1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0 3 * 4 * Copyright (c) 2006 Mellanox Technologies. All rights reserved 5 * 6 * This software is available to you under a choice of one of two 7 * licenses. You may choose to be licensed under the terms of the GNU 8 * General Public License (GPL) Version 2, available from the file 9 * COPYING in the main directory of this source tree, or the 10 * OpenIB.org BSD license below: 11 * 12 * Redistribution and use in source and binary forms, with or 13 * without modification, are permitted provided that the following 14 * conditions are met: 15 * 16 * - Redistributions of source code must retain the above 17 * copyright notice, this list of conditions and the following 18 * disclaimer. 19 * 20 * - Redistributions in binary form must reproduce the above 21 * copyright notice, this list of conditions and the following 22 * disclaimer in the documentation and/or other materials 23 * provided with the distribution. 24 * 25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 32 * SOFTWARE. 33 */ 34 35 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include "ipoib.h" 39 40 #ifdef CONFIG_INFINIBAND_IPOIB_CM 41 42 #include <netinet/ip.h> 43 #include <netinet/ip_icmp.h> 44 #include <netinet/icmp6.h> 45 46 #include <rdma/ib_cm.h> 47 #include <rdma/ib_cache.h> 48 #include <linux/delay.h> 49 50 int ipoib_max_conn_qp = 128; 51 52 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444); 53 MODULE_PARM_DESC(max_nonsrq_conn_qp, 54 "Max number of connected-mode QPs per interface " 55 "(applied only if shared receive queue is not available)"); 56 57 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA 58 static int data_debug_level; 59 60 module_param_named(cm_data_debug_level, data_debug_level, int, 0644); 61 MODULE_PARM_DESC(cm_data_debug_level, 62 "Enable data path debug tracing for connected mode if > 0"); 63 #endif 64 65 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL 66 67 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ) 68 #define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ) 69 #define IPOIB_CM_RX_DELAY (3 * 256 * HZ) 70 #define IPOIB_CM_RX_UPDATE_MASK (0x3) 71 72 static struct ib_qp_attr ipoib_cm_err_attr = { 73 .qp_state = IB_QPS_ERR 74 }; 75 76 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff 77 78 static struct ib_send_wr ipoib_cm_rx_drain_wr = { 79 .wr_id = IPOIB_CM_RX_DRAIN_WRID, 80 .opcode = IB_WR_SEND, 81 }; 82 83 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id, 84 struct ib_cm_event *event); 85 86 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req) 87 { 88 89 ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req); 90 91 } 92 93 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id) 94 { 95 struct ib_recv_wr *bad_wr; 96 struct ipoib_rx_buf *rx_req; 97 struct mbuf *m; 98 int ret; 99 int i; 100 101 rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id]; 102 for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) { 103 priv->cm.rx_sge[i].addr = rx_req->mapping[i]; 104 priv->cm.rx_sge[i].length = m->m_len; 105 } 106 107 priv->cm.rx_wr.num_sge = i; 108 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV; 109 110 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr); 111 if (unlikely(ret)) { 112 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret); 113 ipoib_dma_unmap_rx(priv, rx_req); 114 m_freem(priv->cm.srq_ring[id].mb); 115 priv->cm.srq_ring[id].mb = NULL; 116 } 117 118 return ret; 119 } 120 121 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv, 122 struct ipoib_cm_rx *rx, 123 struct ib_recv_wr *wr, 124 struct ib_sge *sge, int id) 125 { 126 struct ipoib_rx_buf *rx_req; 127 struct ib_recv_wr *bad_wr; 128 struct mbuf *m; 129 int ret; 130 int i; 131 132 rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id]; 133 for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) { 134 sge[i].addr = rx_req->mapping[i]; 135 sge[i].length = m->m_len; 136 } 137 138 wr->num_sge = i; 139 wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV; 140 141 ret = ib_post_recv(rx->qp, wr, &bad_wr); 142 if (unlikely(ret)) { 143 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret); 144 ipoib_dma_unmap_rx(priv, rx_req); 145 m_freem(rx->rx_ring[id].mb); 146 rx->rx_ring[id].mb = NULL; 147 } 148 149 return ret; 150 } 151 152 static struct mbuf * 153 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req) 154 { 155 return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req, 156 priv->cm.max_cm_mtu); 157 } 158 159 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv, 160 struct ipoib_cm_rx_buf *rx_ring) 161 { 162 int i; 163 164 for (i = 0; i < ipoib_recvq_size; ++i) 165 if (rx_ring[i].mb) { 166 ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]); 167 m_freem(rx_ring[i].mb); 168 } 169 170 kfree(rx_ring); 171 } 172 173 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv) 174 { 175 struct ib_send_wr *bad_wr; 176 struct ipoib_cm_rx *p; 177 178 /* We only reserved 1 extra slot in CQ for drain WRs, so 179 * make sure we have at most 1 outstanding WR. */ 180 if (list_empty(&priv->cm.rx_flush_list) || 181 !list_empty(&priv->cm.rx_drain_list)) 182 return; 183 184 /* 185 * QPs on flush list are error state. This way, a "flush 186 * error" WC will be immediately generated for each WR we post. 187 */ 188 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list); 189 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr)) 190 ipoib_warn(priv, "failed to post drain wr\n"); 191 192 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list); 193 } 194 195 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx) 196 { 197 struct ipoib_cm_rx *p = ctx; 198 struct ipoib_dev_priv *priv = p->priv; 199 unsigned long flags; 200 201 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED) 202 return; 203 204 spin_lock_irqsave(&priv->lock, flags); 205 list_move(&p->list, &priv->cm.rx_flush_list); 206 p->state = IPOIB_CM_RX_FLUSH; 207 ipoib_cm_start_rx_drain(priv); 208 spin_unlock_irqrestore(&priv->lock, flags); 209 } 210 211 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv, 212 struct ipoib_cm_rx *p) 213 { 214 struct ib_qp_init_attr attr = { 215 .event_handler = ipoib_cm_rx_event_handler, 216 .send_cq = priv->recv_cq, /* For drain WR */ 217 .recv_cq = priv->recv_cq, 218 .srq = priv->cm.srq, 219 .cap.max_send_wr = 1, /* For drain WR */ 220 .cap.max_send_sge = 1, 221 .sq_sig_type = IB_SIGNAL_ALL_WR, 222 .qp_type = IB_QPT_RC, 223 .qp_context = p, 224 }; 225 226 if (!ipoib_cm_has_srq(priv)) { 227 attr.cap.max_recv_wr = ipoib_recvq_size; 228 attr.cap.max_recv_sge = priv->cm.num_frags; 229 } 230 231 return ib_create_qp(priv->pd, &attr); 232 } 233 234 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv, 235 struct ib_cm_id *cm_id, struct ib_qp *qp, 236 unsigned psn) 237 { 238 struct ib_qp_attr qp_attr; 239 int qp_attr_mask, ret; 240 241 qp_attr.qp_state = IB_QPS_INIT; 242 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 243 if (ret) { 244 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret); 245 return ret; 246 } 247 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 248 if (ret) { 249 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret); 250 return ret; 251 } 252 qp_attr.qp_state = IB_QPS_RTR; 253 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 254 if (ret) { 255 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret); 256 return ret; 257 } 258 qp_attr.rq_psn = psn; 259 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 260 if (ret) { 261 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret); 262 return ret; 263 } 264 265 /* 266 * Current Mellanox HCA firmware won't generate completions 267 * with error for drain WRs unless the QP has been moved to 268 * RTS first. This work-around leaves a window where a QP has 269 * moved to error asynchronously, but this will eventually get 270 * fixed in firmware, so let's not error out if modify QP 271 * fails. 272 */ 273 qp_attr.qp_state = IB_QPS_RTS; 274 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 275 if (ret) { 276 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 277 return 0; 278 } 279 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 280 if (ret) { 281 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 282 return 0; 283 } 284 285 return 0; 286 } 287 288 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv, 289 struct ib_recv_wr *wr, 290 struct ib_sge *sge) 291 { 292 int i; 293 294 for (i = 0; i < IPOIB_CM_RX_SG; i++) 295 sge[i].lkey = priv->pd->local_dma_lkey; 296 297 wr->next = NULL; 298 wr->sg_list = sge; 299 wr->num_sge = 1; 300 } 301 302 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv, 303 struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx) 304 { 305 struct { 306 struct ib_recv_wr wr; 307 struct ib_sge sge[IPOIB_CM_RX_SG]; 308 } *t; 309 int ret; 310 int i; 311 312 rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL); 313 if (!rx->rx_ring) { 314 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n", 315 priv->ca->name, ipoib_recvq_size); 316 return -ENOMEM; 317 } 318 319 memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring); 320 321 t = kmalloc(sizeof *t, GFP_KERNEL); 322 if (!t) { 323 ret = -ENOMEM; 324 goto err_free; 325 } 326 327 ipoib_cm_init_rx_wr(priv, &t->wr, t->sge); 328 329 spin_lock_irq(&priv->lock); 330 331 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) { 332 spin_unlock_irq(&priv->lock); 333 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0); 334 ret = -EINVAL; 335 goto err_free; 336 } else 337 ++priv->cm.nonsrq_conn_qp; 338 339 spin_unlock_irq(&priv->lock); 340 341 for (i = 0; i < ipoib_recvq_size; ++i) { 342 if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) { 343 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 344 ret = -ENOMEM; 345 goto err_count; 346 } 347 ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i); 348 if (ret) { 349 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq " 350 "failed for buf %d\n", i); 351 ret = -EIO; 352 goto err_count; 353 } 354 } 355 356 rx->recv_count = ipoib_recvq_size; 357 358 kfree(t); 359 360 return 0; 361 362 err_count: 363 spin_lock_irq(&priv->lock); 364 --priv->cm.nonsrq_conn_qp; 365 spin_unlock_irq(&priv->lock); 366 367 err_free: 368 kfree(t); 369 ipoib_cm_free_rx_ring(priv, rx->rx_ring); 370 371 return ret; 372 } 373 374 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id, 375 struct ib_qp *qp, struct ib_cm_req_event_param *req, 376 unsigned psn) 377 { 378 struct ipoib_cm_data data = {}; 379 struct ib_cm_rep_param rep = {}; 380 381 data.qpn = cpu_to_be32(priv->qp->qp_num); 382 data.mtu = cpu_to_be32(priv->cm.max_cm_mtu); 383 384 rep.private_data = &data; 385 rep.private_data_len = sizeof data; 386 rep.flow_control = 0; 387 rep.rnr_retry_count = req->rnr_retry_count; 388 rep.srq = ipoib_cm_has_srq(priv); 389 rep.qp_num = qp->qp_num; 390 rep.starting_psn = psn; 391 return ib_send_cm_rep(cm_id, &rep); 392 } 393 394 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 395 { 396 struct ipoib_dev_priv *priv = cm_id->context; 397 struct ipoib_cm_rx *p; 398 unsigned psn; 399 int ret; 400 401 ipoib_dbg(priv, "REQ arrived\n"); 402 p = kzalloc(sizeof *p, GFP_KERNEL); 403 if (!p) 404 return -ENOMEM; 405 p->priv = priv; 406 p->id = cm_id; 407 cm_id->context = p; 408 p->state = IPOIB_CM_RX_LIVE; 409 p->jiffies = jiffies; 410 INIT_LIST_HEAD(&p->list); 411 412 p->qp = ipoib_cm_create_rx_qp(priv, p); 413 if (IS_ERR(p->qp)) { 414 ret = PTR_ERR(p->qp); 415 goto err_qp; 416 } 417 418 psn = random() & 0xffffff; 419 ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn); 420 if (ret) 421 goto err_modify; 422 423 if (!ipoib_cm_has_srq(priv)) { 424 ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p); 425 if (ret) 426 goto err_modify; 427 } 428 429 spin_lock_irq(&priv->lock); 430 queue_delayed_work(ipoib_workqueue, 431 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 432 /* Add this entry to passive ids list head, but do not re-add it 433 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */ 434 p->jiffies = jiffies; 435 if (p->state == IPOIB_CM_RX_LIVE) 436 list_move(&p->list, &priv->cm.passive_ids); 437 spin_unlock_irq(&priv->lock); 438 439 ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn); 440 if (ret) { 441 ipoib_warn(priv, "failed to send REP: %d\n", ret); 442 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 443 ipoib_warn(priv, "unable to move qp to error state\n"); 444 } 445 return 0; 446 447 err_modify: 448 ib_destroy_qp(p->qp); 449 err_qp: 450 kfree(p); 451 return ret; 452 } 453 454 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id, 455 struct ib_cm_event *event) 456 { 457 struct ipoib_cm_rx *p; 458 struct ipoib_dev_priv *priv; 459 460 switch (event->event) { 461 case IB_CM_REQ_RECEIVED: 462 return ipoib_cm_req_handler(cm_id, event); 463 case IB_CM_DREQ_RECEIVED: 464 p = cm_id->context; 465 ib_send_cm_drep(cm_id, NULL, 0); 466 /* Fall through */ 467 case IB_CM_REJ_RECEIVED: 468 p = cm_id->context; 469 priv = p->priv; 470 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 471 ipoib_warn(priv, "unable to move qp to error state\n"); 472 /* Fall through */ 473 default: 474 return 0; 475 } 476 } 477 478 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc) 479 { 480 struct ipoib_cm_rx_buf saverx; 481 struct ipoib_cm_rx_buf *rx_ring; 482 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV); 483 struct ifnet *dev = priv->dev; 484 struct mbuf *mb, *newmb; 485 struct ipoib_cm_rx *p; 486 int has_srq; 487 u_short proto; 488 489 CURVNET_SET_QUIET(dev->if_vnet); 490 491 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n", 492 wr_id, wc->status); 493 494 if (unlikely(wr_id >= ipoib_recvq_size)) { 495 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) { 496 spin_lock(&priv->lock); 497 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list); 498 ipoib_cm_start_rx_drain(priv); 499 if (priv->cm.id != NULL) 500 queue_work(ipoib_workqueue, 501 &priv->cm.rx_reap_task); 502 spin_unlock(&priv->lock); 503 } else 504 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n", 505 wr_id, ipoib_recvq_size); 506 goto done; 507 } 508 509 p = wc->qp->qp_context; 510 511 has_srq = ipoib_cm_has_srq(priv); 512 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring; 513 514 mb = rx_ring[wr_id].mb; 515 516 if (unlikely(wc->status != IB_WC_SUCCESS)) { 517 ipoib_dbg(priv, "cm recv error " 518 "(status=%d, wrid=%d vend_err %x)\n", 519 wc->status, wr_id, wc->vendor_err); 520 if_inc_counter(dev, IFCOUNTER_IERRORS, 1); 521 if (has_srq) 522 goto repost; 523 else { 524 if (!--p->recv_count) { 525 spin_lock(&priv->lock); 526 list_move(&p->list, &priv->cm.rx_reap_list); 527 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task); 528 spin_unlock(&priv->lock); 529 } 530 goto done; 531 } 532 } 533 534 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) { 535 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) { 536 p->jiffies = jiffies; 537 /* Move this entry to list head, but do not re-add it 538 * if it has been moved out of list. */ 539 if (p->state == IPOIB_CM_RX_LIVE) 540 list_move(&p->list, &priv->cm.passive_ids); 541 } 542 } 543 544 memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx)); 545 newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]); 546 if (unlikely(!newmb)) { 547 /* 548 * If we can't allocate a new RX buffer, dump 549 * this packet and reuse the old buffer. 550 */ 551 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id); 552 if_inc_counter(dev, IFCOUNTER_IERRORS, 1); 553 memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx)); 554 goto repost; 555 } 556 557 ipoib_cm_dma_unmap_rx(priv, &saverx); 558 559 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 560 wc->byte_len, wc->slid); 561 562 ipoib_dma_mb(priv, mb, wc->byte_len); 563 564 if_inc_counter(dev, IFCOUNTER_IPACKETS, 1); 565 if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len); 566 567 mb->m_pkthdr.rcvif = dev; 568 proto = *mtod(mb, uint16_t *); 569 m_adj(mb, IPOIB_ENCAP_LEN); 570 571 IPOIB_MTAP_PROTO(dev, mb, proto); 572 ipoib_demux(dev, mb, ntohs(proto)); 573 574 repost: 575 if (has_srq) { 576 if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id))) 577 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed " 578 "for buf %d\n", wr_id); 579 } else { 580 if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p, 581 &priv->cm.rx_wr, 582 priv->cm.rx_sge, 583 wr_id))) { 584 --p->recv_count; 585 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed " 586 "for buf %d\n", wr_id); 587 } 588 } 589 done: 590 CURVNET_RESTORE(); 591 return; 592 } 593 594 static inline int post_send(struct ipoib_dev_priv *priv, 595 struct ipoib_cm_tx *tx, 596 struct ipoib_cm_tx_buf *tx_req, 597 unsigned int wr_id) 598 { 599 struct ib_send_wr *bad_wr; 600 struct mbuf *mb = tx_req->mb; 601 u64 *mapping = tx_req->mapping; 602 struct mbuf *m; 603 int i; 604 605 for (m = mb, i = 0; m != NULL; m = m->m_next, i++) { 606 priv->tx_sge[i].addr = mapping[i]; 607 priv->tx_sge[i].length = m->m_len; 608 } 609 priv->tx_wr.wr.num_sge = i; 610 priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM; 611 priv->tx_wr.wr.opcode = IB_WR_SEND; 612 613 return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr); 614 } 615 616 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx) 617 { 618 struct ipoib_cm_tx_buf *tx_req; 619 struct ifnet *dev = priv->dev; 620 621 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) 622 while (ipoib_poll_tx(priv)); /* nothing */ 623 624 m_adj(mb, sizeof(struct ipoib_pseudoheader)); 625 if (unlikely(mb->m_pkthdr.len > tx->mtu)) { 626 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 627 mb->m_pkthdr.len, tx->mtu); 628 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 629 ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu)); 630 return; 631 } 632 633 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n", 634 tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num); 635 636 637 /* 638 * We put the mb into the tx_ring _before_ we call post_send() 639 * because it's entirely possible that the completion handler will 640 * run before we execute anything after the post_send(). That 641 * means we have to make sure everything is properly recorded and 642 * our state is consistent before we call post_send(). 643 */ 644 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)]; 645 tx_req->mb = mb; 646 if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req, 647 priv->cm.num_frags))) { 648 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 649 if (tx_req->mb) 650 m_freem(tx_req->mb); 651 return; 652 } 653 654 if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) { 655 ipoib_warn(priv, "post_send failed\n"); 656 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 657 ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req); 658 m_freem(mb); 659 } else { 660 ++tx->tx_head; 661 662 if (++priv->tx_outstanding == ipoib_sendq_size) { 663 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n", 664 tx->qp->qp_num); 665 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP)) 666 ipoib_warn(priv, "request notify on send CQ failed\n"); 667 dev->if_drv_flags |= IFF_DRV_OACTIVE; 668 } 669 } 670 671 } 672 673 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc) 674 { 675 struct ipoib_cm_tx *tx = wc->qp->qp_context; 676 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM; 677 struct ifnet *dev = priv->dev; 678 struct ipoib_cm_tx_buf *tx_req; 679 680 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n", 681 wr_id, wc->status); 682 683 if (unlikely(wr_id >= ipoib_sendq_size)) { 684 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n", 685 wr_id, ipoib_sendq_size); 686 return; 687 } 688 689 tx_req = &tx->tx_ring[wr_id]; 690 691 ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req); 692 693 /* FIXME: is this right? Shouldn't we only increment on success? */ 694 if_inc_counter(dev, IFCOUNTER_OPACKETS, 1); 695 696 m_freem(tx_req->mb); 697 698 ++tx->tx_tail; 699 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 700 (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 && 701 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 702 dev->if_drv_flags &= ~IFF_DRV_OACTIVE; 703 704 if (wc->status != IB_WC_SUCCESS && 705 wc->status != IB_WC_WR_FLUSH_ERR) { 706 struct ipoib_path *path; 707 708 ipoib_dbg(priv, "failed cm send event " 709 "(status=%d, wrid=%d vend_err %x)\n", 710 wc->status, wr_id, wc->vendor_err); 711 712 path = tx->path; 713 714 if (path) { 715 path->cm = NULL; 716 rb_erase(&path->rb_node, &priv->path_tree); 717 list_del(&path->list); 718 } 719 720 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 721 list_move(&tx->list, &priv->cm.reap_list); 722 queue_work(ipoib_workqueue, &priv->cm.reap_task); 723 } 724 725 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags); 726 } 727 728 } 729 730 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv) 731 { 732 int ret; 733 734 if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev))) 735 return 0; 736 737 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv); 738 if (IS_ERR(priv->cm.id)) { 739 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name); 740 ret = PTR_ERR(priv->cm.id); 741 goto err_cm; 742 } 743 744 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 0); 745 if (ret) { 746 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name, 747 IPOIB_CM_IETF_ID | priv->qp->qp_num); 748 goto err_listen; 749 } 750 751 return 0; 752 753 err_listen: 754 ib_destroy_cm_id(priv->cm.id); 755 err_cm: 756 priv->cm.id = NULL; 757 return ret; 758 } 759 760 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv) 761 { 762 struct ipoib_cm_rx *rx, *n; 763 LIST_HEAD(list); 764 765 spin_lock_irq(&priv->lock); 766 list_splice_init(&priv->cm.rx_reap_list, &list); 767 spin_unlock_irq(&priv->lock); 768 769 list_for_each_entry_safe(rx, n, &list, list) { 770 ib_destroy_cm_id(rx->id); 771 ib_destroy_qp(rx->qp); 772 if (!ipoib_cm_has_srq(priv)) { 773 ipoib_cm_free_rx_ring(priv, rx->rx_ring); 774 spin_lock_irq(&priv->lock); 775 --priv->cm.nonsrq_conn_qp; 776 spin_unlock_irq(&priv->lock); 777 } 778 kfree(rx); 779 } 780 } 781 782 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv) 783 { 784 struct ipoib_cm_rx *p; 785 unsigned long begin; 786 int ret; 787 788 if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)) || !priv->cm.id) 789 return; 790 791 ib_destroy_cm_id(priv->cm.id); 792 priv->cm.id = NULL; 793 794 cancel_work_sync(&priv->cm.rx_reap_task); 795 796 spin_lock_irq(&priv->lock); 797 while (!list_empty(&priv->cm.passive_ids)) { 798 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list); 799 list_move(&p->list, &priv->cm.rx_error_list); 800 p->state = IPOIB_CM_RX_ERROR; 801 spin_unlock_irq(&priv->lock); 802 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 803 if (ret) 804 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 805 spin_lock_irq(&priv->lock); 806 } 807 808 /* Wait for all RX to be drained */ 809 begin = jiffies; 810 811 while (!list_empty(&priv->cm.rx_error_list) || 812 !list_empty(&priv->cm.rx_flush_list) || 813 !list_empty(&priv->cm.rx_drain_list)) { 814 if (time_after(jiffies, begin + 5 * HZ)) { 815 ipoib_warn(priv, "RX drain timing out\n"); 816 817 /* 818 * assume the HW is wedged and just free up everything. 819 */ 820 list_splice_init(&priv->cm.rx_flush_list, 821 &priv->cm.rx_reap_list); 822 list_splice_init(&priv->cm.rx_error_list, 823 &priv->cm.rx_reap_list); 824 list_splice_init(&priv->cm.rx_drain_list, 825 &priv->cm.rx_reap_list); 826 break; 827 } 828 spin_unlock_irq(&priv->lock); 829 msleep(1); 830 ipoib_drain_cq(priv); 831 spin_lock_irq(&priv->lock); 832 } 833 834 spin_unlock_irq(&priv->lock); 835 836 ipoib_cm_free_rx_reap_list(priv); 837 838 cancel_delayed_work_sync(&priv->cm.stale_task); 839 } 840 841 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 842 { 843 struct ipoib_cm_tx *p = cm_id->context; 844 struct ipoib_dev_priv *priv = p->priv; 845 struct ipoib_cm_data *data = event->private_data; 846 struct ifqueue mbqueue; 847 struct ib_qp_attr qp_attr; 848 int qp_attr_mask, ret; 849 struct mbuf *mb; 850 851 ipoib_dbg(priv, "cm rep handler\n"); 852 p->mtu = be32_to_cpu(data->mtu); 853 854 if (p->mtu <= IPOIB_ENCAP_LEN) { 855 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n", 856 p->mtu, IPOIB_ENCAP_LEN); 857 return -EINVAL; 858 } 859 860 qp_attr.qp_state = IB_QPS_RTR; 861 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 862 if (ret) { 863 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret); 864 return ret; 865 } 866 867 qp_attr.rq_psn = 0 /* FIXME */; 868 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 869 if (ret) { 870 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret); 871 return ret; 872 } 873 874 qp_attr.qp_state = IB_QPS_RTS; 875 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 876 if (ret) { 877 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 878 return ret; 879 } 880 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 881 if (ret) { 882 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 883 return ret; 884 } 885 886 bzero(&mbqueue, sizeof(mbqueue)); 887 888 spin_lock_irq(&priv->lock); 889 set_bit(IPOIB_FLAG_OPER_UP, &p->flags); 890 if (p->path) 891 for (;;) { 892 _IF_DEQUEUE(&p->path->queue, mb); 893 if (mb == NULL) 894 break; 895 _IF_ENQUEUE(&mbqueue, mb); 896 } 897 spin_unlock_irq(&priv->lock); 898 899 for (;;) { 900 struct ifnet *dev = p->priv->dev; 901 _IF_DEQUEUE(&mbqueue, mb); 902 if (mb == NULL) 903 break; 904 mb->m_pkthdr.rcvif = dev; 905 if (dev->if_transmit(dev, mb)) 906 ipoib_warn(priv, "dev_queue_xmit failed " 907 "to requeue packet\n"); 908 } 909 910 ret = ib_send_cm_rtu(cm_id, NULL, 0); 911 if (ret) { 912 ipoib_warn(priv, "failed to send RTU: %d\n", ret); 913 return ret; 914 } 915 return 0; 916 } 917 918 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv, 919 struct ipoib_cm_tx *tx) 920 { 921 struct ib_qp_init_attr attr = { 922 .send_cq = priv->send_cq, 923 .recv_cq = priv->recv_cq, 924 .srq = priv->cm.srq, 925 .cap.max_send_wr = ipoib_sendq_size, 926 .cap.max_send_sge = priv->cm.num_frags, 927 .sq_sig_type = IB_SIGNAL_ALL_WR, 928 .qp_type = IB_QPT_RC, 929 .qp_context = tx 930 }; 931 932 return ib_create_qp(priv->pd, &attr); 933 } 934 935 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv, 936 struct ib_cm_id *id, struct ib_qp *qp, 937 u32 qpn, 938 struct ib_sa_path_rec *pathrec) 939 { 940 struct ipoib_cm_data data = {}; 941 struct ib_cm_req_param req = {}; 942 943 ipoib_dbg(priv, "cm send req\n"); 944 945 data.qpn = cpu_to_be32(priv->qp->qp_num); 946 data.mtu = cpu_to_be32(priv->cm.max_cm_mtu); 947 948 req.primary_path = pathrec; 949 req.alternate_path = NULL; 950 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn); 951 req.qp_num = qp->qp_num; 952 req.qp_type = qp->qp_type; 953 req.private_data = &data; 954 req.private_data_len = sizeof data; 955 req.flow_control = 0; 956 957 req.starting_psn = 0; /* FIXME */ 958 959 /* 960 * Pick some arbitrary defaults here; we could make these 961 * module parameters if anyone cared about setting them. 962 */ 963 req.responder_resources = 4; 964 req.remote_cm_response_timeout = 20; 965 req.local_cm_response_timeout = 20; 966 req.retry_count = 0; /* RFC draft warns against retries */ 967 req.rnr_retry_count = 0; /* RFC draft warns against retries */ 968 req.max_cm_retries = 15; 969 req.srq = ipoib_cm_has_srq(priv); 970 return ib_send_cm_req(id, &req); 971 } 972 973 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv, 974 struct ib_cm_id *cm_id, struct ib_qp *qp) 975 { 976 struct ib_qp_attr qp_attr; 977 int qp_attr_mask, ret; 978 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index); 979 if (ret) { 980 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret); 981 return ret; 982 } 983 984 qp_attr.qp_state = IB_QPS_INIT; 985 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE; 986 qp_attr.port_num = priv->port; 987 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT; 988 989 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 990 if (ret) { 991 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret); 992 return ret; 993 } 994 return 0; 995 } 996 997 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn, 998 struct ib_sa_path_rec *pathrec) 999 { 1000 struct ipoib_dev_priv *priv = p->priv; 1001 int ret; 1002 1003 p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL); 1004 if (!p->tx_ring) { 1005 ipoib_warn(priv, "failed to allocate tx ring\n"); 1006 ret = -ENOMEM; 1007 goto err_tx; 1008 } 1009 memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring); 1010 1011 p->qp = ipoib_cm_create_tx_qp(p->priv, p); 1012 if (IS_ERR(p->qp)) { 1013 ret = PTR_ERR(p->qp); 1014 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret); 1015 goto err_qp; 1016 } 1017 1018 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p); 1019 if (IS_ERR(p->id)) { 1020 ret = PTR_ERR(p->id); 1021 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret); 1022 goto err_id; 1023 } 1024 1025 ret = ipoib_cm_modify_tx_init(p->priv, p->id, p->qp); 1026 if (ret) { 1027 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret); 1028 goto err_modify; 1029 } 1030 1031 ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec); 1032 if (ret) { 1033 ipoib_warn(priv, "failed to send cm req: %d\n", ret); 1034 goto err_send_cm; 1035 } 1036 1037 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n", 1038 p->qp->qp_num, pathrec->dgid.raw, qpn); 1039 1040 return 0; 1041 1042 err_send_cm: 1043 err_modify: 1044 ib_destroy_cm_id(p->id); 1045 err_id: 1046 p->id = NULL; 1047 ib_destroy_qp(p->qp); 1048 err_qp: 1049 p->qp = NULL; 1050 kfree(p->tx_ring); 1051 err_tx: 1052 return ret; 1053 } 1054 1055 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p) 1056 { 1057 struct ipoib_dev_priv *priv = p->priv; 1058 struct ifnet *dev = priv->dev; 1059 struct ipoib_cm_tx_buf *tx_req; 1060 unsigned long begin; 1061 1062 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n", 1063 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail); 1064 1065 if (p->path) 1066 ipoib_path_free(priv, p->path); 1067 1068 if (p->id) 1069 ib_destroy_cm_id(p->id); 1070 1071 if (p->tx_ring) { 1072 /* Wait for all sends to complete */ 1073 begin = jiffies; 1074 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1075 if (time_after(jiffies, begin + 5 * HZ)) { 1076 ipoib_warn(priv, "timing out; %d sends not completed\n", 1077 p->tx_head - p->tx_tail); 1078 goto timeout; 1079 } 1080 1081 msleep(1); 1082 } 1083 } 1084 1085 timeout: 1086 1087 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1088 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)]; 1089 ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req); 1090 m_freem(tx_req->mb); 1091 ++p->tx_tail; 1092 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 1093 (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 && 1094 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 1095 dev->if_drv_flags &= ~IFF_DRV_OACTIVE; 1096 } 1097 1098 if (p->qp) 1099 ib_destroy_qp(p->qp); 1100 1101 kfree(p->tx_ring); 1102 kfree(p); 1103 } 1104 1105 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id, 1106 struct ib_cm_event *event) 1107 { 1108 struct ipoib_cm_tx *tx = cm_id->context; 1109 struct ipoib_dev_priv *priv = tx->priv; 1110 struct ipoib_path *path; 1111 unsigned long flags; 1112 int ret; 1113 1114 switch (event->event) { 1115 case IB_CM_DREQ_RECEIVED: 1116 ipoib_dbg(priv, "DREQ received.\n"); 1117 ib_send_cm_drep(cm_id, NULL, 0); 1118 break; 1119 case IB_CM_REP_RECEIVED: 1120 ipoib_dbg(priv, "REP received.\n"); 1121 ret = ipoib_cm_rep_handler(cm_id, event); 1122 if (ret) 1123 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, 1124 NULL, 0, NULL, 0); 1125 break; 1126 case IB_CM_REQ_ERROR: 1127 case IB_CM_REJ_RECEIVED: 1128 case IB_CM_TIMEWAIT_EXIT: 1129 ipoib_dbg(priv, "CM error %d.\n", event->event); 1130 spin_lock_irqsave(&priv->lock, flags); 1131 path = tx->path; 1132 1133 if (path) { 1134 path->cm = NULL; 1135 tx->path = NULL; 1136 rb_erase(&path->rb_node, &priv->path_tree); 1137 list_del(&path->list); 1138 } 1139 1140 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1141 list_move(&tx->list, &priv->cm.reap_list); 1142 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1143 } 1144 1145 spin_unlock_irqrestore(&priv->lock, flags); 1146 if (path) 1147 ipoib_path_free(tx->priv, path); 1148 break; 1149 default: 1150 break; 1151 } 1152 1153 return 0; 1154 } 1155 1156 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv, 1157 struct ipoib_path *path) 1158 { 1159 struct ipoib_cm_tx *tx; 1160 1161 tx = kzalloc(sizeof *tx, GFP_ATOMIC); 1162 if (!tx) 1163 return NULL; 1164 1165 ipoib_dbg(priv, "Creating cm tx\n"); 1166 path->cm = tx; 1167 tx->path = path; 1168 tx->priv = priv; 1169 list_add(&tx->list, &priv->cm.start_list); 1170 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags); 1171 queue_work(ipoib_workqueue, &priv->cm.start_task); 1172 return tx; 1173 } 1174 1175 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx) 1176 { 1177 struct ipoib_dev_priv *priv = tx->priv; 1178 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1179 spin_lock(&priv->lock); 1180 list_move(&tx->list, &priv->cm.reap_list); 1181 spin_unlock(&priv->lock); 1182 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1183 ipoib_dbg(priv, "Reap connection for gid %pI6\n", 1184 tx->path->pathrec.dgid.raw); 1185 tx->path = NULL; 1186 } 1187 } 1188 1189 static void ipoib_cm_tx_start(struct work_struct *work) 1190 { 1191 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1192 cm.start_task); 1193 struct ipoib_path *path; 1194 struct ipoib_cm_tx *p; 1195 unsigned long flags; 1196 int ret; 1197 1198 struct ib_sa_path_rec pathrec; 1199 u32 qpn; 1200 1201 ipoib_dbg(priv, "cm start task\n"); 1202 spin_lock_irqsave(&priv->lock, flags); 1203 1204 while (!list_empty(&priv->cm.start_list)) { 1205 p = list_entry(priv->cm.start_list.next, typeof(*p), list); 1206 list_del_init(&p->list); 1207 path = p->path; 1208 qpn = IPOIB_QPN(path->hwaddr); 1209 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec); 1210 1211 spin_unlock_irqrestore(&priv->lock, flags); 1212 1213 ret = ipoib_cm_tx_init(p, qpn, &pathrec); 1214 1215 spin_lock_irqsave(&priv->lock, flags); 1216 1217 if (ret) { 1218 path = p->path; 1219 if (path) { 1220 path->cm = NULL; 1221 rb_erase(&path->rb_node, &priv->path_tree); 1222 list_del(&path->list); 1223 ipoib_path_free(priv, path); 1224 } 1225 list_del(&p->list); 1226 kfree(p); 1227 } 1228 } 1229 1230 spin_unlock_irqrestore(&priv->lock, flags); 1231 } 1232 1233 static void ipoib_cm_tx_reap(struct work_struct *work) 1234 { 1235 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1236 cm.reap_task); 1237 struct ipoib_cm_tx *p; 1238 unsigned long flags; 1239 1240 spin_lock_irqsave(&priv->lock, flags); 1241 1242 while (!list_empty(&priv->cm.reap_list)) { 1243 p = list_entry(priv->cm.reap_list.next, typeof(*p), list); 1244 list_del(&p->list); 1245 spin_unlock_irqrestore(&priv->lock, flags); 1246 ipoib_cm_tx_destroy(p); 1247 spin_lock_irqsave(&priv->lock, flags); 1248 } 1249 1250 spin_unlock_irqrestore(&priv->lock, flags); 1251 } 1252 1253 static void ipoib_cm_mb_reap(struct work_struct *work) 1254 { 1255 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1256 cm.mb_task); 1257 struct mbuf *mb; 1258 unsigned long flags; 1259 #if defined(INET) || defined(INET6) 1260 unsigned mtu = priv->mcast_mtu; 1261 #endif 1262 uint16_t proto; 1263 1264 spin_lock_irqsave(&priv->lock, flags); 1265 1266 for (;;) { 1267 IF_DEQUEUE(&priv->cm.mb_queue, mb); 1268 if (mb == NULL) 1269 break; 1270 spin_unlock_irqrestore(&priv->lock, flags); 1271 1272 proto = htons(*mtod(mb, uint16_t *)); 1273 m_adj(mb, IPOIB_ENCAP_LEN); 1274 switch (proto) { 1275 #if defined(INET) 1276 case ETHERTYPE_IP: 1277 icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu); 1278 break; 1279 #endif 1280 #if defined(INET6) 1281 case ETHERTYPE_IPV6: 1282 icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu); 1283 break; 1284 #endif 1285 default: 1286 m_freem(mb); 1287 } 1288 1289 spin_lock_irqsave(&priv->lock, flags); 1290 } 1291 1292 spin_unlock_irqrestore(&priv->lock, flags); 1293 } 1294 1295 void 1296 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu) 1297 { 1298 int e = priv->cm.mb_queue.ifq_len; 1299 1300 IF_ENQUEUE(&priv->cm.mb_queue, mb); 1301 if (e == 0) 1302 queue_work(ipoib_workqueue, &priv->cm.mb_task); 1303 } 1304 1305 static void ipoib_cm_rx_reap(struct work_struct *work) 1306 { 1307 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv, 1308 cm.rx_reap_task)); 1309 } 1310 1311 static void ipoib_cm_stale_task(struct work_struct *work) 1312 { 1313 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1314 cm.stale_task.work); 1315 struct ipoib_cm_rx *p; 1316 int ret; 1317 1318 spin_lock_irq(&priv->lock); 1319 while (!list_empty(&priv->cm.passive_ids)) { 1320 /* List is sorted by LRU, start from tail, 1321 * stop when we see a recently used entry */ 1322 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list); 1323 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT)) 1324 break; 1325 list_move(&p->list, &priv->cm.rx_error_list); 1326 p->state = IPOIB_CM_RX_ERROR; 1327 spin_unlock_irq(&priv->lock); 1328 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 1329 if (ret) 1330 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 1331 spin_lock_irq(&priv->lock); 1332 } 1333 1334 if (!list_empty(&priv->cm.passive_ids)) 1335 queue_delayed_work(ipoib_workqueue, 1336 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 1337 spin_unlock_irq(&priv->lock); 1338 } 1339 1340 1341 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge) 1342 { 1343 struct ib_srq_init_attr srq_init_attr = { 1344 .attr = { 1345 .max_wr = ipoib_recvq_size, 1346 .max_sge = max_sge 1347 } 1348 }; 1349 1350 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr); 1351 if (IS_ERR(priv->cm.srq)) { 1352 if (PTR_ERR(priv->cm.srq) != -ENOSYS) 1353 printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n", 1354 priv->ca->name, PTR_ERR(priv->cm.srq)); 1355 priv->cm.srq = NULL; 1356 return; 1357 } 1358 1359 priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL); 1360 if (!priv->cm.srq_ring) { 1361 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n", 1362 priv->ca->name, ipoib_recvq_size); 1363 ib_destroy_srq(priv->cm.srq); 1364 priv->cm.srq = NULL; 1365 return; 1366 } 1367 1368 memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring); 1369 } 1370 1371 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv) 1372 { 1373 struct ifnet *dev = priv->dev; 1374 int i; 1375 int max_srq_sge; 1376 1377 INIT_LIST_HEAD(&priv->cm.passive_ids); 1378 INIT_LIST_HEAD(&priv->cm.reap_list); 1379 INIT_LIST_HEAD(&priv->cm.start_list); 1380 INIT_LIST_HEAD(&priv->cm.rx_error_list); 1381 INIT_LIST_HEAD(&priv->cm.rx_flush_list); 1382 INIT_LIST_HEAD(&priv->cm.rx_drain_list); 1383 INIT_LIST_HEAD(&priv->cm.rx_reap_list); 1384 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start); 1385 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap); 1386 INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap); 1387 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap); 1388 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task); 1389 1390 bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue)); 1391 mtx_init(&priv->cm.mb_queue.ifq_mtx, 1392 dev->if_xname, "if send queue", MTX_DEF); 1393 1394 max_srq_sge = priv->ca->attrs.max_srq_sge; 1395 1396 ipoib_dbg(priv, "max_srq_sge=%d\n", max_srq_sge); 1397 1398 max_srq_sge = min_t(int, IPOIB_CM_RX_SG, max_srq_sge); 1399 ipoib_cm_create_srq(priv, max_srq_sge); 1400 if (ipoib_cm_has_srq(priv)) { 1401 priv->cm.max_cm_mtu = max_srq_sge * MJUMPAGESIZE; 1402 priv->cm.num_frags = max_srq_sge; 1403 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n", 1404 priv->cm.max_cm_mtu, priv->cm.num_frags); 1405 } else { 1406 priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU; 1407 priv->cm.num_frags = IPOIB_CM_RX_SG; 1408 } 1409 1410 ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge); 1411 1412 if (ipoib_cm_has_srq(priv)) { 1413 for (i = 0; i < ipoib_recvq_size; ++i) { 1414 if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) { 1415 ipoib_warn(priv, "failed to allocate " 1416 "receive buffer %d\n", i); 1417 ipoib_cm_dev_cleanup(priv); 1418 return -ENOMEM; 1419 } 1420 1421 if (ipoib_cm_post_receive_srq(priv, i)) { 1422 ipoib_warn(priv, "ipoib_cm_post_receive_srq " 1423 "failed for buf %d\n", i); 1424 ipoib_cm_dev_cleanup(priv); 1425 return -EIO; 1426 } 1427 } 1428 } 1429 1430 IF_LLADDR(priv->dev)[0] = IPOIB_FLAGS_RC; 1431 return 0; 1432 } 1433 1434 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv) 1435 { 1436 int ret; 1437 1438 if (!priv->cm.srq) 1439 return; 1440 1441 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n"); 1442 1443 ret = ib_destroy_srq(priv->cm.srq); 1444 if (ret) 1445 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret); 1446 1447 priv->cm.srq = NULL; 1448 if (!priv->cm.srq_ring) 1449 return; 1450 1451 ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring); 1452 priv->cm.srq_ring = NULL; 1453 1454 mtx_destroy(&priv->cm.mb_queue.ifq_mtx); 1455 } 1456 1457 #endif /* CONFIG_INFINIBAND_IPOIB_CM */ 1458