0) {
- if (post->buf != NULL) {
- free (post->buf);
- }
+ free (post->buf);
GF_FREE (post);
post = NULL;
@@ -259,22 +188,6 @@ __gf_rdma_quota_get (gf_rdma_peer_t *peer)
return ret;
}
-/*
- static int32_t
- gf_rdma_quota_get (gf_rdma_peer_t *peer)
- {
- int32_t ret = -1;
- gf_rdma_private_t *priv = peer->trans->private;
-
- pthread_mutex_lock (&priv->write_mutex);
- {
- ret = __gf_rdma_quota_get (peer);
- }
- pthread_mutex_unlock (&priv->write_mutex);
-
- return ret;
- }
-*/
static void
__gf_rdma_ioq_entry_free (gf_rdma_ioq_t *entry)
@@ -290,6 +203,7 @@ __gf_rdma_ioq_entry_free (gf_rdma_ioq_t *entry)
iobref_unref (entry->msg.request.rsp_iobref);
entry->msg.request.rsp_iobref = NULL;
}
+
mem_put (entry);
}
@@ -309,26 +223,898 @@ static int32_t
__gf_rdma_disconnect (rpc_transport_t *this)
{
gf_rdma_private_t *priv = NULL;
- int32_t ret = 0;
priv = this->private;
- if (priv->connected || priv->tcp_connected) {
- fcntl (priv->sock, F_SETFL, O_NONBLOCK);
- if (shutdown (priv->sock, SHUT_RDWR) != 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "shutdown () - error: %s",
- strerror (errno));
- ret = -errno;
- priv->tcp_connected = 0;
+ 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_log (this->name, GF_LOG_ERROR,
+ "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 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_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "ibv_query_device on %s returned %d (%s)",
+ priv->device->device_name, ret,
+ (ret > 0) ? strerror (ret) : "");
+ ret = -1;
+ goto out;
+ }
+
+ send_cqe = 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_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "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;
+
+ priv = this->private;
+ options = &priv->options;
+ ctx = this->ctx;
+ rdma_ctx = ctx->ib;
+
+ 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->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->next = rdma_ctx->device;
+ rdma_ctx->device = trav;
+
+ trav->send_chan = ibv_create_comp_channel (trav->context);
+ if (!trav->send_chan) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "could not create CQ for device (%s)",
+ device_name);
+ goto out;
+ }
+
+ /* protection domain */
+ trav->pd = ibv_alloc_pd (trav->context);
+
+ if (!trav->pd) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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_log (this->name, GF_LOG_ERROR,
+ "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);
+
+ if (gf_rdma_create_posts (this) < 0) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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);
+ if (ret) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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);
+ if (ret) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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);
+ if (ret) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "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) {
+ 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);
+ 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->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_log (listener->name, GF_LOG_WARNING,
+ "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_log (listener->name, GF_LOG_WARNING,
+ "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_log (listener->name, GF_LOG_WARNING,
+ "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_log (listener->name, GF_LOG_TRACE,
+ "got a connect request (me:%s peer:%s)",
+ listener->myinfo.identifier, this->peerinfo.identifier);
+
+ ret = gf_rdma_create_qp (this);
+ if (ret < 0) {
+ gf_log (listener->name, GF_LOG_WARNING,
+ "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_log (listener->name, GF_LOG_WARNING, "rdma_accept failed "
+ "peer:%s me:%s (%s)", this->peerinfo.identifier,
+ this->myinfo.identifier, strerror (errno));
+ gf_rdma_cm_handle_disconnect (this);
+ goto out;
+ }
+
+ 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_log (this->name, GF_LOG_WARNING,
+ "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_log (this->name, GF_LOG_WARNING,
+ "rdma_connect failed (%s)", strerror (errno));
+ gf_rdma_cm_handle_disconnect (this);
+ goto out;
+ }
+
+ gf_log (this->name, GF_LOG_TRACE, "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_log (this->name, GF_LOG_WARNING,
+ "rdma_resolve_route failed (me:%s peer:%s) (%s)",
+ this->myinfo.identifier, this->peerinfo.identifier,
+ strerror (errno));
+ gf_rdma_cm_handle_disconnect (this);
+ }
+
+ gf_log (this->name, GF_LOG_TRACE, "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, connected = 0;
+
+ priv = this->private;
+ gf_log (this->name, GF_LOG_DEBUG,
+ "peer disconnected, cleaning up");
+
+ pthread_mutex_lock (&priv->write_mutex);
+ {
+ if (priv->peer.cm_id != NULL) {
+ need_unref = 1;
+ connected = priv->connected;
priv->connected = 0;
}
+
+ __gf_rdma_teardown (this);
}
+ pthread_mutex_unlock (&priv->write_mutex);
+
+ if (connected) {
+ rpc_transport_notify (this, RPC_TRANSPORT_DISCONNECT, this);
+ }
+
+ if (need_unref)
+ rpc_transport_unref (this);
+
+}
+
+
+static int
+gf_rdma_cm_handle_event_established (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;
+
+ 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) {
+ 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_log (this->name, GF_LOG_TRACE,
+ "recieved event RDMA_CM_EVENT_ESTABLISHED (me:%s peer:%s)",
+ this->myinfo.identifier, this->peerinfo.identifier);
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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma_cm_get_event failed (%s)",
+ strerror (errno));
+ 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_event_established (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_log (this->name, GF_LOG_WARNING,
+ "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_log (this->name, GF_LOG_DEBUG,
+ "recieved 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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "device removed");
+ gf_rdma_cm_handle_device_removal (event);
+ break;
+
+ default:
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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)
{
@@ -354,33 +1140,33 @@ gf_rdma_post_send (struct ibv_qp *qp, gf_rdma_post_t *post, int32_t len)
int
__gf_rdma_encode_error(gf_rdma_peer_t *peer, gf_rdma_reply_info_t *reply_info,
- struct iovec *rpchdr, uint32_t *ptr,
+ struct iovec *rpchdr, gf_rdma_header_t *hdr,
gf_rdma_errcode_t err)
{
- uint32_t *startp = NULL;
struct rpc_msg *rpc_msg = NULL;
- startp = ptr;
if (reply_info != NULL) {
- *ptr++ = hton32(reply_info->rm_xid);
+ hdr->rm_xid = hton32(reply_info->rm_xid);
} else {
rpc_msg = rpchdr[0].iov_base; /* assume rpchdr contains
* only one vector.
* (which is true)
*/
- *ptr++ = rpc_msg->rm_xid;
+ hdr->rm_xid = rpc_msg->rm_xid;
}
- *ptr++ = hton32(GF_RDMA_VERSION);
- *ptr++ = hton32(peer->send_count);
- *ptr++ = hton32(GF_RDMA_ERROR);
- *ptr++ = hton32(err);
+ 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) {
- *ptr++ = hton32(GF_RDMA_VERSION);
- *ptr++ = hton32(GF_RDMA_VERSION);
+ 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 (int)((unsigned long)ptr - (unsigned long)startp);
+ return sizeof (*hdr);
}
@@ -392,7 +1178,7 @@ __gf_rdma_send_error (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
int32_t ret = -1, len = 0;
len = __gf_rdma_encode_error (peer, reply_info, entry->rpchdr,
- (uint32_t *)post->buf, err);
+ (gf_rdma_header_t *)post->buf, err);
if (len == -1) {
gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
"encode error returned -1");
@@ -451,8 +1237,10 @@ __gf_rdma_create_read_chunks_from_vector (gf_rdma_peer_t *peer,
vector[i].iov_len,
IBV_ACCESS_REMOTE_READ);
if (!mr) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "memory registration failed");
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "memory registration failed (%s) (peer:%s)",
+ strerror (errno),
+ peer->trans->peerinfo.identifier);
goto out;
}
@@ -501,8 +1289,8 @@ __gf_rdma_create_read_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->rpchdr_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "cannot create read chunks from vector, "
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot create read chunks from vector "
"entry->rpchdr");
goto out;
}
@@ -514,8 +1302,8 @@ __gf_rdma_create_read_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->proghdr_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "cannot create read chunks from vector, "
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot create read chunks from vector "
"entry->proghdr");
}
@@ -527,8 +1315,8 @@ __gf_rdma_create_read_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->prog_payload_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "cannot create read chunks from vector,"
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot create read chunks from vector"
" entry->prog_payload");
}
}
@@ -541,8 +1329,8 @@ __gf_rdma_create_read_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->prog_payload_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "cannot create read chunks from vector, "
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot create read chunks from vector "
"entry->prog_payload");
}
}
@@ -585,8 +1373,10 @@ __gf_rdma_create_write_chunks_from_vector (gf_rdma_peer_t *peer,
IBV_ACCESS_REMOTE_WRITE
| IBV_ACCESS_LOCAL_WRITE);
if (!mr) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "memory registration failed");
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "memory registration failed (%s) (peer:%s)",
+ strerror (errno),
+ peer->trans->peerinfo.identifier);
goto out;
}
@@ -661,7 +1451,7 @@ __gf_rdma_create_write_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->msg.request.rsp_payload_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"cannot create write chunks from vector "
"entry->rpc_payload");
goto out;
@@ -691,7 +1481,7 @@ __gf_rdma_create_write_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->msg.request.rsphdr_count,
request_ctx);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"cannot create write chunks from vector "
"entry->rpchdr");
goto out;
@@ -707,7 +1497,7 @@ out:
}
-inline void
+static inline void
__gf_rdma_deregister_mr (struct ibv_mr **mr, int count)
{
int i = 0;
@@ -828,28 +1618,6 @@ out:
}
-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);
-}
-
-
int
gf_rdma_post_unref (gf_rdma_post_t *post)
{
@@ -1035,10 +1803,11 @@ __gf_rdma_ioq_churn_request (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
chunkptr = &hdr->rm_body.rm_chunks[0];
if (rtype != gf_rdma_noch) {
- ret = __gf_rdma_create_read_chunks (peer, entry, rtype, &chunkptr,
+ ret = __gf_rdma_create_read_chunks (peer, entry, rtype,
+ &chunkptr,
request_ctx);
if (ret != 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"creation of read chunks failed");
goto out;
}
@@ -1047,10 +1816,11 @@ __gf_rdma_ioq_churn_request (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
}
if (wtype != gf_rdma_noch) {
- ret = __gf_rdma_create_write_chunks (peer, entry, wtype, &chunkptr,
+ ret = __gf_rdma_create_write_chunks (peer, entry, wtype,
+ &chunkptr,
request_ctx);
if (ret != 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"creation of write/reply chunk failed");
goto out;
}
@@ -1108,7 +1878,7 @@ out:
}
-inline void
+static inline void
__gf_rdma_fill_reply_header (gf_rdma_header_t *header, struct iovec *rpchdr,
gf_rdma_reply_info_t *reply_info, int credits)
{
@@ -1138,11 +1908,12 @@ __gf_rdma_fill_reply_header (gf_rdma_header_t *header, struct iovec *rpchdr,
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_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;
+ 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)
@@ -1155,6 +1926,10 @@ __gf_rdma_send_reply_inline (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) {
ret = __gf_rdma_send_error (peer, entry, post, reply_info,
ERR_CHUNK);
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "msg size (%d) is greater than maximum size "
+ "of msg that can be sent inlined (%d)",
+ send_size, GLUSTERFS_RDMA_INLINE_THRESHOLD);
goto out;
}
@@ -1189,7 +1964,7 @@ __gf_rdma_send_reply_inline (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
ret = send_size;
} else {
gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
- "gf_rdma_post_send (to %s) failed with ret = %d (%s)",
+ "posting send (to %s) failed with ret = %d (%s)",
peer->trans->peerinfo.identifier, ret,
(ret > 0) ? strerror (ret) : "");
gf_rdma_post_unref (post);
@@ -1262,9 +2037,8 @@ __gf_rdma_register_local_mr_for_rdma (gf_rdma_peer_t *peer,
gf_rdma_private_t *priv = NULL;
gf_rdma_device_t *device = NULL;
- if ((ctx == NULL) || (vector == NULL)) {
- goto out;
- }
+ 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;
@@ -1285,6 +2059,9 @@ __gf_rdma_register_local_mr_for_rdma (gf_rdma_peer_t *peer,
vector[i].iov_len,
IBV_ACCESS_LOCAL_WRITE);
if (ctx->mr[ctx->mr_count] == NULL) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "registering memory for IBV_ACCESS_LOCAL_WRITE "
+ "failed (%s)", strerror (errno));
goto out;
}
@@ -1355,7 +2132,8 @@ __gf_rdma_write (gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *vec,
ret = ibv_post_send(peer->qp, &wr, &bad_wr);
if (ret) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "rdma write to "
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma write to "
"client (%s) failed with ret = %d (%s)",
peer->trans->peerinfo.identifier, ret,
(ret > 0) ? strerror (ret) : "");
@@ -1390,6 +2168,8 @@ __gf_rdma_do_gf_rdma_write (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
ret = __gf_rdma_register_local_mr_for_rdma (peer, vector, count,
&post->ctx);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "registering memory region for rdma failed");
goto out;
}
@@ -1401,9 +2181,13 @@ __gf_rdma_do_gf_rdma_write (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
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,
+ ret = __gf_rdma_write (peer, post, vector, xfer_len,
+ &payload_idx,
&reply_info->wc_array->wc_array[i]);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma write to client (%s) failed",
+ peer->trans->peerinfo.identifier);
goto out;
}
@@ -1422,12 +2206,12 @@ __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];
+ 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;
@@ -1442,9 +2226,10 @@ __gf_rdma_send_reply_type_nomsg (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
/* 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);
+ reply_info,
+ (uint32_t **)&buf);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"encoding write chunks failed");
ret = __gf_rdma_send_error (peer, entry, post, reply_info,
ERR_CHUNK);
@@ -1464,6 +2249,9 @@ __gf_rdma_send_reply_type_nomsg (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
ret = __gf_rdma_do_gf_rdma_write (peer, post, vector, count,
entry->iobref, reply_info);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma write to peer (%s) failed",
+ peer->trans->peerinfo.identifier);
gf_rdma_post_unref (post);
goto out;
}
@@ -1471,7 +2259,7 @@ __gf_rdma_send_reply_type_nomsg (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
ret = gf_rdma_post_send (peer->qp, post, (buf - post->buf));
if (ret) {
gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
- "gf_rdma_post_send to client (%s) failed with "
+ "posting a send request to client (%s) failed with "
"ret = %d (%s)", peer->trans->peerinfo.identifier, ret,
(ret > 0) ? strerror (ret) : "");
ret = -1;
@@ -1487,12 +2275,13 @@ out:
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_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;
+ 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)
@@ -1524,7 +2313,7 @@ __gf_rdma_send_reply_type_msg (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
reply_info,
(uint32_t **)&ptr);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"encoding write chunks failed");
ret = __gf_rdma_send_error (peer, entry, post, reply_info,
ERR_CHUNK);
@@ -1540,6 +2329,8 @@ __gf_rdma_send_reply_type_msg (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
entry->prog_payload_count,
entry->iobref, reply_info);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "rdma write to peer "
+ "(%s) failed", peer->trans->peerinfo.identifier);
gf_rdma_post_unref (post);
goto out;
}
@@ -1614,9 +2405,9 @@ __gf_rdma_ioq_churn_reply (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
int32_t ret = -1;
gf_rdma_chunktype_t type = gf_rdma_noch;
- if ((peer == NULL) || (entry == NULL) || (post == NULL)) {
- goto out;
- }
+ 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) {
@@ -1627,22 +2418,39 @@ __gf_rdma_ioq_churn_reply (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry,
case gf_rdma_noch:
ret = __gf_rdma_send_reply_inline (peer, entry, post,
reply_info);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "failed to send reply with write chunks "
+ "to peer (%s)",
+ peer->trans->peerinfo.identifier);
+ }
break;
default:
gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
- "invalid chunktype (%d) specified for sending reply",
- type);
+ "invalid chunktype (%d) specified for sending reply "
+ " (peer:%s)", type, peer->trans->peerinfo.identifier);
break;
}
@@ -1671,20 +2479,34 @@ __gf_rdma_ioq_churn_entry (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry)
if (quota > 0) {
post = gf_rdma_get_post (&device->sendq);
if (post == NULL) {
- post = gf_rdma_new_post (device,
+ post = gf_rdma_new_post (peer->trans, device,
(options->send_size + 2048),
GF_RDMA_SEND_POST);
}
if (post == NULL) {
ret = -1;
+ gf_log_callingfn (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "failed to process reply ioq entry "
+ "to peer (%s)",
+ peer->trans->peerinfo.identifier);
+ }
}
if (ret != 0) {
@@ -1740,9 +2562,9 @@ gf_rdma_writev (rpc_transport_t *this, gf_rdma_ioq_t *entry)
pthread_mutex_lock (&priv->write_mutex);
{
if (!priv->connected) {
- gf_log (this->name, GF_LOG_DEBUG,
- "rdma is not connected to post a "
- "send request");
+ gf_log (this->name, GF_LOG_WARNING,
+ "rdma is not connected to peer (%s)",
+ this->peerinfo.identifier);
ret = -1;
goto unlock;
}
@@ -1752,6 +2574,13 @@ gf_rdma_writev (rpc_transport_t *this, gf_rdma_ioq_t *entry)
ret = __gf_rdma_ioq_churn_entry (peer, entry);
if (ret != 0) {
need_append = 0;
+
+ if (ret < 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "processing ioq entry destined "
+ "to (%s) failed",
+ this->peerinfo.identifier);
+ }
}
}
@@ -1871,6 +2700,9 @@ gf_rdma_submit_request (rpc_transport_t *this, rpc_transport_req_t *req)
entry = gf_rdma_ioq_new (this, &data);
if (entry == NULL) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "getting a new ioq entry failed (peer:%s)",
+ this->peerinfo.identifier);
goto out;
}
@@ -1879,6 +2711,9 @@ gf_rdma_submit_request (rpc_transport_t *this, rpc_transport_req_t *req)
if (ret > 0) {
ret = 0;
} else if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "sending request to peer (%s) failed",
+ this->peerinfo.identifier);
rpc_transport_disconnect (this);
}
@@ -1901,6 +2736,9 @@ gf_rdma_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply)
entry = gf_rdma_ioq_new (this, &data);
if (entry == NULL) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "getting a new ioq entry failed (peer:%s)",
+ this->peerinfo.identifier);
goto out;
}
@@ -1908,6 +2746,9 @@ gf_rdma_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply)
if (ret > 0) {
ret = 0;
} else if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "sending request to peer (%s) failed",
+ this->peerinfo.identifier);
rpc_transport_disconnect (this);
}
@@ -1915,189 +2756,6 @@ out:
return ret;
}
-#if 0
-static int
-gf_rdma_receive (rpc_transport_t *this, char **hdr_p, size_t *hdrlen_p,
- struct iobuf **iobuf_p)
-{
- gf_rdma_private_t *priv = this->private;
- /* TODO: return error if !priv->connected, check with locks */
- /* TODO: boundry checks for data_ptr/offset */
- char *copy_from = NULL;
- gf_rdma_header_t *header = NULL;
- uint32_t size1, size2, data_len = 0;
- char *hdr = NULL;
- struct iobuf *iobuf = NULL;
- int32_t ret = 0;
-
- pthread_mutex_lock (&priv->recv_mutex);
- {
-/*
- while (!priv->data_ptr)
- pthread_cond_wait (&priv->recv_cond, &priv->recv_mutex);
-*/
-
- copy_from = priv->data_ptr + priv->data_offset;
-
- priv->data_ptr = NULL;
- data_len = priv->data_len;
- pthread_cond_broadcast (&priv->recv_cond);
- }
- pthread_mutex_unlock (&priv->recv_mutex);
-
- header = (gf_rdma_header_t *)copy_from;
- if (strcmp (header->colonO, ":O")) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "%s: corrupt header received", this->name);
- ret = -1;
- goto err;
- }
-
- size1 = ntoh32 (header->size1);
- size2 = ntoh32 (header->size2);
-
- if (data_len != (size1 + size2 + sizeof (*header))) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "%s: sizeof data read from transport is not equal "
- "to the size specified in the header",
- this->name);
- ret = -1;
- goto err;
- }
-
- copy_from += sizeof (*header);
-
- if (size1) {
- hdr = GF_CALLOC (1, size1, gf_common_mt_char);
- if (!hdr) {
- gf_log (this->name, GF_LOG_ERROR,
- "unable to allocate header for peer %s",
- this->peerinfo.identifier);
- ret = -ENOMEM;
- goto err;
- }
- memcpy (hdr, copy_from, size1);
- copy_from += size1;
- *hdr_p = hdr;
- }
- *hdrlen_p = size1;
-
- if (size2) {
- iobuf = iobuf_get2 (this->ctx->iobuf_pool, size2);
- if (!iobuf) {
- gf_log (this->name, GF_LOG_ERROR,
- "unable to allocate IO buffer for peer %s",
- this->peerinfo.identifier);
- ret = -ENOMEM;
- goto err;
- }
- memcpy (iobuf->ptr, copy_from, size2);
- *iobuf_p = iobuf;
- }
-
-err:
- return ret;
-}
-#endif
-
-
-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_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: creation of CQ for device %s failed",
- this->name, device->device_name);
- ret = -1;
- goto out;
- } else if (ibv_req_notify_cq (device->recv_cq, 0)) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: ibv_req_notify_cq on recv CQ of device %s failed",
- this->name, device->device_name);
- ret = -1;
- goto out;
- }
-
- do {
- ret = ibv_query_device (priv->device->context, &device_attr);
- if (ret != 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: ibv_query_device on %s returned %d (%s)",
- this->name, priv->device->device_name, ret,
- (ret > 0) ? strerror (ret) : "");
- ret = -1;
- goto out;
- }
-
- send_cqe = 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_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: creation of send_cq for device %s failed",
- this->name, device->device_name);
- ret = -1;
- goto out;
- }
-
- if (ibv_req_notify_cq (device->send_cq, 0)) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: ibv_req_notify_cq on send_cq for device %s"
- " failed", this->name, device->device_name);
- ret = -1;
- goto out;
- }
- } while (0);
-
-out:
- if (ret != 0)
- gf_rdma_destroy_cq (this);
-
- return ret;
-}
-
static int
gf_rdma_register_peer (gf_rdma_device_t *device, int32_t qp_num,
@@ -2196,25 +2854,6 @@ __gf_rdma_lookup_peer (gf_rdma_device_t *device, int32_t qp_num)
return peer;
}
-/*
- static gf_rdma_peer_t *
- gf_rdma_lookup_peer (gf_rdma_device_t *device,
- int32_t qp_num)
- {
- gf_rdma_qpreg_t *qpreg = NULL;
- gf_rdma_peer_t *peer = NULL;
-
- qpreg = &device->qpreg;
- pthread_mutex_lock (&qpreg->lock);
- {
- peer = __gf_rdma_lookup_peer (device, qp_num);
- }
- pthread_mutex_unlock (&qpreg->lock);
-
- return peer;
- }
-*/
-
static void
__gf_rdma_destroy_qp (rpc_transport_t *this)
@@ -2224,7 +2863,7 @@ __gf_rdma_destroy_qp (rpc_transport_t *this)
priv = this->private;
if (priv->peer.qp) {
gf_rdma_unregister_peer (priv->device, priv->peer.qp->qp_num);
- ibv_destroy_qp (priv->peer.qp);
+ rdma_destroy_qp (priv->peer.cm_id);
}
priv->peer.qp = NULL;
@@ -2235,18 +2874,36 @@ __gf_rdma_destroy_qp (rpc_transport_t *this)
static int32_t
gf_rdma_create_qp (rpc_transport_t *this)
{
- gf_rdma_private_t *priv = NULL;
- gf_rdma_options_t *options = NULL;
- gf_rdma_device_t *device = NULL;
- int32_t ret = 0;
- gf_rdma_peer_t *peer = NULL;
+ 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;
- options = &priv->options;
- device = priv->device;
peer = &priv->peer;
+ device_name = (char *)ibv_get_device_name (peer->cm_id->verbs->device);
+ if (device_name == NULL) {
+ ret = -1;
+ gf_log (this->name, GF_LOG_WARNING, "cannot get device_name");
+ goto out;
+ }
+
+ device = gf_rdma_get_device (this, peer->cm_id->verbs,
+ device_name);
+ if (device == NULL) {
+ ret = -1;
+ gf_log (this->name, GF_LOG_WARNING, "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,
@@ -2260,39 +2917,16 @@ gf_rdma_create_qp (rpc_transport_t *this)
.qp_type = IBV_QPT_RC
};
- struct ibv_qp_attr attr = {
- .qp_state = IBV_QPS_INIT,
- .pkey_index = 0,
- .port_num = options->port,
- .qp_access_flags
- = IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE
- };
-
- peer->qp = ibv_create_qp (device->pd, &init_attr);
- if (!peer->qp) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "%s: could not create QP",
- this->name);
- ret = -1;
- goto out;
- } else if (ibv_modify_qp (peer->qp, &attr,
- IBV_QP_STATE |
- IBV_QP_PKEY_INDEX |
- IBV_QP_PORT |
- IBV_QP_ACCESS_FLAGS)) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_ERROR,
- "%s: failed to modify QP to INIT state",
- this->name);
+ ret = rdma_create_qp(peer->cm_id, device->pd, &init_attr);
+ if (ret != 0) {
+ gf_log (peer->trans->name, GF_LOG_CRITICAL,
+ "%s: could not create QP (%s)", this->name,
+ strerror (errno));
ret = -1;
goto out;
}
- peer->local_lid = gf_rdma_get_local_lid (device->context,
- options->port);
- peer->local_qpn = peer->qp->qp_num;
- peer->local_psn = lrand48 () & 0xffffff;
+ peer->qp = peer->cm_id->qp;
ret = gf_rdma_register_peer (device, peer->qp->qp_num, peer);
@@ -2304,300 +2938,52 @@ out:
}
-static void
-gf_rdma_destroy_posts (rpc_transport_t *this)
-{
-
-}
-
-
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 (device, size + 2048, type);
- if (!post) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: post creation failed",
- this->name);
- ret = -1;
- break;
- }
-
- gf_rdma_put_post (q, post);
- }
- return ret;
-}
-
-
-static int32_t
-gf_rdma_create_posts (rpc_transport_t *this)
+__gf_rdma_teardown (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;
+ gf_rdma_private_t *priv = NULL;
+ gf_rdma_peer_t *peer = 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);
+ peer = &priv->peer;
- 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 (peer->cm_id->qp != NULL) {
+ __gf_rdma_destroy_qp (this);
}
- if (ret)
- gf_rdma_destroy_posts (this);
-
- return ret;
-}
-
-
-static int32_t
-gf_rdma_connect_qp (rpc_transport_t *this)
-{
- gf_rdma_private_t *priv = this->private;
- gf_rdma_options_t *options = &priv->options;
- struct ibv_qp_attr attr = {
- .qp_state = IBV_QPS_RTR,
- .path_mtu = options->mtu,
- .dest_qp_num = priv->peer.remote_qpn,
- .rq_psn = priv->peer.remote_psn,
- .max_dest_rd_atomic = 1,
- .min_rnr_timer = 12,
- .qp_access_flags
- = IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE,
- .ah_attr = {
- .is_global = 0,
- .dlid = priv->peer.remote_lid,
- .sl = 0,
- .src_path_bits = 0,
- .port_num = options->port
- }
- };
- if (ibv_modify_qp (priv->peer.qp, &attr,
- IBV_QP_STATE |
- IBV_QP_AV |
- IBV_QP_PATH_MTU |
- IBV_QP_DEST_QPN |
- IBV_QP_RQ_PSN |
- IBV_QP_MAX_DEST_RD_ATOMIC |
- IBV_QP_MIN_RNR_TIMER)) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "Failed to modify QP to RTR\n");
- return -1;
+ if (!list_empty (&priv->peer.ioq)) {
+ __gf_rdma_ioq_flush (peer);
}
- attr.qp_state = IBV_QPS_RTS;
- attr.timeout = options->attr_timeout;
- attr.retry_cnt = options->attr_retry_cnt;
- attr.rnr_retry = options->attr_rnr_retry;
- attr.sq_psn = priv->peer.local_psn;
- attr.max_rd_atomic = 1;
- if (ibv_modify_qp (priv->peer.qp, &attr,
- IBV_QP_STATE |
- IBV_QP_TIMEOUT |
- IBV_QP_RETRY_CNT |
- IBV_QP_RNR_RETRY |
- IBV_QP_SQ_PSN |
- IBV_QP_MAX_QP_RD_ATOMIC)) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "Failed to modify QP to RTS\n");
- return -1;
+ 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)
+gf_rdma_teardown (rpc_transport_t *this)
{
+ int32_t ret = 0;
gf_rdma_private_t *priv = NULL;
- priv = this->private;
- __gf_rdma_destroy_qp (this);
-
- if (!list_empty (&priv->peer.ioq)) {
- __gf_rdma_ioq_flush (&priv->peer);
+ if (this == NULL) {
+ goto out;
}
- /* TODO: decrement cq size */
- return 0;
-}
-
-/*
- * return value:
- * 0 = success (completed)
- * -1 = error
- * > 0 = incomplete
- */
-
-static int
-__tcp_rwv (rpc_transport_t *this, struct iovec *vector, int count,
- struct iovec **pending_vector, int *pending_count,
- int write)
-{
- gf_rdma_private_t *priv = NULL;
- int sock = -1;
- int ret = -1;
- struct iovec *opvector = NULL;
- int opcount = 0;
- int moved = 0;
-
priv = this->private;
- sock = priv->sock;
- opvector = vector;
- opcount = count;
- while (opcount)
+ pthread_mutex_lock (&priv->write_mutex);
{
- if (write)
- {
- ret = writev (sock, opvector, opcount);
-
- if (ret == 0 || (ret == -1 && errno == EAGAIN))
- {
- /* done for now */
- break;
- }
- }
- else
- {
- ret = readv (sock, opvector, opcount);
-
- if (ret == -1 && errno == EAGAIN)
- {
- /* done for now */
- break;
- }
- }
-
- if (ret == 0)
- {
- gf_log (this->name, GF_LOG_DEBUG,
- "EOF from peer %s", this->peerinfo.identifier);
- opcount = -1;
- errno = ENOTCONN;
- break;
- }
-
- if (ret == -1)
- {
- if (errno == EINTR)
- continue;
-
- gf_log (this->name, GF_LOG_DEBUG,
- "%s failed (%s)", write ? "writev" : "readv",
- strerror (errno));
- if (write && !priv->connected &&
- (errno == ECONNREFUSED))
- gf_log (this->name, GF_LOG_ERROR,
- "possible mismatch of 'rpc-transport-type'"
- " in protocol server and client. "
- "check volume file");
- opcount = -1;
- break;
- }
-
- moved = 0;
-
- while (moved < ret)
- {
- if ((ret - moved) >= opvector[0].iov_len)
- {
- moved += opvector[0].iov_len;
- opvector++;
- opcount--;
- }
- else
- {
- opvector[0].iov_len -= (ret - moved);
- opvector[0].iov_base += (ret - moved);
- moved += (ret - moved);
- }
- while (opcount && !opvector[0].iov_len)
- {
- opvector++;
- opcount--;
- }
- }
- }
-
- if (pending_vector)
- *pending_vector = opvector;
-
- if (pending_count)
- *pending_count = opcount;
-
- return opcount;
-}
-
-
-static int
-__tcp_readv (rpc_transport_t *this, struct iovec *vector, int count,
- struct iovec **pending_vector, int *pending_count)
-{
- int ret = -1;
-
- ret = __tcp_rwv (this, vector, count,
- pending_vector, pending_count, 0);
-
- return ret;
-}
-
-
-static int
-__tcp_writev (rpc_transport_t *this, struct iovec *vector, int count,
- struct iovec **pending_vector, int *pending_count)
-{
- int ret = -1;
- gf_rdma_private_t *priv = NULL;
-
- priv = this->private;
-
- ret = __tcp_rwv (this, vector, count, pending_vector,
- pending_count, 1);
-
- if (ret > 0) {
- /* TODO: Avoid multiple calls when socket is already
- registered for POLLOUT */
- priv->idx = event_select_on (this->ctx->event_pool,
- priv->sock, priv->idx, -1, 1);
- } else if (ret == 0) {
- priv->idx = event_select_on (this->ctx->event_pool,
- priv->sock,
- priv->idx, -1, 0);
+ ret = __gf_rdma_teardown (this);
}
+ pthread_mutex_unlock (&priv->write_mutex);
+out:
return ret;
}
@@ -2692,10 +3078,11 @@ inline 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;
+ 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
@@ -2707,6 +3094,10 @@ gf_rdma_decode_error_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
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;
@@ -2721,15 +3112,15 @@ gf_rdma_decode_error_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
iobref_add (iobref, iobuf);
iobuf_unref (iobuf);
- /*
- * FIXME: construct an appropriate rpc-msg here, what is being sent
- * to rpc is not correct.
- */
- post->ctx.vector[0].iov_base = iobuf_ptr (iobuf);
- post->ctx.vector[0].iov_len = bytes_in_post;
-
- memcpy (post->ctx.vector[0].iov_base, (char *)post->buf,
- post->ctx.vector[0].iov_len);
+
+ ret = rpc_reply_to_xdr (&rpc_msg, iobuf_ptr (iobuf),
+ iobuf_pagesize (iobuf), &post->ctx.vector[0]);
+ if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "Failed to create RPC reply");
+ goto out;
+ }
+
post->ctx.count = 1;
iobuf = NULL;
@@ -2767,6 +3158,8 @@ gf_rdma_decode_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
ret = gf_rdma_get_read_chunklist (&ptr, readch);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot get read chunklist from msg");
goto out;
}
@@ -2775,6 +3168,8 @@ gf_rdma_decode_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
ret = gf_rdma_get_write_chunklist (&ptr, &write_ary);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot get write chunklist from msg");
goto out;
}
@@ -2784,6 +3179,8 @@ gf_rdma_decode_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
if (write_ary != NULL) {
reply_info = gf_rdma_reply_info_alloc (peer);
if (reply_info == NULL) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "reply_info_alloc failed");
ret = -1;
goto out;
}
@@ -2794,12 +3191,16 @@ gf_rdma_decode_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
} else {
ret = gf_rdma_get_write_chunklist (&ptr, &write_ary);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "reply_info_alloc_failed");
ret = -1;
goto out;
}
@@ -2837,9 +3238,7 @@ out:
*readch = NULL;
}
- if (write_ary != NULL) {
- GF_FREE (write_ary);
- }
+ GF_FREE (write_ary);
}
return ret;
@@ -2865,28 +3264,36 @@ gf_rdma_decode_header (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
case GF_RDMA_MSG:
case GF_RDMA_NOMSG:
ret = gf_rdma_decode_msg (peer, post, readch, bytes_in_post);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "cannot decode msg of type (%d)",
+ header->rm_type);
+ }
+
break;
case GF_RDMA_MSGP:
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"rdma msg of msg-type GF_RDMA_MSGP should not have "
"been received");
ret = -1;
break;
case GF_RDMA_DONE:
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"rdma msg of msg-type GF_RDMA_DONE should not have "
"been received");
ret = -1;
break;
case GF_RDMA_ERROR:
- /* ret = gf_rdma_decode_error_msg (peer, post, bytes_in_post); */
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "received a msg of type RDMA_ERROR");
+ ret = gf_rdma_decode_error_msg (peer, post, bytes_in_post);
break;
default:
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"unknown rdma msg-type (%d)", header->rm_type);
}
@@ -2904,6 +3311,8 @@ __gf_rdma_read (gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *to,
ret = __gf_rdma_register_local_mr_for_rdma (peer, to, 1, &post->ctx);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "registering local memory for rdma read failed");
goto out;
}
@@ -2921,7 +3330,8 @@ __gf_rdma_read (gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *to,
ret = ibv_post_send (peer->qp, &wr, &bad_wr);
if (ret) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "rdma read from client "
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma read from client "
"(%s) failed with ret = %d (%s)",
peer->trans->peerinfo.identifier,
ret, (ret > 0) ? strerror (ret) : "");
@@ -2950,7 +3360,7 @@ gf_rdma_do_reads (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
}
if (i == 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"message type specified as rdma-read but there are no "
"rdma read-chunks present");
goto out;
@@ -2980,6 +3390,10 @@ gf_rdma_do_reads (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
pthread_mutex_lock (&priv->write_mutex);
{
if (!priv->connected) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "transport not connected to peer (%s), "
+ "not doing rdma reads",
+ peer->trans->peerinfo.identifier);
goto unlock;
}
@@ -2993,6 +3407,9 @@ gf_rdma_do_reads (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
&post->ctx.vector[count],
&readch[i]);
if (ret == -1) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma read from peer (%s) failed",
+ peer->trans->peerinfo.identifier);
goto unlock;
}
@@ -3099,6 +3516,10 @@ gf_rdma_pollin_notify (gf_rdma_peer_t *peer, gf_rdma_post_t *post)
ret = rpc_transport_notify (peer->trans, RPC_TRANSPORT_MSG_RECEIVED,
pollin);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "transport_notify failed");
+ }
out:
if (pollin != NULL) {
@@ -3159,15 +3580,15 @@ gf_rdma_recv_reply (gf_rdma_peer_t *peer, gf_rdma_post_t *post)
RPC_TRANSPORT_MAP_XID_REQUEST,
&request_info);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "cannot get request information from rpc "
- "layer");
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"rpc request structure not found");
ret = -1;
goto out;
@@ -3185,6 +3606,10 @@ gf_rdma_recv_reply (gf_rdma_peer_t *peer, gf_rdma_post_t *post)
out:
if (ret == 0) {
ret = gf_rdma_pollin_notify (peer, post);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "pollin notify failed");
+ }
}
return ret;
@@ -3199,10 +3624,15 @@ gf_rdma_recv_request (gf_rdma_peer_t *peer, gf_rdma_post_t *post,
if (readch != NULL) {
ret = gf_rdma_do_reads (peer, post, readch);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma read from peer (%s) failed",
+ peer->trans->peerinfo.identifier);
+ }
} else {
ret = gf_rdma_pollin_notify (peer, post);
if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"pollin notification failed");
}
}
@@ -3219,23 +3649,44 @@ gf_rdma_process_recv (gf_rdma_peer_t *peer, struct ibv_wc *wc)
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_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"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_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"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 arithmatic 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;
@@ -3251,30 +3702,48 @@ gf_rdma_process_recv (gf_rdma_peer_t *peer, struct ibv_wc *wc)
break;
case GF_RDMA_ERROR:
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "an error has happened while transmission of msg, "
- "disconnecting the transport");
- rpc_transport_disconnect (peer->trans);
- goto out;
-
-/* ret = gf_rdma_pollin_notify (peer, post);
- if (ret == -1) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "pollin notification failed");
- }
- goto out;
-*/
+ if (header->rm_body.rm_error.rm_type == ERR_CHUNK) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "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_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ "pollin notification failed");
+ }
+ goto out;
+ } else {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
+ "an error has happened while transmission of "
+ "msg, disconnecting the transport");
+ ret = -1;
+ goto out;
+ }
default:
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"invalid rdma msg-type (%d)", header->rm_type);
- break;
+ goto out;
}
if (msg_type == CALL) {
ret = gf_rdma_recv_request (peer, post, readch);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "receiving a request from peer (%s) failed",
+ peer->trans->peerinfo.identifier);
+ }
} else {
ret = gf_rdma_recv_reply (peer, post);
+ if (ret < 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "receiving a reply from peer (%s) failed",
+ peer->trans->peerinfo.identifier);
+ }
}
out:
@@ -3285,6 +3754,42 @@ out:
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_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "Error getting event (%s)",
+ strerror (errno));
+ }
+ } while(ret && errno == EINTR);
+
+ switch (event.event_type) {
+ case IBV_EVENT_SRQ_LIMIT_REACHED:
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "recieved srq_limit reached");
+ break;
+
+ default:
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
+ "event (%d) recieved", event.event_type);
+ break;
+ }
+
+ ibv_ack_async_event(&event);
+ }
+
+ return 0;
+}
+
static void *
gf_rdma_recv_completion_proc (void *data)
@@ -3403,7 +3908,7 @@ gf_rdma_handle_failed_send_completion (gf_rdma_peer_t *peer, struct ibv_wc *wc)
post = (gf_rdma_post_t *) (long) wc->wr_id;
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
"send work request on `%s' returned error "
"wc.status = %d, wc.vendor_err = %d, post->buf = %p, "
"wc.byte_len = %d, post->reused = %d",
@@ -3670,364 +4175,87 @@ gf_rdma_options_init (rpc_transport_t *this)
return;
}
-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 gf_rdma_device_t *
-gf_rdma_get_device (rpc_transport_t *this, struct ibv_context *ibctx)
+gf_rdma_ctx_t *
+__gf_rdma_ctx_create (void)
{
- glusterfs_ctx_t *ctx = NULL;
- gf_rdma_private_t *priv = NULL;
- gf_rdma_options_t *options = NULL;
- char *device_name = NULL;
- uint32_t port = 0;
- uint8_t active_port = 0;
- int32_t ret = 0;
- int32_t i = 0;
- gf_rdma_device_t *trav = NULL;
+ gf_rdma_ctx_t *rdma_ctx = NULL;
+ int ret = -1;
- priv = this->private;
- options = &priv->options;
- device_name = priv->options.device_name;
- ctx = this->ctx;
- trav = ctx->ib;
- port = priv->options.port;
-
- while (trav) {
- if ((!strcmp (trav->device_name, device_name)) &&
- (trav->port == port))
- break;
- trav = trav->next;
+ rdma_ctx = GF_CALLOC (1, sizeof (*rdma_ctx), gf_common_mt_char);
+ if (rdma_ctx == NULL) {
+ goto out;
}
- if (!trav) {
-
- trav = GF_CALLOC (1, sizeof (*trav),
- gf_common_mt_rdma_device_t);
- if (trav == NULL) {
- return NULL;
- }
-
- priv->device = trav;
-
- trav->context = ibctx;
-
- ret = ib_get_active_port (trav->context);
-
- if (ret < 0) {
- if (!port) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "Failed to find any active ports and "
- "none specified in volume file,"
- " exiting");
- GF_FREE (trav);
- return NULL;
- }
- }
-
- trav->request_ctx_pool = mem_pool_new (gf_rdma_request_context_t,
- GF_RDMA_POOL_SIZE);
- if (trav->request_ctx_pool == NULL) {
- return NULL;
- }
-
- trav->ioq_pool = mem_pool_new (gf_rdma_ioq_t, GF_RDMA_POOL_SIZE);
- if (trav->ioq_pool == NULL) {
- mem_pool_destroy (trav->request_ctx_pool);
- return NULL;
- }
-
- trav->reply_info_pool = mem_pool_new (gf_rdma_reply_info_t,
- GF_RDMA_POOL_SIZE);
- if (trav->reply_info_pool == NULL) {
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->ioq_pool);
- return NULL;
- }
-
-
- active_port = ret;
-
- if (port) {
- ret = ib_check_active_port (trav->context, port);
- if (ret < 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
- "On device %s: provided port:%u is "
- "found to be offline, continuing to "
- "use the same port", device_name, port);
- }
- } else {
- priv->options.port = active_port;
- port = active_port;
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_TRACE,
- "Port unspecified in volume file using active "
- "port: %u", port);
- }
-
- trav->device_name = gf_strdup (device_name);
- trav->port = port;
-
- trav->next = ctx->ib;
- ctx->ib = trav;
-
- trav->send_chan = ibv_create_comp_channel (trav->context);
- if (!trav->send_chan) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not create send completion channel",
- device_name);
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- GF_FREE ((char *)trav->device_name);
- GF_FREE (trav);
- return NULL;
- }
-
- trav->recv_chan = ibv_create_comp_channel (trav->context);
- if (!trav->recv_chan) {
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- ibv_destroy_comp_channel (trav->send_chan);
- GF_FREE ((char *)trav->device_name);
- GF_FREE (trav);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "could not create recv completion channel");
- /* TODO: cleanup current mess */
- return NULL;
- }
-
- if (gf_rdma_create_cq (this) < 0) {
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- ibv_destroy_comp_channel (trav->recv_chan);
- ibv_destroy_comp_channel (trav->send_chan);
- GF_FREE ((char *)trav->device_name);
- GF_FREE (trav);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not create CQ",
- this->name);
- return NULL;
- }
-
- /* protection domain */
- trav->pd = ibv_alloc_pd (trav->context);
-
- if (!trav->pd) {
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- 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);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not allocate protection domain",
- this->name);
- return NULL;
- }
-
- struct ibv_srq_init_attr attr = {
- .attr = {
- .max_wr = options->recv_count,
- .max_sge = 1
- }
- };
- trav->srq = ibv_create_srq (trav->pd, &attr);
-
- if (!trav->srq) {
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- 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);
-
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not create SRQ",
- this->name);
- return NULL;
- }
-
- /* queue init */
- gf_rdma_queue_init (&trav->sendq);
- gf_rdma_queue_init (&trav->recvq);
-
- if (gf_rdma_create_posts (this) < 0) {
- mem_pool_destroy (trav->ioq_pool);
- mem_pool_destroy (trav->request_ctx_pool);
- mem_pool_destroy (trav->reply_info_pool);
- 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);
-
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not allocate posts",
- this->name);
- return NULL;
- }
-
- /* completion threads */
- ret = pthread_create (&trav->send_thread,
- NULL,
- gf_rdma_send_completion_proc,
- trav->send_chan);
- if (ret) {
- 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);
- 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);
+ rdma_ctx->rdma_cm_event_channel = rdma_create_event_channel ();
+ if (rdma_ctx->rdma_cm_event_channel == NULL) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "rdma_cm event channel creation failed (%s)",
+ strerror (errno));
+ goto out;
+ }
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "could not create send completion thread");
- return NULL;
- }
+ ret = gf_thread_create (&rdma_ctx->rdma_cm_thread, NULL,
+ gf_rdma_cm_event_handler,
+ rdma_ctx->rdma_cm_event_channel);
+ if (ret != 0) {
+ gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING,
+ "creation of thread to handle rdma-cm events "
+ "failed (%s)", strerror (ret));
+ goto out;
+ }
- ret = pthread_create (&trav->recv_thread,
- NULL,
- gf_rdma_recv_completion_proc,
- trav->recv_chan);
- if (ret) {
- 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);
- 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);
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "could not create recv completion thread");
- return NULL;
+out:
+ if (ret < 0) {
+ if (rdma_ctx->rdma_cm_event_channel != NULL) {
+ rdma_destroy_event_channel (rdma_ctx->rdma_cm_event_channel);
}
- /* 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];
- }
+ GF_FREE (rdma_ctx);
+ rdma_ctx = NULL;
}
- return trav;
+
+ return rdma_ctx;
}
static int32_t
gf_rdma_init (rpc_transport_t *this)
{
gf_rdma_private_t *priv = NULL;
- gf_rdma_options_t *options = NULL;
- struct ibv_device **dev_list;
- struct ibv_context *ib_ctx = NULL;
int32_t ret = 0;
+ glusterfs_ctx_t *ctx = NULL;
+ gf_rdma_options_t *options = NULL;
+
+ ctx= this->ctx;
priv = this->private;
- options = &priv->options;
ibv_fork_init ();
gf_rdma_options_init (this);
- {
- dev_list = ibv_get_device_list (NULL);
-
- if (!dev_list) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "Failed to get IB devices");
- ret = -1;
- goto cleanup;
- }
-
- if (!*dev_list) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "No IB devices found");
- ret = -1;
- goto cleanup;
- }
-
- if (!options->device_name) {
- if (*dev_list) {
- options->device_name =
- gf_strdup (ibv_get_device_name (*dev_list));
- } else {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_CRITICAL,
- "IB device list is empty. Check for "
- "'ib_uverbs' module");
- return -1;
- goto cleanup;
- }
- }
-
- while (*dev_list) {
- if (!strcmp (ibv_get_device_name (*dev_list),
- options->device_name)) {
- ib_ctx = ibv_open_device (*dev_list);
-
- if (!ib_ctx) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_ERROR,
- "Failed to get infiniband"
- "device context");
- ret = -1;
- goto cleanup;
- }
- break;
- }
- ++dev_list;
- }
-
- priv->device = gf_rdma_get_device (this, ib_ctx);
-
- if (!priv->device) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "could not create rdma device for %s",
- options->device_name);
- ret = -1;
- goto cleanup;
- }
- }
+ 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->peer.trans = this;
INIT_LIST_HEAD (&priv->peer.ioq);
- pthread_mutex_init (&priv->read_mutex, NULL);
pthread_mutex_init (&priv->write_mutex, NULL);
pthread_mutex_init (&priv->recv_mutex, NULL);
pthread_cond_init (&priv->recv_cond, NULL);
-cleanup:
- if (-1 == ret) {
- if (ib_ctx)
- ibv_close_device (ib_ctx);
+ pthread_mutex_lock (&ctx->lock);
+ {
+ if (ctx->ib == NULL) {
+ ctx->ib = __gf_rdma_ctx_create ();
+ if (ctx->ib == NULL) {
+ ret = -1;
+ }
+ }
}
-
- if (dev_list)
- ibv_free_device_list (dev_list);
+ pthread_mutex_unlock (&ctx->lock);
return ret;
}
@@ -4040,543 +4268,69 @@ gf_rdma_disconnect (rpc_transport_t *this)
int32_t ret = 0;
priv = this->private;
- pthread_mutex_lock (&priv->write_mutex);
- {
- ret = __gf_rdma_disconnect (this);
- }
- pthread_mutex_unlock (&priv->write_mutex);
-
- return ret;
-}
-
-
-static int32_t
-__tcp_connect_finish (int fd)
-{
- int ret = -1;
- int optval = 0;
- socklen_t optlen = sizeof (int);
-
- ret = getsockopt (fd, SOL_SOCKET, SO_ERROR,
- (void *)&optval, &optlen);
-
- if (ret == 0 && optval)
- {
- errno = optval;
- ret = -1;
- }
-
- return ret;
-}
-
-static inline void
-gf_rdma_fill_handshake_data (char *buf, struct gf_rdma_nbio *nbio,
- gf_rdma_private_t *priv)
-{
- sprintf (buf,
- "QP1:RECV_BLKSIZE=%08x:SEND_BLKSIZE=%08x\n"
- "QP1:LID=%04x:QPN=%06x:PSN=%06x\n",
- priv->peer.recv_size,
- priv->peer.send_size,
- priv->peer.local_lid,
- priv->peer.local_qpn,
- priv->peer.local_psn);
-
- nbio->vector.iov_base = buf;
- nbio->vector.iov_len = strlen (buf) + 1;
- nbio->count = 1;
- return;
-}
-
-static inline void
-gf_rdma_fill_handshake_ack (char *buf, struct gf_rdma_nbio *nbio)
-{
- sprintf (buf, "DONE\n");
- nbio->vector.iov_base = buf;
- nbio->vector.iov_len = strlen (buf) + 1;
- nbio->count = 1;
- return;
-}
-
-static int
-gf_rdma_handshake_pollin (rpc_transport_t *this)
-{
- int ret = 0;
- gf_rdma_private_t *priv = NULL;
- char *buf = NULL;
- int32_t recv_buf_size = 0, send_buf_size;
- socklen_t sock_len = 0;
-
- priv = this->private;
- buf = priv->handshake.incoming.buf;
-
- if (priv->handshake.incoming.state == GF_RDMA_HANDSHAKE_COMPLETE) {
- return -1;
- }
+ gf_log_callingfn (this->name, GF_LOG_WARNING,
+ "disconnect called (peer:%s)",
+ this->peerinfo.identifier);
pthread_mutex_lock (&priv->write_mutex);
{
- while (priv->handshake.incoming.state != GF_RDMA_HANDSHAKE_COMPLETE)
- {
- switch (priv->handshake.incoming.state)
- {
- case GF_RDMA_HANDSHAKE_START:
- buf = priv->handshake.incoming.buf = GF_CALLOC (1, 256, gf_common_mt_char);
- gf_rdma_fill_handshake_data (buf, &priv->handshake.incoming, priv);
- buf[0] = 0;
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_RECEIVING_DATA;
- break;
-
- case GF_RDMA_HANDSHAKE_RECEIVING_DATA:
- ret = __tcp_readv (this,
- &priv->handshake.incoming.vector,
- priv->handshake.incoming.count,
- &priv->handshake.incoming.pending_vector,
- &priv->handshake.incoming.pending_count);
- if (ret == -1) {
- goto unlock;
- }
-
- if (ret > 0) {
- gf_log (this->name, GF_LOG_TRACE,
- "partial header read on NB socket. continue later");
- goto unlock;
- }
-
- if (!ret) {
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_RECEIVED_DATA;
- }
- break;
-
- case GF_RDMA_HANDSHAKE_RECEIVED_DATA:
- ret = sscanf (buf,
- "QP1:RECV_BLKSIZE=%08x:SEND_BLKSIZE=%08x\n"
- "QP1:LID=%04x:QPN=%06x:PSN=%06x\n",
- &recv_buf_size,
- &send_buf_size,
- &priv->peer.remote_lid,
- &priv->peer.remote_qpn,
- &priv->peer.remote_psn);
-
- if ((ret != 5) && (strncmp (buf, "QP1:", 4))) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_CRITICAL,
- "%s: remote-host(%s)'s "
- "transport type is different",
- this->name,
- this->peerinfo.identifier);
- ret = -1;
- goto unlock;
- }
-
- if (recv_buf_size < priv->peer.recv_size)
- priv->peer.recv_size = recv_buf_size;
- if (send_buf_size < priv->peer.send_size)
- priv->peer.send_size = send_buf_size;
-
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_TRACE,
- "%s: transacted recv_size=%d "
- "send_size=%d",
- this->name, priv->peer.recv_size,
- priv->peer.send_size);
-
- priv->peer.quota = priv->peer.send_count;
-
- if (gf_rdma_connect_qp (this)) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_ERROR,
- "%s: failed to connect with "
- "remote QP", this->name);
- ret = -1;
- goto unlock;
- }
- gf_rdma_fill_handshake_ack (buf, &priv->handshake.incoming);
- buf[0] = 0;
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_RECEIVING_ACK;
- break;
-
- case GF_RDMA_HANDSHAKE_RECEIVING_ACK:
- ret = __tcp_readv (this,
- &priv->handshake.incoming.vector,
- priv->handshake.incoming.count,
- &priv->handshake.incoming.pending_vector,
- &priv->handshake.incoming.pending_count);
- if (ret == -1) {
- goto unlock;
- }
-
- if (ret > 0) {
- gf_log (this->name, GF_LOG_TRACE,
- "partial header read on NB "
- "socket. continue later");
- goto unlock;
- }
-
- if (!ret) {
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_RECEIVED_ACK;
- }
- break;
-
- case GF_RDMA_HANDSHAKE_RECEIVED_ACK:
- if (strncmp (buf, "DONE", 4)) {
- gf_log (GF_RDMA_LOG_NAME,
- GF_LOG_DEBUG,
- "%s: handshake-3 did not "
- "return 'DONE' (%s)",
- this->name, buf);
- ret = -1;
- goto unlock;
- }
- ret = 0;
- priv->connected = 1;
- sock_len = sizeof (struct sockaddr_storage);
- getpeername (priv->sock,
- (struct sockaddr *) &this->peerinfo.sockaddr,
- &sock_len);
-
- GF_FREE (priv->handshake.incoming.buf);
- priv->handshake.incoming.buf = NULL;
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_COMPLETE;
- }
- }
- }
-unlock:
- pthread_mutex_unlock (&priv->write_mutex);
-
- if (ret == -1) {
- rpc_transport_disconnect (this);
- } else {
- ret = 0;
- }
-
-
- if (!ret && priv->connected) {
- if (priv->is_server) {
- ret = rpc_transport_notify (priv->listener,
- RPC_TRANSPORT_ACCEPT,
- this);
- } else {
- ret = rpc_transport_notify (this, RPC_TRANSPORT_CONNECT,
- this);
- }
- }
-
- return ret;
-}
-
-static int
-gf_rdma_handshake_pollout (rpc_transport_t *this)
-{
- gf_rdma_private_t *priv = NULL;
- char *buf = NULL;
- int32_t ret = 0;
-
- priv = this->private;
- buf = priv->handshake.outgoing.buf;
-
- if (priv->handshake.outgoing.state == GF_RDMA_HANDSHAKE_COMPLETE) {
- return 0;
- }
-
- pthread_mutex_unlock (&priv->write_mutex);
- {
- while (priv->handshake.outgoing.state
- != GF_RDMA_HANDSHAKE_COMPLETE)
- {
- switch (priv->handshake.outgoing.state)
- {
- case GF_RDMA_HANDSHAKE_START:
- buf = priv->handshake.outgoing.buf
- = GF_CALLOC (1, 256, gf_common_mt_char);
- gf_rdma_fill_handshake_data (buf,
- &priv->handshake.outgoing, priv);
- priv->handshake.outgoing.state
- = GF_RDMA_HANDSHAKE_SENDING_DATA;
- break;
-
- case GF_RDMA_HANDSHAKE_SENDING_DATA:
- ret = __tcp_writev (this,
- &priv->handshake.outgoing.vector,
- priv->handshake.outgoing.count,
- &priv->handshake.outgoing.pending_vector,
- &priv->handshake.outgoing.pending_count);
- if (ret == -1) {
- goto unlock;
- }
-
- if (ret > 0) {
- gf_log (this->name, GF_LOG_TRACE,
- "partial header read on NB "
- "socket. continue later");
- goto unlock;
- }
-
- if (!ret) {
- priv->handshake.outgoing.state
- = GF_RDMA_HANDSHAKE_SENT_DATA;
- }
- break;
-
- case GF_RDMA_HANDSHAKE_SENT_DATA:
- gf_rdma_fill_handshake_ack (buf,
- &priv->handshake.outgoing);
- priv->handshake.outgoing.state
- = GF_RDMA_HANDSHAKE_SENDING_ACK;
- break;
-
- case GF_RDMA_HANDSHAKE_SENDING_ACK:
- ret = __tcp_writev (this,
- &priv->handshake.outgoing.vector,
- priv->handshake.outgoing.count,
- &priv->handshake.outgoing.pending_vector,
- &priv->handshake.outgoing.pending_count);
-
- if (ret == -1) {
- goto unlock;
- }
-
- if (ret > 0) {
- gf_log (this->name, GF_LOG_TRACE,
- "partial header read on NB "
- "socket. continue later");
- goto unlock;
- }
-
- if (!ret) {
- GF_FREE (priv->handshake.outgoing.buf);
- priv->handshake.outgoing.buf = NULL;
- priv->handshake.outgoing.state
- = GF_RDMA_HANDSHAKE_COMPLETE;
- }
- break;
- }
- }
- }
-unlock:
- pthread_mutex_unlock (&priv->write_mutex);
-
- if (ret == -1) {
- rpc_transport_disconnect (this);
- } else {
- ret = 0;
- }
-
- return ret;
-}
-
-static int
-gf_rdma_handshake_pollerr (rpc_transport_t *this)
-{
- gf_rdma_private_t *priv = this->private;
- char need_unref = 0, connected = 0;
-
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG,
- "%s: peer disconnected, cleaning up",
- this->name);
-
- pthread_mutex_lock (&priv->write_mutex);
- {
- __gf_rdma_teardown (this);
-
- connected = priv->connected;
- if (priv->sock != -1) {
- event_unregister (this->ctx->event_pool,
- priv->sock, priv->idx);
- need_unref = 1;
-
- if (close (priv->sock) != 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "close () - error: %s",
- strerror (errno));
- }
- priv->tcp_connected = priv->connected = 0;
- priv->sock = -1;
- }
-
- if (priv->handshake.incoming.buf) {
- GF_FREE (priv->handshake.incoming.buf);
- priv->handshake.incoming.buf = NULL;
- }
-
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_START;
-
- if (priv->handshake.outgoing.buf) {
- GF_FREE (priv->handshake.outgoing.buf);
- priv->handshake.outgoing.buf = NULL;
- }
-
- priv->handshake.outgoing.state = GF_RDMA_HANDSHAKE_START;
- }
- pthread_mutex_unlock (&priv->write_mutex);
-
- if (connected) {
- rpc_transport_notify (this, RPC_TRANSPORT_DISCONNECT, this);
- }
-
- if (need_unref)
- rpc_transport_unref (this);
-
- return 0;
-}
-
-
-static int
-tcp_connect_finish (rpc_transport_t *this)
-{
- gf_rdma_private_t *priv = NULL;
- int error = 0, ret = 0;
-
- priv = this->private;
- pthread_mutex_lock (&priv->write_mutex);
- {
- ret = __tcp_connect_finish (priv->sock);
-
- if (!ret) {
- this->myinfo.sockaddr_len =
- sizeof (this->myinfo.sockaddr);
- ret = getsockname (priv->sock,
- (struct sockaddr *)&this->myinfo.sockaddr,
- &this->myinfo.sockaddr_len);
- if (ret == -1)
- {
- gf_log (this->name, GF_LOG_ERROR,
- "getsockname on new client-socket %d "
- "failed (%s)",
- priv->sock, strerror (errno));
- close (priv->sock);
- error = 1;
- goto unlock;
- }
-
- gf_rdma_get_transport_identifiers (this);
- priv->tcp_connected = 1;
- }
-
- if (ret == -1 && errno != EINPROGRESS) {
- gf_log (this->name, GF_LOG_ERROR,
- "tcp connect to %s failed (%s)",
- this->peerinfo.identifier, strerror (errno));
- error = 1;
- }
+ ret = __gf_rdma_disconnect (this);
}
-unlock:
pthread_mutex_unlock (&priv->write_mutex);
- if (error) {
- rpc_transport_disconnect (this);
- }
-
return ret;
}
-static int
-gf_rdma_event_handler (int fd, int idx, void *data,
- int poll_in, int poll_out, int poll_err)
-{
- rpc_transport_t *this = NULL;
- gf_rdma_private_t *priv = NULL;
- gf_rdma_options_t *options = NULL;
- int ret = 0;
-
- this = data;
- priv = this->private;
- if (!priv->tcp_connected) {
- ret = tcp_connect_finish (this);
- if (priv->tcp_connected) {
- 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;
-
- if ((ret = gf_rdma_create_qp (this)) < 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not create QP",
- this->name);
- rpc_transport_disconnect (this);
- }
- }
- }
-
- if (!ret && poll_out && priv->tcp_connected) {
- ret = gf_rdma_handshake_pollout (this);
- }
-
- if (!ret && !poll_err && poll_in && priv->tcp_connected) {
- if (priv->handshake.incoming.state
- == GF_RDMA_HANDSHAKE_COMPLETE) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: pollin received on tcp socket (peer: %s) "
- "after handshake is complete",
- this->name, this->peerinfo.identifier);
- gf_rdma_handshake_pollerr (this);
- return 0;
- }
- ret = gf_rdma_handshake_pollin (this);
- }
-
- if (ret < 0 || poll_err) {
- ret = gf_rdma_handshake_pollerr (this);
- }
-
- return 0;
-}
-
-static int
-__tcp_nonblock (int fd)
-{
- int flags = 0;
- int ret = -1;
-
- flags = fcntl (fd, F_GETFL);
-
- if (flags != -1)
- ret = fcntl (fd, F_SETFL, flags | O_NONBLOCK);
-
- return ret;
-}
static int32_t
gf_rdma_connect (struct rpc_transport *this, int port)
{
gf_rdma_private_t *priv = NULL;
int32_t ret = 0;
- gf_boolean_t non_blocking = 1;
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_log (this->name, GF_LOG_DEBUG,
"cannot get remote address to connect");
- return ret;
+ goto out;
}
+ rdma_ctx = this->ctx->ib;
pthread_mutex_lock (&priv->write_mutex);
{
- if (priv->sock != -1) {
- ret = 0;
+ if (peer->cm_id != NULL) {
+ ret = -1;
+ errno = EINPROGRESS;
+ connected = _gf_true;
goto unlock;
}
- priv->sock = socket (sock_union.sa.sa_family, SOCK_STREAM, 0);
+ priv->entity = GF_RDMA_CLIENT;
- if (priv->sock == -1) {
+ ret = rdma_create_id (rdma_ctx->rdma_cm_event_channel,
+ &peer->cm_id, this, RDMA_PS_TCP);
+ if (ret != 0) {
gf_log (this->name, GF_LOG_ERROR,
- "socket () - error: %s", strerror (errno));
+ "creation of rdma_cm_id failed (%s)",
+ strerror (errno));
ret = -errno;
goto unlock;
}
- gf_log (this->name, GF_LOG_TRACE,
- "socket fd = %d", priv->sock);
-
memcpy (&this->peerinfo.sockaddr, &sock_union.storage,
sockaddr_len);
this->peerinfo.sockaddr_len = sockaddr_len;
@@ -4587,201 +4341,84 @@ gf_rdma_connect (struct rpc_transport *this, int port)
((struct sockaddr *) &this->myinfo.sockaddr)->sa_family =
((struct sockaddr *)&this->peerinfo.sockaddr)->sa_family;
- if (non_blocking)
- {
- ret = __tcp_nonblock (priv->sock);
-
- if (ret == -1)
- {
- gf_log (this->name, GF_LOG_ERROR,
- "could not set socket %d to non "
- "blocking mode (%s)",
- priv->sock, strerror (errno));
- close (priv->sock);
- priv->sock = -1;
- goto unlock;
- }
- }
-
ret = gf_rdma_client_bind (this,
(struct sockaddr *)&this->myinfo.sockaddr,
&this->myinfo.sockaddr_len,
- priv->sock);
- if (ret == -1)
- {
+ peer->cm_id);
+ if (ret != 0) {
gf_log (this->name, GF_LOG_WARNING,
"client bind failed: %s", strerror (errno));
- close (priv->sock);
- priv->sock = -1;
goto unlock;
}
- ret = connect (priv->sock,
- (struct sockaddr *)&this->peerinfo.sockaddr,
- this->peerinfo.sockaddr_len);
- if (ret == -1 && errno != EINPROGRESS)
- {
- gf_log (this->name, GF_LOG_ERROR,
- "connection attempt failed (%s)",
+ ret = rdma_resolve_addr (peer->cm_id, NULL, &sock_union.sa,
+ 2000);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "rdma_resolve_addr failed (%s)",
strerror (errno));
- close (priv->sock);
- priv->sock = -1;
goto unlock;
}
- priv->tcp_connected = priv->connected = 0;
-
- rpc_transport_ref (this);
-
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_START;
- priv->handshake.outgoing.state = GF_RDMA_HANDSHAKE_START;
-
- priv->idx = event_register (this->ctx->event_pool,
- priv->sock, gf_rdma_event_handler,
- this, 1, 1);
+ priv->connected = 0;
}
unlock:
pthread_mutex_unlock (&priv->write_mutex);
- return ret;
-}
-
-static int
-gf_rdma_server_event_handler (int fd, int idx, void *data,
- int poll_in, int poll_out, int poll_err)
-{
- int32_t main_sock = -1;
- rpc_transport_t *this = NULL, *trans = NULL;
- gf_rdma_private_t *priv = NULL;
- gf_rdma_private_t *trans_priv = NULL;
- gf_rdma_options_t *options = NULL;
-
- if (!poll_in) {
- return 0;
- }
-
- trans = data;
- trans_priv = (gf_rdma_private_t *) trans->private;
-
- this = GF_CALLOC (1, sizeof (rpc_transport_t),
- gf_common_mt_rpc_transport_t);
- if (this == NULL) {
- return -1;
- }
-
- this->listener = trans;
-
- priv = GF_CALLOC (1, sizeof (gf_rdma_private_t),
- gf_common_mt_rdma_private_t);
- if (priv == NULL) {
- GF_FREE (priv);
- return -1;
- }
- this->private = priv;
- /* Copy all the rdma related values in priv, from trans_priv
- as other than QP, all the values remain same */
- priv->device = trans_priv->device;
- priv->options = trans_priv->options;
- priv->is_server = 1;
- priv->listener = trans;
-
- options = &priv->options;
-
- this->ops = trans->ops;
- this->init = trans->init;
- this->fini = trans->fini;
- this->ctx = trans->ctx;
- this->name = gf_strdup (trans->name);
- this->notify = trans->notify;
- this->mydata = trans->mydata;
-
- memcpy (&this->myinfo.sockaddr, &trans->myinfo.sockaddr,
- trans->myinfo.sockaddr_len);
- this->myinfo.sockaddr_len = trans->myinfo.sockaddr_len;
-
- main_sock = (trans_priv)->sock;
- this->peerinfo.sockaddr_len = sizeof (this->peerinfo.sockaddr);
- priv->sock = accept (main_sock,
- (struct sockaddr *)&this->peerinfo.sockaddr,
- &this->peerinfo.sockaddr_len);
- if (priv->sock == -1) {
- gf_log ("rdma/server", GF_LOG_ERROR,
- "accept() failed: %s",
- strerror (errno));
- GF_FREE (this->private);
- GF_FREE (this);
- return -1;
- }
-
- priv->peer.trans = this;
- rpc_transport_ref (this);
-
- gf_rdma_get_transport_identifiers (this);
-
- priv->tcp_connected = 1;
- priv->handshake.incoming.state = GF_RDMA_HANDSHAKE_START;
- priv->handshake.outgoing.state = GF_RDMA_HANDSHAKE_START;
-
- 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;
- INIT_LIST_HEAD (&priv->peer.ioq);
+out:
+ if (ret != 0) {
+ if (!connected) {
+ gf_rdma_teardown (this);
+ }
- if (gf_rdma_create_qp (this) < 0) {
- gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR,
- "%s: could not create QP",
- this->name);
- rpc_transport_disconnect (this);
- return -1;
+ rpc_transport_unref (this);
}
- priv->idx = event_register (this->ctx->event_pool, priv->sock,
- gf_rdma_event_handler, this, 1, 1);
-
- pthread_mutex_init (&priv->read_mutex, NULL);
- pthread_mutex_init (&priv->write_mutex, NULL);
- pthread_mutex_init (&priv->recv_mutex, NULL);
- /* pthread_cond_init (&priv->recv_cond, NULL); */
- return 0;
+ 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;
- int opt = 1, ret = 0;
+ gf_rdma_peer_t *peer = NULL;
+ int ret = 0;
+ gf_rdma_ctx_t *rdma_ctx = NULL;
char service[NI_MAXSERV], host[NI_MAXHOST];
priv = this->private;
- memset (&sock_union, 0, sizeof (sock_union));
- ret = gf_rdma_server_get_local_sockaddr (this,
- &sock_union.sa,
+ 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_log (this->name, GF_LOG_DEBUG,
+ gf_log (this->name, GF_LOG_WARNING,
"cannot find network address of server to bind to");
goto err;
}
- priv->sock = socket (sock_union.sa.sa_family, SOCK_STREAM, 0);
- if (priv->sock == -1) {
- gf_log ("rdma/server", GF_LOG_CRITICAL,
- "init: failed to create socket, error: %s",
+ ret = rdma_create_id (rdma_ctx->rdma_cm_event_channel,
+ &peer->cm_id, this, RDMA_PS_TCP);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "creation of rdma_cm_id failed (%s)",
strerror (errno));
- GF_FREE (this->private);
- ret = -1;
goto err;
}
- memcpy (&this->myinfo.sockaddr, &sock_union.storage, sockaddr_len);
+ 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),
+ this->myinfo.sockaddr_len, host, sizeof (host),
service, sizeof (service),
NI_NUMERICHOST);
if (ret != 0) {
@@ -4789,34 +4426,38 @@ gf_rdma_listen (rpc_transport_t *this)
"getnameinfo failed (%s)", gai_strerror (ret));
goto err;
}
+
sprintf (this->myinfo.identifier, "%s:%s", host, service);
- setsockopt (priv->sock, SOL_SOCKET, SO_REUSEADDR, &opt, sizeof (opt));
- if (bind (priv->sock, &sock_union.sa, sockaddr_len) != 0) {
- ret = -1;
- gf_log ("rdma/server", GF_LOG_ERROR,
- "init: failed to bind to socket for %s (%s)",
- this->myinfo.identifier, strerror (errno));
+ ret = rdma_bind_addr (peer->cm_id, &sock_union.sa);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "rdma_bind_addr failed (%s)", strerror (errno));
goto err;
}
- if (listen (priv->sock, 10) != 0) {
- gf_log ("rdma/server", GF_LOG_ERROR,
- "init: listen () failed on socket for %s (%s)",
- this->myinfo.identifier, strerror (errno));
- ret = -1;
+ ret = rdma_listen (peer->cm_id, 10);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "rdma_listen failed (%s)", strerror (errno));
goto err;
}
- /* Register the main socket */
- priv->idx = event_register (this->ctx->event_pool, priv->sock,
- gf_rdma_server_event_handler,
- rpc_transport_ref (this), 1, 0);
+ 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,
@@ -4835,7 +4476,6 @@ init (rpc_transport_t *this)
return -1;
this->private = priv;
- priv->sock = -1;
if (gf_rdma_init (this)) {
gf_log (this->name, GF_LOG_ERROR,
@@ -4859,13 +4499,6 @@ fini (struct rpc_transport *this)
if (priv) {
pthread_mutex_destroy (&priv->recv_mutex);
pthread_mutex_destroy (&priv->write_mutex);
- pthread_mutex_destroy (&priv->read_mutex);
-
- /* pthread_cond_destroy (&priv->recv_cond); */
- if (priv->sock != -1) {
- event_unregister (this->ctx->event_pool,
- priv->sock, priv->idx);
- }
gf_log (this->name, GF_LOG_TRACE,
"called fini on transport: %p", this);
diff --git a/rpc/rpc-transport/rdma/src/rdma.h b/rpc/rpc-transport/rdma/src/rdma.h
index 687d6005f..7f76244f0 100644
--- a/rpc/rpc-transport/rdma/src/rdma.h
+++ b/rpc/rpc-transport/rdma/src/rdma.h
@@ -29,6 +29,7 @@
#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)
@@ -230,30 +231,33 @@ typedef enum __gf_rdma_send_post_type {
/* represents one communication peer, two per transport_t */
struct __gf_rdma_peer {
- rpc_transport_t *trans;
- struct ibv_qp *qp;
+ 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;
+ int32_t quota;
union {
- struct list_head ioq;
+ struct list_head ioq;
struct {
- gf_rdma_ioq_t *ioq_next;
- gf_rdma_ioq_t *ioq_prev;
+ 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;
+ 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;
@@ -320,33 +324,19 @@ struct __gf_rdma_device {
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;
+ 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;
};
typedef struct __gf_rdma_device gf_rdma_device_t;
-typedef enum {
- GF_RDMA_HANDSHAKE_START = 0,
- GF_RDMA_HANDSHAKE_SENDING_DATA,
- GF_RDMA_HANDSHAKE_RECEIVING_DATA,
- GF_RDMA_HANDSHAKE_SENT_DATA,
- GF_RDMA_HANDSHAKE_RECEIVED_DATA,
- GF_RDMA_HANDSHAKE_SENDING_ACK,
- GF_RDMA_HANDSHAKE_RECEIVING_ACK,
- GF_RDMA_HANDSHAKE_RECEIVED_ACK,
- GF_RDMA_HANDSHAKE_COMPLETE,
-} gf_rdma_handshake_state_t;
-
-struct gf_rdma_nbio {
- int state;
- char *buf;
- int count;
- struct iovec vector;
- struct iovec *pending_vector;
- int pending_count;
+struct __gf_rdma_ctx {
+ gf_rdma_device_t *device;
+ struct rdma_event_channel *rdma_cm_event_channel;
+ pthread_t rdma_cm_thread;
};
+typedef struct __gf_rdma_ctx gf_rdma_ctx_t;
struct __gf_rdma_request_context {
struct ibv_mr *mr[GF_RDMA_MAX_SEGMENTS];
@@ -358,46 +348,35 @@ struct __gf_rdma_request_context {
};
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 sock;
- int32_t idx;
- unsigned char connected;
- unsigned char tcp_connected;
- unsigned char ib_connected;
- in_addr_t addr;
+ int32_t idx;
+ unsigned char connected;
+ in_addr_t addr;
unsigned short port;
/* IB Verbs Driver specific variables, pointers */
- gf_rdma_peer_t peer;
+ gf_rdma_peer_t peer;
struct __gf_rdma_device *device;
- gf_rdma_options_t options;
+ gf_rdma_options_t options;
/* Used by trans->op->receive */
- char *data_ptr;
- int32_t data_offset;
- int32_t data_len;
+ char *data_ptr;
+ int32_t data_offset;
+ int32_t data_len;
/* Mutex */
- pthread_mutex_t read_mutex;
- pthread_mutex_t write_mutex;
- pthread_barrier_t handshake_barrier;
- char handshake_ret;
- char is_server;
- rpc_transport_t *listener;
-
- pthread_mutex_t recv_mutex;
- pthread_cond_t recv_cond;
-
- /* used during gf_rdma_handshake */
- struct {
- struct gf_rdma_nbio incoming;
- struct gf_rdma_nbio outgoing;
- int state;
- gf_rdma_header_t header;
- char *buf;
- size_t size;
- } handshake;
+ 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;
};
-typedef struct __gf_rdma_private gf_rdma_private_t;
+typedef struct __gf_rdma_private gf_rdma_private_t;
#endif /* _XPORT_GF_RDMA_H */
diff --git a/rpc/rpc-transport/socket/src/Makefile.am b/rpc/rpc-transport/socket/src/Makefile.am
index 2c918c7e3..71e6ed6ff 100644
--- a/rpc/rpc-transport/socket/src/Makefile.am
+++ b/rpc/rpc-transport/socket/src/Makefile.am
@@ -3,13 +3,15 @@ noinst_HEADERS = socket.h name.h
rpctransport_LTLIBRARIES = socket.la
rpctransportdir = $(libdir)/glusterfs/$(PACKAGE_VERSION)/rpc-transport
-socket_la_LDFLAGS = -module -avoidversion
+socket_la_LDFLAGS = -module -avoid-version
socket_la_SOURCES = socket.c name.c
-socket_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la
+socket_la_LIBADD = $(top_builddir)/libglusterfs/src/libglusterfs.la -lssl
-AM_CFLAGS = -fPIC -D_FILE_OFFSET_BITS=64 -D_GNU_SOURCE -Wall -D$(GF_HOST_OS)\
+AM_CPPFLAGS = $(GF_CPPFLAGS) \
-I$(top_srcdir)/libglusterfs/src -I$(top_srcdir)/rpc/rpc-lib/src/ \
- -I$(top_srcdir)/rpc/xdr/src/ -shared -nostartfiles $(GF_CFLAGS)
+ -I$(top_srcdir)/rpc/xdr/src/
+
+AM_CFLAGS = -Wall $(GF_CFLAGS)
CLEANFILES = *~
diff --git a/rpc/rpc-transport/socket/src/name.c b/rpc/rpc-transport/socket/src/name.c
index 0b2ca5805..1647d5b6b 100644
--- a/rpc/rpc-transport/socket/src/name.c
+++ b/rpc/rpc-transport/socket/src/name.c
@@ -15,12 +15,6 @@
#include <netdb.h>
#include <string.h>
-#ifdef CLIENT_PORT_CEILING
-#undef CLIENT_PORT_CEILING
-#endif
-
-#define CLIENT_PORT_CEILING 1024
-
#ifndef AF_INET_SDP
#define AF_INET_SDP 27
#endif
@@ -40,9 +34,17 @@ static int32_t
af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
socklen_t sockaddr_len, int ceiling)
{
- int32_t ret = -1;
- /* struct sockaddr_in sin = {0, }; */
- uint16_t port = ceiling - 1;
+ int32_t ret = -1;
+ uint16_t port = ceiling - 1;
+ // by default assume none of the ports are blocked and all are available
+ gf_boolean_t ports[1024] = {_gf_false,};
+ int i = 0;
+
+ ret = gf_process_reserved_ports (ports);
+ if (ret != 0) {
+ for (i = 0; i < 1024; i++)
+ ports[i] = _gf_false;
+ }
while (port)
{
@@ -57,7 +59,11 @@ af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
((struct sockaddr_in *)sockaddr)->sin_port = htons (port);
break;
}
-
+ // ignore the reserved ports
+ if (ports[port] == _gf_true) {
+ port--;
+ continue;
+ }
ret = bind (fd, sockaddr, sockaddr_len);
if (ret == 0)
@@ -134,24 +140,24 @@ client_fill_address_family (rpc_transport_t *this, sa_family_t *sa_family)
if (!(remote_host_data || connect_path_data) ||
(remote_host_data && connect_path_data)) {
gf_log (this->name, GF_LOG_ERROR,
- "transport.address-family not specified and "
- "not able to determine the "
- "same from other options (remote-host:%s and "
- "transport.unix.connect-path:%s)",
+ "transport.address-family not specified. "
+ "Could not guess default value from (remote-host:%s or "
+ "transport.unix.connect-path:%s) options",
data_to_str (remote_host_data),
data_to_str (connect_path_data));
+ *sa_family = AF_UNSPEC;
goto out;
}
if (remote_host_data) {
gf_log (this->name, GF_LOG_DEBUG,
"address-family not specified, guessing it "
- "to be inet/inet6");
- *sa_family = AF_UNSPEC;
+ "to be inet from (remote-host: %s)", data_to_str (remote_host_data));
+ *sa_family = AF_INET;
} else {
gf_log (this->name, GF_LOG_DEBUG,
"address-family not specified, guessing it "
- "to be unix");
+ "to be unix from (transport.unix.connect-path: %s)", data_to_str (connect_path_data));
*sa_family = AF_UNIX;
}
@@ -165,13 +171,11 @@ client_fill_address_family (rpc_transport_t *this, sa_family_t *sa_family)
*sa_family = AF_INET6;
} else if (!strcasecmp (address_family, "inet-sdp")) {
*sa_family = AF_INET_SDP;
- } else if (!strcasecmp (address_family, "inet/inet6")
- || !strcasecmp (address_family, "inet6/inet")) {
- *sa_family = AF_UNSPEC;
} else {
gf_log (this->name, GF_LOG_ERROR,
"unknown address-family (%s) specified",
address_family);
+ *sa_family = AF_UNSPEC;
goto out;
}
}
@@ -442,12 +446,12 @@ client_bind (rpc_transport_t *this,
case AF_INET6:
if (!this->bind_insecure) {
ret = af_inet_bind_to_port_lt_ceiling (sock, sockaddr,
- *sockaddr_len, CLIENT_PORT_CEILING);
+ *sockaddr_len, GF_CLIENT_PORT_CEILING);
}
if (ret == -1) {
gf_log (this->name, GF_LOG_DEBUG,
"cannot bind inet socket (%d) to port less than %d (%s)",
- sock, CLIENT_PORT_CEILING, strerror (errno));
+ sock, GF_CLIENT_PORT_CEILING, strerror (errno));
ret = 0;
}
break;
@@ -542,18 +546,16 @@ server_fill_address_family (rpc_transport_t *this, sa_family_t *sa_family)
*sa_family = AF_INET_SDP;
} else if (!strcasecmp (address_family, "unix")) {
*sa_family = AF_UNIX;
- } else if (!strcasecmp (address_family, "inet/inet6")
- || !strcasecmp (address_family, "inet6/inet")) {
- *sa_family = AF_UNSPEC;
} else {
gf_log (this->name, GF_LOG_ERROR,
"unknown address family (%s) specified", address_family);
+ *sa_family = AF_UNSPEC;
goto out;
}
} else {
gf_log (this->name, GF_LOG_DEBUG,
- "option address-family not specified, defaulting to inet/inet6");
- *sa_family = AF_UNSPEC;
+ "option address-family not specified, defaulting to inet");
+ *sa_family = AF_INET;
}
ret = 0;
diff --git a/rpc/rpc-transport/socket/src/socket.c b/rpc/rpc-transport/socket/src/socket.c
index 6c2d909e4..93da3f296 100644
--- a/rpc/rpc-transport/socket/src/socket.c
+++ b/rpc/rpc-transport/socket/src/socket.c
@@ -35,97 +35,401 @@
#include <errno.h>
#include <netinet/tcp.h>
#include <rpc/xdr.h>
+#include <sys/ioctl.h>
#define GF_LOG_ERRNO(errno) ((errno == ENOTCONN) ? GF_LOG_DEBUG : GF_LOG_ERROR)
#define SA(ptr) ((struct sockaddr *)ptr)
+#define SSL_ENABLED_OPT "transport.socket.ssl-enabled"
+#define SSL_OWN_CERT_OPT "transport.socket.ssl-own-cert"
+#define SSL_PRIVATE_KEY_OPT "transport.socket.ssl-private-key"
+#define SSL_CA_LIST_OPT "transport.socket.ssl-ca-list"
+#define OWN_THREAD_OPT "transport.socket.own-thread"
-#define __socket_proto_reset_pending(priv) do { \
- memset (&priv->incoming.frag.vector, 0, \
- sizeof (priv->incoming.frag.vector)); \
- priv->incoming.frag.pending_vector = \
- &priv->incoming.frag.vector; \
- priv->incoming.frag.pending_vector->iov_base = \
- priv->incoming.frag.fragcurrent; \
- priv->incoming.pending_vector = \
- priv->incoming.frag.pending_vector; \
- } while (0);
+/* TBD: do automake substitutions etc. (ick) to set these. */
+#if !defined(DEFAULT_CERT_PATH)
+#define DEFAULT_CERT_PATH "/etc/ssl/glusterfs.pem"
+#endif
+#if !defined(DEFAULT_KEY_PATH)
+#define DEFAULT_KEY_PATH "/etc/ssl/glusterfs.key"
+#endif
+#if !defined(DEFAULT_CA_PATH)
+#define DEFAULT_CA_PATH "/etc/ssl/glusterfs.ca"
+#endif
+
+#define POLL_MASK_INPUT (POLLIN | POLLPRI)
+#define POLL_MASK_OUTPUT (POLLOUT)
+#define POLL_MASK_ERROR (POLLERR | POLLHUP | POLLNVAL)
+
+typedef int SSL_unary_func (SSL *);
+typedef int SSL_trinary_func (SSL *, void *, int);
+
+#define __socket_proto_reset_pending(priv) do { \
+ struct gf_sock_incoming_frag *frag; \
+ frag = &priv->incoming.frag; \
+ \
+ memset (&frag->vector, 0, sizeof (frag->vector)); \
+ frag->pending_vector = &frag->vector; \
+ frag->pending_vector->iov_base = frag->fragcurrent; \
+ priv->incoming.pending_vector = frag->pending_vector; \
+ } while (0)
#define __socket_proto_update_pending(priv) \
do { \
- uint32_t remaining_fragsize = 0; \
- if (priv->incoming.frag.pending_vector->iov_len == 0) { \
- remaining_fragsize = RPC_FRAGSIZE (priv->incoming.fraghdr) \
- - priv->incoming.frag.bytes_read; \
+ uint32_t remaining; \
+ struct gf_sock_incoming_frag *frag; \
+ frag = &priv->incoming.frag; \
+ if (frag->pending_vector->iov_len == 0) { \
+ remaining = (RPC_FRAGSIZE (priv->incoming.fraghdr) \
+ - frag->bytes_read); \
\
- priv->incoming.frag.pending_vector->iov_len = \
- remaining_fragsize > priv->incoming.frag.remaining_size \
- ? priv->incoming.frag.remaining_size : remaining_fragsize; \
+ frag->pending_vector->iov_len = \
+ (remaining > frag->remaining_size) \
+ ? frag->remaining_size : remaining; \
\
- priv->incoming.frag.remaining_size -= \
- priv->incoming.frag.pending_vector->iov_len; \
+ frag->remaining_size -= \
+ frag->pending_vector->iov_len; \
} \
- } while (0);
+ } while (0)
#define __socket_proto_update_priv_after_read(priv, ret, bytes_read) \
{ \
- priv->incoming.frag.fragcurrent += bytes_read; \
- priv->incoming.frag.bytes_read += bytes_read; \
+ struct gf_sock_incoming_frag *frag; \
+ frag = &priv->incoming.frag; \
\
- if ((ret > 0) || (priv->incoming.frag.remaining_size != 0)) { \
- if (priv->incoming.frag.remaining_size != 0 && ret == 0) { \
+ frag->fragcurrent += bytes_read; \
+ frag->bytes_read += bytes_read; \
+ \
+ if ((ret > 0) || (frag->remaining_size != 0)) { \
+ if (frag->remaining_size != 0 && ret == 0) { \
__socket_proto_reset_pending (priv); \
} \
\
- gf_log (this->name, GF_LOG_TRACE, "partial read on non-blocking socket"); \
+ gf_log (this->name, GF_LOG_TRACE, \
+ "partial read on non-blocking socket"); \
\
break; \
} \
}
-#define __socket_proto_init_pending(priv, size) \
+#define __socket_proto_init_pending(priv,size) \
do { \
- uint32_t remaining_fragsize = 0; \
- remaining_fragsize = RPC_FRAGSIZE (priv->incoming.fraghdr) \
- - priv->incoming.frag.bytes_read; \
+ uint32_t remaining = 0; \
+ struct gf_sock_incoming_frag *frag; \
+ frag = &priv->incoming.frag; \
+ \
+ remaining = (RPC_FRAGSIZE (priv->incoming.fraghdr) \
+ - frag->bytes_read); \
\
- __socket_proto_reset_pending (priv); \
+ __socket_proto_reset_pending (priv); \
\
- priv->incoming.frag.pending_vector->iov_len = \
- remaining_fragsize > size ? size : remaining_fragsize; \
+ frag->pending_vector->iov_len = \
+ (remaining > size) ? size : remaining; \
\
- priv->incoming.frag.remaining_size = \
- size - priv->incoming.frag.pending_vector->iov_len; \
+ frag->remaining_size = (size - frag->pending_vector->iov_len); \
\
- } while (0);
+ } while(0)
/* This will be used in a switch case and breaks from the switch case if all
* the pending data is not read.
*/
#define __socket_proto_read(priv, ret) \
- { \
+ { \
size_t bytes_read = 0; \
+ struct gf_sock_incoming *in; \
+ in = &priv->incoming; \
\
__socket_proto_update_pending (priv); \
\
ret = __socket_readv (this, \
- priv->incoming.pending_vector, 1, \
- &priv->incoming.pending_vector, \
- &priv->incoming.pending_count, \
+ in->pending_vector, 1, \
+ &in->pending_vector, \
+ &in->pending_count, \
&bytes_read); \
- if (ret == -1) { \
- gf_log (this->name, GF_LOG_WARNING, \
- "reading from socket failed. Error (%s), " \
- "peer (%s)", strerror (errno), \
- this->peerinfo.identifier); \
+ if (ret == -1) \
break; \
- } \
__socket_proto_update_priv_after_read (priv, ret, bytes_read); \
}
+static int socket_init (rpc_transport_t *this);
+
+static void
+ssl_dump_error_stack (const char *caller)
+{
+ unsigned long errnum = 0;
+ char errbuf[120] = {0,};
+
+ /* OpenSSL docs explicitly give 120 as the error-string length. */
+
+ while ((errnum = ERR_get_error())) {
+ ERR_error_string(errnum,errbuf);
+ gf_log(caller,GF_LOG_ERROR," %s",errbuf);
+ }
+}
+
+static int
+ssl_do (rpc_transport_t *this, void *buf, size_t len, SSL_trinary_func *func)
+{
+ int r = (-1);
+ struct pollfd pfd = {-1,};
+ socket_private_t *priv = NULL;
+
+ GF_VALIDATE_OR_GOTO(this->name,this->private,out);
+ priv = this->private;
+
+ for (;;) {
+ if (buf) {
+ if (priv->connected == -1) {
+ /*
+ * Fields in the SSL structure (especially
+ * the BIO pointers) are not valid at this
+ * point, so we'll segfault if we pass them
+ * to SSL_read/SSL_write.
+ */
+ gf_log(this->name,GF_LOG_INFO,
+ "lost connection in %s", __func__);
+ break;
+ }
+ 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);
+ }
+ switch (SSL_get_error(priv->ssl_ssl,r)) {
+ case SSL_ERROR_NONE:
+ return r;
+ case SSL_ERROR_WANT_READ:
+ pfd.fd = priv->sock;
+ pfd.events = POLLIN;
+ if (poll(&pfd,1,-1) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,"poll error %d",
+ errno);
+ }
+ break;
+ case SSL_ERROR_WANT_WRITE:
+ pfd.fd = priv->sock;
+ pfd.events = POLLOUT;
+ if (poll(&pfd,1,-1) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,"poll error %d",
+ errno);
+ }
+ break;
+ case SSL_ERROR_SYSCALL:
+ /* This is what we get when remote disconnects. */
+ gf_log(this->name,GF_LOG_DEBUG,
+ "syscall error (probably remote disconnect)");
+ errno = ENODATA;
+ goto out;
+ default:
+ errno = EIO;
+ goto out; /* "break" would just loop again */
+ }
+ }
+out:
+ return -1;
+}
+
+#define ssl_connect_one(t) ssl_do((t),NULL,0,(SSL_trinary_func *)SSL_connect)
+#define ssl_accept_one(t) ssl_do((t),NULL,0,(SSL_trinary_func *)SSL_accept)
+#define ssl_read_one(t,b,l) ssl_do((t),(b),(l),(SSL_trinary_func *)SSL_read)
+#define ssl_write_one(t,b,l) ssl_do((t),(b),(l),(SSL_trinary_func *)SSL_write)
-int socket_init (rpc_transport_t *this);
+static int
+ssl_setup_connection (rpc_transport_t *this, int server)
+{
+ X509 *peer = NULL;
+ char peer_CN[256] = "";
+ int ret = -1;
+ socket_private_t *priv = NULL;
+
+ GF_VALIDATE_OR_GOTO(this->name,this->private,done);
+ priv = this->private;
+
+ priv->ssl_ssl = SSL_new(priv->ssl_ctx);
+ if (!priv->ssl_ssl) {
+ gf_log(this->name,GF_LOG_ERROR,"SSL_new failed");
+ ssl_dump_error_stack(this->name);
+ goto done;
+ }
+ priv->ssl_sbio = BIO_new_socket(priv->sock,BIO_NOCLOSE);
+ if (!priv->ssl_sbio) {
+ gf_log(this->name,GF_LOG_ERROR,"BIO_new_socket failed");
+ ssl_dump_error_stack(this->name);
+ goto free_ssl;
+ }
+ SSL_set_bio(priv->ssl_ssl,priv->ssl_sbio,priv->ssl_sbio);
+
+ if (server) {
+ ret = ssl_accept_one(this);
+ }
+ else {
+ ret = ssl_connect_one(this);
+ }
+
+ /* Make sure _the call_ succeeded. */
+ if (ret < 0) {
+ goto ssl_error;
+ }
+
+ /* Make sure _SSL verification_ succeeded, yielding an identity. */
+ if (SSL_get_verify_result(priv->ssl_ssl) != X509_V_OK) {
+ goto ssl_error;
+ }
+ peer = SSL_get_peer_certificate(priv->ssl_ssl);
+ if (!peer) {
+ goto ssl_error;
+ }
+
+ /* Finally, everything seems OK. */
+ X509_NAME_get_text_by_NID(X509_get_subject_name(peer),
+ NID_commonName, peer_CN, sizeof(peer_CN)-1);
+ peer_CN[sizeof(peer_CN)-1] = '\0';
+ gf_log(this->name,GF_LOG_INFO,"peer CN = %s", peer_CN);
+ return 0;
+
+ /* Error paths. */
+ssl_error:
+ gf_log(this->name,GF_LOG_ERROR,"SSL connect error");
+ ssl_dump_error_stack(this->name);
+free_ssl:
+ SSL_free(priv->ssl_ssl);
+ priv->ssl_ssl = NULL;
+done:
+ return ret;
+}
+
+
+static void
+ssl_teardown_connection (socket_private_t *priv)
+{
+ SSL_shutdown(priv->ssl_ssl);
+ SSL_clear(priv->ssl_ssl);
+ SSL_free(priv->ssl_ssl);
+ priv->ssl_ssl = NULL;
+}
+
+
+static ssize_t
+__socket_ssl_readv (rpc_transport_t *this, struct iovec *opvector, int opcount)
+{
+ socket_private_t *priv = NULL;
+ int sock = -1;
+ int ret = -1;
+
+ priv = this->private;
+ sock = priv->sock;
+
+ if (priv->use_ssl) {
+ ret = ssl_read_one (this, opvector->iov_base, opvector->iov_len);
+ } else {
+ ret = readv (sock, opvector, opcount);
+ }
+
+ return ret;
+}
+
+
+static ssize_t
+__socket_ssl_read (rpc_transport_t *this, void *buf, size_t count)
+{
+ struct iovec iov = {0, };
+ int ret = -1;
+
+ iov.iov_base = buf;
+ iov.iov_len = count;
+
+ ret = __socket_ssl_readv (this, &iov, 1);
+
+ return ret;
+}
+
+
+static int
+__socket_cached_read (rpc_transport_t *this, struct iovec *opvector, int opcount)
+{
+ socket_private_t *priv = NULL;
+ int sock = -1;
+ struct gf_sock_incoming *in = NULL;
+ int req_len = -1;
+ int ret = -1;
+
+ priv = this->private;
+ sock = priv->sock;
+ in = &priv->incoming;
+ req_len = iov_length (opvector, opcount);
+
+ if (in->record_state == SP_STATE_READING_FRAGHDR) {
+ in->ra_read = 0;
+ in->ra_served = 0;
+ in->ra_max = 0;
+ in->ra_buf = NULL;
+ goto uncached;
+ }
+
+ if (!in->ra_max) {
+ /* first call after passing SP_STATE_READING_FRAGHDR */
+ in->ra_max = min (RPC_FRAGSIZE (in->fraghdr), GF_SOCKET_RA_MAX);
+ /* Note that the in->iobuf is the primary iobuf into which
+ headers are read into. By using this itself as our
+ read-ahead cache, we can avoid memory copies in iov_load
+ */
+ in->ra_buf = iobuf_ptr (in->iobuf);
+ }
+
+ /* fill read-ahead */
+ if (in->ra_read < in->ra_max) {
+ ret = __socket_ssl_read (this, &in->ra_buf[in->ra_read],
+ (in->ra_max - in->ra_read));
+ if (ret > 0)
+ in->ra_read += ret;
+
+ /* we proceed to test if there is still cached data to
+ be served even if readahead could not progress */
+ }
+
+ /* serve cached */
+ if (in->ra_served < in->ra_read) {
+ ret = iov_load (opvector, opcount, &in->ra_buf[in->ra_served],
+ min (req_len, (in->ra_read - in->ra_served)));
+
+ in->ra_served += ret;
+ /* Do not read uncached and cached in the same call */
+ goto out;
+ }
+
+ if (in->ra_read < in->ra_max)
+ /* If there was no cached data to be served, (and we are
+ guaranteed to have already performed an attempt to progress
+ readahead above), and we have not yet read out the full
+ readahead capacity, then bail out for now without doing
+ the uncached read below (as that will overtake future cached
+ read)
+ */
+ goto out;
+uncached:
+ ret = __socket_ssl_readv (this, opvector, opcount);
+out:
+ return ret;
+}
+
+static gf_boolean_t
+__does_socket_rwv_error_need_logging (socket_private_t *priv, int write)
+{
+ int read = !write;
+
+ if (priv->connected == -1) /* Didn't even connect, of course it fails */
+ return _gf_false;
+
+ if (read && (priv->read_fail_log == _gf_false))
+ return _gf_false;
+
+ return _gf_true;
+}
/*
* return value:
@@ -134,7 +438,7 @@ int socket_init (rpc_transport_t *this);
* > 0 = incomplete
*/
-int
+static int
__socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
struct iovec **pending_vector, int *pending_count, size_t *bytes,
int write)
@@ -159,9 +463,22 @@ __socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
*bytes = 0;
}
- while (opcount) {
+ while (opcount > 0) {
+ if (opvector->iov_len == 0) {
+ gf_log(this->name,GF_LOG_DEBUG,
+ "would have passed zero length to read/write");
+ ++opvector;
+ --opcount;
+ continue;
+ }
if (write) {
- ret = writev (sock, opvector, opcount);
+ if (priv->use_ssl) {
+ ret = ssl_write_one(this,
+ opvector->iov_base, opvector->iov_len);
+ }
+ else {
+ ret = writev (sock, opvector, opcount);
+ }
if (ret == 0 || (ret == -1 && errno == EAGAIN)) {
/* done for now */
@@ -169,7 +486,13 @@ __socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
}
this->total_bytes_write += ret;
} else {
- ret = readv (sock, opvector, opcount);
+ ret = __socket_cached_read (this, opvector, opcount);
+
+ if (ret == 0) {
+ gf_log(this->name,GF_LOG_DEBUG,"EOF on socket");
+ errno = ENODATA;
+ ret = -1;
+ }
if (ret == -1 && errno == EAGAIN) {
/* done for now */
break;
@@ -190,9 +513,18 @@ __socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
if (errno == EINTR)
continue;
- gf_log (this->name, GF_LOG_WARNING,
- "%s failed (%s)", write ? "writev" : "readv",
- strerror (errno));
+ if (__does_socket_rwv_error_need_logging (priv,
+ write)) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "%s on %s failed (%s)",
+ write ? "writev":"readv",
+ this->peerinfo.identifier,
+ strerror (errno));
+ }
+
+ if (priv->use_ssl) {
+ ssl_dump_error_stack(this->name);
+ }
opcount = -1;
break;
}
@@ -204,6 +536,17 @@ __socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
moved = 0;
while (moved < ret) {
+ if (!opcount) {
+ gf_log(this->name,GF_LOG_DEBUG,
+ "ran out of iov, moved %d/%d",
+ moved, ret);
+ goto ran_out;
+ }
+ if (!opvector[0].iov_len) {
+ opvector++;
+ opcount--;
+ continue;
+ }
if ((ret - moved) >= opvector[0].iov_len) {
moved += opvector[0].iov_len;
opvector++;
@@ -213,13 +556,11 @@ __socket_rwv (rpc_transport_t *this, struct iovec *vector, int count,
opvector[0].iov_base += (ret - moved);
moved += (ret - moved);
}
- while (opcount && !opvector[0].iov_len) {
- opvector++;
- opcount--;
- }
}
}
+ran_out:
+
if (pending_vector)
*pending_vector = opvector;
@@ -231,7 +572,7 @@ out:
}
-int
+static int
__socket_readv (rpc_transport_t *this, struct iovec *vector, int count,
struct iovec **pending_vector, int *pending_count,
size_t *bytes)
@@ -245,7 +586,7 @@ __socket_readv (rpc_transport_t *this, struct iovec *vector, int count,
}
-int
+static int
__socket_writev (rpc_transport_t *this, struct iovec *vector, int count,
struct iovec **pending_vector, int *pending_count)
{
@@ -258,26 +599,55 @@ __socket_writev (rpc_transport_t *this, struct iovec *vector, int count,
}
-int
+static int
+__socket_shutdown (rpc_transport_t *this)
+{
+ int ret = -1;
+ socket_private_t *priv = this->private;
+
+ priv->connected = -1;
+ ret = shutdown (priv->sock, SHUT_RDWR);
+ if (ret) {
+ /* its already disconnected.. no need to understand
+ why it failed to shutdown in normal cases */
+ gf_log (this->name, GF_LOG_DEBUG,
+ "shutdown() returned %d. %s",
+ ret, strerror (errno));
+ }
+
+ return ret;
+}
+
+static int
__socket_disconnect (rpc_transport_t *this)
{
- socket_private_t *priv = NULL;
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, state=%u gen=%u sock=%d", this,
+ priv->ot_state, priv->ot_gen, priv->sock);
+
if (priv->sock != -1) {
- priv->connected = -1;
- ret = shutdown (priv->sock, SHUT_RDWR);
- if (ret) {
- /* its already disconnected.. no need to understand
- why it failed to shutdown in normal cases */
- gf_log (this->name, GF_LOG_DEBUG,
- "shutdown() returned %d. %s",
- ret, strerror (errno));
+ ret = __socket_shutdown(this);
+ if (priv->own_thread) {
+ /*
+ * Without this, reconnect (= disconnect + connect)
+ * won't work except by accident.
+ */
+ close(priv->sock);
+ priv->sock = -1;
+ gf_log (this->name, GF_LOG_TRACE,
+ "OT_PLEASE_DIE on %p", this);
+ priv->ot_state = OT_PLEASE_DIE;
+ }
+ else if (priv->use_ssl) {
+ ssl_teardown_connection(priv);
}
}
@@ -286,7 +656,7 @@ out:
}
-int
+static int
__socket_server_bind (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -314,7 +684,7 @@ __socket_server_bind (rpc_transport_t *this)
memcpy (&unix_addr, SA (&this->myinfo.sockaddr),
this->myinfo.sockaddr_len);
reuse_check_sock = socket (AF_UNIX, SOCK_STREAM, 0);
- if (reuse_check_sock > 0) {
+ if (reuse_check_sock >= 0) {
ret = connect (reuse_check_sock, SA (&unix_addr),
this->myinfo.sockaddr_len);
if ((ret == -1) && (ECONNREFUSED == errno)) {
@@ -342,7 +712,7 @@ out:
}
-int
+static int
__socket_nonblock (int fd)
{
int flags = 0;
@@ -356,8 +726,7 @@ __socket_nonblock (int fd)
return ret;
}
-
-int
+static int
__socket_nodelay (int fd)
{
int on = 1;
@@ -374,7 +743,7 @@ __socket_nodelay (int fd)
static int
-__socket_keepalive (int fd, int keepalive_intvl, int keepalive_idle)
+__socket_keepalive (int fd, int family, int keepalive_intvl, int keepalive_idle)
{
int on = 1;
int ret = -1;
@@ -403,18 +772,23 @@ __socket_keepalive (int fd, int keepalive_intvl, int keepalive_idle)
goto err;
}
#else
+ if (family != AF_INET)
+ goto done;
+
ret = setsockopt (fd, IPPROTO_TCP, TCP_KEEPIDLE, &keepalive_idle,
sizeof (keepalive_intvl));
if (ret == -1) {
gf_log ("socket", GF_LOG_WARNING,
- "failed to set keep idle on socket %d", fd);
+ "failed to set keep idle %d on socket %d, %s",
+ keepalive_idle, fd, strerror(errno));
goto err;
}
- ret = setsockopt (fd, IPPROTO_TCP, TCP_KEEPINTVL, &keepalive_intvl,
+ ret = setsockopt (fd, IPPROTO_TCP , TCP_KEEPINTVL, &keepalive_intvl,
sizeof (keepalive_intvl));
if (ret == -1) {
gf_log ("socket", GF_LOG_WARNING,
- "failed to set keep alive interval on socket %d", fd);
+ "failed to set keep interval %d on socket %d, %s",
+ keepalive_intvl, fd, strerror(errno));
goto err;
}
#endif
@@ -428,7 +802,7 @@ err:
}
-int
+static int
__socket_connect_finish (int fd)
{
int ret = -1;
@@ -446,7 +820,7 @@ __socket_connect_finish (int fd)
}
-void
+static void
__socket_reset (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -467,9 +841,7 @@ __socket_reset (rpc_transport_t *this)
iobuf_unref (priv->incoming.iobuf);
}
- if (priv->incoming.request_info != NULL) {
- GF_FREE (priv->incoming.request_info);
- }
+ GF_FREE (priv->incoming.request_info);
memset (&priv->incoming, 0, sizeof (priv->incoming));
@@ -485,13 +857,13 @@ out:
}
-void
+static void
socket_set_lastfrag (uint32_t *fragsize) {
(*fragsize) |= 0x80000000U;
}
-void
+static void
socket_set_frag_header_size (uint32_t size, char *haddr)
{
size = htonl (size);
@@ -499,14 +871,14 @@ socket_set_frag_header_size (uint32_t size, char *haddr)
}
-void
+static void
socket_set_last_frag_header_size (uint32_t size, char *haddr)
{
socket_set_lastfrag (&size);
socket_set_frag_header_size (size, haddr);
}
-struct ioq *
+static struct ioq *
__socket_ioq_new (rpc_transport_t *this, rpc_transport_msg_t *msg)
{
struct ioq *entry = NULL;
@@ -573,7 +945,7 @@ out:
}
-void
+static void
__socket_ioq_entry_free (struct ioq *entry)
{
GF_VALIDATE_OR_GOTO ("socket", entry, out);
@@ -590,7 +962,7 @@ out:
}
-void
+static void
__socket_ioq_flush (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -611,10 +983,12 @@ out:
}
-int
-__socket_ioq_churn_entry (rpc_transport_t *this, struct ioq *entry)
+static int
+__socket_ioq_churn_entry (rpc_transport_t *this, struct ioq *entry, int direct)
{
- int ret = -1;
+ int ret = -1;
+ socket_private_t *priv = NULL;
+ char a_byte = 0;
ret = __socket_writev (this, entry->pending_vector,
entry->pending_count,
@@ -625,13 +999,25 @@ __socket_ioq_churn_entry (rpc_transport_t *this, struct ioq *entry)
/* current entry was completely written */
GF_ASSERT (entry->pending_count == 0);
__socket_ioq_entry_free (entry);
+ priv = this->private;
+ if (priv->own_thread) {
+ /*
+ * The pipe should only remain readable if there are
+ * more entries after this, so drain the byte
+ * representing this entry.
+ */
+ if (!direct && read(priv->pipe[0],&a_byte,1) < 1) {
+ gf_log(this->name,GF_LOG_WARNING,
+ "read error on pipe");
+ }
+ }
}
return ret;
}
-int
+static int
__socket_ioq_churn (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -647,13 +1033,13 @@ __socket_ioq_churn (rpc_transport_t *this)
/* pick next entry */
entry = priv->ioq_next;
- ret = __socket_ioq_churn_entry (this, entry);
+ ret = __socket_ioq_churn_entry (this, entry, 0);
if (ret != 0)
break;
}
- if (list_empty (&priv->ioq)) {
+ if (!priv->own_thread && 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);
@@ -664,7 +1050,7 @@ out:
}
-int
+static int
socket_event_poll_err (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -689,7 +1075,7 @@ out:
}
-int
+static int
socket_event_poll_out (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -719,43 +1105,45 @@ out:
}
-inline int
+static inline int
__socket_read_simple_msg (rpc_transport_t *this)
{
- socket_private_t *priv = NULL;
- int ret = 0;
- uint32_t remaining_size = 0;
- size_t bytes_read = 0;
+ int ret = 0;
+ uint32_t remaining_size = 0;
+ size_t bytes_read = 0;
+ socket_private_t *priv = NULL;
+ 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;
- switch (priv->incoming.frag.simple_state) {
+ in = &priv->incoming;
+ frag = &in->frag;
+
+ switch (frag->simple_state) {
case SP_STATE_SIMPLE_MSG_INIT:
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
__socket_proto_init_pending (priv, remaining_size);
- priv->incoming.frag.simple_state =
- SP_STATE_READING_SIMPLE_MSG;
+ frag->simple_state = SP_STATE_READING_SIMPLE_MSG;
/* fall through */
case SP_STATE_READING_SIMPLE_MSG:
ret = 0;
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
if (remaining_size > 0) {
ret = __socket_readv (this,
- priv->incoming.pending_vector, 1,
- &priv->incoming.pending_vector,
- &priv->incoming.pending_count,
+ in->pending_vector, 1,
+ &in->pending_vector,
+ &in->pending_count,
&bytes_read);
}
@@ -767,8 +1155,8 @@ __socket_read_simple_msg (rpc_transport_t *this)
break;
}
- priv->incoming.frag.bytes_read += bytes_read;
- priv->incoming.frag.fragcurrent += bytes_read;
+ frag->bytes_read += bytes_read;
+ frag->fragcurrent += bytes_read;
if (ret > 0) {
gf_log (this->name, GF_LOG_TRACE,
@@ -777,8 +1165,7 @@ __socket_read_simple_msg (rpc_transport_t *this)
}
if (ret == 0) {
- priv->incoming.frag.simple_state
- = SP_STATE_SIMPLE_MSG_INIT;
+ frag->simple_state = SP_STATE_SIMPLE_MSG_INIT;
}
}
@@ -787,7 +1174,7 @@ out:
}
-inline int
+static inline int
__socket_read_simple_request (rpc_transport_t *this)
{
return __socket_read_simple_msg (this);
@@ -804,7 +1191,7 @@ __socket_read_simple_request (rpc_transport_t *this)
#define rpc_progver_addr(buf) (buf + RPC_MSGTYPE_SIZE + 8)
#define rpc_procnum_addr(buf) (buf + RPC_MSGTYPE_SIZE + 12)
-inline int
+static inline int
__socket_read_vectored_request (rpc_transport_t *this, rpcsvc_vector_sizer vector_sizer)
{
socket_private_t *priv = NULL;
@@ -814,17 +1201,26 @@ __socket_read_vectored_request (rpc_transport_t *this, rpcsvc_vector_sizer vecto
struct iobuf *iobuf = NULL;
uint32_t remaining_size = 0;
ssize_t readsize = 0;
- size_t size = 0;
+ size_t size = 0;
+ struct gf_sock_incoming *in = NULL;
+ 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;
- switch (priv->incoming.frag.call_body.request.vector_state) {
+ /* used to reduce the indirection */
+ in = &priv->incoming;
+ frag = &in->frag;
+ request = &frag->call_body.request;
+
+ switch (request->vector_state) {
case SP_STATE_VECTORED_REQUEST_INIT:
- priv->incoming.frag.call_body.request.vector_sizer_state = 0;
- addr = rpc_cred_addr (iobuf_ptr (priv->incoming.iobuf));
+ request->vector_sizer_state = 0;
+
+ addr = rpc_cred_addr (iobuf_ptr (in->iobuf));
/* also read verf flavour and verflen */
credlen = ntoh32 (*((uint32_t *)addr))
@@ -832,102 +1228,114 @@ __socket_read_vectored_request (rpc_transport_t *this, rpcsvc_vector_sizer vecto
__socket_proto_init_pending (priv, credlen);
- priv->incoming.frag.call_body.request.vector_state =
- SP_STATE_READING_CREDBYTES;
+ request->vector_state = SP_STATE_READING_CREDBYTES;
/* fall through */
case SP_STATE_READING_CREDBYTES:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.request.vector_state =
- SP_STATE_READ_CREDBYTES;
+ request->vector_state = SP_STATE_READ_CREDBYTES;
/* fall through */
case SP_STATE_READ_CREDBYTES:
- addr = rpc_verf_addr (priv->incoming.frag.fragcurrent);
+ addr = rpc_verf_addr (frag->fragcurrent);
verflen = ntoh32 (*((uint32_t *)addr));
if (verflen == 0) {
- priv->incoming.frag.call_body.request.vector_state
- = SP_STATE_READ_VERFBYTES;
+ request->vector_state = SP_STATE_READ_VERFBYTES;
goto sp_state_read_verfbytes;
}
__socket_proto_init_pending (priv, verflen);
- priv->incoming.frag.call_body.request.vector_state
- = SP_STATE_READING_VERFBYTES;
+ request->vector_state = SP_STATE_READING_VERFBYTES;
/* fall through */
case SP_STATE_READING_VERFBYTES:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.request.vector_state =
- SP_STATE_READ_VERFBYTES;
+ request->vector_state = SP_STATE_READ_VERFBYTES;
/* fall through */
case SP_STATE_READ_VERFBYTES:
sp_state_read_verfbytes:
- priv->incoming.frag.call_body.request.vector_sizer_state =
- vector_sizer (priv->incoming.frag.call_body.request.vector_sizer_state,
- &readsize,
- priv->incoming.frag.fragcurrent);
+ /* set the base_addr 'persistently' across multiple calls
+ into the state machine */
+ in->proghdr_base_addr = frag->fragcurrent;
+
+ request->vector_sizer_state =
+ vector_sizer (request->vector_sizer_state,
+ &readsize, in->proghdr_base_addr,
+ frag->fragcurrent);
__socket_proto_init_pending (priv, readsize);
- priv->incoming.frag.call_body.request.vector_state
- = SP_STATE_READING_PROGHDR;
+
+ request->vector_state = SP_STATE_READING_PROGHDR;
/* fall through */
case SP_STATE_READING_PROGHDR:
__socket_proto_read (priv, ret);
-sp_state_reading_proghdr:
- priv->incoming.frag.call_body.request.vector_sizer_state =
- vector_sizer (priv->incoming.frag.call_body.request.vector_sizer_state,
- &readsize,
- priv->incoming.frag.fragcurrent);
+
+ request->vector_state = SP_STATE_READ_PROGHDR;
+
+ /* fall through */
+
+ case SP_STATE_READ_PROGHDR:
+sp_state_read_proghdr:
+ request->vector_sizer_state =
+ vector_sizer (request->vector_sizer_state,
+ &readsize, in->proghdr_base_addr,
+ frag->fragcurrent);
if (readsize == 0) {
- priv->incoming.frag.call_body.request.vector_state =
- SP_STATE_READ_PROGHDR;
- } else {
- __socket_proto_init_pending (priv, readsize);
- __socket_proto_read (priv, ret);
- goto sp_state_reading_proghdr;
+ request->vector_state = SP_STATE_READ_PROGHDR_XDATA;
+ goto sp_state_read_proghdr_xdata;
}
- case SP_STATE_READ_PROGHDR:
- if (priv->incoming.payload_vector.iov_base == NULL) {
+ __socket_proto_init_pending (priv, readsize);
- size = RPC_FRAGSIZE (priv->incoming.fraghdr) -
- priv->incoming.frag.bytes_read;
+ request->vector_state = SP_STATE_READING_PROGHDR_XDATA;
+
+ /* fall through */
+
+ case SP_STATE_READING_PROGHDR_XDATA:
+ __socket_proto_read (priv, ret);
+
+ request->vector_state = SP_STATE_READ_PROGHDR;
+ /* check if the vector_sizer() has more to say */
+ goto sp_state_read_proghdr;
+
+ case SP_STATE_READ_PROGHDR_XDATA:
+sp_state_read_proghdr_xdata:
+ if (in->payload_vector.iov_base == NULL) {
+
+ size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
iobuf = iobuf_get2 (this->ctx->iobuf_pool, size);
if (!iobuf) {
ret = -1;
break;
}
- if (priv->incoming.iobref == NULL) {
- priv->incoming.iobref = iobref_new ();
- if (priv->incoming.iobref == NULL) {
+ if (in->iobref == NULL) {
+ in->iobref = iobref_new ();
+ if (in->iobref == NULL) {
ret = -1;
iobuf_unref (iobuf);
break;
}
}
- iobref_add (priv->incoming.iobref, iobuf);
+ iobref_add (in->iobref, iobuf);
iobuf_unref (iobuf);
- priv->incoming.payload_vector.iov_base
- = iobuf_ptr (iobuf);
+ in->payload_vector.iov_base = iobuf_ptr (iobuf);
- priv->incoming.frag.fragcurrent = iobuf_ptr (iobuf);
+ frag->fragcurrent = iobuf_ptr (iobuf);
}
- priv->incoming.frag.call_body.request.vector_state =
- SP_STATE_READING_PROG;
+ request->vector_state = SP_STATE_READING_PROG;
/* fall through */
@@ -938,19 +1346,15 @@ sp_state_reading_proghdr:
ret = __socket_read_simple_msg (this);
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
- if ((ret == -1)
- || ((ret == 0)
- && (remaining_size == 0)
- && RPC_LASTFRAG (priv->incoming.fraghdr))) {
- priv->incoming.frag.call_body.request.vector_state
- = SP_STATE_VECTORED_REQUEST_INIT;
- priv->incoming.payload_vector.iov_len
- = (unsigned long)priv->incoming.frag.fragcurrent
- - (unsigned long)
- priv->incoming.payload_vector.iov_base;
+ if ((ret == -1) ||
+ ((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)in->payload_vector.iov_base);
}
break;
}
@@ -959,7 +1363,7 @@ out:
return ret;
}
-inline int
+static inline int
__socket_read_request (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -968,46 +1372,53 @@ __socket_read_request (rpc_transport_t *this)
int ret = -1;
char *buf = NULL;
rpcsvc_vector_sizer vector_sizer = NULL;
+ struct gf_sock_incoming *in = NULL;
+ 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;
- switch (priv->incoming.frag.call_body.request.header_state) {
+ /* used to reduce the indirection */
+ in = &priv->incoming;
+ frag = &in->frag;
+ request = &frag->call_body.request;
+
+ switch (request->header_state) {
case SP_STATE_REQUEST_HEADER_INIT:
__socket_proto_init_pending (priv, RPC_CALL_BODY_SIZE);
- priv->incoming.frag.call_body.request.header_state
- = SP_STATE_READING_RPCHDR1;
+ request->header_state = SP_STATE_READING_RPCHDR1;
/* fall through */
case SP_STATE_READING_RPCHDR1:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.request.header_state =
- SP_STATE_READ_RPCHDR1;
+ request->header_state = SP_STATE_READ_RPCHDR1;
/* fall through */
case SP_STATE_READ_RPCHDR1:
- buf = rpc_prognum_addr (iobuf_ptr (priv->incoming.iobuf));
+ buf = rpc_prognum_addr (iobuf_ptr (in->iobuf));
prognum = ntoh32 (*((uint32_t *)buf));
- buf = rpc_progver_addr (iobuf_ptr (priv->incoming.iobuf));
+ buf = rpc_progver_addr (iobuf_ptr (in->iobuf));
progver = ntoh32 (*((uint32_t *)buf));
- buf = rpc_procnum_addr (iobuf_ptr (priv->incoming.iobuf));
+ buf = rpc_procnum_addr (iobuf_ptr (in->iobuf));
procnum = ntoh32 (*((uint32_t *)buf));
- if (this->listener) {
- /* this check is needed as rpcsvc and rpc-clnt actor structures are
- * not same */
- vector_sizer = rpcsvc_get_program_vector_sizer ((rpcsvc_t *)this->mydata,
- prognum, progver, procnum);
+ if (priv->is_server) {
+ /* this check is needed as rpcsvc and rpc-clnt
+ * actor structures are not same */
+ vector_sizer =
+ rpcsvc_get_program_vector_sizer ((rpcsvc_t *)this->mydata,
+ prognum, progver, procnum);
}
if (vector_sizer) {
@@ -1016,15 +1427,13 @@ __socket_read_request (rpc_transport_t *this)
ret = __socket_read_simple_request (this);
}
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
if ((ret == -1)
|| ((ret == 0)
&& (remaining_size == 0)
- && (RPC_LASTFRAG (priv->incoming.fraghdr)))) {
- priv->incoming.frag.call_body.request.header_state =
- SP_STATE_REQUEST_HEADER_INIT;
+ && (RPC_LASTFRAG (in->fraghdr)))) {
+ request->header_state = SP_STATE_REQUEST_HEADER_INIT;
}
break;
@@ -1035,36 +1444,40 @@ out:
}
-inline int
+static inline int
__socket_read_accepted_successful_reply (rpc_transport_t *this)
{
- socket_private_t *priv = NULL;
- int ret = 0;
- struct iobuf *iobuf = NULL;
- uint32_t gluster_read_rsp_hdr_len = 0;
- gfs3_read_rsp read_rsp = {0, };
- size_t size = 0;
+ socket_private_t *priv = NULL;
+ int ret = 0;
+ struct iobuf *iobuf = NULL;
+ gfs3_read_rsp read_rsp = {0, };
+ ssize_t size = 0;
+ ssize_t default_read_size = 0;
+ char *proghdr_buf = NULL;
+ 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);
priv = this->private;
- switch (priv->incoming.frag.call_body.reply.accepted_success_state) {
+ /* used to reduce the indirection */
+ in = &priv->incoming;
+ frag = &in->frag;
+
+ switch (frag->call_body.reply.accepted_success_state) {
case SP_STATE_ACCEPTED_SUCCESS_REPLY_INIT:
- gluster_read_rsp_hdr_len = xdr_sizeof ((xdrproc_t) xdr_gfs3_read_rsp,
- &read_rsp);
+ default_read_size = xdr_sizeof ((xdrproc_t) xdr_gfs3_read_rsp,
+ &read_rsp);
- if (gluster_read_rsp_hdr_len == 0) {
- gf_log (this->name, GF_LOG_ERROR,
- "xdr_sizeof on gfs3_read_rsp failed");
- ret = -1;
- goto out;
- }
- __socket_proto_init_pending (priv, gluster_read_rsp_hdr_len);
+ proghdr_buf = frag->fragcurrent;
+
+ __socket_proto_init_pending (priv, default_read_size);
- priv->incoming.frag.call_body.reply.accepted_success_state
+ frag->call_body.reply.accepted_success_state
= SP_STATE_READING_PROC_HEADER;
/* fall through */
@@ -1072,13 +1485,42 @@ __socket_read_accepted_successful_reply (rpc_transport_t *this)
case SP_STATE_READING_PROC_HEADER:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.reply.accepted_success_state
- = SP_STATE_READ_PROC_HEADER;
+ /* there can be 'xdata' in read response, figure it out */
+ xdrmem_create (&xdr, proghdr_buf, default_read_size,
+ XDR_DECODE);
- if (priv->incoming.payload_vector.iov_base == NULL) {
+ /* This will fail if there is xdata sent from server, if not,
+ well and good, we don't need to worry about */
+ xdr_gfs3_read_rsp (&xdr, &read_rsp);
- size = (RPC_FRAGSIZE (priv->incoming.fraghdr) -
- priv->incoming.frag.bytes_read);
+ free (read_rsp.xdata.xdata_val);
+
+ /* need to round off to proper roof (%4), as XDR packing pads
+ the end of opaque object with '0' */
+ size = roof (read_rsp.xdata.xdata_len, 4);
+
+ if (!size) {
+ frag->call_body.reply.accepted_success_state
+ = SP_STATE_READ_PROC_OPAQUE;
+ goto read_proc_opaque;
+ }
+
+ __socket_proto_init_pending (priv, size);
+
+ frag->call_body.reply.accepted_success_state
+ = SP_STATE_READING_PROC_OPAQUE;
+
+ case SP_STATE_READING_PROC_OPAQUE:
+ __socket_proto_read (priv, ret);
+
+ frag->call_body.reply.accepted_success_state
+ = SP_STATE_READ_PROC_OPAQUE;
+
+ case SP_STATE_READ_PROC_OPAQUE:
+ read_proc_opaque:
+ if (in->payload_vector.iov_base == NULL) {
+
+ size = (RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read);
iobuf = iobuf_get2 (this->ctx->iobuf_pool, size);
if (iobuf == NULL) {
@@ -1086,26 +1528,27 @@ __socket_read_accepted_successful_reply (rpc_transport_t *this)
goto out;
}
- if (priv->incoming.iobref == NULL) {
- priv->incoming.iobref = iobref_new ();
- if (priv->incoming.iobref == NULL) {
+ if (in->iobref == NULL) {
+ in->iobref = iobref_new ();
+ if (in->iobref == NULL) {
ret = -1;
iobuf_unref (iobuf);
goto out;
}
}
- iobref_add (priv->incoming.iobref, iobuf);
+ iobref_add (in->iobref, iobuf);
iobuf_unref (iobuf);
- priv->incoming.payload_vector.iov_base
- = iobuf_ptr (iobuf);
+ in->payload_vector.iov_base = iobuf_ptr (iobuf);
- priv->incoming.payload_vector.iov_len = size;
+ in->payload_vector.iov_len = size;
}
- priv->incoming.frag.fragcurrent
- = priv->incoming.payload_vector.iov_base;
+ frag->fragcurrent = in->payload_vector.iov_base;
+
+ frag->call_body.reply.accepted_success_state
+ = SP_STATE_READ_PROC_HEADER;
/* fall through */
@@ -1113,9 +1556,8 @@ __socket_read_accepted_successful_reply (rpc_transport_t *this)
/* now read the entire remaining msg into new iobuf */
ret = __socket_read_simple_msg (this);
if ((ret == -1)
- || ((ret == 0)
- && RPC_LASTFRAG (priv->incoming.fraghdr))) {
- priv->incoming.frag.call_body.reply.accepted_success_state
+ || ((ret == 0) && RPC_LASTFRAG (in->fraghdr))) {
+ frag->call_body.reply.accepted_success_state
= SP_STATE_ACCEPTED_SUCCESS_REPLY_INIT;
}
@@ -1129,7 +1571,7 @@ out:
#define rpc_reply_verflen_addr(fragcurrent) ((char *)fragcurrent - 4)
#define rpc_reply_accept_status_addr(fragcurrent) ((char *)fragcurrent - 4)
-inline int
+static inline int
__socket_read_accepted_reply (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -1137,19 +1579,24 @@ __socket_read_accepted_reply (rpc_transport_t *this)
char *buf = NULL;
uint32_t verflen = 0, len = 0;
uint32_t remaining_size = 0;
+ 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;
+ frag = &in->frag;
- switch (priv->incoming.frag.call_body.reply.accepted_state) {
+ switch (frag->call_body.reply.accepted_state) {
case SP_STATE_ACCEPTED_REPLY_INIT:
__socket_proto_init_pending (priv,
RPC_AUTH_FLAVOUR_N_LENGTH_SIZE);
- priv->incoming.frag.call_body.reply.accepted_state
+ frag->call_body.reply.accepted_state
= SP_STATE_READING_REPLY_VERFLEN;
/* fall through */
@@ -1157,13 +1604,13 @@ __socket_read_accepted_reply (rpc_transport_t *this)
case SP_STATE_READING_REPLY_VERFLEN:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.reply.accepted_state
+ frag->call_body.reply.accepted_state
= SP_STATE_READ_REPLY_VERFLEN;
/* fall through */
case SP_STATE_READ_REPLY_VERFLEN:
- buf = rpc_reply_verflen_addr (priv->incoming.frag.fragcurrent);
+ buf = rpc_reply_verflen_addr (frag->fragcurrent);
verflen = ntoh32 (*((uint32_t *) buf));
@@ -1172,7 +1619,7 @@ __socket_read_accepted_reply (rpc_transport_t *this)
__socket_proto_init_pending (priv, len);
- priv->incoming.frag.call_body.reply.accepted_state
+ frag->call_body.reply.accepted_state
= SP_STATE_READING_REPLY_VERFBYTES;
/* fall through */
@@ -1180,19 +1627,19 @@ __socket_read_accepted_reply (rpc_transport_t *this)
case SP_STATE_READING_REPLY_VERFBYTES:
__socket_proto_read (priv, ret);
- priv->incoming.frag.call_body.reply.accepted_state
+ frag->call_body.reply.accepted_state
= SP_STATE_READ_REPLY_VERFBYTES;
- buf = rpc_reply_accept_status_addr (priv->incoming.frag.fragcurrent);
+ buf = rpc_reply_accept_status_addr (frag->fragcurrent);
- priv->incoming.frag.call_body.reply.accept_status
+ frag->call_body.reply.accept_status
= ntoh32 (*(uint32_t *) buf);
/* fall through */
case SP_STATE_READ_REPLY_VERFBYTES:
- if (priv->incoming.frag.call_body.reply.accept_status
+ if (frag->call_body.reply.accept_status
== SUCCESS) {
ret = __socket_read_accepted_successful_reply (this);
} else {
@@ -1202,14 +1649,13 @@ __socket_read_accepted_reply (rpc_transport_t *this)
ret = __socket_read_simple_msg (this);
}
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr)
+ - frag->bytes_read;
if ((ret == -1)
- || ((ret == 0)
- && (remaining_size == 0)
- && (RPC_LASTFRAG (priv->incoming.fraghdr)))) {
- priv->incoming.frag.call_body.reply.accepted_state
+ || ((ret == 0) && (remaining_size == 0)
+ && (RPC_LASTFRAG (in->fraghdr)))) {
+ frag->call_body.reply.accepted_state
= SP_STATE_ACCEPTED_REPLY_INIT;
}
@@ -1221,7 +1667,7 @@ out:
}
-inline int
+static inline int
__socket_read_denied_reply (rpc_transport_t *this)
{
return __socket_read_simple_msg (this);
@@ -1231,25 +1677,29 @@ __socket_read_denied_reply (rpc_transport_t *this)
#define rpc_reply_status_addr(fragcurrent) ((char *)fragcurrent - 4)
-inline int
+static inline int
__socket_read_vectored_reply (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
int ret = 0;
char *buf = NULL;
uint32_t remaining_size = 0;
+ 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;
- switch (priv->incoming.frag.call_body.reply.status_state) {
+ switch (frag->call_body.reply.status_state) {
case SP_STATE_ACCEPTED_REPLY_INIT:
__socket_proto_init_pending (priv, RPC_REPLY_STATUS_SIZE);
- priv->incoming.frag.call_body.reply.status_state
+ frag->call_body.reply.status_state
= SP_STATE_READING_REPLY_STATUS;
/* fall through */
@@ -1257,37 +1707,33 @@ __socket_read_vectored_reply (rpc_transport_t *this)
case SP_STATE_READING_REPLY_STATUS:
__socket_proto_read (priv, ret);
- buf = rpc_reply_status_addr (priv->incoming.frag.fragcurrent);
+ buf = rpc_reply_status_addr (frag->fragcurrent);
- priv->incoming.frag.call_body.reply.accept_status
+ frag->call_body.reply.accept_status
= ntoh32 (*((uint32_t *) buf));
- priv->incoming.frag.call_body.reply.status_state
+ frag->call_body.reply.status_state
= SP_STATE_READ_REPLY_STATUS;
/* fall through */
case SP_STATE_READ_REPLY_STATUS:
- if (priv->incoming.frag.call_body.reply.accept_status
- == MSG_ACCEPTED) {
+ if (frag->call_body.reply.accept_status == MSG_ACCEPTED) {
ret = __socket_read_accepted_reply (this);
} else {
ret = __socket_read_denied_reply (this);
}
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
if ((ret == -1)
- || ((ret == 0)
- && (remaining_size == 0)
- && (RPC_LASTFRAG (priv->incoming.fraghdr)))) {
- priv->incoming.frag.call_body.reply.status_state
+ || ((ret == 0) && (remaining_size == 0)
+ && (RPC_LASTFRAG (in->fraghdr)))) {
+ frag->call_body.reply.status_state
= SP_STATE_ACCEPTED_REPLY_INIT;
- priv->incoming.payload_vector.iov_len
- = (unsigned long)priv->incoming.frag.fragcurrent
- - (unsigned long)
- priv->incoming.payload_vector.iov_base;
+ in->payload_vector.iov_len
+ = (unsigned long)frag->fragcurrent
+ - (unsigned long)in->payload_vector.iov_base;
}
break;
}
@@ -1297,7 +1743,7 @@ out:
}
-inline int
+static inline int
__socket_read_simple_reply (rpc_transport_t *this)
{
return __socket_read_simple_msg (this);
@@ -1305,7 +1751,7 @@ __socket_read_simple_reply (rpc_transport_t *this)
#define rpc_xid_addr(buf) (buf)
-inline int
+static inline int
__socket_read_reply (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -1313,26 +1759,29 @@ __socket_read_reply (rpc_transport_t *this)
int32_t ret = -1;
rpc_request_info_t *request_info = NULL;
char map_xid = 0;
+ 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;
- buf = rpc_xid_addr (iobuf_ptr (priv->incoming.iobuf));
+ buf = rpc_xid_addr (iobuf_ptr (in->iobuf));
- if (priv->incoming.request_info == NULL) {
- priv->incoming.request_info = GF_CALLOC (1,
- sizeof (*request_info),
- gf_common_mt_rpc_trans_reqinfo_t);
- if (priv->incoming.request_info == NULL) {
+ if (in->request_info == NULL) {
+ in->request_info = GF_CALLOC (1, sizeof (*request_info),
+ gf_common_mt_rpc_trans_reqinfo_t);
+ if (in->request_info == NULL) {
goto out;
}
map_xid = 1;
}
- request_info = priv->incoming.request_info;
+ request_info = in->request_info;
if (map_xid) {
request_info->xid = ntoh32 (*((uint32_t *) buf));
@@ -1344,7 +1793,7 @@ __socket_read_reply (rpc_transport_t *this)
{
ret = rpc_transport_notify (this,
RPC_TRANSPORT_MAP_XID_REQUEST,
- priv->incoming.request_info);
+ in->request_info);
}
pthread_mutex_lock (&priv->lock);
@@ -1355,13 +1804,11 @@ __socket_read_reply (rpc_transport_t *this)
}
}
- if ((request_info->prognum == GLUSTER3_1_FOP_PROGRAM)
+ if ((request_info->prognum == GLUSTER_FOP_PROGRAM)
&& (request_info->procnum == GF_FOP_READ)) {
if (map_xid && request_info->rsp.rsp_payload_count != 0) {
- priv->incoming.iobref
- = iobref_ref (request_info->rsp.rsp_iobref);
- priv->incoming.payload_vector
- = *request_info->rsp.rsp_payload;
+ in->iobref = iobref_ref (request_info->rsp.rsp_iobref);
+ in->payload_vector = *request_info->rsp.rsp_payload;
}
ret = __socket_read_vectored_reply (this);
@@ -1374,42 +1821,47 @@ out:
/* returns the number of bytes yet to be read in a fragment */
-inline int
+static inline int
__socket_read_frag (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
int32_t ret = 0;
char *buf = NULL;
uint32_t remaining_size = 0;
+ 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;
+ frag = &in->frag;
- switch (priv->incoming.frag.state) {
+ switch (frag->state) {
case SP_STATE_NADA:
__socket_proto_init_pending (priv, RPC_MSGTYPE_SIZE);
- priv->incoming.frag.state = SP_STATE_READING_MSGTYPE;
+ frag->state = SP_STATE_READING_MSGTYPE;
/* fall through */
case SP_STATE_READING_MSGTYPE:
__socket_proto_read (priv, ret);
- priv->incoming.frag.state = SP_STATE_READ_MSGTYPE;
+ frag->state = SP_STATE_READ_MSGTYPE;
/* fall through */
case SP_STATE_READ_MSGTYPE:
- buf = rpc_msgtype_addr (iobuf_ptr (priv->incoming.iobuf));
- priv->incoming.msg_type = ntoh32 (*((uint32_t *)buf));
+ buf = rpc_msgtype_addr (iobuf_ptr (in->iobuf));
+ in->msg_type = ntoh32 (*((uint32_t *)buf));
- if (priv->incoming.msg_type == CALL) {
+ if (in->msg_type == CALL) {
ret = __socket_read_request (this);
- } else if (priv->incoming.msg_type == REPLY) {
+ } else if (in->msg_type == REPLY) {
ret = __socket_read_reply (this);
- } else if (priv->incoming.msg_type == GF_UNIVERSAL_ANSWER) {
+ } else if (in->msg_type == GF_UNIVERSAL_ANSWER) {
gf_log ("rpc", GF_LOG_ERROR,
"older version of protocol/process trying to "
"connect from %s. use newer version on that node",
@@ -1417,19 +1869,17 @@ __socket_read_frag (rpc_transport_t *this)
} else {
gf_log ("rpc", GF_LOG_ERROR,
"wrong MSG-TYPE (%d) received from %s",
- priv->incoming.msg_type,
+ in->msg_type,
this->peerinfo.identifier);
ret = -1;
}
- remaining_size = RPC_FRAGSIZE (priv->incoming.fraghdr)
- - priv->incoming.frag.bytes_read;
+ remaining_size = RPC_FRAGSIZE (in->fraghdr) - frag->bytes_read;
if ((ret == -1)
- || ((ret == 0)
- && (remaining_size == 0)
- && (RPC_LASTFRAG (priv->incoming.fraghdr)))) {
- priv->incoming.frag.state = SP_STATE_NADA;
+ || ((ret == 0) && (remaining_size == 0)
+ && (RPC_LASTFRAG (in->fraghdr)))) {
+ frag->state = SP_STATE_NADA;
}
break;
@@ -1440,31 +1890,36 @@ out:
}
-inline
+static inline
void __socket_reset_priv (socket_private_t *priv)
{
- if (priv->incoming.iobref) {
- iobref_unref (priv->incoming.iobref);
- priv->incoming.iobref = NULL;
+ struct gf_sock_incoming *in = NULL;
+
+ /* used to reduce the indirection */
+ in = &priv->incoming;
+
+ if (in->iobref) {
+ iobref_unref (in->iobref);
+ in->iobref = NULL;
}
- if (priv->incoming.iobuf) {
- iobuf_unref (priv->incoming.iobuf);
+ if (in->iobuf) {
+ iobuf_unref (in->iobuf);
}
- if (priv->incoming.request_info != NULL) {
- GF_FREE (priv->incoming.request_info);
- priv->incoming.request_info = NULL;
+ if (in->request_info != NULL) {
+ GF_FREE (in->request_info);
+ in->request_info = NULL;
}
- memset (&priv->incoming.payload_vector, 0,
- sizeof (priv->incoming.payload_vector));
+ memset (&in->payload_vector, 0,
+ sizeof (in->payload_vector));
- priv->incoming.iobuf = NULL;
+ in->iobuf = NULL;
}
-int
+static int
__socket_proto_state_machine (rpc_transport_t *this,
rpc_transport_pollin_t **pollin)
{
@@ -1473,46 +1928,40 @@ __socket_proto_state_machine (rpc_transport_t *this,
struct iobuf *iobuf = NULL;
struct iobref *iobref = NULL;
struct iovec vector[2];
+ 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;
- while (priv->incoming.record_state != SP_STATE_COMPLETE) {
- switch (priv->incoming.record_state) {
+ /* used to reduce the indirection */
+ in = &priv->incoming;
+ frag = &in->frag;
+
+ while (in->record_state != SP_STATE_COMPLETE) {
+ switch (in->record_state) {
case SP_STATE_NADA:
- priv->incoming.total_bytes_read = 0;
- priv->incoming.payload_vector.iov_len = 0;
+ in->total_bytes_read = 0;
+ in->payload_vector.iov_len = 0;
- priv->incoming.pending_vector = priv->incoming.vector;
- priv->incoming.pending_vector->iov_base =
- &priv->incoming.fraghdr;
+ in->pending_vector = in->vector;
+ in->pending_vector->iov_base = &in->fraghdr;
- priv->incoming.pending_vector->iov_len =
- sizeof (priv->incoming.fraghdr);
+ in->pending_vector->iov_len = sizeof (in->fraghdr);
- priv->incoming.record_state = SP_STATE_READING_FRAGHDR;
+ in->record_state = SP_STATE_READING_FRAGHDR;
/* fall through */
case SP_STATE_READING_FRAGHDR:
- ret = __socket_readv (this,
- priv->incoming.pending_vector, 1,
- &priv->incoming.pending_vector,
- &priv->incoming.pending_count,
+ ret = __socket_readv (this, in->pending_vector, 1,
+ &in->pending_vector,
+ &in->pending_count,
NULL);
- if (ret == -1) {
- if (priv->read_fail_log == 1) {
- gf_log (this->name,
- ((priv->connected == 1) ?
- GF_LOG_WARNING : GF_LOG_DEBUG),
- "reading from socket failed. Error (%s)"
- ", peer (%s)", strerror (errno),
- this->peerinfo.identifier);
- }
+ if (ret == -1)
goto out;
- }
if (ret > 0) {
gf_log (this->name, GF_LOG_TRACE, "partial "
@@ -1521,44 +1970,40 @@ __socket_proto_state_machine (rpc_transport_t *this,
}
if (ret == 0) {
- priv->incoming.record_state =
- SP_STATE_READ_FRAGHDR;
+ in->record_state = SP_STATE_READ_FRAGHDR;
}
/* fall through */
case SP_STATE_READ_FRAGHDR:
- priv->incoming.fraghdr = ntoh32 (priv->incoming.fraghdr);
- priv->incoming.record_state = SP_STATE_READING_FRAG;
- priv->incoming.total_bytes_read
- += RPC_FRAGSIZE(priv->incoming.fraghdr);
+ in->fraghdr = ntoh32 (in->fraghdr);
+ in->total_bytes_read += RPC_FRAGSIZE(in->fraghdr);
iobuf = iobuf_get2 (this->ctx->iobuf_pool,
- priv->incoming.total_bytes_read +
- sizeof (priv->incoming.fraghdr));
+ (in->total_bytes_read +
+ sizeof (in->fraghdr)));
if (!iobuf) {
ret = -ENOMEM;
goto out;
}
- priv->incoming.iobuf = iobuf;
- priv->incoming.iobuf_size = 0;
- priv->incoming.frag.fragcurrent = iobuf_ptr (iobuf);
+ in->iobuf = iobuf;
+ in->iobuf_size = 0;
+ frag->fragcurrent = iobuf_ptr (iobuf);
+ in->record_state = SP_STATE_READING_FRAG;
/* fall through */
case SP_STATE_READING_FRAG:
ret = __socket_read_frag (this);
- if ((ret == -1)
- || (priv->incoming.frag.bytes_read !=
- RPC_FRAGSIZE (priv->incoming.fraghdr))) {
+ if ((ret == -1) ||
+ (frag->bytes_read != RPC_FRAGSIZE (in->fraghdr))) {
goto out;
}
- priv->incoming.frag.bytes_read = 0;
+ frag->bytes_read = 0;
- if (!RPC_LASTFRAG (priv->incoming.fraghdr)) {
- priv->incoming.record_state =
- SP_STATE_READING_FRAGHDR;
+ if (!RPC_LASTFRAG (in->fraghdr)) {
+ in->record_state = SP_STATE_READING_FRAGHDR;
break;
}
@@ -1567,44 +2012,39 @@ __socket_proto_state_machine (rpc_transport_t *this,
*/
if (pollin != NULL) {
int count = 0;
- priv->incoming.iobuf_size
- = priv->incoming.total_bytes_read
- - priv->incoming.payload_vector.iov_len;
+ in->iobuf_size = (in->total_bytes_read -
+ in->payload_vector.iov_len);
memset (vector, 0, sizeof (vector));
- if (priv->incoming.iobref == NULL) {
- priv->incoming.iobref = iobref_new ();
- if (priv->incoming.iobref == NULL) {
+ if (in->iobref == NULL) {
+ in->iobref = iobref_new ();
+ if (in->iobref == NULL) {
ret = -1;
goto out;
}
}
- vector[count].iov_base
- = iobuf_ptr (priv->incoming.iobuf);
- vector[count].iov_len
- = priv->incoming.iobuf_size;
+ vector[count].iov_base = iobuf_ptr (in->iobuf);
+ vector[count].iov_len = in->iobuf_size;
- iobref = priv->incoming.iobref;
+ iobref = in->iobref;
count++;
- if (priv->incoming.payload_vector.iov_base
- != NULL) {
- vector[count]
- = priv->incoming.payload_vector;
+ if (in->payload_vector.iov_base != NULL) {
+ vector[count] = in->payload_vector;
count++;
}
*pollin = rpc_transport_pollin_alloc (this,
vector,
count,
- priv->incoming.iobuf,
+ in->iobuf,
iobref,
- priv->incoming.request_info);
- iobuf_unref (priv->incoming.iobuf);
- priv->incoming.iobuf = NULL;
+ in->request_info);
+ iobuf_unref (in->iobuf);
+ in->iobuf = NULL;
if (*pollin == NULL) {
gf_log (this->name, GF_LOG_WARNING,
@@ -1612,12 +2052,12 @@ __socket_proto_state_machine (rpc_transport_t *this,
ret = -1;
goto out;
}
- if (priv->incoming.msg_type == REPLY)
+ if (in->msg_type == REPLY)
(*pollin)->is_reply = 1;
- priv->incoming.request_info = NULL;
+ in->request_info = NULL;
}
- priv->incoming.record_state = SP_STATE_COMPLETE;
+ in->record_state = SP_STATE_COMPLETE;
break;
case SP_STATE_COMPLETE:
@@ -1629,8 +2069,8 @@ __socket_proto_state_machine (rpc_transport_t *this,
}
}
- if (priv->incoming.record_state == SP_STATE_COMPLETE) {
- priv->incoming.record_state = SP_STATE_NADA;
+ if (in->record_state == SP_STATE_COMPLETE) {
+ in->record_state = SP_STATE_NADA;
__socket_reset_priv (priv);
}
@@ -1642,7 +2082,7 @@ out:
}
-int
+static int
socket_proto_state_machine (rpc_transport_t *this,
rpc_transport_pollin_t **pollin)
{
@@ -1665,18 +2105,22 @@ out:
}
-int
+static int
socket_event_poll_in (rpc_transport_t *this)
{
int ret = -1;
rpc_transport_pollin_t *pollin = NULL;
+ socket_private_t *priv = this->private;
ret = socket_proto_state_machine (this, &pollin);
if (pollin != NULL) {
+ priv->ot_state = OT_CALLBACK;
ret = rpc_transport_notify (this, RPC_TRANSPORT_MSG_RECEIVED,
pollin);
-
+ if (priv->ot_state == OT_CALLBACK) {
+ priv->ot_state = OT_RUNNING;
+ }
rpc_transport_pollin_destroy (pollin);
}
@@ -1684,7 +2128,7 @@ socket_event_poll_in (rpc_transport_t *this)
}
-int
+static int
socket_connect_finish (rpc_transport_t *this)
{
int ret = -1;
@@ -1699,9 +2143,11 @@ socket_connect_finish (rpc_transport_t *this)
pthread_mutex_lock (&priv->lock);
{
- if (priv->connected)
+ if (priv->connected != 0)
goto unlock;
+ get_transport_identifiers (this);
+
ret = __socket_connect_finish (priv->sock);
if (ret == -1 && errno == EINPROGRESS)
@@ -1716,8 +2162,6 @@ socket_connect_finish (rpc_transport_t *this)
priv->connect_finish_log = 1;
}
__socket_disconnect (this);
- notify_rpc = 1;
- event = RPC_TRANSPORT_DISCONNECT;
goto unlock;
}
@@ -1742,7 +2186,6 @@ socket_connect_finish (rpc_transport_t *this)
priv->connected = 1;
priv->connect_finish_log = 0;
event = RPC_TRANSPORT_CONNECT;
- get_transport_identifiers (this);
}
}
unlock:
@@ -1757,13 +2200,13 @@ out:
/* reads rpc_requests during pollin */
-int
+static int
socket_event_handler (int fd, int idx, void *data,
int poll_in, int poll_out, int poll_err)
{
- rpc_transport_t *this = NULL;
+ rpc_transport_t *this = NULL;
socket_private_t *priv = NULL;
- int ret = 0;
+ int ret = -1;
this = data;
GF_VALIDATE_OR_GOTO ("socket", this, out);
@@ -1773,16 +2216,13 @@ socket_event_handler (int fd, int idx, void *data,
THIS = this->xl;
priv = this->private;
-
pthread_mutex_lock (&priv->lock);
{
priv->idx = idx;
}
pthread_mutex_unlock (&priv->lock);
- if (!priv->connected) {
- ret = socket_connect_finish (this);
- }
+ ret = (priv->connected == 1) ? 0 : socket_connect_finish(this);
if (!ret && poll_out) {
ret = socket_event_poll_out (this);
@@ -1798,14 +2238,198 @@ socket_event_handler (int fd, int idx, void *data,
"disconnecting now");
socket_event_poll_err (this);
rpc_transport_unref (this);
- }
+ }
out:
- return 0;
+ return ret;
}
-int
+static void *
+socket_poller (void *ctx)
+{
+ rpc_transport_t *this = ctx;
+ socket_private_t *priv = this->private;
+ struct pollfd pfd[2] = {{0,},};
+ gf_boolean_t to_write = _gf_false;
+ int ret = 0;
+ uint32_t gen = 0;
+
+ priv->ot_state = OT_RUNNING;
+
+ if (priv->use_ssl) {
+ if (ssl_setup_connection(this,priv->connected) < 0) {
+ gf_log (this->name,GF_LOG_ERROR, "%s setup failed",
+ priv->connected ? "server" : "client");
+ goto err;
+ }
+ }
+
+ if (!priv->bio) {
+ ret = __socket_nonblock (priv->sock);
+ if (ret == -1) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "NBIO on %d failed (%s)",
+ priv->sock, strerror (errno));
+ goto err;
+ }
+ }
+
+ if (priv->connected == 0) {
+ THIS = this->xl;
+ ret = socket_connect_finish (this);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "asynchronous socket_connect_finish failed");
+ }
+ }
+
+ ret = rpc_transport_notify (this->listener,
+ RPC_TRANSPORT_ACCEPT, this);
+ if (ret != 0) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "asynchronous rpc_transport_notify failed");
+ }
+
+ gen = priv->ot_gen;
+ for (;;) {
+ pthread_mutex_lock(&priv->lock);
+ to_write = !list_empty(&priv->ioq);
+ pthread_mutex_unlock(&priv->lock);
+ pfd[0].fd = priv->pipe[0];
+ pfd[0].events = POLL_MASK_ERROR;
+ pfd[0].revents = 0;
+ pfd[1].fd = priv->sock;
+ pfd[1].events = POLL_MASK_INPUT | POLL_MASK_ERROR;
+ pfd[1].revents = 0;
+ if (to_write) {
+ pfd[1].events |= POLL_MASK_OUTPUT;
+ }
+ else {
+ pfd[0].events |= POLL_MASK_INPUT;
+ }
+ if (poll(pfd,2,-1) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,"poll failed");
+ break;
+ }
+ if (pfd[0].revents & POLL_MASK_ERROR) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "poll error on pipe");
+ break;
+ }
+ /* Only glusterd actually seems to need this. */
+ THIS = this->xl;
+ if (pfd[1].revents & POLL_MASK_INPUT) {
+ ret = socket_event_poll_in(this);
+ if (ret >= 0) {
+ /* Suppress errors while making progress. */
+ pfd[1].revents &= ~POLL_MASK_ERROR;
+ }
+ else if (errno == ENOTCONN) {
+ ret = 0;
+ }
+ if (priv->ot_state == OT_PLEASE_DIE) {
+ gf_log (this->name, GF_LOG_TRACE,
+ "OT_IDLE on %p (input request)",
+ this);
+ priv->ot_state = OT_IDLE;
+ break;
+ }
+ }
+ else if (pfd[1].revents & POLL_MASK_OUTPUT) {
+ ret = socket_event_poll_out(this);
+ if (ret >= 0) {
+ /* Suppress errors while making progress. */
+ pfd[1].revents &= ~POLL_MASK_ERROR;
+ }
+ else if (errno == ENOTCONN) {
+ ret = 0;
+ }
+ if (priv->ot_state == OT_PLEASE_DIE) {
+ gf_log (this->name, GF_LOG_TRACE,
+ "OT_IDLE on %p (output request)",
+ this);
+ priv->ot_state = OT_IDLE;
+ break;
+ }
+ }
+ else {
+ /*
+ * This usually means that we left poll() because
+ * somebody pushed a byte onto our pipe. That wakeup
+ * is why the pipe is there, but once awake we can do
+ * all the checking we need on the next iteration.
+ */
+ ret = 0;
+ }
+ if (pfd[1].revents & POLL_MASK_ERROR) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "poll error on socket");
+ break;
+ }
+ if (ret < 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "error in polling loop");
+ break;
+ }
+ if (priv->ot_gen != gen) {
+ gf_log (this->name, GF_LOG_TRACE,
+ "generation mismatch, my %u != %u",
+ gen, priv->ot_gen);
+ return NULL;
+ }
+ }
+
+err:
+ /* All (and only) I/O errors should come here. */
+ pthread_mutex_lock(&priv->lock);
+ if (priv->ssl_ssl) {
+ /*
+ * We're always responsible for this part, but only actually
+ * have to do it if we got far enough for ssl_ssl to be valid
+ * (i.e. errors in ssl_setup_connection don't count).
+ */
+ ssl_teardown_connection(priv);
+ }
+ __socket_shutdown(this);
+ close(priv->sock);
+ priv->sock = -1;
+ priv->ot_state = OT_IDLE;
+ pthread_mutex_unlock(&priv->lock);
+ rpc_transport_notify (this->listener, RPC_TRANSPORT_DISCONNECT,
+ this);
+ rpc_transport_unref (this);
+ return NULL;
+}
+
+
+static void
+socket_spawn (rpc_transport_t *this)
+{
+ socket_private_t *priv = this->private;
+
+ switch (priv->ot_state) {
+ case OT_IDLE:
+ case OT_PLEASE_DIE:
+ break;
+ default:
+ gf_log (this->name, GF_LOG_WARNING,
+ "refusing to start redundant poller");
+ return;
+ }
+
+ priv->ot_gen += 7;
+ priv->ot_state = OT_SPAWNING;
+ gf_log (this->name, GF_LOG_TRACE,
+ "spawning %p with gen %u", this, priv->ot_gen);
+
+ if (gf_thread_create(&priv->thread,NULL,socket_poller,this) != 0) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "could not create poll thread");
+ }
+}
+
+static int
socket_server_event_handler (int fd, int idx, void *data,
int poll_in, int poll_out, int poll_err)
{
@@ -1843,20 +2467,7 @@ socket_server_event_handler (int fd, int idx, void *data,
goto unlock;
}
- if (!priv->bio) {
- ret = __socket_nonblock (new_sock);
-
- if (ret == -1) {
- gf_log (this->name, GF_LOG_WARNING,
- "NBIO on %d failed (%s)",
- new_sock, strerror (errno));
-
- close (new_sock);
- goto unlock;
- }
- }
-
- if (priv->nodelay) {
+ if (priv->nodelay && (new_sockaddr.ss_family != AF_UNIX)) {
ret = __socket_nodelay (new_sock);
if (ret == -1) {
gf_log (this->name, GF_LOG_WARNING,
@@ -1866,8 +2477,10 @@ socket_server_event_handler (int fd, int idx, void *data,
}
}
- if (priv->keepalive) {
+ if (priv->keepalive &&
+ new_sockaddr.ss_family != AF_UNIX) {
ret = __socket_keepalive (new_sock,
+ new_sockaddr.ss_family,
priv->keepaliveintvl,
priv->keepaliveidle);
if (ret == -1)
@@ -1881,6 +2494,15 @@ socket_server_event_handler (int fd, int idx, void *data,
if (!new_trans)
goto unlock;
+ ret = pthread_mutex_init(&new_trans->lock, NULL);
+ if (ret == -1) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "pthread_mutex_init() failed: %s",
+ strerror (errno));
+ close (new_sock);
+ goto unlock;
+ }
+
new_trans->name = gf_strdup (this->name);
memcpy (&new_trans->peerinfo.sockaddr, &new_sockaddr,
@@ -1902,7 +2524,11 @@ socket_server_event_handler (int fd, int idx, void *data,
}
get_transport_identifiers (new_trans);
- socket_init (new_trans);
+ ret = socket_init(new_trans);
+ if (ret != 0) {
+ close(new_sock);
+ goto unlock;
+ }
new_trans->ops = this->ops;
new_trans->init = this->init;
new_trans->fini = this->fini;
@@ -1913,20 +2539,62 @@ socket_server_event_handler (int fd, int idx, void *data,
new_trans->listener = this;
new_priv = new_trans->private;
+ new_priv->use_ssl = priv->use_ssl;
+ new_priv->sock = new_sock;
+ new_priv->own_thread = priv->own_thread;
+
+ new_priv->ssl_ctx = priv->ssl_ctx;
+ if (priv->use_ssl && !priv->own_thread) {
+ if (ssl_setup_connection(new_trans,1) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "server setup failed");
+ close(new_sock);
+ goto unlock;
+ }
+ }
+
+ if (!priv->bio && !priv->own_thread) {
+ ret = __socket_nonblock (new_sock);
+
+ if (ret == -1) {
+ gf_log (this->name, GF_LOG_WARNING,
+ "NBIO on %d failed (%s)",
+ new_sock, strerror (errno));
+
+ close (new_sock);
+ goto unlock;
+ }
+ }
+
pthread_mutex_lock (&new_priv->lock);
{
- new_priv->sock = new_sock;
+ /*
+ * In the own_thread case, this is used to
+ * indicate that we're initializing a server
+ * connection.
+ */
new_priv->connected = 1;
+ new_priv->is_server = _gf_true;
rpc_transport_ref (new_trans);
- new_priv->idx =
- event_register (ctx->event_pool,
- new_sock,
- socket_event_handler,
- new_trans, 1, 0);
+ if (new_priv->own_thread) {
+ if (pipe(new_priv->pipe) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "could not create pipe");
+ }
+ socket_spawn(new_trans);
+ }
+ else {
+ new_priv->idx =
+ event_register (ctx->event_pool,
+ new_sock,
+ socket_event_handler,
+ new_trans,
+ 1, 0);
+ if (new_priv->idx == -1)
+ ret = -1;
+ }
- if (new_priv->idx == -1)
- ret = -1;
}
pthread_mutex_unlock (&new_priv->lock);
if (ret == -1) {
@@ -1935,8 +2603,10 @@ socket_server_event_handler (int fd, int idx, void *data,
goto unlock;
}
- ret = rpc_transport_notify (this, RPC_TRANSPORT_ACCEPT,
- new_trans);
+ if (!priv->own_thread) {
+ ret = rpc_transport_notify (this,
+ RPC_TRANSPORT_ACCEPT, new_trans);
+ }
}
}
unlock:
@@ -1947,7 +2617,7 @@ out:
}
-int
+static int
socket_disconnect (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -1969,7 +2639,7 @@ out:
}
-int
+static int
socket_connect (rpc_transport_t *this, int port)
{
int ret = -1;
@@ -1978,7 +2648,9 @@ socket_connect (rpc_transport_t *this, int port)
socklen_t sockaddr_len = 0;
glusterfs_ctx_t *ctx = NULL;
sa_family_t sa_family = {0, };
+ char *local_addr = NULL;
union gf_sock_union sock_union;
+ struct sockaddr_in *addr = NULL;
GF_VALIDATE_OR_GOTO ("socket", this, err);
GF_VALIDATE_OR_GOTO ("socket", this->private, err);
@@ -2006,6 +2678,10 @@ socket_connect (rpc_transport_t *this, int port)
goto err;
}
+ gf_log (this->name, GF_LOG_TRACE,
+ "connecting %p, state=%u gen=%u sock=%d", this,
+ priv->ot_state, priv->ot_gen, priv->sock);
+
ret = socket_client_get_remote_sockaddr (this, &sock_union.sa,
&sockaddr_len, &sa_family);
if (ret == -1) {
@@ -2016,6 +2692,20 @@ socket_connect (rpc_transport_t *this, int port)
if (port > 0) {
sock_union.sin.sin_port = htons (port);
}
+ if (ntohs(sock_union.sin.sin_port) == GF_DEFAULT_SOCKET_LISTEN_PORT) {
+ if (priv->use_ssl) {
+ gf_log(this->name,GF_LOG_DEBUG,
+ "disabling SSL for portmapper connection");
+ priv->use_ssl = _gf_false;
+ }
+ }
+ else {
+ if (priv->ssl_enabled && !priv->use_ssl) {
+ gf_log(this->name,GF_LOG_DEBUG,
+ "re-enabling SSL for I/O connection");
+ priv->use_ssl = _gf_true;
+ }
+ }
pthread_mutex_lock (&priv->lock);
{
if (priv->sock != -1) {
@@ -2061,7 +2751,7 @@ socket_connect (rpc_transport_t *this, int port)
}
}
- if (priv->nodelay) {
+ if (priv->nodelay && (sa_family != AF_UNIX)) {
ret = __socket_nodelay (priv->sock);
if (ret == -1) {
@@ -2071,21 +2761,9 @@ socket_connect (rpc_transport_t *this, int port)
}
}
- if (!priv->bio) {
- ret = __socket_nonblock (priv->sock);
-
- if (ret == -1) {
- gf_log (this->name, GF_LOG_ERROR,
- "NBIO on %d failed (%s)",
- priv->sock, strerror (errno));
- close (priv->sock);
- priv->sock = -1;
- goto unlock;
- }
- }
-
- if (priv->keepalive) {
+ if (priv->keepalive && sa_family != AF_UNIX) {
ret = __socket_keepalive (priv->sock,
+ sa_family,
priv->keepaliveintvl,
priv->keepaliveidle);
if (ret == -1)
@@ -2097,6 +2775,15 @@ socket_connect (rpc_transport_t *this, int port)
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);
+ 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));
+ }
+
ret = client_bind (this, SA (&this->myinfo.sockaddr),
&this->myinfo.sockaddr_len, priv->sock);
if (ret == -1) {
@@ -2107,29 +2794,86 @@ socket_connect (rpc_transport_t *this, int port)
goto unlock;
}
+ if (!priv->use_ssl && !priv->bio && !priv->own_thread) {
+ ret = __socket_nonblock (priv->sock);
+ if (ret == -1) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "NBIO on %d failed (%s)",
+ priv->sock, strerror (errno));
+ close (priv->sock);
+ priv->sock = -1;
+ goto unlock;
+ }
+ }
+
ret = connect (priv->sock, SA (&this->peerinfo.sockaddr),
this->peerinfo.sockaddr_len);
if (ret == -1 && ((errno != EINPROGRESS) && (errno != ENOENT))) {
- gf_log (this->name, GF_LOG_ERROR,
- "connection attempt failed (%s)",
- strerror (errno));
+ /* 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
+ */
+ gf_log (this->name, ((sa_family == AF_UNIX) ?
+ GF_LOG_DEBUG : GF_LOG_ERROR),
+ "connection attempt on %s failed, (%s)",
+ this->peerinfo.identifier, strerror (errno));
close (priv->sock);
priv->sock = -1;
goto unlock;
}
- priv->connected = 0;
+ if (priv->use_ssl && !priv->own_thread) {
+ ret = ssl_setup_connection(this,0);
+ if (ret < 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "client setup failed");
+ close(priv->sock);
+ priv->sock = -1;
+ goto unlock;
+ }
+ }
- rpc_transport_ref (this);
+ if (!priv->bio && !priv->own_thread) {
+ ret = __socket_nonblock (priv->sock);
- priv->idx = event_register (ctx->event_pool, priv->sock,
- socket_event_handler, this, 1, 1);
- if (priv->idx == -1) {
- gf_log (this->name, GF_LOG_WARNING,
- "failed to register the event");
- ret = -1;
+ if (ret == -1) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "NBIO on %d failed (%s)",
+ priv->sock, strerror (errno));
+ close (priv->sock);
+ priv->sock = -1;
+ goto unlock;
+ }
}
+
+ /*
+ * In the own_thread case, this is used to indicate that we're
+ * initializing a client connection.
+ */
+ priv->connected = 0;
+ priv->is_server = _gf_false;
+ rpc_transport_ref (this);
+
+ if (priv->own_thread) {
+ if (pipe(priv->pipe) < 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "could not create pipe");
+ }
+
+ this->listener = this;
+ socket_spawn(this);
+ }
+ else {
+ priv->idx = event_register (ctx->event_pool, priv->sock,
+ socket_event_handler,
+ this, 1, 1);
+ if (priv->idx == -1) {
+ gf_log ("", GF_LOG_WARNING,
+ "failed to register the event");
+ ret = -1;
+ }
+ }
}
unlock:
pthread_mutex_unlock (&priv->lock);
@@ -2139,7 +2883,7 @@ err:
}
-int
+static int
socket_listen (rpc_transport_t *this)
{
socket_private_t * priv = NULL;
@@ -2221,7 +2965,7 @@ socket_listen (rpc_transport_t *this)
}
}
- if (priv->nodelay) {
+ if (priv->nodelay && (sa_family != AF_UNIX)) {
ret = __socket_nodelay (priv->sock);
if (ret == -1) {
gf_log (this->name, GF_LOG_ERROR,
@@ -2290,7 +3034,7 @@ out:
}
-int32_t
+static int32_t
socket_submit_request (rpc_transport_t *this, rpc_transport_req_t *req)
{
socket_private_t *priv = NULL;
@@ -2299,6 +3043,7 @@ socket_submit_request (rpc_transport_t *this, rpc_transport_req_t *req)
char need_append = 1;
struct ioq *entry = NULL;
glusterfs_ctx_t *ctx = NULL;
+ char a_byte = 'j';
GF_VALIDATE_OR_GOTO ("socket", this, out);
GF_VALIDATE_OR_GOTO ("socket", this->private, out);
@@ -2324,21 +3069,31 @@ socket_submit_request (rpc_transport_t *this, rpc_transport_req_t *req)
goto unlock;
if (list_empty (&priv->ioq)) {
- ret = __socket_ioq_churn_entry (this, entry);
+ ret = __socket_ioq_churn_entry (this, entry, 1);
- if (ret == 0)
+ if (ret == 0) {
need_append = 0;
-
- if (ret > 0)
+ }
+ if (ret > 0) {
need_poll_out = 1;
+ }
}
if (need_append) {
list_add_tail (&entry->list, &priv->ioq);
+ if (priv->own_thread) {
+ /*
+ * Make sure the polling thread wakes up, by
+ * writing a byte to represent this entry.
+ */
+ if (write(priv->pipe[1],&a_byte,1) < 1) {
+ gf_log(this->name,GF_LOG_WARNING,
+ "write error on pipe");
+ }
+ }
ret = 0;
}
-
- if (need_poll_out) {
+ if (!priv->own_thread && need_poll_out) {
/* first entry to wait. continue writing on POLLOUT */
priv->idx = event_select_on (ctx->event_pool,
priv->sock,
@@ -2353,7 +3108,7 @@ out:
}
-int32_t
+static int32_t
socket_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply)
{
socket_private_t *priv = NULL;
@@ -2362,6 +3117,7 @@ socket_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply)
char need_append = 1;
struct ioq *entry = NULL;
glusterfs_ctx_t *ctx = NULL;
+ char a_byte = 'd';
GF_VALIDATE_OR_GOTO ("socket", this, out);
GF_VALIDATE_OR_GOTO ("socket", this->private, out);
@@ -2380,33 +3136,44 @@ socket_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply)
}
goto unlock;
}
+
priv->submit_log = 0;
entry = __socket_ioq_new (this, &reply->msg);
if (!entry)
goto unlock;
+
if (list_empty (&priv->ioq)) {
- ret = __socket_ioq_churn_entry (this, entry);
+ ret = __socket_ioq_churn_entry (this, entry, 1);
- if (ret == 0)
+ if (ret == 0) {
need_append = 0;
-
- if (ret > 0)
+ }
+ if (ret > 0) {
need_poll_out = 1;
+ }
}
if (need_append) {
list_add_tail (&entry->list, &priv->ioq);
+ if (priv->own_thread) {
+ /*
+ * Make sure the polling thread wakes up, by
+ * writing a byte to represent this entry.
+ */
+ if (write(priv->pipe[1],&a_byte,1) < 1) {
+ gf_log(this->name,GF_LOG_WARNING,
+ "write error on pipe");
+ }
+ }
ret = 0;
}
-
- if (need_poll_out) {
+ if (!priv->own_thread && 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);
}
}
-
unlock:
pthread_mutex_unlock (&priv->lock);
@@ -2415,7 +3182,7 @@ out:
}
-int32_t
+static int32_t
socket_getpeername (rpc_transport_t *this, char *hostname, int hostlen)
{
int32_t ret = -1;
@@ -2434,7 +3201,7 @@ out:
}
-int32_t
+static int32_t
socket_getpeeraddr (rpc_transport_t *this, char *peeraddr, int addrlen,
struct sockaddr_storage *sa, socklen_t salen)
{
@@ -2455,7 +3222,7 @@ out:
}
-int32_t
+static int32_t
socket_getmyname (rpc_transport_t *this, char *hostname, int hostlen)
{
int32_t ret = -1;
@@ -2474,7 +3241,7 @@ out:
}
-int32_t
+static int32_t
socket_getmyaddr (rpc_transport_t *this, char *myaddr, int addrlen,
struct sockaddr_storage *sa, socklen_t salen)
{
@@ -2494,6 +3261,25 @@ out:
}
+static int
+socket_throttle (rpc_transport_t *this, gf_boolean_t onoff)
+{
+ socket_private_t *priv = NULL;
+
+ priv = this->private;
+
+ /* The way we implement throttling is by taking off
+ POLLIN event from the polled flags. This way we
+ never get called with the POLLIN event and therefore
+ will never read() any more data until throttling
+ is turned off.
+ */
+ priv->idx = event_select_on (this->ctx->event_pool, priv->sock,
+ priv->idx, (int) !onoff, -1);
+ return 0;
+}
+
+
struct rpc_transport_ops tops = {
.listen = socket_listen,
.connect = socket_connect,
@@ -2504,6 +3290,7 @@ struct rpc_transport_ops tops = {
.get_peeraddr = socket_getpeeraddr,
.get_myname = socket_getmyname,
.get_myaddr = socket_getmyaddr,
+ .throttle = socket_throttle,
};
int
@@ -2560,7 +3347,7 @@ out:
}
-int
+static int
socket_init (rpc_transport_t *this)
{
socket_private_t *priv = NULL;
@@ -2569,6 +3356,7 @@ socket_init (rpc_transport_t *this)
char *optstr = NULL;
uint32_t keepalive = 0;
uint32_t backlog = 0;
+ int session_id = 0;
if (this->private) {
gf_log_callingfn (this->name, GF_LOG_ERROR,
@@ -2580,6 +3368,7 @@ socket_init (rpc_transport_t *this)
if (!priv) {
return -1;
}
+ memset(priv,0,sizeof(*priv));
pthread_mutex_init (&priv->lock, NULL);
@@ -2697,11 +3486,134 @@ socket_init (rpc_transport_t *this)
}
}
- optstr = NULL;
+ priv->windowsize = (int)windowsize;
+
+ priv->ssl_enabled = _gf_false;
+ if (dict_get_str(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");
+ }
+ }
+
+ priv->ssl_own_cert = DEFAULT_CERT_PATH;
+ if (dict_get_str(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, SSL_ENABLED_OPT);
+ }
+ priv->ssl_own_cert = optstr;
+ }
+ 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 (!priv->ssl_enabled) {
+ gf_log(this->name,GF_LOG_WARNING,
+ "%s specified without %s (ignored)",
+ SSL_PRIVATE_KEY_OPT, SSL_ENABLED_OPT);
+ }
+ priv->ssl_private_key = optstr;
+ }
+ 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 (!priv->ssl_enabled) {
+ gf_log(this->name,GF_LOG_WARNING,
+ "%s specified without %s (ignored)",
+ SSL_CA_LIST_OPT, SSL_ENABLED_OPT);
+ }
+ priv->ssl_ca_list = optstr;
+ }
+ priv->ssl_ca_list = gf_strdup(priv->ssl_ca_list);
+
+ gf_log(this->name,GF_LOG_INFO,"SSL support is %s",
+ priv->ssl_enabled ? "ENABLED" : "NOT enabled");
+ /*
+ * This might get overridden temporarily in socket_connect (q.v.)
+ * if we're using the glusterd portmapper.
+ */
+ priv->use_ssl = priv->ssl_enabled;
+
+ priv->own_thread = priv->use_ssl;
+ if (dict_get_str(this->options,OWN_THREAD_OPT,&optstr) == 0) {
+ if (gf_string2boolean (optstr, &priv->own_thread) != 0) {
+ gf_log (this->name, GF_LOG_ERROR,
+ "invalid value given for own-thread boolean");
+ }
+ }
+ gf_log(this->name,GF_LOG_INFO,"using %s polling thread",
+ priv->own_thread ? "private" : "system");
+
+ if (priv->use_ssl) {
+ SSL_library_init();
+ SSL_load_error_strings();
+ priv->ssl_meth = (SSL_METHOD *)TLSv1_method();
+ priv->ssl_ctx = SSL_CTX_new(priv->ssl_meth);
+
+ if (SSL_CTX_set_cipher_list(priv->ssl_ctx,
+ "HIGH:-SSLv2") == 0) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "failed to find any valid ciphers");
+ goto err;
+ }
+
+ if (!SSL_CTX_use_certificate_chain_file(priv->ssl_ctx,
+ priv->ssl_own_cert)) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "could not load our cert");
+ goto err;
+ }
+
+ if (!SSL_CTX_use_PrivateKey_file(priv->ssl_ctx,
+ priv->ssl_private_key,
+ SSL_FILETYPE_PEM)) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "could not load private key");
+ goto err;
+ }
+
+ if (!SSL_CTX_load_verify_locations(priv->ssl_ctx,
+ priv->ssl_ca_list,0)) {
+ gf_log(this->name,GF_LOG_ERROR,
+ "could not load CA list");
+ goto err;
+ }
+
+#if (OPENSSL_VERSION_NUMBER < 0x00905100L)
+ SSL_CTX_set_verify_depth(ctx,1);
+#endif
+
+ priv->ssl_session_id = ++session_id;
+ SSL_CTX_set_session_id_context(priv->ssl_ctx,
+ (void *)&priv->ssl_session_id,
+ sizeof(priv->ssl_session_id));
+
+ SSL_CTX_set_verify(priv->ssl_ctx,SSL_VERIFY_PEER,0);
+ }
+
+ if (priv->own_thread) {
+ priv->ot_state = OT_IDLE;
+ }
+
out:
this->private = priv;
-
return 0;
+
+err:
+ if (priv->ssl_own_cert) {
+ GF_FREE(priv->ssl_own_cert);
+ }
+ if (priv->ssl_private_key) {
+ GF_FREE(priv->ssl_private_key);
+ }
+ if (priv->ssl_ca_list) {
+ GF_FREE(priv->ssl_ca_list);
+ }
+ GF_FREE(priv);
+ return -1;
}
@@ -2727,6 +3639,15 @@ fini (rpc_transport_t *this)
"transport %p destroyed", this);
pthread_mutex_destroy (&priv->lock);
+ if (priv->ssl_private_key) {
+ GF_FREE(priv->ssl_private_key);
+ }
+ if (priv->ssl_own_cert) {
+ GF_FREE(priv->ssl_own_cert);
+ }
+ if (priv->ssl_ca_list) {
+ GF_FREE(priv->ssl_ca_list);
+ }
GF_FREE (priv);
}
@@ -2771,8 +3692,7 @@ struct volume_options options[] = {
},
{ .key = { "transport.address-family",
"address-family" },
- .value = {"inet", "inet6", "inet/inet6", "inet6/inet",
- "unix", "inet-sdp" },
+ .value = {"inet", "inet6", "unix", "inet-sdp" },
.type = GF_OPTION_TYPE_STR
},
@@ -2805,5 +3725,20 @@ struct volume_options options[] = {
{ .key = {"transport.socket.read-fail-log"},
.type = GF_OPTION_TYPE_BOOL
},
+ { .key = {SSL_ENABLED_OPT},
+ .type = GF_OPTION_TYPE_BOOL
+ },
+ { .key = {SSL_OWN_CERT_OPT},
+ .type = GF_OPTION_TYPE_STR
+ },
+ { .key = {SSL_PRIVATE_KEY_OPT},
+ .type = GF_OPTION_TYPE_STR
+ },
+ { .key = {SSL_CA_LIST_OPT},
+ .type = GF_OPTION_TYPE_STR
+ },
+ { .key = {OWN_THREAD_OPT},
+ .type = GF_OPTION_TYPE_BOOL
+ },
{ .key = {NULL} }
};
diff --git a/rpc/rpc-transport/socket/src/socket.h b/rpc/rpc-transport/socket/src/socket.h
index 0c897bd2e..e0b412fcc 100644
--- a/rpc/rpc-transport/socket/src/socket.h
+++ b/rpc/rpc-transport/socket/src/socket.h
@@ -11,6 +11,8 @@
#ifndef _SOCKET_H
#define _SOCKET_H
+#include <openssl/ssl.h>
+#include <openssl/err.h>
#ifndef _CONFIG_H
#define _CONFIG_H
@@ -72,6 +74,12 @@ typedef enum {
SP_STATE_READ_VERFBYTES, /* read verifier data */
SP_STATE_READING_PROGHDR,
SP_STATE_READ_PROGHDR,
+ SP_STATE_READING_PROGHDR_XDATA,
+ SP_STATE_READ_PROGHDR_XDATA, /* It's a bad "name" in the generic
+ RPC state machine, but greatly
+ aids code review (and xdata is
+ the only "consumer" of this state)
+ */
SP_STATE_READING_PROG,
} sp_rpcfrag_vectored_request_state_t;
@@ -115,6 +123,8 @@ typedef enum {
typedef enum {
SP_STATE_ACCEPTED_SUCCESS_REPLY_INIT,
SP_STATE_READING_PROC_HEADER,
+ SP_STATE_READING_PROC_OPAQUE,
+ SP_STATE_READ_PROC_OPAQUE,
SP_STATE_READ_PROC_HEADER,
} sp_rpcfrag_vectored_reply_accepted_success_state_t;
@@ -133,10 +143,60 @@ typedef struct {
sp_rpcfrag_vectored_reply_accepted_success_state_t accepted_success_state;
} sp_rpcfrag_vectored_reply_state_t;
+struct gf_sock_incoming_frag {
+ char *fragcurrent;
+ uint32_t bytes_read;
+ uint32_t remaining_size;
+ struct iovec vector;
+ struct iovec *pending_vector;
+ union {
+ sp_rpcfrag_request_state_t request;
+ sp_rpcfrag_vectored_reply_state_t reply;
+ } call_body;
+
+ sp_rpcfrag_simple_msg_state_t simple_state;
+ sp_rpcfrag_state_t state;
+};
+
+#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 iovec vector[2];
+ int count;
+ struct iovec payload_vector;
+ struct iobref *iobref;
+ rpc_request_info_t *request_info;
+ struct iovec *pending_vector;
+ 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;
+};
+
+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;
- unsigned char connected; // -1 = not connected. 0 = in progress. 1 = connected
+ /* -1 = not connected. 0 = in progress. 1 = connected */
+ char connected;
char bio;
char connect_finish_log;
char submit_log;
@@ -147,36 +207,7 @@ typedef struct {
struct ioq *ioq_prev;
};
};
- struct {
- sp_rpcrecord_state_t record_state;
- struct {
- char *fragcurrent;
- uint32_t bytes_read;
- uint32_t remaining_size;
- struct iovec vector;
- struct iovec *pending_vector;
- union {
- sp_rpcfrag_request_state_t request;
- sp_rpcfrag_vectored_reply_state_t reply;
- } call_body;
-
- sp_rpcfrag_simple_msg_state_t simple_state;
- sp_rpcfrag_state_t state;
- } frag;
- struct iobuf *iobuf;
- size_t iobuf_size;
- struct iovec vector[2];
- int count;
- struct iovec payload_vector;
- struct iobref *iobref;
- rpc_request_info_t *request_info;
- struct iovec *pending_vector;
- int pending_count;
- uint32_t fraghdr;
- char complete_record;
- msg_type_t msg_type;
- size_t total_bytes_read;
- } incoming;
+ struct gf_sock_incoming incoming;
pthread_mutex_t lock;
int windowsize;
char lowlat;
@@ -186,6 +217,22 @@ typedef struct {
int keepaliveintvl;
uint32_t backlog;
gf_boolean_t read_fail_log;
+ gf_boolean_t ssl_enabled;
+ 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];
+ gf_boolean_t own_thread;
+ ot_state_t ot_state;
+ uint32_t ot_gen;
+ gf_boolean_t is_server;
} socket_private_t;
generated by cgit (git 2.34.1) at 2026-01-05 19:52:04 +0000