ipoib_cm.c revision 451f14439847db302e5104c44458b2dbb4b1829d
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 <rdma/ib_cache.h> 35#include <net/dst.h> 36#include <net/icmp.h> 37#include <linux/icmpv6.h> 38#include <linux/delay.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, /* FIXME: 0 Seems not to work */ 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 /* 303 * Current Mellanox HCA firmware won't generate completions 304 * with error for drain WRs unless the QP has been moved to 305 * RTS first. This work-around leaves a window where a QP has 306 * moved to error asynchronously, but this will eventually get 307 * fixed in firmware, so let's not error out if modify QP 308 * fails. 309 */ 310 qp_attr.qp_state = IB_QPS_RTS; 311 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 312 if (ret) { 313 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 314 return 0; 315 } 316 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 317 if (ret) { 318 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 319 return 0; 320 } 321 322 return 0; 323} 324 325static void ipoib_cm_init_rx_wr(struct net_device *dev, 326 struct ib_recv_wr *wr, 327 struct ib_sge *sge) 328{ 329 struct ipoib_dev_priv *priv = netdev_priv(dev); 330 int i; 331 332 for (i = 0; i < priv->cm.num_frags; ++i) 333 sge[i].lkey = priv->mr->lkey; 334 335 sge[0].length = IPOIB_CM_HEAD_SIZE; 336 for (i = 1; i < priv->cm.num_frags; ++i) 337 sge[i].length = PAGE_SIZE; 338 339 wr->next = NULL; 340 wr->sg_list = sge; 341 wr->num_sge = priv->cm.num_frags; 342} 343 344static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id, 345 struct ipoib_cm_rx *rx) 346{ 347 struct ipoib_dev_priv *priv = netdev_priv(dev); 348 struct { 349 struct ib_recv_wr wr; 350 struct ib_sge sge[IPOIB_CM_RX_SG]; 351 } *t; 352 int ret; 353 int i; 354 355 rx->rx_ring = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring); 356 if (!rx->rx_ring) { 357 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n", 358 priv->ca->name, ipoib_recvq_size); 359 return -ENOMEM; 360 } 361 362 memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring); 363 364 t = kmalloc(sizeof *t, GFP_KERNEL); 365 if (!t) { 366 ret = -ENOMEM; 367 goto err_free; 368 } 369 370 ipoib_cm_init_rx_wr(dev, &t->wr, t->sge); 371 372 spin_lock_irq(&priv->lock); 373 374 if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) { 375 spin_unlock_irq(&priv->lock); 376 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0); 377 ret = -EINVAL; 378 goto err_free; 379 } else 380 ++priv->cm.nonsrq_conn_qp; 381 382 spin_unlock_irq(&priv->lock); 383 384 for (i = 0; i < ipoib_recvq_size; ++i) { 385 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1, 386 rx->rx_ring[i].mapping)) { 387 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 388 ret = -ENOMEM; 389 goto err_count; 390 } 391 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i); 392 if (ret) { 393 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq " 394 "failed for buf %d\n", i); 395 ret = -EIO; 396 goto err_count; 397 } 398 } 399 400 rx->recv_count = ipoib_recvq_size; 401 402 kfree(t); 403 404 return 0; 405 406err_count: 407 spin_lock_irq(&priv->lock); 408 --priv->cm.nonsrq_conn_qp; 409 spin_unlock_irq(&priv->lock); 410 411err_free: 412 kfree(t); 413 ipoib_cm_free_rx_ring(dev, rx->rx_ring); 414 415 return ret; 416} 417 418static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id, 419 struct ib_qp *qp, struct ib_cm_req_event_param *req, 420 unsigned psn) 421{ 422 struct ipoib_dev_priv *priv = netdev_priv(dev); 423 struct ipoib_cm_data data = {}; 424 struct ib_cm_rep_param rep = {}; 425 426 data.qpn = cpu_to_be32(priv->qp->qp_num); 427 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE); 428 429 rep.private_data = &data; 430 rep.private_data_len = sizeof data; 431 rep.flow_control = 0; 432 rep.rnr_retry_count = req->rnr_retry_count; 433 rep.srq = ipoib_cm_has_srq(dev); 434 rep.qp_num = qp->qp_num; 435 rep.starting_psn = psn; 436 return ib_send_cm_rep(cm_id, &rep); 437} 438 439static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 440{ 441 struct net_device *dev = cm_id->context; 442 struct ipoib_dev_priv *priv = netdev_priv(dev); 443 struct ipoib_cm_rx *p; 444 unsigned psn; 445 int ret; 446 447 ipoib_dbg(priv, "REQ arrived\n"); 448 p = kzalloc(sizeof *p, GFP_KERNEL); 449 if (!p) 450 return -ENOMEM; 451 p->dev = dev; 452 p->id = cm_id; 453 cm_id->context = p; 454 p->state = IPOIB_CM_RX_LIVE; 455 p->jiffies = jiffies; 456 INIT_LIST_HEAD(&p->list); 457 458 p->qp = ipoib_cm_create_rx_qp(dev, p); 459 if (IS_ERR(p->qp)) { 460 ret = PTR_ERR(p->qp); 461 goto err_qp; 462 } 463 464 psn = random32() & 0xffffff; 465 ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn); 466 if (ret) 467 goto err_modify; 468 469 if (!ipoib_cm_has_srq(dev)) { 470 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p); 471 if (ret) 472 goto err_modify; 473 } 474 475 spin_lock_irq(&priv->lock); 476 queue_delayed_work(ipoib_workqueue, 477 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 478 /* Add this entry to passive ids list head, but do not re-add it 479 * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */ 480 p->jiffies = jiffies; 481 if (p->state == IPOIB_CM_RX_LIVE) 482 list_move(&p->list, &priv->cm.passive_ids); 483 spin_unlock_irq(&priv->lock); 484 485 ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn); 486 if (ret) { 487 ipoib_warn(priv, "failed to send REP: %d\n", ret); 488 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 489 ipoib_warn(priv, "unable to move qp to error state\n"); 490 } 491 return 0; 492 493err_modify: 494 ib_destroy_qp(p->qp); 495err_qp: 496 kfree(p); 497 return ret; 498} 499 500static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id, 501 struct ib_cm_event *event) 502{ 503 struct ipoib_cm_rx *p; 504 struct ipoib_dev_priv *priv; 505 506 switch (event->event) { 507 case IB_CM_REQ_RECEIVED: 508 return ipoib_cm_req_handler(cm_id, event); 509 case IB_CM_DREQ_RECEIVED: 510 p = cm_id->context; 511 ib_send_cm_drep(cm_id, NULL, 0); 512 /* Fall through */ 513 case IB_CM_REJ_RECEIVED: 514 p = cm_id->context; 515 priv = netdev_priv(p->dev); 516 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE)) 517 ipoib_warn(priv, "unable to move qp to error state\n"); 518 /* Fall through */ 519 default: 520 return 0; 521 } 522} 523/* Adjust length of skb with fragments to match received data */ 524static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space, 525 unsigned int length, struct sk_buff *toskb) 526{ 527 int i, num_frags; 528 unsigned int size; 529 530 /* put header into skb */ 531 size = min(length, hdr_space); 532 skb->tail += size; 533 skb->len += size; 534 length -= size; 535 536 num_frags = skb_shinfo(skb)->nr_frags; 537 for (i = 0; i < num_frags; i++) { 538 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 539 540 if (length == 0) { 541 /* don't need this page */ 542 skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE); 543 --skb_shinfo(skb)->nr_frags; 544 } else { 545 size = min(length, (unsigned) PAGE_SIZE); 546 547 frag->size = size; 548 skb->data_len += size; 549 skb->truesize += size; 550 skb->len += size; 551 length -= size; 552 } 553 } 554} 555 556void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc) 557{ 558 struct ipoib_dev_priv *priv = netdev_priv(dev); 559 struct ipoib_cm_rx_buf *rx_ring; 560 unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV); 561 struct sk_buff *skb, *newskb; 562 struct ipoib_cm_rx *p; 563 unsigned long flags; 564 u64 mapping[IPOIB_CM_RX_SG]; 565 int frags; 566 int has_srq; 567 struct sk_buff *small_skb; 568 569 ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n", 570 wr_id, wc->status); 571 572 if (unlikely(wr_id >= ipoib_recvq_size)) { 573 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) { 574 spin_lock_irqsave(&priv->lock, flags); 575 list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list); 576 ipoib_cm_start_rx_drain(priv); 577 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task); 578 spin_unlock_irqrestore(&priv->lock, flags); 579 } else 580 ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n", 581 wr_id, ipoib_recvq_size); 582 return; 583 } 584 585 p = wc->qp->qp_context; 586 587 has_srq = ipoib_cm_has_srq(dev); 588 rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring; 589 590 skb = rx_ring[wr_id].skb; 591 592 if (unlikely(wc->status != IB_WC_SUCCESS)) { 593 ipoib_dbg(priv, "cm recv error " 594 "(status=%d, wrid=%d vend_err %x)\n", 595 wc->status, wr_id, wc->vendor_err); 596 ++dev->stats.rx_dropped; 597 if (has_srq) 598 goto repost; 599 else { 600 if (!--p->recv_count) { 601 spin_lock_irqsave(&priv->lock, flags); 602 list_move(&p->list, &priv->cm.rx_reap_list); 603 spin_unlock_irqrestore(&priv->lock, flags); 604 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task); 605 } 606 return; 607 } 608 } 609 610 if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) { 611 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) { 612 spin_lock_irqsave(&priv->lock, flags); 613 p->jiffies = jiffies; 614 /* Move this entry to list head, but do not re-add it 615 * if it has been moved out of list. */ 616 if (p->state == IPOIB_CM_RX_LIVE) 617 list_move(&p->list, &priv->cm.passive_ids); 618 spin_unlock_irqrestore(&priv->lock, flags); 619 } 620 } 621 622 if (wc->byte_len < IPOIB_CM_COPYBREAK) { 623 int dlen = wc->byte_len; 624 625 small_skb = dev_alloc_skb(dlen + 12); 626 if (small_skb) { 627 skb_reserve(small_skb, 12); 628 ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0], 629 dlen, DMA_FROM_DEVICE); 630 skb_copy_from_linear_data(skb, small_skb->data, dlen); 631 ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0], 632 dlen, DMA_FROM_DEVICE); 633 skb_put(small_skb, dlen); 634 skb = small_skb; 635 goto copied; 636 } 637 } 638 639 frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len, 640 (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE; 641 642 newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping); 643 if (unlikely(!newskb)) { 644 /* 645 * If we can't allocate a new RX buffer, dump 646 * this packet and reuse the old buffer. 647 */ 648 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id); 649 ++dev->stats.rx_dropped; 650 goto repost; 651 } 652 653 ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping); 654 memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping); 655 656 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 657 wc->byte_len, wc->slid); 658 659 skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb); 660 661copied: 662 skb->protocol = ((struct ipoib_header *) skb->data)->proto; 663 skb_reset_mac_header(skb); 664 skb_pull(skb, IPOIB_ENCAP_LEN); 665 666 ++dev->stats.rx_packets; 667 dev->stats.rx_bytes += skb->len; 668 669 skb->dev = dev; 670 /* XXX get correct PACKET_ type here */ 671 skb->pkt_type = PACKET_HOST; 672 netif_receive_skb(skb); 673 674repost: 675 if (has_srq) { 676 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id))) 677 ipoib_warn(priv, "ipoib_cm_post_receive_srq failed " 678 "for buf %d\n", wr_id); 679 } else { 680 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, 681 &priv->cm.rx_wr, 682 priv->cm.rx_sge, 683 wr_id))) { 684 --p->recv_count; 685 ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed " 686 "for buf %d\n", wr_id); 687 } 688 } 689} 690 691static inline int post_send(struct ipoib_dev_priv *priv, 692 struct ipoib_cm_tx *tx, 693 unsigned int wr_id, 694 u64 addr, int len) 695{ 696 struct ib_send_wr *bad_wr; 697 698 priv->tx_sge[0].addr = addr; 699 priv->tx_sge[0].length = len; 700 701 priv->tx_wr.num_sge = 1; 702 priv->tx_wr.wr_id = wr_id | IPOIB_OP_CM; 703 704 return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr); 705} 706 707void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx) 708{ 709 struct ipoib_dev_priv *priv = netdev_priv(dev); 710 struct ipoib_cm_tx_buf *tx_req; 711 u64 addr; 712 713 if (unlikely(skb->len > tx->mtu)) { 714 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 715 skb->len, tx->mtu); 716 ++dev->stats.tx_dropped; 717 ++dev->stats.tx_errors; 718 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN); 719 return; 720 } 721 722 ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n", 723 tx->tx_head, skb->len, tx->qp->qp_num); 724 725 /* 726 * We put the skb into the tx_ring _before_ we call post_send() 727 * because it's entirely possible that the completion handler will 728 * run before we execute anything after the post_send(). That 729 * means we have to make sure everything is properly recorded and 730 * our state is consistent before we call post_send(). 731 */ 732 tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)]; 733 tx_req->skb = skb; 734 addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE); 735 if (unlikely(ib_dma_mapping_error(priv->ca, addr))) { 736 ++dev->stats.tx_errors; 737 dev_kfree_skb_any(skb); 738 return; 739 } 740 741 tx_req->mapping = addr; 742 743 if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), 744 addr, skb->len))) { 745 ipoib_warn(priv, "post_send failed\n"); 746 ++dev->stats.tx_errors; 747 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE); 748 dev_kfree_skb_any(skb); 749 } else { 750 dev->trans_start = jiffies; 751 ++tx->tx_head; 752 753 if (++priv->tx_outstanding == ipoib_sendq_size) { 754 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n", 755 tx->qp->qp_num); 756 netif_stop_queue(dev); 757 } 758 } 759} 760 761void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc) 762{ 763 struct ipoib_dev_priv *priv = netdev_priv(dev); 764 struct ipoib_cm_tx *tx = wc->qp->qp_context; 765 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM; 766 struct ipoib_cm_tx_buf *tx_req; 767 unsigned long flags; 768 769 ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n", 770 wr_id, wc->status); 771 772 if (unlikely(wr_id >= ipoib_sendq_size)) { 773 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n", 774 wr_id, ipoib_sendq_size); 775 return; 776 } 777 778 tx_req = &tx->tx_ring[wr_id]; 779 780 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE); 781 782 /* FIXME: is this right? Shouldn't we only increment on success? */ 783 ++dev->stats.tx_packets; 784 dev->stats.tx_bytes += tx_req->skb->len; 785 786 dev_kfree_skb_any(tx_req->skb); 787 788 netif_tx_lock(dev); 789 790 ++tx->tx_tail; 791 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 792 netif_queue_stopped(dev) && 793 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 794 netif_wake_queue(dev); 795 796 if (wc->status != IB_WC_SUCCESS && 797 wc->status != IB_WC_WR_FLUSH_ERR) { 798 struct ipoib_neigh *neigh; 799 800 ipoib_dbg(priv, "failed cm send event " 801 "(status=%d, wrid=%d vend_err %x)\n", 802 wc->status, wr_id, wc->vendor_err); 803 804 spin_lock_irqsave(&priv->lock, flags); 805 neigh = tx->neigh; 806 807 if (neigh) { 808 neigh->cm = NULL; 809 list_del(&neigh->list); 810 if (neigh->ah) 811 ipoib_put_ah(neigh->ah); 812 ipoib_neigh_free(dev, neigh); 813 814 tx->neigh = NULL; 815 } 816 817 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 818 list_move(&tx->list, &priv->cm.reap_list); 819 queue_work(ipoib_workqueue, &priv->cm.reap_task); 820 } 821 822 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags); 823 824 spin_unlock_irqrestore(&priv->lock, flags); 825 } 826 827 netif_tx_unlock(dev); 828} 829 830int ipoib_cm_dev_open(struct net_device *dev) 831{ 832 struct ipoib_dev_priv *priv = netdev_priv(dev); 833 int ret; 834 835 if (!IPOIB_CM_SUPPORTED(dev->dev_addr)) 836 return 0; 837 838 priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev); 839 if (IS_ERR(priv->cm.id)) { 840 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name); 841 ret = PTR_ERR(priv->cm.id); 842 goto err_cm; 843 } 844 845 ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 846 0, NULL); 847 if (ret) { 848 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name, 849 IPOIB_CM_IETF_ID | priv->qp->qp_num); 850 goto err_listen; 851 } 852 853 return 0; 854 855err_listen: 856 ib_destroy_cm_id(priv->cm.id); 857err_cm: 858 priv->cm.id = NULL; 859 return ret; 860} 861 862static void ipoib_cm_free_rx_reap_list(struct net_device *dev) 863{ 864 struct ipoib_dev_priv *priv = netdev_priv(dev); 865 struct ipoib_cm_rx *rx, *n; 866 LIST_HEAD(list); 867 868 spin_lock_irq(&priv->lock); 869 list_splice_init(&priv->cm.rx_reap_list, &list); 870 spin_unlock_irq(&priv->lock); 871 872 list_for_each_entry_safe(rx, n, &list, list) { 873 ib_destroy_cm_id(rx->id); 874 ib_destroy_qp(rx->qp); 875 if (!ipoib_cm_has_srq(dev)) { 876 ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring); 877 spin_lock_irq(&priv->lock); 878 --priv->cm.nonsrq_conn_qp; 879 spin_unlock_irq(&priv->lock); 880 } 881 kfree(rx); 882 } 883} 884 885void ipoib_cm_dev_stop(struct net_device *dev) 886{ 887 struct ipoib_dev_priv *priv = netdev_priv(dev); 888 struct ipoib_cm_rx *p; 889 unsigned long begin; 890 int ret; 891 892 if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id) 893 return; 894 895 ib_destroy_cm_id(priv->cm.id); 896 priv->cm.id = NULL; 897 898 spin_lock_irq(&priv->lock); 899 while (!list_empty(&priv->cm.passive_ids)) { 900 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list); 901 list_move(&p->list, &priv->cm.rx_error_list); 902 p->state = IPOIB_CM_RX_ERROR; 903 spin_unlock_irq(&priv->lock); 904 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 905 if (ret) 906 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 907 spin_lock_irq(&priv->lock); 908 } 909 910 /* Wait for all RX to be drained */ 911 begin = jiffies; 912 913 while (!list_empty(&priv->cm.rx_error_list) || 914 !list_empty(&priv->cm.rx_flush_list) || 915 !list_empty(&priv->cm.rx_drain_list)) { 916 if (time_after(jiffies, begin + 5 * HZ)) { 917 ipoib_warn(priv, "RX drain timing out\n"); 918 919 /* 920 * assume the HW is wedged and just free up everything. 921 */ 922 list_splice_init(&priv->cm.rx_flush_list, 923 &priv->cm.rx_reap_list); 924 list_splice_init(&priv->cm.rx_error_list, 925 &priv->cm.rx_reap_list); 926 list_splice_init(&priv->cm.rx_drain_list, 927 &priv->cm.rx_reap_list); 928 break; 929 } 930 spin_unlock_irq(&priv->lock); 931 msleep(1); 932 ipoib_drain_cq(dev); 933 spin_lock_irq(&priv->lock); 934 } 935 936 spin_unlock_irq(&priv->lock); 937 938 ipoib_cm_free_rx_reap_list(dev); 939 940 cancel_delayed_work(&priv->cm.stale_task); 941} 942 943static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event) 944{ 945 struct ipoib_cm_tx *p = cm_id->context; 946 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 947 struct ipoib_cm_data *data = event->private_data; 948 struct sk_buff_head skqueue; 949 struct ib_qp_attr qp_attr; 950 int qp_attr_mask, ret; 951 struct sk_buff *skb; 952 953 p->mtu = be32_to_cpu(data->mtu); 954 955 if (p->mtu <= IPOIB_ENCAP_LEN) { 956 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n", 957 p->mtu, IPOIB_ENCAP_LEN); 958 return -EINVAL; 959 } 960 961 qp_attr.qp_state = IB_QPS_RTR; 962 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 963 if (ret) { 964 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret); 965 return ret; 966 } 967 968 qp_attr.rq_psn = 0 /* FIXME */; 969 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 970 if (ret) { 971 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret); 972 return ret; 973 } 974 975 qp_attr.qp_state = IB_QPS_RTS; 976 ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask); 977 if (ret) { 978 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret); 979 return ret; 980 } 981 ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask); 982 if (ret) { 983 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret); 984 return ret; 985 } 986 987 skb_queue_head_init(&skqueue); 988 989 spin_lock_irq(&priv->lock); 990 set_bit(IPOIB_FLAG_OPER_UP, &p->flags); 991 if (p->neigh) 992 while ((skb = __skb_dequeue(&p->neigh->queue))) 993 __skb_queue_tail(&skqueue, skb); 994 spin_unlock_irq(&priv->lock); 995 996 while ((skb = __skb_dequeue(&skqueue))) { 997 skb->dev = p->dev; 998 if (dev_queue_xmit(skb)) 999 ipoib_warn(priv, "dev_queue_xmit failed " 1000 "to requeue packet\n"); 1001 } 1002 1003 ret = ib_send_cm_rtu(cm_id, NULL, 0); 1004 if (ret) { 1005 ipoib_warn(priv, "failed to send RTU: %d\n", ret); 1006 return ret; 1007 } 1008 return 0; 1009} 1010 1011static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx) 1012{ 1013 struct ipoib_dev_priv *priv = netdev_priv(dev); 1014 struct ib_qp_init_attr attr = { 1015 .send_cq = priv->recv_cq, 1016 .recv_cq = priv->recv_cq, 1017 .srq = priv->cm.srq, 1018 .cap.max_send_wr = ipoib_sendq_size, 1019 .cap.max_send_sge = 1, 1020 .sq_sig_type = IB_SIGNAL_ALL_WR, 1021 .qp_type = IB_QPT_RC, 1022 .qp_context = tx 1023 }; 1024 1025 return ib_create_qp(priv->pd, &attr); 1026} 1027 1028static int ipoib_cm_send_req(struct net_device *dev, 1029 struct ib_cm_id *id, struct ib_qp *qp, 1030 u32 qpn, 1031 struct ib_sa_path_rec *pathrec) 1032{ 1033 struct ipoib_dev_priv *priv = netdev_priv(dev); 1034 struct ipoib_cm_data data = {}; 1035 struct ib_cm_req_param req = {}; 1036 1037 data.qpn = cpu_to_be32(priv->qp->qp_num); 1038 data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE); 1039 1040 req.primary_path = pathrec; 1041 req.alternate_path = NULL; 1042 req.service_id = cpu_to_be64(IPOIB_CM_IETF_ID | qpn); 1043 req.qp_num = qp->qp_num; 1044 req.qp_type = qp->qp_type; 1045 req.private_data = &data; 1046 req.private_data_len = sizeof data; 1047 req.flow_control = 0; 1048 1049 req.starting_psn = 0; /* FIXME */ 1050 1051 /* 1052 * Pick some arbitrary defaults here; we could make these 1053 * module parameters if anyone cared about setting them. 1054 */ 1055 req.responder_resources = 4; 1056 req.remote_cm_response_timeout = 20; 1057 req.local_cm_response_timeout = 20; 1058 req.retry_count = 0; /* RFC draft warns against retries */ 1059 req.rnr_retry_count = 0; /* RFC draft warns against retries */ 1060 req.max_cm_retries = 15; 1061 req.srq = ipoib_cm_has_srq(dev); 1062 return ib_send_cm_req(id, &req); 1063} 1064 1065static int ipoib_cm_modify_tx_init(struct net_device *dev, 1066 struct ib_cm_id *cm_id, struct ib_qp *qp) 1067{ 1068 struct ipoib_dev_priv *priv = netdev_priv(dev); 1069 struct ib_qp_attr qp_attr; 1070 int qp_attr_mask, ret; 1071 ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index); 1072 if (ret) { 1073 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret); 1074 return ret; 1075 } 1076 1077 qp_attr.qp_state = IB_QPS_INIT; 1078 qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE; 1079 qp_attr.port_num = priv->port; 1080 qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT; 1081 1082 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask); 1083 if (ret) { 1084 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret); 1085 return ret; 1086 } 1087 return 0; 1088} 1089 1090static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn, 1091 struct ib_sa_path_rec *pathrec) 1092{ 1093 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 1094 int ret; 1095 1096 p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring); 1097 if (!p->tx_ring) { 1098 ipoib_warn(priv, "failed to allocate tx ring\n"); 1099 ret = -ENOMEM; 1100 goto err_tx; 1101 } 1102 memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring); 1103 1104 p->qp = ipoib_cm_create_tx_qp(p->dev, p); 1105 if (IS_ERR(p->qp)) { 1106 ret = PTR_ERR(p->qp); 1107 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret); 1108 goto err_qp; 1109 } 1110 1111 p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p); 1112 if (IS_ERR(p->id)) { 1113 ret = PTR_ERR(p->id); 1114 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret); 1115 goto err_id; 1116 } 1117 1118 ret = ipoib_cm_modify_tx_init(p->dev, p->id, p->qp); 1119 if (ret) { 1120 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret); 1121 goto err_modify; 1122 } 1123 1124 ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec); 1125 if (ret) { 1126 ipoib_warn(priv, "failed to send cm req: %d\n", ret); 1127 goto err_send_cm; 1128 } 1129 1130 ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n", 1131 p->qp->qp_num, pathrec->dgid.raw, qpn); 1132 1133 return 0; 1134 1135err_send_cm: 1136err_modify: 1137 ib_destroy_cm_id(p->id); 1138err_id: 1139 p->id = NULL; 1140 ib_destroy_qp(p->qp); 1141err_qp: 1142 p->qp = NULL; 1143 vfree(p->tx_ring); 1144err_tx: 1145 return ret; 1146} 1147 1148static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p) 1149{ 1150 struct ipoib_dev_priv *priv = netdev_priv(p->dev); 1151 struct ipoib_cm_tx_buf *tx_req; 1152 unsigned long begin; 1153 1154 ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n", 1155 p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail); 1156 1157 if (p->id) 1158 ib_destroy_cm_id(p->id); 1159 1160 if (p->tx_ring) { 1161 /* Wait for all sends to complete */ 1162 begin = jiffies; 1163 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1164 if (time_after(jiffies, begin + 5 * HZ)) { 1165 ipoib_warn(priv, "timing out; %d sends not completed\n", 1166 p->tx_head - p->tx_tail); 1167 goto timeout; 1168 } 1169 1170 msleep(1); 1171 } 1172 } 1173 1174timeout: 1175 1176 while ((int) p->tx_tail - (int) p->tx_head < 0) { 1177 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)]; 1178 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, 1179 DMA_TO_DEVICE); 1180 dev_kfree_skb_any(tx_req->skb); 1181 ++p->tx_tail; 1182 netif_tx_lock_bh(p->dev); 1183 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 1184 netif_queue_stopped(p->dev) && 1185 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 1186 netif_wake_queue(p->dev); 1187 netif_tx_unlock_bh(p->dev); 1188 } 1189 1190 if (p->qp) 1191 ib_destroy_qp(p->qp); 1192 1193 vfree(p->tx_ring); 1194 kfree(p); 1195} 1196 1197static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id, 1198 struct ib_cm_event *event) 1199{ 1200 struct ipoib_cm_tx *tx = cm_id->context; 1201 struct ipoib_dev_priv *priv = netdev_priv(tx->dev); 1202 struct net_device *dev = priv->dev; 1203 struct ipoib_neigh *neigh; 1204 unsigned long flags; 1205 int ret; 1206 1207 switch (event->event) { 1208 case IB_CM_DREQ_RECEIVED: 1209 ipoib_dbg(priv, "DREQ received.\n"); 1210 ib_send_cm_drep(cm_id, NULL, 0); 1211 break; 1212 case IB_CM_REP_RECEIVED: 1213 ipoib_dbg(priv, "REP received.\n"); 1214 ret = ipoib_cm_rep_handler(cm_id, event); 1215 if (ret) 1216 ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, 1217 NULL, 0, NULL, 0); 1218 break; 1219 case IB_CM_REQ_ERROR: 1220 case IB_CM_REJ_RECEIVED: 1221 case IB_CM_TIMEWAIT_EXIT: 1222 ipoib_dbg(priv, "CM error %d.\n", event->event); 1223 netif_tx_lock_bh(dev); 1224 spin_lock_irqsave(&priv->lock, flags); 1225 neigh = tx->neigh; 1226 1227 if (neigh) { 1228 neigh->cm = NULL; 1229 list_del(&neigh->list); 1230 if (neigh->ah) 1231 ipoib_put_ah(neigh->ah); 1232 ipoib_neigh_free(dev, neigh); 1233 1234 tx->neigh = NULL; 1235 } 1236 1237 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1238 list_move(&tx->list, &priv->cm.reap_list); 1239 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1240 } 1241 1242 spin_unlock_irqrestore(&priv->lock, flags); 1243 netif_tx_unlock_bh(dev); 1244 break; 1245 default: 1246 break; 1247 } 1248 1249 return 0; 1250} 1251 1252struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path, 1253 struct ipoib_neigh *neigh) 1254{ 1255 struct ipoib_dev_priv *priv = netdev_priv(dev); 1256 struct ipoib_cm_tx *tx; 1257 1258 tx = kzalloc(sizeof *tx, GFP_ATOMIC); 1259 if (!tx) 1260 return NULL; 1261 1262 neigh->cm = tx; 1263 tx->neigh = neigh; 1264 tx->path = path; 1265 tx->dev = dev; 1266 list_add(&tx->list, &priv->cm.start_list); 1267 set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags); 1268 queue_work(ipoib_workqueue, &priv->cm.start_task); 1269 return tx; 1270} 1271 1272void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx) 1273{ 1274 struct ipoib_dev_priv *priv = netdev_priv(tx->dev); 1275 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) { 1276 list_move(&tx->list, &priv->cm.reap_list); 1277 queue_work(ipoib_workqueue, &priv->cm.reap_task); 1278 ipoib_dbg(priv, "Reap connection for gid %pI6\n", 1279 tx->neigh->dgid.raw); 1280 tx->neigh = NULL; 1281 } 1282} 1283 1284static void ipoib_cm_tx_start(struct work_struct *work) 1285{ 1286 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1287 cm.start_task); 1288 struct net_device *dev = priv->dev; 1289 struct ipoib_neigh *neigh; 1290 struct ipoib_cm_tx *p; 1291 unsigned long flags; 1292 int ret; 1293 1294 struct ib_sa_path_rec pathrec; 1295 u32 qpn; 1296 1297 netif_tx_lock_bh(dev); 1298 spin_lock_irqsave(&priv->lock, flags); 1299 1300 while (!list_empty(&priv->cm.start_list)) { 1301 p = list_entry(priv->cm.start_list.next, typeof(*p), list); 1302 list_del_init(&p->list); 1303 neigh = p->neigh; 1304 qpn = IPOIB_QPN(neigh->neighbour->ha); 1305 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec); 1306 1307 spin_unlock_irqrestore(&priv->lock, flags); 1308 netif_tx_unlock_bh(dev); 1309 1310 ret = ipoib_cm_tx_init(p, qpn, &pathrec); 1311 1312 netif_tx_lock_bh(dev); 1313 spin_lock_irqsave(&priv->lock, flags); 1314 1315 if (ret) { 1316 neigh = p->neigh; 1317 if (neigh) { 1318 neigh->cm = NULL; 1319 list_del(&neigh->list); 1320 if (neigh->ah) 1321 ipoib_put_ah(neigh->ah); 1322 ipoib_neigh_free(dev, neigh); 1323 } 1324 list_del(&p->list); 1325 kfree(p); 1326 } 1327 } 1328 1329 spin_unlock_irqrestore(&priv->lock, flags); 1330 netif_tx_unlock_bh(dev); 1331} 1332 1333static void ipoib_cm_tx_reap(struct work_struct *work) 1334{ 1335 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1336 cm.reap_task); 1337 struct net_device *dev = priv->dev; 1338 struct ipoib_cm_tx *p; 1339 unsigned long flags; 1340 1341 netif_tx_lock_bh(dev); 1342 spin_lock_irqsave(&priv->lock, flags); 1343 1344 while (!list_empty(&priv->cm.reap_list)) { 1345 p = list_entry(priv->cm.reap_list.next, typeof(*p), list); 1346 list_del(&p->list); 1347 spin_unlock_irqrestore(&priv->lock, flags); 1348 netif_tx_unlock_bh(dev); 1349 ipoib_cm_tx_destroy(p); 1350 netif_tx_lock_bh(dev); 1351 spin_lock_irqsave(&priv->lock, flags); 1352 } 1353 1354 spin_unlock_irqrestore(&priv->lock, flags); 1355 netif_tx_unlock_bh(dev); 1356} 1357 1358static void ipoib_cm_skb_reap(struct work_struct *work) 1359{ 1360 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1361 cm.skb_task); 1362 struct net_device *dev = priv->dev; 1363 struct sk_buff *skb; 1364 unsigned long flags; 1365 unsigned mtu = priv->mcast_mtu; 1366 1367 netif_tx_lock_bh(dev); 1368 spin_lock_irqsave(&priv->lock, flags); 1369 1370 while ((skb = skb_dequeue(&priv->cm.skb_queue))) { 1371 spin_unlock_irqrestore(&priv->lock, flags); 1372 netif_tx_unlock_bh(dev); 1373 1374 if (skb->protocol == htons(ETH_P_IP)) 1375 icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu)); 1376#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) 1377 else if (skb->protocol == htons(ETH_P_IPV6)) 1378 icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev); 1379#endif 1380 dev_kfree_skb_any(skb); 1381 1382 netif_tx_lock_bh(dev); 1383 spin_lock_irqsave(&priv->lock, flags); 1384 } 1385 1386 spin_unlock_irqrestore(&priv->lock, flags); 1387 netif_tx_unlock_bh(dev); 1388} 1389 1390void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb, 1391 unsigned int mtu) 1392{ 1393 struct ipoib_dev_priv *priv = netdev_priv(dev); 1394 int e = skb_queue_empty(&priv->cm.skb_queue); 1395 1396 if (skb_dst(skb)) 1397 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu); 1398 1399 skb_queue_tail(&priv->cm.skb_queue, skb); 1400 if (e) 1401 queue_work(ipoib_workqueue, &priv->cm.skb_task); 1402} 1403 1404static void ipoib_cm_rx_reap(struct work_struct *work) 1405{ 1406 ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv, 1407 cm.rx_reap_task)->dev); 1408} 1409 1410static void ipoib_cm_stale_task(struct work_struct *work) 1411{ 1412 struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv, 1413 cm.stale_task.work); 1414 struct ipoib_cm_rx *p; 1415 int ret; 1416 1417 spin_lock_irq(&priv->lock); 1418 while (!list_empty(&priv->cm.passive_ids)) { 1419 /* List is sorted by LRU, start from tail, 1420 * stop when we see a recently used entry */ 1421 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list); 1422 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT)) 1423 break; 1424 list_move(&p->list, &priv->cm.rx_error_list); 1425 p->state = IPOIB_CM_RX_ERROR; 1426 spin_unlock_irq(&priv->lock); 1427 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE); 1428 if (ret) 1429 ipoib_warn(priv, "unable to move qp to error state: %d\n", ret); 1430 spin_lock_irq(&priv->lock); 1431 } 1432 1433 if (!list_empty(&priv->cm.passive_ids)) 1434 queue_delayed_work(ipoib_workqueue, 1435 &priv->cm.stale_task, IPOIB_CM_RX_DELAY); 1436 spin_unlock_irq(&priv->lock); 1437} 1438 1439 1440static ssize_t show_mode(struct device *d, struct device_attribute *attr, 1441 char *buf) 1442{ 1443 struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d)); 1444 1445 if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags)) 1446 return sprintf(buf, "connected\n"); 1447 else 1448 return sprintf(buf, "datagram\n"); 1449} 1450 1451static ssize_t set_mode(struct device *d, struct device_attribute *attr, 1452 const char *buf, size_t count) 1453{ 1454 struct net_device *dev = to_net_dev(d); 1455 struct ipoib_dev_priv *priv = netdev_priv(dev); 1456 1457 if (!rtnl_trylock()) 1458 return restart_syscall(); 1459 1460 /* flush paths if we switch modes so that connections are restarted */ 1461 if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) { 1462 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 1463 ipoib_warn(priv, "enabling connected mode " 1464 "will cause multicast packet drops\n"); 1465 1466 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO); 1467 rtnl_unlock(); 1468 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM; 1469 1470 ipoib_flush_paths(dev); 1471 return count; 1472 } 1473 1474 if (!strcmp(buf, "datagram\n")) { 1475 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags); 1476 1477 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) { 1478 dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG; 1479 if (priv->hca_caps & IB_DEVICE_UD_TSO) 1480 dev->features |= NETIF_F_TSO; 1481 } 1482 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu)); 1483 rtnl_unlock(); 1484 ipoib_flush_paths(dev); 1485 1486 return count; 1487 } 1488 rtnl_unlock(); 1489 1490 return -EINVAL; 1491} 1492 1493static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode); 1494 1495int ipoib_cm_add_mode_attr(struct net_device *dev) 1496{ 1497 return device_create_file(&dev->dev, &dev_attr_mode); 1498} 1499 1500static void ipoib_cm_create_srq(struct net_device *dev, int max_sge) 1501{ 1502 struct ipoib_dev_priv *priv = netdev_priv(dev); 1503 struct ib_srq_init_attr srq_init_attr = { 1504 .attr = { 1505 .max_wr = ipoib_recvq_size, 1506 .max_sge = max_sge 1507 } 1508 }; 1509 1510 priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr); 1511 if (IS_ERR(priv->cm.srq)) { 1512 if (PTR_ERR(priv->cm.srq) != -ENOSYS) 1513 printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n", 1514 priv->ca->name, PTR_ERR(priv->cm.srq)); 1515 priv->cm.srq = NULL; 1516 return; 1517 } 1518 1519 priv->cm.srq_ring = vmalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring); 1520 if (!priv->cm.srq_ring) { 1521 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n", 1522 priv->ca->name, ipoib_recvq_size); 1523 ib_destroy_srq(priv->cm.srq); 1524 priv->cm.srq = NULL; 1525 return; 1526 } 1527 1528 memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring); 1529} 1530 1531int ipoib_cm_dev_init(struct net_device *dev) 1532{ 1533 struct ipoib_dev_priv *priv = netdev_priv(dev); 1534 int i, ret; 1535 struct ib_device_attr attr; 1536 1537 INIT_LIST_HEAD(&priv->cm.passive_ids); 1538 INIT_LIST_HEAD(&priv->cm.reap_list); 1539 INIT_LIST_HEAD(&priv->cm.start_list); 1540 INIT_LIST_HEAD(&priv->cm.rx_error_list); 1541 INIT_LIST_HEAD(&priv->cm.rx_flush_list); 1542 INIT_LIST_HEAD(&priv->cm.rx_drain_list); 1543 INIT_LIST_HEAD(&priv->cm.rx_reap_list); 1544 INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start); 1545 INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap); 1546 INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap); 1547 INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap); 1548 INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task); 1549 1550 skb_queue_head_init(&priv->cm.skb_queue); 1551 1552 ret = ib_query_device(priv->ca, &attr); 1553 if (ret) { 1554 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret); 1555 return ret; 1556 } 1557 1558 ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge); 1559 1560 attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge); 1561 ipoib_cm_create_srq(dev, attr.max_srq_sge); 1562 if (ipoib_cm_has_srq(dev)) { 1563 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10; 1564 priv->cm.num_frags = attr.max_srq_sge; 1565 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n", 1566 priv->cm.max_cm_mtu, priv->cm.num_frags); 1567 } else { 1568 priv->cm.max_cm_mtu = IPOIB_CM_MTU; 1569 priv->cm.num_frags = IPOIB_CM_RX_SG; 1570 } 1571 1572 ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge); 1573 1574 if (ipoib_cm_has_srq(dev)) { 1575 for (i = 0; i < ipoib_recvq_size; ++i) { 1576 if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i, 1577 priv->cm.num_frags - 1, 1578 priv->cm.srq_ring[i].mapping)) { 1579 ipoib_warn(priv, "failed to allocate " 1580 "receive buffer %d\n", i); 1581 ipoib_cm_dev_cleanup(dev); 1582 return -ENOMEM; 1583 } 1584 1585 if (ipoib_cm_post_receive_srq(dev, i)) { 1586 ipoib_warn(priv, "ipoib_cm_post_receive_srq " 1587 "failed for buf %d\n", i); 1588 ipoib_cm_dev_cleanup(dev); 1589 return -EIO; 1590 } 1591 } 1592 } 1593 1594 priv->dev->dev_addr[0] = IPOIB_FLAGS_RC; 1595 return 0; 1596} 1597 1598void ipoib_cm_dev_cleanup(struct net_device *dev) 1599{ 1600 struct ipoib_dev_priv *priv = netdev_priv(dev); 1601 int ret; 1602 1603 if (!priv->cm.srq) 1604 return; 1605 1606 ipoib_dbg(priv, "Cleanup ipoib connected mode.\n"); 1607 1608 ret = ib_destroy_srq(priv->cm.srq); 1609 if (ret) 1610 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret); 1611 1612 priv->cm.srq = NULL; 1613 if (!priv->cm.srq_ring) 1614 return; 1615 1616 ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring); 1617 priv->cm.srq_ring = NULL; 1618} 1619