summaryrefslogtreecommitdiffstats
path: root/xlators/protocol/server/src/server-resolve.c
diff options
context:
space:
mode:
authorshishir gowda <shishirng@gluster.com>2012-02-28 18:03:29 +0530
committerVijay Bellur <vijay@gluster.com>2012-02-29 02:09:59 -0800
commit5ad1c2caed743fc10a0645055a825252dfbf77ba (patch)
treed8ed477083a1a8a64089a243e1ab452bde035d85 /xlators/protocol/server/src/server-resolve.c
parentc2c82c360fda20533093c27c82ce0725cf40cab5 (diff)
storage/posix: In lookup check for buf->gfid being NULL
There are few cases where create and lookup race. Lookup ends up getting a valid struct iatt, but with no gfid set. We need to check for gfid being 0, and handle it as an error. Signed-off-by: shishir gowda <shishirng@gluster.com> Change-Id: I36ae1978b325aff964cbc3b24730c1e993666267 BUG: 797167 Reviewed-on: http://review.gluster.com/2832 Tested-by: Gluster Build System <jenkins@build.gluster.com> Reviewed-by: Amar Tumballi <amarts@redhat.com> Reviewed-by: Vijay Bellur <vijay@gluster.com>
Diffstat (limited to 'xlators/protocol/server/src/server-resolve.c')
0 files changed, 0 insertions, 0 deletions
generated by cgit (git 2.34.1) at 2026-01-18 20:02:03 +0000
e.h
-I$(top_srcdir)/libglusterfs/src -I$(top_srcdir)/rpc/rpc-lib/src/ \
diff --git a/rpc/rpc-transport/rdma/src/name.c b/rpc/rpc-transport/rdma/src/name.c
index f6ae818b10b..c57428ad62e 100644
--- a/rpc/rpc-transport/rdma/src/name.c
+++ b/rpc/rpc-transport/rdma/src/name.c
@@ -13,6 +13,7 @@
#include <errno.h>
#include <netdb.h>
#include <string.h>
+#include <rdma/rdma_cma.h>
#ifndef AF_INET_SDP
#define AF_INET_SDP 27
@@ -31,7 +32,8 @@ 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;
@@ -51,12 +53,14 @@ af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
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;
}
// ignore the reserved ports
@@ -64,7 +68,7 @@ af_inet_bind_to_port_lt_ceiling (int fd, struct sockaddr *sockaddr,
port--;
continue;
}
- ret = bind (fd, sockaddr, sockaddr_len);
+ ret = rdma_bind_addr (cm_id, sockaddr);
if (ret == 0)
break;
@@ -78,11 +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;
@@ -114,6 +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)
@@ -412,10 +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;
@@ -427,22 +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,
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, GF_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:
diff --git a/rpc/rpc-transport/rdma/src/name.h b/rpc/rpc-transport/rdma/src/name.h
index 114ed1661a3..742fc5fc3f2 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 a44e8995f11..135fbdf2868 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"
@@ -35,99 +34,26 @@ 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;
-
- if (ibv_query_port (context, port, &attr))
- return 0;
-
- return attr.lid;
-}
+static void *
+gf_rdma_send_completion_proc (void *data);
-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";
- }
-}
+static void *
+gf_rdma_recv_completion_proc (void *data);
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
@@ -157,7 +83,7 @@ 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;
@@ -184,7 +110,7 @@ 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_WARNING,
+ gf_log (this->name, GF_LOG_WARNING,
"memory registration failed (%s)",
strerror (errno));
goto out;
@@ -259,22 +185,6 @@ __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 +200,7 @@ __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 +220,890 @@ 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
+ }
+ };
+ 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 = pthread_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 = pthread_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;
+ }
+
+ /* 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)
{
@@ -832,28 +1607,6 @@ 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)
{
@@ -1263,7 +2016,7 @@ out:
}
-static inline int32_t
+inline int32_t
__gf_rdma_register_local_mr_for_rdma (gf_rdma_peer_t *peer,
struct iovec *vector, int count,
gf_rdma_post_context_t *ctx)
@@ -1715,7 +2468,7 @@ __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);
}
@@ -1992,189 +2745,6 @@ 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,
@@ -2273,25 +2843,6 @@ __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)
@@ -2301,7 +2852,7 @@ __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;
@@ -2312,18 +2863,36 @@ __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,
@@ -2337,39 +2906,16 @@ 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);
@@ -2381,300 +2927,52 @@ 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);
-