err.no Git - linux-2.6/blob - drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  * $Id$
  33  */
  34
  35 #include <rdma/ib_cm.h>
  36 #include <rdma/ib_cache.h>
  37 #include <net/dst.h>
  38 #include <net/icmp.h>
  39 #include <linux/icmpv6.h>
  40 #include <linux/delay.h>
  41 #include <linux/vmalloc.h>
  42
  43 #include "ipoib.h"
  44
  45 int ipoib_max_conn_qp = 128;
  46
  47 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  48 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  49                  "Max number of connected-mode QPs per interface "
  50                  "(applied only if shared receive queue is not available)");
  51
  52 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  53 static int data_debug_level;
  54
  55 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  56 MODULE_PARM_DESC(cm_data_debug_level,
  57                  "Enable data path debug tracing for connected mode if > 0");
  58 #endif
  59
  60 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  61
  62 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  63 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  64 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  65 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  66
  67 static struct ib_qp_attr ipoib_cm_err_attr = {
  68         .qp_state = IB_QPS_ERR
  69 };
  70
  71 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  72
  73 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  74         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  75         .opcode = IB_WR_SEND,
  76 };
  77
  78 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  79                                struct ib_cm_event *event);
  80
  81 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  82                                   u64 mapping[IPOIB_CM_RX_SG])
  83 {
  84         int i;
  85
  86         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  87
  88         for (i = 0; i < frags; ++i)
  89                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  90 }
  91
  92 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  93 {
  94         struct ipoib_dev_priv *priv = netdev_priv(dev);
  95         struct ib_recv_wr *bad_wr;
  96         int i, ret;
  97
  98         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  99
 100         for (i = 0; i < priv->cm.num_frags; ++i)
 101                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 102
 103         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 104         if (unlikely(ret)) {
 105                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 106                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 107                                       priv->cm.srq_ring[id].mapping);
 108                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 109                 priv->cm.srq_ring[id].skb = NULL;
 110         }
 111
 112         return ret;
 113 }
 114
 115 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 116                                         struct ipoib_cm_rx *rx, int id)
 117 {
 118         struct ipoib_dev_priv *priv = netdev_priv(dev);
 119         struct ib_recv_wr *bad_wr;
 120         int i, ret;
 121
 122         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 123
 124         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 125                 priv->cm.rx_sge[i].addr = rx->rx_ring[id].mapping[i];
 126
 127         ret = ib_post_recv(rx->qp, &priv->cm.rx_wr, &bad_wr);
 128         if (unlikely(ret)) {
 129                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 130                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 131                                       rx->rx_ring[id].mapping);
 132                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 133                 rx->rx_ring[id].skb = NULL;
 134         }
 135
 136         return ret;
 137 }
 138
 139 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 140                                              struct ipoib_cm_rx_buf *rx_ring,
 141                                              int id, int frags,
 142                                              u64 mapping[IPOIB_CM_RX_SG])
 143 {
 144         struct ipoib_dev_priv *priv = netdev_priv(dev);
 145         struct sk_buff *skb;
 146         int i;
 147
 148         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 149         if (unlikely(!skb))
 150                 return NULL;
 151
 152         /*
 153          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 154          * IP header to a multiple of 16.
 155          */
 156         skb_reserve(skb, 12);
 157
 158         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 159                                        DMA_FROM_DEVICE);
 160         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 161                 dev_kfree_skb_any(skb);
 162                 return NULL;
 163         }
 164
 165         for (i = 0; i < frags; i++) {
 166                 struct page *page = alloc_page(GFP_ATOMIC);
 167
 168                 if (!page)
 169                         goto partial_error;
 170                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 171
 172                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 173                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 174                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 175                         goto partial_error;
 176         }
 177
 178         rx_ring[id].skb = skb;
 179         return skb;
 180
 181 partial_error:
 182
 183         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 184
 185         for (; i > 0; --i)
 186                 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 187
 188         dev_kfree_skb_any(skb);
 189         return NULL;
 190 }
 191
 192 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 193                                   struct ipoib_cm_rx_buf *rx_ring)
 194 {
 195         struct ipoib_dev_priv *priv = netdev_priv(dev);
 196         int i;
 197
 198         for (i = 0; i < ipoib_recvq_size; ++i)
 199                 if (rx_ring[i].skb) {
 200                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 201                                               rx_ring[i].mapping);
 202                         dev_kfree_skb_any(rx_ring[i].skb);
 203                 }
 204
 205         kfree(rx_ring);
 206 }
 207
 208 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 209 {
 210         struct ib_send_wr *bad_wr;
 211         struct ipoib_cm_rx *p;
 212
 213         /* We only reserved 1 extra slot in CQ for drain WRs, so
 214          * make sure we have at most 1 outstanding WR. */
 215         if (list_empty(&priv->cm.rx_flush_list) ||
 216             !list_empty(&priv->cm.rx_drain_list))
 217                 return;
 218
 219         /*
 220          * QPs on flush list are error state.  This way, a "flush
 221          * error" WC will be immediately generated for each WR we post.
 222          */
 223         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 224         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 225                 ipoib_warn(priv, "failed to post drain wr\n");
 226
 227         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 228 }
 229
 230 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 231 {
 232         struct ipoib_cm_rx *p = ctx;
 233         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 234         unsigned long flags;
 235
 236         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 237                 return;
 238
 239         spin_lock_irqsave(&priv->lock, flags);
 240         list_move(&p->list, &priv->cm.rx_flush_list);
 241         p->state = IPOIB_CM_RX_FLUSH;
 242         ipoib_cm_start_rx_drain(priv);
 243         spin_unlock_irqrestore(&priv->lock, flags);
 244 }
 245
 246 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 247                                            struct ipoib_cm_rx *p)
 248 {
 249         struct ipoib_dev_priv *priv = netdev_priv(dev);
 250         struct ib_qp_init_attr attr = {
 251                 .event_handler = ipoib_cm_rx_event_handler,
 252                 .send_cq = priv->cq, /* For drain WR */
 253                 .recv_cq = priv->cq,
 254                 .srq = priv->cm.srq,
 255                 .cap.max_send_wr = 1, /* For drain WR */
 256                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 257                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 258                 .qp_type = IB_QPT_RC,
 259                 .qp_context = p,
 260         };
 261
 262         if (!ipoib_cm_has_srq(dev)) {
 263                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 264                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 265         }
 266
 267         return ib_create_qp(priv->pd, &attr);
 268 }
 269
 270 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 271                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 272                                  unsigned psn)
 273 {
 274         struct ipoib_dev_priv *priv = netdev_priv(dev);
 275         struct ib_qp_attr qp_attr;
 276         int qp_attr_mask, ret;
 277
 278         qp_attr.qp_state = IB_QPS_INIT;
 279         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 280         if (ret) {
 281                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 282                 return ret;
 283         }
 284         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 285         if (ret) {
 286                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 287                 return ret;
 288         }
 289         qp_attr.qp_state = IB_QPS_RTR;
 290         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 291         if (ret) {
 292                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 293                 return ret;
 294         }
 295         qp_attr.rq_psn = psn;
 296         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 297         if (ret) {
 298                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 299                 return ret;
 300         }
 301
 302         /*
 303          * Current Mellanox HCA firmware won't generate completions
 304          * with error for drain WRs unless the QP has been moved to
 305          * RTS first. This work-around leaves a window where a QP has
 306          * moved to error asynchronously, but this will eventually get
 307          * fixed in firmware, so let's not error out if modify QP
 308          * fails.
 309          */
 310         qp_attr.qp_state = IB_QPS_RTS;
 311         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 312         if (ret) {
 313                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 314                 return 0;
 315         }
 316         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 317         if (ret) {
 318                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 319                 return 0;
 320         }
 321
 322         return 0;
 323 }
 324
 325 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 326                                    struct ipoib_cm_rx *rx)
 327 {
 328         struct ipoib_dev_priv *priv = netdev_priv(dev);
 329         int ret;
 330         int i;
 331
 332         rx->rx_ring = kcalloc(ipoib_recvq_size, sizeof *rx->rx_ring, GFP_KERNEL);
 333         if (!rx->rx_ring)
 334                 return -ENOMEM;
 335
 336         spin_lock_irq(&priv->lock);
 337
 338         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 339                 spin_unlock_irq(&priv->lock);
 340                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 341                 ret = -EINVAL;
 342                 goto err_free;
 343         } else
 344                 ++priv->cm.nonsrq_conn_qp;
 345
 346         spin_unlock_irq(&priv->lock);
 347
 348         for (i = 0; i < ipoib_recvq_size; ++i) {
 349                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 350                                            rx->rx_ring[i].mapping)) {
 351                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 352                                 ret = -ENOMEM;
 353                                 goto err_count;
 354                         }
 355                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, i);
 356                 if (ret) {
 357                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 358                                    "failed for buf %d\n", i);
 359                         ret = -EIO;
 360                         goto err_count;
 361                 }
 362         }
 363
 364         rx->recv_count = ipoib_recvq_size;
 365
 366         return 0;
 367
 368 err_count:
 369         spin_lock_irq(&priv->lock);
 370         --priv->cm.nonsrq_conn_qp;
 371         spin_unlock_irq(&priv->lock);
 372
 373 err_free:
 374         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 375
 376         return ret;
 377 }
 378
 379 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 380                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 381                              unsigned psn)
 382 {
 383         struct ipoib_dev_priv *priv = netdev_priv(dev);
 384         struct ipoib_cm_data data = {};
 385         struct ib_cm_rep_param rep = {};
 386
 387         data.qpn = cpu_to_be32(priv->qp->qp_num);
 388         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 389
 390         rep.private_data = &data;
 391         rep.private_data_len = sizeof data;
 392         rep.flow_control = 0;
 393         rep.rnr_retry_count = req->rnr_retry_count;
 394         rep.srq = ipoib_cm_has_srq(dev);
 395         rep.qp_num = qp->qp_num;
 396         rep.starting_psn = psn;
 397         return ib_send_cm_rep(cm_id, &rep);
 398 }
 399
 400 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 401 {
 402         struct net_device *dev = cm_id->context;
 403         struct ipoib_dev_priv *priv = netdev_priv(dev);
 404         struct ipoib_cm_rx *p;
 405         unsigned psn;
 406         int ret;
 407
 408         ipoib_dbg(priv, "REQ arrived\n");
 409         p = kzalloc(sizeof *p, GFP_KERNEL);
 410         if (!p)
 411                 return -ENOMEM;
 412         p->dev = dev;
 413         p->id = cm_id;
 414         cm_id->context = p;
 415         p->state = IPOIB_CM_RX_LIVE;
 416         p->jiffies = jiffies;
 417         INIT_LIST_HEAD(&p->list);
 418
 419         p->qp = ipoib_cm_create_rx_qp(dev, p);
 420         if (IS_ERR(p->qp)) {
 421                 ret = PTR_ERR(p->qp);
 422                 goto err_qp;
 423         }
 424
 425         psn = random32() & 0xffffff;
 426         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 427         if (ret)
 428                 goto err_modify;
 429
 430         if (!ipoib_cm_has_srq(dev)) {
 431                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 432                 if (ret)
 433                         goto err_modify;
 434         }
 435
 436         spin_lock_irq(&priv->lock);
 437         queue_delayed_work(ipoib_workqueue,
 438                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 439         /* Add this entry to passive ids list head, but do not re-add it
 440          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 441         p->jiffies = jiffies;
 442         if (p->state == IPOIB_CM_RX_LIVE)
 443                 list_move(&p->list, &priv->cm.passive_ids);
 444         spin_unlock_irq(&priv->lock);
 445
 446         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 447         if (ret) {
 448                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 449                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 450                         ipoib_warn(priv, "unable to move qp to error state\n");
 451         }
 452         return 0;
 453
 454 err_modify:
 455         ib_destroy_qp(p->qp);
 456 err_qp:
 457         kfree(p);
 458         return ret;
 459 }
 460
 461 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 462                                struct ib_cm_event *event)
 463 {
 464         struct ipoib_cm_rx *p;
 465         struct ipoib_dev_priv *priv;
 466
 467         switch (event->event) {
 468         case IB_CM_REQ_RECEIVED:
 469                 return ipoib_cm_req_handler(cm_id, event);
 470         case IB_CM_DREQ_RECEIVED:
 471                 p = cm_id->context;
 472                 ib_send_cm_drep(cm_id, NULL, 0);
 473                 /* Fall through */
 474         case IB_CM_REJ_RECEIVED:
 475                 p = cm_id->context;
 476                 priv = netdev_priv(p->dev);
 477                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 478                         ipoib_warn(priv, "unable to move qp to error state\n");
 479                 /* Fall through */
 480         default:
 481                 return 0;
 482         }
 483 }
 484 /* Adjust length of skb with fragments to match received data */
 485 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 486                           unsigned int length, struct sk_buff *toskb)
 487 {
 488         int i, num_frags;
 489         unsigned int size;
 490
 491         /* put header into skb */
 492         size = min(length, hdr_space);
 493         skb->tail += size;
 494         skb->len += size;
 495         length -= size;
 496
 497         num_frags = skb_shinfo(skb)->nr_frags;
 498         for (i = 0; i < num_frags; i++) {
 499                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 500
 501                 if (length == 0) {
 502                         /* don't need this page */
 503                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 504                         --skb_shinfo(skb)->nr_frags;
 505                 } else {
 506                         size = min(length, (unsigned) PAGE_SIZE);
 507
 508                         frag->size = size;
 509                         skb->data_len += size;
 510                         skb->truesize += size;
 511                         skb->len += size;
 512                         length -= size;
 513                 }
 514         }
 515 }
 516
 517 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 518 {
 519         struct ipoib_dev_priv *priv = netdev_priv(dev);
 520         struct ipoib_cm_rx_buf *rx_ring;
 521         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 522         struct sk_buff *skb, *newskb;
 523         struct ipoib_cm_rx *p;
 524         unsigned long flags;
 525         u64 mapping[IPOIB_CM_RX_SG];
 526         int frags;
 527         int has_srq;
 528
 529         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 530                        wr_id, wc->status);
 531
 532         if (unlikely(wr_id >= ipoib_recvq_size)) {
 533                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 534                         spin_lock_irqsave(&priv->lock, flags);
 535                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 536                         ipoib_cm_start_rx_drain(priv);
 537                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 538                         spin_unlock_irqrestore(&priv->lock, flags);
 539                 } else
 540                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 541                                    wr_id, ipoib_recvq_size);
 542                 return;
 543         }
 544
 545         p = wc->qp->qp_context;
 546
 547         has_srq = ipoib_cm_has_srq(dev);
 548         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 549
 550         skb = rx_ring[wr_id].skb;
 551
 552         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 553                 ipoib_dbg(priv, "cm recv error "
 554                            "(status=%d, wrid=%d vend_err %x)\n",
 555                            wc->status, wr_id, wc->vendor_err);
 556                 ++dev->stats.rx_dropped;
 557                 if (has_srq)
 558                         goto repost;
 559                 else {
 560                         if (!--p->recv_count) {
 561                                 spin_lock_irqsave(&priv->lock, flags);
 562                                 list_move(&p->list, &priv->cm.rx_reap_list);
 563                                 spin_unlock_irqrestore(&priv->lock, flags);
 564                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 565                         }
 566                         return;
 567                 }
 568         }
 569
 570         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 571                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 572                         spin_lock_irqsave(&priv->lock, flags);
 573                         p->jiffies = jiffies;
 574                         /* Move this entry to list head, but do not re-add it
 575                          * if it has been moved out of list. */
 576                         if (p->state == IPOIB_CM_RX_LIVE)
 577                                 list_move(&p->list, &priv->cm.passive_ids);
 578                         spin_unlock_irqrestore(&priv->lock, flags);
 579                 }
 580         }
 581
 582         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 583                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 584
 585         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 586         if (unlikely(!newskb)) {
 587                 /*
 588                  * If we can't allocate a new RX buffer, dump
 589                  * this packet and reuse the old buffer.
 590                  */
 591                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 592                 ++dev->stats.rx_dropped;
 593                 goto repost;
 594         }
 595
 596         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 597         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 598
 599         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 600                        wc->byte_len, wc->slid);
 601
 602         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 603
 604         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 605         skb_reset_mac_header(skb);
 606         skb_pull(skb, IPOIB_ENCAP_LEN);
 607
 608         dev->last_rx = jiffies;
 609         ++dev->stats.rx_packets;
 610         dev->stats.rx_bytes += skb->len;
 611
 612         skb->dev = dev;
 613         /* XXX get correct PACKET_ type here */
 614         skb->pkt_type = PACKET_HOST;
 615         netif_receive_skb(skb);
 616
 617 repost:
 618         if (has_srq) {
 619                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 620                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 621                                    "for buf %d\n", wr_id);
 622         } else {
 623                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p, wr_id))) {
 624                         --p->recv_count;
 625                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 626                                    "for buf %d\n", wr_id);
 627                 }
 628         }
 629 }
 630
 631 static inline int post_send(struct ipoib_dev_priv *priv,
 632                             struct ipoib_cm_tx *tx,
 633                             unsigned int wr_id,
 634                             u64 addr, int len)
 635 {
 636         struct ib_send_wr *bad_wr;
 637
 638         priv->tx_sge[0].addr          = addr;
 639         priv->tx_sge[0].length        = len;
 640
 641         priv->tx_wr.num_sge     = 1;
 642         priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 643
 644         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 645 }
 646
 647 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 648 {
 649         struct ipoib_dev_priv *priv = netdev_priv(dev);
 650         struct ipoib_tx_buf *tx_req;
 651         u64 addr;
 652
 653         if (unlikely(skb->len > tx->mtu)) {
 654                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 655                            skb->len, tx->mtu);
 656                 ++dev->stats.tx_dropped;
 657                 ++dev->stats.tx_errors;
 658                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 659                 return;
 660         }
 661
 662         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 663                        tx->tx_head, skb->len, tx->qp->qp_num);
 664
 665         /*
 666          * We put the skb into the tx_ring _before_ we call post_send()
 667          * because it's entirely possible that the completion handler will
 668          * run before we execute anything after the post_send().  That
 669          * means we have to make sure everything is properly recorded and
 670          * our state is consistent before we call post_send().
 671          */
 672         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 673         tx_req->skb = skb;
 674         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 675         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 676                 ++dev->stats.tx_errors;
 677                 dev_kfree_skb_any(skb);
 678                 return;
 679         }
 680
 681         tx_req->mapping[0] = addr;
 682
 683         if (unlikely(post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 684                                addr, skb->len))) {
 685                 ipoib_warn(priv, "post_send failed\n");
 686                 ++dev->stats.tx_errors;
 687                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 688                 dev_kfree_skb_any(skb);
 689         } else {
 690                 dev->trans_start = jiffies;
 691                 ++tx->tx_head;
 692
 693                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 694                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 695                                   tx->qp->qp_num);
 696                         netif_stop_queue(dev);
 697                 }
 698         }
 699 }
 700
 701 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 702 {
 703         struct ipoib_dev_priv *priv = netdev_priv(dev);
 704         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 705         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 706         struct ipoib_tx_buf *tx_req;
 707         unsigned long flags;
 708
 709         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 710                        wr_id, wc->status);
 711
 712         if (unlikely(wr_id >= ipoib_sendq_size)) {
 713                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 714                            wr_id, ipoib_sendq_size);
 715                 return;
 716         }
 717
 718         tx_req = &tx->tx_ring[wr_id];
 719
 720         ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len, DMA_TO_DEVICE);
 721
 722         /* FIXME: is this right? Shouldn't we only increment on success? */
 723         ++dev->stats.tx_packets;
 724         dev->stats.tx_bytes += tx_req->skb->len;
 725
 726         dev_kfree_skb_any(tx_req->skb);
 727
 728         spin_lock_irqsave(&priv->tx_lock, flags);
 729         ++tx->tx_tail;
 730         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 731             netif_queue_stopped(dev) &&
 732             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 733                 netif_wake_queue(dev);
 734
 735         if (wc->status != IB_WC_SUCCESS &&
 736             wc->status != IB_WC_WR_FLUSH_ERR) {
 737                 struct ipoib_neigh *neigh;
 738
 739                 ipoib_dbg(priv, "failed cm send event "
 740                            "(status=%d, wrid=%d vend_err %x)\n",
 741                            wc->status, wr_id, wc->vendor_err);
 742
 743                 spin_lock(&priv->lock);
 744                 neigh = tx->neigh;
 745
 746                 if (neigh) {
 747                         neigh->cm = NULL;
 748                         list_del(&neigh->list);
 749                         if (neigh->ah)
 750                                 ipoib_put_ah(neigh->ah);
 751                         ipoib_neigh_free(dev, neigh);
 752
 753                         tx->neigh = NULL;
 754                 }
 755
 756                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 757                         list_move(&tx->list, &priv->cm.reap_list);
 758                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 759                 }
 760
 761                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 762
 763                 spin_unlock(&priv->lock);
 764         }
 765
 766         spin_unlock_irqrestore(&priv->tx_lock, flags);
 767 }
 768
 769 int ipoib_cm_dev_open(struct net_device *dev)
 770 {
 771         struct ipoib_dev_priv *priv = netdev_priv(dev);
 772         int ret;
 773
 774         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 775                 return 0;
 776
 777         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 778         if (IS_ERR(priv->cm.id)) {
 779                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 780                 ret = PTR_ERR(priv->cm.id);
 781                 goto err_cm;
 782         }
 783
 784         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 785                            0, NULL);
 786         if (ret) {
 787                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 788                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 789                 goto err_listen;
 790         }
 791
 792         return 0;
 793
 794 err_listen:
 795         ib_destroy_cm_id(priv->cm.id);
 796 err_cm:
 797         priv->cm.id = NULL;
 798         return ret;
 799 }
 800
 801 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 802 {
 803         struct ipoib_dev_priv *priv = netdev_priv(dev);
 804         struct ipoib_cm_rx *rx, *n;
 805         LIST_HEAD(list);
 806
 807         spin_lock_irq(&priv->lock);
 808         list_splice_init(&priv->cm.rx_reap_list, &list);
 809         spin_unlock_irq(&priv->lock);
 810
 811         list_for_each_entry_safe(rx, n, &list, list) {
 812                 ib_destroy_cm_id(rx->id);
 813                 ib_destroy_qp(rx->qp);
 814                 if (!ipoib_cm_has_srq(dev)) {
 815                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 816                         spin_lock_irq(&priv->lock);
 817                         --priv->cm.nonsrq_conn_qp;
 818                         spin_unlock_irq(&priv->lock);
 819                 }
 820                 kfree(rx);
 821         }
 822 }
 823
 824 void ipoib_cm_dev_stop(struct net_device *dev)
 825 {
 826         struct ipoib_dev_priv *priv = netdev_priv(dev);
 827         struct ipoib_cm_rx *p;
 828         unsigned long begin;
 829         int ret;
 830
 831         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 832                 return;
 833
 834         ib_destroy_cm_id(priv->cm.id);
 835         priv->cm.id = NULL;
 836
 837         spin_lock_irq(&priv->lock);
 838         while (!list_empty(&priv->cm.passive_ids)) {
 839                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 840                 list_move(&p->list, &priv->cm.rx_error_list);
 841                 p->state = IPOIB_CM_RX_ERROR;
 842                 spin_unlock_irq(&priv->lock);
 843                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 844                 if (ret)
 845                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 846                 spin_lock_irq(&priv->lock);
 847         }
 848
 849         /* Wait for all RX to be drained */
 850         begin = jiffies;
 851
 852         while (!list_empty(&priv->cm.rx_error_list) ||
 853                !list_empty(&priv->cm.rx_flush_list) ||
 854                !list_empty(&priv->cm.rx_drain_list)) {
 855                 if (time_after(jiffies, begin + 5 * HZ)) {
 856                         ipoib_warn(priv, "RX drain timing out\n");
 857
 858                         /*
 859                          * assume the HW is wedged and just free up everything.
 860                          */
 861                         list_splice_init(&priv->cm.rx_flush_list,
 862                                          &priv->cm.rx_reap_list);
 863                         list_splice_init(&priv->cm.rx_error_list,
 864                                          &priv->cm.rx_reap_list);
 865                         list_splice_init(&priv->cm.rx_drain_list,
 866                                          &priv->cm.rx_reap_list);
 867                         break;
 868                 }
 869                 spin_unlock_irq(&priv->lock);
 870                 msleep(1);
 871                 ipoib_drain_cq(dev);
 872                 spin_lock_irq(&priv->lock);
 873         }
 874
 875         spin_unlock_irq(&priv->lock);
 876
 877         ipoib_cm_free_rx_reap_list(dev);
 878
 879         cancel_delayed_work(&priv->cm.stale_task);
 880 }
 881
 882 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 883 {
 884         struct ipoib_cm_tx *p = cm_id->context;
 885         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 886         struct ipoib_cm_data *data = event->private_data;
 887         struct sk_buff_head skqueue;
 888         struct ib_qp_attr qp_attr;
 889         int qp_attr_mask, ret;
 890         struct sk_buff *skb;
 891
 892         p->mtu = be32_to_cpu(data->mtu);
 893
 894         if (p->mtu <= IPOIB_ENCAP_LEN) {
 895                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 896                            p->mtu, IPOIB_ENCAP_LEN);
 897                 return -EINVAL;
 898         }
 899
 900         qp_attr.qp_state = IB_QPS_RTR;
 901         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 902         if (ret) {
 903                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 904                 return ret;
 905         }
 906
 907         qp_attr.rq_psn = 0 /* FIXME */;
 908         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 909         if (ret) {
 910                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 911                 return ret;
 912         }
 913
 914         qp_attr.qp_state = IB_QPS_RTS;
 915         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 916         if (ret) {
 917                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 918                 return ret;
 919         }
 920         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 921         if (ret) {
 922                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 923                 return ret;
 924         }
 925
 926         skb_queue_head_init(&skqueue);
 927
 928         spin_lock_irq(&priv->lock);
 929         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 930         if (p->neigh)
 931                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 932                         __skb_queue_tail(&skqueue, skb);
 933         spin_unlock_irq(&priv->lock);
 934
 935         while ((skb = __skb_dequeue(&skqueue))) {
 936                 skb->dev = p->dev;
 937                 if (dev_queue_xmit(skb))
 938                         ipoib_warn(priv, "dev_queue_xmit failed "
 939                                    "to requeue packet\n");
 940         }
 941
 942         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 943         if (ret) {
 944                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 945                 return ret;
 946         }
 947         return 0;
 948 }
 949
 950 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
 951 {
 952         struct ipoib_dev_priv *priv = netdev_priv(dev);
 953         struct ib_qp_init_attr attr = {
 954                 .send_cq                = priv->cq,
 955                 .recv_cq                = priv->cq,
 956                 .srq                    = priv->cm.srq,
 957                 .cap.max_send_wr        = ipoib_sendq_size,
 958                 .cap.max_send_sge       = 1,
 959                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
 960                 .qp_type                = IB_QPT_RC,
 961                 .qp_context             = tx
 962         };
 963
 964         return ib_create_qp(priv->pd, &attr);
 965 }
 966
 967 static int ipoib_cm_send_req(struct net_device *dev,
 968                              struct ib_cm_id *id, struct ib_qp *qp,
 969                              u32 qpn,
 970                              struct ib_sa_path_rec *pathrec)
 971 {
 972         struct ipoib_dev_priv *priv = netdev_priv(dev);
 973         struct ipoib_cm_data data = {};
 974         struct ib_cm_req_param req = {};
 975
 976         data.qpn = cpu_to_be32(priv->qp->qp_num);
 977         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 978
 979         req.primary_path                = pathrec;
 980         req.alternate_path              = NULL;
 981         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 982         req.qp_num                      = qp->qp_num;
 983         req.qp_type                     = qp->qp_type;
 984         req.private_data                = &data;
 985         req.private_data_len            = sizeof data;
 986         req.flow_control                = 0;
 987
 988         req.starting_psn                = 0; /* FIXME */
 989
 990         /*
 991          * Pick some arbitrary defaults here; we could make these
 992          * module parameters if anyone cared about setting them.
 993          */
 994         req.responder_resources         = 4;
 995         req.remote_cm_response_timeout  = 20;
 996         req.local_cm_response_timeout   = 20;
 997         req.retry_count                 = 0; /* RFC draft warns against retries */
 998         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
 999         req.max_cm_retries              = 15;
1000         req.srq                         = ipoib_cm_has_srq(dev);
1001         return ib_send_cm_req(id, &req);
1002 }
1003
1004 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1005                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1006 {
1007         struct ipoib_dev_priv *priv = netdev_priv(dev);
1008         struct ib_qp_attr qp_attr;
1009         int qp_attr_mask, ret;
1010         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1011         if (ret) {
1012                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1013                 return ret;
1014         }
1015
1016         qp_attr.qp_state = IB_QPS_INIT;
1017         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1018         qp_attr.port_num = priv->port;
1019         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1020
1021         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1022         if (ret) {
1023                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1024                 return ret;
1025         }
1026         return 0;
1027 }
1028
1029 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1030                             struct ib_sa_path_rec *pathrec)
1031 {
1032         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1033         int ret;
1034
1035         p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1036         if (!p->tx_ring) {
1037                 ipoib_warn(priv, "failed to allocate tx ring\n");
1038                 ret = -ENOMEM;
1039                 goto err_tx;
1040         }
1041         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1042
1043         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1044         if (IS_ERR(p->qp)) {
1045                 ret = PTR_ERR(p->qp);
1046                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1047                 goto err_qp;
1048         }
1049
1050         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1051         if (IS_ERR(p->id)) {
1052                 ret = PTR_ERR(p->id);
1053                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1054                 goto err_id;
1055         }
1056
1057         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1058         if (ret) {
1059                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1060                 goto err_modify;
1061         }
1062
1063         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1064         if (ret) {
1065                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1066                 goto err_send_cm;
1067         }
1068
1069         ipoib_dbg(priv, "Request connection 0x%x for gid " IPOIB_GID_FMT " qpn 0x%x\n",
1070                   p->qp->qp_num, IPOIB_GID_ARG(pathrec->dgid), qpn);
1071
1072         return 0;
1073
1074 err_send_cm:
1075 err_modify:
1076         ib_destroy_cm_id(p->id);
1077 err_id:
1078         p->id = NULL;
1079         ib_destroy_qp(p->qp);
1080 err_qp:
1081         p->qp = NULL;
1082         vfree(p->tx_ring);
1083 err_tx:
1084         return ret;
1085 }
1086
1087 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1088 {
1089         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1090         struct ipoib_tx_buf *tx_req;
1091         unsigned long flags;
1092         unsigned long begin;
1093
1094         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1095                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1096
1097         if (p->id)
1098                 ib_destroy_cm_id(p->id);
1099
1100         if (p->tx_ring) {
1101                 /* Wait for all sends to complete */
1102                 begin = jiffies;
1103                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1104                         if (time_after(jiffies, begin + 5 * HZ)) {
1105                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1106                                            p->tx_head - p->tx_tail);
1107                                 goto timeout;
1108                         }
1109
1110                         msleep(1);
1111                 }
1112         }
1113
1114 timeout:
1115
1116         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1117                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1118                 ib_dma_unmap_single(priv->ca, tx_req->mapping[0], tx_req->skb->len,
1119                                     DMA_TO_DEVICE);
1120                 dev_kfree_skb_any(tx_req->skb);
1121                 ++p->tx_tail;
1122                 spin_lock_irqsave(&priv->tx_lock, flags);
1123                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1124                     netif_queue_stopped(p->dev) &&
1125                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1126                         netif_wake_queue(p->dev);
1127                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1128         }
1129
1130         if (p->qp)
1131                 ib_destroy_qp(p->qp);
1132
1133         vfree(p->tx_ring);
1134         kfree(p);
1135 }
1136
1137 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1138                                struct ib_cm_event *event)
1139 {
1140         struct ipoib_cm_tx *tx = cm_id->context;
1141         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1142         struct net_device *dev = priv->dev;
1143         struct ipoib_neigh *neigh;
1144         int ret;
1145
1146         switch (event->event) {
1147         case IB_CM_DREQ_RECEIVED:
1148                 ipoib_dbg(priv, "DREQ received.\n");
1149                 ib_send_cm_drep(cm_id, NULL, 0);
1150                 break;
1151         case IB_CM_REP_RECEIVED:
1152                 ipoib_dbg(priv, "REP received.\n");
1153                 ret = ipoib_cm_rep_handler(cm_id, event);
1154                 if (ret)
1155                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1156                                        NULL, 0, NULL, 0);
1157                 break;
1158         case IB_CM_REQ_ERROR:
1159         case IB_CM_REJ_RECEIVED:
1160         case IB_CM_TIMEWAIT_EXIT:
1161                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1162                 spin_lock_irq(&priv->tx_lock);
1163                 spin_lock(&priv->lock);
1164                 neigh = tx->neigh;
1165
1166                 if (neigh) {
1167                         neigh->cm = NULL;
1168                         list_del(&neigh->list);
1169                         if (neigh->ah)
1170                                 ipoib_put_ah(neigh->ah);
1171                         ipoib_neigh_free(dev, neigh);
1172
1173                         tx->neigh = NULL;
1174                 }
1175
1176                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1177                         list_move(&tx->list, &priv->cm.reap_list);
1178                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1179                 }
1180
1181                 spin_unlock(&priv->lock);
1182                 spin_unlock_irq(&priv->tx_lock);
1183                 break;
1184         default:
1185                 break;
1186         }
1187
1188         return 0;
1189 }
1190
1191 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1192                                        struct ipoib_neigh *neigh)
1193 {
1194         struct ipoib_dev_priv *priv = netdev_priv(dev);
1195         struct ipoib_cm_tx *tx;
1196
1197         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1198         if (!tx)
1199                 return NULL;
1200
1201         neigh->cm = tx;
1202         tx->neigh = neigh;
1203         tx->path = path;
1204         tx->dev = dev;
1205         list_add(&tx->list, &priv->cm.start_list);
1206         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1207         queue_work(ipoib_workqueue, &priv->cm.start_task);
1208         return tx;
1209 }
1210
1211 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1212 {
1213         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1214         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1215                 list_move(&tx->list, &priv->cm.reap_list);
1216                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1217                 ipoib_dbg(priv, "Reap connection for gid " IPOIB_GID_FMT "\n",
1218                           IPOIB_GID_ARG(tx->neigh->dgid));
1219                 tx->neigh = NULL;
1220         }
1221 }
1222
1223 static void ipoib_cm_tx_start(struct work_struct *work)
1224 {
1225         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1226                                                    cm.start_task);
1227         struct net_device *dev = priv->dev;
1228         struct ipoib_neigh *neigh;
1229         struct ipoib_cm_tx *p;
1230         unsigned long flags;
1231         int ret;
1232
1233         struct ib_sa_path_rec pathrec;
1234         u32 qpn;
1235
1236         spin_lock_irqsave(&priv->tx_lock, flags);
1237         spin_lock(&priv->lock);
1238         while (!list_empty(&priv->cm.start_list)) {
1239                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1240                 list_del_init(&p->list);
1241                 neigh = p->neigh;
1242                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1243                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1244                 spin_unlock(&priv->lock);
1245                 spin_unlock_irqrestore(&priv->tx_lock, flags);
1246                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1247                 spin_lock_irqsave(&priv->tx_lock, flags);
1248                 spin_lock(&priv->lock);
1249                 if (ret) {
1250                         neigh = p->neigh;
1251                         if (neigh) {
1252                                 neigh->cm = NULL;
1253                                 list_del(&neigh->list);
1254                                 if (neigh->ah)
1255                                         ipoib_put_ah(neigh->ah);
1256                                 ipoib_neigh_free(dev, neigh);
1257                         }
1258                         list_del(&p->list);
1259                         kfree(p);
1260                 }
1261         }
1262         spin_unlock(&priv->lock);
1263         spin_unlock_irqrestore(&priv->tx_lock, flags);
1264 }
1265
1266 static void ipoib_cm_tx_reap(struct work_struct *work)
1267 {
1268         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1269                                                    cm.reap_task);
1270         struct ipoib_cm_tx *p;
1271
1272         spin_lock_irq(&priv->tx_lock);
1273         spin_lock(&priv->lock);
1274         while (!list_empty(&priv->cm.reap_list)) {
1275                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1276                 list_del(&p->list);
1277                 spin_unlock(&priv->lock);
1278                 spin_unlock_irq(&priv->tx_lock);
1279                 ipoib_cm_tx_destroy(p);
1280                 spin_lock_irq(&priv->tx_lock);
1281                 spin_lock(&priv->lock);
1282         }
1283         spin_unlock(&priv->lock);
1284         spin_unlock_irq(&priv->tx_lock);
1285 }
1286
1287 static void ipoib_cm_skb_reap(struct work_struct *work)
1288 {
1289         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1290                                                    cm.skb_task);
1291         struct sk_buff *skb;
1292
1293         unsigned mtu = priv->mcast_mtu;
1294
1295         spin_lock_irq(&priv->tx_lock);
1296         spin_lock(&priv->lock);
1297         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1298                 spin_unlock(&priv->lock);
1299                 spin_unlock_irq(&priv->tx_lock);
1300                 if (skb->protocol == htons(ETH_P_IP))
1301                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1302 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1303                 else if (skb->protocol == htons(ETH_P_IPV6))
1304                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu, priv->dev);
1305 #endif
1306                 dev_kfree_skb_any(skb);
1307                 spin_lock_irq(&priv->tx_lock);
1308                 spin_lock(&priv->lock);
1309         }
1310         spin_unlock(&priv->lock);
1311         spin_unlock_irq(&priv->tx_lock);
1312 }
1313
1314 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1315                            unsigned int mtu)
1316 {
1317         struct ipoib_dev_priv *priv = netdev_priv(dev);
1318         int e = skb_queue_empty(&priv->cm.skb_queue);
1319
1320         if (skb->dst)
1321                 skb->dst->ops->update_pmtu(skb->dst, mtu);
1322
1323         skb_queue_tail(&priv->cm.skb_queue, skb);
1324         if (e)
1325                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1326 }
1327
1328 static void ipoib_cm_rx_reap(struct work_struct *work)
1329 {
1330         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1331                                                 cm.rx_reap_task)->dev);
1332 }
1333
1334 static void ipoib_cm_stale_task(struct work_struct *work)
1335 {
1336         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1337                                                    cm.stale_task.work);
1338         struct ipoib_cm_rx *p;
1339         int ret;
1340
1341         spin_lock_irq(&priv->lock);
1342         while (!list_empty(&priv->cm.passive_ids)) {
1343                 /* List is sorted by LRU, start from tail,
1344                  * stop when we see a recently used entry */
1345                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1346                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1347                         break;
1348                 list_move(&p->list, &priv->cm.rx_error_list);
1349                 p->state = IPOIB_CM_RX_ERROR;
1350                 spin_unlock_irq(&priv->lock);
1351                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1352                 if (ret)
1353                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1354                 spin_lock_irq(&priv->lock);
1355         }
1356
1357         if (!list_empty(&priv->cm.passive_ids))
1358                 queue_delayed_work(ipoib_workqueue,
1359                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1360         spin_unlock_irq(&priv->lock);
1361 }
1362
1363
1364 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1365                          char *buf)
1366 {
1367         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1368
1369         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1370                 return sprintf(buf, "connected\n");
1371         else
1372                 return sprintf(buf, "datagram\n");
1373 }
1374
1375 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1376                         const char *buf, size_t count)
1377 {
1378         struct net_device *dev = to_net_dev(d);
1379         struct ipoib_dev_priv *priv = netdev_priv(dev);
1380
1381         /* flush paths if we switch modes so that connections are restarted */
1382         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1383                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1384                 ipoib_warn(priv, "enabling connected mode "
1385                            "will cause multicast packet drops\n");
1386
1387                 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1388                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1389
1390                 ipoib_flush_paths(dev);
1391                 return count;
1392         }
1393
1394         if (!strcmp(buf, "datagram\n")) {
1395                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1396                 dev->mtu = min(priv->mcast_mtu, dev->mtu);
1397                 ipoib_flush_paths(dev);
1398
1399                 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1400                         dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1401                         if (priv->hca_caps & IB_DEVICE_UD_TSO)
1402                                 dev->features |= NETIF_F_TSO;
1403                 }
1404
1405                 return count;
1406         }
1407
1408         return -EINVAL;
1409 }
1410
1411 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1412
1413 int ipoib_cm_add_mode_attr(struct net_device *dev)
1414 {
1415         return device_create_file(&dev->dev, &dev_attr_mode);
1416 }
1417
1418 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1419 {
1420         struct ipoib_dev_priv *priv = netdev_priv(dev);
1421         struct ib_srq_init_attr srq_init_attr = {
1422                 .attr = {
1423                         .max_wr  = ipoib_recvq_size,
1424                         .max_sge = max_sge
1425                 }
1426         };
1427
1428         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1429         if (IS_ERR(priv->cm.srq)) {
1430                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1431                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1432                                priv->ca->name, PTR_ERR(priv->cm.srq));
1433                 priv->cm.srq = NULL;
1434                 return;
1435         }
1436
1437         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring,
1438                                     GFP_KERNEL);
1439         if (!priv->cm.srq_ring) {
1440                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1441                        priv->ca->name, ipoib_recvq_size);
1442                 ib_destroy_srq(priv->cm.srq);
1443                 priv->cm.srq = NULL;
1444         }
1445 }
1446
1447 int ipoib_cm_dev_init(struct net_device *dev)
1448 {
1449         struct ipoib_dev_priv *priv = netdev_priv(dev);
1450         int i, ret;
1451         struct ib_device_attr attr;
1452
1453         INIT_LIST_HEAD(&priv->cm.passive_ids);
1454         INIT_LIST_HEAD(&priv->cm.reap_list);
1455         INIT_LIST_HEAD(&priv->cm.start_list);
1456         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1457         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1458         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1459         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1460         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1461         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1462         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1463         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1464         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1465
1466         skb_queue_head_init(&priv->cm.skb_queue);
1467
1468         ret = ib_query_device(priv->ca, &attr);
1469         if (ret) {
1470                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1471                 return ret;
1472         }
1473
1474         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1475
1476         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1477         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1478         if (ipoib_cm_has_srq(dev)) {
1479                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1480                 priv->cm.num_frags  = attr.max_srq_sge;
1481                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1482                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1483         } else {
1484                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1485                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1486         }
1487
1488         for (i = 0; i < priv->cm.num_frags; ++i)
1489                 priv->cm.rx_sge[i].lkey = priv->mr->lkey;
1490
1491         priv->cm.rx_sge[0].length = IPOIB_CM_HEAD_SIZE;
1492         for (i = 1; i < priv->cm.num_frags; ++i)
1493                 priv->cm.rx_sge[i].length = PAGE_SIZE;
1494         priv->cm.rx_wr.next = NULL;
1495         priv->cm.rx_wr.sg_list = priv->cm.rx_sge;
1496         priv->cm.rx_wr.num_sge = priv->cm.num_frags;
1497
1498         if (ipoib_cm_has_srq(dev)) {
1499                 for (i = 0; i < ipoib_recvq_size; ++i) {
1500                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1501                                                    priv->cm.num_frags - 1,
1502                                                    priv->cm.srq_ring[i].mapping)) {
1503                                 ipoib_warn(priv, "failed to allocate "
1504                                            "receive buffer %d\n", i);
1505                                 ipoib_cm_dev_cleanup(dev);
1506                                 return -ENOMEM;
1507                         }
1508
1509                         if (ipoib_cm_post_receive_srq(dev, i)) {
1510                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1511                                            "failed for buf %d\n", i);
1512                                 ipoib_cm_dev_cleanup(dev);
1513                                 return -EIO;
1514                         }
1515                 }
1516         }
1517
1518         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1519         return 0;
1520 }
1521
1522 void ipoib_cm_dev_cleanup(struct net_device *dev)
1523 {
1524         struct ipoib_dev_priv *priv = netdev_priv(dev);
1525         int ret;
1526
1527         if (!priv->cm.srq)
1528                 return;
1529
1530         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1531
1532         ret = ib_destroy_srq(priv->cm.srq);
1533         if (ret)
1534                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1535
1536         priv->cm.srq = NULL;
1537         if (!priv->cm.srq_ring)
1538                 return;
1539
1540         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1541         priv->cm.srq_ring = NULL;
1542 }