ipoib_ib.c revision 331769
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 "ipoib.h" 37 38#include <rdma/ib_cache.h> 39 40#include <security/mac/mac_framework.h> 41 42#include <linux/delay.h> 43#include <linux/dma-mapping.h> 44 45#ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA 46static int data_debug_level; 47 48module_param(data_debug_level, int, 0644); 49MODULE_PARM_DESC(data_debug_level, 50 "Enable data path debug tracing if > 0"); 51#endif 52 53static DEFINE_MUTEX(pkey_mutex); 54 55struct ipoib_ah *ipoib_create_ah(struct ipoib_dev_priv *priv, 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->priv = priv; 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(priv, "Created ah %p\n", ah->ah); 74 75 return ah; 76} 77 78void 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 = ah->priv; 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 90void 91ipoib_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req) 92{ 93 struct mbuf *m; 94 int i; 95 96 for (i = 0, m = rx_req->mb; m != NULL; m = m->m_next, i++) 97 ib_dma_unmap_single(priv->ca, rx_req->mapping[i], m->m_len, 98 DMA_FROM_DEVICE); 99} 100 101void 102ipoib_dma_mb(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int length) 103{ 104 105 m_adj(mb, -(mb->m_pkthdr.len - length)); 106} 107 108struct mbuf * 109ipoib_alloc_map_mb(struct ipoib_dev_priv *priv, struct ipoib_rx_buf *rx_req, 110 int size) 111{ 112 struct mbuf *mb, *m; 113 int i, j; 114 115 rx_req->mb = NULL; 116 mb = m_getm2(NULL, size, M_NOWAIT, MT_DATA, M_PKTHDR); 117 if (mb == NULL) 118 return (NULL); 119 for (i = 0, m = mb; m != NULL; m = m->m_next, i++) { 120 m->m_len = M_SIZE(m); 121 mb->m_pkthdr.len += m->m_len; 122 rx_req->mapping[i] = ib_dma_map_single(priv->ca, 123 mtod(m, void *), m->m_len, DMA_FROM_DEVICE); 124 if (unlikely(ib_dma_mapping_error(priv->ca, 125 rx_req->mapping[i]))) 126 goto error; 127 128 } 129 rx_req->mb = mb; 130 return (mb); 131error: 132 for (j = 0, m = mb; j < i; m = m->m_next, j++) 133 ib_dma_unmap_single(priv->ca, rx_req->mapping[j], m->m_len, 134 DMA_FROM_DEVICE); 135 m_freem(mb); 136 return (NULL); 137 138} 139 140static int ipoib_ib_post_receive(struct ipoib_dev_priv *priv, int id) 141{ 142 struct ipoib_rx_buf *rx_req; 143 struct ib_recv_wr *bad_wr; 144 struct mbuf *m; 145 int ret; 146 int i; 147 148 rx_req = &priv->rx_ring[id]; 149 for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) { 150 priv->rx_sge[i].addr = rx_req->mapping[i]; 151 priv->rx_sge[i].length = m->m_len; 152 } 153 priv->rx_wr.num_sge = i; 154 priv->rx_wr.wr_id = id | IPOIB_OP_RECV; 155 156 ret = ib_post_recv(priv->qp, &priv->rx_wr, &bad_wr); 157 if (unlikely(ret)) { 158 ipoib_warn(priv, "receive failed for buf %d (%d)\n", id, ret); 159 ipoib_dma_unmap_rx(priv, &priv->rx_ring[id]); 160 m_freem(priv->rx_ring[id].mb); 161 priv->rx_ring[id].mb = NULL; 162 } 163 164 return ret; 165} 166 167static struct mbuf * 168ipoib_alloc_rx_mb(struct ipoib_dev_priv *priv, int id) 169{ 170 171 return ipoib_alloc_map_mb(priv, &priv->rx_ring[id], 172 priv->max_ib_mtu + IB_GRH_BYTES); 173} 174 175static int ipoib_ib_post_receives(struct ipoib_dev_priv *priv) 176{ 177 int i; 178 179 for (i = 0; i < ipoib_recvq_size; ++i) { 180 if (!ipoib_alloc_rx_mb(priv, i)) { 181 ipoib_warn(priv, "failed to allocate receive buffer %d\n", i); 182 return -ENOMEM; 183 } 184 if (ipoib_ib_post_receive(priv, i)) { 185 ipoib_warn(priv, "ipoib_ib_post_receive failed for buf %d\n", i); 186 return -EIO; 187 } 188 } 189 190 return 0; 191} 192 193static void 194ipoib_ib_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc) 195{ 196 struct ipoib_rx_buf saverx; 197 unsigned int wr_id = wc->wr_id & ~IPOIB_OP_RECV; 198 struct ifnet *dev = priv->dev; 199 struct ipoib_header *eh; 200 struct mbuf *mb; 201 202 ipoib_dbg_data(priv, "recv completion: id %d, status: %d\n", 203 wr_id, wc->status); 204 205 if (unlikely(wr_id >= ipoib_recvq_size)) { 206 ipoib_warn(priv, "recv completion event with wrid %d (> %d)\n", 207 wr_id, ipoib_recvq_size); 208 return; 209 } 210 211 mb = priv->rx_ring[wr_id].mb; 212 213 if (unlikely(wc->status != IB_WC_SUCCESS)) { 214 if (wc->status != IB_WC_WR_FLUSH_ERR) { 215 ipoib_warn(priv, "failed recv event " 216 "(status=%d, wrid=%d vend_err %x)\n", 217 wc->status, wr_id, wc->vendor_err); 218 goto repost; 219 } 220 if (mb) { 221 ipoib_dma_unmap_rx(priv, &priv->rx_ring[wr_id]); 222 m_freem(mb); 223 priv->rx_ring[wr_id].mb = NULL; 224 } 225 return; 226 } 227 228 /* 229 * Drop packets that this interface sent, ie multicast packets 230 * that the HCA has replicated. 231 */ 232 if (wc->slid == priv->local_lid && wc->src_qp == priv->qp->qp_num) 233 goto repost; 234 235 memcpy(&saverx, &priv->rx_ring[wr_id], sizeof(saverx)); 236 /* 237 * If we can't allocate a new RX buffer, dump 238 * this packet and reuse the old buffer. 239 */ 240 if (unlikely(!ipoib_alloc_rx_mb(priv, wr_id))) { 241 memcpy(&priv->rx_ring[wr_id], &saverx, sizeof(saverx)); 242 if_inc_counter(dev, IFCOUNTER_IQDROPS, 1); 243 goto repost; 244 } 245 246 ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n", 247 wc->byte_len, wc->slid); 248 249 ipoib_dma_unmap_rx(priv, &saverx); 250 ipoib_dma_mb(priv, mb, wc->byte_len); 251 252 if_inc_counter(dev, IFCOUNTER_IPACKETS, 1); 253 if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len); 254 mb->m_pkthdr.rcvif = dev; 255 m_adj(mb, sizeof(struct ib_grh) - INFINIBAND_ALEN); 256 eh = mtod(mb, struct ipoib_header *); 257 bzero(eh->hwaddr, 4); /* Zero the queue pair, only dgid is in grh */ 258 259 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags) && likely(wc->wc_flags & IB_WC_IP_CSUM_OK)) 260 mb->m_pkthdr.csum_flags = CSUM_IP_CHECKED | CSUM_IP_VALID; 261 262 dev->if_input(dev, mb); 263 264repost: 265 if (unlikely(ipoib_ib_post_receive(priv, wr_id))) 266 ipoib_warn(priv, "ipoib_ib_post_receive failed " 267 "for buf %d\n", wr_id); 268} 269 270int ipoib_dma_map_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req, int max) 271{ 272 struct mbuf *mb = tx_req->mb; 273 u64 *mapping = tx_req->mapping; 274 struct mbuf *m, *p; 275 int error; 276 int i; 277 278 for (m = mb, p = NULL, i = 0; m != NULL; p = m, m = m->m_next, i++) { 279 if (m->m_len != 0) 280 continue; 281 if (p == NULL) 282 panic("ipoib_dma_map_tx: First mbuf empty\n"); 283 p->m_next = m_free(m); 284 m = p; 285 i--; 286 } 287 i--; 288 if (i >= max) { 289 tx_req->mb = mb = m_defrag(mb, M_NOWAIT); 290 if (mb == NULL) 291 return -EIO; 292 for (m = mb, i = 0; m != NULL; m = m->m_next, i++); 293 if (i >= max) 294 return -EIO; 295 } 296 error = 0; 297 for (m = mb, i = 0; m != NULL; m = m->m_next, i++) { 298 mapping[i] = ib_dma_map_single(ca, mtod(m, void *), 299 m->m_len, DMA_TO_DEVICE); 300 if (unlikely(ib_dma_mapping_error(ca, mapping[i]))) { 301 error = -EIO; 302 break; 303 } 304 } 305 if (error) { 306 int end; 307 308 end = i; 309 for (m = mb, i = 0; i < end; m = m->m_next, i++) 310 ib_dma_unmap_single(ca, mapping[i], m->m_len, 311 DMA_TO_DEVICE); 312 } 313 return error; 314} 315 316void ipoib_dma_unmap_tx(struct ib_device *ca, struct ipoib_tx_buf *tx_req) 317{ 318 struct mbuf *mb = tx_req->mb; 319 u64 *mapping = tx_req->mapping; 320 struct mbuf *m; 321 int i; 322 323 for (m = mb, i = 0; m != NULL; m = m->m_next, i++) 324 ib_dma_unmap_single(ca, mapping[i], m->m_len, DMA_TO_DEVICE); 325} 326 327static void ipoib_ib_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc) 328{ 329 struct ifnet *dev = priv->dev; 330 unsigned int wr_id = wc->wr_id; 331 struct ipoib_tx_buf *tx_req; 332 333 ipoib_dbg_data(priv, "send completion: id %d, status: %d\n", 334 wr_id, wc->status); 335 336 if (unlikely(wr_id >= ipoib_sendq_size)) { 337 ipoib_warn(priv, "send completion event with wrid %d (> %d)\n", 338 wr_id, ipoib_sendq_size); 339 return; 340 } 341 342 tx_req = &priv->tx_ring[wr_id]; 343 344 ipoib_dma_unmap_tx(priv->ca, tx_req); 345 346 if_inc_counter(dev, IFCOUNTER_OPACKETS, 1); 347 348 m_freem(tx_req->mb); 349 350 ++priv->tx_tail; 351 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) && 352 (dev->if_drv_flags & IFF_DRV_OACTIVE) && 353 test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) 354 dev->if_drv_flags &= ~IFF_DRV_OACTIVE; 355 356 if (wc->status != IB_WC_SUCCESS && 357 wc->status != IB_WC_WR_FLUSH_ERR) 358 ipoib_warn(priv, "failed send event " 359 "(status=%d, wrid=%d vend_err %x)\n", 360 wc->status, wr_id, wc->vendor_err); 361} 362 363int 364ipoib_poll_tx(struct ipoib_dev_priv *priv) 365{ 366 int n, i; 367 368 n = ib_poll_cq(priv->send_cq, MAX_SEND_CQE, priv->send_wc); 369 for (i = 0; i < n; ++i) { 370 struct ib_wc *wc = priv->send_wc + i; 371 if (wc->wr_id & IPOIB_OP_CM) 372 ipoib_cm_handle_tx_wc(priv, wc); 373 else 374 ipoib_ib_handle_tx_wc(priv, wc); 375 } 376 377 return n == MAX_SEND_CQE; 378} 379 380static void 381ipoib_poll(struct ipoib_dev_priv *priv) 382{ 383 int n, i; 384 385poll_more: 386 spin_lock(&priv->drain_lock); 387 for (;;) { 388 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc); 389 for (i = 0; i < n; i++) { 390 struct ib_wc *wc = priv->ibwc + i; 391 392 if ((wc->wr_id & IPOIB_OP_RECV) == 0) 393 panic("ipoib_poll: Bad wr_id 0x%jX\n", 394 (intmax_t)wc->wr_id); 395 if (wc->wr_id & IPOIB_OP_CM) 396 ipoib_cm_handle_rx_wc(priv, wc); 397 else 398 ipoib_ib_handle_rx_wc(priv, wc); 399 } 400 401 if (n != IPOIB_NUM_WC) 402 break; 403 } 404 spin_unlock(&priv->drain_lock); 405 406 if (ib_req_notify_cq(priv->recv_cq, 407 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS)) 408 goto poll_more; 409} 410 411void ipoib_ib_completion(struct ib_cq *cq, void *dev_ptr) 412{ 413 struct ipoib_dev_priv *priv = dev_ptr; 414 415 ipoib_poll(priv); 416} 417 418static void drain_tx_cq(struct ipoib_dev_priv *priv) 419{ 420 struct ifnet *dev = priv->dev; 421 422 spin_lock(&priv->lock); 423 while (ipoib_poll_tx(priv)) 424 ; /* nothing */ 425 426 if (dev->if_drv_flags & IFF_DRV_OACTIVE) 427 mod_timer(&priv->poll_timer, jiffies + 1); 428 429 spin_unlock(&priv->lock); 430} 431 432void ipoib_send_comp_handler(struct ib_cq *cq, void *dev_ptr) 433{ 434 struct ipoib_dev_priv *priv = dev_ptr; 435 436 mod_timer(&priv->poll_timer, jiffies); 437} 438 439static inline int 440post_send(struct ipoib_dev_priv *priv, unsigned int wr_id, 441 struct ib_ah *address, u32 qpn, struct ipoib_tx_buf *tx_req, void *head, 442 int hlen) 443{ 444 struct ib_send_wr *bad_wr; 445 struct mbuf *mb = tx_req->mb; 446 u64 *mapping = tx_req->mapping; 447 struct mbuf *m; 448 int i; 449 450 for (m = mb, i = 0; m != NULL; m = m->m_next, i++) { 451 priv->tx_sge[i].addr = mapping[i]; 452 priv->tx_sge[i].length = m->m_len; 453 } 454 priv->tx_wr.wr.num_sge = i; 455 priv->tx_wr.wr.wr_id = wr_id; 456 priv->tx_wr.remote_qpn = qpn; 457 priv->tx_wr.ah = address; 458 459 if (head) { 460 priv->tx_wr.mss = 0; /* XXX mb_shinfo(mb)->gso_size; */ 461 priv->tx_wr.header = head; 462 priv->tx_wr.hlen = hlen; 463 priv->tx_wr.wr.opcode = IB_WR_LSO; 464 } else 465 priv->tx_wr.wr.opcode = IB_WR_SEND; 466 467 return ib_post_send(priv->qp, &priv->tx_wr.wr, &bad_wr); 468} 469 470void 471ipoib_send(struct ipoib_dev_priv *priv, struct mbuf *mb, 472 struct ipoib_ah *address, u32 qpn) 473{ 474 struct ifnet *dev = priv->dev; 475 struct ipoib_tx_buf *tx_req; 476 int hlen; 477 void *phead; 478 479 if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) 480 while (ipoib_poll_tx(priv)) 481 ; /* nothing */ 482 483 m_adj(mb, sizeof (struct ipoib_pseudoheader)); 484 if (0 /* XXX segment offload mb_is_gso(mb) */) { 485 /* XXX hlen = mb_transport_offset(mb) + tcp_hdrlen(mb); */ 486 phead = mtod(mb, void *); 487 if (mb->m_len < hlen) { 488 ipoib_warn(priv, "linear data too small\n"); 489 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 490 m_freem(mb); 491 return; 492 } 493 m_adj(mb, hlen); 494 } else { 495 if (unlikely(mb->m_pkthdr.len - IPOIB_ENCAP_LEN > priv->mcast_mtu)) { 496 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n", 497 mb->m_pkthdr.len, priv->mcast_mtu); 498 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 499 ipoib_cm_mb_too_long(priv, mb, priv->mcast_mtu); 500 return; 501 } 502 phead = NULL; 503 hlen = 0; 504 } 505 506 ipoib_dbg_data(priv, "sending packet, length=%d address=%p qpn=0x%06x\n", 507 mb->m_pkthdr.len, address, qpn); 508 509 /* 510 * We put the mb into the tx_ring _before_ we call post_send() 511 * because it's entirely possible that the completion handler will 512 * run before we execute anything after the post_send(). That 513 * means we have to make sure everything is properly recorded and 514 * our state is consistent before we call post_send(). 515 */ 516 tx_req = &priv->tx_ring[priv->tx_head & (ipoib_sendq_size - 1)]; 517 tx_req->mb = mb; 518 if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req, IPOIB_UD_TX_SG))) { 519 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 520 if (tx_req->mb) 521 m_freem(tx_req->mb); 522 return; 523 } 524 525 if (mb->m_pkthdr.csum_flags & (CSUM_IP|CSUM_TCP|CSUM_UDP)) 526 priv->tx_wr.wr.send_flags |= IB_SEND_IP_CSUM; 527 else 528 priv->tx_wr.wr.send_flags &= ~IB_SEND_IP_CSUM; 529 530 if (++priv->tx_outstanding == ipoib_sendq_size) { 531 ipoib_dbg(priv, "TX ring full, stopping kernel net queue\n"); 532 if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP)) 533 ipoib_warn(priv, "request notify on send CQ failed\n"); 534 dev->if_drv_flags |= IFF_DRV_OACTIVE; 535 } 536 537 if (unlikely(post_send(priv, 538 priv->tx_head & (ipoib_sendq_size - 1), address->ah, qpn, 539 tx_req, phead, hlen))) { 540 ipoib_warn(priv, "post_send failed\n"); 541 if_inc_counter(dev, IFCOUNTER_OERRORS, 1); 542 --priv->tx_outstanding; 543 ipoib_dma_unmap_tx(priv->ca, tx_req); 544 m_freem(mb); 545 if (dev->if_drv_flags & IFF_DRV_OACTIVE) 546 dev->if_drv_flags &= ~IFF_DRV_OACTIVE; 547 } else { 548 address->last_send = priv->tx_head; 549 ++priv->tx_head; 550 } 551} 552 553static void __ipoib_reap_ah(struct ipoib_dev_priv *priv) 554{ 555 struct ipoib_ah *ah, *tah; 556 LIST_HEAD(remove_list); 557 unsigned long flags; 558 559 spin_lock_irqsave(&priv->lock, flags); 560 561 list_for_each_entry_safe(ah, tah, &priv->dead_ahs, list) 562 if ((int) priv->tx_tail - (int) ah->last_send >= 0) { 563 list_del(&ah->list); 564 ib_destroy_ah(ah->ah); 565 kfree(ah); 566 } 567 568 spin_unlock_irqrestore(&priv->lock, flags); 569} 570 571void ipoib_reap_ah(struct work_struct *work) 572{ 573 struct ipoib_dev_priv *priv = 574 container_of(work, struct ipoib_dev_priv, ah_reap_task.work); 575 576 __ipoib_reap_ah(priv); 577 578 if (!test_bit(IPOIB_STOP_REAPER, &priv->flags)) 579 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, 580 HZ); 581} 582 583static void ipoib_ah_dev_cleanup(struct ipoib_dev_priv *priv) 584{ 585 unsigned long begin; 586 587 begin = jiffies; 588 589 while (!list_empty(&priv->dead_ahs)) { 590 __ipoib_reap_ah(priv); 591 592 if (time_after(jiffies, begin + HZ)) { 593 ipoib_warn(priv, "timing out; will leak address handles\n"); 594 break; 595 } 596 597 msleep(1); 598 } 599} 600 601static void ipoib_ib_tx_timer_func(unsigned long ctx) 602{ 603 drain_tx_cq((struct ipoib_dev_priv *)ctx); 604} 605 606int ipoib_ib_dev_open(struct ipoib_dev_priv *priv) 607{ 608 int ret; 609 610 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &priv->pkey_index)) { 611 ipoib_warn(priv, "P_Key 0x%04x not found\n", priv->pkey); 612 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 613 return -1; 614 } 615 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 616 617 ret = ipoib_init_qp(priv); 618 if (ret) { 619 ipoib_warn(priv, "ipoib_init_qp returned %d\n", ret); 620 return -1; 621 } 622 623 ret = ipoib_ib_post_receives(priv); 624 if (ret) { 625 ipoib_warn(priv, "ipoib_ib_post_receives returned %d\n", ret); 626 ipoib_ib_dev_stop(priv, 1); 627 return -1; 628 } 629 630 ret = ipoib_cm_dev_open(priv); 631 if (ret) { 632 ipoib_warn(priv, "ipoib_cm_dev_open returned %d\n", ret); 633 ipoib_ib_dev_stop(priv, 1); 634 return -1; 635 } 636 637 clear_bit(IPOIB_STOP_REAPER, &priv->flags); 638 queue_delayed_work(ipoib_workqueue, &priv->ah_reap_task, HZ); 639 640 set_bit(IPOIB_FLAG_INITIALIZED, &priv->flags); 641 642 return 0; 643} 644 645static void ipoib_pkey_dev_check_presence(struct ipoib_dev_priv *priv) 646{ 647 u16 pkey_index = 0; 648 649 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &pkey_index)) 650 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 651 else 652 set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 653} 654 655int ipoib_ib_dev_up(struct ipoib_dev_priv *priv) 656{ 657 658 ipoib_pkey_dev_check_presence(priv); 659 660 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 661 ipoib_dbg(priv, "PKEY is not assigned.\n"); 662 return 0; 663 } 664 665 set_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 666 667 return ipoib_mcast_start_thread(priv); 668} 669 670int ipoib_ib_dev_down(struct ipoib_dev_priv *priv, int flush) 671{ 672 673 ipoib_dbg(priv, "downing ib_dev\n"); 674 675 clear_bit(IPOIB_FLAG_OPER_UP, &priv->flags); 676 if_link_state_change(priv->dev, LINK_STATE_DOWN); 677 678 /* Shutdown the P_Key thread if still active */ 679 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 680 mutex_lock(&pkey_mutex); 681 set_bit(IPOIB_PKEY_STOP, &priv->flags); 682 cancel_delayed_work(&priv->pkey_poll_task); 683 mutex_unlock(&pkey_mutex); 684 if (flush) 685 flush_workqueue(ipoib_workqueue); 686 } 687 688 ipoib_mcast_stop_thread(priv, flush); 689 ipoib_mcast_dev_flush(priv); 690 691 ipoib_flush_paths(priv); 692 693 return 0; 694} 695 696static int recvs_pending(struct ipoib_dev_priv *priv) 697{ 698 int pending = 0; 699 int i; 700 701 for (i = 0; i < ipoib_recvq_size; ++i) 702 if (priv->rx_ring[i].mb) 703 ++pending; 704 705 return pending; 706} 707 708void ipoib_drain_cq(struct ipoib_dev_priv *priv) 709{ 710 int i, n; 711 712 spin_lock(&priv->drain_lock); 713 do { 714 n = ib_poll_cq(priv->recv_cq, IPOIB_NUM_WC, priv->ibwc); 715 for (i = 0; i < n; ++i) { 716 /* 717 * Convert any successful completions to flush 718 * errors to avoid passing packets up the 719 * stack after bringing the device down. 720 */ 721 if (priv->ibwc[i].status == IB_WC_SUCCESS) 722 priv->ibwc[i].status = IB_WC_WR_FLUSH_ERR; 723 724 if ((priv->ibwc[i].wr_id & IPOIB_OP_RECV) == 0) 725 panic("ipoib_drain_cq: Bad wrid 0x%jX\n", 726 (intmax_t)priv->ibwc[i].wr_id); 727 if (priv->ibwc[i].wr_id & IPOIB_OP_CM) 728 ipoib_cm_handle_rx_wc(priv, priv->ibwc + i); 729 else 730 ipoib_ib_handle_rx_wc(priv, priv->ibwc + i); 731 } 732 } while (n == IPOIB_NUM_WC); 733 spin_unlock(&priv->drain_lock); 734 735 spin_lock(&priv->lock); 736 while (ipoib_poll_tx(priv)) 737 ; /* nothing */ 738 739 spin_unlock(&priv->lock); 740} 741 742int ipoib_ib_dev_stop(struct ipoib_dev_priv *priv, int flush) 743{ 744 struct ib_qp_attr qp_attr; 745 unsigned long begin; 746 struct ipoib_tx_buf *tx_req; 747 int i; 748 749 clear_bit(IPOIB_FLAG_INITIALIZED, &priv->flags); 750 751 ipoib_cm_dev_stop(priv); 752 753 /* 754 * Move our QP to the error state and then reinitialize in 755 * when all work requests have completed or have been flushed. 756 */ 757 qp_attr.qp_state = IB_QPS_ERR; 758 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 759 ipoib_warn(priv, "Failed to modify QP to ERROR state\n"); 760 761 /* Wait for all sends and receives to complete */ 762 begin = jiffies; 763 764 while (priv->tx_head != priv->tx_tail || recvs_pending(priv)) { 765 if (time_after(jiffies, begin + 5 * HZ)) { 766 ipoib_warn(priv, "timing out; %d sends %d receives not completed\n", 767 priv->tx_head - priv->tx_tail, recvs_pending(priv)); 768 769 /* 770 * assume the HW is wedged and just free up 771 * all our pending work requests. 772 */ 773 while ((int) priv->tx_tail - (int) priv->tx_head < 0) { 774 tx_req = &priv->tx_ring[priv->tx_tail & 775 (ipoib_sendq_size - 1)]; 776 ipoib_dma_unmap_tx(priv->ca, tx_req); 777 m_freem(tx_req->mb); 778 ++priv->tx_tail; 779 --priv->tx_outstanding; 780 } 781 782 for (i = 0; i < ipoib_recvq_size; ++i) { 783 struct ipoib_rx_buf *rx_req; 784 785 rx_req = &priv->rx_ring[i]; 786 if (!rx_req->mb) 787 continue; 788 ipoib_dma_unmap_rx(priv, &priv->rx_ring[i]); 789 m_freem(rx_req->mb); 790 rx_req->mb = NULL; 791 } 792 793 goto timeout; 794 } 795 796 ipoib_drain_cq(priv); 797 798 msleep(1); 799 } 800 801 ipoib_dbg(priv, "All sends and receives done.\n"); 802 803timeout: 804 del_timer_sync(&priv->poll_timer); 805 qp_attr.qp_state = IB_QPS_RESET; 806 if (ib_modify_qp(priv->qp, &qp_attr, IB_QP_STATE)) 807 ipoib_warn(priv, "Failed to modify QP to RESET state\n"); 808 809 /* Wait for all AHs to be reaped */ 810 set_bit(IPOIB_STOP_REAPER, &priv->flags); 811 cancel_delayed_work(&priv->ah_reap_task); 812 if (flush) 813 flush_workqueue(ipoib_workqueue); 814 815 ipoib_ah_dev_cleanup(priv); 816 817 ib_req_notify_cq(priv->recv_cq, IB_CQ_NEXT_COMP); 818 819 return 0; 820} 821 822int ipoib_ib_dev_init(struct ipoib_dev_priv *priv, struct ib_device *ca, int port) 823{ 824 struct ifnet *dev = priv->dev; 825 826 priv->ca = ca; 827 priv->port = port; 828 priv->qp = NULL; 829 830 if (ipoib_transport_dev_init(priv, ca)) { 831 printk(KERN_WARNING "%s: ipoib_transport_dev_init failed\n", ca->name); 832 return -ENODEV; 833 } 834 835 setup_timer(&priv->poll_timer, ipoib_ib_tx_timer_func, 836 (unsigned long) priv); 837 838 if (dev->if_flags & IFF_UP) { 839 if (ipoib_ib_dev_open(priv)) { 840 ipoib_transport_dev_cleanup(priv); 841 return -ENODEV; 842 } 843 } 844 845 return 0; 846} 847 848static void __ipoib_ib_dev_flush(struct ipoib_dev_priv *priv, 849 enum ipoib_flush_level level) 850{ 851 struct ipoib_dev_priv *cpriv; 852 u16 new_index; 853 854 mutex_lock(&priv->vlan_mutex); 855 856 /* 857 * Flush any child interfaces too -- they might be up even if 858 * the parent is down. 859 */ 860 list_for_each_entry(cpriv, &priv->child_intfs, list) 861 __ipoib_ib_dev_flush(cpriv, level); 862 863 mutex_unlock(&priv->vlan_mutex); 864 865 if (!test_bit(IPOIB_FLAG_INITIALIZED, &priv->flags)) { 866 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_INITIALIZED not set.\n"); 867 return; 868 } 869 870 if (!test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 871 ipoib_dbg(priv, "Not flushing - IPOIB_FLAG_ADMIN_UP not set.\n"); 872 return; 873 } 874 875 if (level == IPOIB_FLUSH_HEAVY) { 876 if (ib_find_pkey(priv->ca, priv->port, priv->pkey, &new_index)) { 877 clear_bit(IPOIB_PKEY_ASSIGNED, &priv->flags); 878 ipoib_ib_dev_down(priv, 0); 879 ipoib_ib_dev_stop(priv, 0); 880 if (ipoib_pkey_dev_delay_open(priv)) 881 return; 882 } 883 884 /* restart QP only if P_Key index is changed */ 885 if (test_and_set_bit(IPOIB_PKEY_ASSIGNED, &priv->flags) && 886 new_index == priv->pkey_index) { 887 ipoib_dbg(priv, "Not flushing - P_Key index not changed.\n"); 888 return; 889 } 890 priv->pkey_index = new_index; 891 } 892 893 if (level == IPOIB_FLUSH_LIGHT) { 894 ipoib_mark_paths_invalid(priv); 895 ipoib_mcast_dev_flush(priv); 896 } 897 898 if (level >= IPOIB_FLUSH_NORMAL) 899 ipoib_ib_dev_down(priv, 0); 900 901 if (level == IPOIB_FLUSH_HEAVY) { 902 ipoib_ib_dev_stop(priv, 0); 903 ipoib_ib_dev_open(priv); 904 } 905 906 /* 907 * The device could have been brought down between the start and when 908 * we get here, don't bring it back up if it's not configured up 909 */ 910 if (test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)) { 911 if (level >= IPOIB_FLUSH_NORMAL) 912 ipoib_ib_dev_up(priv); 913 ipoib_mcast_restart_task(&priv->restart_task); 914 } 915} 916 917void ipoib_ib_dev_flush_light(struct work_struct *work) 918{ 919 struct ipoib_dev_priv *priv = 920 container_of(work, struct ipoib_dev_priv, flush_light); 921 922 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_LIGHT); 923} 924 925void ipoib_ib_dev_flush_normal(struct work_struct *work) 926{ 927 struct ipoib_dev_priv *priv = 928 container_of(work, struct ipoib_dev_priv, flush_normal); 929 930 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_NORMAL); 931} 932 933void ipoib_ib_dev_flush_heavy(struct work_struct *work) 934{ 935 struct ipoib_dev_priv *priv = 936 container_of(work, struct ipoib_dev_priv, flush_heavy); 937 938 __ipoib_ib_dev_flush(priv, IPOIB_FLUSH_HEAVY); 939} 940 941void ipoib_ib_dev_cleanup(struct ipoib_dev_priv *priv) 942{ 943 944 ipoib_dbg(priv, "cleaning up ib_dev\n"); 945 946 ipoib_mcast_stop_thread(priv, 1); 947 ipoib_mcast_dev_flush(priv); 948 949 ipoib_ah_dev_cleanup(priv); 950 ipoib_transport_dev_cleanup(priv); 951} 952 953/* 954 * Delayed P_Key Assigment Interim Support 955 * 956 * The following is initial implementation of delayed P_Key assigment 957 * mechanism. It is using the same approach implemented for the multicast 958 * group join. The single goal of this implementation is to quickly address 959 * Bug #2507. This implementation will probably be removed when the P_Key 960 * change async notification is available. 961 */ 962 963void ipoib_pkey_poll(struct work_struct *work) 964{ 965 struct ipoib_dev_priv *priv = 966 container_of(work, struct ipoib_dev_priv, pkey_poll_task.work); 967 968 ipoib_pkey_dev_check_presence(priv); 969 970 if (test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) 971 ipoib_open(priv); 972 else { 973 mutex_lock(&pkey_mutex); 974 if (!test_bit(IPOIB_PKEY_STOP, &priv->flags)) 975 queue_delayed_work(ipoib_workqueue, 976 &priv->pkey_poll_task, 977 HZ); 978 mutex_unlock(&pkey_mutex); 979 } 980} 981 982int ipoib_pkey_dev_delay_open(struct ipoib_dev_priv *priv) 983{ 984 985 /* Look for the interface pkey value in the IB Port P_Key table and */ 986 /* set the interface pkey assigment flag */ 987 ipoib_pkey_dev_check_presence(priv); 988 989 /* P_Key value not assigned yet - start polling */ 990 if (!test_bit(IPOIB_PKEY_ASSIGNED, &priv->flags)) { 991 mutex_lock(&pkey_mutex); 992 clear_bit(IPOIB_PKEY_STOP, &priv->flags); 993 queue_delayed_work(ipoib_workqueue, 994 &priv->pkey_poll_task, 995 HZ); 996 mutex_unlock(&pkey_mutex); 997 return 1; 998 } 999 1000 return 0; 1001} 1002