1/* 2 * Copyright (c) 2006 Mellanox Technologies. 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 <rdma/ib_cm.h> 34#include <net/dst.h> 35#include <net/icmp.h> 36#include <linux/icmpv6.h> 37#include <linux/delay.h> 38#include <linux/slab.h> 39#include <linux/vmalloc.h> 40 41#include "ipoib.h" 42 43int ipoib_max_conn_qp = 128; 44 45module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444); 46MODULE_PARM_DESC(max_nonsrq_conn_qp, 47 "Max number of connected-mode QPs per interface " 48 "(applied only if shared receive queue is not available)"); 49 50#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA 51static int data_debug_level; 52 53module_param_named(cm_data_debug_level, data_debug_level, int, 0644); 54MODULE_PARM_DESC(cm_data_debug_level, 55 "Enable data path debug tracing for connected mode if > 0"); 56#endif 57 58#define IPOIB_CM_IETF_ID 0x1000000000000000ULL 59 60#define IPOIB_CM_RX_UPDATE_TIME (256 * HZ) 61#define IPOIB_CM_RX_TIMEOUT (2 * 256 * HZ) 62#define IPOIB_CM_RX_DELAY (3 * 256 * HZ) 63#define IPOIB_CM_RX_UPDATE_MASK (0x3) 64 65static struct ib_qp_attr ipoib_cm_err_attr = { 66 .qp_state = IB_QPS_ERR 67}; 68 69#define IPOIB_CM_RX_DRAIN_WRID 0xffffffff 70 71static struct ib_send_wr ipoib_cm_rx_drain_wr = { 72 .wr_id = IPOIB_CM_RX_DRAIN_WRID, 73 .opcode = IB_WR_SEND, 74}; 75 76static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id, 77 struct ib_cm_event *event); 78 79static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags, 80 u64 mapping[IPOIB_CM_RX_SG]) 81{ 82 int i; 83 84 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE); 85 86 for (i = 0; i < frags; ++i) 87 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE); 88} 89 90static int ipoib_cm_post_receive_srq(struct net_device *dev, int id) 91{ 92 struct ipoib_dev_priv *priv = netdev_priv(dev); 93 struct ib_recv_wr *bad_wr; 94 int i, ret; 95 96 priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV; 97 98 for (i = 0; i < priv->cm.num_frags; ++i) 99 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i]; 100 101 ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr); 102 if (unlikely(ret)) { 103 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret); 104 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1, 105 priv->cm.srq_ring[id].mapping); 106 dev_kfree_skb_any(priv->cm.srq_ring[id].skb); 107 priv->cm.srq_ring[id].skb = NULL; 108 } 109 110 return ret; 111} 112 113static int ipoib_cm_post_receive_nonsrq(struct net_device *dev, 114 struct ipoib_cm_rx *rx, 115 struct ib_recv_wr *wr, 116 struct ib_sge *sge, int id) 117{ 118 struct ipoib_dev_priv *priv = netdev_priv(dev); 119 struct ib_recv_wr *bad_wr; 120 int i, ret; 121 122 wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV; 123 124 for (i = 0; i < IPOIB_CM_RX_SG; ++i) 125 sge[i].addr = rx->rx_ring[id].mapping[i]; 126 127 ret = ib_post_recv(rx->qp, wr, &bad_wr); 128 if (unlikely(ret)) { 129 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret); 130 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1, 131 rx->rx_ring[id].mapping); 132 dev_kfree_skb_any(rx->rx_ring[id].skb); 133 rx->rx_ring[id].skb = NULL; 134 } 135 136 return ret; 137} 138 139static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev, 140 struct ipoib_cm_rx_buf *rx_ring, 141 int id, int frags, 142 u64 mapping[IPOIB_CM_RX_SG]) 143{ 144 struct ipoib_dev_priv *priv = netdev_priv(dev); 145 struct sk_buff *skb; 146 int i; 147 148 skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12); 149 if (unlikely(!skb)) 150 return NULL; 151 152 /* 153 * IPoIB adds a 4 byte header. So we need 12 more bytes to align the 154 * IP header to a multiple of 16. 155 */ 156 skb_reserve(skb, 12); 157 158 mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE, 159 DMA_FROM_DEVICE); 160 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) { 161 dev_kfree_skb_any(skb); 162 return NULL; 163 } 164 165 for (i = 0; i < frags; i++) { 166 struct page *page = alloc_page(GFP_ATOMIC); 167 168 if (!page) 169 goto partial_error; 170 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE); 171 172 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page, 173 0, PAGE_SIZE, DMA_FROM_DEVICE); 174 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1]))) 175 goto partial_error; 176 } 177 178 rx_ring[id].skb = skb; 179 return skb; 180 181partial_error: 182 183 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE); 184 185 for (; i > 0; --i) 186 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE); 187 188 dev_kfree_skb_any(skb); 189 return NULL; 190} 191 192static void ipoib_cm_free_rx_ring(struct net_device *dev, 193 struct ipoib_cm_rx_buf *rx_ring) 194{ 195 struct ipoib_dev_priv *priv = netdev_priv(dev); 196 int i; 197 198 for (i = 0; i < ipoib_recvq_size; ++i) 199 if (rx_ring[i].skb) { 200 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1, 201 rx_ring[i].mapping); 202 dev_kfree_skb_any(rx_ring[i].skb); 203 } 204 205 vfree(rx_ring); 206} 207 208static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv) 209{ 210 struct ib_send_wr *bad_wr; 211 struct ipoib_cm_rx *p; 212 213 /* We only reserved 1 extra slot in CQ for drain WRs, so 214 * make sure we have at most 1 outstanding WR. */ 215 if (list_empty(&priv->cm.rx_flush_list) || 216 !list_empty(&priv->cm.rx_drain_list)) 217 return; 218 219 /* 220 * QPs on flush list are error state. This way, a "flush 221 * error" WC will be immediately generated for each WR we post. 222 */ 223 p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list); 224 if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr)) 225 ipoib_warn(priv, "failed to post drain wr\n"); 226 227 list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list); 228} 229 230static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx) 231{ 232 struct ipoib_cm_rx *p = ctx; 233 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 234 unsigned long flags; 235 236 if (event->event != IB_EVENT_QP_LAST_WQE_REACHED) 237 return; 238 239 spin_lock_irqsave(&priv->lock, flags); 240 list_move(&p->list, &priv->cm.rx_flush_list); 241 p->state = IPOIB_CM_RX_FLUSH; 242 ipoib_cm_start_rx_drain(priv); 243 spin_unlock_irqrestore(&priv->lock, flags); 244} 245 246static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev, 247 struct ipoib_cm_rx *p) 248{ 249 struct ipoib_dev_priv *priv = netdev_priv(dev); 250 struct ib_qp_init_attr attr = { 251 .event_handler = ipoib_cm_rx_event_handler, 252 .send_cq = priv->recv_cq, /* For drain WR */ 253 .recv_cq = priv->recv_cq, 254 .srq = priv->cm.srq, 255 .cap.max_send_wr = 1, /* For drain WR */ 256 .cap.max_send_sge = 1, 257 .sq_sig_type = IB_SIGNAL_ALL_WR, 258 .qp_type = IB_QPT_RC, 259 .qp_context = p, 260 }; 261 262 if (!ipoib_cm_has_srq(dev)) { 263 attr.cap.max_recv_wr = ipoib_recvq_size; 264 attr.cap.max_recv_sge = IPOIB_CM_RX_SG; 265 } 266 267 return ib_create_qp(priv->pd, &attr); 268} 269 270static int ipoib_cm_modify_rx_qp(struct net_device *dev, 271 struct ib_cm_id *cm_id, struct ib_qp *qp, 272 unsigned psn) 273{ 274 struct ipoib_dev_priv *priv = netdev_priv(dev); 275 struct ib_qp_attr qp_attr; 276 int qp_attr_mask, ret; 277 278 qp_attr.qp_state = IB_QPS_INIT; 279 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 280 if (ret) { 281 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret); 282 return ret; 283 } 284 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 285 if (ret) { 286 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret); 287 return ret; 288 } 289 qp_attr.qp_state = IB_QPS_RTR; 290 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 291 if (ret) { 292 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret); 293 return ret; 294 } 295 qp_attr.rq_psn = psn; 296 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 297 if (ret) { 298 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret); 299 return ret; 300 } 301 302 qp_attr.qp_state = IB_QPS_RTS; 303 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 304 if (ret) { 305 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 306 return 0; 307 } 308 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 309 if (ret) { 310 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 311 return 0; 312 } 313 314 return 0; 315} 316 317static void ipoib_cm_init_rx_wr(struct net_device *dev, 318 struct ib_recv_wr *wr, 319 struct ib_sge *sge) 320{ 321 struct ipoib_dev_priv *priv = netdev_priv(dev); 322 int i; 323 324 for (i = 0; i < priv->cm.num_frags; ++i) 325 sge[i].lkey = priv->mr->lkey; 326 327 sge[0].length = IPOIB_CM_HEAD_SIZE; 328 for (i = 1; i < priv->cm.num_frags; ++i) 329 sge[i].length = PAGE_SIZE; 330 331 wr->next = NULL; 332 wr->sg_list = sge; 333 wr->num_sge = priv->cm.num_frags; 334} 335 336static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id, 337 struct ipoib_cm_rx *rx) 338{ 339 struct ipoib_dev_priv *priv = netdev_priv(dev); 340 struct { 341 struct ib_recv_wr wr; 342 struct ib_sge sge[IPOIB_CM_RX_SG]; 343 } *t; 344 int ret; 345 int i; 346 347 rx->rx_ring = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring); 348 if (!rx->rx_ring) { 349 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n", 350 priv->ca->name, ipoib_recvq_size); 351 return -ENOMEM; 352 } 353 354 memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring); 355 356 t = kmalloc(sizeof *t, GFP_KERNEL); 357 if (!t) { 358 ret = -ENOMEM; 359 goto err_free; 360 } 361 362 ipoib_cm_init_rx_wr(dev, &t->wr, t->sge); 363 364 spin_lock_irq(&priv->lock); 365 366 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) { 367 spin_unlock_irq(&priv->lock); 368 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0); 369 ret = -EINVAL; 370 goto err_free; 371 } else 372 ++priv->cm.nonsrq_conn_qp; 373 374 spin_unlock_irq(&priv->lock); 375 376 for (i = 0; i < ipoib_recvq_size; ++i) { 377 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1, 378 rx->rx_ring[i].mapping)) { 379 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 380 ret = -ENOMEM; 381 goto err_count; 382 } 383 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i); 384 if (ret) { 385 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq " 386 "failed for buf %d\n", i); 387 ret = -EIO; 388 goto err_count; 389 } 390 } 391 392 rx->recv_count = ipoib_recvq_size; 393 394 kfree(t); 395 396 return 0; 397 398err_count: 399 spin_lock_irq(&priv->lock); 400 --priv->cm.nonsrq_conn_qp; 401 spin_unlock_irq(&priv->lock); 402 403err_free: 404 kfree(t); 405 ipoib_cm_free_rx_ring(dev, rx->rx_ring); 406 407 return ret; 408} 409 410static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id, 411 struct ib_qp *qp, struct ib_cm_req_event_param *req, 412 unsigned psn) 413{ 414 struct ipoib_dev_priv *priv = netdev_priv(dev); 415 struct ipoib_cm_data data = {}; 416 struct ib_cm_rep_param rep = {}; 417 418 data.qpn = cpu_to_be32(priv->qp->qp_num); 419 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE); 420 421 rep.private_data = &data; 422 rep.private_data_len = sizeof data; 423 rep.flow_control = 0; 424 rep.rnr_retry_count = req->rnr_retry_count; 425 rep.srq = ipoib_cm_has_srq(dev); 426 rep.qp_num = qp->qp_num; 427 rep.starting_psn = psn; 428 return ib_send_cm_rep(cm_id, &rep); 429} 430 431static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 432{ 433 struct net_device *dev = cm_id->context; 434 struct ipoib_dev_priv *priv = netdev_priv(dev); 435 struct ipoib_cm_rx *p; 436 unsigned psn; 437 int ret; 438 439 ipoib_dbg(priv, "REQ arrived\n"); 440 p = kzalloc(sizeof *p, GFP_KERNEL); 441 if (!p) 442 return -ENOMEM; 443 p->dev = dev; 444 p->id = cm_id; 445 cm_id->context = p; 446 p->state = IPOIB_CM_RX_LIVE; 447 p->jiffies = jiffies; 448 INIT_LIST_HEAD(&p->list); 449 450 p->qp = ipoib_cm_create_rx_qp(dev, p); 451 if (IS_ERR(p->qp)) { 452 ret = PTR_ERR(p->qp); 453 goto err_qp; 454 } 455 456 psn = random32() & 0xffffff; 457 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn); 458 if (ret) 459 goto err_modify; 460 461 if (!ipoib_cm_has_srq(dev)) { 462 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p); 463 if (ret) 464 goto err_modify; 465 } 466 467 spin_lock_irq(&priv->lock); 468 queue_delayed_work(ipoib_workqueue, 469 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 470 /* Add this entry to passive ids list head, but do not re-add it 471 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */ 472 p->jiffies = jiffies; 473 if (p->state == IPOIB_CM_RX_LIVE) 474 list_move(&p->list, &priv->cm.passive_ids); 475 spin_unlock_irq(&priv->lock); 476 477 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn); 478 if (ret) { 479 ipoib_warn(priv, "failed to send REP: %d\n", ret); 480 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 481 ipoib_warn(priv, "unable to move qp to error state\n"); 482 } 483 return 0; 484 485err_modify: 486 ib_destroy_qp(p->qp); 487err_qp: 488 kfree(p); 489 return ret; 490} 491 492static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id, 493 struct ib_cm_event *event) 494{ 495 struct ipoib_cm_rx *p; 496 struct ipoib_dev_priv *priv; 497 498 switch (event->event) { 499 case IB_CM_REQ_RECEIVED: 500 return ipoib_cm_req_handler(cm_id, event); 501 case IB_CM_DREQ_RECEIVED: 502 p = cm_id->context; 503 ib_send_cm_drep(cm_id, NULL, 0); 504 /* Fall through */ 505 case IB_CM_REJ_RECEIVED: 506 p = cm_id->context; 507 priv = netdev_priv(p->dev); 508 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 509 ipoib_warn(priv, "unable to move qp to error state\n"); 510 /* Fall through */ 511 default: 512 return 0; 513 } 514} 515/* Adjust length of skb with fragments to match received data */ 516static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space, 517 unsigned int length, struct sk_buff *toskb) 518{ 519 int i, num_frags; 520 unsigned int size; 521 522 /* put header into skb */ 523 size = min(length, hdr_space); 524 skb->tail += size; 525 skb->len += size; 526 length -= size; 527 528 num_frags = skb_shinfo(skb)->nr_frags; 529 for (i = 0; i < num_frags; i++) { 530 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 531 532 if (length == 0) { 533 /* don't need this page */ 534 skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE); 535 --skb_shinfo(skb)->nr_frags; 536 } else { 537 size = min(length, (unsigned) PAGE_SIZE); 538 539 frag->size = size; 540 skb->data_len += size; 541 skb->truesize += size; 542 skb->len += size; 543 length -= size; 544 } 545 } 546} 547 548void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc) 549{ 550 struct ipoib_dev_priv *priv = netdev_priv(dev); 551 struct ipoib_cm_rx_buf *rx_ring; 552 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV); 553 struct sk_buff *skb, *newskb; 554 struct ipoib_cm_rx *p; 555 unsigned long flags; 556 u64 mapping[IPOIB_CM_RX_SG]; 557 int frags; 558 int has_srq; 559 struct sk_buff *small_skb; 560 561 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n", 562 wr_id, wc->status); 563 564 if (unlikely(wr_id >= ipoib_recvq_size)) { 565 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) { 566 spin_lock_irqsave(&priv->lock, flags); 567 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list); 568 ipoib_cm_start_rx_drain(priv); 569 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task); 570 spin_unlock_irqrestore(&priv->lock, flags); 571 } else 572 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n", 573 wr_id, ipoib_recvq_size); 574 return; 575 } 576 577 p = wc->qp->qp_context; 578 579 has_srq = ipoib_cm_has_srq(dev); 580 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring; 581 582 skb = rx_ring[wr_id].skb; 583 584 if (unlikely(wc->status != IB_WC_SUCCESS)) { 585 ipoib_dbg(priv, "cm recv error " 586 "(status=%d, wrid=%d vend_err %x)\n", 587 wc->status, wr_id, wc->vendor_err); 588 ++dev->stats.rx_dropped; 589 if (has_srq) 590 goto repost; 591 else { 592 if (!--p->recv_count) { 593 spin_lock_irqsave(&priv->lock, flags); 594 list_move(&p->list, &priv->cm.rx_reap_list); 595 spin_unlock_irqrestore(&priv->lock, flags); 596 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task); 597 } 598 return; 599 } 600 } 601 602 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) { 603 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) { 604 spin_lock_irqsave(&priv->lock, flags); 605 p->jiffies = jiffies; 606 /* Move this entry to list head, but do not re-add it 607 * if it has been moved out of list. */ 608 if (p->state == IPOIB_CM_RX_LIVE) 609 list_move(&p->list, &priv->cm.passive_ids); 610 spin_unlock_irqrestore(&priv->lock, flags); 611 } 612 } 613 614 if (wc->byte_len < IPOIB_CM_COPYBREAK) { 615 int dlen = wc->byte_len; 616 617 small_skb = dev_alloc_skb(dlen + 12); 618 if (small_skb) { 619 skb_reserve(small_skb, 12); 620 ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0], 621 dlen, DMA_FROM_DEVICE); 622 skb_copy_from_linear_data(skb, small_skb->data, dlen); 623 ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0], 624 dlen, DMA_FROM_DEVICE); 625 skb_put(small_skb, dlen); 626 skb = small_skb; 627 goto copied; 628 } 629 } 630 631 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len, 632 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE; 633 634 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping); 635 if (unlikely(!newskb)) { 636 /* 637 * If we can't allocate a new RX buffer, dump 638 * this packet and reuse the old buffer. 639 */ 640 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id); 641 ++dev->stats.rx_dropped; 642 goto repost; 643 } 644 645 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping); 646 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping); 647 648 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 649 wc->byte_len, wc->slid); 650 651 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb); 652 653copied: 654 skb->protocol = ((struct ipoib_header *) skb->data)->proto; 655 skb_reset_mac_header(skb); 656 skb_pull(skb, IPOIB_ENCAP_LEN); 657 658 ++dev->stats.rx_packets; 659 dev->stats.rx_bytes += skb->len; 660 661 skb->dev = dev; 662 skb->pkt_type = PACKET_HOST; 663 netif_receive_skb(skb); 664 665repost: 666 if (has_srq) { 667 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id))) 668 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed " 669 "for buf %d\n", wr_id); 670 } else { 671 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, 672 &priv->cm.rx_wr, 673 priv->cm.rx_sge, 674 wr_id))) { 675 --p->recv_count; 676 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed " 677 "for buf %d\n", wr_id); 678 } 679 } 680} 681 682static inline int post_send(struct ipoib_dev_priv *priv, 683 struct ipoib_cm_tx *tx, 684 unsigned int wr_id, 685 u64 addr, int len) 686{ 687 struct ib_send_wr *bad_wr; 688 689 priv->tx_sge[0].addr = addr; 690 priv->tx_sge[0].length = len; 691 692 priv->tx_wr.num_sge = 1; 693 priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM; 694 695 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr); 696} 697 698void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx) 699{ 700 struct ipoib_dev_priv *priv = netdev_priv(dev); 701 struct ipoib_cm_tx_buf *tx_req; 702 u64 addr; 703 int rc; 704 705 if (unlikely(skb->len > tx->mtu)) { 706 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 707 skb->len, tx->mtu); 708 ++dev->stats.tx_dropped; 709 ++dev->stats.tx_errors; 710 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN); 711 return; 712 } 713 714 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n", 715 tx->tx_head, skb->len, tx->qp->qp_num); 716 717 /* 718 * We put the skb into the tx_ring _before_ we call post_send() 719 * because it's entirely possible that the completion handler will 720 * run before we execute anything after the post_send(). That 721 * means we have to make sure everything is properly recorded and 722 * our state is consistent before we call post_send(). 723 */ 724 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)]; 725 tx_req->skb = skb; 726 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE); 727 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) { 728 ++dev->stats.tx_errors; 729 dev_kfree_skb_any(skb); 730 return; 731 } 732 733 tx_req->mapping = addr; 734 735 rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), 736 addr, skb->len); 737 if (unlikely(rc)) { 738 ipoib_warn(priv, "post_send failed, error %d\n", rc); 739 ++dev->stats.tx_errors; 740 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE); 741 dev_kfree_skb_any(skb); 742 } else { 743 dev->trans_start = jiffies; 744 ++tx->tx_head; 745 746 if (++priv->tx_outstanding == ipoib_sendq_size) { 747 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n", 748 tx->qp->qp_num); 749 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP)) 750 ipoib_warn(priv, "request notify on send CQ failed\n"); 751 netif_stop_queue(dev); 752 } 753 } 754} 755 756void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc) 757{ 758 struct ipoib_dev_priv *priv = netdev_priv(dev); 759 struct ipoib_cm_tx *tx = wc->qp->qp_context; 760 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM; 761 struct ipoib_cm_tx_buf *tx_req; 762 unsigned long flags; 763 764 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n", 765 wr_id, wc->status); 766 767 if (unlikely(wr_id >= ipoib_sendq_size)) { 768 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n", 769 wr_id, ipoib_sendq_size); 770 return; 771 } 772 773 tx_req = &tx->tx_ring[wr_id]; 774 775 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE); 776 777 ++dev->stats.tx_packets; 778 dev->stats.tx_bytes += tx_req->skb->len; 779 780 dev_kfree_skb_any(tx_req->skb); 781 782 netif_tx_lock(dev); 783 784 ++tx->tx_tail; 785 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 786 netif_queue_stopped(dev) && 787 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 788 netif_wake_queue(dev); 789 790 if (wc->status != IB_WC_SUCCESS && 791 wc->status != IB_WC_WR_FLUSH_ERR) { 792 struct ipoib_neigh *neigh; 793 794 ipoib_dbg(priv, "failed cm send event " 795 "(status=%d, wrid=%d vend_err %x)\n", 796 wc->status, wr_id, wc->vendor_err); 797 798 spin_lock_irqsave(&priv->lock, flags); 799 neigh = tx->neigh; 800 801 if (neigh) { 802 neigh->cm = NULL; 803 list_del(&neigh->list); 804 if (neigh->ah) 805 ipoib_put_ah(neigh->ah); 806 ipoib_neigh_free(dev, neigh); 807 808 tx->neigh = NULL; 809 } 810 811 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 812 list_move(&tx->list, &priv->cm.reap_list); 813 queue_work(ipoib_workqueue, &priv->cm.reap_task); 814 } 815 816 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags); 817 818 spin_unlock_irqrestore(&priv->lock, flags); 819 } 820 821 netif_tx_unlock(dev); 822} 823 824int ipoib_cm_dev_open(struct net_device *dev) 825{ 826 struct ipoib_dev_priv *priv = netdev_priv(dev); 827 int ret; 828 829 if (!IPOIB_CM_SUPPORTED(dev->dev_addr)) 830 return 0; 831 832 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev); 833 if (IS_ERR(priv->cm.id)) { 834 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name); 835 ret = PTR_ERR(priv->cm.id); 836 goto err_cm; 837 } 838 839 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 840 0, NULL); 841 if (ret) { 842 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name, 843 IPOIB_CM_IETF_ID | priv->qp->qp_num); 844 goto err_listen; 845 } 846 847 return 0; 848 849err_listen: 850 ib_destroy_cm_id(priv->cm.id); 851err_cm: 852 priv->cm.id = NULL; 853 return ret; 854} 855 856static void ipoib_cm_free_rx_reap_list(struct net_device *dev) 857{ 858 struct ipoib_dev_priv *priv = netdev_priv(dev); 859 struct ipoib_cm_rx *rx, *n; 860 LIST_HEAD(list); 861 862 spin_lock_irq(&priv->lock); 863 list_splice_init(&priv->cm.rx_reap_list, &list); 864 spin_unlock_irq(&priv->lock); 865 866 list_for_each_entry_safe(rx, n, &list, list) { 867 ib_destroy_cm_id(rx->id); 868 ib_destroy_qp(rx->qp); 869 if (!ipoib_cm_has_srq(dev)) { 870 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring); 871 spin_lock_irq(&priv->lock); 872 --priv->cm.nonsrq_conn_qp; 873 spin_unlock_irq(&priv->lock); 874 } 875 kfree(rx); 876 } 877} 878 879void ipoib_cm_dev_stop(struct net_device *dev) 880{ 881 struct ipoib_dev_priv *priv = netdev_priv(dev); 882 struct ipoib_cm_rx *p; 883 unsigned long begin; 884 int ret; 885 886 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id) 887 return; 888 889 ib_destroy_cm_id(priv->cm.id); 890 priv->cm.id = NULL; 891 892 spin_lock_irq(&priv->lock); 893 while (!list_empty(&priv->cm.passive_ids)) { 894 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list); 895 list_move(&p->list, &priv->cm.rx_error_list); 896 p->state = IPOIB_CM_RX_ERROR; 897 spin_unlock_irq(&priv->lock); 898 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 899 if (ret) 900 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 901 spin_lock_irq(&priv->lock); 902 } 903 904 /* Wait for all RX to be drained */ 905 begin = jiffies; 906 907 while (!list_empty(&priv->cm.rx_error_list) || 908 !list_empty(&priv->cm.rx_flush_list) || 909 !list_empty(&priv->cm.rx_drain_list)) { 910 if (time_after(jiffies, begin + 5 * HZ)) { 911 ipoib_warn(priv, "RX drain timing out\n"); 912 913 /* 914 * assume the HW is wedged and just free up everything. 915 */ 916 list_splice_init(&priv->cm.rx_flush_list, 917 &priv->cm.rx_reap_list); 918 list_splice_init(&priv->cm.rx_error_list, 919 &priv->cm.rx_reap_list); 920 list_splice_init(&priv->cm.rx_drain_list, 921 &priv->cm.rx_reap_list); 922 break; 923 } 924 spin_unlock_irq(&priv->lock); 925 msleep(1); 926 ipoib_drain_cq(dev); 927 spin_lock_irq(&priv->lock); 928 } 929 930 spin_unlock_irq(&priv->lock); 931 932 ipoib_cm_free_rx_reap_list(dev); 933 934 cancel_delayed_work(&priv->cm.stale_task); 935} 936 937static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 938{ 939 struct ipoib_cm_tx *p = cm_id->context; 940 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 941 struct ipoib_cm_data *data = event->private_data; 942 struct sk_buff_head skqueue; 943 struct ib_qp_attr qp_attr; 944 int qp_attr_mask, ret; 945 struct sk_buff *skb; 946 947 p->mtu = be32_to_cpu(data->mtu); 948 949 if (p->mtu <= IPOIB_ENCAP_LEN) { 950 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n", 951 p->mtu, IPOIB_ENCAP_LEN); 952 return -EINVAL; 953 } 954 955 qp_attr.qp_state = IB_QPS_RTR; 956 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 957 if (ret) { 958 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret); 959 return ret; 960 } 961 962 qp_attr.rq_psn = 0; 963 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 964 if (ret) { 965 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret); 966 return ret; 967 } 968 969 qp_attr.qp_state = IB_QPS_RTS; 970 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 971 if (ret) { 972 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 973 return ret; 974 } 975 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 976 if (ret) { 977 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 978 return ret; 979 } 980 981 skb_queue_head_init(&skqueue); 982 983 spin_lock_irq(&priv->lock); 984 set_bit(IPOIB_FLAG_OPER_UP, &p->flags); 985 if (p->neigh) 986 while ((skb = __skb_dequeue(&p->neigh->queue))) 987 __skb_queue_tail(&skqueue, skb); 988 spin_unlock_irq(&priv->lock); 989 990 while ((skb = __skb_dequeue(&skqueue))) { 991 skb->dev = p->dev; 992 if (dev_queue_xmit(skb)) 993 ipoib_warn(priv, "dev_queue_xmit failed " 994 "to requeue packet\n"); 995 } 996 997 ret = ib_send_cm_rtu(cm_id, NULL, 0); 998 if (ret) { 999 ipoib_warn(priv, "failed to send RTU: %d\n", ret); 1000 return ret; 1001 } 1002 return 0; 1003} 1004 1005static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx) 1006{ 1007 struct ipoib_dev_priv *priv = netdev_priv(dev); 1008 struct ib_qp_init_attr attr = { 1009 .send_cq = priv->recv_cq, 1010 .recv_cq = priv->recv_cq, 1011 .srq = priv->cm.srq, 1012 .cap.max_send_wr = ipoib_sendq_size, 1013 .cap.max_send_sge = 1, 1014 .sq_sig_type = IB_SIGNAL_ALL_WR, 1015 .qp_type = IB_QPT_RC, 1016 .qp_context = tx 1017 }; 1018 1019 return ib_create_qp(priv->pd, &attr); 1020} 1021 1022static int ipoib_cm_send_req(struct net_device *dev, 1023 struct ib_cm_id *id, struct ib_qp *qp, 1024 u32 qpn, 1025 struct ib_sa_path_rec *pathrec) 1026{ 1027 struct ipoib_dev_priv *priv = netdev_priv(dev); 1028 struct ipoib_cm_data data = {}; 1029 struct ib_cm_req_param req = {}; 1030 1031 data.qpn = cpu_to_be32(priv->qp->qp_num); 1032 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE); 1033 1034 req.primary_path = pathrec; 1035 req.alternate_path = NULL; 1036 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn); 1037 req.qp_num = qp->qp_num; 1038 req.qp_type = qp->qp_type; 1039 req.private_data = &data; 1040 req.private_data_len = sizeof data; 1041 req.flow_control = 0; 1042 1043 req.starting_psn = 0; 1044 1045 /* 1046 * Pick some arbitrary defaults here; we could make these 1047 * module parameters if anyone cared about setting them. 1048 */ 1049 req.responder_resources = 4; 1050 req.remote_cm_response_timeout = 20; 1051 req.local_cm_response_timeout = 20; 1052 req.retry_count = 0; /* RFC draft warns against retries */ 1053 req.rnr_retry_count = 0; /* RFC draft warns against retries */ 1054 req.max_cm_retries = 15; 1055 req.srq = ipoib_cm_has_srq(dev); 1056 return ib_send_cm_req(id, &req); 1057} 1058 1059static int ipoib_cm_modify_tx_init(struct net_device *dev, 1060 struct ib_cm_id *cm_id, struct ib_qp *qp) 1061{ 1062 struct ipoib_dev_priv *priv = netdev_priv(dev); 1063 struct ib_qp_attr qp_attr; 1064 int qp_attr_mask, ret; 1065 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index); 1066 if (ret) { 1067 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret); 1068 return ret; 1069 } 1070 1071 qp_attr.qp_state = IB_QPS_INIT; 1072 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE; 1073 qp_attr.port_num = priv->port; 1074 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT; 1075 1076 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 1077 if (ret) { 1078 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret); 1079 return ret; 1080 } 1081 return 0; 1082} 1083 1084static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn, 1085 struct ib_sa_path_rec *pathrec) 1086{ 1087 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 1088 int ret; 1089 1090 p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring); 1091 if (!p->tx_ring) { 1092 ipoib_warn(priv, "failed to allocate tx ring\n"); 1093 ret = -ENOMEM; 1094 goto err_tx; 1095 } 1096 memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring); 1097 1098 p->qp = ipoib_cm_create_tx_qp(p->dev, p); 1099 if (IS_ERR(p->qp)) { 1100 ret = PTR_ERR(p->qp); 1101 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret); 1102 goto err_qp; 1103 } 1104 1105 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p); 1106 if (IS_ERR(p->id)) { 1107 ret = PTR_ERR(p->id); 1108 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret); 1109 goto err_id; 1110 } 1111 1112 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp); 1113 if (ret) { 1114 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret); 1115 goto err_modify; 1116 } 1117 1118 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec); 1119 if (ret) { 1120 ipoib_warn(priv, "failed to send cm req: %d\n", ret); 1121 goto err_send_cm; 1122 } 1123 1124 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n", 1125 p->qp->qp_num, pathrec->dgid.raw, qpn); 1126 1127 return 0; 1128 1129err_send_cm: 1130err_modify: 1131 ib_destroy_cm_id(p->id); 1132err_id: 1133 p->id = NULL; 1134 ib_destroy_qp(p->qp); 1135err_qp: 1136 p->qp = NULL; 1137 vfree(p->tx_ring); 1138err_tx: 1139 return ret; 1140} 1141 1142static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p) 1143{ 1144 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 1145 struct ipoib_cm_tx_buf *tx_req; 1146 unsigned long begin; 1147 1148 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n", 1149 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail); 1150 1151 if (p->id) 1152 ib_destroy_cm_id(p->id); 1153 1154 if (p->tx_ring) { 1155 /* Wait for all sends to complete */ 1156 begin = jiffies; 1157 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1158 if (time_after(jiffies, begin + 5 * HZ)) { 1159 ipoib_warn(priv, "timing out; %d sends not completed\n", 1160 p->tx_head - p->tx_tail); 1161 goto timeout; 1162 } 1163 1164 msleep(1); 1165 } 1166 } 1167 1168timeout: 1169 1170 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1171 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)]; 1172 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, 1173 DMA_TO_DEVICE); 1174 dev_kfree_skb_any(tx_req->skb); 1175 ++p->tx_tail; 1176 netif_tx_lock_bh(p->dev); 1177 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 1178 netif_queue_stopped(p->dev) && 1179 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 1180 netif_wake_queue(p->dev); 1181 netif_tx_unlock_bh(p->dev); 1182 } 1183 1184 if (p->qp) 1185 ib_destroy_qp(p->qp); 1186 1187 vfree(p->tx_ring); 1188 kfree(p); 1189} 1190 1191static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id, 1192 struct ib_cm_event *event) 1193{ 1194 struct ipoib_cm_tx *tx = cm_id->context; 1195 struct ipoib_dev_priv *priv = netdev_priv(tx->dev); 1196 struct net_device *dev = priv->dev; 1197 struct ipoib_neigh *neigh; 1198 unsigned long flags; 1199 int ret; 1200 1201 switch (event->event) { 1202 case IB_CM_DREQ_RECEIVED: 1203 ipoib_dbg(priv, "DREQ received.\n"); 1204 ib_send_cm_drep(cm_id, NULL, 0); 1205 break; 1206 case IB_CM_REP_RECEIVED: 1207 ipoib_dbg(priv, "REP received.\n"); 1208 ret = ipoib_cm_rep_handler(cm_id, event); 1209 if (ret) 1210 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, 1211 NULL, 0, NULL, 0); 1212 break; 1213 case IB_CM_REQ_ERROR: 1214 case IB_CM_REJ_RECEIVED: 1215 case IB_CM_TIMEWAIT_EXIT: 1216 ipoib_dbg(priv, "CM error %d.\n", event->event); 1217 netif_tx_lock_bh(dev); 1218 spin_lock_irqsave(&priv->lock, flags); 1219 neigh = tx->neigh; 1220 1221 if (neigh) { 1222 neigh->cm = NULL; 1223 list_del(&neigh->list); 1224 if (neigh->ah) 1225 ipoib_put_ah(neigh->ah); 1226 ipoib_neigh_free(dev, neigh); 1227 1228 tx->neigh = NULL; 1229 } 1230 1231 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1232 list_move(&tx->list, &priv->cm.reap_list); 1233 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1234 } 1235 1236 spin_unlock_irqrestore(&priv->lock, flags); 1237 netif_tx_unlock_bh(dev); 1238 break; 1239 default: 1240 break; 1241 } 1242 1243 return 0; 1244} 1245 1246struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path, 1247 struct ipoib_neigh *neigh) 1248{ 1249 struct ipoib_dev_priv *priv = netdev_priv(dev); 1250 struct ipoib_cm_tx *tx; 1251 1252 tx = kzalloc(sizeof *tx, GFP_ATOMIC); 1253 if (!tx) 1254 return NULL; 1255 1256 neigh->cm = tx; 1257 tx->neigh = neigh; 1258 tx->path = path; 1259 tx->dev = dev; 1260 list_add(&tx->list, &priv->cm.start_list); 1261 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags); 1262 queue_work(ipoib_workqueue, &priv->cm.start_task); 1263 return tx; 1264} 1265 1266void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx) 1267{ 1268 struct ipoib_dev_priv *priv = netdev_priv(tx->dev); 1269 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1270 list_move(&tx->list, &priv->cm.reap_list); 1271 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1272 ipoib_dbg(priv, "Reap connection for gid %pI6\n", 1273 tx->neigh->dgid.raw); 1274 tx->neigh = NULL; 1275 } 1276} 1277 1278static void ipoib_cm_tx_start(struct work_struct *work) 1279{ 1280 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1281 cm.start_task); 1282 struct net_device *dev = priv->dev; 1283 struct ipoib_neigh *neigh; 1284 struct ipoib_cm_tx *p; 1285 unsigned long flags; 1286 int ret; 1287 1288 struct ib_sa_path_rec pathrec; 1289 u32 qpn; 1290 1291 netif_tx_lock_bh(dev); 1292 spin_lock_irqsave(&priv->lock, flags); 1293 1294 while (!list_empty(&priv->cm.start_list)) { 1295 p = list_entry(priv->cm.start_list.next, typeof(*p), list); 1296 list_del_init(&p->list); 1297 neigh = p->neigh; 1298 qpn = IPOIB_QPN(neigh->neighbour->ha); 1299 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec); 1300 1301 spin_unlock_irqrestore(&priv->lock, flags); 1302 netif_tx_unlock_bh(dev); 1303 1304 ret = ipoib_cm_tx_init(p, qpn, &pathrec); 1305 1306 netif_tx_lock_bh(dev); 1307 spin_lock_irqsave(&priv->lock, flags); 1308 1309 if (ret) { 1310 neigh = p->neigh; 1311 if (neigh) { 1312 neigh->cm = NULL; 1313 list_del(&neigh->list); 1314 if (neigh->ah) 1315 ipoib_put_ah(neigh->ah); 1316 ipoib_neigh_free(dev, neigh); 1317 } 1318 list_del(&p->list); 1319 kfree(p); 1320 } 1321 } 1322 1323 spin_unlock_irqrestore(&priv->lock, flags); 1324 netif_tx_unlock_bh(dev); 1325} 1326 1327static void ipoib_cm_tx_reap(struct work_struct *work) 1328{ 1329 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1330 cm.reap_task); 1331 struct net_device *dev = priv->dev; 1332 struct ipoib_cm_tx *p; 1333 unsigned long flags; 1334 1335 netif_tx_lock_bh(dev); 1336 spin_lock_irqsave(&priv->lock, flags); 1337 1338 while (!list_empty(&priv->cm.reap_list)) { 1339 p = list_entry(priv->cm.reap_list.next, typeof(*p), list); 1340 list_del(&p->list); 1341 spin_unlock_irqrestore(&priv->lock, flags); 1342 netif_tx_unlock_bh(dev); 1343 ipoib_cm_tx_destroy(p); 1344 netif_tx_lock_bh(dev); 1345 spin_lock_irqsave(&priv->lock, flags); 1346 } 1347 1348 spin_unlock_irqrestore(&priv->lock, flags); 1349 netif_tx_unlock_bh(dev); 1350} 1351 1352static void ipoib_cm_skb_reap(struct work_struct *work) 1353{ 1354 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1355 cm.skb_task); 1356 struct net_device *dev = priv->dev; 1357 struct sk_buff *skb; 1358 unsigned long flags; 1359 unsigned mtu = priv->mcast_mtu; 1360 1361 netif_tx_lock_bh(dev); 1362 spin_lock_irqsave(&priv->lock, flags); 1363 1364 while ((skb = skb_dequeue(&priv->cm.skb_queue))) { 1365 spin_unlock_irqrestore(&priv->lock, flags); 1366 netif_tx_unlock_bh(dev); 1367 1368 if (skb->protocol == htons(ETH_P_IP)) 1369 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu)); 1370#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 1371 else if (skb->protocol == htons(ETH_P_IPV6)) 1372 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu); 1373#endif 1374 dev_kfree_skb_any(skb); 1375 1376 netif_tx_lock_bh(dev); 1377 spin_lock_irqsave(&priv->lock, flags); 1378 } 1379 1380 spin_unlock_irqrestore(&priv->lock, flags); 1381 netif_tx_unlock_bh(dev); 1382} 1383 1384void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb, 1385 unsigned int mtu) 1386{ 1387 struct ipoib_dev_priv *priv = netdev_priv(dev); 1388 int e = skb_queue_empty(&priv->cm.skb_queue); 1389 1390 if (skb_dst(skb)) 1391 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu); 1392 1393 skb_queue_tail(&priv->cm.skb_queue, skb); 1394 if (e) 1395 queue_work(ipoib_workqueue, &priv->cm.skb_task); 1396} 1397 1398static void ipoib_cm_rx_reap(struct work_struct *work) 1399{ 1400 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv, 1401 cm.rx_reap_task)->dev); 1402} 1403 1404static void ipoib_cm_stale_task(struct work_struct *work) 1405{ 1406 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1407 cm.stale_task.work); 1408 struct ipoib_cm_rx *p; 1409 int ret; 1410 1411 spin_lock_irq(&priv->lock); 1412 while (!list_empty(&priv->cm.passive_ids)) { 1413 /* List is sorted by LRU, start from tail, 1414 * stop when we see a recently used entry */ 1415 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list); 1416 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT)) 1417 break; 1418 list_move(&p->list, &priv->cm.rx_error_list); 1419 p->state = IPOIB_CM_RX_ERROR; 1420 spin_unlock_irq(&priv->lock); 1421 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 1422 if (ret) 1423 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 1424 spin_lock_irq(&priv->lock); 1425 } 1426 1427 if (!list_empty(&priv->cm.passive_ids)) 1428 queue_delayed_work(ipoib_workqueue, 1429 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 1430 spin_unlock_irq(&priv->lock); 1431} 1432 1433 1434static ssize_t show_mode(struct device *d, struct device_attribute *attr, 1435 char *buf) 1436{ 1437 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d)); 1438 1439 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags)) 1440 return sprintf(buf, "connected\n"); 1441 else 1442 return sprintf(buf, "datagram\n"); 1443} 1444 1445static ssize_t set_mode(struct device *d, struct device_attribute *attr, 1446 const char *buf, size_t count) 1447{ 1448 struct net_device *dev = to_net_dev(d); 1449 struct ipoib_dev_priv *priv = netdev_priv(dev); 1450 1451 if (!rtnl_trylock()) 1452 return restart_syscall(); 1453 1454 /* flush paths if we switch modes so that connections are restarted */ 1455 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) { 1456 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 1457 ipoib_warn(priv, "enabling connected mode " 1458 "will cause multicast packet drops\n"); 1459 1460 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO); 1461 rtnl_unlock(); 1462 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM; 1463 1464 ipoib_flush_paths(dev); 1465 return count; 1466 } 1467 1468 if (!strcmp(buf, "datagram\n")) { 1469 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 1470 1471 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) { 1472 dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG; 1473 if (priv->hca_caps & IB_DEVICE_UD_TSO) 1474 dev->features |= NETIF_F_TSO; 1475 } 1476 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu)); 1477 rtnl_unlock(); 1478 ipoib_flush_paths(dev); 1479 1480 return count; 1481 } 1482 rtnl_unlock(); 1483 1484 return -EINVAL; 1485} 1486 1487static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode); 1488 1489int ipoib_cm_add_mode_attr(struct net_device *dev) 1490{ 1491 return device_create_file(&dev->dev, &dev_attr_mode); 1492} 1493 1494static void ipoib_cm_create_srq(struct net_device *dev, int max_sge) 1495{ 1496 struct ipoib_dev_priv *priv = netdev_priv(dev); 1497 struct ib_srq_init_attr srq_init_attr = { 1498 .attr = { 1499 .max_wr = ipoib_recvq_size, 1500 .max_sge = max_sge 1501 } 1502 }; 1503 1504 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr); 1505 if (IS_ERR(priv->cm.srq)) { 1506 if (PTR_ERR(priv->cm.srq) != -ENOSYS) 1507 printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n", 1508 priv->ca->name, PTR_ERR(priv->cm.srq)); 1509 priv->cm.srq = NULL; 1510 return; 1511 } 1512 1513 priv->cm.srq_ring = vmalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring); 1514 if (!priv->cm.srq_ring) { 1515 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n", 1516 priv->ca->name, ipoib_recvq_size); 1517 ib_destroy_srq(priv->cm.srq); 1518 priv->cm.srq = NULL; 1519 return; 1520 } 1521 1522 memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring); 1523} 1524 1525int ipoib_cm_dev_init(struct net_device *dev) 1526{ 1527 struct ipoib_dev_priv *priv = netdev_priv(dev); 1528 int i, ret; 1529 struct ib_device_attr attr; 1530 1531 INIT_LIST_HEAD(&priv->cm.passive_ids); 1532 INIT_LIST_HEAD(&priv->cm.reap_list); 1533 INIT_LIST_HEAD(&priv->cm.start_list); 1534 INIT_LIST_HEAD(&priv->cm.rx_error_list); 1535 INIT_LIST_HEAD(&priv->cm.rx_flush_list); 1536 INIT_LIST_HEAD(&priv->cm.rx_drain_list); 1537 INIT_LIST_HEAD(&priv->cm.rx_reap_list); 1538 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start); 1539 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap); 1540 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap); 1541 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap); 1542 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task); 1543 1544 skb_queue_head_init(&priv->cm.skb_queue); 1545 1546 ret = ib_query_device(priv->ca, &attr); 1547 if (ret) { 1548 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret); 1549 return ret; 1550 } 1551 1552 ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge); 1553 1554 attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge); 1555 ipoib_cm_create_srq(dev, attr.max_srq_sge); 1556 if (ipoib_cm_has_srq(dev)) { 1557 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10; 1558 priv->cm.num_frags = attr.max_srq_sge; 1559 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n", 1560 priv->cm.max_cm_mtu, priv->cm.num_frags); 1561 } else { 1562 priv->cm.max_cm_mtu = IPOIB_CM_MTU; 1563 priv->cm.num_frags = IPOIB_CM_RX_SG; 1564 } 1565 1566 ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge); 1567 1568 if (ipoib_cm_has_srq(dev)) { 1569 for (i = 0; i < ipoib_recvq_size; ++i) { 1570 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i, 1571 priv->cm.num_frags - 1, 1572 priv->cm.srq_ring[i].mapping)) { 1573 ipoib_warn(priv, "failed to allocate " 1574 "receive buffer %d\n", i); 1575 ipoib_cm_dev_cleanup(dev); 1576 return -ENOMEM; 1577 } 1578 1579 if (ipoib_cm_post_receive_srq(dev, i)) { 1580 ipoib_warn(priv, "ipoib_cm_post_receive_srq " 1581 "failed for buf %d\n", i); 1582 ipoib_cm_dev_cleanup(dev); 1583 return -EIO; 1584 } 1585 } 1586 } 1587 1588 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC; 1589 return 0; 1590} 1591 1592void ipoib_cm_dev_cleanup(struct net_device *dev) 1593{ 1594 struct ipoib_dev_priv *priv = netdev_priv(dev); 1595 int ret; 1596 1597 if (!priv->cm.srq) 1598 return; 1599 1600 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n"); 1601 1602 ret = ib_destroy_srq(priv->cm.srq); 1603 if (ret) 1604 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret); 1605 1606 priv->cm.srq = NULL; 1607 if (!priv->cm.srq_ring) 1608 return; 1609 1610 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring); 1611 priv->cm.srq_ring = NULL; 1612} 1613