/* Copyright (c) 2006-2010 Gluster, Inc. This file is part of GlusterFS. GlusterFS is free software; you can redistribute it and/or modify it under the terms of the GNU 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 General Public License for more details. You should have received a copy of the GNU General Public License along with this program. If not, see . */ #ifndef _CONFIG_H #define _CONFIG_H #include "config.h" #endif #include "dict.h" #include "glusterfs.h" #include "logging.h" #include "rdma.h" #include "name.h" #include "byte-order.h" #include "xlator.h" #include #define RDMA_LOG_NAME "rpc-transport/rdma" static int32_t __rdma_ioq_churn (rdma_peer_t *peer); rdma_post_t * rdma_post_ref (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; if (ibv_query_port (context, port, &attr)) return 0; return attr.lid; } static const char * get_port_state_str(enum ibv_port_state pstate) { switch (pstate) { case IBV_PORT_DOWN: return "PORT_DOWN"; case IBV_PORT_INIT: return "PORT_INIT"; case IBV_PORT_ARMED: return "PORT_ARMED"; case IBV_PORT_ACTIVE: return "PORT_ACTIVE"; case IBV_PORT_ACTIVE_DEFER: return "PORT_ACTIVE_DEFER"; default: return "invalid state"; } } 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 (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; } 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; if (!ib_ctx) { gf_log (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; } 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; gf_log (RDMA_LOG_NAME, GF_LOG_TRACE, "Port:(%u) not active", ib_port); continue; } return ret; } static void rdma_put_post (rdma_queue_t *queue, rdma_post_t *post) { post->ctx.is_request = 0; pthread_mutex_lock (&queue->lock); { if (post->prev) { queue->active_count--; post->prev->next = post->next; } if (post->next) { post->next->prev = post->prev; } post->prev = &queue->passive_posts; post->next = post->prev->next; post->prev->next = post; post->next->prev = post; queue->passive_count++; } pthread_mutex_unlock (&queue->lock); } static rdma_post_t * rdma_new_post (rdma_device_t *device, int32_t len, rdma_post_type_t type) { rdma_post_t *post = NULL; int ret = -1; post = (rdma_post_t *) GF_CALLOC (1, sizeof (*post), gf_common_mt_rdma_post_t); if (post == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } pthread_mutex_init (&post->lock, NULL); post->buf_size = len; post->buf = valloc (len); if (!post->buf) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } post->mr = ibv_reg_mr (device->pd, post->buf, post->buf_size, IBV_ACCESS_LOCAL_WRITE); if (!post->mr) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "memory registration failed"); goto out; } post->device = device; post->type = type; ret = 0; out: if (ret != 0) { if (post->buf != NULL) { free (post->buf); } GF_FREE (post); post = NULL; } return post; } static rdma_post_t * rdma_get_post (rdma_queue_t *queue) { rdma_post_t *post; pthread_mutex_lock (&queue->lock); { post = queue->passive_posts.next; if (post == &queue->passive_posts) post = NULL; if (post) { if (post->prev) post->prev->next = post->next; if (post->next) post->next->prev = post->prev; post->prev = &queue->active_posts; post->next = post->prev->next; post->prev->next = post; post->next->prev = post; post->reused++; queue->active_count++; } } pthread_mutex_unlock (&queue->lock); return post; } void rdma_destroy_post (rdma_post_t *post) { ibv_dereg_mr (post->mr); free (post->buf); GF_FREE (post); } static int32_t __rdma_quota_get (rdma_peer_t *peer) { int32_t ret = -1; rdma_private_t *priv = peer->trans->private; if (priv->connected && peer->quota > 0) { ret = peer->quota--; } 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) { list_del_init (&entry->list); if (entry->iobref) { iobref_unref (entry->iobref); entry->iobref = NULL; } if (entry->msg.request.rsp_iobref) { iobref_unref (entry->msg.request.rsp_iobref); entry->msg.request.rsp_iobref = NULL; } /* TODO: use mem-pool */ mem_put (entry->pool, entry); } static void __rdma_ioq_flush (rdma_peer_t *peer) { rdma_ioq_t *entry = NULL, *dummy = NULL; list_for_each_entry_safe (entry, dummy, &peer->ioq, list) { __rdma_ioq_entry_free (entry); } } static int32_t __rdma_disconnect (rpc_transport_t *this) { rdma_private_t *priv = this->private; int32_t ret = 0; 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->connected = 0; } } return ret; } static int32_t rdma_post_send (struct ibv_qp *qp, rdma_post_t *post, int32_t len) { struct ibv_sge list = { .addr = (unsigned long) post->buf, .length = len, .lkey = post->mr->lkey }; struct ibv_send_wr wr = { .wr_id = (unsigned long) post, .sg_list = &list, .num_sge = 1, .opcode = IBV_WR_SEND, .send_flags = IBV_SEND_SIGNALED, }, *bad_wr; if (!qp) return EINVAL; return ibv_post_send (qp, &wr, &bad_wr); } int __rdma_encode_error(rdma_peer_t *peer, rdma_reply_info_t *reply_info, struct iovec *rpchdr, uint32_t *ptr, 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); } else { rpc_msg = rpchdr[0].iov_base; /* assume rpchdr contains * only one vector. * (which is true) */ *ptr++ = rpc_msg->rm_xid; } *ptr++ = hton32(RDMA_VERSION); *ptr++ = hton32(peer->send_count); *ptr++ = hton32(RDMA_ERROR); *ptr++ = hton32(err); if (err == ERR_VERS) { *ptr++ = hton32(RDMA_VERSION); *ptr++ = hton32(RDMA_VERSION); } return (int)((unsigned long)ptr - (unsigned long)startp); } 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) { int32_t ret = -1, len; len = __rdma_encode_error (peer, reply_info, entry->rpchdr, (uint32_t *)post->buf, err); if (len == -1) { goto out; } rdma_post_ref (post); ret = rdma_post_send (peer->qp, post, len); if (!ret) { ret = len; } else { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); ret = -1; } out: return ret; } 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) { 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; if ((peer == NULL) || (readch_ptr == NULL) || (*readch_ptr == NULL) || (request_ctx == NULL) || (vector == NULL)) { goto out; } priv = peer->trans->private; device = priv->device; readch = *readch_ptr; for (i = 0; i < count; i++) { readch->rc_discrim = hton32 (1); readch->rc_position = hton32 (*pos); mr = ibv_reg_mr (device->pd, vector[i].iov_base, vector[i].iov_len, IBV_ACCESS_REMOTE_READ); if (!mr) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "memory registration failed"); goto out; } request_ctx->mr[request_ctx->mr_count++] = mr; readch->rc_target.rs_handle = hton32 (mr->rkey); readch->rc_target.rs_length = hton32 (vector[i].iov_len); readch->rc_target.rs_offset = hton64 ((uint64_t)(unsigned long)vector[i].iov_base); *pos = *pos + vector[i].iov_len; readch++; } *readch_ptr = readch; ret = 0; out: return ret; } int32_t __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) { int32_t ret = -1; rdma_device_t *device = NULL; rdma_private_t *priv = NULL; int pos = 0; if ((peer == NULL) || (entry == NULL) || (ptr == NULL) || (*ptr == NULL) || (request_ctx == NULL)) { goto out; } priv = peer->trans->private; device = priv->device; request_ctx->iobref = iobref_ref (entry->iobref); if (type == 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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot create read chunks from vector, " "entry->prog_payload"); } } /* terminate read-chunk list*/ **ptr = 0; *ptr = *ptr + 1; out: return ret; } 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) { 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; if ((peer == NULL) || (writech_ptr == NULL) || (*writech_ptr == NULL) || (request_ctx == NULL) || (vector == NULL)) { goto out; } writech = *writech_ptr; priv = peer->trans->private; device = priv->device; for (i = 0; i < count; i++) { mr = ibv_reg_mr (device->pd, vector[i].iov_base, vector[i].iov_len, IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_LOCAL_WRITE); if (!mr) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "memory registration failed"); goto out; } request_ctx->mr[request_ctx->mr_count++] = mr; writech->wc_target.rs_handle = hton32 (mr->rkey); writech->wc_target.rs_length = hton32 (vector[i].iov_len); writech->wc_target.rs_offset = hton64 (((uint64_t)(unsigned long)vector[i].iov_base)); writech++; } *writech_ptr = writech; ret = 0; out: return ret; } int32_t __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) { int32_t ret = -1; rdma_write_array_t *warray = NULL; if ((peer == NULL) || (entry == NULL) || (ptr == NULL) || (*ptr == NULL) || (request_ctx == NULL)) { goto out; } if ((chunk_type == rdma_replych) && ((entry->msg.request.rsphdr_count != 1) || (entry->msg.request.rsphdr_vec[0].iov_base == NULL))) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, (entry->msg.request.rsphdr_count == 1) ? "chunktype specified as reply chunk but the vector " "specifying the buffer to be used for holding reply" " header is not correct" : "chunktype specified as reply chunk, but more than one " "buffer provided for holding reply"); goto out; } /* if ((chunk_type == 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, "vector specifying buffer to hold the program's reply " "header should also be provided when buffers are " "provided for holding the program's payload in reply"); goto out; } */ if (chunk_type == rdma_writech) { warray = (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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot create write chunks from vector " "entry->rpc_payload"); goto out; } /* terminate write chunklist */ **ptr = 0; *ptr = *ptr + 1; /* no reply chunklist */ **ptr = 0; *ptr = *ptr + 1; } else { /* no write chunklist */ **ptr = 0; *ptr = *ptr + 1; warray = (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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot create write chunks from vector " "entry->rpchdr"); goto out; } /* terminate reply chunklist */ **ptr = 0; *ptr = *ptr + 1; } out: return ret; } inline void __rdma_deregister_mr (struct ibv_mr **mr, int count) { int i = 0; if (mr == NULL) { goto out; } for (i = 0; i < count; i++) { ibv_dereg_mr (mr[i]); } out: return; } static int32_t __rdma_quota_put (rdma_peer_t *peer) { int32_t ret; peer->quota++; ret = peer->quota; if (!list_empty (&peer->ioq)) { ret = __rdma_ioq_churn (peer); } return ret; } static int32_t rdma_quota_put (rdma_peer_t *peer) { int32_t ret; rdma_private_t *priv = peer->trans->private; pthread_mutex_lock (&priv->write_mutex); { ret = __rdma_quota_put (peer); } pthread_mutex_unlock (&priv->write_mutex); return ret; } /* to be called with priv->mutex held */ void __rdma_request_context_destroy (rdma_request_context_t *context) { rdma_peer_t *peer = NULL; rdma_private_t *priv = NULL; int32_t ret = 0; if (context == NULL) { goto out; } peer = context->peer; __rdma_deregister_mr (context->mr, context->mr_count); priv = peer->trans->private; if (priv->connected) { ret = __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); goto out; } } if (context->iobref != NULL) { iobref_unref (context->iobref); context->iobref = NULL; } if (context->rsp_iobref != NULL) { iobref_unref (context->rsp_iobref); context->rsp_iobref = NULL; } mem_put (context->pool, context); out: return; } void rdma_post_context_destroy (rdma_post_context_t *ctx) { if (ctx == NULL) { goto out; } __rdma_deregister_mr (ctx->mr, ctx->mr_count); if (ctx->iobref != NULL) { iobref_unref (ctx->iobref); } if (ctx->hdr_iobuf != NULL) { iobuf_unref (ctx->hdr_iobuf); } memset (ctx, 0, sizeof (*ctx)); out: return; } 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) { int refcount = -1; if (post == NULL) { goto out; } pthread_mutex_lock (&post->lock); { refcount = --post->refcount; } 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); } else { rdma_post_recv (post->device->srq, post); } } out: return refcount; } int rdma_post_get_refcount (rdma_post_t *post) { int refcount = -1; if (post == NULL) { goto out; } pthread_mutex_lock (&post->lock); { refcount = post->refcount; } pthread_mutex_unlock (&post->lock); out: return refcount; } rdma_post_t * rdma_post_ref (rdma_post_t *post) { if (post == NULL) { goto out; } pthread_mutex_lock (&post->lock); { post->refcount++; } pthread_mutex_unlock (&post->lock); out: return post; } int32_t __rdma_ioq_churn_request (rdma_peer_t *peer, rdma_ioq_t *entry, 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; if ((peer == NULL) || (entry == NULL) || (post == NULL)) { goto out; } if ((entry->msg.request.rsphdr_count != 0) && (entry->msg.request.rsp_payload_count != 0)) { ret = -1; gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "both write-chunklist and reply-chunk cannot be " "present"); goto out; } post->ctx.is_request = 1; priv = peer->trans->private; device = priv->device; hdr = (rdma_header_t *)post->buf; send_size = iov_length (entry->rpchdr, entry->rpchdr_count) + iov_length (entry->proghdr, entry->proghdr_count) + GLUSTERFS_RDMA_MAX_HEADER_SIZE; if (entry->prog_payload_count != 0) { prog_payload_length = iov_length (entry->prog_payload, entry->prog_payload_count); } if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { rtype = rdma_areadch; } else if ((send_size + prog_payload_length) < GLUSTERFS_RDMA_INLINE_THRESHOLD) { rtype = rdma_noch; } else if (entry->prog_payload_count != 0) { rtype = rdma_readch; } if (entry->msg.request.rsphdr_count != 0) { wtype = rdma_replych; } else if (entry->msg.request.rsp_payload_count != 0) { wtype = rdma_writech; } if (rtype == rdma_readch) { chunk_count += entry->prog_payload_count; } else if (rtype == rdma_areadch) { chunk_count += entry->rpchdr_count; chunk_count += entry->proghdr_count; } if (wtype == rdma_writech) { chunk_count += entry->msg.request.rsp_payload_count; } else if (wtype == rdma_replych) { chunk_count += entry->msg.request.rsphdr_count; } if (chunk_count > RDMA_MAX_SEGMENTS) { ret = -1; gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "chunk count(%d) exceeding maximum allowed RDMA " "segment count(%d)", chunk_count, RDMA_MAX_SEGMENTS); goto out; } if ((wtype != rdma_noch) || (rtype != rdma_noch)) { request_ctx = mem_get (priv->device->request_ctx_pool); if (request_ctx == NULL) { ret = -1; gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } memset (request_ctx, 0, sizeof (*request_ctx)); request_ctx->pool = priv->device->request_ctx_pool; request_ctx->peer = peer; entry->msg.request.rpc_req->conn_private = request_ctx; if (entry->msg.request.rsp_iobref != NULL) { request_ctx->rsp_iobref = iobref_ref (entry->msg.request.rsp_iobref); } } rpc_msg = (struct rpc_msg *) entry->rpchdr[0].iov_base; hdr->rm_xid = rpc_msg->rm_xid; /* no need of hton32(rpc_msg->rm_xid), * since rpc_msg->rm_xid is already * hton32ed value of actual xid */ hdr->rm_vers = hton32 (RDMA_VERSION); hdr->rm_credit = hton32 (peer->send_count); if (rtype == rdma_areadch) { hdr->rm_type = hton32 (RDMA_NOMSG); } else { hdr->rm_type = hton32 (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 (ret != 0) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "creation of read chunks failed"); goto out; } } else { *chunkptr++ = 0; /* no read chunks */ } if (wtype != rdma_noch) { ret = __rdma_create_write_chunks (peer, entry, wtype, &chunkptr, request_ctx); if (ret != 0) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "creation of write/reply chunk failed"); goto out; } } else { *chunkptr++ = 0; /* no write chunks */ *chunkptr++ = 0; /* no reply chunk */ } buf = (char *)chunkptr; if (rtype != 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) { iov_unload (buf, entry->prog_payload, entry->prog_payload_count); buf += iov_length (entry->prog_payload, entry->prog_payload_count); } } len = buf - post->buf; rdma_post_ref (post); ret = rdma_post_send (peer->qp, post, len); if (!ret) { ret = len; } else { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); ret = -1; } out: if (ret == -1) { rpc_req = entry->msg.request.rpc_req; if (request_ctx != NULL) { __rdma_request_context_destroy (rpc_req->conn_private); } rpc_req->conn_private = NULL; } return ret; } inline void __rdma_fill_reply_header (rdma_header_t *header, struct iovec *rpchdr, rdma_reply_info_t *reply_info, int credits) { struct rpc_msg *rpc_msg = NULL; if (reply_info != NULL) { header->rm_xid = hton32 (reply_info->rm_xid); } else { rpc_msg = rpchdr[0].iov_base; /* assume rpchdr contains * only one vector. * (which is true) */ header->rm_xid = rpc_msg->rm_xid; } header->rm_type = hton32 (RDMA_MSG); header->rm_vers = hton32 (RDMA_VERSION); header->rm_credit = hton32 (credits); header->rm_body.rm_chunks[0] = 0; /* no read chunks */ header->rm_body.rm_chunks[1] = 0; /* no write chunks */ header->rm_body.rm_chunks[2] = 0; /* no reply chunks */ return; } int32_t __rdma_send_reply_inline (rdma_peer_t *peer, rdma_ioq_t *entry, rdma_post_t *post, 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; 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 */ if (send_size > GLUSTERFS_RDMA_INLINE_THRESHOLD) { ret = __rdma_send_error (peer, entry, post, reply_info, ERR_CHUNK); goto out; } header = (rdma_header_t *)post->buf; __rdma_fill_reply_header (header, entry->rpchdr, reply_info, peer->send_count); buf = (char *)&header->rm_body.rm_chunks[3]; if (entry->rpchdr_count != 0) { iov_unload (buf, entry->rpchdr, entry->rpchdr_count); buf += iov_length (entry->rpchdr, entry->rpchdr_count); } if (entry->proghdr_count != 0) { iov_unload (buf, entry->proghdr, entry->proghdr_count); buf += iov_length (entry->proghdr, entry->proghdr_count); } if (entry->prog_payload_count != 0) { iov_unload (buf, entry->prog_payload, entry->prog_payload_count); buf += iov_length (entry->prog_payload, entry->prog_payload_count); } rdma_post_ref (post); ret = rdma_post_send (peer->qp, post, (buf - post->buf)); if (!ret) { ret = send_size; } else { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); ret = -1; } out: return ret; } 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) { uint32_t chunk_size = 0; int32_t ret = -1; rdma_write_array_t *target_array = NULL; int i = 0; target_array = (rdma_write_array_t *)*ptr; for (i = 0; i < reply_info->wc_array->wc_nchunks; i++) { chunk_size += reply_info->wc_array->wc_array[i].wc_target.rs_length; } if (chunk_size < payload_size) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "length of payload (%d) is exceeding the total " "write chunk length (%d)", payload_size, chunk_size); goto out; } target_array->wc_discrim = hton32 (1); for (i = 0; (i < reply_info->wc_array->wc_nchunks) && (payload_size != 0); i++) { target_array->wc_array[i].wc_target.rs_offset = hton64 (reply_info->wc_array->wc_array[i].wc_target.rs_offset); target_array->wc_array[i].wc_target.rs_length = hton32 (min (payload_size, reply_info->wc_array->wc_array[i].wc_target.rs_length)); } target_array->wc_nchunks = hton32 (i); target_array->wc_array[i].wc_target.rs_handle = 0; /* terminate chunklist */ ret = 0; *ptr = &target_array->wc_array[i].wc_target.rs_length; out: return ret; } inline int32_t __rdma_register_local_mr_for_rdma (rdma_peer_t *peer, struct iovec *vector, int count, rdma_post_context_t *ctx) { int i = 0; int32_t ret = -1; rdma_private_t *priv = NULL; rdma_device_t *device = NULL; if ((ctx == NULL) || (vector == NULL)) { goto out; } priv = peer->trans->private; device = priv->device; for (i = 0; i < count; i++) { /* what if the memory is registered more than once? * Assume that a single write buffer is passed to afr, which * then passes it to its children. If more than one children * happen to use rdma, then the buffer is registered more than * once. * Ib-verbs specification says that multiple registrations of * same memory location is allowed. Refer to 10.6.3.8 of * Infiniband Architecture Specification Volume 1 * (Release 1.2.1) */ ctx->mr[ctx->mr_count] = ibv_reg_mr (device->pd, vector[i].iov_base, vector[i].iov_len, IBV_ACCESS_LOCAL_WRITE); if (ctx->mr[ctx->mr_count] == NULL) { goto out; } ctx->mr_count++; } ret = 0; out: return ret; } /* 1. assumes xfer_len of data is pointed by vector(s) starting from vec[*idx] * 2. modifies vec */ int32_t __rdma_write (rdma_peer_t *peer, rdma_post_t *post, struct iovec *vec, uint32_t xfer_len, int *idx, rdma_write_chunk_t *writech) { int size = 0, num_sge = 0, i = 0; int32_t ret = -1; struct ibv_sge *sg_list = NULL; struct ibv_send_wr wr = { .opcode = IBV_WR_RDMA_WRITE, .send_flags = IBV_SEND_SIGNALED, }, *bad_wr; if ((peer == NULL) || (writech == NULL) || (idx == NULL) || (post == NULL) || (vec == NULL) || (xfer_len == 0)) { goto out; } for (i = *idx; size < xfer_len; i++) { size += vec[i].iov_len; } num_sge = i - *idx; sg_list = GF_CALLOC (num_sge, sizeof (struct ibv_sge), gf_common_mt_sge); if (sg_list == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); ret = -1; goto out; } for ((i = *idx), (num_sge = 0); (xfer_len != 0); i++, num_sge++) { size = min (xfer_len, vec[i].iov_len); sg_list [num_sge].addr = (unsigned long)vec[i].iov_base; sg_list [num_sge].length = size; sg_list [num_sge].lkey = post->ctx.mr[i]->lkey; xfer_len -= size; } *idx = i; if (size < vec[i - 1].iov_len) { vec[i - 1].iov_base += size; vec[i - 1].iov_len -= size; *idx = i - 1; } wr.sg_list = sg_list; wr.num_sge = num_sge; wr.wr_id = (unsigned long) 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_DEBUG, "rdma write to " "client (%s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); ret = -1; } GF_FREE (sg_list); out: return ret; } int32_t __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) { int i = 0, payload_idx = 0; uint32_t payload_size = 0, xfer_len = 0; int32_t ret = -1; if (count != 0) { payload_size = iov_length (vector, count); } if (payload_size == 0) { ret = 0; goto out; } ret = __rdma_register_local_mr_for_rdma (peer, vector, count, &post->ctx); if (ret == -1) { goto out; } post->ctx.iobref = iobref_ref (iobref); for (i = 0; (i < reply_info->wc_array->wc_nchunks) && (payload_size != 0); i++) { xfer_len = min (payload_size, reply_info->wc_array->wc_array[i].wc_target.rs_length); ret = __rdma_write (peer, post, vector, xfer_len, &payload_idx, &reply_info->wc_array->wc_array[i]); if (ret == -1) { goto out; } payload_size -= xfer_len; } ret = 0; out: return ret; } 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) { 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]; priv = peer->trans->private; device = priv->device; header = (rdma_header_t *)post->buf; __rdma_fill_reply_header (header, entry->rpchdr, reply_info, peer->send_count); header->rm_type = hton32 (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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "encoding write chunks failed"); ret = __rdma_send_error (peer, entry, post, reply_info, ERR_CHUNK); goto out; } rdma_post_ref (post); for (i = 0; i < entry->rpchdr_count; i++) { vector[count++] = entry->rpchdr[i]; } for (i = 0; i < entry->proghdr_count; i++) { vector[count++] = entry->proghdr[i]; } ret = __rdma_do_rdma_write (peer, post, vector, count, entry->iobref, reply_info); if (ret == -1) { rdma_post_unref (post); goto out; } ret = rdma_post_send (peer->qp, post, (buf - post->buf)); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "rdma_post_send to client (%s) failed with " "ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); ret = -1; rdma_post_unref (post); } else { ret = payload_size; } out: return ret; } 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) { 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; 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_DEBUG, "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); goto out; } header = (rdma_header_t *)post->buf; __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); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "encoding write chunks failed"); ret = __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); ret = __rdma_do_rdma_write (peer, post, entry->prog_payload, entry->prog_payload_count, entry->iobref, reply_info); if (ret == -1) { rdma_post_unref (post); goto out; } iov_unload (ptr, entry->rpchdr, entry->rpchdr_count); ptr += iov_length (entry->rpchdr, entry->rpchdr_count); iov_unload (ptr, entry->proghdr, entry->proghdr_count); ptr += iov_length (entry->proghdr, entry->proghdr_count); ret = rdma_post_send (peer->qp, post, (ptr - post->buf)); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "rdma send to client (%s) failed with ret = %d (%s)", peer->trans->peerinfo.identifier, ret, (ret > 0) ? strerror (ret) : ""); rdma_post_unref (post); ret = -1; } else { ret = send_size + payload_size; } out: return ret; } void rdma_reply_info_destroy (rdma_reply_info_t *reply_info) { if (reply_info == NULL) { goto out; } if (reply_info->wc_array != NULL) { GF_FREE (reply_info->wc_array); reply_info->wc_array = NULL; } mem_put (reply_info->pool, reply_info); out: return; } rdma_reply_info_t * rdma_reply_info_alloc (rdma_peer_t *peer) { rdma_reply_info_t *reply_info = NULL; rdma_private_t *priv = NULL; priv = peer->trans->private; reply_info = mem_get (priv->device->reply_info_pool); if (reply_info == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } memset (reply_info, 0, sizeof (*reply_info)); reply_info->pool = priv->device->reply_info_pool; out: return reply_info; } int32_t __rdma_ioq_churn_reply (rdma_peer_t *peer, rdma_ioq_t *entry, rdma_post_t *post) { rdma_reply_info_t *reply_info = NULL; int32_t ret = -1; rdma_chunktype_t type = rdma_noch; if ((peer == NULL) || (entry == NULL) || (post == NULL)) { goto out; } reply_info = entry->msg.reply_info; if (reply_info != NULL) { type = reply_info->type; } switch (type) { case rdma_noch: ret = __rdma_send_reply_inline (peer, entry, post, reply_info); break; case rdma_replych: ret = __rdma_send_reply_type_nomsg (peer, entry, post, reply_info); break; case rdma_writech: ret = __rdma_send_reply_type_msg (peer, entry, post, reply_info); break; default: gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "invalid chunktype (%d) specified for sending reply", type); break; } if (reply_info != NULL) { rdma_reply_info_destroy (reply_info); } out: return ret; } int32_t __rdma_ioq_churn_entry (rdma_peer_t *peer, 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; priv = peer->trans->private; options = &priv->options; device = priv->device; quota = __rdma_quota_get (peer); if (quota > 0) { post = rdma_get_post (&device->sendq); if (post == NULL) { post = rdma_new_post (device, (options->send_size + 2048), RDMA_SEND_POST); } if (post == NULL) { ret = -1; goto out; } if (entry->is_request) { ret = __rdma_ioq_churn_request (peer, entry, post); } else { ret = __rdma_ioq_churn_reply (peer, entry, post); } if (ret != 0) { __rdma_ioq_entry_free (entry); } } else { ret = 0; } out: return ret; } static int32_t __rdma_ioq_churn (rdma_peer_t *peer) { 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); if (ret <= 0) break; } /* list_for_each_entry_safe (entry, dummy, &peer->ioq, list) { ret = __rdma_ioq_churn_entry (peer, entry); if (ret <= 0) { break; } } */ return ret; } static int32_t rdma_writev (rpc_transport_t *this, rdma_ioq_t *entry) { int32_t ret = 0, need_append = 1; rdma_private_t *priv = this->private; rdma_peer_t *peer = NULL; 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"); ret = -1; goto unlock; } peer = &priv->peer; if (list_empty (&peer->ioq)) { ret = __rdma_ioq_churn_entry (peer, entry); if (ret != 0) { need_append = 0; } } if (need_append) { list_add_tail (&entry->list, &peer->ioq); } } unlock: pthread_mutex_unlock (&priv->write_mutex); return ret; } rdma_ioq_t * 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; if ((data == NULL) || (this == NULL)) { goto out; } priv = this->private; /* TODO: use mem-pool */ entry = mem_get (priv->device->ioq_pool); if (entry == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } memset (entry, 0, sizeof (*entry)); entry->pool = priv->device->ioq_pool; if (data->is_request) { msg = &data->data.req.msg; if (data->data.req.rsp.rsphdr_count != 0) { for (i = 0; i < data->data.req.rsp.rsphdr_count; i++) { entry->msg.request.rsphdr_vec[i] = data->data.req.rsp.rsphdr[i]; } entry->msg.request.rsphdr_count = data->data.req.rsp.rsphdr_count; } if (data->data.req.rsp.rsp_payload_count != 0) { for (i = 0; i < data->data.req.rsp.rsp_payload_count; i++) { entry->msg.request.rsp_payload[i] = data->data.req.rsp.rsp_payload[i]; } entry->msg.request.rsp_payload_count = data->data.req.rsp.rsp_payload_count; } entry->msg.request.rpc_req = data->data.req.rpc_req; if (data->data.req.rsp.rsp_iobref != NULL) { entry->msg.request.rsp_iobref = iobref_ref (data->data.req.rsp.rsp_iobref); } } else { msg = &data->data.reply.msg; entry->msg.reply_info = data->data.reply.private; } entry->is_request = data->is_request; count = msg->rpchdrcount + msg->proghdrcount + msg->progpayloadcount; GF_ASSERT (count <= MAX_IOVEC); if (msg->rpchdr != NULL) { memcpy (&entry->rpchdr[0], msg->rpchdr, sizeof (struct iovec) * msg->rpchdrcount); entry->rpchdr_count = msg->rpchdrcount; } if (msg->proghdr != NULL) { memcpy (&entry->proghdr[0], msg->proghdr, sizeof (struct iovec) * msg->proghdrcount); entry->proghdr_count = msg->proghdrcount; } if (msg->progpayload != NULL) { memcpy (&entry->prog_payload[0], msg->progpayload, sizeof (struct iovec) * msg->progpayloadcount); entry->prog_payload_count = msg->progpayloadcount; } if (msg->iobref != NULL) { entry->iobref = iobref_ref (msg->iobref); } INIT_LIST_HEAD (&entry->list); out: return entry; } int32_t 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, }; if (req == NULL) { goto out; } data.is_request = 1; data.data.req = *req; entry = rdma_ioq_new (this, &data); if (entry == NULL) { goto out; } ret = rdma_writev (this, entry); if (ret > 0) { ret = 0; } else if (ret < 0) { rpc_transport_disconnect (this); } out: return ret; } int32_t 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, }; if (reply == NULL) { goto out; } data.data.reply = *reply; entry = rdma_ioq_new (this, &data); if (entry == NULL) { goto out; } ret = rdma_writev (this, entry); if (ret > 0) { ret = 0; } else if (ret < 0) { rpc_transport_disconnect (this); } 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) { struct _qpent *ent = NULL; rdma_qpreg_t *qpreg = NULL; int32_t hash = 0; int ret = -1; qpreg = &device->qpreg; hash = qp_num % 42; pthread_mutex_lock (&qpreg->lock); { ent = qpreg->ents[hash].next; while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) { ent = ent->next; } if (ent->qp_num == qp_num) { ret = 0; goto unlock; } ent = (struct _qpent *) GF_CALLOC (1, sizeof (*ent), gf_common_mt_qpent); if (ent == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto unlock; } /* TODO: ref reg->peer */ ent->peer = peer; ent->next = &qpreg->ents[hash]; ent->prev = ent->next->prev; ent->next->prev = ent; ent->prev->next = ent; ent->qp_num = qp_num; qpreg->count++; ret = 0; } unlock: pthread_mutex_unlock (&qpreg->lock); return ret; } static void rdma_unregister_peer (rdma_device_t *device, int32_t qp_num) { struct _qpent *ent; rdma_qpreg_t *qpreg = &device->qpreg; int32_t hash = qp_num % 42; pthread_mutex_lock (&qpreg->lock); ent = qpreg->ents[hash].next; while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) ent = ent->next; if (ent->qp_num != qp_num) { pthread_mutex_unlock (&qpreg->lock); return; } ent->prev->next = ent->next; ent->next->prev = ent->prev; /* TODO: unref reg->peer */ GF_FREE (ent); qpreg->count--; pthread_mutex_unlock (&qpreg->lock); } static rdma_peer_t * __rdma_lookup_peer (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; qpreg = &device->qpreg; hash = qp_num % 42; ent = qpreg->ents[hash].next; while ((ent != &qpreg->ents[hash]) && (ent->qp_num != qp_num)) ent = ent->next; if (ent != &qpreg->ents[hash]) { peer = ent->peer; } return peer; } /* static 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) { rdma_private_t *priv = this->private; if (priv->peer.qp) { rdma_unregister_peer (priv->device, priv->peer.qp->qp_num); ibv_destroy_qp (priv->peer.qp); } priv->peer.qp = NULL; return; } static int32_t 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; peer = &priv->peer; struct ibv_qp_init_attr init_attr = { .send_cq = device->send_cq, .recv_cq = device->recv_cq, .srq = device->srq, .cap = { .max_send_wr = peer->send_count, .max_recv_wr = peer->recv_count, .max_send_sge = 2, .max_recv_sge = 1 }, .qp_type = IBV_QPT_RC }; 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 = -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; ret = rdma_register_peer (device, peer->qp->qp_num, peer); out: if (ret == -1) __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) { int32_t i; int32_t ret = 0; rdma_private_t *priv = this->private; rdma_device_t *device = priv->device; for (i=0 ; iname); ret = -1; break; } rdma_put_post (q, post); } return ret; } 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); if (!ret) { for (i=0 ; irecv_count ; i++) { post = rdma_get_post (&device->recvq); if (rdma_post_recv (device->srq, post) != 0) { ret = -1; break; } } } 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; } /* 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; } return 0; } static int32_t __rdma_teardown (rpc_transport_t *this) { rdma_private_t *priv = this->private; __rdma_destroy_qp (this); if (!list_empty (&priv->peer.ioq)) { __rdma_ioq_flush (&priv->peer); } /* 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) { 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); } return ret; } /* * allocates new memory to hold write-chunklist. New memory is needed since * write-chunklist will be used while sending reply and the post holding initial * write-chunklist sent from client will be put back to srq before a pollin * event is sent to upper layers. */ int32_t rdma_get_write_chunklist (char **ptr, rdma_write_array_t **write_ary) { rdma_write_array_t *from = NULL, *to = NULL; int32_t ret = -1, size = 0, i = 0; from = (rdma_write_array_t *) *ptr; if (from->wc_discrim == 0) { ret = 0; goto out; } from->wc_nchunks = ntoh32 (from->wc_nchunks); size = sizeof (*from) + (sizeof (rdma_write_chunk_t) * from->wc_nchunks); to = GF_CALLOC (1, size, gf_common_mt_char); if (to == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); ret = -1; goto out; } to->wc_discrim = ntoh32 (from->wc_discrim); to->wc_nchunks = from->wc_nchunks; for (i = 0; i < to->wc_nchunks; i++) { to->wc_array[i].wc_target.rs_handle = ntoh32 (from->wc_array[i].wc_target.rs_handle); to->wc_array[i].wc_target.rs_length = ntoh32 (from->wc_array[i].wc_target.rs_length); to->wc_array[i].wc_target.rs_offset = ntoh64 (from->wc_array[i].wc_target.rs_offset); } *write_ary = to; ret = 0; *ptr = (char *)&from->wc_array[i].wc_target.rs_handle; out: return ret; } /* * does not allocate new memory to hold read-chunklist. New memory is not * needed, since post is not put back to srq till we've completed all the * rdma-reads and hence readchunk-list can point to memory held by post. */ int32_t rdma_get_read_chunklist (char **ptr, rdma_read_chunk_t **readch) { int32_t ret = -1; rdma_read_chunk_t *chunk = NULL; int i = 0; chunk = (rdma_read_chunk_t *)*ptr; if (chunk[0].rc_discrim == 0) { ret = 0; goto out; } for (i = 0; chunk[i].rc_discrim != 0; i++) { chunk[i].rc_discrim = ntoh32 (chunk[i].rc_discrim); chunk[i].rc_position = ntoh32 (chunk[i].rc_position); chunk[i].rc_target.rs_handle = ntoh32 (chunk[i].rc_target.rs_handle); chunk[i].rc_target.rs_length = ntoh32 (chunk[i].rc_target.rs_length); chunk[i].rc_target.rs_offset = ntoh64 (chunk[i].rc_target.rs_offset); } *readch = &chunk[0]; ret = 0; *ptr = (char *)&chunk[i].rc_discrim; out: return ret; } inline int32_t rdma_decode_error_msg (rdma_peer_t *peer, 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; header = (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); } iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); if (iobuf == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); ret = -1; goto out; } post->ctx.iobref = iobref = iobref_new (); if (iobref == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); ret = -1; goto out; } 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); post->ctx.count = 1; iobuf = NULL; iobref = NULL; out: if (ret == -1) { if (iobuf != NULL) { iobuf_unref (iobuf); } if (iobref != NULL) { iobref_unref (iobref); } } return 0; } int32_t rdma_decode_msg (rdma_peer_t *peer, rdma_post_t *post, 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; header = (rdma_header_t *)post->buf; ptr = (char *)&header->rm_body.rm_chunks[0]; ret = rdma_get_read_chunklist (&ptr, readch); if (ret == -1) { goto out; } /* skip terminator of read-chunklist */ ptr = ptr + sizeof (uint32_t); ret = rdma_get_write_chunklist (&ptr, &write_ary); if (ret == -1) { goto out; } /* skip terminator of write-chunklist */ ptr = ptr + sizeof (uint32_t); if (write_ary != NULL) { reply_info = rdma_reply_info_alloc (peer); if (reply_info == NULL) { ret = -1; goto out; } reply_info->type = rdma_writech; reply_info->wc_array = write_ary; reply_info->rm_xid = header->rm_xid; } else { ret = rdma_get_write_chunklist (&ptr, &write_ary); if (ret == -1) { goto out; } if (write_ary != NULL) { reply_info = rdma_reply_info_alloc (peer); if (reply_info == NULL) { ret = -1; goto out; } reply_info->type = rdma_replych; reply_info->wc_array = write_ary; reply_info->rm_xid = header->rm_xid; } } /* skip terminator of reply chunk */ ptr = ptr + sizeof (uint32_t); if (header->rm_type != RDMA_NOMSG) { post->ctx.hdr_iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); if (post->ctx.hdr_iobuf == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); 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; } post->ctx.reply_info = reply_info; out: if (ret == -1) { if (*readch != NULL) { GF_FREE (*readch); *readch = NULL; } if (write_ary != NULL) { GF_FREE (write_ary); } } return ret; } /* 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) { int32_t ret = -1; rdma_header_t *header = NULL; header = (rdma_header_t *)post->buf; header->rm_xid = ntoh32 (header->rm_xid); header->rm_vers = ntoh32 (header->rm_vers); header->rm_credit = ntoh32 (header->rm_credit); header->rm_type = ntoh32 (header->rm_type); switch (header->rm_type) { case RDMA_MSG: case RDMA_NOMSG: ret = rdma_decode_msg (peer, post, readch, bytes_in_post); break; case RDMA_MSGP: gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "rdma msg of msg-type RDMA_MSGP should not have been " "recieved"); 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 " "recieved"); ret = -1; break; case RDMA_ERROR: /* ret = rdma_decode_error_msg (peer, post, bytes_in_post); */ break; default: gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "unknown rdma msg-type (%d)", header->rm_type); } return ret; } int32_t __rdma_read (rdma_peer_t *peer, rdma_post_t *post, struct iovec *to, rdma_read_chunk_t *readch) { int32_t ret = -1; struct ibv_sge list = {0, }; struct ibv_send_wr wr = {0, }, *bad_wr = NULL; ret = __rdma_register_local_mr_for_rdma (peer, to, 1, &post->ctx); if (ret == -1) { goto out; } list.addr = (unsigned long) to->iov_base; 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.sg_list = &list; wr.num_sge = 1; wr.opcode = IBV_WR_RDMA_READ; wr.send_flags = IBV_SEND_SIGNALED; wr.wr.rdma.remote_addr = readch->rc_target.rs_offset; wr.wr.rdma.rkey = readch->rc_target.rs_handle; ret = ibv_post_send (peer->qp, &wr, &bad_wr); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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); } out: return ret; } int32_t rdma_do_reads (rdma_peer_t *peer, rdma_post_t *post, 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; priv = peer->trans->private; for (i = 0; readch[i].rc_discrim != 0; i++) { size += readch[i].rc_target.rs_length; } if (i == 0) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "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; } iobuf = iobuf_get (peer->trans->ctx->iobuf_pool); if (iobuf == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } if (post->ctx.iobref == NULL) { post->ctx.iobref = iobref_new (); if (post->ctx.iobref == NULL) { iobuf_unref (iobuf); gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } } iobref_add (post->ctx.iobref, iobuf); iobuf_unref (iobuf); ptr = iobuf_ptr (iobuf); iobuf = NULL; pthread_mutex_lock (&priv->write_mutex); { if (!priv->connected) { goto unlock; } for (i = 0; readch[i].rc_discrim != 0; i++) { count = post->ctx.count++; post->ctx.vector[count].iov_base = ptr; post->ctx.vector[count].iov_len = readch[i].rc_target.rs_length; ret = __rdma_read (peer, post, &post->ctx.vector[count], &readch[i]); if (ret == -1) { goto unlock; } ptr += readch[i].rc_target.rs_length; } ret = 0; } unlock: pthread_mutex_unlock (&priv->write_mutex); out: if (ret == -1) { if (iobuf != NULL) { iobuf_unref (iobuf); } } return ret; } int32_t rdma_pollin_notify (rdma_peer_t *peer, 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; if ((peer == NULL) || (post == NULL)) { goto out; } if (post->ctx.iobref == NULL) { post->ctx.iobref = iobref_new (); if (post->ctx.iobref == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } /* handling the case where both hdr and payload of * GF_FOP_READ_CBK were recieved in a single iobuf * because of server sending entire msg as inline without * doing rdma writes. */ iobref_add (post->ctx.iobref, post->ctx.hdr_iobuf); } pollin = rpc_transport_pollin_alloc (peer->trans, post->ctx.vector, post->ctx.count, post->ctx.hdr_iobuf, post->ctx.iobref, post->ctx.reply_info); if (pollin == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); goto out; } ptr = (uint32_t *)pollin->vector[0].iov_base; request_info.xid = ntoh32 (*ptr); msg_type = ntoh32 (*(ptr + 1)); if (msg_type == REPLY) { ret = rpc_transport_notify (peer->trans, RPC_TRANSPORT_MAP_XID_REQUEST, &request_info); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot get request information from rpc " "layer"); goto out; } rpc_req = request_info.rpc_req; if (rpc_req == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "rpc request structure not found"); ret = -1; goto out; } request_context = rpc_req->conn_private; rpc_req->conn_private = NULL; priv = peer->trans->private; if (request_context != NULL) { pthread_mutex_lock (&priv->write_mutex); { __rdma_request_context_destroy (request_context); } pthread_mutex_unlock (&priv->write_mutex); } else { rdma_quota_put (peer); } pollin->is_reply = 1; } ret = rpc_transport_notify (peer->trans, RPC_TRANSPORT_MSG_RECEIVED, pollin); out: if (pollin != NULL) { pollin->private = NULL; rpc_transport_pollin_destroy (pollin); } return ret; } int32_t rdma_recv_reply (rdma_peer_t *peer, 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; reply_info = post->ctx.reply_info; /* no write chunklist, just notify upper layers */ if (reply_info == NULL) { ret = 0; goto out; } wc_array = reply_info->wc_array; if (header->rm_type == RDMA_NOMSG) { post->ctx.vector[0].iov_base = (void *)(long)wc_array->wc_array[0].wc_target.rs_offset; post->ctx.vector[0].iov_len = wc_array->wc_array[0].wc_target.rs_length; post->ctx.count = 1; } else { for (i = 0; i < wc_array->wc_nchunks; i++) { post->ctx.vector[i + 1].iov_base = (void *)(long)wc_array->wc_array[i].wc_target.rs_offset; post->ctx.vector[i + 1].iov_len = wc_array->wc_array[i].wc_target.rs_length; } post->ctx.count += wc_array->wc_nchunks; } ptr = (uint32_t *)post->ctx.vector[0].iov_base; request_info.xid = ntoh32 (*ptr); ret = rpc_transport_notify (peer->trans, RPC_TRANSPORT_MAP_XID_REQUEST, &request_info); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "cannot get request information from rpc " "layer"); goto out; } rpc_req = request_info.rpc_req; if (rpc_req == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "rpc request structure not found"); ret = -1; goto out; } ctx = rpc_req->conn_private; if (post->ctx.iobref == NULL) { post->ctx.iobref = iobref_ref (ctx->rsp_iobref); } ret = 0; rdma_reply_info_destroy (reply_info); out: if (ret == 0) { ret = rdma_pollin_notify (peer, post); } return ret; } inline int32_t rdma_recv_request (rdma_peer_t *peer, rdma_post_t *post, rdma_read_chunk_t *readch) { int32_t ret = -1; if (readch != NULL) { ret = rdma_do_reads (peer, post, readch); } else { ret = rdma_pollin_notify (peer, post); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "pollin notification failed"); } } return ret; } void rdma_process_recv (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; if (post == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "no post found in successful work completion element"); goto out; } ret = rdma_decode_header (peer, post, &readch, wc->byte_len); if (ret == -1) { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "decoding of header failed"); goto out; } header = (rdma_header_t *)post->buf; switch (header->rm_type) { case RDMA_MSG: ptr = (uint32_t *)post->ctx.vector[0].iov_base; msg_type = ntoh32 (*(ptr + 1)); break; case RDMA_NOMSG: if (readch != NULL) { msg_type = CALL; } else { msg_type = REPLY; } 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; */ default: gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "invalid rdma msg-type (%d)", header->rm_type); break; } if (msg_type == CALL) { ret = rdma_recv_request (peer, post, readch); } else { ret = rdma_recv_reply (peer, post); } out: if (ret == -1) { rpc_transport_disconnect (peer->trans); } return; } static void * 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; struct ibv_cq *event_cq = NULL; struct ibv_wc wc = {0, }; void *event_ctx = NULL; int32_t ret = 0; chan = data; while (1) { ret = ibv_get_cq_event (chan, &event_cq, &event_ctx); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_get_cq_event failed, terminating recv " "thread %d (%d)", ret, errno); continue; } device = event_ctx; ret = ibv_req_notify_cq (event_cq, 0); if (ret) { gf_log (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; while ((ret = ibv_poll_cq (event_cq, 1, &wc)) > 0) { post = (rdma_post_t *) (long) wc.wr_id; pthread_mutex_lock (&device->qpreg.lock); { peer = __rdma_lookup_peer (device, wc.qp_num); /* * keep a refcount on transport so that it * doesnot get freed because of some error * indicated by wc.status till we are done * with usage of peer and thereby that of trans. */ if (peer != NULL) { rpc_transport_ref (peer->trans); } } pthread_mutex_unlock (&device->qpreg.lock); if (wc.status != IBV_WC_SUCCESS) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "recv work request on `%s' returned " "error (%d)", device->device_name, wc.status); if (peer) { rpc_transport_unref (peer->trans); rpc_transport_disconnect (peer->trans); } if (post) { rdma_post_unref (post); } continue; } if (peer) { rdma_process_recv (peer, &wc); rpc_transport_unref (peer->trans); } else { gf_log (RDMA_LOG_NAME, GF_LOG_DEBUG, "could not lookup peer for qp_num: %d", wc.qp_num); } rdma_post_unref (post); } if (ret < 0) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_poll_cq on `%s' returned error " "(ret = %d, errno = %d)", device->device_name, ret, errno); continue; } ibv_ack_cq_events (event_cq, 1); } return NULL; } void rdma_handle_failed_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) { rdma_post_t *post = NULL; rdma_device_t *device = NULL; rdma_private_t *priv = NULL; if (peer != NULL) { priv = peer->trans->private; if (priv != NULL) { device = priv->device; } } post = (rdma_post_t *) (long) wc->wr_id; gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "send work request on `%s' returned error " "wc.status = %d, wc.vendor_err = %d, post->buf = %p, " "wc.byte_len = %d, post->reused = %d", (device != NULL) ? device->device_name : NULL, wc->status, wc->vendor_err, post->buf, wc->byte_len, post->reused); if (wc->status == IBV_WC_RETRY_EXC_ERR) { gf_log ("rdma", GF_LOG_ERROR, "connection between client and" " server not working. check by running " "'ibv_srq_pingpong'. also make sure subnet manager" " is running (eg: 'opensm'), or check if rdma port is " "valid (or active) by running 'ibv_devinfo'. contact " "Gluster Support Team if the problem persists."); } if (peer) { rpc_transport_disconnect (peer->trans); } return; } void rdma_handle_successful_send_completion (rdma_peer_t *peer, struct ibv_wc *wc) { rdma_post_t *post = NULL; int reads = 0, ret = 0; if (wc->opcode != IBV_WC_RDMA_READ) { goto out; } post = (rdma_post_t *)(long) wc->wr_id; pthread_mutex_lock (&post->lock); { reads = --post->ctx.rdma_reads; } pthread_mutex_unlock (&post->lock); if (reads != 0) { /* if it is not the last rdma read, we've got nothing to do */ goto out; } ret = rdma_pollin_notify (peer, post); if ((ret == -1) && (peer != NULL)) { rpc_transport_disconnect (peer->trans); } out: return; } static void * rdma_send_completion_proc (void *data) { struct ibv_comp_channel *chan = NULL; rdma_post_t *post = NULL; rdma_peer_t *peer = NULL; struct ibv_cq *event_cq = NULL; void *event_ctx = NULL; rdma_device_t *device = NULL; struct ibv_wc wc = {0, }; char is_request = 0; int32_t ret = 0, quota_ret = 0; chan = data; while (1) { ret = ibv_get_cq_event (chan, &event_cq, &event_ctx); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_get_cq_event on failed, terminating " "send thread: %d (%d)", ret, errno); continue; } device = event_ctx; ret = ibv_req_notify_cq (event_cq, 0); if (ret) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "ibv_req_notify_cq on %s failed, terminating " "send thread: %d (%d)", device->device_name, ret, errno); continue; } while ((ret = ibv_poll_cq (event_cq, 1, &wc)) > 0) { post = (rdma_post_t *) (long) wc.wr_id; pthread_mutex_lock (&device->qpreg.lock); { peer = __rdma_lookup_peer (device, wc.qp_num); /* * keep a refcount on transport so that it * does not get freed because of some error * indicated by wc.status, till we are done * with usage of peer and thereby that of trans. */ if (peer != NULL) { rpc_transport_ref (peer->trans); } } pthread_mutex_unlock (&device->qpreg.lock); if (wc.status != IBV_WC_SUCCESS) { rdma_handle_failed_send_completion (peer, &wc); } else { rdma_handle_successful_send_completion (peer, &wc); } if (post) { is_request = post->ctx.is_request; ret = rdma_post_unref (post); if ((ret == 0) && (wc.status == IBV_WC_SUCCESS) && !is_request && (post->type == RDMA_SEND_POST) && (peer != NULL)) { /* An RDMA_RECV_POST can end up in * rdma_send_completion_proc for * rdma-reads, and we do not take * quota for getting an RDMA_RECV_POST. */ /* * if it is request, quota is returned * after reply has come. */ quota_ret = rdma_quota_put (peer); if (quota_ret < 0) { gf_log ("rdma", GF_LOG_DEBUG, "failed to send " "message"); } } } if (peer) { rpc_transport_unref (peer->trans); } else { gf_log (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, "ibv_poll_cq on `%s' returned error (ret = %d," " errno = %d)", device->device_name, ret, errno); continue; } ibv_ack_cq_events (event_cq, 1); } return NULL; } static void 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; /* TODO: validate arguments from options below */ 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; temp = dict_get (this->options, "transport.rdma.work-request-send-count"); if (temp) options->send_count = data_to_int32 (temp); temp = dict_get (this->options, "transport.rdma.work-request-recv-count"); if (temp) options->recv_count = data_to_int32 (temp); options->port = 1; temp = dict_get (this->options, "transport.rdma.port"); if (temp) options->port = data_to_uint64 (temp); options->mtu = mtu = IBV_MTU_2048; temp = dict_get (this->options, "transport.rdma.mtu"); if (temp) mtu = data_to_int32 (temp); switch (mtu) { case 256: options->mtu = IBV_MTU_256; break; case 512: options->mtu = IBV_MTU_512; break; case 1024: options->mtu = IBV_MTU_1024; break; case 2048: options->mtu = IBV_MTU_2048; break; case 4096: options->mtu = IBV_MTU_4096; break; default: if (temp) gf_log (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, "%s: defaulting MTU to '2048'", this->name); options->mtu = IBV_MTU_2048; break; } temp = dict_get (this->options, "transport.rdma.device-name"); if (temp) options->device_name = gf_strdup (temp->data); return; } 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) { 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; 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; } if (!trav) { trav = GF_CALLOC (1, sizeof (*trav), gf_common_mt_rdma_device_t); if (trav == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); 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) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); 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); gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); return NULL; } trav->reply_info_pool = mem_pool_new (rdma_reply_info_t, RDMA_POOL_SIZE); if (trav->reply_info_pool == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); 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; } /* 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 = 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; } /* 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]; } } return trav; } static int32_t 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); ibv_fork_init (); { 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; } if (!*dev_list) { gf_log (RDMA_LOG_NAME, GF_LOG_CRITICAL, "No IB devices found"); ret = -1; goto cleanup; } if (!options->device_name) { if (*dev_list) { options->device_name = gf_strdup (ibv_get_device_name (*dev_list)); } else { gf_log (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->device = rdma_get_device (this, ib_ctx); 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; } } 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); { 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; } } } unlock: pthread_mutex_unlock (&priv->write_mutex); if (ret == -1) { rpc_transport_disconnect (this); } else { ret = 0; } if (!ret && priv->connected) { if (priv->is_server) { ret = rpc_transport_notify (priv->listener, RPC_TRANSPORT_ACCEPT, this); } else { ret = rpc_transport_notify (this, RPC_TRANSPORT_CONNECT, this); } } return ret; } static int 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) { 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); 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; 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; } } 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) { dict_t *options = this->options; rdma_private_t *priv = this->private; int32_t ret = 0; gf_boolean_t non_blocking = 1; struct sockaddr_storage sockaddr; socklen_t sockaddr_len = 0; if (dict_get (options, "non-blocking-io")) { char *nb_connect = data_to_str (dict_get (this->options, "non-blocking-io")); 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; } } ret = gf_rdma_client_get_remote_sockaddr (this, (struct sockaddr *)&sockaddr, &sockaddr_len, port); if (ret != 0) { gf_log (this->name, GF_LOG_DEBUG, "cannot get remote address to connect"); return ret; } pthread_mutex_lock (&priv->write_mutex); { if (priv->sock != -1) { ret = 0; goto unlock; } priv->sock = socket (((struct sockaddr *)&sockaddr)->sa_family, SOCK_STREAM, 0); if (priv->sock == -1) { gf_log (this->name, GF_LOG_ERROR, "socket () - error: %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); this->peerinfo.sockaddr_len = sockaddr_len; if (port > 0) ((struct sockaddr_in *) (&sockaddr))->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) { 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)", 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); } 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) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); return -1; } this->listener = trans; priv = GF_CALLOC (1, sizeof (rdma_private_t), gf_common_mt_rdma_private_t); if (priv == NULL) { gf_log (RDMA_LOG_NAME, GF_LOG_ERROR, "out of memory"); GF_FREE (priv); return -1; } this->private = priv; /* Copy all the rdma related values in priv, from trans_priv as other than QP, all the values remain same */ priv->device = trans_priv->device; priv->options = trans_priv->options; priv->is_server = 1; priv->listener = trans; options = &priv->options; this->ops = trans->ops; this->init = trans->init; this->fini = trans->fini; this->ctx = trans->ctx; this->name = gf_strdup (trans->name); this->notify = trans->notify; this->mydata = trans->mydata; memcpy (&this->myinfo.sockaddr, &trans->myinfo.sockaddr, trans->myinfo.sockaddr_len); this->myinfo.sockaddr_len = trans->myinfo.sockaddr_len; main_sock = (trans_priv)->sock; this->peerinfo.sockaddr_len = sizeof (this->peerinfo.sockaddr); priv->sock = accept (main_sock, (struct sockaddr *)&this->peerinfo.sockaddr, &this->peerinfo.sockaddr_len); if (priv->sock == -1) { gf_log ("rdma/server", GF_LOG_ERROR, "accept() failed: %s", strerror (errno)); GF_FREE (this->private); GF_FREE (this); return -1; } priv->peer.trans = this; rpc_transport_ref (this); gf_rdma_get_transport_identifiers (this); priv->tcp_connected = 1; priv->handshake.incoming.state = 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); 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->idx = event_register (this->ctx->event_pool, priv->sock, rdma_event_handler, this, 1, 1); pthread_mutex_init (&priv->read_mutex, NULL); pthread_mutex_init (&priv->write_mutex, NULL); pthread_mutex_init (&priv->recv_mutex, NULL); /* pthread_cond_init (&priv->recv_cond, NULL); */ return 0; } 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, &sockaddr_len); if (ret != 0) { gf_log (this->name, GF_LOG_DEBUG, "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", strerror (errno)); GF_FREE (this->private); ret = -1; goto err; } memcpy (&this->myinfo.sockaddr, &sockaddr, sockaddr_len); this->myinfo.sockaddr_len = sockaddr_len; ret = getnameinfo ((struct sockaddr *)&this->myinfo.sockaddr, this->myinfo.sockaddr_len, host, sizeof (host), service, sizeof (service), NI_NUMERICHOST); if (ret != 0) { gf_log (this->name, GF_LOG_ERROR, "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)); 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; 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); err: 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, }; int32_t init (rpc_transport_t *this) { 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)) { gf_log (this->name, GF_LOG_ERROR, "Failed to initialize IB Device"); return -1; } return 0; } void fini (struct rpc_transport *this) { /* TODO: verify this function does graceful finish */ rdma_private_t *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); GF_FREE (priv); } return; } /* TODO: expand each option */ struct volume_options options[] = { { .key = {"transport.rdma.port", "rdma-port"}, .type = GF_OPTION_TYPE_INT, .min = 1, .max = 4, .description = "check the option by 'ibv_devinfo'" }, { .key = {"transport.rdma.mtu", "rdma-mtu"}, .type = GF_OPTION_TYPE_INT, }, { .key = {"transport.rdma.device-name", "rdma-device-name"}, .type = GF_OPTION_TYPE_ANY, .description = "check by 'ibv_devinfo'" }, { .key = {"transport.rdma.work-request-send-count", "rdma-work-request-send-count"}, .type = GF_OPTION_TYPE_INT, }, { .key = {"transport.rdma.work-request-recv-count", "rdma-work-request-recv-count"}, .type = GF_OPTION_TYPE_INT, }, { .key = {"remote-port", "transport.remote-port", "transport.rdma.remote-port"}, .type = GF_OPTION_TYPE_INT }, { .key = {"transport.rdma.listen-port", "listen-port"}, .type = GF_OPTION_TYPE_INT }, { .key = {"transport.rdma.connect-path", "connect-path"}, .type = GF_OPTION_TYPE_ANY }, { .key = {"transport.rdma.bind-path", "bind-path"}, .type = GF_OPTION_TYPE_ANY }, { .key = {"transport.rdma.listen-path", "listen-path"}, .type = GF_OPTION_TYPE_ANY }, { .key = {"transport.address-family", "address-family"}, .value = {"inet", "inet6", "inet/inet6", "inet6/inet", "unix", "inet-sdp" }, .type = GF_OPTION_TYPE_STR }, { .key = {"transport.socket.lowlat"}, .type = GF_OPTION_TYPE_BOOL }, { .key = {NULL} } };