]> git.openfabrics.org - ~shefty/libibverbs.git/commitdiff
libibverbs: Add xrc pingpong example
authorJay Sternberg <jay.e.sternberg@intel.com>
Thu, 22 Dec 2011 08:09:12 +0000 (00:09 -0800)
committerSean Hefty <sean.hefty@intel.com>
Thu, 22 Dec 2011 08:09:12 +0000 (00:09 -0800)
Signed-off-by: Jay Sternberg <jay.e.sternberg@intel.com>
Signed-off-by: Sean Hefty <sean.hefty@intel.com>
Makefile.am
examples/xsrq_pingpong.c [new file with mode: 0644]

index 0552e89108009d3daea42e5d20cad2a5f97a11d7..ee86c70c5a343ec75a59d838d2c91962339b1fdc 100644 (file)
@@ -18,7 +18,7 @@ src_libibverbs_la_DEPENDENCIES = $(srcdir)/src/libibverbs.map
 
 bin_PROGRAMS = examples/ibv_devices examples/ibv_devinfo \
     examples/ibv_asyncwatch examples/ibv_rc_pingpong examples/ibv_uc_pingpong \
-    examples/ibv_ud_pingpong examples/ibv_srq_pingpong
+    examples/ibv_ud_pingpong examples/ibv_srq_pingpong examples/ibv_xsrq_pingpong
 examples_ibv_devices_SOURCES = examples/device_list.c
 examples_ibv_devices_LDADD = $(top_builddir)/src/libibverbs.la
 examples_ibv_devinfo_SOURCES = examples/devinfo.c
@@ -31,6 +31,8 @@ examples_ibv_ud_pingpong_SOURCES = examples/ud_pingpong.c examples/pingpong.c
 examples_ibv_ud_pingpong_LDADD = $(top_builddir)/src/libibverbs.la
 examples_ibv_srq_pingpong_SOURCES = examples/srq_pingpong.c examples/pingpong.c
 examples_ibv_srq_pingpong_LDADD = $(top_builddir)/src/libibverbs.la
+examples_ibv_xsrq_pingpong_SOURCES = examples/xsrq_pingpong.c examples/pingpong.c
+examples_ibv_xsrq_pingpong_LDADD = $(top_builddir)/src/libibverbs.la
 examples_ibv_asyncwatch_SOURCES = examples/asyncwatch.c
 examples_ibv_asyncwatch_LDADD = $(top_builddir)/src/libibverbs.la
 
diff --git a/examples/xsrq_pingpong.c b/examples/xsrq_pingpong.c
new file mode 100644 (file)
index 0000000..c18b636
--- /dev/null
@@ -0,0 +1,873 @@
+/*\r
+ * Copyright (c) 2005 Topspin Communications.  All rights reserved.\r
+ * Copyright (c) 2011 Intel Corporation, Inc.  All rights reserved.\r
+ *\r
+ * This software is available to you under a choice of one of two\r
+ * licenses.  You may choose to be licensed under the terms of the GNU\r
+ * General Public License (GPL) Version 2, available from the file\r
+ * COPYING in the main directory of this source tree, or the\r
+ * OpenIB.org BSD license below:\r
+ *\r
+ *     Redistribution and use in source and binary forms, with or\r
+ *     without modification, are permitted provided that the following\r
+ *     conditions are met:\r
+ *\r
+ *      - Redistributions of source code must retain the above\r
+ *        copyright notice, this list of conditions and the following\r
+ *        disclaimer.\r
+ *\r
+ *      - Redistributions in binary form must reproduce the above\r
+ *        copyright notice, this list of conditions and the following\r
+ *        disclaimer in the documentation and/or other materials\r
+ *        provided with the distribution.\r
+ *\r
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,\r
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF\r
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND\r
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS\r
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN\r
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN\r
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\r
+ * SOFTWARE.\r
+ */\r
+\r
+#if HAVE_CONFIG_H\r
+#  include <config.h>\r
+#endif /* HAVE_CONFIG_H */\r
+\r
+#include <stdio.h>\r
+#include <fcntl.h>\r
+#include <errno.h>\r
+#include <stdlib.h>\r
+#include <unistd.h>\r
+#include <string.h>\r
+#include <sys/types.h>\r
+#include <sys/socket.h>\r
+#include <sys/time.h>\r
+#include <netdb.h>\r
+#include <malloc.h>\r
+#include <getopt.h>\r
+#include <arpa/inet.h>\r
+#include <time.h>\r
+\r
+#include "pingpong.h"\r
+\r
+#define MSG_FORMAT "%04x:%06x:%06x:%06x:%04x"\r
+#define MSG_SIZE   30\r
+#define MSG_SSCAN  "%x:%x:%x:%x:%x"\r
+#define ADDR_FORMAT \\r
+       "%8s: LID %04x, QPN RECV %06x SEND %06x, PSN %06x, SRQN %04x\n"\r
+static int page_size;\r
+\r
+struct pingpong_dest {\r
+       int lid;\r
+       int recv_qpn;\r
+       int send_qpn;\r
+       int recv_psn;\r
+       int send_psn;\r
+       int srqn;\r
+       int pp_cnt;\r
+};\r
+\r
+struct pingpong_context {\r
+       struct ibv_context      *context;\r
+       struct ibv_comp_channel *channel;\r
+       struct ibv_pd           *pd;\r
+       struct ibv_mr           *mr;\r
+       struct ibv_cq           *send_cq;\r
+       struct ibv_cq           *recv_cq;\r
+       struct ibv_srq          *srq;\r
+       struct ibv_xrcd         *xrcd;\r
+       struct ibv_qp           **recv_qp;\r
+       struct ibv_qp           **send_qp;\r
+       struct pingpong_dest    *rem_dest;\r
+       void                    *buf;\r
+       int                      lid;\r
+       int                      sl;\r
+       enum ibv_mtu             mtu;\r
+       int                      ib_port;\r
+       int                      fd;\r
+       int                      size;\r
+       int                      num_clients;\r
+       int                      num_tests;\r
+       int                      use_event;\r
+};\r
+\r
+struct pingpong_context ctx;\r
+\r
+\r
+static int open_device(char *ib_devname)\r
+{\r
+       struct ibv_device **dev_list;\r
+       int i = 0;\r
+\r
+       dev_list = ibv_get_device_list(NULL);\r
+       if (!dev_list) {\r
+               fprintf(stderr, "Failed to get IB devices list");\r
+               return -1;\r
+       }\r
+\r
+       if (ib_devname) {\r
+               for (; dev_list[i]; ++i) {\r
+                       if (!strcmp(ibv_get_device_name(dev_list[i]), ib_devname))\r
+                               break;\r
+               }\r
+       }\r
+       if (!dev_list[i]) {\r
+               fprintf(stderr, "IB device %s not found\n",\r
+                       ib_devname ? ib_devname : "");\r
+               return -1;\r
+       }\r
+\r
+       ctx.context = ibv_open_device(dev_list[i]);\r
+       if (!ctx.context) {\r
+               fprintf(stderr, "Couldn't get context for %s\n",\r
+                       ibv_get_device_name(dev_list[i]));\r
+               return -1;\r
+       }\r
+\r
+       ibv_free_device_list(dev_list);\r
+       return 0;\r
+}\r
+\r
+static int create_qps(void)\r
+{\r
+       struct ibv_qp_init_attr init;\r
+       struct ibv_qp_attr mod;\r
+       int i;\r
+\r
+       for (i = 0; i < ctx.num_clients; ++i) {\r
+\r
+               memset(&init, 0, sizeof init);\r
+               init.qp_type = IBV_QPT_XRC_RECV;\r
+               init.ext.xrc_recv.xrcd = ctx.xrcd;\r
+\r
+               ctx.recv_qp[i] = ibv_create_qp(ctx.pd, &init);\r
+               if (!ctx.recv_qp[i])  {\r
+                       fprintf(stderr, "Couldn't create recv QP[%d] errno %d\n",\r
+                               i, errno);\r
+                       return 1;\r
+               }\r
+\r
+               mod.qp_state        = IBV_QPS_INIT;\r
+               mod.pkey_index      = 0;\r
+               mod.port_num        = ctx.ib_port;\r
+               mod.qp_access_flags = IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ;\r
+\r
+               if (ibv_modify_qp(ctx.recv_qp[i], &mod,\r
+                                 IBV_QP_STATE | IBV_QP_PKEY_INDEX |\r
+                                 IBV_QP_PORT | IBV_QP_ACCESS_FLAGS)) {\r
+                       fprintf(stderr, "Failed to modify recv QP[%d] to INIT\n", i);\r
+                       return 1;\r
+               }\r
+\r
+               memset(&init, 0, sizeof init);\r
+               init.qp_type          = IBV_QPT_XRC_SEND;\r
+               init.send_cq          = ctx.send_cq;\r
+               init.cap.max_send_wr  = ctx.num_clients * ctx.num_tests;\r
+               init.cap.max_send_sge = 1;\r
+\r
+               ctx.send_qp[i] = ibv_create_qp(ctx.pd, &init);\r
+               if (!ctx.send_qp[i])  {\r
+                       fprintf(stderr, "Couldn't create send QP[%d] errno %d\n",\r
+                               i, errno);\r
+                       return 1;\r
+               }\r
+\r
+               mod.qp_state        = IBV_QPS_INIT;\r
+               mod.pkey_index      = 0;\r
+               mod.port_num        = ctx.ib_port;\r
+               mod.qp_access_flags = 0;\r
+\r
+               if (ibv_modify_qp(ctx.send_qp[i], &mod,\r
+                                 IBV_QP_STATE | IBV_QP_PKEY_INDEX |\r
+                                 IBV_QP_PORT | IBV_QP_ACCESS_FLAGS)) {\r
+                       fprintf(stderr, "Failed to modify send QP[%d] to INIT\n", i);\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       return 0;\r
+}\r
+\r
+static int pp_init_ctx(char *ib_devname)\r
+{\r
+       struct ibv_srq_init_attr attr;\r
+\r
+       ctx.recv_qp = calloc(ctx.num_clients, sizeof *ctx.recv_qp);\r
+       ctx.send_qp = calloc(ctx.num_clients, sizeof *ctx.send_qp);\r
+       ctx.rem_dest = calloc(ctx.num_clients, sizeof *ctx.rem_dest);\r
+       if (!ctx.recv_qp || !ctx.send_qp || !ctx.rem_dest)\r
+               return 1;\r
+\r
+       if (open_device(ib_devname)) {\r
+               fprintf(stderr, "Failed to open device\n");\r
+               return 1;\r
+       }\r
+\r
+       ctx.lid = pp_get_local_lid(ctx.context, ctx.ib_port);\r
+       if (ctx.lid < 0) {\r
+               fprintf(stderr, "Failed to get SLID\n");\r
+               return 1;\r
+       }\r
+\r
+       ctx.buf = memalign(page_size, ctx.size);\r
+       if (!ctx.buf) {\r
+               fprintf(stderr, "Couldn't allocate work buf.\n");\r
+               return 1;\r
+       }\r
+\r
+       memset(ctx.buf, 0, ctx.size);\r
+\r
+       if (ctx.use_event) {\r
+               ctx.channel = ibv_create_comp_channel(ctx.context);\r
+               if (!ctx.channel) {\r
+                       fprintf(stderr, "Couldn't create completion channel\n");\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       ctx.pd = ibv_alloc_pd(ctx.context);\r
+       if (!ctx.pd) {\r
+               fprintf(stderr, "Couldn't allocate PD\n");\r
+               return 1;\r
+       }\r
+\r
+       ctx.mr = ibv_reg_mr(ctx.pd, ctx.buf, ctx.size, IBV_ACCESS_LOCAL_WRITE);\r
+       if (!ctx.mr) {\r
+               fprintf(stderr, "Couldn't register MR\n");\r
+               return 1;\r
+       }\r
+\r
+       ctx.fd = open("/tmp/xrc_domain", O_CREAT);\r
+       if (ctx.fd < 0) {\r
+               fprintf(stderr,\r
+                       "Couldn't create the file for the XRC Domain "\r
+                       "but not stopping %d\n", errno);\r
+               ctx.fd = -1;\r
+       }\r
+\r
+       ctx.xrcd = ibv_open_xrcd(ctx.context, ctx.fd, O_CREAT);\r
+       if (!ctx.xrcd) {\r
+               fprintf(stderr, "Couldn't Open the XRC Domain %d\n", errno);\r
+               return 1;\r
+       }\r
+\r
+       ctx.recv_cq = ibv_create_cq(ctx.context, ctx.num_clients, &ctx.recv_cq,\r
+                                   ctx.channel, 0);\r
+       if (!ctx.recv_cq) {\r
+               fprintf(stderr, "Couldn't create recv CQ\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ctx.use_event) {\r
+               if (ibv_req_notify_cq(ctx.recv_cq, 0)) {\r
+                       fprintf(stderr, "Couldn't request CQ notification\n");\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       ctx.send_cq = ibv_create_cq(ctx.context, ctx.num_clients, NULL, NULL, 0);\r
+       if (!ctx.send_cq) {\r
+               fprintf(stderr, "Couldn't create send CQ\n");\r
+               return 1;\r
+       }\r
+\r
+       memset(&attr, 0, sizeof attr);\r
+       attr.srq_type = IBV_SRQT_XRC;\r
+       attr.attr.max_wr = ctx.num_clients;\r
+       attr.attr.max_sge = 1;\r
+       attr.ext.xrc.xrcd = ctx.xrcd;\r
+       attr.ext.xrc.cq = ctx.recv_cq;\r
+\r
+       ctx.srq = ibv_create_xsrq(ctx.pd, &attr);\r
+       if (!ctx.srq)  {\r
+               fprintf(stderr, "Couldn't create SRQ\n");\r
+               return 1;\r
+       }\r
+\r
+       if (create_qps())\r
+               return 1;\r
+\r
+       return 0;\r
+}\r
+\r
+static int send_local_dest(int sockfd, int index)\r
+{\r
+       char msg[MSG_SIZE];\r
+\r
+       ctx.rem_dest[index].recv_psn = lrand48() & 0xffffff;\r
+\r
+       printf(ADDR_FORMAT, "local", ctx.lid, ctx.recv_qp[index]->qp_num,\r
+               ctx.send_qp[index]->qp_num, ctx.rem_dest[index].recv_psn,\r
+               ctx.srq->ext.xrc.srq_num);\r
+\r
+       sprintf(msg, MSG_FORMAT, ctx.lid, ctx.recv_qp[index]->qp_num,\r
+               ctx.send_qp[index]->qp_num, ctx.rem_dest[index].recv_psn,\r
+               ctx.srq->ext.xrc.srq_num);\r
+\r
+       if (write(sockfd, msg, MSG_SIZE) != MSG_SIZE) {\r
+               fprintf(stderr, "Couldn't send local address\n");\r
+               return -1;\r
+       }\r
+\r
+       return 0;\r
+}\r
+\r
+static int recv_remote_dest(int sockfd, int index)\r
+{\r
+       struct pingpong_dest *rem_dest;\r
+       char msg[MSG_SIZE];\r
+       int n = 0, r;\r
+\r
+       while (n < MSG_SIZE) {\r
+               r = read(sockfd, msg + n, MSG_SIZE - n);\r
+               if (r < 0) {\r
+                       perror("client read");\r
+                       fprintf(stderr,\r
+                               "%d/%d: Couldn't read remote address [%d]\n",\r
+                               n, MSG_SIZE, index);\r
+                       return -1;\r
+               }\r
+               n += r;\r
+       }\r
+\r
+       rem_dest = &ctx.rem_dest[index];\r
+       sscanf(msg, MSG_SSCAN, &rem_dest->lid, &rem_dest->recv_qpn,\r
+              &rem_dest->send_qpn, &rem_dest->send_psn, &rem_dest->srqn);\r
+\r
+       printf(ADDR_FORMAT, "remote", rem_dest->lid, rem_dest->recv_qpn,\r
+              rem_dest->send_qpn, rem_dest->send_psn, rem_dest->srqn);\r
+\r
+       return 0;\r
+}\r
+\r
+static int connect_qps(int index)\r
+{\r
+       struct ibv_qp_attr attr;\r
+\r
+       memset(&attr, 0, sizeof attr);\r
+       attr.qp_state         = IBV_QPS_RTR;\r
+       attr.dest_qp_num      = ctx.rem_dest[index].send_qpn;\r
+       attr.path_mtu         = ctx.mtu;\r
+       attr.rq_psn           = ctx.rem_dest[index].send_psn;\r
+       attr.min_rnr_timer    = 12;\r
+       attr.ah_attr.dlid     = ctx.rem_dest[index].lid;\r
+       attr.ah_attr.sl       = ctx.sl;\r
+       attr.ah_attr.port_num = ctx.ib_port;\r
+\r
+       if (ibv_modify_qp(ctx.recv_qp[index], &attr,\r
+                         IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU |\r
+                         IBV_QP_DEST_QPN | IBV_QP_RQ_PSN |\r
+                         IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER)) {\r
+               fprintf(stderr, "Failed to modify send QP[%d] to RTR\n", index);\r
+               return 1;\r
+       }\r
+\r
+       memset(&attr, 0, sizeof attr);\r
+       attr.qp_state = IBV_QPS_RTS;\r
+       attr.timeout = 14;\r
+       attr.sq_psn = ctx.rem_dest[index].recv_psn;\r
+\r
+       if (ibv_modify_qp(ctx.recv_qp[index], &attr,\r
+                         IBV_QP_STATE | IBV_QP_TIMEOUT | IBV_QP_SQ_PSN)) {\r
+               fprintf(stderr, "Failed to modify recv QP[%d] to RTS\n", index);\r
+               return 1;\r
+       }\r
+\r
+       memset(&attr, 0, sizeof attr);\r
+       attr.qp_state         = IBV_QPS_RTR;\r
+       attr.dest_qp_num      = ctx.rem_dest[index].recv_qpn;\r
+       attr.path_mtu         = ctx.mtu;\r
+       attr.rq_psn           = ctx.rem_dest[index].send_psn;\r
+       attr.ah_attr.dlid     = ctx.rem_dest[index].lid;\r
+       attr.ah_attr.sl       = ctx.sl;\r
+       attr.ah_attr.port_num = ctx.ib_port;\r
+\r
+       if (ibv_modify_qp(ctx.send_qp[index], &attr,\r
+                         IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU |\r
+                         IBV_QP_DEST_QPN | IBV_QP_RQ_PSN)) {\r
+               fprintf(stderr, "Failed to modify send QP[%d] to RTR\n", index);\r
+               return 1;\r
+       }\r
+\r
+       memset(&attr, 0, sizeof attr);\r
+       attr.qp_state = IBV_QPS_RTS;\r
+       attr.timeout = 14;\r
+       attr.retry_cnt = 7;\r
+       attr.rnr_retry = 7;\r
+       attr.sq_psn = ctx.rem_dest[index].recv_psn;\r
+\r
+       if (ibv_modify_qp(ctx.send_qp[index], &attr,\r
+                         IBV_QP_STATE | IBV_QP_TIMEOUT | IBV_QP_SQ_PSN |\r
+                         IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY | IBV_QP_MAX_QP_RD_ATOMIC)) {\r
+               fprintf(stderr, "Failed to modify send QP[%d] to RTS\n", index);\r
+               return 1;\r
+       }\r
+\r
+       return 0;\r
+}\r
+\r
+static int pp_client_connect(const char *servername, int port)\r
+{\r
+       struct addrinfo *res, *t;\r
+       char *service;\r
+       int ret;\r
+       int sockfd = -1;\r
+       struct addrinfo hints = {\r
+               .ai_family   = AF_UNSPEC,\r
+               .ai_socktype = SOCK_STREAM\r
+       };\r
+\r
+       if (asprintf(&service, "%d", port) < 0)\r
+               return 1;\r
+\r
+       ret = getaddrinfo(servername, service, &hints, &res);\r
+       if (ret < 0) {\r
+               fprintf(stderr, "%s for %s:%d\n", gai_strerror(ret), servername, port);\r
+               free(service);\r
+               return 1;\r
+       }\r
+\r
+       for (t = res; t; t = t->ai_next) {\r
+               sockfd = socket(t->ai_family, t->ai_socktype, t->ai_protocol);\r
+               if (sockfd >= 0) {\r
+                       if (!connect(sockfd, t->ai_addr, t->ai_addrlen))\r
+                               break;\r
+                       close(sockfd);\r
+                       sockfd = -1;\r
+               }\r
+       }\r
+\r
+       freeaddrinfo(res);\r
+       free(service);\r
+\r
+       if (sockfd < 0) {\r
+               fprintf(stderr, "Couldn't connect to %s:%d\n", servername, port);\r
+               return 1;\r
+       }\r
+\r
+       if (send_local_dest(sockfd, 0))\r
+               return 1;\r
+\r
+       if (recv_remote_dest(sockfd, 0))\r
+               return 1;\r
+\r
+       if (connect_qps(0))\r
+               return 1;\r
+\r
+       shutdown(sockfd, SHUT_RDWR);\r
+       close(sockfd);\r
+       return 0;\r
+}\r
+\r
+static int pp_server_connect(int port)\r
+{\r
+       struct addrinfo *res, *t;\r
+       char *service;\r
+       int ret, i, n;\r
+       int sockfd = -1, connfd;\r
+       struct addrinfo hints = {\r
+               .ai_flags    = AI_PASSIVE,\r
+               .ai_family   = AF_UNSPEC,\r
+               .ai_socktype = SOCK_STREAM\r
+       };\r
+\r
+       if (asprintf(&service, "%d", port) < 0)\r
+               return 1;\r
+\r
+       ret = getaddrinfo(NULL, service, &hints, &res);\r
+       if (ret < 0) {\r
+               fprintf(stderr, "%s for port %d\n", gai_strerror(ret), port);\r
+               free(service);\r
+               return 1;\r
+       }\r
+\r
+       for (t = res; t; t = t->ai_next) {\r
+               sockfd = socket(t->ai_family, t->ai_socktype, t->ai_protocol);\r
+               if (sockfd >= 0) {\r
+                       n = 1;\r
+                       setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &n, sizeof n);\r
+                       if (!bind(sockfd, t->ai_addr, t->ai_addrlen))\r
+                               break;\r
+                       close(sockfd);\r
+                       sockfd = -1;\r
+               }\r
+       }\r
+\r
+       freeaddrinfo(res);\r
+       free(service);\r
+\r
+       if (sockfd < 0) {\r
+               fprintf(stderr, "Couldn't listen to port %d\n", port);\r
+               return 1;\r
+       }\r
+\r
+       listen(sockfd, ctx.num_clients);\r
+\r
+       for (i = 0; i < ctx.num_clients; i++) {\r
+               connfd = accept(sockfd, NULL, 0);\r
+               if (connfd < 0) {\r
+                       fprintf(stderr, "accept() failed for client %d\n", i);\r
+                       return 1;\r
+               }\r
+\r
+               if (recv_remote_dest(connfd, i))\r
+                       return 1;\r
+\r
+               if (send_local_dest(connfd, i))\r
+                       return 1;\r
+\r
+               if (connect_qps(i))\r
+                       return 1;\r
+\r
+               shutdown(connfd, SHUT_RDWR);\r
+               close(connfd);\r
+       }\r
+\r
+       close(sockfd);\r
+       return 0;\r
+}\r
+\r
+\r
+static int pp_close_ctx(void)\r
+{\r
+       int i;\r
+\r
+       for (i = 0; i < ctx.num_clients; ++i) {\r
+               /*\r
+                * Receive QPs may be used by other clients.  Leave them, they\r
+                * will be cleaned up when the xrcd is destroyed.\r
+                */\r
+               if (ibv_destroy_qp(ctx.send_qp[i])) {\r
+                       fprintf(stderr, "Couldn't destroy INI QP[%d]\n", i);\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       if (ibv_destroy_srq(ctx.srq)) {\r
+               fprintf(stderr, "Couldn't destroy SRQ\n");\r
+               return 1;\r
+       }\r
+\r
+       if (!ctx.xrcd && ibv_close_xrcd(ctx.xrcd)) {\r
+               fprintf(stderr, "Couldn't close the XRC Domain\n");\r
+               return 1;\r
+       }\r
+       if (ctx.fd >= 0 && close(ctx.fd)) {\r
+               fprintf(stderr, "Couldn't close the file for the XRC Domain\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ibv_destroy_cq(ctx.send_cq)) {\r
+               fprintf(stderr, "Couldn't destroy send CQ\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ibv_destroy_cq(ctx.recv_cq)) {\r
+               fprintf(stderr, "Couldn't destroy recv CQ\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ibv_dereg_mr(ctx.mr)) {\r
+               fprintf(stderr, "Couldn't deregister MR\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ibv_dealloc_pd(ctx.pd)) {\r
+               fprintf(stderr, "Couldn't deallocate PD\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ctx.channel) {\r
+               if (ibv_destroy_comp_channel(ctx.channel)) {\r
+                       fprintf(stderr,\r
+                               "Couldn't destroy completion channel\n");\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       if (ibv_close_device(ctx.context)) {\r
+               fprintf(stderr, "Couldn't release context\n");\r
+               return 1;\r
+       }\r
+\r
+       free(ctx.buf);\r
+       free(ctx.rem_dest);\r
+       free(ctx.send_qp);\r
+       free(ctx.recv_qp);\r
+       return 0;\r
+}\r
+\r
+static int pp_post_recv(int cnt)\r
+{\r
+       struct ibv_sge sge;\r
+       struct ibv_recv_wr wr, *bad_wr;\r
+\r
+       sge.addr = (uintptr_t) ctx.buf;\r
+       sge.length = ctx.size;\r
+       sge.lkey = ctx.mr->lkey;\r
+\r
+       wr.next       = NULL;\r
+       wr.wr_id      = (uintptr_t) &ctx;\r
+       wr.sg_list    = &sge;\r
+       wr.num_sge    = 1;\r
+\r
+       while (cnt--) {\r
+               if (ibv_post_srq_recv(ctx.srq, &wr, &bad_wr)) {\r
+                       fprintf(stderr, "Failed to post receive to SRQ\n");\r
+                       return 1;\r
+               }\r
+       }\r
+       return 0;\r
+}\r
+\r
+/*\r
+ * Send to each client round robin on each set of xrc send/recv qp.\r
+ * Generate a completion on the last send.\r
+ */\r
+static int pp_post_send(int index)\r
+{\r
+       struct ibv_sge sge;\r
+       struct ibv_send_wr wr, *bad_wr;\r
+       int qpi;\r
+\r
+       sge.addr = (uintptr_t) ctx.buf;\r
+       sge.length = ctx.size;\r
+       sge.lkey = ctx.mr->lkey;\r
+\r
+       wr.wr_id   = (uintptr_t) index;\r
+       wr.next    = NULL;\r
+       wr.sg_list = &sge;\r
+       wr.num_sge = 1;\r
+       wr.opcode  = IBV_WR_SEND;\r
+       wr.wr.xrc.remote_srqn = ctx.rem_dest[index].srqn;\r
+\r
+       qpi = (index + ctx.rem_dest[index].pp_cnt) % ctx.num_clients;\r
+       wr.send_flags = (++ctx.rem_dest[index].pp_cnt >= ctx.num_tests) ?\r
+                       IBV_SEND_SIGNALED : 0;\r
+\r
+       return ibv_post_send(ctx.send_qp[qpi], &wr, &bad_wr);\r
+}\r
+\r
+static int find_qp(int qpn)\r
+{\r
+       int i;\r
+\r
+       if (ctx.num_clients == 1)\r
+               return 0;\r
+\r
+       for (i = 0; i < ctx.num_clients; ++i)\r
+               if (ctx.recv_qp[i]->qp_num == qpn)\r
+                       return i;\r
+\r
+       fprintf(stderr, "Unable to find qp %x\n", qpn);\r
+       return 0;\r
+}\r
+\r
+static int get_cq_event(void)\r
+{\r
+       struct ibv_cq *ev_cq;\r
+       void          *ev_ctx;\r
+\r
+       if (ibv_get_cq_event(ctx.channel, &ev_cq, &ev_ctx)) {\r
+               fprintf(stderr, "Failed to get cq_event\n");\r
+               return 1;\r
+       }\r
+\r
+       if (ev_cq != ctx.recv_cq) {\r
+               fprintf(stderr, "CQ event for unknown CQ %p\n", ev_cq);\r
+               return 1;\r
+       }\r
+\r
+       if (ibv_req_notify_cq(ctx.recv_cq, 0)) {\r
+               fprintf(stderr, "Couldn't request CQ notification\n");\r
+               return 1;\r
+       }\r
+\r
+       return 0;\r
+}\r
+\r
+static void init(void)\r
+{\r
+       srand48(getpid() * time(NULL));\r
+\r
+       ctx.size = 4096;\r
+       ctx.ib_port = 1;\r
+       ctx.num_clients  = 1;\r
+       ctx.num_tests = 5;\r
+       ctx.mtu = IBV_MTU_2048;\r
+       ctx.sl = 0;\r
+}\r
+\r
+static void usage(const char *argv0)\r
+{\r
+       printf("Usage:\n");\r
+       printf("  %s            start a server and wait for connection\n", argv0);\r
+       printf("  %s <host>     connect to server at <host>\n", argv0);\r
+       printf("\n");\r
+       printf("Options:\n");\r
+       printf("  -p, --port=<port>      listen on/connect to port <port> (default 18515)\n");\r
+       printf("  -d, --ib-dev=<dev>     use IB device <dev> (default first device found)\n");\r
+       printf("  -i, --ib-port=<port>   use port <port> of IB device (default 1)\n");\r
+       printf("  -s, --size=<size>      size of message to exchange (default 4096)\n");\r
+       printf("  -m, --mtu=<size>       path MTU (default 2048)\n");\r
+       printf("  -c, --clients=<n>      number of clients (on server only, default 1)\n");\r
+       printf("  -n, --num_tests=<n>    number of tests per client (default 5)\n");\r
+       printf("  -l, --sl=<sl>          service level value\n");\r
+       printf("  -e, --events           sleep on CQ events (default poll)\n");\r
+}\r
+\r
+int main(int argc, char *argv[])\r
+{\r
+       char          *ib_devname = NULL;\r
+       char          *servername = NULL;\r
+       int           port = 18515;\r
+       int           i, total, cnt = 0;\r
+       int           ne, qpi, num_cq_events = 0;\r
+       struct ibv_wc wc;\r
+\r
+       init();\r
+       while (1) {\r
+               int c;\r
+\r
+               static struct option long_options[] = {\r
+                       { .name = "port",      .has_arg = 1, .val = 'p' },\r
+                       { .name = "ib-dev",    .has_arg = 1, .val = 'd' },\r
+                       { .name = "ib-port",   .has_arg = 1, .val = 'i' },\r
+                       { .name = "size",      .has_arg = 1, .val = 's' },\r
+                       { .name = "mtu",       .has_arg = 1, .val = 'm' },\r
+                       { .name = "clients",   .has_arg = 1, .val = 'c' },\r
+                       { .name = "num_tests", .has_arg = 1, .val = 'n' },\r
+                       { .name = "sl",        .has_arg = 1, .val = 'l' },\r
+                       { .name = "events",    .has_arg = 0, .val = 'e' },\r
+                       { 0 }\r
+               };\r
+\r
+               c = getopt_long(argc, argv, "p:d:i:s:m:c:n:l:e", long_options,\r
+                               NULL);\r
+               if (c == -1)\r
+                       break;\r
+\r
+               switch (c) {\r
+               case 'p':\r
+                       port = strtol(optarg, NULL, 0);\r
+                       if (port < 0 || port > 65535) {\r
+                               usage(argv[0]);\r
+                               return 1;\r
+                       }\r
+                       break;\r
+               case 'd':\r
+                       ib_devname = strdupa(optarg);\r
+                       break;\r
+               case 'i':\r
+                       ctx.ib_port = strtol(optarg, NULL, 0);\r
+                       if (ctx.ib_port < 0) {\r
+                               usage(argv[0]);\r
+                               return 1;\r
+                       }\r
+                       break;\r
+               case 's':\r
+                       ctx.size = strtol(optarg, NULL, 0);\r
+                       break;\r
+               case 'm':\r
+                       ctx.mtu = pp_mtu_to_enum(strtol(optarg, NULL, 0));\r
+                       if (ctx.mtu < 0) {\r
+                               usage(argv[0]);\r
+                               return 1;\r
+                       }\r
+                       break;\r
+               case 'c':\r
+                       ctx.num_clients = strtol(optarg, NULL, 0);\r
+                       break;\r
+               case 'n':\r
+                       ctx.num_tests = strtol(optarg, NULL, 0);\r
+                       break;\r
+               case 'l':\r
+                       ctx.sl = strtol(optarg, NULL, 0);\r
+                       break;\r
+               case 'e':\r
+                       ctx.use_event = 1;\r
+                       break;\r
+               default:\r
+                       usage(argv[0]);\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       if (optind == argc - 1) {\r
+               servername = strdupa(argv[optind]);\r
+               ctx.num_clients = 1;\r
+       } else if (optind < argc) {\r
+               usage(argv[0]);\r
+               return 1;\r
+       }\r
+\r
+       page_size = sysconf(_SC_PAGESIZE);\r
+\r
+       if (pp_init_ctx(ib_devname))\r
+               return 1;\r
+\r
+       if (pp_post_recv(ctx.num_clients)) {\r
+               fprintf(stderr, "Couldn't post receives\n");\r
+               return 1;\r
+       }\r
+\r
+       if (servername) {\r
+               if (pp_client_connect(servername, port))\r
+                       return 1;\r
+       } else {\r
+               if (pp_server_connect(port))\r
+                       return 1;\r
+\r
+               for (i = 0; i < ctx.num_clients; i++)\r
+                       pp_post_send(i);\r
+       }\r
+\r
+       total = ctx.num_clients * ctx.num_tests;\r
+       while (cnt < total) {\r
+               if (ctx.use_event) {\r
+                       if (get_cq_event())\r
+                               return 1;\r
+\r
+                       ++num_cq_events;\r
+               }\r
+\r
+               do {\r
+                       ne = ibv_poll_cq(ctx.recv_cq, 1, &wc);\r
+                       if (ne < 0) {\r
+                               fprintf(stderr, "Error polling cq %d\n", ne);\r
+                               return 1;\r
+                       } else if (ne == 0) {\r
+                               break;\r
+                       }\r
+\r
+                       if (wc.status) {\r
+                               fprintf(stderr, "Work completion error %d\n", wc.status);\r
+                               return 1;\r
+                       }\r
+\r
+                       pp_post_recv(ne);\r
+                       qpi = find_qp(wc.qp_num);\r
+                       if (ctx.rem_dest[qpi].pp_cnt < ctx.num_tests)\r
+                               pp_post_send(qpi);\r
+                       cnt += ne;\r
+               } while (ne > 0);\r
+       }\r
+\r
+       for (cnt = 0; cnt < ctx.num_clients; cnt += ne) {\r
+               ne = ibv_poll_cq(ctx.send_cq, 1, &wc);\r
+               if (ne < 0) {\r
+                       fprintf(stderr, "Error polling cq %d\n", ne);\r
+                       return 1;\r
+               }\r
+       }\r
+\r
+       if (ctx.use_event)\r
+               ibv_ack_cq_events(ctx.recv_cq, num_cq_events);\r
+\r
+       if (pp_close_ctx())\r
+               return 1;\r
+\r
+       return 0;\r
+}\r