1 /* 2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved. 3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved. 4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved. 5 * Copyright (c) 2004, 2005 Voltaire, Inc. All rights reserved. 6 * 7 * This software is available to you under a choice of one of two 8 * licenses. You may choose to be licensed under the terms of the GNU 9 * General Public License (GPL) Version 2, available from the file 10 * COPYING in the main directory of this source tree, or the 11 * OpenIB.org BSD license below: 12 * 13 * Redistribution and use in source and binary forms, with or 14 * without modification, are permitted provided that the following 15 * conditions are met: 16 * 17 * - Redistributions of source code must retain the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer. 20 * 21 * - Redistributions in binary form must reproduce the above 22 * copyright notice, this list of conditions and the following 23 * disclaimer in the documentation and/or other materials 24 * provided with the distribution. 25 * 26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 33 * SOFTWARE. 34 */ 35 36 #include <linux/delay.h> 37 #include <linux/moduleparam.h> 38 #include <linux/dma-mapping.h> 39 #include <linux/slab.h> 40 41 #include <linux/ip.h> 42 #include <linux/tcp.h> 43 44 #include "ipoib.h" 45 46 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA 47 static int data_debug_level; 48 49 module_param(data_debug_level, int, 0644); 50 MODULE_PARM_DESC(data_debug_level, 51 "Enable data path debug tracing if > 0"); 52 #endif 53 54 struct ipoib_ah *ipoib_create_ah(struct net_device *dev, 55 struct ib_pd *pd, struct rdma_ah_attr *attr) 56 { 57 struct ipoib_ah *ah; 58 struct ib_ah *vah; 59 60 ah = kmalloc(sizeof *ah, GFP_KERNEL); 61 if (!ah) 62 return ERR_PTR(-ENOMEM); 63 64 ah->dev = dev; 65 ah->last_send = 0; 66 kref_init(&ah->ref); 67 68 vah = rdma_create_ah(pd, attr); 69 if (IS_ERR(vah)) { 70 kfree(ah); 71 ah = (struct ipoib_ah *)vah; 72 } else { 73 ah->ah = vah; 74 ipoib_dbg(ipoib_priv(dev), "Created ah %p\n", ah->ah); 75 } 76 77 return ah; 78 } 79 80 void ipoib_free_ah(struct kref *kref) 81 { 82 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref); 83 struct ipoib_dev_priv *priv = ipoib_priv(ah->dev); 84 85 unsigned long flags; 86 87 spin_lock_irqsave(&priv->lock, flags); 88 list_add_tail(&ah->list, &priv->dead_ahs); 89 spin_unlock_irqrestore(&priv->lock, flags); 90 } 91 92 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv, 93 u64 mapping[IPOIB_UD_RX_SG]) 94 { 95 ib_dma_unmap_single(priv->ca, mapping[0], 96 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu), 97 DMA_FROM_DEVICE); 98 } 99 100 static int ipoib_ib_post_receive(struct net_device *dev, int id) 101 { 102 struct ipoib_dev_priv *priv = ipoib_priv(dev); 103 struct ib_recv_wr *bad_wr; 104 int ret; 105 106 priv->rx_wr.wr_id = id | IPOIB_OP_RECV; 107 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0]; 108 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1]; 109 110 111 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr); 112 if (unlikely(ret)) { 113 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret); 114 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping); 115 dev_kfree_skb_any(priv->rx_ring[id].skb); 116 priv->rx_ring[id].skb = NULL; 117 } 118 119 return ret; 120 } 121 122 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id) 123 { 124 struct ipoib_dev_priv *priv = ipoib_priv(dev); 125 struct sk_buff *skb; 126 int buf_size; 127 u64 *mapping; 128 129 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu); 130 131 skb = dev_alloc_skb(buf_size + IPOIB_HARD_LEN); 132 if (unlikely(!skb)) 133 return NULL; 134 135 /* 136 * the IP header will be at IPOIP_HARD_LEN + IB_GRH_BYTES, that is 137 * 64 bytes aligned 138 */ 139 skb_reserve(skb, sizeof(struct ipoib_pseudo_header)); 140 141 mapping = priv->rx_ring[id].mapping; 142 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size, 143 DMA_FROM_DEVICE); 144 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) 145 goto error; 146 147 priv->rx_ring[id].skb = skb; 148 return skb; 149 error: 150 dev_kfree_skb_any(skb); 151 return NULL; 152 } 153 154 static int ipoib_ib_post_receives(struct net_device *dev) 155 { 156 struct ipoib_dev_priv *priv = ipoib_priv(dev); 157 int i; 158 159 for (i = 0; i < ipoib_recvq_size; ++i) { 160 if (!ipoib_alloc_rx_skb(dev, i)) { 161 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 162 return -ENOMEM; 163 } 164 if (ipoib_ib_post_receive(dev, i)) { 165 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i); 166 return -EIO; 167 } 168 } 169 170 return 0; 171 } 172 173 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc) 174 { 175 struct ipoib_dev_priv *priv = ipoib_priv(dev); 176 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV; 177 struct sk_buff *skb; 178 u64 mapping[IPOIB_UD_RX_SG]; 179 union ib_gid *dgid; 180 union ib_gid *sgid; 181 182 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n", 183 wr_id, wc->status); 184 185 if (unlikely(wr_id >= ipoib_recvq_size)) { 186 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n", 187 wr_id, ipoib_recvq_size); 188 return; 189 } 190 191 skb = priv->rx_ring[wr_id].skb; 192 193 if (unlikely(wc->status != IB_WC_SUCCESS)) { 194 if (wc->status != IB_WC_WR_FLUSH_ERR) 195 ipoib_warn(priv, "failed recv event " 196 "(status=%d, wrid=%d vend_err %x)\n", 197 wc->status, wr_id, wc->vendor_err); 198 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping); 199 dev_kfree_skb_any(skb); 200 priv->rx_ring[wr_id].skb = NULL; 201 return; 202 } 203 204 memcpy(mapping, priv->rx_ring[wr_id].mapping, 205 IPOIB_UD_RX_SG * sizeof *mapping); 206 207 /* 208 * If we can't allocate a new RX buffer, dump 209 * this packet and reuse the old buffer. 210 */ 211 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) { 212 ++dev->stats.rx_dropped; 213 goto repost; 214 } 215 216 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 217 wc->byte_len, wc->slid); 218 219 ipoib_ud_dma_unmap_rx(priv, mapping); 220 221 skb_put(skb, wc->byte_len); 222 223 /* First byte of dgid signals multicast when 0xff */ 224 dgid = &((struct ib_grh *)skb->data)->dgid; 225 226 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff) 227 skb->pkt_type = PACKET_HOST; 228 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0) 229 skb->pkt_type = PACKET_BROADCAST; 230 else 231 skb->pkt_type = PACKET_MULTICAST; 232 233 sgid = &((struct ib_grh *)skb->data)->sgid; 234 235 /* 236 * Drop packets that this interface sent, ie multicast packets 237 * that the HCA has replicated. 238 */ 239 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) { 240 int need_repost = 1; 241 242 if ((wc->wc_flags & IB_WC_GRH) && 243 sgid->global.interface_id != priv->local_gid.global.interface_id) 244 need_repost = 0; 245 246 if (need_repost) { 247 dev_kfree_skb_any(skb); 248 goto repost; 249 } 250 } 251 252 skb_pull(skb, IB_GRH_BYTES); 253 254 skb->protocol = ((struct ipoib_header *) skb->data)->proto; 255 skb_add_pseudo_hdr(skb); 256 257 ++dev->stats.rx_packets; 258 dev->stats.rx_bytes += skb->len; 259 if (skb->pkt_type == PACKET_MULTICAST) 260 dev->stats.multicast++; 261 262 skb->dev = dev; 263 if ((dev->features & NETIF_F_RXCSUM) && 264 likely(wc->wc_flags & IB_WC_IP_CSUM_OK)) 265 skb->ip_summed = CHECKSUM_UNNECESSARY; 266 267 napi_gro_receive(&priv->napi, skb); 268 269 repost: 270 if (unlikely(ipoib_ib_post_receive(dev, wr_id))) 271 ipoib_warn(priv, "ipoib_ib_post_receive failed " 272 "for buf %d\n", wr_id); 273 } 274 275 int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req) 276 { 277 struct sk_buff *skb = tx_req->skb; 278 u64 *mapping = tx_req->mapping; 279 int i; 280 int off; 281 282 if (skb_headlen(skb)) { 283 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb), 284 DMA_TO_DEVICE); 285 if (unlikely(ib_dma_mapping_error(ca, mapping[0]))) 286 return -EIO; 287 288 off = 1; 289 } else 290 off = 0; 291 292 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { 293 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 294 mapping[i + off] = ib_dma_map_page(ca, 295 skb_frag_page(frag), 296 frag->page_offset, skb_frag_size(frag), 297 DMA_TO_DEVICE); 298 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off]))) 299 goto partial_error; 300 } 301 return 0; 302 303 partial_error: 304 for (; i > 0; --i) { 305 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1]; 306 307 ib_dma_unmap_page(ca, mapping[i - !off], skb_frag_size(frag), DMA_TO_DEVICE); 308 } 309 310 if (off) 311 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE); 312 313 return -EIO; 314 } 315 316 void ipoib_dma_unmap_tx(struct ipoib_dev_priv *priv, 317 struct ipoib_tx_buf *tx_req) 318 { 319 struct sk_buff *skb = tx_req->skb; 320 u64 *mapping = tx_req->mapping; 321 int i; 322 int off; 323 324 if (skb_headlen(skb)) { 325 ib_dma_unmap_single(priv->ca, mapping[0], skb_headlen(skb), 326 DMA_TO_DEVICE); 327 off = 1; 328 } else 329 off = 0; 330 331 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { 332 const skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 333 334 ib_dma_unmap_page(priv->ca, mapping[i + off], 335 skb_frag_size(frag), DMA_TO_DEVICE); 336 } 337 } 338 339 /* 340 * As the result of a completion error the QP Can be transferred to SQE states. 341 * The function checks if the (send)QP is in SQE state and 342 * moves it back to RTS state, that in order to have it functional again. 343 */ 344 static void ipoib_qp_state_validate_work(struct work_struct *work) 345 { 346 struct ipoib_qp_state_validate *qp_work = 347 container_of(work, struct ipoib_qp_state_validate, work); 348 349 struct ipoib_dev_priv *priv = qp_work->priv; 350 struct ib_qp_attr qp_attr; 351 struct ib_qp_init_attr query_init_attr; 352 int ret; 353 354 ret = ib_query_qp(priv->qp, &qp_attr, IB_QP_STATE, &query_init_attr); 355 if (ret) { 356 ipoib_warn(priv, "%s: Failed to query QP ret: %d\n", 357 __func__, ret); 358 goto free_res; 359 } 360 pr_info("%s: QP: 0x%x is in state: %d\n", 361 __func__, priv->qp->qp_num, qp_attr.qp_state); 362 363 /* currently support only in SQE->RTS transition*/ 364 if (qp_attr.qp_state == IB_QPS_SQE) { 365 qp_attr.qp_state = IB_QPS_RTS; 366 367 ret = ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE); 368 if (ret) { 369 pr_warn("failed(%d) modify QP:0x%x SQE->RTS\n", 370 ret, priv->qp->qp_num); 371 goto free_res; 372 } 373 pr_info("%s: QP: 0x%x moved from IB_QPS_SQE to IB_QPS_RTS\n", 374 __func__, priv->qp->qp_num); 375 } else { 376 pr_warn("QP (%d) will stay in state: %d\n", 377 priv->qp->qp_num, qp_attr.qp_state); 378 } 379 380 free_res: 381 kfree(qp_work); 382 } 383 384 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc) 385 { 386 struct ipoib_dev_priv *priv = ipoib_priv(dev); 387 unsigned int wr_id = wc->wr_id; 388 struct ipoib_tx_buf *tx_req; 389 390 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n", 391 wr_id, wc->status); 392 393 if (unlikely(wr_id >= ipoib_sendq_size)) { 394 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n", 395 wr_id, ipoib_sendq_size); 396 return; 397 } 398 399 tx_req = &priv->tx_ring[wr_id]; 400 401 ipoib_dma_unmap_tx(priv, tx_req); 402 403 ++dev->stats.tx_packets; 404 dev->stats.tx_bytes += tx_req->skb->len; 405 406 dev_kfree_skb_any(tx_req->skb); 407 408 ++priv->tx_tail; 409 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 410 netif_queue_stopped(dev) && 411 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 412 netif_wake_queue(dev); 413 414 if (wc->status != IB_WC_SUCCESS && 415 wc->status != IB_WC_WR_FLUSH_ERR) { 416 struct ipoib_qp_state_validate *qp_work; 417 ipoib_warn(priv, "failed send event " 418 "(status=%d, wrid=%d vend_err %x)\n", 419 wc->status, wr_id, wc->vendor_err); 420 qp_work = kzalloc(sizeof(*qp_work), GFP_ATOMIC); 421 if (!qp_work) 422 return; 423 424 INIT_WORK(&qp_work->work, ipoib_qp_state_validate_work); 425 qp_work->priv = priv; 426 queue_work(priv->wq, &qp_work->work); 427 } 428 } 429 430 static int poll_tx(struct ipoib_dev_priv *priv) 431 { 432 int n, i; 433 434 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc); 435 for (i = 0; i < n; ++i) 436 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i); 437 438 return n == MAX_SEND_CQE; 439 } 440 441 int ipoib_poll(struct napi_struct *napi, int budget) 442 { 443 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi); 444 struct net_device *dev = priv->dev; 445 int done; 446 int t; 447 int n, i; 448 449 done = 0; 450 451 poll_more: 452 while (done < budget) { 453 int max = (budget - done); 454 455 t = min(IPOIB_NUM_WC, max); 456 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc); 457 458 for (i = 0; i < n; i++) { 459 struct ib_wc *wc = priv->ibwc + i; 460 461 if (wc->wr_id & IPOIB_OP_RECV) { 462 ++done; 463 if (wc->wr_id & IPOIB_OP_CM) 464 ipoib_cm_handle_rx_wc(dev, wc); 465 else 466 ipoib_ib_handle_rx_wc(dev, wc); 467 } else 468 ipoib_cm_handle_tx_wc(priv->dev, wc); 469 } 470 471 if (n != t) 472 break; 473 } 474 475 if (done < budget) { 476 napi_complete(napi); 477 if (unlikely(ib_req_notify_cq(priv->recv_cq, 478 IB_CQ_NEXT_COMP | 479 IB_CQ_REPORT_MISSED_EVENTS)) && 480 napi_reschedule(napi)) 481 goto poll_more; 482 } 483 484 return done; 485 } 486 487 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr) 488 { 489 struct net_device *dev = dev_ptr; 490 struct ipoib_dev_priv *priv = ipoib_priv(dev); 491 492 napi_schedule(&priv->napi); 493 } 494 495 static void drain_tx_cq(struct net_device *dev) 496 { 497 struct ipoib_dev_priv *priv = ipoib_priv(dev); 498 499 netif_tx_lock(dev); 500 while (poll_tx(priv)) 501 ; /* nothing */ 502 503 if (netif_queue_stopped(dev)) 504 mod_timer(&priv->poll_timer, jiffies + 1); 505 506 netif_tx_unlock(dev); 507 } 508 509 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr) 510 { 511 struct ipoib_dev_priv *priv = ipoib_priv(dev_ptr); 512 513 mod_timer(&priv->poll_timer, jiffies); 514 } 515 516 static inline int post_send(struct ipoib_dev_priv *priv, 517 unsigned int wr_id, 518 struct ib_ah *address, u32 dqpn, 519 struct ipoib_tx_buf *tx_req, 520 void *head, int hlen) 521 { 522 struct ib_send_wr *bad_wr; 523 struct sk_buff *skb = tx_req->skb; 524 525 ipoib_build_sge(priv, tx_req); 526 527 priv->tx_wr.wr.wr_id = wr_id; 528 priv->tx_wr.remote_qpn = dqpn; 529 priv->tx_wr.ah = address; 530 531 if (head) { 532 priv->tx_wr.mss = skb_shinfo(skb)->gso_size; 533 priv->tx_wr.header = head; 534 priv->tx_wr.hlen = hlen; 535 priv->tx_wr.wr.opcode = IB_WR_LSO; 536 } else 537 priv->tx_wr.wr.opcode = IB_WR_SEND; 538 539 return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr); 540 } 541 542 int ipoib_send(struct net_device *dev, struct sk_buff *skb, 543 struct ib_ah *address, u32 dqpn) 544 { 545 struct ipoib_dev_priv *priv = ipoib_priv(dev); 546 struct ipoib_tx_buf *tx_req; 547 int hlen, rc; 548 void *phead; 549 unsigned usable_sge = priv->max_send_sge - !!skb_headlen(skb); 550 551 if (skb_is_gso(skb)) { 552 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb); 553 phead = skb->data; 554 if (unlikely(!skb_pull(skb, hlen))) { 555 ipoib_warn(priv, "linear data too small\n"); 556 ++dev->stats.tx_dropped; 557 ++dev->stats.tx_errors; 558 dev_kfree_skb_any(skb); 559 return -1; 560 } 561 } else { 562 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) { 563 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 564 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN); 565 ++dev->stats.tx_dropped; 566 ++dev->stats.tx_errors; 567 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu); 568 return -1; 569 } 570 phead = NULL; 571 hlen = 0; 572 } 573 if (skb_shinfo(skb)->nr_frags > usable_sge) { 574 if (skb_linearize(skb) < 0) { 575 ipoib_warn(priv, "skb could not be linearized\n"); 576 ++dev->stats.tx_dropped; 577 ++dev->stats.tx_errors; 578 dev_kfree_skb_any(skb); 579 return -1; 580 } 581 /* Does skb_linearize return ok without reducing nr_frags? */ 582 if (skb_shinfo(skb)->nr_frags > usable_sge) { 583 ipoib_warn(priv, "too many frags after skb linearize\n"); 584 ++dev->stats.tx_dropped; 585 ++dev->stats.tx_errors; 586 dev_kfree_skb_any(skb); 587 return -1; 588 } 589 } 590 591 ipoib_dbg_data(priv, 592 "sending packet, length=%d address=%p dqpn=0x%06x\n", 593 skb->len, address, dqpn); 594 595 /* 596 * We put the skb into the tx_ring _before_ we call post_send() 597 * because it's entirely possible that the completion handler will 598 * run before we execute anything after the post_send(). That 599 * means we have to make sure everything is properly recorded and 600 * our state is consistent before we call post_send(). 601 */ 602 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)]; 603 tx_req->skb = skb; 604 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) { 605 ++dev->stats.tx_errors; 606 dev_kfree_skb_any(skb); 607 return -1; 608 } 609 610 if (skb->ip_summed == CHECKSUM_PARTIAL) 611 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM; 612 else 613 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM; 614 615 if (++priv->tx_outstanding == ipoib_sendq_size) { 616 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n"); 617 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP)) 618 ipoib_warn(priv, "request notify on send CQ failed\n"); 619 netif_stop_queue(dev); 620 } 621 622 skb_orphan(skb); 623 skb_dst_drop(skb); 624 625 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1), 626 address, dqpn, tx_req, phead, hlen); 627 if (unlikely(rc)) { 628 ipoib_warn(priv, "post_send failed, error %d\n", rc); 629 ++dev->stats.tx_errors; 630 --priv->tx_outstanding; 631 ipoib_dma_unmap_tx(priv, tx_req); 632 dev_kfree_skb_any(skb); 633 if (netif_queue_stopped(dev)) 634 netif_wake_queue(dev); 635 rc = 0; 636 } else { 637 netif_trans_update(dev); 638 639 rc = priv->tx_head; 640 ++priv->tx_head; 641 } 642 643 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) 644 while (poll_tx(priv)) 645 ; /* nothing */ 646 647 return rc; 648 } 649 650 static void __ipoib_reap_ah(struct net_device *dev) 651 { 652 struct ipoib_dev_priv *priv = ipoib_priv(dev); 653 struct ipoib_ah *ah, *tah; 654 LIST_HEAD(remove_list); 655 unsigned long flags; 656 657 netif_tx_lock_bh(dev); 658 spin_lock_irqsave(&priv->lock, flags); 659 660 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list) 661 if ((int) priv->tx_tail - (int) ah->last_send >= 0) { 662 list_del(&ah->list); 663 rdma_destroy_ah(ah->ah); 664 kfree(ah); 665 } 666 667 spin_unlock_irqrestore(&priv->lock, flags); 668 netif_tx_unlock_bh(dev); 669 } 670 671 void ipoib_reap_ah(struct work_struct *work) 672 { 673 struct ipoib_dev_priv *priv = 674 container_of(work, struct ipoib_dev_priv, ah_reap_task.work); 675 struct net_device *dev = priv->dev; 676 677 __ipoib_reap_ah(dev); 678 679 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags)) 680 queue_delayed_work(priv->wq, &priv->ah_reap_task, 681 round_jiffies_relative(HZ)); 682 } 683 684 static void ipoib_flush_ah(struct net_device *dev) 685 { 686 struct ipoib_dev_priv *priv = ipoib_priv(dev); 687 688 cancel_delayed_work(&priv->ah_reap_task); 689 flush_workqueue(priv->wq); 690 ipoib_reap_ah(&priv->ah_reap_task.work); 691 } 692 693 static void ipoib_stop_ah(struct net_device *dev) 694 { 695 struct ipoib_dev_priv *priv = ipoib_priv(dev); 696 697 set_bit(IPOIB_STOP_REAPER, &priv->flags); 698 ipoib_flush_ah(dev); 699 } 700 701 static int recvs_pending(struct net_device *dev) 702 { 703 struct ipoib_dev_priv *priv = ipoib_priv(dev); 704 int pending = 0; 705 int i; 706 707 for (i = 0; i < ipoib_recvq_size; ++i) 708 if (priv->rx_ring[i].skb) 709 ++pending; 710 711 return pending; 712 } 713 714 static void check_qp_movement_and_print(struct ipoib_dev_priv *priv, 715 struct ib_qp *qp, 716 enum ib_qp_state new_state) 717 { 718 struct ib_qp_attr qp_attr; 719 struct ib_qp_init_attr query_init_attr; 720 int ret; 721 722 ret = ib_query_qp(qp, &qp_attr, IB_QP_STATE, &query_init_attr); 723 if (ret) { 724 ipoib_warn(priv, "%s: Failed to query QP\n", __func__); 725 return; 726 } 727 /* print according to the new-state and the previous state.*/ 728 if (new_state == IB_QPS_ERR && qp_attr.qp_state == IB_QPS_RESET) 729 ipoib_dbg(priv, "Failed modify QP, IB_QPS_RESET to IB_QPS_ERR, acceptable\n"); 730 else 731 ipoib_warn(priv, "Failed to modify QP to state: %d from state: %d\n", 732 new_state, qp_attr.qp_state); 733 } 734 735 int ipoib_ib_dev_stop_default(struct net_device *dev) 736 { 737 struct ipoib_dev_priv *priv = ipoib_priv(dev); 738 struct ib_qp_attr qp_attr; 739 unsigned long begin; 740 struct ipoib_tx_buf *tx_req; 741 int i; 742 743 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) 744 napi_disable(&priv->napi); 745 746 ipoib_cm_dev_stop(dev); 747 748 /* 749 * Move our QP to the error state and then reinitialize in 750 * when all work requests have completed or have been flushed. 751 */ 752 qp_attr.qp_state = IB_QPS_ERR; 753 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 754 check_qp_movement_and_print(priv, priv->qp, IB_QPS_ERR); 755 756 /* Wait for all sends and receives to complete */ 757 begin = jiffies; 758 759 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) { 760 if (time_after(jiffies, begin + 5 * HZ)) { 761 ipoib_warn(priv, 762 "timing out; %d sends %d receives not completed\n", 763 priv->tx_head - priv->tx_tail, 764 recvs_pending(dev)); 765 766 /* 767 * assume the HW is wedged and just free up 768 * all our pending work requests. 769 */ 770 while ((int)priv->tx_tail - (int)priv->tx_head < 0) { 771 tx_req = &priv->tx_ring[priv->tx_tail & 772 (ipoib_sendq_size - 1)]; 773 ipoib_dma_unmap_tx(priv, tx_req); 774 dev_kfree_skb_any(tx_req->skb); 775 ++priv->tx_tail; 776 --priv->tx_outstanding; 777 } 778 779 for (i = 0; i < ipoib_recvq_size; ++i) { 780 struct ipoib_rx_buf *rx_req; 781 782 rx_req = &priv->rx_ring[i]; 783 if (!rx_req->skb) 784 continue; 785 ipoib_ud_dma_unmap_rx(priv, 786 priv->rx_ring[i].mapping); 787 dev_kfree_skb_any(rx_req->skb); 788 rx_req->skb = NULL; 789 } 790 791 goto timeout; 792 } 793 794 ipoib_drain_cq(dev); 795 796 usleep_range(1000, 2000); 797 } 798 799 ipoib_dbg(priv, "All sends and receives done.\n"); 800 801 timeout: 802 del_timer_sync(&priv->poll_timer); 803 qp_attr.qp_state = IB_QPS_RESET; 804 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 805 ipoib_warn(priv, "Failed to modify QP to RESET state\n"); 806 807 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP); 808 809 return 0; 810 } 811 812 int ipoib_ib_dev_stop(struct net_device *dev) 813 { 814 struct ipoib_dev_priv *priv = ipoib_priv(dev); 815 816 priv->rn_ops->ndo_stop(dev); 817 818 clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags); 819 ipoib_flush_ah(dev); 820 821 return 0; 822 } 823 824 void ipoib_ib_tx_timer_func(unsigned long ctx) 825 { 826 drain_tx_cq((struct net_device *)ctx); 827 } 828 829 int ipoib_ib_dev_open_default(struct net_device *dev) 830 { 831 struct ipoib_dev_priv *priv = ipoib_priv(dev); 832 int ret; 833 834 ret = ipoib_init_qp(dev); 835 if (ret) { 836 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret); 837 return -1; 838 } 839 840 ret = ipoib_ib_post_receives(dev); 841 if (ret) { 842 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret); 843 goto out; 844 } 845 846 ret = ipoib_cm_dev_open(dev); 847 if (ret) { 848 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret); 849 goto out; 850 } 851 852 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) 853 napi_enable(&priv->napi); 854 855 return 0; 856 out: 857 return -1; 858 } 859 860 int ipoib_ib_dev_open(struct net_device *dev) 861 { 862 struct ipoib_dev_priv *priv = ipoib_priv(dev); 863 864 ipoib_pkey_dev_check_presence(dev); 865 866 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 867 ipoib_warn(priv, "P_Key 0x%04x is %s\n", priv->pkey, 868 (!(priv->pkey & 0x7fff) ? "Invalid" : "not found")); 869 return -1; 870 } 871 872 clear_bit(IPOIB_STOP_REAPER, &priv->flags); 873 queue_delayed_work(priv->wq, &priv->ah_reap_task, 874 round_jiffies_relative(HZ)); 875 876 if (priv->rn_ops->ndo_open(dev)) { 877 pr_warn("%s: Failed to open dev\n", dev->name); 878 goto dev_stop; 879 } 880 881 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags); 882 883 return 0; 884 885 dev_stop: 886 set_bit(IPOIB_STOP_REAPER, &priv->flags); 887 cancel_delayed_work(&priv->ah_reap_task); 888 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags); 889 ipoib_ib_dev_stop(dev); 890 return -1; 891 } 892 893 void ipoib_pkey_dev_check_presence(struct net_device *dev) 894 { 895 struct ipoib_dev_priv *priv = ipoib_priv(dev); 896 897 if (!(priv->pkey & 0x7fff) || 898 ib_find_pkey(priv->ca, priv->port, priv->pkey, 899 &priv->pkey_index)) 900 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 901 else 902 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 903 } 904 905 void ipoib_ib_dev_up(struct net_device *dev) 906 { 907 struct ipoib_dev_priv *priv = ipoib_priv(dev); 908 909 ipoib_pkey_dev_check_presence(dev); 910 911 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 912 ipoib_dbg(priv, "PKEY is not assigned.\n"); 913 return; 914 } 915 916 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 917 918 ipoib_mcast_start_thread(dev); 919 } 920 921 void ipoib_ib_dev_down(struct net_device *dev) 922 { 923 struct ipoib_dev_priv *priv = ipoib_priv(dev); 924 925 ipoib_dbg(priv, "downing ib_dev\n"); 926 927 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 928 netif_carrier_off(dev); 929 930 ipoib_mcast_stop_thread(dev); 931 ipoib_mcast_dev_flush(dev); 932 933 ipoib_flush_paths(dev); 934 } 935 936 void ipoib_drain_cq(struct net_device *dev) 937 { 938 struct ipoib_dev_priv *priv = ipoib_priv(dev); 939 int i, n; 940 941 /* 942 * We call completion handling routines that expect to be 943 * called from the BH-disabled NAPI poll context, so disable 944 * BHs here too. 945 */ 946 local_bh_disable(); 947 948 do { 949 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc); 950 for (i = 0; i < n; ++i) { 951 /* 952 * Convert any successful completions to flush 953 * errors to avoid passing packets up the 954 * stack after bringing the device down. 955 */ 956 if (priv->ibwc[i].status == IB_WC_SUCCESS) 957 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR; 958 959 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) { 960 if (priv->ibwc[i].wr_id & IPOIB_OP_CM) 961 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i); 962 else 963 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i); 964 } else 965 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i); 966 } 967 } while (n == IPOIB_NUM_WC); 968 969 while (poll_tx(priv)) 970 ; /* nothing */ 971 972 local_bh_enable(); 973 } 974 975 /* 976 * Takes whatever value which is in pkey index 0 and updates priv->pkey 977 * returns 0 if the pkey value was changed. 978 */ 979 static inline int update_parent_pkey(struct ipoib_dev_priv *priv) 980 { 981 int result; 982 u16 prev_pkey; 983 984 prev_pkey = priv->pkey; 985 result = ib_query_pkey(priv->ca, priv->port, 0, &priv->pkey); 986 if (result) { 987 ipoib_warn(priv, "ib_query_pkey port %d failed (ret = %d)\n", 988 priv->port, result); 989 return result; 990 } 991 992 priv->pkey |= 0x8000; 993 994 if (prev_pkey != priv->pkey) { 995 ipoib_dbg(priv, "pkey changed from 0x%x to 0x%x\n", 996 prev_pkey, priv->pkey); 997 /* 998 * Update the pkey in the broadcast address, while making sure to set 999 * the full membership bit, so that we join the right broadcast group. 1000 */ 1001 priv->dev->broadcast[8] = priv->pkey >> 8; 1002 priv->dev->broadcast[9] = priv->pkey & 0xff; 1003 return 0; 1004 } 1005 1006 return 1; 1007 } 1008 /* 1009 * returns 0 if pkey value was found in a different slot. 1010 */ 1011 static inline int update_child_pkey(struct ipoib_dev_priv *priv) 1012 { 1013 u16 old_index = priv->pkey_index; 1014 1015 priv->pkey_index = 0; 1016 ipoib_pkey_dev_check_presence(priv->dev); 1017 1018 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) && 1019 (old_index == priv->pkey_index)) 1020 return 1; 1021 return 0; 1022 } 1023 1024 /* 1025 * returns true if the device address of the ipoib interface has changed and the 1026 * new address is a valid one (i.e in the gid table), return false otherwise. 1027 */ 1028 static bool ipoib_dev_addr_changed_valid(struct ipoib_dev_priv *priv) 1029 { 1030 union ib_gid search_gid; 1031 union ib_gid gid0; 1032 union ib_gid *netdev_gid; 1033 int err; 1034 u16 index; 1035 u8 port; 1036 bool ret = false; 1037 1038 netdev_gid = (union ib_gid *)(priv->dev->dev_addr + 4); 1039 if (ib_query_gid(priv->ca, priv->port, 0, &gid0, NULL)) 1040 return false; 1041 1042 netif_addr_lock_bh(priv->dev); 1043 1044 /* The subnet prefix may have changed, update it now so we won't have 1045 * to do it later 1046 */ 1047 priv->local_gid.global.subnet_prefix = gid0.global.subnet_prefix; 1048 netdev_gid->global.subnet_prefix = gid0.global.subnet_prefix; 1049 search_gid.global.subnet_prefix = gid0.global.subnet_prefix; 1050 1051 search_gid.global.interface_id = priv->local_gid.global.interface_id; 1052 1053 netif_addr_unlock_bh(priv->dev); 1054 1055 err = ib_find_gid(priv->ca, &search_gid, IB_GID_TYPE_IB, 1056 priv->dev, &port, &index); 1057 1058 netif_addr_lock_bh(priv->dev); 1059 1060 if (search_gid.global.interface_id != 1061 priv->local_gid.global.interface_id) 1062 /* There was a change while we were looking up the gid, bail 1063 * here and let the next work sort this out 1064 */ 1065 goto out; 1066 1067 /* The next section of code needs some background: 1068 * Per IB spec the port GUID can't change if the HCA is powered on. 1069 * port GUID is the basis for GID at index 0 which is the basis for 1070 * the default device address of a ipoib interface. 1071 * 1072 * so it seems the flow should be: 1073 * if user_changed_dev_addr && gid in gid tbl 1074 * set bit dev_addr_set 1075 * return true 1076 * else 1077 * return false 1078 * 1079 * The issue is that there are devices that don't follow the spec, 1080 * they change the port GUID when the HCA is powered, so in order 1081 * not to break userspace applications, We need to check if the 1082 * user wanted to control the device address and we assume that 1083 * if he sets the device address back to be based on GID index 0, 1084 * he no longer wishs to control it. 1085 * 1086 * If the user doesn't control the the device address, 1087 * IPOIB_FLAG_DEV_ADDR_SET is set and ib_find_gid failed it means 1088 * the port GUID has changed and GID at index 0 has changed 1089 * so we need to change priv->local_gid and priv->dev->dev_addr 1090 * to reflect the new GID. 1091 */ 1092 if (!test_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags)) { 1093 if (!err && port == priv->port) { 1094 set_bit(IPOIB_FLAG_DEV_ADDR_SET, &priv->flags); 1095 if (index == 0) 1096 clear_bit(IPOIB_FLAG_DEV_ADDR_CTRL, 1097 &priv->flags); 1098 else 1099 set_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags); 1100 ret = true; 1101 } else { 1102 ret = false; 1103 } 1104 } else { 1105 if (!err && port == priv->port) { 1106 ret = true; 1107 } else { 1108 if (!test_bit(IPOIB_FLAG_DEV_ADDR_CTRL, &priv->flags)) { 1109 memcpy(&priv->local_gid, &gid0, 1110 sizeof(priv->local_gid)); 1111 memcpy(priv->dev->dev_addr + 4, &gid0, 1112 sizeof(priv->local_gid)); 1113 ret = true; 1114 } 1115 } 1116 } 1117 1118 out: 1119 netif_addr_unlock_bh(priv->dev); 1120 1121 return ret; 1122 } 1123 1124 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, 1125 enum ipoib_flush_level level, 1126 int nesting) 1127 { 1128 struct ipoib_dev_priv *cpriv; 1129 struct net_device *dev = priv->dev; 1130 int result; 1131 1132 down_read_nested(&priv->vlan_rwsem, nesting); 1133 1134 /* 1135 * Flush any child interfaces too -- they might be up even if 1136 * the parent is down. 1137 */ 1138 list_for_each_entry(cpriv, &priv->child_intfs, list) 1139 __ipoib_ib_dev_flush(cpriv, level, nesting + 1); 1140 1141 up_read(&priv->vlan_rwsem); 1142 1143 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags) && 1144 level != IPOIB_FLUSH_HEAVY) { 1145 /* Make sure the dev_addr is set even if not flushing */ 1146 if (level == IPOIB_FLUSH_LIGHT) 1147 ipoib_dev_addr_changed_valid(priv); 1148 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n"); 1149 return; 1150 } 1151 1152 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 1153 /* interface is down. update pkey and leave. */ 1154 if (level == IPOIB_FLUSH_HEAVY) { 1155 if (!test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) 1156 update_parent_pkey(priv); 1157 else 1158 update_child_pkey(priv); 1159 } else if (level == IPOIB_FLUSH_LIGHT) 1160 ipoib_dev_addr_changed_valid(priv); 1161 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n"); 1162 return; 1163 } 1164 1165 if (level == IPOIB_FLUSH_HEAVY) { 1166 /* child devices chase their origin pkey value, while non-child 1167 * (parent) devices should always takes what present in pkey index 0 1168 */ 1169 if (test_bit(IPOIB_FLAG_SUBINTERFACE, &priv->flags)) { 1170 result = update_child_pkey(priv); 1171 if (result) { 1172 /* restart QP only if P_Key index is changed */ 1173 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n"); 1174 return; 1175 } 1176 1177 } else { 1178 result = update_parent_pkey(priv); 1179 /* restart QP only if P_Key value changed */ 1180 if (result) { 1181 ipoib_dbg(priv, "Not flushing - P_Key value not changed.\n"); 1182 return; 1183 } 1184 } 1185 } 1186 1187 if (level == IPOIB_FLUSH_LIGHT) { 1188 int oper_up; 1189 ipoib_mark_paths_invalid(dev); 1190 /* Set IPoIB operation as down to prevent races between: 1191 * the flush flow which leaves MCG and on the fly joins 1192 * which can happen during that time. mcast restart task 1193 * should deal with join requests we missed. 1194 */ 1195 oper_up = test_and_clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 1196 ipoib_mcast_dev_flush(dev); 1197 if (oper_up) 1198 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 1199 ipoib_flush_ah(dev); 1200 } 1201 1202 if (level >= IPOIB_FLUSH_NORMAL) 1203 ipoib_ib_dev_down(dev); 1204 1205 if (level == IPOIB_FLUSH_HEAVY) { 1206 if (test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) 1207 ipoib_ib_dev_stop(dev); 1208 if (ipoib_ib_dev_open(dev) != 0) 1209 return; 1210 if (netif_queue_stopped(dev)) 1211 netif_start_queue(dev); 1212 } 1213 1214 /* 1215 * The device could have been brought down between the start and when 1216 * we get here, don't bring it back up if it's not configured up 1217 */ 1218 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 1219 if (level >= IPOIB_FLUSH_NORMAL) 1220 ipoib_ib_dev_up(dev); 1221 if (ipoib_dev_addr_changed_valid(priv)) 1222 ipoib_mcast_restart_task(&priv->restart_task); 1223 } 1224 } 1225 1226 void ipoib_ib_dev_flush_light(struct work_struct *work) 1227 { 1228 struct ipoib_dev_priv *priv = 1229 container_of(work, struct ipoib_dev_priv, flush_light); 1230 1231 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT, 0); 1232 } 1233 1234 void ipoib_ib_dev_flush_normal(struct work_struct *work) 1235 { 1236 struct ipoib_dev_priv *priv = 1237 container_of(work, struct ipoib_dev_priv, flush_normal); 1238 1239 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL, 0); 1240 } 1241 1242 void ipoib_ib_dev_flush_heavy(struct work_struct *work) 1243 { 1244 struct ipoib_dev_priv *priv = 1245 container_of(work, struct ipoib_dev_priv, flush_heavy); 1246 1247 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY, 0); 1248 } 1249 1250 void ipoib_ib_dev_cleanup(struct net_device *dev) 1251 { 1252 struct ipoib_dev_priv *priv = ipoib_priv(dev); 1253 1254 ipoib_dbg(priv, "cleaning up ib_dev\n"); 1255 /* 1256 * We must make sure there are no more (path) completions 1257 * that may wish to touch priv fields that are no longer valid 1258 */ 1259 ipoib_flush_paths(dev); 1260 1261 ipoib_mcast_stop_thread(dev); 1262 ipoib_mcast_dev_flush(dev); 1263 1264 /* 1265 * All of our ah references aren't free until after 1266 * ipoib_mcast_dev_flush(), ipoib_flush_paths, and 1267 * the neighbor garbage collection is stopped and reaped. 1268 * That should all be done now, so make a final ah flush. 1269 */ 1270 ipoib_stop_ah(dev); 1271 1272 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 1273 1274 priv->rn_ops->ndo_uninit(dev); 1275 1276 if (priv->pd) { 1277 ib_dealloc_pd(priv->pd); 1278 priv->pd = NULL; 1279 } 1280 } 1281 1282