diff options
Diffstat (limited to 'rpc/rpc-transport/rdma/src/rdma.c')
| -rw-r--r-- | rpc/rpc-transport/rdma/src/rdma.c | 4471 |
1 files changed, 2075 insertions, 2396 deletions
diff --git a/rpc/rpc-transport/rdma/src/rdma.c b/rpc/rpc-transport/rdma/src/rdma.c index 7d2bfc0ad..6e6099a98 100644 --- a/rpc/rpc-transport/rdma/src/rdma.c +++ b/rpc/rpc-transport/rdma/src/rdma.c @@ -1,23 +1,13 @@ /* - Copyright (c) 2006-2010 Gluster, Inc. <http://www.gluster.com> + Copyright (c) 2008-2012 Red Hat, Inc. <http://www.redhat.com> This file is part of GlusterFS. - GlusterFS is free software; you can redistribute it and/or modify - it under the terms of the GNU Affero General Public License as published - by the Free Software Foundation; either version 3 of the License, - or (at your option) any later version. - - GlusterFS is distributed in the hope that it will be useful, but - WITHOUT ANY WARRANTY; without even the implied warranty of - MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - Affero General Public License for more details. - - You should have received a copy of the GNU Affero General Public License - along with this program. If not, see - <http://www.gnu.org/licenses/>. + This file is licensed to you under your choice of the GNU Lesser + General Public License, version 3 or any later version (LGPLv3 or + later), or the GNU General Public License, version 2 (GPLv2), in all + cases as published by the Free Software Foundation. */ - #ifndef _CONFIG_H #define _CONFIG_H #include "config.h" @@ -30,119 +20,47 @@ #include "name.h" #include "byte-order.h" #include "xlator.h" +#include "xdr-rpc.h" #include <signal.h> -#define RDMA_LOG_NAME "rpc-transport/rdma" +#define GF_RDMA_LOG_NAME "rpc-transport/rdma" static int32_t -__rdma_ioq_churn (rdma_peer_t *peer); +__gf_rdma_ioq_churn (gf_rdma_peer_t *peer); -rdma_post_t * -rdma_post_ref (rdma_post_t *post); +gf_rdma_post_t * +gf_rdma_post_ref (gf_rdma_post_t *post); int -rdma_post_unref (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 -rdma_get_local_lid (struct ibv_context *context, - int32_t port) -{ - struct ibv_port_attr attr; +gf_rdma_post_unref (gf_rdma_post_t *post); - if (ibv_query_port (context, port, &attr)) - return 0; +static void * +gf_rdma_send_completion_proc (void *data); - return attr.lid; -} +static void * +gf_rdma_recv_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"; - } -} +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; - - int32_t ret = 0; - const char *state_str = NULL; - - if (!ctx) { - gf_log_callingfn (RDMA_LOG_NAME, GF_LOG_ERROR, - "Error in supplied context"); - return -1; - } - - ret = ibv_query_port (ctx, port, &port_attr); - - if (ret) { - gf_log (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 (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; - - int32_t ret = -1; - uint8_t ib_port = 0; +__gf_rdma_teardown (rpc_transport_t *this); - if (!ib_ctx) { - gf_log_callingfn (RDMA_LOG_NAME, GF_LOG_ERROR, - "Error in supplied context"); - return -1; - } - if (ibv_query_device (ib_ctx, &ib_device_attr)) { - gf_log (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 (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 -rdma_put_post (rdma_queue_t *queue, - rdma_post_t *post) +gf_rdma_put_post (gf_rdma_queue_t *queue, gf_rdma_post_t *post) { post->ctx.is_request = 0; @@ -167,14 +85,15 @@ rdma_put_post (rdma_queue_t *queue, } -static rdma_post_t * -rdma_new_post (rdma_device_t *device, int32_t len, rdma_post_type_t type) +static gf_rdma_post_t * +gf_rdma_new_post (rpc_transport_t *this, gf_rdma_device_t *device, int32_t len, + gf_rdma_post_type_t type) { - rdma_post_t *post = NULL; - int ret = -1; + gf_rdma_post_t *post = NULL; + int ret = -1; - post = (rdma_post_t *) GF_CALLOC (1, sizeof (*post), - gf_common_mt_rdma_post_t); + post = (gf_rdma_post_t *) GF_CALLOC (1, sizeof (*post), + gf_common_mt_rdma_post_t); if (post == NULL) { goto out; } @@ -185,7 +104,7 @@ rdma_new_post (rdma_device_t *device, int32_t len, rdma_post_type_t type) post->buf = valloc (len); if (!post->buf) { - gf_log_nomem (RDMA_LOG_NAME, GF_LOG_ERROR, len); + gf_log_nomem (GF_RDMA_LOG_NAME, GF_LOG_ERROR, len); goto out; } @@ -194,8 +113,9 @@ rdma_new_post (rdma_device_t *device, int32_t len, rdma_post_type_t type) post->buf_size, IBV_ACCESS_LOCAL_WRITE); if (!post->mr) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "memory registration failed"); + gf_log (this->name, GF_LOG_WARNING, + "memory registration failed (%s)", + strerror (errno)); goto out; } @@ -205,9 +125,7 @@ rdma_new_post (rdma_device_t *device, int32_t len, rdma_post_type_t type) ret = 0; out: if (ret != 0) { - if (post->buf != NULL) { - free (post->buf); - } + free (post->buf); GF_FREE (post); post = NULL; @@ -217,10 +135,10 @@ out: } -static rdma_post_t * -rdma_get_post (rdma_queue_t *queue) +static gf_rdma_post_t * +gf_rdma_get_post (gf_rdma_queue_t *queue) { - rdma_post_t *post; + gf_rdma_post_t *post = NULL; pthread_mutex_lock (&queue->lock); { @@ -247,7 +165,7 @@ rdma_get_post (rdma_queue_t *queue) } void -rdma_destroy_post (rdma_post_t *post) +gf_rdma_destroy_post (gf_rdma_post_t *post) { ibv_dereg_mr (post->mr); free (post->buf); @@ -256,10 +174,12 @@ rdma_destroy_post (rdma_post_t *post) static int32_t -__rdma_quota_get (rdma_peer_t *peer) +__gf_rdma_quota_get (gf_rdma_peer_t *peer) { - int32_t ret = -1; - rdma_private_t *priv = peer->trans->private; + int32_t ret = -1; + gf_rdma_private_t *priv = NULL; + + priv = peer->trans->private; if (priv->connected && peer->quota > 0) { ret = peer->quota--; @@ -268,25 +188,9 @@ __rdma_quota_get (rdma_peer_t *peer) return ret; } -/* - static int32_t - rdma_quota_get (rdma_peer_t *peer) - { - int32_t ret = -1; - rdma_private_t *priv = peer->trans->private; - - pthread_mutex_lock (&priv->write_mutex); - { - ret = __rdma_quota_get (peer); - } - pthread_mutex_unlock (&priv->write_mutex); - - return ret; - } -*/ static void -__rdma_ioq_entry_free (rdma_ioq_t *entry) +__gf_rdma_ioq_entry_free (gf_rdma_ioq_t *entry) { list_del_init (&entry->list); @@ -299,47 +203,920 @@ __rdma_ioq_entry_free (rdma_ioq_t *entry) iobref_unref (entry->msg.request.rsp_iobref); entry->msg.request.rsp_iobref = NULL; } - /* TODO: use mem-pool */ - mem_put (entry->pool, entry); + + mem_put (entry); } static void -__rdma_ioq_flush (rdma_peer_t *peer) +__gf_rdma_ioq_flush (gf_rdma_peer_t *peer) { - rdma_ioq_t *entry = NULL, *dummy = NULL; + gf_rdma_ioq_t *entry = NULL, *dummy = NULL; list_for_each_entry_safe (entry, dummy, &peer->ioq, list) { - __rdma_ioq_entry_free (entry); + __gf_rdma_ioq_entry_free (entry); } } static int32_t -__rdma_disconnect (rpc_transport_t *this) +__gf_rdma_disconnect (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - int32_t ret = 0; + gf_rdma_private_t *priv = NULL; - if (priv->connected || priv->tcp_connected) { - fcntl (priv->sock, F_SETFL, O_NONBLOCK); - if (shutdown (priv->sock, SHUT_RDWR) != 0) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, - "shutdown () - error: %s", - strerror (errno)); - ret = -errno; - priv->tcp_connected = 0; + priv = this->private; + + if (priv->connected) { + rdma_disconnect (priv->peer.cm_id); + } + + return 0; +} + + +static void +gf_rdma_queue_init (gf_rdma_queue_t *queue) +{ + pthread_mutex_init (&queue->lock, NULL); + + queue->active_posts.next = &queue->active_posts; + queue->active_posts.prev = &queue->active_posts; + queue->passive_posts.next = &queue->passive_posts; + queue->passive_posts.prev = &queue->passive_posts; +} + + +static void +__gf_rdma_destroy_queue (gf_rdma_post_t *post) +{ + gf_rdma_post_t *tmp = NULL; + + while (post->next != post) { + tmp = post->next; + + post->next = post->next->next; + post->next->prev = post; + + gf_rdma_destroy_post (tmp); + } +} + + +static void +gf_rdma_destroy_queue (gf_rdma_queue_t *queue) +{ + if (queue == NULL) { + goto out; + } + + pthread_mutex_lock (&queue->lock); + { + if (queue->passive_count > 0) { + __gf_rdma_destroy_queue (&queue->passive_posts); + queue->passive_count = 0; + } + + if (queue->active_count > 0) { + __gf_rdma_destroy_queue (&queue->active_posts); + queue->active_count = 0; + } + } + pthread_mutex_unlock (&queue->lock); + +out: + return; +} + + +static void +gf_rdma_destroy_posts (rpc_transport_t *this) +{ + gf_rdma_device_t *device = NULL; + gf_rdma_private_t *priv = NULL; + + if (this == NULL) { + goto out; + } + + priv = this->private; + device = priv->device; + + gf_rdma_destroy_queue (&device->sendq); + gf_rdma_destroy_queue (&device->recvq); + +out: + return; +} + + +static int32_t +__gf_rdma_create_posts (rpc_transport_t *this, int32_t count, int32_t size, + gf_rdma_queue_t *q, gf_rdma_post_type_t type) +{ + int32_t i = 0; + int32_t ret = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + + priv = this->private; + device = priv->device; + + for (i=0 ; i<count ; i++) { + gf_rdma_post_t *post = NULL; + + post = gf_rdma_new_post (this, device, size + 2048, type); + if (!post) { + gf_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 -rdma_post_send (struct ibv_qp *qp, - rdma_post_t *post, - int32_t len) +gf_rdma_post_send (struct ibv_qp *qp, gf_rdma_post_t *post, int32_t len) { struct ibv_sge list = { .addr = (unsigned long) post->buf, @@ -362,63 +1139,64 @@ rdma_post_send (struct ibv_qp *qp, } int -__rdma_encode_error(rdma_peer_t *peer, rdma_reply_info_t *reply_info, - struct iovec *rpchdr, uint32_t *ptr, - rdma_errcode_t err) +__gf_rdma_encode_error(gf_rdma_peer_t *peer, gf_rdma_reply_info_t *reply_info, + struct iovec *rpchdr, gf_rdma_header_t *hdr, + gf_rdma_errcode_t err) { - 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(RDMA_VERSION); - *ptr++ = hton32(peer->send_count); - *ptr++ = hton32(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(RDMA_VERSION); - *ptr++ = hton32(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); } int32_t -__rdma_send_error (rdma_peer_t *peer, rdma_ioq_t *entry, rdma_post_t *post, - rdma_reply_info_t *reply_info, rdma_errcode_t err) +__gf_rdma_send_error (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, + gf_rdma_post_t *post, gf_rdma_reply_info_t *reply_info, + gf_rdma_errcode_t err) { - int32_t ret = -1, len; + int32_t ret = -1, len = 0; - len = __rdma_encode_error (peer, reply_info, entry->rpchdr, - (uint32_t *)post->buf, err); + len = __gf_rdma_encode_error (peer, reply_info, entry->rpchdr, + (gf_rdma_header_t *)post->buf, err); if (len == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "encode error returned -1"); goto out; } - rdma_post_ref (post); + gf_rdma_post_ref (post); - ret = rdma_post_send (peer->qp, post, len); + ret = gf_rdma_post_send (peer->qp, post, len); if (!ret) { ret = len; } else { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, - "rdma_post_send (to %s) failed with ret = %d (%s)", + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "gf_rdma_post_send (to %s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); - rdma_post_unref (post); - __rdma_disconnect (peer->trans); + gf_rdma_post_unref (post); + __gf_rdma_disconnect (peer->trans); ret = -1; } @@ -428,24 +1206,24 @@ out: int32_t -__rdma_create_read_chunks_from_vector (rdma_peer_t *peer, - rdma_read_chunk_t **readch_ptr, - int32_t *pos, struct iovec *vector, - int count, - rdma_request_context_t *request_ctx) +__gf_rdma_create_read_chunks_from_vector (gf_rdma_peer_t *peer, + gf_rdma_read_chunk_t **readch_ptr, + int32_t *pos, struct iovec *vector, + int count, + gf_rdma_request_context_t *request_ctx) { - int i = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - struct ibv_mr *mr = NULL; - rdma_read_chunk_t *readch = NULL; - int32_t ret = -1; - - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, readch_ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, *readch_ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, vector, out); + int i = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + struct ibv_mr *mr = NULL; + gf_rdma_read_chunk_t *readch = NULL; + int32_t ret = -1; + + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, peer, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, readch_ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, *readch_ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, request_ctx, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, vector, out); priv = peer->trans->private; device = priv->device; @@ -459,8 +1237,10 @@ __rdma_create_read_chunks_from_vector (rdma_peer_t *peer, vector[i].iov_len, IBV_ACCESS_REMOTE_READ); if (!mr) { - gf_log (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; } @@ -485,77 +1265,72 @@ out: int32_t -__rdma_create_read_chunks (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_chunktype_t type, uint32_t **ptr, - rdma_request_context_t *request_ctx) +__gf_rdma_create_read_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, + gf_rdma_chunktype_t type, uint32_t **ptr, + gf_rdma_request_context_t *request_ctx) { int32_t ret = -1; - rdma_device_t *device = NULL; - rdma_private_t *priv = NULL; int pos = 0; - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, entry, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, *ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, request_ctx, out); - - priv = peer->trans->private; - device = priv->device; + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, peer, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, entry, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, *ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, request_ctx, out); request_ctx->iobref = iobref_ref (entry->iobref); - if (type == rdma_areadch) { + if (type == gf_rdma_areadch) { pos = 0; - ret = __rdma_create_read_chunks_from_vector (peer, - (rdma_read_chunk_t **)ptr, - &pos, - entry->rpchdr, - entry->rpchdr_count, - request_ctx); + ret = __gf_rdma_create_read_chunks_from_vector (peer, + (gf_rdma_read_chunk_t **)ptr, + &pos, + entry->rpchdr, + entry->rpchdr_count, + request_ctx); if (ret == -1) { - gf_log (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; } - ret = __rdma_create_read_chunks_from_vector (peer, - (rdma_read_chunk_t **)ptr, - &pos, - entry->proghdr, - entry->proghdr_count, - request_ctx); + ret = __gf_rdma_create_read_chunks_from_vector (peer, + (gf_rdma_read_chunk_t **)ptr, + &pos, + entry->proghdr, + entry->proghdr_count, + request_ctx); if (ret == -1) { - gf_log (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"); } if (entry->prog_payload_count != 0) { - ret = __rdma_create_read_chunks_from_vector (peer, - (rdma_read_chunk_t **)ptr, - &pos, - entry->prog_payload, - entry->prog_payload_count, - request_ctx); + ret = __gf_rdma_create_read_chunks_from_vector (peer, + (gf_rdma_read_chunk_t **)ptr, + &pos, + entry->prog_payload, + entry->prog_payload_count, + request_ctx); if (ret == -1) { - gf_log (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"); } } } else { pos = iov_length (entry->rpchdr, entry->rpchdr_count); - ret = __rdma_create_read_chunks_from_vector (peer, - (rdma_read_chunk_t **)ptr, - &pos, - entry->prog_payload, - entry->prog_payload_count, - request_ctx); + ret = __gf_rdma_create_read_chunks_from_vector (peer, + (gf_rdma_read_chunk_t **)ptr, + &pos, + entry->prog_payload, + entry->prog_payload_count, + request_ctx); if (ret == -1) { - gf_log (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"); } } @@ -569,23 +1344,23 @@ out: int32_t -__rdma_create_write_chunks_from_vector (rdma_peer_t *peer, - rdma_write_chunk_t **writech_ptr, - struct iovec *vector, int count, - rdma_request_context_t *request_ctx) +__gf_rdma_create_write_chunks_from_vector (gf_rdma_peer_t *peer, + gf_rdma_write_chunk_t **writech_ptr, + struct iovec *vector, int count, + gf_rdma_request_context_t *request_ctx) { - int i = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - struct ibv_mr *mr = NULL; - rdma_write_chunk_t *writech = NULL; - int32_t ret = -1; - - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, writech_ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, *writech_ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, vector, out); + int i = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + struct ibv_mr *mr = NULL; + gf_rdma_write_chunk_t *writech = NULL; + int32_t ret = -1; + + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, peer, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, writech_ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, *writech_ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, request_ctx, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, vector, out); writech = *writech_ptr; @@ -598,8 +1373,10 @@ __rdma_create_write_chunks_from_vector (rdma_peer_t *peer, IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_LOCAL_WRITE); if (!mr) { - gf_log (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; } @@ -622,23 +1399,23 @@ out: int32_t -__rdma_create_write_chunks (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_chunktype_t chunk_type, uint32_t **ptr, - rdma_request_context_t *request_ctx) +__gf_rdma_create_write_chunks (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, + gf_rdma_chunktype_t chunk_type, uint32_t **ptr, + gf_rdma_request_context_t *request_ctx) { - int32_t ret = -1; - rdma_write_array_t *warray = NULL; + int32_t ret = -1; + gf_rdma_write_array_t *warray = NULL; - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, *ptr, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, request_ctx, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, entry, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, peer, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, *ptr, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, request_ctx, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, entry, out); - if ((chunk_type == rdma_replych) + if ((chunk_type == gf_rdma_replych) && ((entry->msg.request.rsphdr_count != 1) || (entry->msg.request.rsphdr_vec[0].iov_base == NULL))) { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, (entry->msg.request.rsphdr_count == 1) ? "chunktype specified as reply chunk but the vector " "specifying the buffer to be used for holding reply" @@ -649,10 +1426,10 @@ __rdma_create_write_chunks (rdma_peer_t *peer, rdma_ioq_t *entry, } /* - if ((chunk_type == rdma_writech) + if ((chunk_type == gf_rdma_writech) && ((entry->msg.request.rsphdr_count == 0) || (entry->msg.request.rsphdr_vec[0].iov_base == NULL))) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "vector specifying buffer to hold the program's reply " "header should also be provided when buffers are " "provided for holding the program's payload in reply"); @@ -660,21 +1437,21 @@ __rdma_create_write_chunks (rdma_peer_t *peer, rdma_ioq_t *entry, } */ - if (chunk_type == rdma_writech) { - warray = (rdma_write_array_t *)*ptr; + if (chunk_type == gf_rdma_writech) { + warray = (gf_rdma_write_array_t *)*ptr; warray->wc_discrim = hton32 (1); warray->wc_nchunks = hton32 (entry->msg.request.rsp_payload_count); *ptr = (uint32_t *)&warray->wc_array[0]; - ret = __rdma_create_write_chunks_from_vector (peer, - (rdma_write_chunk_t **)ptr, - entry->msg.request.rsp_payload, - entry->msg.request.rsp_payload_count, - request_ctx); + ret = __gf_rdma_create_write_chunks_from_vector (peer, + (gf_rdma_write_chunk_t **)ptr, + entry->msg.request.rsp_payload, + entry->msg.request.rsp_payload_count, + request_ctx); if (ret == -1) { - gf_log (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; @@ -692,19 +1469,19 @@ __rdma_create_write_chunks (rdma_peer_t *peer, rdma_ioq_t *entry, **ptr = 0; *ptr = *ptr + 1; - warray = (rdma_write_array_t *)*ptr; + warray = (gf_rdma_write_array_t *)*ptr; warray->wc_discrim = hton32 (1); warray->wc_nchunks = hton32 (entry->msg.request.rsphdr_count); *ptr = (uint32_t *)&warray->wc_array[0]; - ret = __rdma_create_write_chunks_from_vector (peer, - (rdma_write_chunk_t **)ptr, - entry->msg.request.rsphdr_vec, - entry->msg.request.rsphdr_count, - request_ctx); + ret = __gf_rdma_create_write_chunks_from_vector (peer, + (gf_rdma_write_chunk_t **)ptr, + entry->msg.request.rsphdr_vec, + entry->msg.request.rsphdr_count, + request_ctx); if (ret == -1) { - gf_log (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; @@ -720,8 +1497,8 @@ out: } -inline void -__rdma_deregister_mr (struct ibv_mr **mr, int count) +static inline void +__gf_rdma_deregister_mr (struct ibv_mr **mr, int count) { int i = 0; @@ -739,15 +1516,15 @@ out: static int32_t -__rdma_quota_put (rdma_peer_t *peer) +__gf_rdma_quota_put (gf_rdma_peer_t *peer) { - int32_t ret; + int32_t ret = 0; peer->quota++; ret = peer->quota; if (!list_empty (&peer->ioq)) { - ret = __rdma_ioq_churn (peer); + ret = __gf_rdma_ioq_churn (peer); } return ret; @@ -755,14 +1532,15 @@ __rdma_quota_put (rdma_peer_t *peer) static int32_t -rdma_quota_put (rdma_peer_t *peer) +gf_rdma_quota_put (gf_rdma_peer_t *peer) { - int32_t ret; - rdma_private_t *priv = peer->trans->private; + int32_t ret = 0; + gf_rdma_private_t *priv = NULL; + priv = peer->trans->private; pthread_mutex_lock (&priv->write_mutex); { - ret = __rdma_quota_put (peer); + ret = __gf_rdma_quota_put (peer); } pthread_mutex_unlock (&priv->write_mutex); @@ -772,11 +1550,11 @@ rdma_quota_put (rdma_peer_t *peer) /* to be called with priv->mutex held */ void -__rdma_request_context_destroy (rdma_request_context_t *context) +__gf_rdma_request_context_destroy (gf_rdma_request_context_t *context) { - rdma_peer_t *peer = NULL; - rdma_private_t *priv = NULL; - int32_t ret = 0; + gf_rdma_peer_t *peer = NULL; + gf_rdma_private_t *priv = NULL; + int32_t ret = 0; if (context == NULL) { goto out; @@ -784,18 +1562,18 @@ __rdma_request_context_destroy (rdma_request_context_t *context) peer = context->peer; - __rdma_deregister_mr (context->mr, context->mr_count); + __gf_rdma_deregister_mr (context->mr, context->mr_count); priv = peer->trans->private; if (priv->connected) { - ret = __rdma_quota_put (peer); + ret = __gf_rdma_quota_put (peer); if (ret < 0) { gf_log ("rdma", GF_LOG_DEBUG, "failed to send " "message"); - mem_put (context->pool, context); - __rdma_disconnect (peer->trans); + mem_put (context); + __gf_rdma_disconnect (peer->trans); goto out; } } @@ -810,22 +1588,21 @@ __rdma_request_context_destroy (rdma_request_context_t *context) context->rsp_iobref = NULL; } - mem_put (context->pool, context); + mem_put (context); out: return; } - void -rdma_post_context_destroy (rdma_post_context_t *ctx) +gf_rdma_post_context_destroy (gf_rdma_post_context_t *ctx) { if (ctx == NULL) { goto out; } - __rdma_deregister_mr (ctx->mr, ctx->mr_count); + __gf_rdma_deregister_mr (ctx->mr, ctx->mr_count); if (ctx->iobref != NULL) { iobref_unref (ctx->iobref); @@ -841,30 +1618,8 @@ out: } -static int32_t -rdma_post_recv (struct ibv_srq *srq, - 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; - - rdma_post_ref (post); - - return ibv_post_srq_recv (srq, &wr, &bad_wr); -} - - int -rdma_post_unref (rdma_post_t *post) +gf_rdma_post_unref (gf_rdma_post_t *post) { int refcount = -1; @@ -879,11 +1634,11 @@ rdma_post_unref (rdma_post_t *post) pthread_mutex_unlock (&post->lock); if (refcount == 0) { - rdma_post_context_destroy (&post->ctx); - if (post->type == RDMA_SEND_POST) { - rdma_put_post (&post->device->sendq, post); + gf_rdma_post_context_destroy (&post->ctx); + if (post->type == GF_RDMA_SEND_POST) { + gf_rdma_put_post (&post->device->sendq, post); } else { - rdma_post_recv (post->device->srq, post); + gf_rdma_post_recv (post->device->srq, post); } } out: @@ -892,7 +1647,7 @@ out: int -rdma_post_get_refcount (rdma_post_t *post) +gf_rdma_post_get_refcount (gf_rdma_post_t *post) { int refcount = -1; @@ -910,8 +1665,8 @@ out: return refcount; } -rdma_post_t * -rdma_post_ref (rdma_post_t *post) +gf_rdma_post_t * +gf_rdma_post_ref (gf_rdma_post_t *post) { if (post == NULL) { goto out; @@ -929,31 +1684,32 @@ out: int32_t -__rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_post_t *post) +__gf_rdma_ioq_churn_request (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, + gf_rdma_post_t *post) { - rdma_chunktype_t rtype = rdma_noch, wtype = rdma_noch; - uint64_t send_size = 0; - rdma_header_t *hdr = NULL; - struct rpc_msg *rpc_msg = NULL; - uint32_t *chunkptr = NULL; - char *buf = NULL; - int32_t ret = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - int chunk_count = 0; - rdma_request_context_t *request_ctx = NULL; - uint32_t prog_payload_length = 0, len = 0; - struct rpc_req *rpc_req = NULL; - - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, peer, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, entry, out); - GF_VALIDATE_OR_GOTO (RDMA_LOG_NAME, post, out); + gf_rdma_chunktype_t rtype = gf_rdma_noch; + gf_rdma_chunktype_t wtype = gf_rdma_noch; + uint64_t send_size = 0; + gf_rdma_header_t *hdr = NULL; + struct rpc_msg *rpc_msg = NULL; + uint32_t *chunkptr = NULL; + char *buf = NULL; + int32_t ret = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + int chunk_count = 0; + gf_rdma_request_context_t *request_ctx = NULL; + uint32_t prog_payload_length = 0, len = 0; + struct rpc_req *rpc_req = NULL; + + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, peer, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, entry, out); + GF_VALIDATE_OR_GOTO (GF_RDMA_LOG_NAME, post, out); if ((entry->msg.request.rsphdr_count != 0) && (entry->msg.request.rsp_payload_count != 0)) { ret = -1; - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "both write-chunklist and reply-chunk cannot be " "present"); goto out; @@ -963,7 +1719,7 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, priv = peer->trans->private; device = priv->device; - hdr = (rdma_header_t *)post->buf; + hdr = (gf_rdma_header_t *)post->buf; send_size = iov_length (entry->rpchdr, entry->rpchdr_count) + iov_length (entry->proghdr, entry->proghdr_count) @@ -976,43 +1732,43 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, } if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - rtype = rdma_areadch; + rtype = gf_rdma_areadch; } else if ((send_size + prog_payload_length) < GLUSTERFS_RDMA_INLINE_THRESHOLD) { - rtype = rdma_noch; + rtype = gf_rdma_noch; } else if (entry->prog_payload_count != 0) { - rtype = rdma_readch; + rtype = gf_rdma_readch; } if (entry->msg.request.rsphdr_count != 0) { - wtype = rdma_replych; + wtype = gf_rdma_replych; } else if (entry->msg.request.rsp_payload_count != 0) { - wtype = rdma_writech; + wtype = gf_rdma_writech; } - if (rtype == rdma_readch) { + if (rtype == gf_rdma_readch) { chunk_count += entry->prog_payload_count; - } else if (rtype == rdma_areadch) { + } else if (rtype == gf_rdma_areadch) { chunk_count += entry->rpchdr_count; chunk_count += entry->proghdr_count; } - if (wtype == rdma_writech) { + if (wtype == gf_rdma_writech) { chunk_count += entry->msg.request.rsp_payload_count; - } else if (wtype == rdma_replych) { + } else if (wtype == gf_rdma_replych) { chunk_count += entry->msg.request.rsphdr_count; } - if (chunk_count > RDMA_MAX_SEGMENTS) { + if (chunk_count > GF_RDMA_MAX_SEGMENTS) { ret = -1; - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "chunk count(%d) exceeding maximum allowed RDMA " - "segment count(%d)", chunk_count, RDMA_MAX_SEGMENTS); + "segment count(%d)", chunk_count, GF_RDMA_MAX_SEGMENTS); goto out; } - if ((wtype != rdma_noch) || (rtype != rdma_noch)) { - request_ctx = mem_get (priv->device->request_ctx_pool); + if ((wtype != gf_rdma_noch) || (rtype != gf_rdma_noch)) { + request_ctx = mem_get (device->request_ctx_pool); if (request_ctx == NULL) { ret = -1; goto out; @@ -1020,7 +1776,7 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, memset (request_ctx, 0, sizeof (*request_ctx)); - request_ctx->pool = priv->device->request_ctx_pool; + request_ctx->pool = device->request_ctx_pool; request_ctx->peer = peer; entry->msg.request.rpc_req->conn_private = request_ctx; @@ -1037,20 +1793,21 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, * since rpc_msg->rm_xid is already * hton32ed value of actual xid */ - hdr->rm_vers = hton32 (RDMA_VERSION); + hdr->rm_vers = hton32 (GF_RDMA_VERSION); hdr->rm_credit = hton32 (peer->send_count); - if (rtype == rdma_areadch) { - hdr->rm_type = hton32 (RDMA_NOMSG); + if (rtype == gf_rdma_areadch) { + hdr->rm_type = hton32 (GF_RDMA_NOMSG); } else { - hdr->rm_type = hton32 (RDMA_MSG); + hdr->rm_type = hton32 (GF_RDMA_MSG); } chunkptr = &hdr->rm_body.rm_chunks[0]; - if (rtype != rdma_noch) { - ret = __rdma_create_read_chunks (peer, entry, rtype, &chunkptr, - request_ctx); + if (rtype != gf_rdma_noch) { + ret = __gf_rdma_create_read_chunks (peer, entry, rtype, + &chunkptr, + request_ctx); if (ret != 0) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "creation of read chunks failed"); goto out; } @@ -1058,11 +1815,12 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, *chunkptr++ = 0; /* no read chunks */ } - if (wtype != rdma_noch) { - ret = __rdma_create_write_chunks (peer, entry, wtype, &chunkptr, - request_ctx); + if (wtype != gf_rdma_noch) { + ret = __gf_rdma_create_write_chunks (peer, entry, wtype, + &chunkptr, + request_ctx); if (ret != 0) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "creation of write/reply chunk failed"); goto out; } @@ -1073,14 +1831,14 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, buf = (char *)chunkptr; - if (rtype != rdma_areadch) { + if (rtype != gf_rdma_areadch) { iov_unload (buf, entry->rpchdr, entry->rpchdr_count); buf += iov_length (entry->rpchdr, entry->rpchdr_count); iov_unload (buf, entry->proghdr, entry->proghdr_count); buf += iov_length (entry->proghdr, entry->proghdr_count); - if (rtype != rdma_readch) { + if (rtype != gf_rdma_readch) { iov_unload (buf, entry->prog_payload, entry->prog_payload_count); buf += iov_length (entry->prog_payload, @@ -1090,18 +1848,18 @@ __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, len = buf - post->buf; - rdma_post_ref (post); + gf_rdma_post_ref (post); - ret = rdma_post_send (peer->qp, post, len); + ret = gf_rdma_post_send (peer->qp, post, len); if (!ret) { ret = len; } else { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, - "rdma_post_send (to %s) failed with ret = %d (%s)", + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "gf_rdma_post_send (to %s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); - rdma_post_unref (post); - __rdma_disconnect (peer->trans); + gf_rdma_post_unref (post); + __gf_rdma_disconnect (peer->trans); ret = -1; } @@ -1110,7 +1868,7 @@ out: rpc_req = entry->msg.request.rpc_req; if (request_ctx != NULL) { - __rdma_request_context_destroy (rpc_req->conn_private); + __gf_rdma_request_context_destroy (rpc_req->conn_private); } rpc_req->conn_private = NULL; @@ -1120,9 +1878,9 @@ out: } -inline void -__rdma_fill_reply_header (rdma_header_t *header, struct iovec *rpchdr, - rdma_reply_info_t *reply_info, int credits) +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) { struct rpc_msg *rpc_msg = NULL; @@ -1136,8 +1894,8 @@ __rdma_fill_reply_header (rdma_header_t *header, struct iovec *rpchdr, header->rm_xid = rpc_msg->rm_xid; } - header->rm_type = hton32 (RDMA_MSG); - header->rm_vers = hton32 (RDMA_VERSION); + header->rm_type = hton32 (GF_RDMA_MSG); + header->rm_vers = hton32 (GF_RDMA_VERSION); header->rm_credit = hton32 (credits); header->rm_body.rm_chunks[0] = 0; /* no read chunks */ @@ -1149,36 +1907,36 @@ __rdma_fill_reply_header (rdma_header_t *header, struct iovec *rpchdr, int32_t -__rdma_send_reply_inline (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_post_t *post, rdma_reply_info_t *reply_info) +__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) { - rdma_header_t *header = NULL; - int32_t send_size = 0, ret = 0; - char *buf = NULL; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - - priv = peer->trans->private; - device = priv->device; + gf_rdma_header_t *header = NULL; + int32_t send_size = 0, ret = 0; + char *buf = NULL; send_size = iov_length (entry->rpchdr, entry->rpchdr_count) + iov_length (entry->proghdr, entry->proghdr_count) + iov_length (entry->prog_payload, entry->prog_payload_count) - + sizeof (rdma_header_t); /* - * remember, no chunklists in the - * reply - */ + + sizeof (gf_rdma_header_t); /* + * remember, no chunklists in the + * reply + */ if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - ret = __rdma_send_error (peer, entry, post, reply_info, - ERR_CHUNK); + 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; } - header = (rdma_header_t *)post->buf; + header = (gf_rdma_header_t *)post->buf; - __rdma_fill_reply_header (header, entry->rpchdr, reply_info, - peer->send_count); + __gf_rdma_fill_reply_header (header, entry->rpchdr, reply_info, + peer->send_count); buf = (char *)&header->rm_body.rm_chunks[3]; @@ -1199,18 +1957,18 @@ __rdma_send_reply_inline (rdma_peer_t *peer, rdma_ioq_t *entry, entry->prog_payload_count); } - rdma_post_ref (post); + gf_rdma_post_ref (post); - ret = rdma_post_send (peer->qp, post, (buf - post->buf)); + ret = gf_rdma_post_send (peer->qp, post, (buf - post->buf)); if (!ret) { ret = send_size; } else { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, - "rdma_post_send (to %s) failed with ret = %d (%s)", + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "posting send (to %s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); - rdma_post_unref (post); - __rdma_disconnect (peer->trans); + gf_rdma_post_unref (post); + __gf_rdma_disconnect (peer->trans); ret = -1; } @@ -1220,17 +1978,18 @@ out: int32_t -__rdma_reply_encode_write_chunks (rdma_peer_t *peer, uint32_t payload_size, - rdma_post_t *post, - rdma_reply_info_t *reply_info, - uint32_t **ptr) +__gf_rdma_reply_encode_write_chunks (gf_rdma_peer_t *peer, + uint32_t payload_size, + gf_rdma_post_t *post, + gf_rdma_reply_info_t *reply_info, + uint32_t **ptr) { - uint32_t chunk_size = 0; - int32_t ret = -1; - rdma_write_array_t *target_array = NULL; - int i = 0; + uint32_t chunk_size = 0; + int32_t ret = -1; + gf_rdma_write_array_t *target_array = NULL; + int i = 0; - target_array = (rdma_write_array_t *)*ptr; + target_array = (gf_rdma_write_array_t *)*ptr; for (i = 0; i < reply_info->wc_array->wc_nchunks; i++) { chunk_size += @@ -1238,7 +1997,7 @@ __rdma_reply_encode_write_chunks (rdma_peer_t *peer, uint32_t payload_size, } if (chunk_size < payload_size) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "length of payload (%d) is exceeding the total " "write chunk length (%d)", payload_size, chunk_size); goto out; @@ -1269,18 +2028,17 @@ out: inline int32_t -__rdma_register_local_mr_for_rdma (rdma_peer_t *peer, - struct iovec *vector, int count, - rdma_post_context_t *ctx) +__gf_rdma_register_local_mr_for_rdma (gf_rdma_peer_t *peer, + struct iovec *vector, int count, + gf_rdma_post_context_t *ctx) { - int i = 0; - int32_t ret = -1; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; + int i = 0; + int32_t ret = -1; + 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; @@ -1301,6 +2059,9 @@ __rdma_register_local_mr_for_rdma (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; } @@ -1316,15 +2077,15 @@ out: * 2. modifies vec */ int32_t -__rdma_write (rdma_peer_t *peer, rdma_post_t *post, struct iovec *vec, - uint32_t xfer_len, int *idx, rdma_write_chunk_t *writech) +__gf_rdma_write (gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *vec, + uint32_t xfer_len, int *idx, gf_rdma_write_chunk_t *writech) { - int size = 0, num_sge = 0, i = 0; - int32_t ret = -1; - struct ibv_sge *sg_list = NULL; - struct ibv_send_wr wr = { - .opcode = IBV_WR_RDMA_WRITE, - .send_flags = IBV_SEND_SIGNALED, + int size = 0, num_sge = 0, i = 0; + int32_t ret = -1; + struct ibv_sge *sg_list = NULL; + struct ibv_send_wr wr = { + .opcode = IBV_WR_RDMA_WRITE, + .send_flags = IBV_SEND_SIGNALED, }, *bad_wr; if ((peer == NULL) || (writech == NULL) || (idx == NULL) @@ -1338,7 +2099,8 @@ __rdma_write (rdma_peer_t *peer, rdma_post_t *post, struct iovec *vec, num_sge = i - *idx; - sg_list = GF_CALLOC (num_sge, sizeof (struct ibv_sge), gf_common_mt_sge); + sg_list = GF_CALLOC (num_sge, sizeof (struct ibv_sge), + gf_common_mt_sge); if (sg_list == NULL) { ret = -1; goto out; @@ -1364,13 +2126,14 @@ __rdma_write (rdma_peer_t *peer, rdma_post_t *post, struct iovec *vec, wr.sg_list = sg_list; wr.num_sge = num_sge; - wr.wr_id = (unsigned long) rdma_post_ref (post); + wr.wr_id = (unsigned long) gf_rdma_post_ref (post); wr.wr.rdma.rkey = writech->wc_target.rs_handle; wr.wr.rdma.remote_addr = writech->wc_target.rs_offset; ret = ibv_post_send(peer->qp, &wr, &bad_wr); if (ret) { - gf_log (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) : ""); @@ -1384,13 +2147,14 @@ out: int32_t -__rdma_do_rdma_write (rdma_peer_t *peer, rdma_post_t *post, - struct iovec *vector, int count, struct iobref *iobref, - rdma_reply_info_t *reply_info) +__gf_rdma_do_gf_rdma_write (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + struct iovec *vector, int count, + struct iobref *iobref, + gf_rdma_reply_info_t *reply_info) { - int i = 0, payload_idx = 0; - uint32_t payload_size = 0, xfer_len = 0; - int32_t ret = -1; + int i = 0, payload_idx = 0; + uint32_t payload_size = 0, xfer_len = 0; + int32_t ret = -1; if (count != 0) { payload_size = iov_length (vector, count); @@ -1401,9 +2165,11 @@ __rdma_do_rdma_write (rdma_peer_t *peer, rdma_post_t *post, goto out; } - ret = __rdma_register_local_mr_for_rdma (peer, vector, count, - &post->ctx); + 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; } @@ -1415,9 +2181,13 @@ __rdma_do_rdma_write (rdma_peer_t *peer, rdma_post_t *post, xfer_len = min (payload_size, reply_info->wc_array->wc_array[i].wc_target.rs_length); - ret = __rdma_write (peer, post, vector, xfer_len, &payload_idx, - &reply_info->wc_array->wc_array[i]); + 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; } @@ -1432,44 +2202,41 @@ out: int32_t -__rdma_send_reply_type_nomsg (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_post_t *post, rdma_reply_info_t *reply_info) +__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) { - rdma_header_t *header = NULL; - char *buf = NULL; - uint32_t payload_size = 0; - int count = 0, i = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - 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]; - priv = peer->trans->private; - device = priv->device; + header = (gf_rdma_header_t *)post->buf; - header = (rdma_header_t *)post->buf; + __gf_rdma_fill_reply_header (header, entry->rpchdr, reply_info, + peer->send_count); - __rdma_fill_reply_header (header, entry->rpchdr, reply_info, - peer->send_count); - - header->rm_type = hton32 (RDMA_NOMSG); + header->rm_type = hton32 (GF_RDMA_NOMSG); payload_size = iov_length (entry->rpchdr, entry->rpchdr_count) + iov_length (entry->proghdr, entry->proghdr_count); /* encode reply chunklist */ buf = (char *)&header->rm_body.rm_chunks[2]; - ret = __rdma_reply_encode_write_chunks (peer, payload_size, post, - reply_info, (uint32_t **)&buf); + ret = __gf_rdma_reply_encode_write_chunks (peer, payload_size, post, + reply_info, + (uint32_t **)&buf); if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "encoding write chunks failed"); - ret = __rdma_send_error (peer, entry, post, reply_info, - ERR_CHUNK); + ret = __gf_rdma_send_error (peer, entry, post, reply_info, + ERR_CHUNK); goto out; } - rdma_post_ref (post); + gf_rdma_post_ref (post); for (i = 0; i < entry->rpchdr_count; i++) { vector[count++] = entry->rpchdr[i]; @@ -1479,21 +2246,24 @@ __rdma_send_reply_type_nomsg (rdma_peer_t *peer, rdma_ioq_t *entry, vector[count++] = entry->proghdr[i]; } - ret = __rdma_do_rdma_write (peer, post, vector, count, entry->iobref, - reply_info); + ret = __gf_rdma_do_gf_rdma_write (peer, post, vector, count, + entry->iobref, reply_info); if (ret == -1) { - rdma_post_unref (post); + 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; } - ret = rdma_post_send (peer->qp, post, (buf - post->buf)); + ret = gf_rdma_post_send (peer->qp, post, (buf - post->buf)); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, - "rdma_post_send to client (%s) failed with " + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "posting a send request to client (%s) failed with " "ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); ret = -1; - rdma_post_unref (post); + gf_rdma_post_unref (post); } else { ret = payload_size; } @@ -1504,65 +2274,64 @@ out: int32_t -__rdma_send_reply_type_msg (rdma_peer_t *peer, rdma_ioq_t *entry, - rdma_post_t *post, rdma_reply_info_t *reply_info) +__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) { - rdma_header_t *header = NULL; - int32_t send_size = 0, ret = 0; - char *ptr = NULL; - uint32_t payload_size = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - - priv = peer->trans->private; - device = priv->device; + gf_rdma_header_t *header = NULL; + int32_t send_size = 0, ret = 0; + char *ptr = NULL; + uint32_t payload_size = 0; send_size = iov_length (entry->rpchdr, entry->rpchdr_count) + iov_length (entry->proghdr, entry->proghdr_count) + GLUSTERFS_RDMA_MAX_HEADER_SIZE; if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "client has provided only write chunks, but the " "combined size of rpc and program header (%d) is " "exceeding the size of msg that can be sent using " "RDMA send (%d)", send_size, GLUSTERFS_RDMA_INLINE_THRESHOLD); - ret = __rdma_send_error (peer, entry, post, reply_info, - ERR_CHUNK); + ret = __gf_rdma_send_error (peer, entry, post, reply_info, + ERR_CHUNK); goto out; } - header = (rdma_header_t *)post->buf; + header = (gf_rdma_header_t *)post->buf; - __rdma_fill_reply_header (header, entry->rpchdr, reply_info, - peer->send_count); + __gf_rdma_fill_reply_header (header, entry->rpchdr, reply_info, + peer->send_count); payload_size = iov_length (entry->prog_payload, entry->prog_payload_count); ptr = (char *)&header->rm_body.rm_chunks[1]; - ret = __rdma_reply_encode_write_chunks (peer, payload_size, post, - reply_info, (uint32_t **)&ptr); + ret = __gf_rdma_reply_encode_write_chunks (peer, payload_size, post, + reply_info, + (uint32_t **)&ptr); if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "encoding write chunks failed"); - ret = __rdma_send_error (peer, entry, post, reply_info, - ERR_CHUNK); + ret = __gf_rdma_send_error (peer, entry, post, reply_info, + ERR_CHUNK); goto out; } *(uint32_t *)ptr = 0; /* terminate reply chunklist */ ptr += sizeof (uint32_t); - rdma_post_ref (post); + gf_rdma_post_ref (post); - ret = __rdma_do_rdma_write (peer, post, entry->prog_payload, - entry->prog_payload_count, entry->iobref, - reply_info); + ret = __gf_rdma_do_gf_rdma_write (peer, post, entry->prog_payload, + entry->prog_payload_count, + entry->iobref, reply_info); if (ret == -1) { - rdma_post_unref (post); + 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; } @@ -1572,13 +2341,13 @@ __rdma_send_reply_type_msg (rdma_peer_t *peer, rdma_ioq_t *entry, iov_unload (ptr, entry->proghdr, entry->proghdr_count); ptr += iov_length (entry->proghdr, entry->proghdr_count); - ret = rdma_post_send (peer->qp, post, (ptr - post->buf)); + ret = gf_rdma_post_send (peer->qp, post, (ptr - post->buf)); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "rdma send to client (%s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); - rdma_post_unref (post); + gf_rdma_post_unref (post); ret = -1; } else { ret = send_size + payload_size; @@ -1590,7 +2359,7 @@ out: void -rdma_reply_info_destroy (rdma_reply_info_t *reply_info) +gf_rdma_reply_info_destroy (gf_rdma_reply_info_t *reply_info) { if (reply_info == NULL) { goto out; @@ -1601,17 +2370,17 @@ rdma_reply_info_destroy (rdma_reply_info_t *reply_info) reply_info->wc_array = NULL; } - mem_put (reply_info->pool, reply_info); + mem_put (reply_info); out: return; } -rdma_reply_info_t * -rdma_reply_info_alloc (rdma_peer_t *peer) +gf_rdma_reply_info_t * +gf_rdma_reply_info_alloc (gf_rdma_peer_t *peer) { - rdma_reply_info_t *reply_info = NULL; - rdma_private_t *priv = NULL; + gf_rdma_reply_info_t *reply_info = NULL; + gf_rdma_private_t *priv = NULL; priv = peer->trans->private; @@ -1629,15 +2398,16 @@ out: int32_t -__rdma_ioq_churn_reply (rdma_peer_t *peer, rdma_ioq_t *entry, rdma_post_t *post) +__gf_rdma_ioq_churn_reply (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry, + gf_rdma_post_t *post) { - rdma_reply_info_t *reply_info = NULL; - int32_t ret = -1; - rdma_chunktype_t type = rdma_noch; + gf_rdma_reply_info_t *reply_info = NULL; + 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) { @@ -1645,29 +2415,47 @@ __rdma_ioq_churn_reply (rdma_peer_t *peer, rdma_ioq_t *entry, rdma_post_t *post) } switch (type) { - case rdma_noch: - ret = __rdma_send_reply_inline (peer, entry, post, reply_info); + 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 rdma_replych: - ret = __rdma_send_reply_type_nomsg (peer, entry, post, - reply_info); + 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 rdma_writech: - ret = __rdma_send_reply_type_msg (peer, entry, post, - reply_info); + 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 (RDMA_LOG_NAME, GF_LOG_WARNING, - "invalid chunktype (%d) specified for sending reply", - type); + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "invalid chunktype (%d) specified for sending reply " + " (peer:%s)", type, peer->trans->peerinfo.identifier); break; } if (reply_info != NULL) { - rdma_reply_info_destroy (reply_info); + gf_rdma_reply_info_destroy (reply_info); } out: return ret; @@ -1675,40 +2463,54 @@ out: int32_t -__rdma_ioq_churn_entry (rdma_peer_t *peer, rdma_ioq_t *entry) +__gf_rdma_ioq_churn_entry (gf_rdma_peer_t *peer, gf_rdma_ioq_t *entry) { - int32_t ret = 0, quota = 0; - rdma_private_t *priv = NULL; - rdma_device_t *device = NULL; - rdma_options_t *options = NULL; - rdma_post_t *post = NULL; + int32_t ret = 0, quota = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + gf_rdma_options_t *options = NULL; + gf_rdma_post_t *post = NULL; priv = peer->trans->private; options = &priv->options; device = priv->device; - quota = __rdma_quota_get (peer); + quota = __gf_rdma_quota_get (peer); if (quota > 0) { - post = rdma_get_post (&device->sendq); + post = gf_rdma_get_post (&device->sendq); if (post == NULL) { - post = rdma_new_post (device, - (options->send_size + 2048), - RDMA_SEND_POST); + 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 = __rdma_ioq_churn_request (peer, entry, post); + 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 = __rdma_ioq_churn_reply (peer, entry, post); + 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) { - __rdma_ioq_entry_free (entry); + __gf_rdma_ioq_entry_free (entry); } } else { ret = 0; @@ -1720,17 +2522,17 @@ out: static int32_t -__rdma_ioq_churn (rdma_peer_t *peer) +__gf_rdma_ioq_churn (gf_rdma_peer_t *peer) { - rdma_ioq_t *entry = NULL; - int32_t ret = 0; + gf_rdma_ioq_t *entry = NULL; + int32_t ret = 0; while (!list_empty (&peer->ioq)) { /* pick next entry */ entry = peer->ioq_next; - ret = __rdma_ioq_churn_entry (peer, entry); + ret = __gf_rdma_ioq_churn_entry (peer, entry); if (ret <= 0) break; @@ -1738,7 +2540,7 @@ __rdma_ioq_churn (rdma_peer_t *peer) /* list_for_each_entry_safe (entry, dummy, &peer->ioq, list) { - ret = __rdma_ioq_churn_entry (peer, entry); + ret = __gf_rdma_ioq_churn_entry (peer, entry); if (ret <= 0) { break; } @@ -1750,28 +2552,35 @@ __rdma_ioq_churn (rdma_peer_t *peer) static int32_t -rdma_writev (rpc_transport_t *this, - rdma_ioq_t *entry) +gf_rdma_writev (rpc_transport_t *this, gf_rdma_ioq_t *entry) { - int32_t ret = 0, need_append = 1; - rdma_private_t *priv = this->private; - rdma_peer_t *peer = NULL; + int32_t ret = 0, need_append = 1; + gf_rdma_private_t *priv = NULL; + gf_rdma_peer_t *peer = NULL; + priv = this->private; pthread_mutex_lock (&priv->write_mutex); { if (!priv->connected) { - gf_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; } peer = &priv->peer; if (list_empty (&peer->ioq)) { - ret = __rdma_ioq_churn_entry (peer, 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); + } } } @@ -1785,20 +2594,20 @@ unlock: } -rdma_ioq_t * -rdma_ioq_new (rpc_transport_t *this, rpc_transport_data_t *data) +gf_rdma_ioq_t * +gf_rdma_ioq_new (rpc_transport_t *this, rpc_transport_data_t *data) { - rdma_ioq_t *entry = NULL; - int count = 0, i = 0; - rpc_transport_msg_t *msg = NULL; - rdma_private_t *priv = NULL; + gf_rdma_ioq_t *entry = NULL; + int count = 0, i = 0; + rpc_transport_msg_t *msg = NULL; + gf_rdma_private_t *priv = NULL; if ((data == NULL) || (this == NULL)) { goto out; } priv = this->private; - /* TODO: use mem-pool */ + entry = mem_get (priv->device->ioq_pool); if (entry == NULL) { goto out; @@ -1876,12 +2685,11 @@ out: int32_t -rdma_submit_request (rpc_transport_t *this, - rpc_transport_req_t *req) +gf_rdma_submit_request (rpc_transport_t *this, rpc_transport_req_t *req) { - int32_t ret = 0; - rdma_ioq_t *entry = NULL; - rpc_transport_data_t data = {0, }; + int32_t ret = 0; + gf_rdma_ioq_t *entry = NULL; + rpc_transport_data_t data = {0, }; if (req == NULL) { goto out; @@ -1890,16 +2698,22 @@ rdma_submit_request (rpc_transport_t *this, data.is_request = 1; data.data.req = *req; - entry = rdma_ioq_new (this, &data); + 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; } - ret = rdma_writev (this, entry); + ret = gf_rdma_writev (this, entry); 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); } @@ -1908,11 +2722,11 @@ out: } int32_t -rdma_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply) +gf_rdma_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply) { - int32_t ret = 0; - rdma_ioq_t *entry = NULL; - rpc_transport_data_t data = {0, }; + int32_t ret = 0; + gf_rdma_ioq_t *entry = NULL; + rpc_transport_data_t data = {0, }; if (reply == NULL) { goto out; @@ -1920,15 +2734,21 @@ rdma_submit_reply (rpc_transport_t *this, rpc_transport_reply_t *reply) data.data.reply = *reply; - entry = rdma_ioq_new (this, &data); + 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; } - ret = rdma_writev (this, entry); + ret = gf_rdma_writev (this, entry); 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); } @@ -1936,196 +2756,15 @@ out: return ret; } -#if 0 -static int -rdma_receive (rpc_transport_t *this, char **hdr_p, size_t *hdrlen_p, - struct iobuf **iobuf_p) -{ - 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; - 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 = (rdma_header_t *)copy_from; - if (strcmp (header->colonO, ":O")) { - gf_log (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 (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_get (this->ctx->iobuf_pool); - 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 -rdma_destroy_cq (rpc_transport_t *this) -{ - rdma_private_t *priv = this->private; - rdma_device_t *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 -rdma_create_cq (rpc_transport_t *this) -{ - rdma_private_t *priv = NULL; - rdma_options_t *options = NULL; - 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 (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 (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 (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 (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 (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) - rdma_destroy_cq (this); - - return ret; -} - static int -rdma_register_peer (rdma_device_t *device, - int32_t qp_num, - rdma_peer_t *peer) +gf_rdma_register_peer (gf_rdma_device_t *device, int32_t qp_num, + gf_rdma_peer_t *peer) { - struct _qpent *ent = NULL; - rdma_qpreg_t *qpreg = NULL; - int32_t hash = 0; - int ret = -1; + struct _qpent *ent = NULL; + gf_rdma_qpreg_t *qpreg = NULL; + int32_t hash = 0; + int ret = -1; qpreg = &device->qpreg; hash = qp_num % 42; @@ -2166,37 +2805,41 @@ unlock: static void -rdma_unregister_peer (rdma_device_t *device, - int32_t qp_num) +gf_rdma_unregister_peer (gf_rdma_device_t *device, int32_t qp_num) { - struct _qpent *ent; - rdma_qpreg_t *qpreg = &device->qpreg; - int32_t hash = qp_num % 42; + struct _qpent *ent = NULL; + gf_rdma_qpreg_t *qpreg = NULL; + int32_t hash = 0; + + qpreg = &device->qpreg; + hash = qp_num % 42; pthread_mutex_lock (&qpreg->lock); - ent = qpreg->ents[hash].next; - while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) - ent = ent->next; - if (ent->qp_num != qp_num) { - pthread_mutex_unlock (&qpreg->lock); - return; - } - ent->prev->next = ent->next; - ent->next->prev = ent->prev; - /* TODO: unref reg->peer */ - GF_FREE (ent); - qpreg->count--; + { + ent = qpreg->ents[hash].next; + while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) + ent = ent->next; + if (ent->qp_num != qp_num) { + pthread_mutex_unlock (&qpreg->lock); + return; + } + ent->prev->next = ent->next; + ent->next->prev = ent->prev; + /* TODO: unref reg->peer */ + GF_FREE (ent); + qpreg->count--; + } pthread_mutex_unlock (&qpreg->lock); } -static rdma_peer_t * -__rdma_lookup_peer (rdma_device_t *device, int32_t qp_num) +static gf_rdma_peer_t * +__gf_rdma_lookup_peer (gf_rdma_device_t *device, int32_t qp_num) { - struct _qpent *ent = NULL; - rdma_peer_t *peer = NULL; - rdma_qpreg_t *qpreg = NULL; - int32_t hash = 0; + struct _qpent *ent = NULL; + gf_rdma_peer_t *peer = NULL; + gf_rdma_qpreg_t *qpreg = NULL; + int32_t hash = 0; qpreg = &device->qpreg; hash = qp_num % 42; @@ -2211,34 +2854,16 @@ __rdma_lookup_peer (rdma_device_t *device, int32_t qp_num) return peer; } -/* - static rdma_peer_t * - rdma_lookup_peer (rdma_device_t *device, - int32_t qp_num) - { - rdma_qpreg_t *qpreg = NULL; - rdma_peer_t *peer = NULL; - - qpreg = &device->qpreg; - pthread_mutex_lock (&qpreg->lock); - { - peer = __rdma_lookup_peer (device, qp_num); - } - pthread_mutex_unlock (&qpreg->lock); - - return peer; - } -*/ - static void -__rdma_destroy_qp (rpc_transport_t *this) +__gf_rdma_destroy_qp (rpc_transport_t *this) { - rdma_private_t *priv = this->private; + gf_rdma_private_t *priv = NULL; + priv = this->private; if (priv->peer.qp) { - rdma_unregister_peer (priv->device, priv->peer.qp->qp_num); - ibv_destroy_qp (priv->peer.qp); + gf_rdma_unregister_peer (priv->device, priv->peer.qp->qp_num); + rdma_destroy_qp (priv->peer.cm_id); } priv->peer.qp = NULL; @@ -2247,15 +2872,38 @@ __rdma_destroy_qp (rpc_transport_t *this) static int32_t -rdma_create_qp (rpc_transport_t *this) +gf_rdma_create_qp (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - rdma_options_t *options = &priv->options; - rdma_device_t *device = priv->device; - int32_t ret = 0; - rdma_peer_t *peer; + gf_rdma_private_t *priv = NULL; + gf_rdma_device_t *device = NULL; + int32_t ret = 0; + gf_rdma_peer_t *peer = NULL; + char *device_name = NULL; + + priv = this->private; peer = &priv->peer; + + device_name = (char *)ibv_get_device_name (peer->cm_id->verbs->device); + if (device_name == NULL) { + ret = -1; + gf_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, @@ -2269,333 +2917,73 @@ 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 (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 (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 = 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 = rdma_register_peer (device, peer->qp->qp_num, peer); + ret = gf_rdma_register_peer (device, peer->qp->qp_num, peer); out: if (ret == -1) - __rdma_destroy_qp (this); + __gf_rdma_destroy_qp (this); return ret; } -static void -rdma_destroy_posts (rpc_transport_t *this) -{ - -} - - static int32_t -__rdma_create_posts (rpc_transport_t *this, int32_t count, int32_t size, - rdma_queue_t *q, rdma_post_type_t type) +__gf_rdma_teardown (rpc_transport_t *this) { - int32_t i; - int32_t ret = 0; - rdma_private_t *priv = this->private; - rdma_device_t *device = priv->device; - - for (i=0 ; i<count ; i++) { - rdma_post_t *post; - - post = rdma_new_post (device, size + 2048, type); - if (!post) { - gf_log (RDMA_LOG_NAME, - GF_LOG_ERROR, - "%s: post creation failed", - this->name); - ret = -1; - break; - } - - rdma_put_post (q, post); - } - return ret; -} - + gf_rdma_private_t *priv = NULL; + gf_rdma_peer_t *peer = NULL; -static int32_t -rdma_create_posts (rpc_transport_t *this) -{ - int32_t i, ret; - rdma_post_t *post = NULL; - rdma_private_t *priv = this->private; - rdma_options_t *options = &priv->options; - rdma_device_t *device = priv->device; - - ret = __rdma_create_posts (this, options->send_count, - options->send_size, - &device->sendq, RDMA_SEND_POST); - if (!ret) - ret = __rdma_create_posts (this, options->recv_count, - options->recv_size, - &device->recvq, RDMA_RECV_POST); + priv = this->private; + peer = &priv->peer; - if (!ret) { - for (i=0 ; i<options->recv_count ; i++) { - post = rdma_get_post (&device->recvq); - if (rdma_post_recv (device->srq, post) != 0) { - ret = -1; - break; - } - } + if (peer->cm_id->qp != NULL) { + __gf_rdma_destroy_qp (this); } - if (ret) - rdma_destroy_posts (this); - - return ret; -} - - -static int32_t -rdma_connect_qp (rpc_transport_t *this) -{ - rdma_private_t *priv = this->private; - 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 (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); } - /* TODO: make timeout and retry_cnt configurable from options */ - attr.qp_state = IBV_QPS_RTS; - attr.timeout = 14; - attr.retry_cnt = 7; - attr.rnr_retry = 7; - 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 (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 -__rdma_teardown (rpc_transport_t *this) +gf_rdma_teardown (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - - __rdma_destroy_qp (this); + int32_t ret = 0; + gf_rdma_private_t *priv = NULL; - if (!list_empty (&priv->peer.ioq)) { - __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) -{ - rdma_private_t *priv = NULL; - int sock = -1; - int ret = -1; - struct iovec *opvector = vector; - int opcount = count; - int moved = 0; - priv = this->private; - sock = priv->sock; - 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; - rdma_private_t *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; } @@ -2607,12 +2995,12 @@ __tcp_writev (rpc_transport_t *this, struct iovec *vector, int count, * event is sent to upper layers. */ int32_t -rdma_get_write_chunklist (char **ptr, rdma_write_array_t **write_ary) +gf_rdma_get_write_chunklist (char **ptr, gf_rdma_write_array_t **write_ary) { - rdma_write_array_t *from = NULL, *to = NULL; - int32_t ret = -1, size = 0, i = 0; + gf_rdma_write_array_t *from = NULL, *to = NULL; + int32_t ret = -1, size = 0, i = 0; - from = (rdma_write_array_t *) *ptr; + from = (gf_rdma_write_array_t *) *ptr; if (from->wc_discrim == 0) { ret = 0; goto out; @@ -2621,7 +3009,7 @@ rdma_get_write_chunklist (char **ptr, rdma_write_array_t **write_ary) from->wc_nchunks = ntoh32 (from->wc_nchunks); size = sizeof (*from) - + (sizeof (rdma_write_chunk_t) * from->wc_nchunks); + + (sizeof (gf_rdma_write_chunk_t) * from->wc_nchunks); to = GF_CALLOC (1, size, gf_common_mt_char); if (to == NULL) { @@ -2655,13 +3043,13 @@ out: * rdma-reads and hence readchunk-list can point to memory held by post. */ int32_t -rdma_get_read_chunklist (char **ptr, rdma_read_chunk_t **readch) +gf_rdma_get_read_chunklist (char **ptr, gf_rdma_read_chunk_t **readch) { - int32_t ret = -1; - rdma_read_chunk_t *chunk = NULL; - int i = 0; + int32_t ret = -1; + gf_rdma_read_chunk_t *chunk = NULL; + int i = 0; - chunk = (rdma_read_chunk_t *)*ptr; + chunk = (gf_rdma_read_chunk_t *)*ptr; if (chunk[0].rc_discrim == 0) { ret = 0; goto out; @@ -2687,25 +3075,30 @@ out: inline int32_t -rdma_decode_error_msg (rdma_peer_t *peer, rdma_post_t *post, - size_t bytes_in_post) +gf_rdma_decode_error_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + size_t bytes_in_post) { - 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 = (rdma_header_t *)post->buf; + header = (gf_rdma_header_t *)post->buf; header->rm_body.rm_error.rm_type = ntoh32 (header->rm_body.rm_error.rm_type); if (header->rm_body.rm_error.rm_type == ERR_VERS) { - header->rm_body.rm_error.rm_version.rdma_vers_low = - ntoh32 (header->rm_body.rm_error.rm_version.rdma_vers_low); - header->rm_body.rm_error.rm_version.rdma_vers_high = - ntoh32 (header->rm_body.rm_error.rm_version.rdma_vers_high); + header->rm_body.rm_error.rm_version.gf_rdma_vers_low = + ntoh32 (header->rm_body.rm_error.rm_version.gf_rdma_vers_low); + header->rm_body.rm_error.rm_version.gf_rdma_vers_high = + ntoh32 (header->rm_body.rm_error.rm_version.gf_rdma_vers_high); } - iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); + rpc_msg.rm_xid = header->rm_xid; + rpc_msg.rm_direction = REPLY; + rpc_msg.rm_reply.rp_stat = MSG_DENIED; + + iobuf = iobuf_get2 (peer->trans->ctx->iobuf_pool, bytes_in_post); if (iobuf == NULL) { ret = -1; goto out; @@ -2719,15 +3112,15 @@ rdma_decode_error_msg (rdma_peer_t *peer, 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; @@ -2749,30 +3142,34 @@ out: int32_t -rdma_decode_msg (rdma_peer_t *peer, rdma_post_t *post, - rdma_read_chunk_t **readch, size_t bytes_in_post) +gf_rdma_decode_msg (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + gf_rdma_read_chunk_t **readch, size_t bytes_in_post) { - int32_t ret = -1; - rdma_header_t *header = NULL; - rdma_reply_info_t *reply_info = NULL; - char *ptr = NULL; - rdma_write_array_t *write_ary = NULL; - size_t header_len = 0; + int32_t ret = -1; + gf_rdma_header_t *header = NULL; + gf_rdma_reply_info_t *reply_info = NULL; + char *ptr = NULL; + gf_rdma_write_array_t *write_ary = NULL; + size_t header_len = 0; - header = (rdma_header_t *)post->buf; + header = (gf_rdma_header_t *)post->buf; ptr = (char *)&header->rm_body.rm_chunks[0]; - ret = rdma_get_read_chunklist (&ptr, readch); + 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; } /* skip terminator of read-chunklist */ ptr = ptr + sizeof (uint32_t); - ret = rdma_get_write_chunklist (&ptr, &write_ary); + 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; } @@ -2780,29 +3177,35 @@ rdma_decode_msg (rdma_peer_t *peer, rdma_post_t *post, ptr = ptr + sizeof (uint32_t); if (write_ary != NULL) { - reply_info = rdma_reply_info_alloc (peer); + 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; } - reply_info->type = rdma_writech; + reply_info->type = gf_rdma_writech; reply_info->wc_array = write_ary; reply_info->rm_xid = header->rm_xid; } else { - ret = rdma_get_write_chunklist (&ptr, &write_ary); + 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 = rdma_reply_info_alloc (peer); + 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; } - reply_info->type = rdma_replych; + reply_info->type = gf_rdma_replych; reply_info->wc_array = write_ary; reply_info->rm_xid = header->rm_xid; } @@ -2810,16 +3213,18 @@ rdma_decode_msg (rdma_peer_t *peer, rdma_post_t *post, /* skip terminator of reply chunk */ ptr = ptr + sizeof (uint32_t); - if (header->rm_type != RDMA_NOMSG) { - post->ctx.hdr_iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); + if (header->rm_type != GF_RDMA_NOMSG) { + header_len = (long)ptr - (long)post->buf; + post->ctx.vector[0].iov_len = (bytes_in_post - header_len); + + post->ctx.hdr_iobuf = iobuf_get2 (peer->trans->ctx->iobuf_pool, + (bytes_in_post - header_len)); if (post->ctx.hdr_iobuf == NULL) { ret = -1; goto out; } - header_len = (long)ptr - (long)post->buf; post->ctx.vector[0].iov_base = iobuf_ptr (post->ctx.hdr_iobuf); - post->ctx.vector[0].iov_len = bytes_in_post - header_len; memcpy (post->ctx.vector[0].iov_base, ptr, post->ctx.vector[0].iov_len); post->ctx.count = 1; @@ -2833,9 +3238,7 @@ out: *readch = NULL; } - if (write_ary != NULL) { - GF_FREE (write_ary); - } + GF_FREE (write_ary); } return ret; @@ -2844,13 +3247,13 @@ out: /* Assumes only one of either write-chunklist or a reply chunk is present */ int32_t -rdma_decode_header (rdma_peer_t *peer, rdma_post_t *post, - rdma_read_chunk_t **readch, size_t bytes_in_post) +gf_rdma_decode_header (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + gf_rdma_read_chunk_t **readch, size_t bytes_in_post) { - int32_t ret = -1; - rdma_header_t *header = NULL; + int32_t ret = -1; + gf_rdma_header_t *header = NULL; - header = (rdma_header_t *)post->buf; + header = (gf_rdma_header_t *)post->buf; header->rm_xid = ntoh32 (header->rm_xid); header->rm_vers = ntoh32 (header->rm_vers); @@ -2858,31 +3261,39 @@ rdma_decode_header (rdma_peer_t *peer, rdma_post_t *post, header->rm_type = ntoh32 (header->rm_type); switch (header->rm_type) { - case RDMA_MSG: - case RDMA_NOMSG: - ret = rdma_decode_msg (peer, post, readch, bytes_in_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 RDMA_MSGP: - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "rdma msg of msg-type RDMA_MSGP should not have been " - "received"); + case GF_RDMA_MSGP: + 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 RDMA_DONE: - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "rdma msg of msg-type RDMA_DONE should not have been " - "received"); + case GF_RDMA_DONE: + 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 RDMA_ERROR: - /* ret = rdma_decode_error_msg (peer, post, bytes_in_post); */ + case GF_RDMA_ERROR: + 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 (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "unknown rdma msg-type (%d)", header->rm_type); } @@ -2891,15 +3302,17 @@ rdma_decode_header (rdma_peer_t *peer, rdma_post_t *post, int32_t -__rdma_read (rdma_peer_t *peer, rdma_post_t *post, struct iovec *to, - rdma_read_chunk_t *readch) +__gf_rdma_read (gf_rdma_peer_t *peer, gf_rdma_post_t *post, struct iovec *to, + gf_rdma_read_chunk_t *readch) { - int32_t ret = -1; + int32_t ret = -1; struct ibv_sge list = {0, }; - struct ibv_send_wr wr = {0, }, *bad_wr = NULL; + struct ibv_send_wr wr = {0, }, *bad_wr = NULL; - ret = __rdma_register_local_mr_for_rdma (peer, to, 1, &post->ctx); + 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; } @@ -2907,7 +3320,7 @@ __rdma_read (rdma_peer_t *peer, rdma_post_t *post, struct iovec *to, list.length = to->iov_len; list.lkey = post->ctx.mr[post->ctx.mr_count - 1]->lkey; - wr.wr_id = (unsigned long) rdma_post_ref (post); + wr.wr_id = (unsigned long) gf_rdma_post_ref (post); wr.sg_list = &list; wr.num_sge = 1; wr.opcode = IBV_WR_RDMA_READ; @@ -2917,12 +3330,13 @@ __rdma_read (rdma_peer_t *peer, rdma_post_t *post, struct iovec *to, ret = ibv_post_send (peer->qp, &wr, &bad_wr); if (ret) { - gf_log (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) : ""); ret = -1; - rdma_post_unref (post); + gf_rdma_post_unref (post); } out: return ret; @@ -2930,13 +3344,14 @@ out: int32_t -rdma_do_reads (rdma_peer_t *peer, rdma_post_t *post, rdma_read_chunk_t *readch) +gf_rdma_do_reads (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + gf_rdma_read_chunk_t *readch) { - int32_t ret = -1, i = 0, count = 0; - size_t size = 0; - char *ptr = NULL; - struct iobuf *iobuf = NULL; - rdma_private_t *priv = NULL; + int32_t ret = -1, i = 0, count = 0; + size_t size = 0; + char *ptr = NULL; + struct iobuf *iobuf = NULL; + gf_rdma_private_t *priv = NULL; priv = peer->trans->private; @@ -2945,24 +3360,15 @@ rdma_do_reads (rdma_peer_t *peer, rdma_post_t *post, rdma_read_chunk_t *readch) } if (i == 0) { - gf_log (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; } - post->ctx.rdma_reads = i; - - if (size > peer->trans->ctx->page_size) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "total size of rdma-read (%lu) is greater than " - "page-size (%lu). This is not supported till variable " - "sized iobufs are implemented", (unsigned long)size, - (unsigned long)peer->trans->ctx->page_size); - goto out; - } + post->ctx.gf_rdma_reads = i; - iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); + iobuf = iobuf_get2 (peer->trans->ctx->iobuf_pool, size); if (iobuf == NULL) { goto out; } @@ -2984,6 +3390,10 @@ rdma_do_reads (rdma_peer_t *peer, rdma_post_t *post, rdma_read_chunk_t *readch) 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; } @@ -2993,10 +3403,13 @@ rdma_do_reads (rdma_peer_t *peer, rdma_post_t *post, rdma_read_chunk_t *readch) post->ctx.vector[count].iov_len = readch[i].rc_target.rs_length; - ret = __rdma_read (peer, post, - &post->ctx.vector[count], - &readch[i]); + ret = __gf_rdma_read (peer, 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; } @@ -3020,16 +3433,16 @@ out: int32_t -rdma_pollin_notify (rdma_peer_t *peer, rdma_post_t *post) +gf_rdma_pollin_notify (gf_rdma_peer_t *peer, gf_rdma_post_t *post) { - int32_t ret = -1; - enum msg_type msg_type = 0; - struct rpc_req *rpc_req = NULL; - rdma_request_context_t *request_context = NULL; - rpc_request_info_t request_info = {0, }; - rdma_private_t *priv = NULL; - uint32_t *ptr = NULL; - rpc_transport_pollin_t *pollin = NULL; + int32_t ret = -1; + enum msg_type msg_type = 0; + struct rpc_req *rpc_req = NULL; + gf_rdma_request_context_t *request_context = NULL; + rpc_request_info_t request_info = {0, }; + gf_rdma_private_t *priv = NULL; + uint32_t *ptr = NULL; + rpc_transport_pollin_t *pollin = NULL; if ((peer == NULL) || (post == NULL)) { goto out; @@ -3046,7 +3459,8 @@ rdma_pollin_notify (rdma_peer_t *peer, rdma_post_t *post) * because of server sending entire msg as inline without * doing rdma writes. */ - iobref_add (post->ctx.iobref, post->ctx.hdr_iobuf); + if (post->ctx.hdr_iobuf) + iobref_add (post->ctx.iobref, post->ctx.hdr_iobuf); } pollin = rpc_transport_pollin_alloc (peer->trans, @@ -3069,7 +3483,7 @@ rdma_pollin_notify (rdma_peer_t *peer, rdma_post_t *post) RPC_TRANSPORT_MAP_XID_REQUEST, &request_info); if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot get request information from rpc " "layer"); goto out; @@ -3077,7 +3491,7 @@ rdma_pollin_notify (rdma_peer_t *peer, rdma_post_t *post) rpc_req = request_info.rpc_req; if (rpc_req == NULL) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "rpc request structure not found"); ret = -1; goto out; @@ -3090,11 +3504,11 @@ rdma_pollin_notify (rdma_peer_t *peer, rdma_post_t *post) if (request_context != NULL) { pthread_mutex_lock (&priv->write_mutex); { - __rdma_request_context_destroy (request_context); + __gf_rdma_request_context_destroy (request_context); } pthread_mutex_unlock (&priv->write_mutex); } else { - rdma_quota_put (peer); + gf_rdma_quota_put (peer); } pollin->is_reply = 1; @@ -3102,6 +3516,10 @@ rdma_pollin_notify (rdma_peer_t *peer, 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) { @@ -3114,19 +3532,19 @@ out: int32_t -rdma_recv_reply (rdma_peer_t *peer, rdma_post_t *post) +gf_rdma_recv_reply (gf_rdma_peer_t *peer, gf_rdma_post_t *post) { - int32_t ret = -1; - rdma_header_t *header = NULL; - rdma_reply_info_t *reply_info = NULL; - rdma_write_array_t *wc_array = NULL; - int i = 0; - uint32_t *ptr = NULL; - rdma_request_context_t *ctx = NULL; - rpc_request_info_t request_info = {0, }; - struct rpc_req *rpc_req = NULL; - - header = (rdma_header_t *)post->buf; + int32_t ret = -1; + gf_rdma_header_t *header = NULL; + gf_rdma_reply_info_t *reply_info = NULL; + gf_rdma_write_array_t *wc_array = NULL; + int i = 0; + uint32_t *ptr = NULL; + gf_rdma_request_context_t *ctx = NULL; + rpc_request_info_t request_info = {0, }; + struct rpc_req *rpc_req = NULL; + + header = (gf_rdma_header_t *)post->buf; reply_info = post->ctx.reply_info; /* no write chunklist, just notify upper layers */ @@ -3137,7 +3555,7 @@ rdma_recv_reply (rdma_peer_t *peer, rdma_post_t *post) wc_array = reply_info->wc_array; - if (header->rm_type == RDMA_NOMSG) { + if (header->rm_type == GF_RDMA_NOMSG) { post->ctx.vector[0].iov_base = (void *)(long)wc_array->wc_array[0].wc_target.rs_offset; post->ctx.vector[0].iov_len @@ -3162,32 +3580,36 @@ rdma_recv_reply (rdma_peer_t *peer, rdma_post_t *post) RPC_TRANSPORT_MAP_XID_REQUEST, &request_info); if (ret == -1) { - gf_log (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 (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "rpc request structure not found"); ret = -1; goto out; } ctx = rpc_req->conn_private; - if (post->ctx.iobref == NULL) { + if ((post->ctx.iobref == NULL) && ctx->rsp_iobref) { post->ctx.iobref = iobref_ref (ctx->rsp_iobref); } ret = 0; - rdma_reply_info_destroy (reply_info); + gf_rdma_reply_info_destroy (reply_info); out: if (ret == 0) { - ret = rdma_pollin_notify (peer, post); + ret = gf_rdma_pollin_notify (peer, post); + if (ret < 0) { + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, + "pollin notify failed"); + } } return ret; @@ -3195,17 +3617,22 @@ out: inline int32_t -rdma_recv_request (rdma_peer_t *peer, rdma_post_t *post, - rdma_read_chunk_t *readch) +gf_rdma_recv_request (gf_rdma_peer_t *peer, gf_rdma_post_t *post, + gf_rdma_read_chunk_t *readch) { int32_t ret = -1; if (readch != NULL) { - ret = rdma_do_reads (peer, post, readch); + 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 = rdma_pollin_notify (peer, post); + ret = gf_rdma_pollin_notify (peer, post); if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "pollin notification failed"); } } @@ -3214,38 +3641,59 @@ rdma_recv_request (rdma_peer_t *peer, rdma_post_t *post, } void -rdma_process_recv (rdma_peer_t *peer, struct ibv_wc *wc) +gf_rdma_process_recv (gf_rdma_peer_t *peer, struct ibv_wc *wc) { - rdma_post_t *post = NULL; - rdma_read_chunk_t *readch = NULL; - int ret = -1; - uint32_t *ptr = NULL; - enum msg_type msg_type = 0; - rdma_header_t *header = NULL; - - post = (rdma_post_t *) (long) wc->wr_id; + gf_rdma_post_t *post = NULL; + gf_rdma_read_chunk_t *readch = NULL; + int ret = -1; + uint32_t *ptr = NULL; + enum msg_type msg_type = 0; + gf_rdma_header_t *header = NULL; + gf_rdma_private_t *priv = NULL; + + post = (gf_rdma_post_t *) (long) wc->wr_id; if (post == NULL) { - gf_log (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 = rdma_decode_header (peer, post, &readch, wc->byte_len); + ret = gf_rdma_decode_header (peer, post, &readch, wc->byte_len); if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "decoding of header failed"); goto out; } - header = (rdma_header_t *)post->buf; + 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 RDMA_MSG: + case GF_RDMA_MSG: ptr = (uint32_t *)post->ctx.vector[0].iov_base; msg_type = ntoh32 (*(ptr + 1)); break; - case RDMA_NOMSG: + case GF_RDMA_NOMSG: if (readch != NULL) { msg_type = CALL; } else { @@ -3253,31 +3701,49 @@ rdma_process_recv (rdma_peer_t *peer, struct ibv_wc *wc) } break; - case RDMA_ERROR: - gf_log (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 = rdma_pollin_notify (peer, post); - if (ret == -1) { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, - "pollin notification failed"); - } - goto out; -*/ + case GF_RDMA_ERROR: + 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 (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 = rdma_recv_request (peer, post, readch); + 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 = rdma_recv_reply (peer, post); + 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: @@ -3288,14 +3754,50 @@ 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 * -rdma_recv_completion_proc (void *data) +gf_rdma_recv_completion_proc (void *data) { struct ibv_comp_channel *chan = NULL; - rdma_device_t *device = NULL;; - rdma_post_t *post = NULL; - rdma_peer_t *peer = NULL; + gf_rdma_device_t *device = NULL;; + gf_rdma_post_t *post = NULL; + gf_rdma_peer_t *peer = NULL; struct ibv_cq *event_cq = NULL; struct ibv_wc wc = {0, }; void *event_ctx = NULL; @@ -3306,7 +3808,7 @@ rdma_recv_completion_proc (void *data) while (1) { ret = ibv_get_cq_event (chan, &event_cq, &event_ctx); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_get_cq_event failed, terminating recv " "thread %d (%d)", ret, errno); continue; @@ -3316,26 +3818,26 @@ rdma_recv_completion_proc (void *data) ret = ibv_req_notify_cq (event_cq, 0); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_req_notify_cq on %s failed, terminating " "recv thread: %d (%d)", device->device_name, ret, errno); continue; } - device = (rdma_device_t *) event_ctx; + device = (gf_rdma_device_t *) event_ctx; while ((ret = ibv_poll_cq (event_cq, 1, &wc)) > 0) { - post = (rdma_post_t *) (long) wc.wr_id; + post = (gf_rdma_post_t *) (long) wc.wr_id; pthread_mutex_lock (&device->qpreg.lock); { - peer = __rdma_lookup_peer (device, - wc.qp_num); + peer = __gf_rdma_lookup_peer (device, + wc.qp_num); /* * keep a refcount on transport so that it - * doesnot get freed because of some error + * does not get freed because of some error * indicated by wc.status till we are done * with usage of peer and thereby that of trans. */ @@ -3346,7 +3848,7 @@ rdma_recv_completion_proc (void *data) pthread_mutex_unlock (&device->qpreg.lock); if (wc.status != IBV_WC_SUCCESS) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "recv work request on `%s' returned " "error (%d)", device->device_name, wc.status); @@ -3356,26 +3858,26 @@ rdma_recv_completion_proc (void *data) } if (post) { - rdma_post_unref (post); + gf_rdma_post_unref (post); } continue; } if (peer) { - rdma_process_recv (peer, &wc); + gf_rdma_process_recv (peer, &wc); rpc_transport_unref (peer->trans); } else { - gf_log (RDMA_LOG_NAME, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "could not lookup peer for qp_num: %d", wc.qp_num); } - rdma_post_unref (post); + gf_rdma_post_unref (post); } if (ret < 0) { - gf_log (RDMA_LOG_NAME, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_poll_cq on `%s' returned error " "(ret = %d, errno = %d)", @@ -3390,11 +3892,11 @@ rdma_recv_completion_proc (void *data) void -rdma_handle_failed_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) +gf_rdma_handle_failed_send_completion (gf_rdma_peer_t *peer, struct ibv_wc *wc) { - rdma_post_t *post = NULL; - rdma_device_t *device = NULL; - rdma_private_t *priv = NULL; + gf_rdma_post_t *post = NULL; + gf_rdma_device_t *device = NULL; + gf_rdma_private_t *priv = NULL; if (peer != NULL) { priv = peer->trans->private; @@ -3404,9 +3906,9 @@ rdma_handle_failed_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) } - post = (rdma_post_t *) (long) wc->wr_id; + post = (gf_rdma_post_t *) (long) wc->wr_id; - gf_log (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", @@ -3431,20 +3933,22 @@ rdma_handle_failed_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) void -rdma_handle_successful_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) +gf_rdma_handle_successful_send_completion (gf_rdma_peer_t *peer, + struct ibv_wc *wc) { - rdma_post_t *post = NULL; - int reads = 0, ret = 0; + gf_rdma_post_t *post = NULL; + int reads = 0, ret = 0; + gf_rdma_header_t *header = NULL; if (wc->opcode != IBV_WC_RDMA_READ) { goto out; } - post = (rdma_post_t *)(long) wc->wr_id; + post = (gf_rdma_post_t *)(long) wc->wr_id; pthread_mutex_lock (&post->lock); { - reads = --post->ctx.rdma_reads; + reads = --post->ctx.gf_rdma_reads; } pthread_mutex_unlock (&post->lock); @@ -3453,7 +3957,14 @@ rdma_handle_successful_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) goto out; } - ret = rdma_pollin_notify (peer, post); + header = (gf_rdma_header_t *)post->buf; + + if (header->rm_type == GF_RDMA_NOMSG) { + post->ctx.count = 1; + post->ctx.vector[0].iov_len += post->ctx.vector[1].iov_len; + } + + ret = gf_rdma_pollin_notify (peer, post); if ((ret == -1) && (peer != NULL)) { rpc_transport_disconnect (peer->trans); } @@ -3464,14 +3975,14 @@ out: static void * -rdma_send_completion_proc (void *data) +gf_rdma_send_completion_proc (void *data) { struct ibv_comp_channel *chan = NULL; - rdma_post_t *post = NULL; - rdma_peer_t *peer = NULL; + gf_rdma_post_t *post = NULL; + gf_rdma_peer_t *peer = NULL; struct ibv_cq *event_cq = NULL; void *event_ctx = NULL; - rdma_device_t *device = NULL; + gf_rdma_device_t *device = NULL; struct ibv_wc wc = {0, }; char is_request = 0; int32_t ret = 0, quota_ret = 0; @@ -3480,7 +3991,7 @@ rdma_send_completion_proc (void *data) while (1) { ret = ibv_get_cq_event (chan, &event_cq, &event_ctx); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_get_cq_event on failed, terminating " "send thread: %d (%d)", ret, errno); continue; @@ -3490,7 +4001,7 @@ rdma_send_completion_proc (void *data) ret = ibv_req_notify_cq (event_cq, 0); if (ret) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_req_notify_cq on %s failed, terminating " "send thread: %d (%d)", device->device_name, ret, errno); @@ -3498,11 +4009,11 @@ rdma_send_completion_proc (void *data) } while ((ret = ibv_poll_cq (event_cq, 1, &wc)) > 0) { - post = (rdma_post_t *) (long) wc.wr_id; + post = (gf_rdma_post_t *) (long) wc.wr_id; pthread_mutex_lock (&device->qpreg.lock); { - peer = __rdma_lookup_peer (device, wc.qp_num); + peer = __gf_rdma_lookup_peer (device, wc.qp_num); /* * keep a refcount on transport so that it @@ -3517,32 +4028,32 @@ rdma_send_completion_proc (void *data) pthread_mutex_unlock (&device->qpreg.lock); if (wc.status != IBV_WC_SUCCESS) { - rdma_handle_failed_send_completion (peer, &wc); + gf_rdma_handle_failed_send_completion (peer, &wc); } else { - rdma_handle_successful_send_completion (peer, - &wc); + gf_rdma_handle_successful_send_completion (peer, + &wc); } if (post) { is_request = post->ctx.is_request; - ret = rdma_post_unref (post); + ret = gf_rdma_post_unref (post); if ((ret == 0) && (wc.status == IBV_WC_SUCCESS) && !is_request - && (post->type == RDMA_SEND_POST) + && (post->type == GF_RDMA_SEND_POST) && (peer != NULL)) { - /* An RDMA_RECV_POST can end up in - * rdma_send_completion_proc for + /* An GF_RDMA_RECV_POST can end up in + * gf_rdma_send_completion_proc for * rdma-reads, and we do not take - * quota for getting an RDMA_RECV_POST. + * quota for getting an GF_RDMA_RECV_POST. */ /* * if it is request, quota is returned * after reply has come. */ - quota_ret = rdma_quota_put (peer); + quota_ret = gf_rdma_quota_put (peer); if (quota_ret < 0) { gf_log ("rdma", GF_LOG_DEBUG, "failed to send " @@ -3554,14 +4065,14 @@ rdma_send_completion_proc (void *data) if (peer) { rpc_transport_unref (peer->trans); } else { - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_DEBUG, "could not lookup peer for qp_num: %d", wc.qp_num); } } if (ret < 0) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_poll_cq on `%s' returned error (ret = %d," " errno = %d)", device->device_name, ret, errno); @@ -3576,19 +4087,24 @@ rdma_send_completion_proc (void *data) static void -rdma_options_init (rpc_transport_t *this) +gf_rdma_options_init (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - rdma_options_t *options = &priv->options; - int32_t mtu; - data_t *temp; + gf_rdma_private_t *priv = NULL; + gf_rdma_options_t *options = NULL; + int32_t mtu = 0; + data_t *temp = NULL; /* TODO: validate arguments from options below */ + priv = this->private; + options = &priv->options; options->send_size = GLUSTERFS_RDMA_INLINE_THRESHOLD;/*this->ctx->page_size * 4; 512 KB*/ options->recv_size = GLUSTERFS_RDMA_INLINE_THRESHOLD;/*this->ctx->page_size * 4; 512 KB*/ options->send_count = 4096; options->recv_count = 4096; + options->attr_timeout = GF_RDMA_TIMEOUT; + options->attr_retry_cnt = GF_RDMA_RETRY_CNT; + options->attr_rnr_retry = GF_RDMA_RNR_RETRY; temp = dict_get (this->options, "transport.rdma.work-request-send-count"); @@ -3598,7 +4114,22 @@ rdma_options_init (rpc_transport_t *this) temp = dict_get (this->options, "transport.rdma.work-request-recv-count"); if (temp) - options->recv_count = data_to_int32 (temp); + options->recv_count = data_to_int32 (temp); + + temp = dict_get (this->options, "transport.rdma.attr-timeout"); + + if (temp) + options->attr_timeout = data_to_uint8 (temp); + + temp = dict_get (this->options, "transport.rdma.attr-retry-cnt"); + + if (temp) + options->attr_retry_cnt = data_to_uint8 (temp); + + temp = dict_get (this->options, "transport.rdma.attr-rnr-retry"); + + if (temp) + options->attr_rnr_retry = data_to_uint8 (temp); options->port = 1; temp = dict_get (this->options, @@ -3624,12 +4155,12 @@ rdma_options_init (rpc_transport_t *this) break; default: if (temp) - gf_log (RDMA_LOG_NAME, GF_LOG_WARNING, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_WARNING, "%s: unrecognized MTU value '%s', defaulting " "to '2048'", this->name, data_to_str (temp)); else - gf_log (RDMA_LOG_NAME, GF_LOG_TRACE, + gf_log (GF_RDMA_LOG_NAME, GF_LOG_TRACE, "%s: defaulting MTU to '2048'", this->name); options->mtu = IBV_MTU_2048; @@ -3644,1111 +4175,250 @@ rdma_options_init (rpc_transport_t *this) return; } -static void -rdma_queue_init (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 rdma_device_t * -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; - rdma_private_t *priv = NULL; - 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_ctx_t *rdma_ctx = NULL; + int ret = -1; - rdma_device_t *trav = NULL; - - 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 (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 (rdma_request_context_t, - RDMA_POOL_SIZE); - if (trav->request_ctx_pool == NULL) { - return NULL; - } - - trav->ioq_pool = mem_pool_new (rdma_ioq_t, RDMA_POOL_SIZE); - if (trav->ioq_pool == NULL) { - mem_pool_destroy (trav->request_ctx_pool); - return NULL; - } - - trav->reply_info_pool = mem_pool_new (rdma_reply_info_t, - 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 (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 (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 (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 (RDMA_LOG_NAME, GF_LOG_ERROR, - "could not create recv completion channel"); - /* TODO: cleanup current mess */ - return NULL; - } - - if (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 (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); - 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 (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); - 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 (RDMA_LOG_NAME, GF_LOG_ERROR, - "%s: could not create SRQ", - this->name); - return NULL; - } - - /* queue init */ - rdma_queue_init (&trav->sendq); - rdma_queue_init (&trav->recvq); - - if (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); - 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 (RDMA_LOG_NAME, GF_LOG_ERROR, - "%s: could not allocate posts", - this->name); - return NULL; - } + 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; + } - /* completion threads */ - ret = pthread_create (&trav->send_thread, - NULL, - rdma_send_completion_proc, - trav->send_chan); - if (ret) { - 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); - 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 (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, - rdma_recv_completion_proc, - trav->recv_chan); - if (ret) { - 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); - 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 (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 -rdma_init (rpc_transport_t *this) +gf_rdma_init (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - rdma_options_t *options = &priv->options; - struct ibv_device **dev_list; - struct ibv_context *ib_ctx = NULL; - int32_t ret = 0; - - rdma_options_init (this); - - { - dev_list = ibv_get_device_list (NULL); - - if (!dev_list) { - gf_log (RDMA_LOG_NAME, - GF_LOG_CRITICAL, - "Failed to get IB devices"); - ret = -1; - goto cleanup; - } + gf_rdma_private_t *priv = NULL; + int32_t ret = 0; + glusterfs_ctx_t *ctx = NULL; + gf_rdma_options_t *options = NULL; - if (!*dev_list) { - gf_log (RDMA_LOG_NAME, - GF_LOG_CRITICAL, - "No IB devices found"); - ret = -1; - goto cleanup; - } + ctx= this->ctx; - if (!options->device_name) { - if (*dev_list) { - options->device_name = - gf_strdup (ibv_get_device_name (*dev_list)); - } else { - gf_log (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 (RDMA_LOG_NAME, - GF_LOG_ERROR, - "Failed to get infiniband" - "device context"); - ret = -1; - goto cleanup; - } - break; - } - ++dev_list; - } + priv = this->private; - priv->device = rdma_get_device (this, ib_ctx); + ibv_fork_init (); + gf_rdma_options_init (this); - if (!priv->device) { - gf_log (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); - } - - if (dev_list) - ibv_free_device_list (dev_list); - - return ret; -} - - -static int32_t -rdma_disconnect (rpc_transport_t *this) -{ - rdma_private_t *priv = this->private; - int32_t ret = 0; - - pthread_mutex_lock (&priv->write_mutex); + pthread_mutex_lock (&ctx->lock); { - ret = __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 -rdma_fill_handshake_data (char *buf, struct rdma_nbio *nbio, - 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 -rdma_fill_handshake_ack (char *buf, struct 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 -rdma_handshake_pollin (rpc_transport_t *this) -{ - int ret = 0; - rdma_private_t *priv = this->private; - char *buf = priv->handshake.incoming.buf; - int32_t recv_buf_size, send_buf_size; - socklen_t sock_len; - - if (priv->handshake.incoming.state == RDMA_HANDSHAKE_COMPLETE) { - return -1; - } - - pthread_mutex_lock (&priv->write_mutex); - { - while (priv->handshake.incoming.state != RDMA_HANDSHAKE_COMPLETE) - { - switch (priv->handshake.incoming.state) - { - case RDMA_HANDSHAKE_START: - buf = priv->handshake.incoming.buf = GF_CALLOC (1, 256, gf_common_mt_char); - rdma_fill_handshake_data (buf, &priv->handshake.incoming, priv); - buf[0] = 0; - priv->handshake.incoming.state = RDMA_HANDSHAKE_RECEIVING_DATA; - break; - - case 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 = RDMA_HANDSHAKE_RECEIVED_DATA; - } - break; - - case 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 (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 (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 (rdma_connect_qp (this)) { - gf_log (RDMA_LOG_NAME, - GF_LOG_ERROR, - "%s: failed to connect with " - "remote QP", this->name); - ret = -1; - goto unlock; - } - rdma_fill_handshake_ack (buf, &priv->handshake.incoming); - buf[0] = 0; - priv->handshake.incoming.state = RDMA_HANDSHAKE_RECEIVING_ACK; - break; - - case 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 = RDMA_HANDSHAKE_RECEIVED_ACK; - } - break; - - case RDMA_HANDSHAKE_RECEIVED_ACK: - if (strncmp (buf, "DONE", 4)) { - gf_log (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 = RDMA_HANDSHAKE_COMPLETE; + if (ctx->ib == NULL) { + ctx->ib = __gf_rdma_ctx_create (); + if (ctx->ib == NULL) { + ret = -1; } } } -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); - } - } + pthread_mutex_unlock (&ctx->lock); return ret; } -static int -rdma_handshake_pollout (rpc_transport_t *this) -{ - rdma_private_t *priv = this->private; - char *buf = priv->handshake.outgoing.buf; - int32_t ret = 0; - - if (priv->handshake.outgoing.state == RDMA_HANDSHAKE_COMPLETE) { - return 0; - } - - pthread_mutex_unlock (&priv->write_mutex); - { - while (priv->handshake.outgoing.state != RDMA_HANDSHAKE_COMPLETE) - { - switch (priv->handshake.outgoing.state) - { - case RDMA_HANDSHAKE_START: - buf = priv->handshake.outgoing.buf = GF_CALLOC (1, 256, gf_common_mt_char); - rdma_fill_handshake_data (buf, &priv->handshake.outgoing, priv); - priv->handshake.outgoing.state = RDMA_HANDSHAKE_SENDING_DATA; - break; - - case 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 = RDMA_HANDSHAKE_SENT_DATA; - } - break; - - case RDMA_HANDSHAKE_SENT_DATA: - rdma_fill_handshake_ack (buf, &priv->handshake.outgoing); - priv->handshake.outgoing.state = RDMA_HANDSHAKE_SENDING_ACK; - break; - - case 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 = 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 -rdma_handshake_pollerr (rpc_transport_t *this) +static int32_t +gf_rdma_disconnect (rpc_transport_t *this) { - rdma_private_t *priv = this->private; - int32_t ret = 0; - char need_unref = 0, connected = 0; - - gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, - "%s: peer disconnected, cleaning up", - this->name); + gf_rdma_private_t *priv = NULL; + int32_t ret = 0; - pthread_mutex_lock (&priv->write_mutex); - { - __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 (RDMA_LOG_NAME, GF_LOG_ERROR, - "close () - error: %s", - strerror (errno)); - ret = -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 = RDMA_HANDSHAKE_START; - - if (priv->handshake.outgoing.buf) { - GF_FREE (priv->handshake.outgoing.buf); - priv->handshake.outgoing.buf = NULL; - } - - priv->handshake.outgoing.state = 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) -{ - rdma_private_t *priv = this->private; - int error = 0, ret = 0; + priv = this->private; + gf_log_callingfn (this->name, GF_LOG_WARNING, + "disconnect called (peer:%s)", + this->peerinfo.identifier); 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 -rdma_event_handler (int fd, int idx, void *data, - int poll_in, int poll_out, int poll_err) -{ - rpc_transport_t *this = data; - rdma_private_t *priv = this->private; - rdma_options_t *options = NULL; - int ret = 0; - - 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 = rdma_create_qp (this)) < 0) { - gf_log (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 = rdma_handshake_pollout (this); - } - - if (!ret && !poll_err && poll_in && priv->tcp_connected) { - if (priv->handshake.incoming.state == RDMA_HANDSHAKE_COMPLETE) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "%s: pollin received on tcp socket (peer: %s) " - "after handshake is complete", - this->name, this->peerinfo.identifier); - rdma_handshake_pollerr (this); - return 0; - } - ret = rdma_handshake_pollin (this); - } - - if (ret < 0 || poll_err) { - ret = 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 -rdma_connect (struct rpc_transport *this, int port) +gf_rdma_connect (struct rpc_transport *this, int port) { - dict_t *options = this->options; - - rdma_private_t *priv = this->private; + gf_rdma_private_t *priv = NULL; + int32_t ret = 0; + union gf_sock_union sock_union = {{0, }, }; + socklen_t sockaddr_len = 0; + gf_rdma_peer_t *peer = NULL; + gf_rdma_ctx_t *rdma_ctx = NULL; + gf_boolean_t connected = _gf_false; - int32_t ret = 0; - gf_boolean_t non_blocking = 1; - struct sockaddr_storage sockaddr; - socklen_t sockaddr_len = 0; - - if (priv->connected) { - return 0; - } + priv = this->private; - if (dict_get (options, "non-blocking-io")) { - char *nb_connect = data_to_str (dict_get (this->options, - "non-blocking-io")); + peer = &priv->peer; - if (gf_string2boolean (nb_connect, &non_blocking) == -1) { - gf_log (this->name, GF_LOG_ERROR, - "'non-blocking-io' takes only boolean " - "options, not taking any action"); - non_blocking = 1; - } - } + rpc_transport_ref (this); ret = gf_rdma_client_get_remote_sockaddr (this, - (struct sockaddr *)&sockaddr, + &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 (((struct sockaddr *)&sockaddr)->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, &sockaddr, sockaddr_len); + memcpy (&this->peerinfo.sockaddr, &sock_union.storage, + sockaddr_len); this->peerinfo.sockaddr_len = sockaddr_len; if (port > 0) - ((struct sockaddr_in *) (&sockaddr))->sin_port - = htons (port); + sock_union.sin.sin_port = htons (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) - { + &this->myinfo.sockaddr_len, + 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 = RDMA_HANDSHAKE_START; - priv->handshake.outgoing.state = RDMA_HANDSHAKE_START; - - priv->idx = event_register (this->ctx->event_pool, - priv->sock, rdma_event_handler, - this, 1, 1); + priv->connected = 0; } unlock: pthread_mutex_unlock (&priv->write_mutex); - return ret; -} - -static int -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, *trans = data; - rdma_private_t *priv = NULL; - rdma_private_t *trans_priv = (rdma_private_t *) trans->private; - rdma_options_t *options = NULL; - - if (!poll_in) { - return 0; - } - - 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 (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; +out: + if (ret != 0) { + if (!connected) { + gf_rdma_teardown (this); + } - 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; + rpc_transport_unref (this); } - priv->peer.trans = this; - rpc_transport_ref (this); - - gf_rdma_get_transport_identifiers (this); + return ret; +} - priv->tcp_connected = 1; - priv->handshake.incoming.state = RDMA_HANDSHAKE_START; - priv->handshake.outgoing.state = 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); +static int32_t +gf_rdma_listen (rpc_transport_t *this) +{ + union gf_sock_union sock_union = {{0, }, }; + socklen_t sockaddr_len = 0; + gf_rdma_private_t *priv = NULL; + gf_rdma_peer_t *peer = NULL; + int ret = 0; + gf_rdma_ctx_t *rdma_ctx = NULL; + char service[NI_MAXSERV], host[NI_MAXHOST]; - if (rdma_create_qp (this) < 0) { - gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, - "%s: could not create QP", - this->name); - rpc_transport_disconnect (this); - return -1; - } + priv = this->private; + peer = &priv->peer; - priv->idx = event_register (this->ctx->event_pool, priv->sock, - rdma_event_handler, this, 1, 1); + priv->entity = GF_RDMA_SERVER_LISTENER; - 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; -} + rdma_ctx = this->ctx->ib; -static int32_t -rdma_listen (rpc_transport_t *this) -{ - struct sockaddr_storage sockaddr; - socklen_t sockaddr_len; - rdma_private_t *priv = this->private; - int opt = 1, ret = 0; - char service[NI_MAXSERV], host[NI_MAXHOST]; - - memset (&sockaddr, 0, sizeof (sockaddr)); - ret = gf_rdma_server_get_local_sockaddr (this, - (struct sockaddr *)&sockaddr, + 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 (((struct sockaddr *)&sockaddr)->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, &sockaddr, 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) { @@ -4756,57 +4426,58 @@ 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, - (struct sockaddr *)&sockaddr, - 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, - 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 = rdma_submit_request, - .submit_reply = rdma_submit_reply, - .connect = rdma_connect, - .disconnect = rdma_disconnect, - .listen = rdma_listen, + .submit_request = gf_rdma_submit_request, + .submit_reply = gf_rdma_submit_reply, + .connect = gf_rdma_connect, + .disconnect = gf_rdma_disconnect, + .listen = gf_rdma_listen, }; int32_t init (rpc_transport_t *this) { - rdma_private_t *priv = NULL; + gf_rdma_private_t *priv = NULL; priv = GF_CALLOC (1, sizeof (*priv), gf_common_mt_rdma_private_t); if (!priv) return -1; this->private = priv; - priv->sock = -1; - if (rdma_init (this)) { + if (gf_rdma_init (this)) { gf_log (this->name, GF_LOG_ERROR, "Failed to initialize IB Device"); return -1; @@ -4819,19 +4490,15 @@ void fini (struct rpc_transport *this) { /* TODO: verify this function does graceful finish */ - rdma_private_t *priv = this->private; + gf_rdma_private_t *priv = NULL; + + priv = this->private; + this->private = NULL; 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); @@ -4871,6 +4538,18 @@ struct volume_options options[] = { "transport.rdma.remote-port"}, .type = GF_OPTION_TYPE_INT }, + { .key = {"transport.rdma.attr-timeout", + "rdma-attr-timeout"}, + .type = GF_OPTION_TYPE_INT + }, + { .key = {"transport.rdma.attr-retry-cnt", + "rdma-attr-retry-cnt"}, + .type = GF_OPTION_TYPE_INT + }, + { .key = {"transport.rdma.attr-rnr-retry", + "rdma-attr-rnr-retry"}, + .type = GF_OPTION_TYPE_INT + }, { .key = {"transport.rdma.listen-port", "listen-port"}, .type = GF_OPTION_TYPE_INT }, |