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
  33 #include <rdma/ib_cm.h>
  34 #include <net/dst.h>
  35 #include <net/icmp.h>
  36 #include <linux/icmpv6.h>
  37 #include <linux/delay.h>
  38 #include <linux/slab.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "ipoib.h"
  42
  43 int ipoib_max_conn_qp = 128;
  44
  45 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  46 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  47                  "Max number of connected-mode QPs per interface "
  48                  "(applied only if shared receive queue is not available)");
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  51 static int data_debug_level;
  52
  53 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  54 MODULE_PARM_DESC(cm_data_debug_level,
  55                  "Enable data path debug tracing for connected mode if > 0");
  56 #endif
  57
  58 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  59
  60 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  61 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  62 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  63 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  64
  65 static struct ib_qp_attr ipoib_cm_err_attr = {
  66         .qp_state = IB_QPS_ERR
  67 };
  68
  69 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  70
  71 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  72         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  73         .opcode = IB_WR_SEND,
  74 };
  75
  76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  77                                struct ib_cm_event *event);
  78
  79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  80                                   u64 mapping[IPOIB_CM_RX_SG])
  81 {
  82         int i;
  83
  84         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  85
  86         for (i = 0; i < frags; ++i)
  87                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  88 }
  89
  90 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_recv_wr *bad_wr;
  94         int i, ret;
  95
  96         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  97
  98         for (i = 0; i < priv->cm.num_frags; ++i)
  99                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 100
 101         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 102         if (unlikely(ret)) {
 103                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 104                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 105                                       priv->cm.srq_ring[id].mapping);
 106                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 107                 priv->cm.srq_ring[id].skb = NULL;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 114                                         struct ipoib_cm_rx *rx,
 115                                         struct ib_recv_wr *wr,
 116                                         struct ib_sge *sge, int id)
 117 {
 118         struct ipoib_dev_priv *priv = netdev_priv(dev);
 119         struct ib_recv_wr *bad_wr;
 120         int i, ret;
 121
 122         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 123
 124         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 125                 sge[i].addr = rx->rx_ring[id].mapping[i];
 126
 127         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 128         if (unlikely(ret)) {
 129                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 130                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 131                                       rx->rx_ring[id].mapping);
 132                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 133                 rx->rx_ring[id].skb = NULL;
 134         }
 135
 136         return ret;
 137 }
 138
 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         vfree(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->recv_cq, /* For drain WR */
 253                 .recv_cq = priv->recv_cq,
 254                 .srq = priv->cm.srq,
 255                 .cap.max_send_wr = 1, /* For drain WR */
 256                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 257                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 258                 .qp_type = IB_QPT_RC,
 259                 .qp_context = p,
 260         };
 261
 262         if (!ipoib_cm_has_srq(dev)) {
 263                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 264                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 265         }
 266
 267         return ib_create_qp(priv->pd, &attr);
 268 }
 269
 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 void ipoib_cm_init_rx_wr(struct net_device *dev,
 326                                 struct ib_recv_wr *wr,
 327                                 struct ib_sge *sge)
 328 {
 329         struct ipoib_dev_priv *priv = netdev_priv(dev);
 330         int i;
 331
 332         for (i = 0; i < priv->cm.num_frags; ++i)
 333                 sge[i].lkey = priv->mr->lkey;
 334
 335         sge[0].length = IPOIB_CM_HEAD_SIZE;
 336         for (i = 1; i < priv->cm.num_frags; ++i)
 337                 sge[i].length = PAGE_SIZE;
 338
 339         wr->next    = NULL;
 340         wr->sg_list = sge;
 341         wr->num_sge = priv->cm.num_frags;
 342 }
 343
 344 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 345                                    struct ipoib_cm_rx *rx)
 346 {
 347         struct ipoib_dev_priv *priv = netdev_priv(dev);
 348         struct {
 349                 struct ib_recv_wr wr;
 350                 struct ib_sge sge[IPOIB_CM_RX_SG];
 351         } *t;
 352         int ret;
 353         int i;
 354
 355         rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 356         if (!rx->rx_ring) {
 357                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 358                        priv->ca->name, ipoib_recvq_size);
 359                 return -ENOMEM;
 360         }
 361
 362         t = kmalloc(sizeof *t, GFP_KERNEL);
 363         if (!t) {
 364                 ret = -ENOMEM;
 365                 goto err_free;
 366         }
 367
 368         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 369
 370         spin_lock_irq(&priv->lock);
 371
 372         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 373                 spin_unlock_irq(&priv->lock);
 374                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 375                 ret = -EINVAL;
 376                 goto err_free;
 377         } else
 378                 ++priv->cm.nonsrq_conn_qp;
 379
 380         spin_unlock_irq(&priv->lock);
 381
 382         for (i = 0; i < ipoib_recvq_size; ++i) {
 383                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 384                                            rx->rx_ring[i].mapping)) {
 385                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 386                                 ret = -ENOMEM;
 387                                 goto err_count;
 388                 }
 389                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 390                 if (ret) {
 391                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 392                                    "failed for buf %d\n", i);
 393                         ret = -EIO;
 394                         goto err_count;
 395                 }
 396         }
 397
 398         rx->recv_count = ipoib_recvq_size;
 399
 400         kfree(t);
 401
 402         return 0;
 403
 404 err_count:
 405         spin_lock_irq(&priv->lock);
 406         --priv->cm.nonsrq_conn_qp;
 407         spin_unlock_irq(&priv->lock);
 408
 409 err_free:
 410         kfree(t);
 411         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 412
 413         return ret;
 414 }
 415
 416 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 417                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 418                              unsigned psn)
 419 {
 420         struct ipoib_dev_priv *priv = netdev_priv(dev);
 421         struct ipoib_cm_data data = {};
 422         struct ib_cm_rep_param rep = {};
 423
 424         data.qpn = cpu_to_be32(priv->qp->qp_num);
 425         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 426
 427         rep.private_data = &data;
 428         rep.private_data_len = sizeof data;
 429         rep.flow_control = 0;
 430         rep.rnr_retry_count = req->rnr_retry_count;
 431         rep.srq = ipoib_cm_has_srq(dev);
 432         rep.qp_num = qp->qp_num;
 433         rep.starting_psn = psn;
 434         return ib_send_cm_rep(cm_id, &rep);
 435 }
 436
 437 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 438 {
 439         struct net_device *dev = cm_id->context;
 440         struct ipoib_dev_priv *priv = netdev_priv(dev);
 441         struct ipoib_cm_rx *p;
 442         unsigned psn;
 443         int ret;
 444
 445         ipoib_dbg(priv, "REQ arrived\n");
 446         p = kzalloc(sizeof *p, GFP_KERNEL);
 447         if (!p)
 448                 return -ENOMEM;
 449         p->dev = dev;
 450         p->id = cm_id;
 451         cm_id->context = p;
 452         p->state = IPOIB_CM_RX_LIVE;
 453         p->jiffies = jiffies;
 454         INIT_LIST_HEAD(&p->list);
 455
 456         p->qp = ipoib_cm_create_rx_qp(dev, p);
 457         if (IS_ERR(p->qp)) {
 458                 ret = PTR_ERR(p->qp);
 459                 goto err_qp;
 460         }
 461
 462         psn = random32() & 0xffffff;
 463         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 464         if (ret)
 465                 goto err_modify;
 466
 467         if (!ipoib_cm_has_srq(dev)) {
 468                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 469                 if (ret)
 470                         goto err_modify;
 471         }
 472
 473         spin_lock_irq(&priv->lock);
 474         queue_delayed_work(ipoib_workqueue,
 475                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 476         /* Add this entry to passive ids list head, but do not re-add it
 477          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 478         p->jiffies = jiffies;
 479         if (p->state == IPOIB_CM_RX_LIVE)
 480                 list_move(&p->list, &priv->cm.passive_ids);
 481         spin_unlock_irq(&priv->lock);
 482
 483         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 484         if (ret) {
 485                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 486                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 487                         ipoib_warn(priv, "unable to move qp to error state\n");
 488         }
 489         return 0;
 490
 491 err_modify:
 492         ib_destroy_qp(p->qp);
 493 err_qp:
 494         kfree(p);
 495         return ret;
 496 }
 497
 498 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 499                                struct ib_cm_event *event)
 500 {
 501         struct ipoib_cm_rx *p;
 502         struct ipoib_dev_priv *priv;
 503
 504         switch (event->event) {
 505         case IB_CM_REQ_RECEIVED:
 506                 return ipoib_cm_req_handler(cm_id, event);
 507         case IB_CM_DREQ_RECEIVED:
 508                 p = cm_id->context;
 509                 ib_send_cm_drep(cm_id, NULL, 0);
 510                 /* Fall through */
 511         case IB_CM_REJ_RECEIVED:
 512                 p = cm_id->context;
 513                 priv = netdev_priv(p->dev);
 514                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 515                         ipoib_warn(priv, "unable to move qp to error state\n");
 516                 /* Fall through */
 517         default:
 518                 return 0;
 519         }
 520 }
 521 /* Adjust length of skb with fragments to match received data */
 522 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 523                           unsigned int length, struct sk_buff *toskb)
 524 {
 525         int i, num_frags;
 526         unsigned int size;
 527
 528         /* put header into skb */
 529         size = min(length, hdr_space);
 530         skb->tail += size;
 531         skb->len += size;
 532         length -= size;
 533
 534         num_frags = skb_shinfo(skb)->nr_frags;
 535         for (i = 0; i < num_frags; i++) {
 536                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 537
 538                 if (length == 0) {
 539                         /* don't need this page */
 540                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 541                         --skb_shinfo(skb)->nr_frags;
 542                 } else {
 543                         size = min(length, (unsigned) PAGE_SIZE);
 544
 545                         frag->size = size;
 546                         skb->data_len += size;
 547                         skb->truesize += size;
 548                         skb->len += size;
 549                         length -= size;
 550                 }
 551         }
 552 }
 553
 554 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 555 {
 556         struct ipoib_dev_priv *priv = netdev_priv(dev);
 557         struct ipoib_cm_rx_buf *rx_ring;
 558         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 559         struct sk_buff *skb, *newskb;
 560         struct ipoib_cm_rx *p;
 561         unsigned long flags;
 562         u64 mapping[IPOIB_CM_RX_SG];
 563         int frags;
 564         int has_srq;
 565         struct sk_buff *small_skb;
 566
 567         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 568                        wr_id, wc->status);
 569
 570         if (unlikely(wr_id >= ipoib_recvq_size)) {
 571                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 572                         spin_lock_irqsave(&priv->lock, flags);
 573                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 574                         ipoib_cm_start_rx_drain(priv);
 575                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 576                         spin_unlock_irqrestore(&priv->lock, flags);
 577                 } else
 578                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 579                                    wr_id, ipoib_recvq_size);
 580                 return;
 581         }
 582
 583         p = wc->qp->qp_context;
 584
 585         has_srq = ipoib_cm_has_srq(dev);
 586         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 587
 588         skb = rx_ring[wr_id].skb;
 589
 590         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 591                 ipoib_dbg(priv, "cm recv error "
 592                            "(status=%d, wrid=%d vend_err %x)\n",
 593                            wc->status, wr_id, wc->vendor_err);
 594                 ++dev->stats.rx_dropped;
 595                 if (has_srq)
 596                         goto repost;
 597                 else {
 598                         if (!--p->recv_count) {
 599                                 spin_lock_irqsave(&priv->lock, flags);
 600                                 list_move(&p->list, &priv->cm.rx_reap_list);
 601                                 spin_unlock_irqrestore(&priv->lock, flags);
 602                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 603                         }
 604                         return;
 605                 }
 606         }
 607
 608         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 609                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 610                         spin_lock_irqsave(&priv->lock, flags);
 611                         p->jiffies = jiffies;
 612                         /* Move this entry to list head, but do not re-add it
 613                          * if it has been moved out of list. */
 614                         if (p->state == IPOIB_CM_RX_LIVE)
 615                                 list_move(&p->list, &priv->cm.passive_ids);
 616                         spin_unlock_irqrestore(&priv->lock, flags);
 617                 }
 618         }
 619
 620         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 621                 int dlen = wc->byte_len;
 622
 623                 small_skb = dev_alloc_skb(dlen + 12);
 624                 if (small_skb) {
 625                         skb_reserve(small_skb, 12);
 626                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 627                                                    dlen, DMA_FROM_DEVICE);
 628                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 629                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 630                                                       dlen, DMA_FROM_DEVICE);
 631                         skb_put(small_skb, dlen);
 632                         skb = small_skb;
 633                         goto copied;
 634                 }
 635         }
 636
 637         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 638                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 639
 640         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 641         if (unlikely(!newskb)) {
 642                 /*
 643                  * If we can't allocate a new RX buffer, dump
 644                  * this packet and reuse the old buffer.
 645                  */
 646                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 647                 ++dev->stats.rx_dropped;
 648                 goto repost;
 649         }
 650
 651         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 652         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 653
 654         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 655                        wc->byte_len, wc->slid);
 656
 657         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 658
 659 copied:
 660         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 661         skb_reset_mac_header(skb);
 662         skb_pull(skb, IPOIB_ENCAP_LEN);
 663
 664         ++dev->stats.rx_packets;
 665         dev->stats.rx_bytes += skb->len;
 666
 667         skb->dev = dev;
 668         /* XXX get correct PACKET_ type here */
 669         skb->pkt_type = PACKET_HOST;
 670         netif_receive_skb(skb);
 671
 672 repost:
 673         if (has_srq) {
 674                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 675                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 676                                    "for buf %d\n", wr_id);
 677         } else {
 678                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 679                                                           &priv->cm.rx_wr,
 680                                                           priv->cm.rx_sge,
 681                                                           wr_id))) {
 682                         --p->recv_count;
 683                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 684                                    "for buf %d\n", wr_id);
 685                 }
 686         }
 687 }
 688
 689 static inline int post_send(struct ipoib_dev_priv *priv,
 690                             struct ipoib_cm_tx *tx,
 691                             unsigned int wr_id,
 692                             u64 addr, int len)
 693 {
 694         struct ib_send_wr *bad_wr;
 695
 696         priv->tx_sge[0].addr          = addr;
 697         priv->tx_sge[0].length        = len;
 698
 699         priv->tx_wr.num_sge     = 1;
 700         priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 701
 702         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 703 }
 704
 705 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 706 {
 707         struct ipoib_dev_priv *priv = netdev_priv(dev);
 708         struct ipoib_cm_tx_buf *tx_req;
 709         u64 addr;
 710         int rc;
 711
 712         if (unlikely(skb->len > tx->mtu)) {
 713                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 714                            skb->len, tx->mtu);
 715                 ++dev->stats.tx_dropped;
 716                 ++dev->stats.tx_errors;
 717                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 718                 return;
 719         }
 720
 721         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 722                        tx->tx_head, skb->len, tx->qp->qp_num);
 723
 724         /*
 725          * We put the skb into the tx_ring _before_ we call post_send()
 726          * because it's entirely possible that the completion handler will
 727          * run before we execute anything after the post_send().  That
 728          * means we have to make sure everything is properly recorded and
 729          * our state is consistent before we call post_send().
 730          */
 731         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 732         tx_req->skb = skb;
 733         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 734         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 735                 ++dev->stats.tx_errors;
 736                 dev_kfree_skb_any(skb);
 737                 return;
 738         }
 739
 740         tx_req->mapping = addr;
 741
 742         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 743                        addr, skb->len);
 744         if (unlikely(rc)) {
 745                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 746                 ++dev->stats.tx_errors;
 747                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 748                 dev_kfree_skb_any(skb);
 749         } else {
 750                 dev->trans_start = jiffies;
 751                 ++tx->tx_head;
 752
 753                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 754                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 755                                   tx->qp->qp_num);
 756                         if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 757                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 758                         netif_stop_queue(dev);
 759                 }
 760         }
 761 }
 762
 763 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 764 {
 765         struct ipoib_dev_priv *priv = netdev_priv(dev);
 766         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 767         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 768         struct ipoib_cm_tx_buf *tx_req;
 769         unsigned long flags;
 770
 771         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 772                        wr_id, wc->status);
 773
 774         if (unlikely(wr_id >= ipoib_sendq_size)) {
 775                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 776                            wr_id, ipoib_sendq_size);
 777                 return;
 778         }
 779
 780         tx_req = &tx->tx_ring[wr_id];
 781
 782         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 783
 784         /* FIXME: is this right? Shouldn't we only increment on success? */
 785         ++dev->stats.tx_packets;
 786         dev->stats.tx_bytes += tx_req->skb->len;
 787
 788         dev_kfree_skb_any(tx_req->skb);
 789
 790         netif_tx_lock(dev);
 791
 792         ++tx->tx_tail;
 793         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 794             netif_queue_stopped(dev) &&
 795             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 796                 netif_wake_queue(dev);
 797
 798         if (wc->status != IB_WC_SUCCESS &&
 799             wc->status != IB_WC_WR_FLUSH_ERR) {
 800                 struct ipoib_neigh *neigh;
 801
 802                 ipoib_dbg(priv, "failed cm send event "
 803                            "(status=%d, wrid=%d vend_err %x)\n",
 804                            wc->status, wr_id, wc->vendor_err);
 805
 806                 spin_lock_irqsave(&priv->lock, flags);
 807                 neigh = tx->neigh;
 808
 809                 if (neigh) {
 810                         neigh->cm = NULL;
 811                         list_del(&neigh->list);
 812                         if (neigh->ah)
 813                                 ipoib_put_ah(neigh->ah);
 814                         ipoib_neigh_free(dev, neigh);
 815
 816                         tx->neigh = NULL;
 817                 }
 818
 819                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 820                         list_move(&tx->list, &priv->cm.reap_list);
 821                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 822                 }
 823
 824                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 825
 826                 spin_unlock_irqrestore(&priv->lock, flags);
 827         }
 828
 829         netif_tx_unlock(dev);
 830 }
 831
 832 int ipoib_cm_dev_open(struct net_device *dev)
 833 {
 834         struct ipoib_dev_priv *priv = netdev_priv(dev);
 835         int ret;
 836
 837         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 838                 return 0;
 839
 840         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 841         if (IS_ERR(priv->cm.id)) {
 842                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 843                 ret = PTR_ERR(priv->cm.id);
 844                 goto err_cm;
 845         }
 846
 847         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 848                            0, NULL);
 849         if (ret) {
 850                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 851                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 852                 goto err_listen;
 853         }
 854
 855         return 0;
 856
 857 err_listen:
 858         ib_destroy_cm_id(priv->cm.id);
 859 err_cm:
 860         priv->cm.id = NULL;
 861         return ret;
 862 }
 863
 864 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 865 {
 866         struct ipoib_dev_priv *priv = netdev_priv(dev);
 867         struct ipoib_cm_rx *rx, *n;
 868         LIST_HEAD(list);
 869
 870         spin_lock_irq(&priv->lock);
 871         list_splice_init(&priv->cm.rx_reap_list, &list);
 872         spin_unlock_irq(&priv->lock);
 873
 874         list_for_each_entry_safe(rx, n, &list, list) {
 875                 ib_destroy_cm_id(rx->id);
 876                 ib_destroy_qp(rx->qp);
 877                 if (!ipoib_cm_has_srq(dev)) {
 878                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 879                         spin_lock_irq(&priv->lock);
 880                         --priv->cm.nonsrq_conn_qp;
 881                         spin_unlock_irq(&priv->lock);
 882                 }
 883                 kfree(rx);
 884         }
 885 }
 886
 887 void ipoib_cm_dev_stop(struct net_device *dev)
 888 {
 889         struct ipoib_dev_priv *priv = netdev_priv(dev);
 890         struct ipoib_cm_rx *p;
 891         unsigned long begin;
 892         int ret;
 893
 894         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 895                 return;
 896
 897         ib_destroy_cm_id(priv->cm.id);
 898         priv->cm.id = NULL;
 899
 900         spin_lock_irq(&priv->lock);
 901         while (!list_empty(&priv->cm.passive_ids)) {
 902                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 903                 list_move(&p->list, &priv->cm.rx_error_list);
 904                 p->state = IPOIB_CM_RX_ERROR;
 905                 spin_unlock_irq(&priv->lock);
 906                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 907                 if (ret)
 908                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 909                 spin_lock_irq(&priv->lock);
 910         }
 911
 912         /* Wait for all RX to be drained */
 913         begin = jiffies;
 914
 915         while (!list_empty(&priv->cm.rx_error_list) ||
 916                !list_empty(&priv->cm.rx_flush_list) ||
 917                !list_empty(&priv->cm.rx_drain_list)) {
 918                 if (time_after(jiffies, begin + 5 * HZ)) {
 919                         ipoib_warn(priv, "RX drain timing out\n");
 920
 921                         /*
 922                          * assume the HW is wedged and just free up everything.
 923                          */
 924                         list_splice_init(&priv->cm.rx_flush_list,
 925                                          &priv->cm.rx_reap_list);
 926                         list_splice_init(&priv->cm.rx_error_list,
 927                                          &priv->cm.rx_reap_list);
 928                         list_splice_init(&priv->cm.rx_drain_list,
 929                                          &priv->cm.rx_reap_list);
 930                         break;
 931                 }
 932                 spin_unlock_irq(&priv->lock);
 933                 msleep(1);
 934                 ipoib_drain_cq(dev);
 935                 spin_lock_irq(&priv->lock);
 936         }
 937
 938         spin_unlock_irq(&priv->lock);
 939
 940         ipoib_cm_free_rx_reap_list(dev);
 941
 942         cancel_delayed_work(&priv->cm.stale_task);
 943 }
 944
 945 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 946 {
 947         struct ipoib_cm_tx *p = cm_id->context;
 948         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 949         struct ipoib_cm_data *data = event->private_data;
 950         struct sk_buff_head skqueue;
 951         struct ib_qp_attr qp_attr;
 952         int qp_attr_mask, ret;
 953         struct sk_buff *skb;
 954
 955         p->mtu = be32_to_cpu(data->mtu);
 956
 957         if (p->mtu <= IPOIB_ENCAP_LEN) {
 958                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 959                            p->mtu, IPOIB_ENCAP_LEN);
 960                 return -EINVAL;
 961         }
 962
 963         qp_attr.qp_state = IB_QPS_RTR;
 964         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 965         if (ret) {
 966                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 967                 return ret;
 968         }
 969
 970         qp_attr.rq_psn = 0 /* FIXME */;
 971         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 972         if (ret) {
 973                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 974                 return ret;
 975         }
 976
 977         qp_attr.qp_state = IB_QPS_RTS;
 978         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 979         if (ret) {
 980                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 981                 return ret;
 982         }
 983         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 984         if (ret) {
 985                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 986                 return ret;
 987         }
 988
 989         skb_queue_head_init(&skqueue);
 990
 991         spin_lock_irq(&priv->lock);
 992         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 993         if (p->neigh)
 994                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 995                         __skb_queue_tail(&skqueue, skb);
 996         spin_unlock_irq(&priv->lock);
 997
 998         while ((skb = __skb_dequeue(&skqueue))) {
 999                 skb->dev = p->dev;
1000                 if (dev_queue_xmit(skb))
1001                         ipoib_warn(priv, "dev_queue_xmit failed "
1002                                    "to requeue packet\n");
1003         }
1004
1005         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1006         if (ret) {
1007                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1008                 return ret;
1009         }
1010         return 0;
1011 }
1012
1013 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1014 {
1015         struct ipoib_dev_priv *priv = netdev_priv(dev);
1016         struct ib_qp_init_attr attr = {
1017                 .send_cq                = priv->recv_cq,
1018                 .recv_cq                = priv->recv_cq,
1019                 .srq                    = priv->cm.srq,
1020                 .cap.max_send_wr        = ipoib_sendq_size,
1021                 .cap.max_send_sge       = 1,
1022                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1023                 .qp_type                = IB_QPT_RC,
1024                 .qp_context             = tx
1025         };
1026
1027         return ib_create_qp(priv->pd, &attr);
1028 }
1029
1030 static int ipoib_cm_send_req(struct net_device *dev,
1031                              struct ib_cm_id *id, struct ib_qp *qp,
1032                              u32 qpn,
1033                              struct ib_sa_path_rec *pathrec)
1034 {
1035         struct ipoib_dev_priv *priv = netdev_priv(dev);
1036         struct ipoib_cm_data data = {};
1037         struct ib_cm_req_param req = {};
1038
1039         data.qpn = cpu_to_be32(priv->qp->qp_num);
1040         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1041
1042         req.primary_path                = pathrec;
1043         req.alternate_path              = NULL;
1044         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1045         req.qp_num                      = qp->qp_num;
1046         req.qp_type                     = qp->qp_type;
1047         req.private_data                = &data;
1048         req.private_data_len            = sizeof data;
1049         req.flow_control                = 0;
1050
1051         req.starting_psn                = 0; /* FIXME */
1052
1053         /*
1054          * Pick some arbitrary defaults here; we could make these
1055          * module parameters if anyone cared about setting them.
1056          */
1057         req.responder_resources         = 4;
1058         req.remote_cm_response_timeout  = 20;
1059         req.local_cm_response_timeout   = 20;
1060         req.retry_count                 = 0; /* RFC draft warns against retries */
1061         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1062         req.max_cm_retries              = 15;
1063         req.srq                         = ipoib_cm_has_srq(dev);
1064         return ib_send_cm_req(id, &req);
1065 }
1066
1067 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1068                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1069 {
1070         struct ipoib_dev_priv *priv = netdev_priv(dev);
1071         struct ib_qp_attr qp_attr;
1072         int qp_attr_mask, ret;
1073         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1074         if (ret) {
1075                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1076                 return ret;
1077         }
1078
1079         qp_attr.qp_state = IB_QPS_INIT;
1080         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1081         qp_attr.port_num = priv->port;
1082         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1083
1084         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1085         if (ret) {
1086                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1087                 return ret;
1088         }
1089         return 0;
1090 }
1091
1092 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1093                             struct ib_sa_path_rec *pathrec)
1094 {
1095         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1096         int ret;
1097
1098         p->tx_ring = vzalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1099         if (!p->tx_ring) {
1100                 ipoib_warn(priv, "failed to allocate tx ring\n");
1101                 ret = -ENOMEM;
1102                 goto err_tx;
1103         }
1104
1105         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1106         if (IS_ERR(p->qp)) {
1107                 ret = PTR_ERR(p->qp);
1108                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1109                 goto err_qp;
1110         }
1111
1112         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1113         if (IS_ERR(p->id)) {
1114                 ret = PTR_ERR(p->id);
1115                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1116                 goto err_id;
1117         }
1118
1119         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1120         if (ret) {
1121                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1122                 goto err_modify;
1123         }
1124
1125         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1126         if (ret) {
1127                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1128                 goto err_send_cm;
1129         }
1130
1131         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1132                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1133
1134         return 0;
1135
1136 err_send_cm:
1137 err_modify:
1138         ib_destroy_cm_id(p->id);
1139 err_id:
1140         p->id = NULL;
1141         ib_destroy_qp(p->qp);
1142 err_qp:
1143         p->qp = NULL;
1144         vfree(p->tx_ring);
1145 err_tx:
1146         return ret;
1147 }
1148
1149 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1150 {
1151         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1152         struct ipoib_cm_tx_buf *tx_req;
1153         unsigned long begin;
1154
1155         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1156                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1157
1158         if (p->id)
1159                 ib_destroy_cm_id(p->id);
1160
1161         if (p->tx_ring) {
1162                 /* Wait for all sends to complete */
1163                 begin = jiffies;
1164                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1165                         if (time_after(jiffies, begin + 5 * HZ)) {
1166                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1167                                            p->tx_head - p->tx_tail);
1168                                 goto timeout;
1169                         }
1170
1171                         msleep(1);
1172                 }
1173         }
1174
1175 timeout:
1176
1177         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1178                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1179                 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1180                                     DMA_TO_DEVICE);
1181                 dev_kfree_skb_any(tx_req->skb);
1182                 ++p->tx_tail;
1183                 netif_tx_lock_bh(p->dev);
1184                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1185                     netif_queue_stopped(p->dev) &&
1186                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1187                         netif_wake_queue(p->dev);
1188                 netif_tx_unlock_bh(p->dev);
1189         }
1190
1191         if (p->qp)
1192                 ib_destroy_qp(p->qp);
1193
1194         vfree(p->tx_ring);
1195         kfree(p);
1196 }
1197
1198 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1199                                struct ib_cm_event *event)
1200 {
1201         struct ipoib_cm_tx *tx = cm_id->context;
1202         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1203         struct net_device *dev = priv->dev;
1204         struct ipoib_neigh *neigh;
1205         unsigned long flags;
1206         int ret;
1207
1208         switch (event->event) {
1209         case IB_CM_DREQ_RECEIVED:
1210                 ipoib_dbg(priv, "DREQ received.\n");
1211                 ib_send_cm_drep(cm_id, NULL, 0);
1212                 break;
1213         case IB_CM_REP_RECEIVED:
1214                 ipoib_dbg(priv, "REP received.\n");
1215                 ret = ipoib_cm_rep_handler(cm_id, event);
1216                 if (ret)
1217                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1218                                        NULL, 0, NULL, 0);
1219                 break;
1220         case IB_CM_REQ_ERROR:
1221         case IB_CM_REJ_RECEIVED:
1222         case IB_CM_TIMEWAIT_EXIT:
1223                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1224                 netif_tx_lock_bh(dev);
1225                 spin_lock_irqsave(&priv->lock, flags);
1226                 neigh = tx->neigh;
1227
1228                 if (neigh) {
1229                         neigh->cm = NULL;
1230                         list_del(&neigh->list);
1231                         if (neigh->ah)
1232                                 ipoib_put_ah(neigh->ah);
1233                         ipoib_neigh_free(dev, neigh);
1234
1235                         tx->neigh = NULL;
1236                 }
1237
1238                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1239                         list_move(&tx->list, &priv->cm.reap_list);
1240                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1241                 }
1242
1243                 spin_unlock_irqrestore(&priv->lock, flags);
1244                 netif_tx_unlock_bh(dev);
1245                 break;
1246         default:
1247                 break;
1248         }
1249
1250         return 0;
1251 }
1252
1253 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1254                                        struct ipoib_neigh *neigh)
1255 {
1256         struct ipoib_dev_priv *priv = netdev_priv(dev);
1257         struct ipoib_cm_tx *tx;
1258
1259         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1260         if (!tx)
1261                 return NULL;
1262
1263         neigh->cm = tx;
1264         tx->neigh = neigh;
1265         tx->path = path;
1266         tx->dev = dev;
1267         list_add(&tx->list, &priv->cm.start_list);
1268         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1269         queue_work(ipoib_workqueue, &priv->cm.start_task);
1270         return tx;
1271 }
1272
1273 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1274 {
1275         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1276         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1277                 list_move(&tx->list, &priv->cm.reap_list);
1278                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1279                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1280                           tx->neigh->dgid.raw);
1281                 tx->neigh = NULL;
1282         }
1283 }
1284
1285 static void ipoib_cm_tx_start(struct work_struct *work)
1286 {
1287         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1288                                                    cm.start_task);
1289         struct net_device *dev = priv->dev;
1290         struct ipoib_neigh *neigh;
1291         struct ipoib_cm_tx *p;
1292         unsigned long flags;
1293         int ret;
1294
1295         struct ib_sa_path_rec pathrec;
1296         u32 qpn;
1297
1298         netif_tx_lock_bh(dev);
1299         spin_lock_irqsave(&priv->lock, flags);
1300
1301         while (!list_empty(&priv->cm.start_list)) {
1302                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1303                 list_del_init(&p->list);
1304                 neigh = p->neigh;
1305                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1306                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1307
1308                 spin_unlock_irqrestore(&priv->lock, flags);
1309                 netif_tx_unlock_bh(dev);
1310
1311                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1312
1313                 netif_tx_lock_bh(dev);
1314                 spin_lock_irqsave(&priv->lock, flags);
1315
1316                 if (ret) {
1317                         neigh = p->neigh;
1318                         if (neigh) {
1319                                 neigh->cm = NULL;
1320                                 list_del(&neigh->list);
1321                                 if (neigh->ah)
1322                                         ipoib_put_ah(neigh->ah);
1323                                 ipoib_neigh_free(dev, neigh);
1324                         }
1325                         list_del(&p->list);
1326                         kfree(p);
1327                 }
1328         }
1329
1330         spin_unlock_irqrestore(&priv->lock, flags);
1331         netif_tx_unlock_bh(dev);
1332 }
1333
1334 static void ipoib_cm_tx_reap(struct work_struct *work)
1335 {
1336         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1337                                                    cm.reap_task);
1338         struct net_device *dev = priv->dev;
1339         struct ipoib_cm_tx *p;
1340         unsigned long flags;
1341
1342         netif_tx_lock_bh(dev);
1343         spin_lock_irqsave(&priv->lock, flags);
1344
1345         while (!list_empty(&priv->cm.reap_list)) {
1346                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1347                 list_del(&p->list);
1348                 spin_unlock_irqrestore(&priv->lock, flags);
1349                 netif_tx_unlock_bh(dev);
1350                 ipoib_cm_tx_destroy(p);
1351                 netif_tx_lock_bh(dev);
1352                 spin_lock_irqsave(&priv->lock, flags);
1353         }
1354
1355         spin_unlock_irqrestore(&priv->lock, flags);
1356         netif_tx_unlock_bh(dev);
1357 }
1358
1359 static void ipoib_cm_skb_reap(struct work_struct *work)
1360 {
1361         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1362                                                    cm.skb_task);
1363         struct net_device *dev = priv->dev;
1364         struct sk_buff *skb;
1365         unsigned long flags;
1366         unsigned mtu = priv->mcast_mtu;
1367
1368         netif_tx_lock_bh(dev);
1369         spin_lock_irqsave(&priv->lock, flags);
1370
1371         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1372                 spin_unlock_irqrestore(&priv->lock, flags);
1373                 netif_tx_unlock_bh(dev);
1374
1375                 if (skb->protocol == htons(ETH_P_IP))
1376                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1377 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1378                 else if (skb->protocol == htons(ETH_P_IPV6))
1379                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1380 #endif
1381                 dev_kfree_skb_any(skb);
1382
1383                 netif_tx_lock_bh(dev);
1384                 spin_lock_irqsave(&priv->lock, flags);
1385         }
1386
1387         spin_unlock_irqrestore(&priv->lock, flags);
1388         netif_tx_unlock_bh(dev);
1389 }
1390
1391 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1392                            unsigned int mtu)
1393 {
1394         struct ipoib_dev_priv *priv = netdev_priv(dev);
1395         int e = skb_queue_empty(&priv->cm.skb_queue);
1396
1397         if (skb_dst(skb))
1398                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1399
1400         skb_queue_tail(&priv->cm.skb_queue, skb);
1401         if (e)
1402                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1403 }
1404
1405 static void ipoib_cm_rx_reap(struct work_struct *work)
1406 {
1407         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1408                                                 cm.rx_reap_task)->dev);
1409 }
1410
1411 static void ipoib_cm_stale_task(struct work_struct *work)
1412 {
1413         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1414                                                    cm.stale_task.work);
1415         struct ipoib_cm_rx *p;
1416         int ret;
1417
1418         spin_lock_irq(&priv->lock);
1419         while (!list_empty(&priv->cm.passive_ids)) {
1420                 /* List is sorted by LRU, start from tail,
1421                  * stop when we see a recently used entry */
1422                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1423                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1424                         break;
1425                 list_move(&p->list, &priv->cm.rx_error_list);
1426                 p->state = IPOIB_CM_RX_ERROR;
1427                 spin_unlock_irq(&priv->lock);
1428                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1429                 if (ret)
1430                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1431                 spin_lock_irq(&priv->lock);
1432         }
1433
1434         if (!list_empty(&priv->cm.passive_ids))
1435                 queue_delayed_work(ipoib_workqueue,
1436                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1437         spin_unlock_irq(&priv->lock);
1438 }
1439
1440
1441 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1442                          char *buf)
1443 {
1444         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1445
1446         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1447                 return sprintf(buf, "connected\n");
1448         else
1449                 return sprintf(buf, "datagram\n");
1450 }
1451
1452 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1453                         const char *buf, size_t count)
1454 {
1455         struct net_device *dev = to_net_dev(d);
1456         struct ipoib_dev_priv *priv = netdev_priv(dev);
1457
1458         if (!rtnl_trylock())
1459                 return restart_syscall();
1460
1461         /* flush paths if we switch modes so that connections are restarted */
1462         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1463                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1464                 ipoib_warn(priv, "enabling connected mode "
1465                            "will cause multicast packet drops\n");
1466                 netdev_update_features(dev);
1467                 rtnl_unlock();
1468                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1469
1470                 ipoib_flush_paths(dev);
1471                 return count;
1472         }
1473
1474         if (!strcmp(buf, "datagram\n")) {
1475                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1476                 netdev_update_features(dev);
1477                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1478                 rtnl_unlock();
1479                 ipoib_flush_paths(dev);
1480
1481                 return count;
1482         }
1483         rtnl_unlock();
1484
1485         return -EINVAL;
1486 }
1487
1488 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1489
1490 int ipoib_cm_add_mode_attr(struct net_device *dev)
1491 {
1492         return device_create_file(&dev->dev, &dev_attr_mode);
1493 }
1494
1495 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1496 {
1497         struct ipoib_dev_priv *priv = netdev_priv(dev);
1498         struct ib_srq_init_attr srq_init_attr = {
1499                 .attr = {
1500                         .max_wr  = ipoib_recvq_size,
1501                         .max_sge = max_sge
1502                 }
1503         };
1504
1505         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1506         if (IS_ERR(priv->cm.srq)) {
1507                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1508                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1509                                priv->ca->name, PTR_ERR(priv->cm.srq));
1510                 priv->cm.srq = NULL;
1511                 return;
1512         }
1513
1514         priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1515         if (!priv->cm.srq_ring) {
1516                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1517                        priv->ca->name, ipoib_recvq_size);
1518                 ib_destroy_srq(priv->cm.srq);
1519                 priv->cm.srq = NULL;
1520                 return;
1521         }
1522
1523 }
1524
1525 int ipoib_cm_dev_init(struct net_device *dev)
1526 {
1527         struct ipoib_dev_priv *priv = netdev_priv(dev);
1528         int i, ret;
1529         struct ib_device_attr attr;
1530
1531         INIT_LIST_HEAD(&priv->cm.passive_ids);
1532         INIT_LIST_HEAD(&priv->cm.reap_list);
1533         INIT_LIST_HEAD(&priv->cm.start_list);
1534         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1535         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1536         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1537         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1538         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1539         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1540         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1541         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1542         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1543
1544         skb_queue_head_init(&priv->cm.skb_queue);
1545
1546         ret = ib_query_device(priv->ca, &attr);
1547         if (ret) {
1548                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1549                 return ret;
1550         }
1551
1552         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1553
1554         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1555         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1556         if (ipoib_cm_has_srq(dev)) {
1557                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1558                 priv->cm.num_frags  = attr.max_srq_sge;
1559                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1560                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1561         } else {
1562                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1563                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1564         }
1565
1566         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1567
1568         if (ipoib_cm_has_srq(dev)) {
1569                 for (i = 0; i < ipoib_recvq_size; ++i) {
1570                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1571                                                    priv->cm.num_frags - 1,
1572                                                    priv->cm.srq_ring[i].mapping)) {
1573                                 ipoib_warn(priv, "failed to allocate "
1574                                            "receive buffer %d\n", i);
1575                                 ipoib_cm_dev_cleanup(dev);
1576                                 return -ENOMEM;
1577                         }
1578
1579                         if (ipoib_cm_post_receive_srq(dev, i)) {
1580                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1581                                            "failed for buf %d\n", i);
1582                                 ipoib_cm_dev_cleanup(dev);
1583                                 return -EIO;
1584                         }
1585                 }
1586         }
1587
1588         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1589         return 0;
1590 }
1591
1592 void ipoib_cm_dev_cleanup(struct net_device *dev)
1593 {
1594         struct ipoib_dev_priv *priv = netdev_priv(dev);
1595         int ret;
1596
1597         if (!priv->cm.srq)
1598                 return;
1599
1600         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1601
1602         ret = ib_destroy_srq(priv->cm.srq);
1603         if (ret)
1604                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1605
1606         priv->cm.srq = NULL;
1607         if (!priv->cm.srq_ring)
1608                 return;
1609
1610         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1611         priv->cm.srq_ring = NULL;
1612 }