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 #include <linux/ratelimit.h> 38 39 #include "rds.h" 40 #include "ib.h" 41 42 /* 43 * Set the selected protocol version 44 */ 45 static void rds_ib_set_protocol(struct rds_connection *conn, unsigned int version) 46 { 47 conn->c_version = version; 48 } 49 50 /* 51 * Set up flow control 52 */ 53 static void rds_ib_set_flow_control(struct rds_connection *conn, u32 credits) 54 { 55 struct rds_ib_connection *ic = conn->c_transport_data; 56 57 if (rds_ib_sysctl_flow_control && credits != 0) { 58 /* We're doing flow control */ 59 ic->i_flowctl = 1; 60 rds_ib_send_add_credits(conn, credits); 61 } else { 62 ic->i_flowctl = 0; 63 } 64 } 65 66 /* 67 * Tune RNR behavior. Without flow control, we use a rather 68 * low timeout, but not the absolute minimum - this should 69 * be tunable. 70 * 71 * We already set the RNR retry count to 7 (which is the 72 * smallest infinite number :-) above. 73 * If flow control is off, we want to change this back to 0 74 * so that we learn quickly when our credit accounting is 75 * buggy. 76 * 77 * Caller passes in a qp_attr pointer - don't waste stack spacv 78 * by allocation this twice. 79 */ 80 static void 81 rds_ib_tune_rnr(struct rds_ib_connection *ic, struct ib_qp_attr *attr) 82 { 83 int ret; 84 85 attr->min_rnr_timer = IB_RNR_TIMER_000_32; 86 ret = ib_modify_qp(ic->i_cm_id->qp, attr, IB_QP_MIN_RNR_TIMER); 87 if (ret) 88 printk(KERN_NOTICE "ib_modify_qp(IB_QP_MIN_RNR_TIMER): err=%d\n", -ret); 89 } 90 91 /* 92 * Connection established. 93 * We get here for both outgoing and incoming connection. 94 */ 95 void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_event *event) 96 { 97 const struct rds_ib_connect_private *dp = NULL; 98 struct rds_ib_connection *ic = conn->c_transport_data; 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 if (conn->c_version < RDS_PROTOCOL(3,1)) { 115 printk(KERN_NOTICE "RDS/IB: Connection to %pI4 version %u.%u failed," 116 " no longer supported\n", 117 &conn->c_faddr, 118 RDS_PROTOCOL_MAJOR(conn->c_version), 119 RDS_PROTOCOL_MINOR(conn->c_version)); 120 rds_conn_destroy(conn); 121 return; 122 } else { 123 printk(KERN_NOTICE "RDS/IB: connected to %pI4 version %u.%u%s\n", 124 &conn->c_faddr, 125 RDS_PROTOCOL_MAJOR(conn->c_version), 126 RDS_PROTOCOL_MINOR(conn->c_version), 127 ic->i_flowctl ? ", flow control" : ""); 128 } 129 130 /* 131 * Init rings and fill recv. this needs to wait until protocol negotiation 132 * is complete, since ring layout is different from 3.0 to 3.1. 133 */ 134 rds_ib_send_init_ring(ic); 135 rds_ib_recv_init_ring(ic); 136 /* Post receive buffers - as a side effect, this will update 137 * the posted credit count. */ 138 rds_ib_recv_refill(conn, 1, GFP_KERNEL); 139 140 /* Tune RNR behavior */ 141 rds_ib_tune_rnr(ic, &qp_attr); 142 143 qp_attr.qp_state = IB_QPS_RTS; 144 err = ib_modify_qp(ic->i_cm_id->qp, &qp_attr, IB_QP_STATE); 145 if (err) 146 printk(KERN_NOTICE "ib_modify_qp(IB_QP_STATE, RTS): err=%d\n", err); 147 148 /* update ib_device with this local ipaddr */ 149 err = rds_ib_update_ipaddr(ic->rds_ibdev, conn->c_laddr); 150 if (err) 151 printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", 152 err); 153 154 /* If the peer gave us the last packet it saw, process this as if 155 * we had received a regular ACK. */ 156 if (dp) { 157 /* dp structure start is not guaranteed to be 8 bytes aligned. 158 * Since dp_ack_seq is 64-bit extended load operations can be 159 * used so go through get_unaligned to avoid unaligned errors. 160 */ 161 __be64 dp_ack_seq = get_unaligned(&dp->dp_ack_seq); 162 163 if (dp_ack_seq) 164 rds_send_drop_acked(conn, be64_to_cpu(dp_ack_seq), 165 NULL); 166 } 167 168 rds_connect_complete(conn); 169 } 170 171 static void rds_ib_cm_fill_conn_param(struct rds_connection *conn, 172 struct rdma_conn_param *conn_param, 173 struct rds_ib_connect_private *dp, 174 u32 protocol_version, 175 u32 max_responder_resources, 176 u32 max_initiator_depth) 177 { 178 struct rds_ib_connection *ic = conn->c_transport_data; 179 struct rds_ib_device *rds_ibdev = ic->rds_ibdev; 180 181 memset(conn_param, 0, sizeof(struct rdma_conn_param)); 182 183 conn_param->responder_resources = 184 min_t(u32, rds_ibdev->max_responder_resources, max_responder_resources); 185 conn_param->initiator_depth = 186 min_t(u32, rds_ibdev->max_initiator_depth, max_initiator_depth); 187 conn_param->retry_count = min_t(unsigned int, rds_ib_retry_count, 7); 188 conn_param->rnr_retry_count = 7; 189 190 if (dp) { 191 memset(dp, 0, sizeof(*dp)); 192 dp->dp_saddr = conn->c_laddr; 193 dp->dp_daddr = conn->c_faddr; 194 dp->dp_protocol_major = RDS_PROTOCOL_MAJOR(protocol_version); 195 dp->dp_protocol_minor = RDS_PROTOCOL_MINOR(protocol_version); 196 dp->dp_protocol_minor_mask = cpu_to_be16(RDS_IB_SUPPORTED_PROTOCOLS); 197 dp->dp_ack_seq = cpu_to_be64(rds_ib_piggyb_ack(ic)); 198 199 /* Advertise flow control */ 200 if (ic->i_flowctl) { 201 unsigned int credits; 202 203 credits = IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)); 204 dp->dp_credit = cpu_to_be32(credits); 205 atomic_sub(IB_SET_POST_CREDITS(credits), &ic->i_credits); 206 } 207 208 conn_param->private_data = dp; 209 conn_param->private_data_len = sizeof(*dp); 210 } 211 } 212 213 static void rds_ib_cq_event_handler(struct ib_event *event, void *data) 214 { 215 rdsdebug("event %u (%s) data %p\n", 216 event->event, ib_event_msg(event->event), data); 217 } 218 219 /* Plucking the oldest entry from the ring can be done concurrently with 220 * the thread refilling the ring. Each ring operation is protected by 221 * spinlocks and the transient state of refilling doesn't change the 222 * recording of which entry is oldest. 223 * 224 * This relies on IB only calling one cq comp_handler for each cq so that 225 * there will only be one caller of rds_recv_incoming() per RDS connection. 226 */ 227 static void rds_ib_cq_comp_handler_recv(struct ib_cq *cq, void *context) 228 { 229 struct rds_connection *conn = context; 230 struct rds_ib_connection *ic = conn->c_transport_data; 231 232 rdsdebug("conn %p cq %p\n", conn, cq); 233 234 rds_ib_stats_inc(s_ib_evt_handler_call); 235 236 tasklet_schedule(&ic->i_recv_tasklet); 237 } 238 239 static void poll_scq(struct rds_ib_connection *ic, struct ib_cq *cq, 240 struct ib_wc *wcs) 241 { 242 int nr, i; 243 struct ib_wc *wc; 244 245 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { 246 for (i = 0; i < nr; i++) { 247 wc = wcs + i; 248 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 249 (unsigned long long)wc->wr_id, wc->status, 250 wc->byte_len, be32_to_cpu(wc->ex.imm_data)); 251 252 if (wc->wr_id <= ic->i_send_ring.w_nr || 253 wc->wr_id == RDS_IB_ACK_WR_ID) 254 rds_ib_send_cqe_handler(ic, wc); 255 else 256 rds_ib_mr_cqe_handler(ic, wc); 257 258 } 259 } 260 } 261 262 static void rds_ib_tasklet_fn_send(unsigned long data) 263 { 264 struct rds_ib_connection *ic = (struct rds_ib_connection *)data; 265 struct rds_connection *conn = ic->conn; 266 267 rds_ib_stats_inc(s_ib_tasklet_call); 268 269 poll_scq(ic, ic->i_send_cq, ic->i_send_wc); 270 ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); 271 poll_scq(ic, ic->i_send_cq, ic->i_send_wc); 272 273 if (rds_conn_up(conn) && 274 (!test_bit(RDS_LL_SEND_FULL, &conn->c_flags) || 275 test_bit(0, &conn->c_map_queued))) 276 rds_send_xmit(ic->conn); 277 } 278 279 static void poll_rcq(struct rds_ib_connection *ic, struct ib_cq *cq, 280 struct ib_wc *wcs, 281 struct rds_ib_ack_state *ack_state) 282 { 283 int nr, i; 284 struct ib_wc *wc; 285 286 while ((nr = ib_poll_cq(cq, RDS_IB_WC_MAX, wcs)) > 0) { 287 for (i = 0; i < nr; i++) { 288 wc = wcs + i; 289 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n", 290 (unsigned long long)wc->wr_id, wc->status, 291 wc->byte_len, be32_to_cpu(wc->ex.imm_data)); 292 293 rds_ib_recv_cqe_handler(ic, wc, ack_state); 294 } 295 } 296 } 297 298 static void rds_ib_tasklet_fn_recv(unsigned long data) 299 { 300 struct rds_ib_connection *ic = (struct rds_ib_connection *)data; 301 struct rds_connection *conn = ic->conn; 302 struct rds_ib_device *rds_ibdev = ic->rds_ibdev; 303 struct rds_ib_ack_state state; 304 305 if (!rds_ibdev) 306 rds_conn_drop(conn); 307 308 rds_ib_stats_inc(s_ib_tasklet_call); 309 310 memset(&state, 0, sizeof(state)); 311 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); 312 ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 313 poll_rcq(ic, ic->i_recv_cq, ic->i_recv_wc, &state); 314 315 if (state.ack_next_valid) 316 rds_ib_set_ack(ic, state.ack_next, state.ack_required); 317 if (state.ack_recv_valid && state.ack_recv > ic->i_ack_recv) { 318 rds_send_drop_acked(conn, state.ack_recv, NULL); 319 ic->i_ack_recv = state.ack_recv; 320 } 321 322 if (rds_conn_up(conn)) 323 rds_ib_attempt_ack(ic); 324 } 325 326 static void rds_ib_qp_event_handler(struct ib_event *event, void *data) 327 { 328 struct rds_connection *conn = data; 329 struct rds_ib_connection *ic = conn->c_transport_data; 330 331 rdsdebug("conn %p ic %p event %u (%s)\n", conn, ic, event->event, 332 ib_event_msg(event->event)); 333 334 switch (event->event) { 335 case IB_EVENT_COMM_EST: 336 rdma_notify(ic->i_cm_id, IB_EVENT_COMM_EST); 337 break; 338 default: 339 rdsdebug("Fatal QP Event %u (%s) " 340 "- connection %pI4->%pI4, reconnecting\n", 341 event->event, ib_event_msg(event->event), 342 &conn->c_laddr, &conn->c_faddr); 343 rds_conn_drop(conn); 344 break; 345 } 346 } 347 348 static void rds_ib_cq_comp_handler_send(struct ib_cq *cq, void *context) 349 { 350 struct rds_connection *conn = context; 351 struct rds_ib_connection *ic = conn->c_transport_data; 352 353 rdsdebug("conn %p cq %p\n", conn, cq); 354 355 rds_ib_stats_inc(s_ib_evt_handler_call); 356 357 tasklet_schedule(&ic->i_send_tasklet); 358 } 359 360 /* 361 * This needs to be very careful to not leave IS_ERR pointers around for 362 * cleanup to trip over. 363 */ 364 static int rds_ib_setup_qp(struct rds_connection *conn) 365 { 366 struct rds_ib_connection *ic = conn->c_transport_data; 367 struct ib_device *dev = ic->i_cm_id->device; 368 struct ib_qp_init_attr attr; 369 struct ib_cq_init_attr cq_attr = {}; 370 struct rds_ib_device *rds_ibdev; 371 int ret, fr_queue_space; 372 373 /* 374 * It's normal to see a null device if an incoming connection races 375 * with device removal, so we don't print a warning. 376 */ 377 rds_ibdev = rds_ib_get_client_data(dev); 378 if (!rds_ibdev) 379 return -EOPNOTSUPP; 380 381 /* The fr_queue_space is currently set to 512, to add extra space on 382 * completion queue and send queue. This extra space is used for FRMR 383 * registration and invalidation work requests 384 */ 385 fr_queue_space = (rds_ibdev->use_fastreg ? RDS_IB_DEFAULT_FR_WR : 0); 386 387 /* add the conn now so that connection establishment has the dev */ 388 rds_ib_add_conn(rds_ibdev, conn); 389 390 if (rds_ibdev->max_wrs < ic->i_send_ring.w_nr + 1) 391 rds_ib_ring_resize(&ic->i_send_ring, rds_ibdev->max_wrs - 1); 392 if (rds_ibdev->max_wrs < ic->i_recv_ring.w_nr + 1) 393 rds_ib_ring_resize(&ic->i_recv_ring, rds_ibdev->max_wrs - 1); 394 395 /* Protection domain and memory range */ 396 ic->i_pd = rds_ibdev->pd; 397 398 cq_attr.cqe = ic->i_send_ring.w_nr + fr_queue_space + 1; 399 400 ic->i_send_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_send, 401 rds_ib_cq_event_handler, conn, 402 &cq_attr); 403 if (IS_ERR(ic->i_send_cq)) { 404 ret = PTR_ERR(ic->i_send_cq); 405 ic->i_send_cq = NULL; 406 rdsdebug("ib_create_cq send failed: %d\n", ret); 407 goto out; 408 } 409 410 cq_attr.cqe = ic->i_recv_ring.w_nr; 411 ic->i_recv_cq = ib_create_cq(dev, rds_ib_cq_comp_handler_recv, 412 rds_ib_cq_event_handler, conn, 413 &cq_attr); 414 if (IS_ERR(ic->i_recv_cq)) { 415 ret = PTR_ERR(ic->i_recv_cq); 416 ic->i_recv_cq = NULL; 417 rdsdebug("ib_create_cq recv failed: %d\n", ret); 418 goto out; 419 } 420 421 ret = ib_req_notify_cq(ic->i_send_cq, IB_CQ_NEXT_COMP); 422 if (ret) { 423 rdsdebug("ib_req_notify_cq send failed: %d\n", ret); 424 goto out; 425 } 426 427 ret = ib_req_notify_cq(ic->i_recv_cq, IB_CQ_SOLICITED); 428 if (ret) { 429 rdsdebug("ib_req_notify_cq recv failed: %d\n", ret); 430 goto out; 431 } 432 433 /* XXX negotiate max send/recv with remote? */ 434 memset(&attr, 0, sizeof(attr)); 435 attr.event_handler = rds_ib_qp_event_handler; 436 attr.qp_context = conn; 437 /* + 1 to allow for the single ack message */ 438 attr.cap.max_send_wr = ic->i_send_ring.w_nr + fr_queue_space + 1; 439 attr.cap.max_recv_wr = ic->i_recv_ring.w_nr + 1; 440 attr.cap.max_send_sge = rds_ibdev->max_sge; 441 attr.cap.max_recv_sge = RDS_IB_RECV_SGE; 442 attr.sq_sig_type = IB_SIGNAL_REQ_WR; 443 attr.qp_type = IB_QPT_RC; 444 attr.send_cq = ic->i_send_cq; 445 attr.recv_cq = ic->i_recv_cq; 446 atomic_set(&ic->i_fastreg_wrs, RDS_IB_DEFAULT_FR_WR); 447 448 /* 449 * XXX this can fail if max_*_wr is too large? Are we supposed 450 * to back off until we get a value that the hardware can support? 451 */ 452 ret = rdma_create_qp(ic->i_cm_id, ic->i_pd, &attr); 453 if (ret) { 454 rdsdebug("rdma_create_qp failed: %d\n", ret); 455 goto out; 456 } 457 458 ic->i_send_hdrs = ib_dma_alloc_coherent(dev, 459 ic->i_send_ring.w_nr * 460 sizeof(struct rds_header), 461 &ic->i_send_hdrs_dma, GFP_KERNEL); 462 if (!ic->i_send_hdrs) { 463 ret = -ENOMEM; 464 rdsdebug("ib_dma_alloc_coherent send failed\n"); 465 goto out; 466 } 467 468 ic->i_recv_hdrs = ib_dma_alloc_coherent(dev, 469 ic->i_recv_ring.w_nr * 470 sizeof(struct rds_header), 471 &ic->i_recv_hdrs_dma, GFP_KERNEL); 472 if (!ic->i_recv_hdrs) { 473 ret = -ENOMEM; 474 rdsdebug("ib_dma_alloc_coherent recv failed\n"); 475 goto out; 476 } 477 478 ic->i_ack = ib_dma_alloc_coherent(dev, sizeof(struct rds_header), 479 &ic->i_ack_dma, GFP_KERNEL); 480 if (!ic->i_ack) { 481 ret = -ENOMEM; 482 rdsdebug("ib_dma_alloc_coherent ack failed\n"); 483 goto out; 484 } 485 486 ic->i_sends = vzalloc_node(ic->i_send_ring.w_nr * sizeof(struct rds_ib_send_work), 487 ibdev_to_node(dev)); 488 if (!ic->i_sends) { 489 ret = -ENOMEM; 490 rdsdebug("send allocation failed\n"); 491 goto out; 492 } 493 494 ic->i_recvs = vzalloc_node(ic->i_recv_ring.w_nr * sizeof(struct rds_ib_recv_work), 495 ibdev_to_node(dev)); 496 if (!ic->i_recvs) { 497 ret = -ENOMEM; 498 rdsdebug("recv allocation failed\n"); 499 goto out; 500 } 501 502 rds_ib_recv_init_ack(ic); 503 504 rdsdebug("conn %p pd %p cq %p %p\n", conn, ic->i_pd, 505 ic->i_send_cq, ic->i_recv_cq); 506 507 out: 508 rds_ib_dev_put(rds_ibdev); 509 return ret; 510 } 511 512 static u32 rds_ib_protocol_compatible(struct rdma_cm_event *event) 513 { 514 const struct rds_ib_connect_private *dp = event->param.conn.private_data; 515 u16 common; 516 u32 version = 0; 517 518 /* 519 * rdma_cm private data is odd - when there is any private data in the 520 * request, we will be given a pretty large buffer without telling us the 521 * original size. The only way to tell the difference is by looking at 522 * the contents, which are initialized to zero. 523 * If the protocol version fields aren't set, this is a connection attempt 524 * from an older version. This could could be 3.0 or 2.0 - we can't tell. 525 * We really should have changed this for OFED 1.3 :-( 526 */ 527 528 /* Be paranoid. RDS always has privdata */ 529 if (!event->param.conn.private_data_len) { 530 printk(KERN_NOTICE "RDS incoming connection has no private data, " 531 "rejecting\n"); 532 return 0; 533 } 534 535 /* Even if len is crap *now* I still want to check it. -ASG */ 536 if (event->param.conn.private_data_len < sizeof (*dp) || 537 dp->dp_protocol_major == 0) 538 return RDS_PROTOCOL_3_0; 539 540 common = be16_to_cpu(dp->dp_protocol_minor_mask) & RDS_IB_SUPPORTED_PROTOCOLS; 541 if (dp->dp_protocol_major == 3 && common) { 542 version = RDS_PROTOCOL_3_0; 543 while ((common >>= 1) != 0) 544 version++; 545 } else 546 printk_ratelimited(KERN_NOTICE "RDS: Connection from %pI4 using incompatible protocol version %u.%u\n", 547 &dp->dp_saddr, 548 dp->dp_protocol_major, 549 dp->dp_protocol_minor); 550 return version; 551 } 552 553 int rds_ib_cm_handle_connect(struct rdma_cm_id *cm_id, 554 struct rdma_cm_event *event) 555 { 556 __be64 lguid = cm_id->route.path_rec->sgid.global.interface_id; 557 __be64 fguid = cm_id->route.path_rec->dgid.global.interface_id; 558 const struct rds_ib_connect_private *dp = event->param.conn.private_data; 559 struct rds_ib_connect_private dp_rep; 560 struct rds_connection *conn = NULL; 561 struct rds_ib_connection *ic = NULL; 562 struct rdma_conn_param conn_param; 563 u32 version; 564 int err = 1, destroy = 1; 565 566 /* Check whether the remote protocol version matches ours. */ 567 version = rds_ib_protocol_compatible(event); 568 if (!version) 569 goto out; 570 571 rdsdebug("saddr %pI4 daddr %pI4 RDSv%u.%u lguid 0x%llx fguid " 572 "0x%llx\n", &dp->dp_saddr, &dp->dp_daddr, 573 RDS_PROTOCOL_MAJOR(version), RDS_PROTOCOL_MINOR(version), 574 (unsigned long long)be64_to_cpu(lguid), 575 (unsigned long long)be64_to_cpu(fguid)); 576 577 /* RDS/IB is not currently netns aware, thus init_net */ 578 conn = rds_conn_create(&init_net, dp->dp_daddr, dp->dp_saddr, 579 &rds_ib_transport, GFP_KERNEL); 580 if (IS_ERR(conn)) { 581 rdsdebug("rds_conn_create failed (%ld)\n", PTR_ERR(conn)); 582 conn = NULL; 583 goto out; 584 } 585 586 /* 587 * The connection request may occur while the 588 * previous connection exist, e.g. in case of failover. 589 * But as connections may be initiated simultaneously 590 * by both hosts, we have a random backoff mechanism - 591 * see the comment above rds_queue_reconnect() 592 */ 593 mutex_lock(&conn->c_cm_lock); 594 if (!rds_conn_transition(conn, RDS_CONN_DOWN, RDS_CONN_CONNECTING)) { 595 if (rds_conn_state(conn) == RDS_CONN_UP) { 596 rdsdebug("incoming connect while connecting\n"); 597 rds_conn_drop(conn); 598 rds_ib_stats_inc(s_ib_listen_closed_stale); 599 } else 600 if (rds_conn_state(conn) == RDS_CONN_CONNECTING) { 601 /* Wait and see - our connect may still be succeeding */ 602 rds_ib_stats_inc(s_ib_connect_raced); 603 } 604 goto out; 605 } 606 607 ic = conn->c_transport_data; 608 609 rds_ib_set_protocol(conn, version); 610 rds_ib_set_flow_control(conn, be32_to_cpu(dp->dp_credit)); 611 612 /* If the peer gave us the last packet it saw, process this as if 613 * we had received a regular ACK. */ 614 if (dp->dp_ack_seq) 615 rds_send_drop_acked(conn, be64_to_cpu(dp->dp_ack_seq), NULL); 616 617 BUG_ON(cm_id->context); 618 BUG_ON(ic->i_cm_id); 619 620 ic->i_cm_id = cm_id; 621 cm_id->context = conn; 622 623 /* We got halfway through setting up the ib_connection, if we 624 * fail now, we have to take the long route out of this mess. */ 625 destroy = 0; 626 627 err = rds_ib_setup_qp(conn); 628 if (err) { 629 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", err); 630 goto out; 631 } 632 633 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp_rep, version, 634 event->param.conn.responder_resources, 635 event->param.conn.initiator_depth); 636 637 /* rdma_accept() calls rdma_reject() internally if it fails */ 638 err = rdma_accept(cm_id, &conn_param); 639 if (err) 640 rds_ib_conn_error(conn, "rdma_accept failed (%d)\n", err); 641 642 out: 643 if (conn) 644 mutex_unlock(&conn->c_cm_lock); 645 if (err) 646 rdma_reject(cm_id, NULL, 0); 647 return destroy; 648 } 649 650 651 int rds_ib_cm_initiate_connect(struct rdma_cm_id *cm_id) 652 { 653 struct rds_connection *conn = cm_id->context; 654 struct rds_ib_connection *ic = conn->c_transport_data; 655 struct rdma_conn_param conn_param; 656 struct rds_ib_connect_private dp; 657 int ret; 658 659 /* If the peer doesn't do protocol negotiation, we must 660 * default to RDSv3.0 */ 661 rds_ib_set_protocol(conn, RDS_PROTOCOL_3_0); 662 ic->i_flowctl = rds_ib_sysctl_flow_control; /* advertise flow control */ 663 664 ret = rds_ib_setup_qp(conn); 665 if (ret) { 666 rds_ib_conn_error(conn, "rds_ib_setup_qp failed (%d)\n", ret); 667 goto out; 668 } 669 670 rds_ib_cm_fill_conn_param(conn, &conn_param, &dp, RDS_PROTOCOL_VERSION, 671 UINT_MAX, UINT_MAX); 672 ret = rdma_connect(cm_id, &conn_param); 673 if (ret) 674 rds_ib_conn_error(conn, "rdma_connect failed (%d)\n", ret); 675 676 out: 677 /* Beware - returning non-zero tells the rdma_cm to destroy 678 * the cm_id. We should certainly not do it as long as we still 679 * "own" the cm_id. */ 680 if (ret) { 681 if (ic->i_cm_id == cm_id) 682 ret = 0; 683 } 684 return ret; 685 } 686 687 int rds_ib_conn_connect(struct rds_connection *conn) 688 { 689 struct rds_ib_connection *ic = conn->c_transport_data; 690 struct sockaddr_in src, dest; 691 int ret; 692 693 /* XXX I wonder what affect the port space has */ 694 /* delegate cm event handler to rdma_transport */ 695 ic->i_cm_id = rdma_create_id(&init_net, rds_rdma_cm_event_handler, conn, 696 RDMA_PS_TCP, IB_QPT_RC); 697 if (IS_ERR(ic->i_cm_id)) { 698 ret = PTR_ERR(ic->i_cm_id); 699 ic->i_cm_id = NULL; 700 rdsdebug("rdma_create_id() failed: %d\n", ret); 701 goto out; 702 } 703 704 rdsdebug("created cm id %p for conn %p\n", ic->i_cm_id, conn); 705 706 src.sin_family = AF_INET; 707 src.sin_addr.s_addr = (__force u32)conn->c_laddr; 708 src.sin_port = (__force u16)htons(0); 709 710 dest.sin_family = AF_INET; 711 dest.sin_addr.s_addr = (__force u32)conn->c_faddr; 712 dest.sin_port = (__force u16)htons(RDS_PORT); 713 714 ret = rdma_resolve_addr(ic->i_cm_id, (struct sockaddr *)&src, 715 (struct sockaddr *)&dest, 716 RDS_RDMA_RESOLVE_TIMEOUT_MS); 717 if (ret) { 718 rdsdebug("addr resolve failed for cm id %p: %d\n", ic->i_cm_id, 719 ret); 720 rdma_destroy_id(ic->i_cm_id); 721 ic->i_cm_id = NULL; 722 } 723 724 out: 725 return ret; 726 } 727 728 /* 729 * This is so careful about only cleaning up resources that were built up 730 * so that it can be called at any point during startup. In fact it 731 * can be called multiple times for a given connection. 732 */ 733 void rds_ib_conn_shutdown(struct rds_connection *conn) 734 { 735 struct rds_ib_connection *ic = conn->c_transport_data; 736 int err = 0; 737 738 rdsdebug("cm %p pd %p cq %p %p qp %p\n", ic->i_cm_id, 739 ic->i_pd, ic->i_send_cq, ic->i_recv_cq, 740 ic->i_cm_id ? ic->i_cm_id->qp : NULL); 741 742 if (ic->i_cm_id) { 743 struct ib_device *dev = ic->i_cm_id->device; 744 745 rdsdebug("disconnecting cm %p\n", ic->i_cm_id); 746 err = rdma_disconnect(ic->i_cm_id); 747 if (err) { 748 /* Actually this may happen quite frequently, when 749 * an outgoing connect raced with an incoming connect. 750 */ 751 rdsdebug("failed to disconnect, cm: %p err %d\n", 752 ic->i_cm_id, err); 753 } 754 755 /* 756 * We want to wait for tx and rx completion to finish 757 * before we tear down the connection, but we have to be 758 * careful not to get stuck waiting on a send ring that 759 * only has unsignaled sends in it. We've shutdown new 760 * sends before getting here so by waiting for signaled 761 * sends to complete we're ensured that there will be no 762 * more tx processing. 763 */ 764 wait_event(rds_ib_ring_empty_wait, 765 rds_ib_ring_empty(&ic->i_recv_ring) && 766 (atomic_read(&ic->i_signaled_sends) == 0) && 767 (atomic_read(&ic->i_fastreg_wrs) == RDS_IB_DEFAULT_FR_WR)); 768 tasklet_kill(&ic->i_send_tasklet); 769 tasklet_kill(&ic->i_recv_tasklet); 770 771 /* first destroy the ib state that generates callbacks */ 772 if (ic->i_cm_id->qp) 773 rdma_destroy_qp(ic->i_cm_id); 774 if (ic->i_send_cq) 775 ib_destroy_cq(ic->i_send_cq); 776 if (ic->i_recv_cq) 777 ib_destroy_cq(ic->i_recv_cq); 778 779 /* then free the resources that ib callbacks use */ 780 if (ic->i_send_hdrs) 781 ib_dma_free_coherent(dev, 782 ic->i_send_ring.w_nr * 783 sizeof(struct rds_header), 784 ic->i_send_hdrs, 785 ic->i_send_hdrs_dma); 786 787 if (ic->i_recv_hdrs) 788 ib_dma_free_coherent(dev, 789 ic->i_recv_ring.w_nr * 790 sizeof(struct rds_header), 791 ic->i_recv_hdrs, 792 ic->i_recv_hdrs_dma); 793 794 if (ic->i_ack) 795 ib_dma_free_coherent(dev, sizeof(struct rds_header), 796 ic->i_ack, ic->i_ack_dma); 797 798 if (ic->i_sends) 799 rds_ib_send_clear_ring(ic); 800 if (ic->i_recvs) 801 rds_ib_recv_clear_ring(ic); 802 803 rdma_destroy_id(ic->i_cm_id); 804 805 /* 806 * Move connection back to the nodev list. 807 */ 808 if (ic->rds_ibdev) 809 rds_ib_remove_conn(ic->rds_ibdev, conn); 810 811 ic->i_cm_id = NULL; 812 ic->i_pd = NULL; 813 ic->i_send_cq = NULL; 814 ic->i_recv_cq = NULL; 815 ic->i_send_hdrs = NULL; 816 ic->i_recv_hdrs = NULL; 817 ic->i_ack = NULL; 818 } 819 BUG_ON(ic->rds_ibdev); 820 821 /* Clear pending transmit */ 822 if (ic->i_data_op) { 823 struct rds_message *rm; 824 825 rm = container_of(ic->i_data_op, struct rds_message, data); 826 rds_message_put(rm); 827 ic->i_data_op = NULL; 828 } 829 830 /* Clear the ACK state */ 831 clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags); 832 #ifdef KERNEL_HAS_ATOMIC64 833 atomic64_set(&ic->i_ack_next, 0); 834 #else 835 ic->i_ack_next = 0; 836 #endif 837 ic->i_ack_recv = 0; 838 839 /* Clear flow control state */ 840 ic->i_flowctl = 0; 841 atomic_set(&ic->i_credits, 0); 842 843 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); 844 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); 845 846 if (ic->i_ibinc) { 847 rds_inc_put(&ic->i_ibinc->ii_inc); 848 ic->i_ibinc = NULL; 849 } 850 851 vfree(ic->i_sends); 852 ic->i_sends = NULL; 853 vfree(ic->i_recvs); 854 ic->i_recvs = NULL; 855 } 856 857 int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp) 858 { 859 struct rds_ib_connection *ic; 860 unsigned long flags; 861 int ret; 862 863 /* XXX too lazy? */ 864 ic = kzalloc(sizeof(struct rds_ib_connection), gfp); 865 if (!ic) 866 return -ENOMEM; 867 868 ret = rds_ib_recv_alloc_caches(ic); 869 if (ret) { 870 kfree(ic); 871 return ret; 872 } 873 874 INIT_LIST_HEAD(&ic->ib_node); 875 tasklet_init(&ic->i_send_tasklet, rds_ib_tasklet_fn_send, 876 (unsigned long)ic); 877 tasklet_init(&ic->i_recv_tasklet, rds_ib_tasklet_fn_recv, 878 (unsigned long)ic); 879 mutex_init(&ic->i_recv_mutex); 880 #ifndef KERNEL_HAS_ATOMIC64 881 spin_lock_init(&ic->i_ack_lock); 882 #endif 883 atomic_set(&ic->i_signaled_sends, 0); 884 885 /* 886 * rds_ib_conn_shutdown() waits for these to be emptied so they 887 * must be initialized before it can be called. 888 */ 889 rds_ib_ring_init(&ic->i_send_ring, rds_ib_sysctl_max_send_wr); 890 rds_ib_ring_init(&ic->i_recv_ring, rds_ib_sysctl_max_recv_wr); 891 892 ic->conn = conn; 893 conn->c_transport_data = ic; 894 895 spin_lock_irqsave(&ib_nodev_conns_lock, flags); 896 list_add_tail(&ic->ib_node, &ib_nodev_conns); 897 spin_unlock_irqrestore(&ib_nodev_conns_lock, flags); 898 899 900 rdsdebug("conn %p conn ic %p\n", conn, conn->c_transport_data); 901 return 0; 902 } 903 904 /* 905 * Free a connection. Connection must be shut down and not set for reconnect. 906 */ 907 void rds_ib_conn_free(void *arg) 908 { 909 struct rds_ib_connection *ic = arg; 910 spinlock_t *lock_ptr; 911 912 rdsdebug("ic %p\n", ic); 913 914 /* 915 * Conn is either on a dev's list or on the nodev list. 916 * A race with shutdown() or connect() would cause problems 917 * (since rds_ibdev would change) but that should never happen. 918 */ 919 lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock; 920 921 spin_lock_irq(lock_ptr); 922 list_del(&ic->ib_node); 923 spin_unlock_irq(lock_ptr); 924 925 rds_ib_recv_free_caches(ic); 926 927 kfree(ic); 928 } 929 930 931 /* 932 * An error occurred on the connection 933 */ 934 void 935 __rds_ib_conn_error(struct rds_connection *conn, const char *fmt, ...) 936 { 937 va_list ap; 938 939 rds_conn_drop(conn); 940 941 va_start(ap, fmt); 942 vprintk(fmt, ap); 943 va_end(ap); 944 } 945