1/* 2 * Copyright (c) 2006 Oracle. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 * 32 */ 33#include <linux/kernel.h> 34#include <linux/in.h> 35#include <linux/slab.h> 36#include <linux/vmalloc.h> 37 38#include "rds.h" 39#include "ib.h" 40 41/* 42 * Set the selected protocol version 43 */ 44static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) 45{ 46 conn->c_version = version; 47} 48 49/* 50 * Set up flow control 51 */ 52static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) 53{ 54 struct rds_ib_connection *ic = conn->c_transport_data; 55 56 if (rds_ib_sysctl_flow_control && credits != 0) { 57 /* We're doing flow control */ 58 ic->i_flowctl = 1; 59 rds_ib_send_add_credits(conn, credits); 60 } else { 61 ic->i_flowctl = 0; 62 } 63} 64 65/* 66 * Tune RNR behavior. Without flow control, we use a rather 67 * low timeout, but not the absolute minimum - this should 68 * be tunable. 69 * 70 * We already set the RNR retry count to 7 (which is the 71 * smallest infinite number :-) above. 72 * If flow control is off, we want to change this back to 0 73 * so that we learn quickly when our credit accounting is 74 * buggy. 75 * 76 * Caller passes in a qp_attr pointer - don't waste stack spacv 77 * by allocation this twice. 78 */ 79static void 80rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) 81{ 82 int ret; 83 84 attr->min_rnr_timer = IB_RNR_TIMER_000_32; 85 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); 86 if (ret) 87 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); 88} 89 90/* 91 * Connection established. 92 * We get here for both outgoing and incoming connection. 93 */ 94void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) 95{ 96 const struct rds_ib_connect_private *dp = NULL; 97 struct rds_ib_connection *ic = conn->c_transport_data; 98 struct rds_ib_device *rds_ibdev; 99 struct ib_qp_attr qp_attr; 100 int err; 101 102 if (event->param.conn.private_data_len >= sizeof(*dp)) { 103 dp = event->param.conn.private_data; 104 105 /* make sure it isn't empty data */ 106 if (dp->dp_protocol_major) { 107 rds_ib_set_protocol(conn, 108 RDS_PROTOCOL(dp->dp_protocol_major, 109 dp->dp_protocol_minor)); 110 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); 111 } 112 } 113 114 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", 115 &conn->c_faddr, 116 RDS_PROTOCOL_MAJOR(conn->c_version), 117 RDS_PROTOCOL_MINOR(conn->c_version), 118 ic->i_flowctl ? ", flow control" : ""); 119 120 /* 121 * Init rings and fill recv. this needs to wait until protocol negotiation 122 * is complete, since ring layout is different from 3.0 to 3.1. 123 */ 124 rds_ib_send_init_ring(ic); 125 rds_ib_recv_init_ring(ic); 126 /* Post receive buffers - as a side effect, this will update 127 * the posted credit count. */ 128 rds_ib_recv_refill(conn, GFP_KERNEL, GFP_HIGHUSER, 1); 129 130 /* Tune RNR behavior */ 131 rds_ib_tune_rnr(ic, &qp_attr); 132 133 qp_attr.qp_state = IB_QPS_RTS; 134 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); 135 if (err) 136 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); 137 138 /* update ib_device with this local ipaddr & conn */ 139 rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client); 140 err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr); 141 if (err) 142 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err); 143 rds_ib_add_conn(rds_ibdev, conn); 144 145 /* If the peer gave us the last packet it saw, process this as if 146 * we had received a regular ACK. */ 147 if (dp && dp->dp_ack_seq) 148 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); 149 150 rds_connect_complete(conn); 151} 152 153static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, 154 struct rdma_conn_param *conn_param, 155 struct rds_ib_connect_private *dp, 156 u32 protocol_version) 157{ 158 memset(conn_param, 0, sizeof(struct rdma_conn_param)); 159 conn_param->responder_resources = 1; 160 conn_param->initiator_depth = 1; 161 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); 162 conn_param->rnr_retry_count = 7; 163 164 if (dp) { 165 struct rds_ib_connection *ic = conn->c_transport_data; 166 167 memset(dp, 0, sizeof(*dp)); 168 dp->dp_saddr = conn->c_laddr; 169 dp->dp_daddr = conn->c_faddr; 170 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); 171 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); 172 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); 173 dp->dp_ack_seq = rds_ib_piggyb_ack(ic); 174 175 /* Advertise flow control */ 176 if (ic->i_flowctl) { 177 unsigned int credits; 178 179 credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); 180 dp->dp_credit = cpu_to_be32(credits); 181 atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); 182 } 183 184 conn_param->private_data = dp; 185 conn_param->private_data_len = sizeof(*dp); 186 } 187} 188 189static void rds_ib_cq_event_handler(struct ib_event *event, void *data) 190{ 191 rdsdebug("event %u data %p\n", event->event, data); 192} 193 194static void rds_ib_qp_event_handler(struct ib_event *event, void *data) 195{ 196 struct rds_connection *conn = data; 197 struct rds_ib_connection *ic = conn->c_transport_data; 198 199 rdsdebug("conn %p ic %p event %u\n", conn, ic, event->event); 200 201 switch (event->event) { 202 case IB_EVENT_COMM_EST: 203 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); 204 break; 205 default: 206 rdsdebug("Fatal QP Event %u " 207 "- connection %pI4->%pI4, reconnecting\n", 208 event->event, &conn->c_laddr, &conn->c_faddr); 209 rds_conn_drop(conn); 210 break; 211 } 212} 213 214/* 215 * This needs to be very careful to not leave IS_ERR pointers around for 216 * cleanup to trip over. 217 */ 218static int rds_ib_setup_qp(struct rds_connection *conn) 219{ 220 struct rds_ib_connection *ic = conn->c_transport_data; 221 struct ib_device *dev = ic->i_cm_id->device; 222 struct ib_qp_init_attr attr; 223 struct rds_ib_device *rds_ibdev; 224 int ret; 225 226 /* rds_ib_add_one creates a rds_ib_device object per IB device, 227 * and allocates a protection domain, memory range and FMR pool 228 * for each. If that fails for any reason, it will not register 229 * the rds_ibdev at all. 230 */ 231 rds_ibdev = ib_get_client_data(dev, &rds_ib_client); 232 if (rds_ibdev == NULL) { 233 if (printk_ratelimit()) 234 printk(KERN_NOTICE "RDS/IB: No client_data for device %s\n", 235 dev->name); 236 return -EOPNOTSUPP; 237 } 238 239 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) 240 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); 241 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) 242 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); 243 244 /* Protection domain and memory range */ 245 ic->i_pd = rds_ibdev->pd; 246 ic->i_mr = rds_ibdev->mr; 247 248 ic->i_send_cq = ib_create_cq(dev, rds_ib_send_cq_comp_handler, 249 rds_ib_cq_event_handler, conn, 250 ic->i_send_ring.w_nr + 1, 0); 251 if (IS_ERR(ic->i_send_cq)) { 252 ret = PTR_ERR(ic->i_send_cq); 253 ic->i_send_cq = NULL; 254 rdsdebug("ib_create_cq send failed: %d\n", ret); 255 goto out; 256 } 257 258 ic->i_recv_cq = ib_create_cq(dev, rds_ib_recv_cq_comp_handler, 259 rds_ib_cq_event_handler, conn, 260 ic->i_recv_ring.w_nr, 0); 261 if (IS_ERR(ic->i_recv_cq)) { 262 ret = PTR_ERR(ic->i_recv_cq); 263 ic->i_recv_cq = NULL; 264 rdsdebug("ib_create_cq recv failed: %d\n", ret); 265 goto out; 266 } 267 268 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); 269 if (ret) { 270 rdsdebug("ib_req_notify_cq send failed: %d\n", ret); 271 goto out; 272 } 273 274 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 275 if (ret) { 276 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); 277 goto out; 278 } 279 280 memset(&attr, 0, sizeof(attr)); 281 attr.event_handler = rds_ib_qp_event_handler; 282 attr.qp_context = conn; 283 /* + 1 to allow for the single ack message */ 284 attr.cap.max_send_wr = ic->i_send_ring.w_nr + 1; 285 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; 286 attr.cap.max_send_sge = rds_ibdev->max_sge; 287 attr.cap.max_recv_sge = RDS_IB_RECV_SGE; 288 attr.sq_sig_type = IB_SIGNAL_REQ_WR; 289 attr.qp_type = IB_QPT_RC; 290 attr.send_cq = ic->i_send_cq; 291 attr.recv_cq = ic->i_recv_cq; 292 293 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); 294 if (ret) { 295 rdsdebug("rdma_create_qp failed: %d\n", ret); 296 goto out; 297 } 298 299 ic->i_send_hdrs = ib_dma_alloc_coherent(dev, 300 ic->i_send_ring.w_nr * 301 sizeof(struct rds_header), 302 &ic->i_send_hdrs_dma, GFP_KERNEL); 303 if (ic->i_send_hdrs == NULL) { 304 ret = -ENOMEM; 305 rdsdebug("ib_dma_alloc_coherent send failed\n"); 306 goto out; 307 } 308 309 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, 310 ic->i_recv_ring.w_nr * 311 sizeof(struct rds_header), 312 &ic->i_recv_hdrs_dma, GFP_KERNEL); 313 if (ic->i_recv_hdrs == NULL) { 314 ret = -ENOMEM; 315 rdsdebug("ib_dma_alloc_coherent recv failed\n"); 316 goto out; 317 } 318 319 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), 320 &ic->i_ack_dma, GFP_KERNEL); 321 if (ic->i_ack == NULL) { 322 ret = -ENOMEM; 323 rdsdebug("ib_dma_alloc_coherent ack failed\n"); 324 goto out; 325 } 326 327 ic->i_sends = vmalloc(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); 328 if (ic->i_sends == NULL) { 329 ret = -ENOMEM; 330 rdsdebug("send allocation failed\n"); 331 goto out; 332 } 333 memset(ic->i_sends, 0, ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work)); 334 335 ic->i_recvs = vmalloc(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); 336 if (ic->i_recvs == NULL) { 337 ret = -ENOMEM; 338 rdsdebug("recv allocation failed\n"); 339 goto out; 340 } 341 memset(ic->i_recvs, 0, ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work)); 342 343 rds_ib_recv_init_ack(ic); 344 345 rdsdebug("conn %p pd %p mr %p cq %p %p\n", conn, ic->i_pd, ic->i_mr, 346 ic->i_send_cq, ic->i_recv_cq); 347 348out: 349 return ret; 350} 351 352static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) 353{ 354 const struct rds_ib_connect_private *dp = event->param.conn.private_data; 355 u16 common; 356 u32 version = 0; 357 358 /* 359 * rdma_cm private data is odd - when there is any private data in the 360 * request, we will be given a pretty large buffer without telling us the 361 * original size. The only way to tell the difference is by looking at 362 * the contents, which are initialized to zero. 363 * If the protocol version fields aren't set, this is a connection attempt 364 * from an older version. This could could be 3.0 or 2.0 - we can't tell. 365 * We really should have changed this for OFED 1.3 :-( 366 */ 367 368 /* Be paranoid. RDS always has privdata */ 369 if (!event->param.conn.private_data_len) { 370 printk(KERN_NOTICE "RDS incoming connection has no private data, " 371 "rejecting\n"); 372 return 0; 373 } 374 375 /* Even if len is crap *now* I still want to check it. -ASG */ 376 if (event->param.conn.private_data_len < sizeof (*dp) || 377 dp->dp_protocol_major == 0) 378 return RDS_PROTOCOL_3_0; 379 380 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; 381 if (dp->dp_protocol_major == 3 && common) { 382 version = RDS_PROTOCOL_3_0; 383 while ((common >>= 1) != 0) 384 version++; 385 } else if (printk_ratelimit()) { 386 printk(KERN_NOTICE "RDS: Connection from %pI4 using " 387 "incompatible protocol version %u.%u\n", 388 &dp->dp_saddr, 389 dp->dp_protocol_major, 390 dp->dp_protocol_minor); 391 } 392 return version; 393} 394 395int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 396 struct rdma_cm_event *event) 397{ 398 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; 399 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; 400 const struct rds_ib_connect_private *dp = event->param.conn.private_data; 401 struct rds_ib_connect_private dp_rep; 402 struct rds_connection *conn = NULL; 403 struct rds_ib_connection *ic = NULL; 404 struct rdma_conn_param conn_param; 405 u32 version; 406 int err, destroy = 1; 407 408 /* Check whether the remote protocol version matches ours. */ 409 version = rds_ib_protocol_compatible(event); 410 if (!version) 411 goto out; 412 413 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " 414 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, 415 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), 416 (unsigned long long)be64_to_cpu(lguid), 417 (unsigned long long)be64_to_cpu(fguid)); 418 419 conn = rds_conn_create(dp->dp_daddr, dp->dp_saddr, &rds_ib_transport, 420 GFP_KERNEL); 421 if (IS_ERR(conn)) { 422 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); 423 conn = NULL; 424 goto out; 425 } 426 427 /* 428 * The connection request may occur while the 429 * previous connection exist, e.g. in case of failover. 430 * But as connections may be initiated simultaneously 431 * by both hosts, we have a random backoff mechanism - 432 * see the comment above rds_queue_reconnect() 433 */ 434 mutex_lock(&conn->c_cm_lock); 435 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { 436 if (rds_conn_state(conn) == RDS_CONN_UP) { 437 rdsdebug("incoming connect while connecting\n"); 438 rds_conn_drop(conn); 439 rds_ib_stats_inc(s_ib_listen_closed_stale); 440 } else 441 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { 442 /* Wait and see - our connect may still be succeeding */ 443 rds_ib_stats_inc(s_ib_connect_raced); 444 } 445 mutex_unlock(&conn->c_cm_lock); 446 goto out; 447 } 448 449 ic = conn->c_transport_data; 450 451 rds_ib_set_protocol(conn, version); 452 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); 453 454 /* If the peer gave us the last packet it saw, process this as if 455 * we had received a regular ACK. */ 456 if (dp->dp_ack_seq) 457 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); 458 459 BUG_ON(cm_id->context); 460 BUG_ON(ic->i_cm_id); 461 462 ic->i_cm_id = cm_id; 463 cm_id->context = conn; 464 465 /* We got halfway through setting up the ib_connection, if we 466 * fail now, we have to take the long route out of this mess. */ 467 destroy = 0; 468 469 err = rds_ib_setup_qp(conn); 470 if (err) { 471 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); 472 mutex_unlock(&conn->c_cm_lock); 473 goto out; 474 } 475 476 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version); 477 478 /* rdma_accept() calls rdma_reject() internally if it fails */ 479 err = rdma_accept(cm_id, &conn_param); 480 mutex_unlock(&conn->c_cm_lock); 481 if (err) { 482 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); 483 goto out; 484 } 485 486 return 0; 487 488out: 489 rdma_reject(cm_id, NULL, 0); 490 return destroy; 491} 492 493 494int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) 495{ 496 struct rds_connection *conn = cm_id->context; 497 struct rds_ib_connection *ic = conn->c_transport_data; 498 struct rdma_conn_param conn_param; 499 struct rds_ib_connect_private dp; 500 int ret; 501 502 /* If the peer doesn't do protocol negotiation, we must 503 * default to RDSv3.0 */ 504 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); 505 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ 506 507 ret = rds_ib_setup_qp(conn); 508 if (ret) { 509 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); 510 goto out; 511 } 512 513 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION); 514 515 ret = rdma_connect(cm_id, &conn_param); 516 if (ret) 517 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); 518 519out: 520 /* Beware - returning non-zero tells the rdma_cm to destroy 521 * the cm_id. We should certainly not do it as long as we still 522 * "own" the cm_id. */ 523 if (ret) { 524 if (ic->i_cm_id == cm_id) 525 ret = 0; 526 } 527 return ret; 528} 529 530int rds_ib_conn_connect(struct rds_connection *conn) 531{ 532 struct rds_ib_connection *ic = conn->c_transport_data; 533 struct sockaddr_in src, dest; 534 int ret; 535 536 /* delegate cm event handler to rdma_transport */ 537 ic->i_cm_id = rdma_create_id(rds_rdma_cm_event_handler, conn, 538 RDMA_PS_TCP); 539 if (IS_ERR(ic->i_cm_id)) { 540 ret = PTR_ERR(ic->i_cm_id); 541 ic->i_cm_id = NULL; 542 rdsdebug("rdma_create_id() failed: %d\n", ret); 543 goto out; 544 } 545 546 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); 547 548 src.sin_family = AF_INET; 549 src.sin_addr.s_addr = (__force u32)conn->c_laddr; 550 src.sin_port = (__force u16)htons(0); 551 552 dest.sin_family = AF_INET; 553 dest.sin_addr.s_addr = (__force u32)conn->c_faddr; 554 dest.sin_port = (__force u16)htons(RDS_PORT); 555 556 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, 557 (struct sockaddr *)&dest, 558 RDS_RDMA_RESOLVE_TIMEOUT_MS); 559 if (ret) { 560 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, 561 ret); 562 rdma_destroy_id(ic->i_cm_id); 563 ic->i_cm_id = NULL; 564 } 565 566out: 567 return ret; 568} 569 570/* 571 * This is so careful about only cleaning up resources that were built up 572 * so that it can be called at any point during startup. In fact it 573 * can be called multiple times for a given connection. 574 */ 575void rds_ib_conn_shutdown(struct rds_connection *conn) 576{ 577 struct rds_ib_connection *ic = conn->c_transport_data; 578 int err = 0; 579 580 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, 581 ic->i_pd, ic->i_send_cq, ic->i_recv_cq, 582 ic->i_cm_id ? ic->i_cm_id->qp : NULL); 583 584 if (ic->i_cm_id) { 585 struct ib_device *dev = ic->i_cm_id->device; 586 587 rdsdebug("disconnecting cm %p\n", ic->i_cm_id); 588 err = rdma_disconnect(ic->i_cm_id); 589 if (err) { 590 /* Actually this may happen quite frequently, when 591 * an outgoing connect raced with an incoming connect. 592 */ 593 rdsdebug("failed to disconnect, cm: %p err %d\n", 594 ic->i_cm_id, err); 595 } 596 597 wait_event(rds_ib_ring_empty_wait, 598 rds_ib_ring_empty(&ic->i_send_ring) && 599 rds_ib_ring_empty(&ic->i_recv_ring)); 600 601 if (ic->i_send_hdrs) 602 ib_dma_free_coherent(dev, 603 ic->i_send_ring.w_nr * 604 sizeof(struct rds_header), 605 ic->i_send_hdrs, 606 ic->i_send_hdrs_dma); 607 608 if (ic->i_recv_hdrs) 609 ib_dma_free_coherent(dev, 610 ic->i_recv_ring.w_nr * 611 sizeof(struct rds_header), 612 ic->i_recv_hdrs, 613 ic->i_recv_hdrs_dma); 614 615 if (ic->i_ack) 616 ib_dma_free_coherent(dev, sizeof(struct rds_header), 617 ic->i_ack, ic->i_ack_dma); 618 619 if (ic->i_sends) 620 rds_ib_send_clear_ring(ic); 621 if (ic->i_recvs) 622 rds_ib_recv_clear_ring(ic); 623 624 if (ic->i_cm_id->qp) 625 rdma_destroy_qp(ic->i_cm_id); 626 if (ic->i_send_cq) 627 ib_destroy_cq(ic->i_send_cq); 628 if (ic->i_recv_cq) 629 ib_destroy_cq(ic->i_recv_cq); 630 rdma_destroy_id(ic->i_cm_id); 631 632 /* 633 * Move connection back to the nodev list. 634 */ 635 if (ic->rds_ibdev) 636 rds_ib_remove_conn(ic->rds_ibdev, conn); 637 638 ic->i_cm_id = NULL; 639 ic->i_pd = NULL; 640 ic->i_mr = NULL; 641 ic->i_send_cq = NULL; 642 ic->i_recv_cq = NULL; 643 ic->i_send_hdrs = NULL; 644 ic->i_recv_hdrs = NULL; 645 ic->i_ack = NULL; 646 } 647 BUG_ON(ic->rds_ibdev); 648 649 /* Clear pending transmit */ 650 if (ic->i_rm) { 651 rds_message_put(ic->i_rm); 652 ic->i_rm = NULL; 653 } 654 655 /* Clear the ACK state */ 656 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 657#ifdef KERNEL_HAS_ATOMIC64 658 atomic64_set(&ic->i_ack_next, 0); 659#else 660 ic->i_ack_next = 0; 661#endif 662 ic->i_ack_recv = 0; 663 664 /* Clear flow control state */ 665 ic->i_flowctl = 0; 666 atomic_set(&ic->i_credits, 0); 667 668 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); 669 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); 670 671 if (ic->i_ibinc) { 672 rds_inc_put(&ic->i_ibinc->ii_inc); 673 ic->i_ibinc = NULL; 674 } 675 676 vfree(ic->i_sends); 677 ic->i_sends = NULL; 678 vfree(ic->i_recvs); 679 ic->i_recvs = NULL; 680} 681 682int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) 683{ 684 struct rds_ib_connection *ic; 685 unsigned long flags; 686 687 ic = kzalloc(sizeof(struct rds_ib_connection), GFP_KERNEL); 688 if (ic == NULL) 689 return -ENOMEM; 690 691 INIT_LIST_HEAD(&ic->ib_node); 692 tasklet_init(&ic->i_recv_tasklet, rds_ib_recv_tasklet_fn, 693 (unsigned long) ic); 694 mutex_init(&ic->i_recv_mutex); 695#ifndef KERNEL_HAS_ATOMIC64 696 spin_lock_init(&ic->i_ack_lock); 697#endif 698 699 /* 700 * rds_ib_conn_shutdown() waits for these to be emptied so they 701 * must be initialized before it can be called. 702 */ 703 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); 704 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); 705 706 ic->conn = conn; 707 conn->c_transport_data = ic; 708 709 spin_lock_irqsave(&ib_nodev_conns_lock, flags); 710 list_add_tail(&ic->ib_node, &ib_nodev_conns); 711 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); 712 713 714 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); 715 return 0; 716} 717 718/* 719 * Free a connection. Connection must be shut down and not set for reconnect. 720 */ 721void rds_ib_conn_free(void *arg) 722{ 723 struct rds_ib_connection *ic = arg; 724 spinlock_t *lock_ptr; 725 726 rdsdebug("ic %p\n", ic); 727 728 /* 729 * Conn is either on a dev's list or on the nodev list. 730 * A race with shutdown() or connect() would cause problems 731 * (since rds_ibdev would change) but that should never happen. 732 */ 733 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; 734 735 spin_lock_irq(lock_ptr); 736 list_del(&ic->ib_node); 737 spin_unlock_irq(lock_ptr); 738 739 kfree(ic); 740} 741 742 743/* 744 * An error occurred on the connection 745 */ 746void 747__rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) 748{ 749 va_list ap; 750 751 rds_conn_drop(conn); 752 753 va_start(ap, fmt); 754 vprintk(fmt, ap); 755 va_end(ap); 756} 757