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/dma-mapping.h> 38 #include <linux/slab.h> 39 40 #include <linux/ip.h> 41 #include <linux/tcp.h> 42 43 #include "ipoib.h" 44 45 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA 46 static int data_debug_level; 47 48 module_param(data_debug_level, int, 0644); 49 MODULE_PARM_DESC(data_debug_level, 50 "Enable data path debug tracing if > 0"); 51 #endif 52 53 static DEFINE_MUTEX(pkey_mutex); 54 55 struct ipoib_ah *ipoib_create_ah(struct net_device *dev, 56 struct ib_pd *pd, struct ib_ah_attr *attr) 57 { 58 struct ipoib_ah *ah; 59 60 ah = kmalloc(sizeof *ah, GFP_KERNEL); 61 if (!ah) 62 return NULL; 63 64 ah->dev = dev; 65 ah->last_send = 0; 66 kref_init(&ah->ref); 67 68 ah->ah = ib_create_ah(pd, attr); 69 if (IS_ERR(ah->ah)) { 70 kfree(ah); 71 ah = NULL; 72 } else 73 ipoib_dbg(netdev_priv(dev), "Created ah %p\n", ah->ah); 74 75 return ah; 76 } 77 78 void ipoib_free_ah(struct kref *kref) 79 { 80 struct ipoib_ah *ah = container_of(kref, struct ipoib_ah, ref); 81 struct ipoib_dev_priv *priv = netdev_priv(ah->dev); 82 83 unsigned long flags; 84 85 spin_lock_irqsave(&priv->lock, flags); 86 list_add_tail(&ah->list, &priv->dead_ahs); 87 spin_unlock_irqrestore(&priv->lock, flags); 88 } 89 90 static void ipoib_ud_dma_unmap_rx(struct ipoib_dev_priv *priv, 91 u64 mapping[IPOIB_UD_RX_SG]) 92 { 93 if (ipoib_ud_need_sg(priv->max_ib_mtu)) { 94 ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_UD_HEAD_SIZE, 95 DMA_FROM_DEVICE); 96 ib_dma_unmap_page(priv->ca, mapping[1], PAGE_SIZE, 97 DMA_FROM_DEVICE); 98 } else 99 ib_dma_unmap_single(priv->ca, mapping[0], 100 IPOIB_UD_BUF_SIZE(priv->max_ib_mtu), 101 DMA_FROM_DEVICE); 102 } 103 104 static void ipoib_ud_skb_put_frags(struct ipoib_dev_priv *priv, 105 struct sk_buff *skb, 106 unsigned int length) 107 { 108 if (ipoib_ud_need_sg(priv->max_ib_mtu)) { 109 skb_frag_t *frag = &skb_shinfo(skb)->frags[0]; 110 unsigned int size; 111 /* 112 * There is only two buffers needed for max_payload = 4K, 113 * first buf size is IPOIB_UD_HEAD_SIZE 114 */ 115 skb->tail += IPOIB_UD_HEAD_SIZE; 116 skb->len += length; 117 118 size = length - IPOIB_UD_HEAD_SIZE; 119 120 frag->size = size; 121 skb->data_len += size; 122 skb->truesize += size; 123 } else 124 skb_put(skb, length); 125 126 } 127 128 static int ipoib_ib_post_receive(struct net_device *dev, int id) 129 { 130 struct ipoib_dev_priv *priv = netdev_priv(dev); 131 struct ib_recv_wr *bad_wr; 132 int ret; 133 134 priv->rx_wr.wr_id = id | IPOIB_OP_RECV; 135 priv->rx_sge[0].addr = priv->rx_ring[id].mapping[0]; 136 priv->rx_sge[1].addr = priv->rx_ring[id].mapping[1]; 137 138 139 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr); 140 if (unlikely(ret)) { 141 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret); 142 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[id].mapping); 143 dev_kfree_skb_any(priv->rx_ring[id].skb); 144 priv->rx_ring[id].skb = NULL; 145 } 146 147 return ret; 148 } 149 150 static struct sk_buff *ipoib_alloc_rx_skb(struct net_device *dev, int id) 151 { 152 struct ipoib_dev_priv *priv = netdev_priv(dev); 153 struct sk_buff *skb; 154 int buf_size; 155 u64 *mapping; 156 157 if (ipoib_ud_need_sg(priv->max_ib_mtu)) 158 buf_size = IPOIB_UD_HEAD_SIZE; 159 else 160 buf_size = IPOIB_UD_BUF_SIZE(priv->max_ib_mtu); 161 162 skb = dev_alloc_skb(buf_size + 4); 163 if (unlikely(!skb)) 164 return NULL; 165 166 /* 167 * IB will leave a 40 byte gap for a GRH and IPoIB adds a 4 byte 168 * header. So we need 4 more bytes to get to 48 and align the 169 * IP header to a multiple of 16. 170 */ 171 skb_reserve(skb, 4); 172 173 mapping = priv->rx_ring[id].mapping; 174 mapping[0] = ib_dma_map_single(priv->ca, skb->data, buf_size, 175 DMA_FROM_DEVICE); 176 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) 177 goto error; 178 179 if (ipoib_ud_need_sg(priv->max_ib_mtu)) { 180 struct page *page = alloc_page(GFP_ATOMIC); 181 if (!page) 182 goto partial_error; 183 skb_fill_page_desc(skb, 0, page, 0, PAGE_SIZE); 184 mapping[1] = 185 ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[0].page, 186 0, PAGE_SIZE, DMA_FROM_DEVICE); 187 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[1]))) 188 goto partial_error; 189 } 190 191 priv->rx_ring[id].skb = skb; 192 return skb; 193 194 partial_error: 195 ib_dma_unmap_single(priv->ca, mapping[0], buf_size, DMA_FROM_DEVICE); 196 error: 197 dev_kfree_skb_any(skb); 198 return NULL; 199 } 200 201 static int ipoib_ib_post_receives(struct net_device *dev) 202 { 203 struct ipoib_dev_priv *priv = netdev_priv(dev); 204 int i; 205 206 for (i = 0; i < ipoib_recvq_size; ++i) { 207 if (!ipoib_alloc_rx_skb(dev, i)) { 208 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 209 return -ENOMEM; 210 } 211 if (ipoib_ib_post_receive(dev, i)) { 212 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i); 213 return -EIO; 214 } 215 } 216 217 return 0; 218 } 219 220 static void ipoib_ib_handle_rx_wc(struct net_device *dev, struct ib_wc *wc) 221 { 222 struct ipoib_dev_priv *priv = netdev_priv(dev); 223 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV; 224 struct sk_buff *skb; 225 u64 mapping[IPOIB_UD_RX_SG]; 226 union ib_gid *dgid; 227 228 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n", 229 wr_id, wc->status); 230 231 if (unlikely(wr_id >= ipoib_recvq_size)) { 232 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n", 233 wr_id, ipoib_recvq_size); 234 return; 235 } 236 237 skb = priv->rx_ring[wr_id].skb; 238 239 if (unlikely(wc->status != IB_WC_SUCCESS)) { 240 if (wc->status != IB_WC_WR_FLUSH_ERR) 241 ipoib_warn(priv, "failed recv event " 242 "(status=%d, wrid=%d vend_err %x)\n", 243 wc->status, wr_id, wc->vendor_err); 244 ipoib_ud_dma_unmap_rx(priv, priv->rx_ring[wr_id].mapping); 245 dev_kfree_skb_any(skb); 246 priv->rx_ring[wr_id].skb = NULL; 247 return; 248 } 249 250 /* 251 * Drop packets that this interface sent, ie multicast packets 252 * that the HCA has replicated. 253 */ 254 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) 255 goto repost; 256 257 memcpy(mapping, priv->rx_ring[wr_id].mapping, 258 IPOIB_UD_RX_SG * sizeof *mapping); 259 260 /* 261 * If we can't allocate a new RX buffer, dump 262 * this packet and reuse the old buffer. 263 */ 264 if (unlikely(!ipoib_alloc_rx_skb(dev, wr_id))) { 265 ++dev->stats.rx_dropped; 266 goto repost; 267 } 268 269 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 270 wc->byte_len, wc->slid); 271 272 ipoib_ud_dma_unmap_rx(priv, mapping); 273 ipoib_ud_skb_put_frags(priv, skb, wc->byte_len); 274 275 /* First byte of dgid signals multicast when 0xff */ 276 dgid = &((struct ib_grh *)skb->data)->dgid; 277 278 if (!(wc->wc_flags & IB_WC_GRH) || dgid->raw[0] != 0xff) 279 skb->pkt_type = PACKET_HOST; 280 else if (memcmp(dgid, dev->broadcast + 4, sizeof(union ib_gid)) == 0) 281 skb->pkt_type = PACKET_BROADCAST; 282 else 283 skb->pkt_type = PACKET_MULTICAST; 284 285 skb_pull(skb, IB_GRH_BYTES); 286 287 skb->protocol = ((struct ipoib_header *) skb->data)->proto; 288 skb_reset_mac_header(skb); 289 skb_pull(skb, IPOIB_ENCAP_LEN); 290 291 ++dev->stats.rx_packets; 292 dev->stats.rx_bytes += skb->len; 293 294 skb->dev = dev; 295 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->csum_ok)) 296 skb->ip_summed = CHECKSUM_UNNECESSARY; 297 298 napi_gro_receive(&priv->napi, skb); 299 300 repost: 301 if (unlikely(ipoib_ib_post_receive(dev, wr_id))) 302 ipoib_warn(priv, "ipoib_ib_post_receive failed " 303 "for buf %d\n", wr_id); 304 } 305 306 static int ipoib_dma_map_tx(struct ib_device *ca, 307 struct ipoib_tx_buf *tx_req) 308 { 309 struct sk_buff *skb = tx_req->skb; 310 u64 *mapping = tx_req->mapping; 311 int i; 312 int off; 313 314 if (skb_headlen(skb)) { 315 mapping[0] = ib_dma_map_single(ca, skb->data, skb_headlen(skb), 316 DMA_TO_DEVICE); 317 if (unlikely(ib_dma_mapping_error(ca, mapping[0]))) 318 return -EIO; 319 320 off = 1; 321 } else 322 off = 0; 323 324 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { 325 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 326 mapping[i + off] = ib_dma_map_page(ca, frag->page, 327 frag->page_offset, frag->size, 328 DMA_TO_DEVICE); 329 if (unlikely(ib_dma_mapping_error(ca, mapping[i + off]))) 330 goto partial_error; 331 } 332 return 0; 333 334 partial_error: 335 for (; i > 0; --i) { 336 skb_frag_t *frag = &skb_shinfo(skb)->frags[i - 1]; 337 ib_dma_unmap_page(ca, mapping[i - !off], frag->size, DMA_TO_DEVICE); 338 } 339 340 if (off) 341 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE); 342 343 return -EIO; 344 } 345 346 static void ipoib_dma_unmap_tx(struct ib_device *ca, 347 struct ipoib_tx_buf *tx_req) 348 { 349 struct sk_buff *skb = tx_req->skb; 350 u64 *mapping = tx_req->mapping; 351 int i; 352 int off; 353 354 if (skb_headlen(skb)) { 355 ib_dma_unmap_single(ca, mapping[0], skb_headlen(skb), DMA_TO_DEVICE); 356 off = 1; 357 } else 358 off = 0; 359 360 for (i = 0; i < skb_shinfo(skb)->nr_frags; ++i) { 361 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 362 ib_dma_unmap_page(ca, mapping[i + off], frag->size, 363 DMA_TO_DEVICE); 364 } 365 } 366 367 static void ipoib_ib_handle_tx_wc(struct net_device *dev, struct ib_wc *wc) 368 { 369 struct ipoib_dev_priv *priv = netdev_priv(dev); 370 unsigned int wr_id = wc->wr_id; 371 struct ipoib_tx_buf *tx_req; 372 373 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n", 374 wr_id, wc->status); 375 376 if (unlikely(wr_id >= ipoib_sendq_size)) { 377 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n", 378 wr_id, ipoib_sendq_size); 379 return; 380 } 381 382 tx_req = &priv->tx_ring[wr_id]; 383 384 ipoib_dma_unmap_tx(priv->ca, tx_req); 385 386 ++dev->stats.tx_packets; 387 dev->stats.tx_bytes += tx_req->skb->len; 388 389 dev_kfree_skb_any(tx_req->skb); 390 391 ++priv->tx_tail; 392 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 393 netif_queue_stopped(dev) && 394 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 395 netif_wake_queue(dev); 396 397 if (wc->status != IB_WC_SUCCESS && 398 wc->status != IB_WC_WR_FLUSH_ERR) 399 ipoib_warn(priv, "failed send event " 400 "(status=%d, wrid=%d vend_err %x)\n", 401 wc->status, wr_id, wc->vendor_err); 402 } 403 404 static int poll_tx(struct ipoib_dev_priv *priv) 405 { 406 int n, i; 407 408 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc); 409 for (i = 0; i < n; ++i) 410 ipoib_ib_handle_tx_wc(priv->dev, priv->send_wc + i); 411 412 return n == MAX_SEND_CQE; 413 } 414 415 int ipoib_poll(struct napi_struct *napi, int budget) 416 { 417 struct ipoib_dev_priv *priv = container_of(napi, struct ipoib_dev_priv, napi); 418 struct net_device *dev = priv->dev; 419 int done; 420 int t; 421 int n, i; 422 423 done = 0; 424 425 poll_more: 426 while (done < budget) { 427 int max = (budget - done); 428 429 t = min(IPOIB_NUM_WC, max); 430 n = ib_poll_cq(priv->recv_cq, t, priv->ibwc); 431 432 for (i = 0; i < n; i++) { 433 struct ib_wc *wc = priv->ibwc + i; 434 435 if (wc->wr_id & IPOIB_OP_RECV) { 436 ++done; 437 if (wc->wr_id & IPOIB_OP_CM) 438 ipoib_cm_handle_rx_wc(dev, wc); 439 else 440 ipoib_ib_handle_rx_wc(dev, wc); 441 } else 442 ipoib_cm_handle_tx_wc(priv->dev, wc); 443 } 444 445 if (n != t) 446 break; 447 } 448 449 if (done < budget) { 450 napi_complete(napi); 451 if (unlikely(ib_req_notify_cq(priv->recv_cq, 452 IB_CQ_NEXT_COMP | 453 IB_CQ_REPORT_MISSED_EVENTS)) && 454 napi_reschedule(napi)) 455 goto poll_more; 456 } 457 458 return done; 459 } 460 461 void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr) 462 { 463 struct net_device *dev = dev_ptr; 464 struct ipoib_dev_priv *priv = netdev_priv(dev); 465 466 napi_schedule(&priv->napi); 467 } 468 469 static void drain_tx_cq(struct net_device *dev) 470 { 471 struct ipoib_dev_priv *priv = netdev_priv(dev); 472 473 netif_tx_lock(dev); 474 while (poll_tx(priv)) 475 ; /* nothing */ 476 477 if (netif_queue_stopped(dev)) 478 mod_timer(&priv->poll_timer, jiffies + 1); 479 480 netif_tx_unlock(dev); 481 } 482 483 void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr) 484 { 485 struct ipoib_dev_priv *priv = netdev_priv(dev_ptr); 486 487 mod_timer(&priv->poll_timer, jiffies); 488 } 489 490 static inline int post_send(struct ipoib_dev_priv *priv, 491 unsigned int wr_id, 492 struct ib_ah *address, u32 qpn, 493 struct ipoib_tx_buf *tx_req, 494 void *head, int hlen) 495 { 496 struct ib_send_wr *bad_wr; 497 int i, off; 498 struct sk_buff *skb = tx_req->skb; 499 skb_frag_t *frags = skb_shinfo(skb)->frags; 500 int nr_frags = skb_shinfo(skb)->nr_frags; 501 u64 *mapping = tx_req->mapping; 502 503 if (skb_headlen(skb)) { 504 priv->tx_sge[0].addr = mapping[0]; 505 priv->tx_sge[0].length = skb_headlen(skb); 506 off = 1; 507 } else 508 off = 0; 509 510 for (i = 0; i < nr_frags; ++i) { 511 priv->tx_sge[i + off].addr = mapping[i + off]; 512 priv->tx_sge[i + off].length = frags[i].size; 513 } 514 priv->tx_wr.num_sge = nr_frags + off; 515 priv->tx_wr.wr_id = wr_id; 516 priv->tx_wr.wr.ud.remote_qpn = qpn; 517 priv->tx_wr.wr.ud.ah = address; 518 519 if (head) { 520 priv->tx_wr.wr.ud.mss = skb_shinfo(skb)->gso_size; 521 priv->tx_wr.wr.ud.header = head; 522 priv->tx_wr.wr.ud.hlen = hlen; 523 priv->tx_wr.opcode = IB_WR_LSO; 524 } else 525 priv->tx_wr.opcode = IB_WR_SEND; 526 527 return ib_post_send(priv->qp, &priv->tx_wr, &bad_wr); 528 } 529 530 void ipoib_send(struct net_device *dev, struct sk_buff *skb, 531 struct ipoib_ah *address, u32 qpn) 532 { 533 struct ipoib_dev_priv *priv = netdev_priv(dev); 534 struct ipoib_tx_buf *tx_req; 535 int hlen, rc; 536 void *phead; 537 538 if (skb_is_gso(skb)) { 539 hlen = skb_transport_offset(skb) + tcp_hdrlen(skb); 540 phead = skb->data; 541 if (unlikely(!skb_pull(skb, hlen))) { 542 ipoib_warn(priv, "linear data too small\n"); 543 ++dev->stats.tx_dropped; 544 ++dev->stats.tx_errors; 545 dev_kfree_skb_any(skb); 546 return; 547 } 548 } else { 549 if (unlikely(skb->len > priv->mcast_mtu + IPOIB_ENCAP_LEN)) { 550 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 551 skb->len, priv->mcast_mtu + IPOIB_ENCAP_LEN); 552 ++dev->stats.tx_dropped; 553 ++dev->stats.tx_errors; 554 ipoib_cm_skb_too_long(dev, skb, priv->mcast_mtu); 555 return; 556 } 557 phead = NULL; 558 hlen = 0; 559 } 560 561 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n", 562 skb->len, address, qpn); 563 564 /* 565 * We put the skb into the tx_ring _before_ we call post_send() 566 * because it's entirely possible that the completion handler will 567 * run before we execute anything after the post_send(). That 568 * means we have to make sure everything is properly recorded and 569 * our state is consistent before we call post_send(). 570 */ 571 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)]; 572 tx_req->skb = skb; 573 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) { 574 ++dev->stats.tx_errors; 575 dev_kfree_skb_any(skb); 576 return; 577 } 578 579 if (skb->ip_summed == CHECKSUM_PARTIAL) 580 priv->tx_wr.send_flags |= IB_SEND_IP_CSUM; 581 else 582 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM; 583 584 if (++priv->tx_outstanding == ipoib_sendq_size) { 585 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n"); 586 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP)) 587 ipoib_warn(priv, "request notify on send CQ failed\n"); 588 netif_stop_queue(dev); 589 } 590 591 rc = post_send(priv, priv->tx_head & (ipoib_sendq_size - 1), 592 address->ah, qpn, tx_req, phead, hlen); 593 if (unlikely(rc)) { 594 ipoib_warn(priv, "post_send failed, error %d\n", rc); 595 ++dev->stats.tx_errors; 596 --priv->tx_outstanding; 597 ipoib_dma_unmap_tx(priv->ca, tx_req); 598 dev_kfree_skb_any(skb); 599 if (netif_queue_stopped(dev)) 600 netif_wake_queue(dev); 601 } else { 602 dev->trans_start = jiffies; 603 604 address->last_send = priv->tx_head; 605 ++priv->tx_head; 606 skb_orphan(skb); 607 608 } 609 610 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) 611 while (poll_tx(priv)) 612 ; /* nothing */ 613 } 614 615 static void __ipoib_reap_ah(struct net_device *dev) 616 { 617 struct ipoib_dev_priv *priv = netdev_priv(dev); 618 struct ipoib_ah *ah, *tah; 619 LIST_HEAD(remove_list); 620 unsigned long flags; 621 622 netif_tx_lock_bh(dev); 623 spin_lock_irqsave(&priv->lock, flags); 624 625 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list) 626 if ((int) priv->tx_tail - (int) ah->last_send >= 0) { 627 list_del(&ah->list); 628 ib_destroy_ah(ah->ah); 629 kfree(ah); 630 } 631 632 spin_unlock_irqrestore(&priv->lock, flags); 633 netif_tx_unlock_bh(dev); 634 } 635 636 void ipoib_reap_ah(struct work_struct *work) 637 { 638 struct ipoib_dev_priv *priv = 639 container_of(work, struct ipoib_dev_priv, ah_reap_task.work); 640 struct net_device *dev = priv->dev; 641 642 __ipoib_reap_ah(dev); 643 644 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags)) 645 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, 646 round_jiffies_relative(HZ)); 647 } 648 649 static void ipoib_ib_tx_timer_func(unsigned long ctx) 650 { 651 drain_tx_cq((struct net_device *)ctx); 652 } 653 654 int ipoib_ib_dev_open(struct net_device *dev) 655 { 656 struct ipoib_dev_priv *priv = netdev_priv(dev); 657 int ret; 658 659 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) { 660 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey); 661 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 662 return -1; 663 } 664 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 665 666 ret = ipoib_init_qp(dev); 667 if (ret) { 668 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret); 669 return -1; 670 } 671 672 ret = ipoib_ib_post_receives(dev); 673 if (ret) { 674 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret); 675 ipoib_ib_dev_stop(dev, 1); 676 return -1; 677 } 678 679 ret = ipoib_cm_dev_open(dev); 680 if (ret) { 681 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret); 682 ipoib_ib_dev_stop(dev, 1); 683 return -1; 684 } 685 686 clear_bit(IPOIB_STOP_REAPER, &priv->flags); 687 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, 688 round_jiffies_relative(HZ)); 689 690 if (!test_and_set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) 691 napi_enable(&priv->napi); 692 693 return 0; 694 } 695 696 static void ipoib_pkey_dev_check_presence(struct net_device *dev) 697 { 698 struct ipoib_dev_priv *priv = netdev_priv(dev); 699 u16 pkey_index = 0; 700 701 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index)) 702 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 703 else 704 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 705 } 706 707 int ipoib_ib_dev_up(struct net_device *dev) 708 { 709 struct ipoib_dev_priv *priv = netdev_priv(dev); 710 711 ipoib_pkey_dev_check_presence(dev); 712 713 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 714 ipoib_dbg(priv, "PKEY is not assigned.\n"); 715 return 0; 716 } 717 718 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 719 720 return ipoib_mcast_start_thread(dev); 721 } 722 723 int ipoib_ib_dev_down(struct net_device *dev, int flush) 724 { 725 struct ipoib_dev_priv *priv = netdev_priv(dev); 726 727 ipoib_dbg(priv, "downing ib_dev\n"); 728 729 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 730 netif_carrier_off(dev); 731 732 /* Shutdown the P_Key thread if still active */ 733 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 734 mutex_lock(&pkey_mutex); 735 set_bit(IPOIB_PKEY_STOP, &priv->flags); 736 cancel_delayed_work(&priv->pkey_poll_task); 737 mutex_unlock(&pkey_mutex); 738 if (flush) 739 flush_workqueue(ipoib_workqueue); 740 } 741 742 ipoib_mcast_stop_thread(dev, flush); 743 ipoib_mcast_dev_flush(dev); 744 745 ipoib_flush_paths(dev); 746 747 return 0; 748 } 749 750 static int recvs_pending(struct net_device *dev) 751 { 752 struct ipoib_dev_priv *priv = netdev_priv(dev); 753 int pending = 0; 754 int i; 755 756 for (i = 0; i < ipoib_recvq_size; ++i) 757 if (priv->rx_ring[i].skb) 758 ++pending; 759 760 return pending; 761 } 762 763 void ipoib_drain_cq(struct net_device *dev) 764 { 765 struct ipoib_dev_priv *priv = netdev_priv(dev); 766 int i, n; 767 768 /* 769 * We call completion handling routines that expect to be 770 * called from the BH-disabled NAPI poll context, so disable 771 * BHs here too. 772 */ 773 local_bh_disable(); 774 775 do { 776 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc); 777 for (i = 0; i < n; ++i) { 778 /* 779 * Convert any successful completions to flush 780 * errors to avoid passing packets up the 781 * stack after bringing the device down. 782 */ 783 if (priv->ibwc[i].status == IB_WC_SUCCESS) 784 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR; 785 786 if (priv->ibwc[i].wr_id & IPOIB_OP_RECV) { 787 if (priv->ibwc[i].wr_id & IPOIB_OP_CM) 788 ipoib_cm_handle_rx_wc(dev, priv->ibwc + i); 789 else 790 ipoib_ib_handle_rx_wc(dev, priv->ibwc + i); 791 } else 792 ipoib_cm_handle_tx_wc(dev, priv->ibwc + i); 793 } 794 } while (n == IPOIB_NUM_WC); 795 796 while (poll_tx(priv)) 797 ; /* nothing */ 798 799 local_bh_enable(); 800 } 801 802 int ipoib_ib_dev_stop(struct net_device *dev, int flush) 803 { 804 struct ipoib_dev_priv *priv = netdev_priv(dev); 805 struct ib_qp_attr qp_attr; 806 unsigned long begin; 807 struct ipoib_tx_buf *tx_req; 808 int i; 809 810 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) 811 napi_disable(&priv->napi); 812 813 ipoib_cm_dev_stop(dev); 814 815 /* 816 * Move our QP to the error state and then reinitialize in 817 * when all work requests have completed or have been flushed. 818 */ 819 qp_attr.qp_state = IB_QPS_ERR; 820 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 821 ipoib_warn(priv, "Failed to modify QP to ERROR state\n"); 822 823 /* Wait for all sends and receives to complete */ 824 begin = jiffies; 825 826 while (priv->tx_head != priv->tx_tail || recvs_pending(dev)) { 827 if (time_after(jiffies, begin + 5 * HZ)) { 828 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n", 829 priv->tx_head - priv->tx_tail, recvs_pending(dev)); 830 831 /* 832 * assume the HW is wedged and just free up 833 * all our pending work requests. 834 */ 835 while ((int) priv->tx_tail - (int) priv->tx_head < 0) { 836 tx_req = &priv->tx_ring[priv->tx_tail & 837 (ipoib_sendq_size - 1)]; 838 ipoib_dma_unmap_tx(priv->ca, tx_req); 839 dev_kfree_skb_any(tx_req->skb); 840 ++priv->tx_tail; 841 --priv->tx_outstanding; 842 } 843 844 for (i = 0; i < ipoib_recvq_size; ++i) { 845 struct ipoib_rx_buf *rx_req; 846 847 rx_req = &priv->rx_ring[i]; 848 if (!rx_req->skb) 849 continue; 850 ipoib_ud_dma_unmap_rx(priv, 851 priv->rx_ring[i].mapping); 852 dev_kfree_skb_any(rx_req->skb); 853 rx_req->skb = NULL; 854 } 855 856 goto timeout; 857 } 858 859 ipoib_drain_cq(dev); 860 861 msleep(1); 862 } 863 864 ipoib_dbg(priv, "All sends and receives done.\n"); 865 866 timeout: 867 del_timer_sync(&priv->poll_timer); 868 qp_attr.qp_state = IB_QPS_RESET; 869 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 870 ipoib_warn(priv, "Failed to modify QP to RESET state\n"); 871 872 /* Wait for all AHs to be reaped */ 873 set_bit(IPOIB_STOP_REAPER, &priv->flags); 874 cancel_delayed_work(&priv->ah_reap_task); 875 if (flush) 876 flush_workqueue(ipoib_workqueue); 877 878 begin = jiffies; 879 880 while (!list_empty(&priv->dead_ahs)) { 881 __ipoib_reap_ah(dev); 882 883 if (time_after(jiffies, begin + HZ)) { 884 ipoib_warn(priv, "timing out; will leak address handles\n"); 885 break; 886 } 887 888 msleep(1); 889 } 890 891 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP); 892 893 return 0; 894 } 895 896 int ipoib_ib_dev_init(struct net_device *dev, struct ib_device *ca, int port) 897 { 898 struct ipoib_dev_priv *priv = netdev_priv(dev); 899 900 priv->ca = ca; 901 priv->port = port; 902 priv->qp = NULL; 903 904 if (ipoib_transport_dev_init(dev, ca)) { 905 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name); 906 return -ENODEV; 907 } 908 909 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func, 910 (unsigned long) dev); 911 912 if (dev->flags & IFF_UP) { 913 if (ipoib_ib_dev_open(dev)) { 914 ipoib_transport_dev_cleanup(dev); 915 return -ENODEV; 916 } 917 } 918 919 return 0; 920 } 921 922 static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, 923 enum ipoib_flush_level level) 924 { 925 struct ipoib_dev_priv *cpriv; 926 struct net_device *dev = priv->dev; 927 u16 new_index; 928 929 mutex_lock(&priv->vlan_mutex); 930 931 /* 932 * Flush any child interfaces too -- they might be up even if 933 * the parent is down. 934 */ 935 list_for_each_entry(cpriv, &priv->child_intfs, list) 936 __ipoib_ib_dev_flush(cpriv, level); 937 938 mutex_unlock(&priv->vlan_mutex); 939 940 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) { 941 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n"); 942 return; 943 } 944 945 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 946 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n"); 947 return; 948 } 949 950 if (level == IPOIB_FLUSH_HEAVY) { 951 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) { 952 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 953 ipoib_ib_dev_down(dev, 0); 954 ipoib_ib_dev_stop(dev, 0); 955 if (ipoib_pkey_dev_delay_open(dev)) 956 return; 957 } 958 959 /* restart QP only if P_Key index is changed */ 960 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) && 961 new_index == priv->pkey_index) { 962 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n"); 963 return; 964 } 965 priv->pkey_index = new_index; 966 } 967 968 if (level == IPOIB_FLUSH_LIGHT) { 969 ipoib_mark_paths_invalid(dev); 970 ipoib_mcast_dev_flush(dev); 971 } 972 973 if (level >= IPOIB_FLUSH_NORMAL) 974 ipoib_ib_dev_down(dev, 0); 975 976 if (level == IPOIB_FLUSH_HEAVY) { 977 ipoib_ib_dev_stop(dev, 0); 978 ipoib_ib_dev_open(dev); 979 } 980 981 /* 982 * The device could have been brought down between the start and when 983 * we get here, don't bring it back up if it's not configured up 984 */ 985 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 986 if (level >= IPOIB_FLUSH_NORMAL) 987 ipoib_ib_dev_up(dev); 988 ipoib_mcast_restart_task(&priv->restart_task); 989 } 990 } 991 992 void ipoib_ib_dev_flush_light(struct work_struct *work) 993 { 994 struct ipoib_dev_priv *priv = 995 container_of(work, struct ipoib_dev_priv, flush_light); 996 997 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT); 998 } 999 1000 void ipoib_ib_dev_flush_normal(struct work_struct *work) 1001 { 1002 struct ipoib_dev_priv *priv = 1003 container_of(work, struct ipoib_dev_priv, flush_normal); 1004 1005 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL); 1006 } 1007 1008 void ipoib_ib_dev_flush_heavy(struct work_struct *work) 1009 { 1010 struct ipoib_dev_priv *priv = 1011 container_of(work, struct ipoib_dev_priv, flush_heavy); 1012 1013 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY); 1014 } 1015 1016 void ipoib_ib_dev_cleanup(struct net_device *dev) 1017 { 1018 struct ipoib_dev_priv *priv = netdev_priv(dev); 1019 1020 ipoib_dbg(priv, "cleaning up ib_dev\n"); 1021 1022 ipoib_mcast_stop_thread(dev, 1); 1023 ipoib_mcast_dev_flush(dev); 1024 1025 ipoib_transport_dev_cleanup(dev); 1026 } 1027 1028 /* 1029 * Delayed P_Key Assigment Interim Support 1030 * 1031 * The following is initial implementation of delayed P_Key assigment 1032 * mechanism. It is using the same approach implemented for the multicast 1033 * group join. The single goal of this implementation is to quickly address 1034 * Bug #2507. This implementation will probably be removed when the P_Key 1035 * change async notification is available. 1036 */ 1037 1038 void ipoib_pkey_poll(struct work_struct *work) 1039 { 1040 struct ipoib_dev_priv *priv = 1041 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work); 1042 struct net_device *dev = priv->dev; 1043 1044 ipoib_pkey_dev_check_presence(dev); 1045 1046 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) 1047 ipoib_open(dev); 1048 else { 1049 mutex_lock(&pkey_mutex); 1050 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags)) 1051 queue_delayed_work(ipoib_workqueue, 1052 &priv->pkey_poll_task, 1053 HZ); 1054 mutex_unlock(&pkey_mutex); 1055 } 1056 } 1057 1058 int ipoib_pkey_dev_delay_open(struct net_device *dev) 1059 { 1060 struct ipoib_dev_priv *priv = netdev_priv(dev); 1061 1062 /* Look for the interface pkey value in the IB Port P_Key table and */ 1063 /* set the interface pkey assigment flag */ 1064 ipoib_pkey_dev_check_presence(dev); 1065 1066 /* P_Key value not assigned yet - start polling */ 1067 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 1068 mutex_lock(&pkey_mutex); 1069 clear_bit(IPOIB_PKEY_STOP, &priv->flags); 1070 queue_delayed_work(ipoib_workqueue, 1071 &priv->pkey_poll_task, 1072 HZ); 1073 mutex_unlock(&pkey_mutex); 1074 return 1; 1075 } 1076 1077 return 0; 1078 } 1079