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/vmalloc.h>
  39
  40 #include "ipoib.h"
  41
  42 int ipoib_max_conn_qp = 128;
  43
  44 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  45 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  46                  "Max number of connected-mode QPs per interface "
  47                  "(applied only if shared receive queue is not available)");
  48
  49 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  50 static int data_debug_level;
  51
  52 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  53 MODULE_PARM_DESC(cm_data_debug_level,
  54                  "Enable data path debug tracing for connected mode if > 0");
  55 #endif
  56
  57 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  58
  59 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  60 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  61 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  62 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  63
  64 static struct ib_qp_attr ipoib_cm_err_attr = {
  65         .qp_state = IB_QPS_ERR
  66 };
  67
  68 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  69
  70 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  71         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  72         .opcode = IB_WR_SEND,
  73 };
  74
  75 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  76                                struct ib_cm_event *event);
  77
  78 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  79                                   u64 mapping[IPOIB_CM_RX_SG])
  80 {
  81         int i;
  82
  83         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  84
  85         for (i = 0; i < frags; ++i)
  86                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  87 }
  88
  89 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  90 {
  91         struct ipoib_dev_priv *priv = netdev_priv(dev);
  92         struct ib_recv_wr *bad_wr;
  93         int i, ret;
  94
  95         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  96
  97         for (i = 0; i < priv->cm.num_frags; ++i)
  98                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
  99
 100         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 101         if (unlikely(ret)) {
 102                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 103                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 104                                       priv->cm.srq_ring[id].mapping);
 105                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 106                 priv->cm.srq_ring[id].skb = NULL;
 107         }
 108
 109         return ret;
 110 }
 111
 112 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 113                                         struct ipoib_cm_rx *rx,
 114                                         struct ib_recv_wr *wr,
 115                                         struct ib_sge *sge, int id)
 116 {
 117         struct ipoib_dev_priv *priv = netdev_priv(dev);
 118         struct ib_recv_wr *bad_wr;
 119         int i, ret;
 120
 121         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 122
 123         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 124                 sge[i].addr = rx->rx_ring[id].mapping[i];
 125
 126         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 127         if (unlikely(ret)) {
 128                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 129                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 130                                       rx->rx_ring[id].mapping);
 131                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 132                 rx->rx_ring[id].skb = NULL;
 133         }
 134
 135         return ret;
 136 }
 137
 138 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 139                                              struct ipoib_cm_rx_buf *rx_ring,
 140                                              int id, int frags,
 141                                              u64 mapping[IPOIB_CM_RX_SG])
 142 {
 143         struct ipoib_dev_priv *priv = netdev_priv(dev);
 144         struct sk_buff *skb;
 145         int i;
 146
 147         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 148         if (unlikely(!skb))
 149                 return NULL;
 150
 151         /*
 152          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 153          * IP header to a multiple of 16.
 154          */
 155         skb_reserve(skb, 12);
 156
 157         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 158                                        DMA_FROM_DEVICE);
 159         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 160                 dev_kfree_skb_any(skb);
 161                 return NULL;
 162         }
 163
 164         for (i = 0; i < frags; i++) {
 165                 struct page *page = alloc_page(GFP_ATOMIC);
 166
 167                 if (!page)
 168                         goto partial_error;
 169                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 170
 171                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 172                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 173                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 174                         goto partial_error;
 175         }
 176
 177         rx_ring[id].skb = skb;
 178         return skb;
 179
 180 partial_error:
 181
 182         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 183
 184         for (; i > 0; --i)
 185                 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 186
 187         dev_kfree_skb_any(skb);
 188         return NULL;
 189 }
 190
 191 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 192                                   struct ipoib_cm_rx_buf *rx_ring)
 193 {
 194         struct ipoib_dev_priv *priv = netdev_priv(dev);
 195         int i;
 196
 197         for (i = 0; i < ipoib_recvq_size; ++i)
 198                 if (rx_ring[i].skb) {
 199                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 200                                               rx_ring[i].mapping);
 201                         dev_kfree_skb_any(rx_ring[i].skb);
 202                 }
 203
 204         vfree(rx_ring);
 205 }
 206
 207 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 208 {
 209         struct ib_send_wr *bad_wr;
 210         struct ipoib_cm_rx *p;
 211
 212         /* We only reserved 1 extra slot in CQ for drain WRs, so
 213          * make sure we have at most 1 outstanding WR. */
 214         if (list_empty(&priv->cm.rx_flush_list) ||
 215             !list_empty(&priv->cm.rx_drain_list))
 216                 return;
 217
 218         /*
 219          * QPs on flush list are error state.  This way, a "flush
 220          * error" WC will be immediately generated for each WR we post.
 221          */
 222         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 223         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 224                 ipoib_warn(priv, "failed to post drain wr\n");
 225
 226         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 227 }
 228
 229 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 230 {
 231         struct ipoib_cm_rx *p = ctx;
 232         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 233         unsigned long flags;
 234
 235         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 236                 return;
 237
 238         spin_lock_irqsave(&priv->lock, flags);
 239         list_move(&p->list, &priv->cm.rx_flush_list);
 240         p->state = IPOIB_CM_RX_FLUSH;
 241         ipoib_cm_start_rx_drain(priv);
 242         spin_unlock_irqrestore(&priv->lock, flags);
 243 }
 244
 245 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 246                                            struct ipoib_cm_rx *p)
 247 {
 248         struct ipoib_dev_priv *priv = netdev_priv(dev);
 249         struct ib_qp_init_attr attr = {
 250                 .event_handler = ipoib_cm_rx_event_handler,
 251                 .send_cq = priv->recv_cq, /* For drain WR */
 252                 .recv_cq = priv->recv_cq,
 253                 .srq = priv->cm.srq,
 254                 .cap.max_send_wr = 1, /* For drain WR */
 255                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 256                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 257                 .qp_type = IB_QPT_RC,
 258                 .qp_context = p,
 259         };
 260
 261         if (!ipoib_cm_has_srq(dev)) {
 262                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 263                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 264         }
 265
 266         return ib_create_qp(priv->pd, &attr);
 267 }
 268
 269 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 270                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 271                                  unsigned psn)
 272 {
 273         struct ipoib_dev_priv *priv = netdev_priv(dev);
 274         struct ib_qp_attr qp_attr;
 275         int qp_attr_mask, ret;
 276
 277         qp_attr.qp_state = IB_QPS_INIT;
 278         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 279         if (ret) {
 280                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 281                 return ret;
 282         }
 283         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 284         if (ret) {
 285                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 286                 return ret;
 287         }
 288         qp_attr.qp_state = IB_QPS_RTR;
 289         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 290         if (ret) {
 291                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 292                 return ret;
 293         }
 294         qp_attr.rq_psn = psn;
 295         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 296         if (ret) {
 297                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 298                 return ret;
 299         }
 300
 301         /*
 302          * Current Mellanox HCA firmware won't generate completions
 303          * with error for drain WRs unless the QP has been moved to
 304          * RTS first. This work-around leaves a window where a QP has
 305          * moved to error asynchronously, but this will eventually get
 306          * fixed in firmware, so let's not error out if modify QP
 307          * fails.
 308          */
 309         qp_attr.qp_state = IB_QPS_RTS;
 310         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 311         if (ret) {
 312                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 313                 return 0;
 314         }
 315         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 316         if (ret) {
 317                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 318                 return 0;
 319         }
 320
 321         return 0;
 322 }
 323
 324 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 325                                 struct ib_recv_wr *wr,
 326                                 struct ib_sge *sge)
 327 {
 328         struct ipoib_dev_priv *priv = netdev_priv(dev);
 329         int i;
 330
 331         for (i = 0; i < priv->cm.num_frags; ++i)
 332                 sge[i].lkey = priv->mr->lkey;
 333
 334         sge[0].length = IPOIB_CM_HEAD_SIZE;
 335         for (i = 1; i < priv->cm.num_frags; ++i)
 336                 sge[i].length = PAGE_SIZE;
 337
 338         wr->next    = NULL;
 339         wr->sg_list = sge;
 340         wr->num_sge = priv->cm.num_frags;
 341 }
 342
 343 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 344                                    struct ipoib_cm_rx *rx)
 345 {
 346         struct ipoib_dev_priv *priv = netdev_priv(dev);
 347         struct {
 348                 struct ib_recv_wr wr;
 349                 struct ib_sge sge[IPOIB_CM_RX_SG];
 350         } *t;
 351         int ret;
 352         int i;
 353
 354         rx->rx_ring = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 355         if (!rx->rx_ring) {
 356                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 357                        priv->ca->name, ipoib_recvq_size);
 358                 return -ENOMEM;
 359         }
 360
 361         memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
 362
 363         t = kmalloc(sizeof *t, GFP_KERNEL);
 364         if (!t) {
 365                 ret = -ENOMEM;
 366                 goto err_free;
 367         }
 368
 369         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 370
 371         spin_lock_irq(&priv->lock);
 372
 373         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 374                 spin_unlock_irq(&priv->lock);
 375                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 376                 ret = -EINVAL;
 377                 goto err_free;
 378         } else
 379                 ++priv->cm.nonsrq_conn_qp;
 380
 381         spin_unlock_irq(&priv->lock);
 382
 383         for (i = 0; i < ipoib_recvq_size; ++i) {
 384                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 385                                            rx->rx_ring[i].mapping)) {
 386                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 387                                 ret = -ENOMEM;
 388                                 goto err_count;
 389                 }
 390                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 391                 if (ret) {
 392                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 393                                    "failed for buf %d\n", i);
 394                         ret = -EIO;
 395                         goto err_count;
 396                 }
 397         }
 398
 399         rx->recv_count = ipoib_recvq_size;
 400
 401         kfree(t);
 402
 403         return 0;
 404
 405 err_count:
 406         spin_lock_irq(&priv->lock);
 407         --priv->cm.nonsrq_conn_qp;
 408         spin_unlock_irq(&priv->lock);
 409
 410 err_free:
 411         kfree(t);
 412         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 413
 414         return ret;
 415 }
 416
 417 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 418                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 419                              unsigned psn)
 420 {
 421         struct ipoib_dev_priv *priv = netdev_priv(dev);
 422         struct ipoib_cm_data data = {};
 423         struct ib_cm_rep_param rep = {};
 424
 425         data.qpn = cpu_to_be32(priv->qp->qp_num);
 426         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 427
 428         rep.private_data = &data;
 429         rep.private_data_len = sizeof data;
 430         rep.flow_control = 0;
 431         rep.rnr_retry_count = req->rnr_retry_count;
 432         rep.srq = ipoib_cm_has_srq(dev);
 433         rep.qp_num = qp->qp_num;
 434         rep.starting_psn = psn;
 435         return ib_send_cm_rep(cm_id, &rep);
 436 }
 437
 438 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 439 {
 440         struct net_device *dev = cm_id->context;
 441         struct ipoib_dev_priv *priv = netdev_priv(dev);
 442         struct ipoib_cm_rx *p;
 443         unsigned psn;
 444         int ret;
 445
 446         ipoib_dbg(priv, "REQ arrived\n");
 447         p = kzalloc(sizeof *p, GFP_KERNEL);
 448         if (!p)
 449                 return -ENOMEM;
 450         p->dev = dev;
 451         p->id = cm_id;
 452         cm_id->context = p;
 453         p->state = IPOIB_CM_RX_LIVE;
 454         p->jiffies = jiffies;
 455         INIT_LIST_HEAD(&p->list);
 456
 457         p->qp = ipoib_cm_create_rx_qp(dev, p);
 458         if (IS_ERR(p->qp)) {
 459                 ret = PTR_ERR(p->qp);
 460                 goto err_qp;
 461         }
 462
 463         psn = random32() & 0xffffff;
 464         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 465         if (ret)
 466                 goto err_modify;
 467
 468         if (!ipoib_cm_has_srq(dev)) {
 469                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 470                 if (ret)
 471                         goto err_modify;
 472         }
 473
 474         spin_lock_irq(&priv->lock);
 475         queue_delayed_work(ipoib_workqueue,
 476                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 477         /* Add this entry to passive ids list head, but do not re-add it
 478          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 479         p->jiffies = jiffies;
 480         if (p->state == IPOIB_CM_RX_LIVE)
 481                 list_move(&p->list, &priv->cm.passive_ids);
 482         spin_unlock_irq(&priv->lock);
 483
 484         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 485         if (ret) {
 486                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 487                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 488                         ipoib_warn(priv, "unable to move qp to error state\n");
 489         }
 490         return 0;
 491
 492 err_modify:
 493         ib_destroy_qp(p->qp);
 494 err_qp:
 495         kfree(p);
 496         return ret;
 497 }
 498
 499 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 500                                struct ib_cm_event *event)
 501 {
 502         struct ipoib_cm_rx *p;
 503         struct ipoib_dev_priv *priv;
 504
 505         switch (event->event) {
 506         case IB_CM_REQ_RECEIVED:
 507                 return ipoib_cm_req_handler(cm_id, event);
 508         case IB_CM_DREQ_RECEIVED:
 509                 p = cm_id->context;
 510                 ib_send_cm_drep(cm_id, NULL, 0);
 511                 /* Fall through */
 512         case IB_CM_REJ_RECEIVED:
 513                 p = cm_id->context;
 514                 priv = netdev_priv(p->dev);
 515                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 516                         ipoib_warn(priv, "unable to move qp to error state\n");
 517                 /* Fall through */
 518         default:
 519                 return 0;
 520         }
 521 }
 522 /* Adjust length of skb with fragments to match received data */
 523 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 524                           unsigned int length, struct sk_buff *toskb)
 525 {
 526         int i, num_frags;
 527         unsigned int size;
 528
 529         /* put header into skb */
 530         size = min(length, hdr_space);
 531         skb->tail += size;
 532         skb->len += size;
 533         length -= size;
 534
 535         num_frags = skb_shinfo(skb)->nr_frags;
 536         for (i = 0; i < num_frags; i++) {
 537                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 538
 539                 if (length == 0) {
 540                         /* don't need this page */
 541                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 542                         --skb_shinfo(skb)->nr_frags;
 543                 } else {
 544                         size = min(length, (unsigned) PAGE_SIZE);
 545
 546                         frag->size = size;
 547                         skb->data_len += size;
 548                         skb->truesize += size;
 549                         skb->len += size;
 550                         length -= size;
 551                 }
 552         }
 553 }
 554
 555 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 556 {
 557         struct ipoib_dev_priv *priv = netdev_priv(dev);
 558         struct ipoib_cm_rx_buf *rx_ring;
 559         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 560         struct sk_buff *skb, *newskb;
 561         struct ipoib_cm_rx *p;
 562         unsigned long flags;
 563         u64 mapping[IPOIB_CM_RX_SG];
 564         int frags;
 565         int has_srq;
 566         struct sk_buff *small_skb;
 567
 568         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 569                        wr_id, wc->status);
 570
 571         if (unlikely(wr_id >= ipoib_recvq_size)) {
 572                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 573                         spin_lock_irqsave(&priv->lock, flags);
 574                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 575                         ipoib_cm_start_rx_drain(priv);
 576                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 577                         spin_unlock_irqrestore(&priv->lock, flags);
 578                 } else
 579                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 580                                    wr_id, ipoib_recvq_size);
 581                 return;
 582         }
 583
 584         p = wc->qp->qp_context;
 585
 586         has_srq = ipoib_cm_has_srq(dev);
 587         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 588
 589         skb = rx_ring[wr_id].skb;
 590
 591         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 592                 ipoib_dbg(priv, "cm recv error "
 593                            "(status=%d, wrid=%d vend_err %x)\n",
 594                            wc->status, wr_id, wc->vendor_err);
 595                 ++dev->stats.rx_dropped;
 596                 if (has_srq)
 597                         goto repost;
 598                 else {
 599                         if (!--p->recv_count) {
 600                                 spin_lock_irqsave(&priv->lock, flags);
 601                                 list_move(&p->list, &priv->cm.rx_reap_list);
 602                                 spin_unlock_irqrestore(&priv->lock, flags);
 603                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 604                         }
 605                         return;
 606                 }
 607         }
 608
 609         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 610                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 611                         spin_lock_irqsave(&priv->lock, flags);
 612                         p->jiffies = jiffies;
 613                         /* Move this entry to list head, but do not re-add it
 614                          * if it has been moved out of list. */
 615                         if (p->state == IPOIB_CM_RX_LIVE)
 616                                 list_move(&p->list, &priv->cm.passive_ids);
 617                         spin_unlock_irqrestore(&priv->lock, flags);
 618                 }
 619         }
 620
 621         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 622                 int dlen = wc->byte_len;
 623
 624                 small_skb = dev_alloc_skb(dlen + 12);
 625                 if (small_skb) {
 626                         skb_reserve(small_skb, 12);
 627                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 628                                                    dlen, DMA_FROM_DEVICE);
 629                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 630                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 631                                                       dlen, DMA_FROM_DEVICE);
 632                         skb_put(small_skb, dlen);
 633                         skb = small_skb;
 634                         goto copied;
 635                 }
 636         }
 637
 638         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 639                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 640
 641         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 642         if (unlikely(!newskb)) {
 643                 /*
 644                  * If we can't allocate a new RX buffer, dump
 645                  * this packet and reuse the old buffer.
 646                  */
 647                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 648                 ++dev->stats.rx_dropped;
 649                 goto repost;
 650         }
 651
 652         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 653         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 654
 655         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 656                        wc->byte_len, wc->slid);
 657
 658         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 659
 660 copied:
 661         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 662         skb_reset_mac_header(skb);
 663         skb_pull(skb, IPOIB_ENCAP_LEN);
 664
 665         ++dev->stats.rx_packets;
 666         dev->stats.rx_bytes += skb->len;
 667
 668         skb->dev = dev;
 669         /* XXX get correct PACKET_ type here */
 670         skb->pkt_type = PACKET_HOST;
 671         netif_receive_skb(skb);
 672
 673 repost:
 674         if (has_srq) {
 675                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 676                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 677                                    "for buf %d\n", wr_id);
 678         } else {
 679                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 680                                                           &priv->cm.rx_wr,
 681                                                           priv->cm.rx_sge,
 682                                                           wr_id))) {
 683                         --p->recv_count;
 684                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 685                                    "for buf %d\n", wr_id);
 686                 }
 687         }
 688 }
 689
 690 static inline int post_send(struct ipoib_dev_priv *priv,
 691                             struct ipoib_cm_tx *tx,
 692                             unsigned int wr_id,
 693                             u64 addr, int len)
 694 {
 695         struct ib_send_wr *bad_wr;
 696
 697         priv->tx_sge[0].addr          = addr;
 698         priv->tx_sge[0].length        = len;
 699
 700         priv->tx_wr.num_sge     = 1;
 701         priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 702
 703         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 704 }
 705
 706 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 707 {
 708         struct ipoib_dev_priv *priv = netdev_priv(dev);
 709         struct ipoib_cm_tx_buf *tx_req;
 710         u64 addr;
 711         int rc;
 712
 713         if (unlikely(skb->len > tx->mtu)) {
 714                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 715                            skb->len, tx->mtu);
 716                 ++dev->stats.tx_dropped;
 717                 ++dev->stats.tx_errors;
 718                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 719                 return;
 720         }
 721
 722         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 723                        tx->tx_head, skb->len, tx->qp->qp_num);
 724
 725         /*
 726          * We put the skb into the tx_ring _before_ we call post_send()
 727          * because it's entirely possible that the completion handler will
 728          * run before we execute anything after the post_send().  That
 729          * means we have to make sure everything is properly recorded and
 730          * our state is consistent before we call post_send().
 731          */
 732         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 733         tx_req->skb = skb;
 734         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 735         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 736                 ++dev->stats.tx_errors;
 737                 dev_kfree_skb_any(skb);
 738                 return;
 739         }
 740
 741         tx_req->mapping = addr;
 742
 743         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 744                        addr, skb->len);
 745         if (unlikely(rc)) {
 746                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 747                 ++dev->stats.tx_errors;
 748                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 749                 dev_kfree_skb_any(skb);
 750         } else {
 751                 dev->trans_start = jiffies;
 752                 ++tx->tx_head;
 753
 754                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 755                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 756                                   tx->qp->qp_num);
 757                         if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 758                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 759                         netif_stop_queue(dev);
 760                 }
 761         }
 762 }
 763
 764 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 765 {
 766         struct ipoib_dev_priv *priv = netdev_priv(dev);
 767         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 768         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 769         struct ipoib_cm_tx_buf *tx_req;
 770         unsigned long flags;
 771
 772         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 773                        wr_id, wc->status);
 774
 775         if (unlikely(wr_id >= ipoib_sendq_size)) {
 776                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 777                            wr_id, ipoib_sendq_size);
 778                 return;
 779         }
 780
 781         tx_req = &tx->tx_ring[wr_id];
 782
 783         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 784
 785         /* FIXME: is this right? Shouldn't we only increment on success? */
 786         ++dev->stats.tx_packets;
 787         dev->stats.tx_bytes += tx_req->skb->len;
 788
 789         dev_kfree_skb_any(tx_req->skb);
 790
 791         netif_tx_lock(dev);
 792
 793         ++tx->tx_tail;
 794         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 795             netif_queue_stopped(dev) &&
 796             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 797                 netif_wake_queue(dev);
 798
 799         if (wc->status != IB_WC_SUCCESS &&
 800             wc->status != IB_WC_WR_FLUSH_ERR) {
 801                 struct ipoib_neigh *neigh;
 802
 803                 ipoib_dbg(priv, "failed cm send event "
 804                            "(status=%d, wrid=%d vend_err %x)\n",
 805                            wc->status, wr_id, wc->vendor_err);
 806
 807                 spin_lock_irqsave(&priv->lock, flags);
 808                 neigh = tx->neigh;
 809
 810                 if (neigh) {
 811                         neigh->cm = NULL;
 812                         list_del(&neigh->list);
 813                         if (neigh->ah)
 814                                 ipoib_put_ah(neigh->ah);
 815                         ipoib_neigh_free(dev, neigh);
 816
 817                         tx->neigh = NULL;
 818                 }
 819
 820                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 821                         list_move(&tx->list, &priv->cm.reap_list);
 822                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 823                 }
 824
 825                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 826
 827                 spin_unlock_irqrestore(&priv->lock, flags);
 828         }
 829
 830         netif_tx_unlock(dev);
 831 }
 832
 833 int ipoib_cm_dev_open(struct net_device *dev)
 834 {
 835         struct ipoib_dev_priv *priv = netdev_priv(dev);
 836         int ret;
 837
 838         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 839                 return 0;
 840
 841         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 842         if (IS_ERR(priv->cm.id)) {
 843                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 844                 ret = PTR_ERR(priv->cm.id);
 845                 goto err_cm;
 846         }
 847
 848         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 849                            0, NULL);
 850         if (ret) {
 851                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 852                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 853                 goto err_listen;
 854         }
 855
 856         return 0;
 857
 858 err_listen:
 859         ib_destroy_cm_id(priv->cm.id);
 860 err_cm:
 861         priv->cm.id = NULL;
 862         return ret;
 863 }
 864
 865 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 866 {
 867         struct ipoib_dev_priv *priv = netdev_priv(dev);
 868         struct ipoib_cm_rx *rx, *n;
 869         LIST_HEAD(list);
 870
 871         spin_lock_irq(&priv->lock);
 872         list_splice_init(&priv->cm.rx_reap_list, &list);
 873         spin_unlock_irq(&priv->lock);
 874
 875         list_for_each_entry_safe(rx, n, &list, list) {
 876                 ib_destroy_cm_id(rx->id);
 877                 ib_destroy_qp(rx->qp);
 878                 if (!ipoib_cm_has_srq(dev)) {
 879                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 880                         spin_lock_irq(&priv->lock);
 881                         --priv->cm.nonsrq_conn_qp;
 882                         spin_unlock_irq(&priv->lock);
 883                 }
 884                 kfree(rx);
 885         }
 886 }
 887
 888 void ipoib_cm_dev_stop(struct net_device *dev)
 889 {
 890         struct ipoib_dev_priv *priv = netdev_priv(dev);
 891         struct ipoib_cm_rx *p;
 892         unsigned long begin;
 893         int ret;
 894
 895         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 896                 return;
 897
 898         ib_destroy_cm_id(priv->cm.id);
 899         priv->cm.id = NULL;
 900
 901         spin_lock_irq(&priv->lock);
 902         while (!list_empty(&priv->cm.passive_ids)) {
 903                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 904                 list_move(&p->list, &priv->cm.rx_error_list);
 905                 p->state = IPOIB_CM_RX_ERROR;
 906                 spin_unlock_irq(&priv->lock);
 907                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 908                 if (ret)
 909                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 910                 spin_lock_irq(&priv->lock);
 911         }
 912
 913         /* Wait for all RX to be drained */
 914         begin = jiffies;
 915
 916         while (!list_empty(&priv->cm.rx_error_list) ||
 917                !list_empty(&priv->cm.rx_flush_list) ||
 918                !list_empty(&priv->cm.rx_drain_list)) {
 919                 if (time_after(jiffies, begin + 5 * HZ)) {
 920                         ipoib_warn(priv, "RX drain timing out\n");
 921
 922                         /*
 923                          * assume the HW is wedged and just free up everything.
 924                          */
 925                         list_splice_init(&priv->cm.rx_flush_list,
 926                                          &priv->cm.rx_reap_list);
 927                         list_splice_init(&priv->cm.rx_error_list,
 928                                          &priv->cm.rx_reap_list);
 929                         list_splice_init(&priv->cm.rx_drain_list,
 930                                          &priv->cm.rx_reap_list);
 931                         break;
 932                 }
 933                 spin_unlock_irq(&priv->lock);
 934                 msleep(1);
 935                 ipoib_drain_cq(dev);
 936                 spin_lock_irq(&priv->lock);
 937         }
 938
 939         spin_unlock_irq(&priv->lock);
 940
 941         ipoib_cm_free_rx_reap_list(dev);
 942
 943         cancel_delayed_work(&priv->cm.stale_task);
 944 }
 945
 946 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 947 {
 948         struct ipoib_cm_tx *p = cm_id->context;
 949         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 950         struct ipoib_cm_data *data = event->private_data;
 951         struct sk_buff_head skqueue;
 952         struct ib_qp_attr qp_attr;
 953         int qp_attr_mask, ret;
 954         struct sk_buff *skb;
 955
 956         p->mtu = be32_to_cpu(data->mtu);
 957
 958         if (p->mtu <= IPOIB_ENCAP_LEN) {
 959                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 960                            p->mtu, IPOIB_ENCAP_LEN);
 961                 return -EINVAL;
 962         }
 963
 964         qp_attr.qp_state = IB_QPS_RTR;
 965         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 966         if (ret) {
 967                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 968                 return ret;
 969         }
 970
 971         qp_attr.rq_psn = 0 /* FIXME */;
 972         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 973         if (ret) {
 974                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 975                 return ret;
 976         }
 977
 978         qp_attr.qp_state = IB_QPS_RTS;
 979         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 980         if (ret) {
 981                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 982                 return ret;
 983         }
 984         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 985         if (ret) {
 986                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 987                 return ret;
 988         }
 989
 990         skb_queue_head_init(&skqueue);
 991
 992         spin_lock_irq(&priv->lock);
 993         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 994         if (p->neigh)
 995                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 996                         __skb_queue_tail(&skqueue, skb);
 997         spin_unlock_irq(&priv->lock);
 998
 999         while ((skb = __skb_dequeue(&skqueue))) {
1000                 skb->dev = p->dev;
1001                 if (dev_queue_xmit(skb))
1002                         ipoib_warn(priv, "dev_queue_xmit failed "
1003                                    "to requeue packet\n");
1004         }
1005
1006         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1007         if (ret) {
1008                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1009                 return ret;
1010         }
1011         return 0;
1012 }
1013
1014 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1015 {
1016         struct ipoib_dev_priv *priv = netdev_priv(dev);
1017         struct ib_qp_init_attr attr = {
1018                 .send_cq                = priv->recv_cq,
1019                 .recv_cq                = priv->recv_cq,
1020                 .srq                    = priv->cm.srq,
1021                 .cap.max_send_wr        = ipoib_sendq_size,
1022                 .cap.max_send_sge       = 1,
1023                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1024                 .qp_type                = IB_QPT_RC,
1025                 .qp_context             = tx
1026         };
1027
1028         return ib_create_qp(priv->pd, &attr);
1029 }
1030
1031 static int ipoib_cm_send_req(struct net_device *dev,
1032                              struct ib_cm_id *id, struct ib_qp *qp,
1033                              u32 qpn,
1034                              struct ib_sa_path_rec *pathrec)
1035 {
1036         struct ipoib_dev_priv *priv = netdev_priv(dev);
1037         struct ipoib_cm_data data = {};
1038         struct ib_cm_req_param req = {};
1039
1040         data.qpn = cpu_to_be32(priv->qp->qp_num);
1041         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1042
1043         req.primary_path                = pathrec;
1044         req.alternate_path              = NULL;
1045         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1046         req.qp_num                      = qp->qp_num;
1047         req.qp_type                     = qp->qp_type;
1048         req.private_data                = &data;
1049         req.private_data_len            = sizeof data;
1050         req.flow_control                = 0;
1051
1052         req.starting_psn                = 0; /* FIXME */
1053
1054         /*
1055          * Pick some arbitrary defaults here; we could make these
1056          * module parameters if anyone cared about setting them.
1057          */
1058         req.responder_resources         = 4;
1059         req.remote_cm_response_timeout  = 20;
1060         req.local_cm_response_timeout   = 20;
1061         req.retry_count                 = 0; /* RFC draft warns against retries */
1062         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1063         req.max_cm_retries              = 15;
1064         req.srq                         = ipoib_cm_has_srq(dev);
1065         return ib_send_cm_req(id, &req);
1066 }
1067
1068 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1069                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1070 {
1071         struct ipoib_dev_priv *priv = netdev_priv(dev);
1072         struct ib_qp_attr qp_attr;
1073         int qp_attr_mask, ret;
1074         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1075         if (ret) {
1076                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1077                 return ret;
1078         }
1079
1080         qp_attr.qp_state = IB_QPS_INIT;
1081         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1082         qp_attr.port_num = priv->port;
1083         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1084
1085         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1086         if (ret) {
1087                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1088                 return ret;
1089         }
1090         return 0;
1091 }
1092
1093 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1094                             struct ib_sa_path_rec *pathrec)
1095 {
1096         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1097         int ret;
1098
1099         p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1100         if (!p->tx_ring) {
1101                 ipoib_warn(priv, "failed to allocate tx ring\n");
1102                 ret = -ENOMEM;
1103                 goto err_tx;
1104         }
1105         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1106
1107         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1108         if (IS_ERR(p->qp)) {
1109                 ret = PTR_ERR(p->qp);
1110                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1111                 goto err_qp;
1112         }
1113
1114         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1115         if (IS_ERR(p->id)) {
1116                 ret = PTR_ERR(p->id);
1117                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1118                 goto err_id;
1119         }
1120
1121         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1122         if (ret) {
1123                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1124                 goto err_modify;
1125         }
1126
1127         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1128         if (ret) {
1129                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1130                 goto err_send_cm;
1131         }
1132
1133         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1134                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1135
1136         return 0;
1137
1138 err_send_cm:
1139 err_modify:
1140         ib_destroy_cm_id(p->id);
1141 err_id:
1142         p->id = NULL;
1143         ib_destroy_qp(p->qp);
1144 err_qp:
1145         p->qp = NULL;
1146         vfree(p->tx_ring);
1147 err_tx:
1148         return ret;
1149 }
1150
1151 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1152 {
1153         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1154         struct ipoib_cm_tx_buf *tx_req;
1155         unsigned long begin;
1156
1157         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1158                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1159
1160         if (p->id)
1161                 ib_destroy_cm_id(p->id);
1162
1163         if (p->tx_ring) {
1164                 /* Wait for all sends to complete */
1165                 begin = jiffies;
1166                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1167                         if (time_after(jiffies, begin + 5 * HZ)) {
1168                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1169                                            p->tx_head - p->tx_tail);
1170                                 goto timeout;
1171                         }
1172
1173                         msleep(1);
1174                 }
1175         }
1176
1177 timeout:
1178
1179         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1180                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1181                 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1182                                     DMA_TO_DEVICE);
1183                 dev_kfree_skb_any(tx_req->skb);
1184                 ++p->tx_tail;
1185                 netif_tx_lock_bh(p->dev);
1186                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1187                     netif_queue_stopped(p->dev) &&
1188                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1189                         netif_wake_queue(p->dev);
1190                 netif_tx_unlock_bh(p->dev);
1191         }
1192
1193         if (p->qp)
1194                 ib_destroy_qp(p->qp);
1195
1196         vfree(p->tx_ring);
1197         kfree(p);
1198 }
1199
1200 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1201                                struct ib_cm_event *event)
1202 {
1203         struct ipoib_cm_tx *tx = cm_id->context;
1204         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1205         struct net_device *dev = priv->dev;
1206         struct ipoib_neigh *neigh;
1207         unsigned long flags;
1208         int ret;
1209
1210         switch (event->event) {
1211         case IB_CM_DREQ_RECEIVED:
1212                 ipoib_dbg(priv, "DREQ received.\n");
1213                 ib_send_cm_drep(cm_id, NULL, 0);
1214                 break;
1215         case IB_CM_REP_RECEIVED:
1216                 ipoib_dbg(priv, "REP received.\n");
1217                 ret = ipoib_cm_rep_handler(cm_id, event);
1218                 if (ret)
1219                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1220                                        NULL, 0, NULL, 0);
1221                 break;
1222         case IB_CM_REQ_ERROR:
1223         case IB_CM_REJ_RECEIVED:
1224         case IB_CM_TIMEWAIT_EXIT:
1225                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1226                 netif_tx_lock_bh(dev);
1227                 spin_lock_irqsave(&priv->lock, flags);
1228                 neigh = tx->neigh;
1229
1230                 if (neigh) {
1231                         neigh->cm = NULL;
1232                         list_del(&neigh->list);
1233                         if (neigh->ah)
1234                                 ipoib_put_ah(neigh->ah);
1235                         ipoib_neigh_free(dev, neigh);
1236
1237                         tx->neigh = NULL;
1238                 }
1239
1240                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1241                         list_move(&tx->list, &priv->cm.reap_list);
1242                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1243                 }
1244
1245                 spin_unlock_irqrestore(&priv->lock, flags);
1246                 netif_tx_unlock_bh(dev);
1247                 break;
1248         default:
1249                 break;
1250         }
1251
1252         return 0;
1253 }
1254
1255 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1256                                        struct ipoib_neigh *neigh)
1257 {
1258         struct ipoib_dev_priv *priv = netdev_priv(dev);
1259         struct ipoib_cm_tx *tx;
1260
1261         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1262         if (!tx)
1263                 return NULL;
1264
1265         neigh->cm = tx;
1266         tx->neigh = neigh;
1267         tx->path = path;
1268         tx->dev = dev;
1269         list_add(&tx->list, &priv->cm.start_list);
1270         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1271         queue_work(ipoib_workqueue, &priv->cm.start_task);
1272         return tx;
1273 }
1274
1275 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1276 {
1277         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1278         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1279                 list_move(&tx->list, &priv->cm.reap_list);
1280                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1281                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1282                           tx->neigh->dgid.raw);
1283                 tx->neigh = NULL;
1284         }
1285 }
1286
1287 static void ipoib_cm_tx_start(struct work_struct *work)
1288 {
1289         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1290                                                    cm.start_task);
1291         struct net_device *dev = priv->dev;
1292         struct ipoib_neigh *neigh;
1293         struct ipoib_cm_tx *p;
1294         unsigned long flags;
1295         int ret;
1296
1297         struct ib_sa_path_rec pathrec;
1298         u32 qpn;
1299
1300         netif_tx_lock_bh(dev);
1301         spin_lock_irqsave(&priv->lock, flags);
1302
1303         while (!list_empty(&priv->cm.start_list)) {
1304                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1305                 list_del_init(&p->list);
1306                 neigh = p->neigh;
1307                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1308                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1309
1310                 spin_unlock_irqrestore(&priv->lock, flags);
1311                 netif_tx_unlock_bh(dev);
1312
1313                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1314
1315                 netif_tx_lock_bh(dev);
1316                 spin_lock_irqsave(&priv->lock, flags);
1317
1318                 if (ret) {
1319                         neigh = p->neigh;
1320                         if (neigh) {
1321                                 neigh->cm = NULL;
1322                                 list_del(&neigh->list);
1323                                 if (neigh->ah)
1324                                         ipoib_put_ah(neigh->ah);
1325                                 ipoib_neigh_free(dev, neigh);
1326                         }
1327                         list_del(&p->list);
1328                         kfree(p);
1329                 }
1330         }
1331
1332         spin_unlock_irqrestore(&priv->lock, flags);
1333         netif_tx_unlock_bh(dev);
1334 }
1335
1336 static void ipoib_cm_tx_reap(struct work_struct *work)
1337 {
1338         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1339                                                    cm.reap_task);
1340         struct net_device *dev = priv->dev;
1341         struct ipoib_cm_tx *p;
1342         unsigned long flags;
1343
1344         netif_tx_lock_bh(dev);
1345         spin_lock_irqsave(&priv->lock, flags);
1346
1347         while (!list_empty(&priv->cm.reap_list)) {
1348                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1349                 list_del(&p->list);
1350                 spin_unlock_irqrestore(&priv->lock, flags);
1351                 netif_tx_unlock_bh(dev);
1352                 ipoib_cm_tx_destroy(p);
1353                 netif_tx_lock_bh(dev);
1354                 spin_lock_irqsave(&priv->lock, flags);
1355         }
1356
1357         spin_unlock_irqrestore(&priv->lock, flags);
1358         netif_tx_unlock_bh(dev);
1359 }
1360
1361 static void ipoib_cm_skb_reap(struct work_struct *work)
1362 {
1363         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1364                                                    cm.skb_task);
1365         struct net_device *dev = priv->dev;
1366         struct sk_buff *skb;
1367         unsigned long flags;
1368         unsigned mtu = priv->mcast_mtu;
1369
1370         netif_tx_lock_bh(dev);
1371         spin_lock_irqsave(&priv->lock, flags);
1372
1373         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1374                 spin_unlock_irqrestore(&priv->lock, flags);
1375                 netif_tx_unlock_bh(dev);
1376
1377                 if (skb->protocol == htons(ETH_P_IP))
1378                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1379 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1380                 else if (skb->protocol == htons(ETH_P_IPV6))
1381                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1382 #endif
1383                 dev_kfree_skb_any(skb);
1384
1385                 netif_tx_lock_bh(dev);
1386                 spin_lock_irqsave(&priv->lock, flags);
1387         }
1388
1389         spin_unlock_irqrestore(&priv->lock, flags);
1390         netif_tx_unlock_bh(dev);
1391 }
1392
1393 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1394                            unsigned int mtu)
1395 {
1396         struct ipoib_dev_priv *priv = netdev_priv(dev);
1397         int e = skb_queue_empty(&priv->cm.skb_queue);
1398
1399         if (skb_dst(skb))
1400                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1401
1402         skb_queue_tail(&priv->cm.skb_queue, skb);
1403         if (e)
1404                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1405 }
1406
1407 static void ipoib_cm_rx_reap(struct work_struct *work)
1408 {
1409         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1410                                                 cm.rx_reap_task)->dev);
1411 }
1412
1413 static void ipoib_cm_stale_task(struct work_struct *work)
1414 {
1415         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1416                                                    cm.stale_task.work);
1417         struct ipoib_cm_rx *p;
1418         int ret;
1419
1420         spin_lock_irq(&priv->lock);
1421         while (!list_empty(&priv->cm.passive_ids)) {
1422                 /* List is sorted by LRU, start from tail,
1423                  * stop when we see a recently used entry */
1424                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1425                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1426                         break;
1427                 list_move(&p->list, &priv->cm.rx_error_list);
1428                 p->state = IPOIB_CM_RX_ERROR;
1429                 spin_unlock_irq(&priv->lock);
1430                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1431                 if (ret)
1432                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1433                 spin_lock_irq(&priv->lock);
1434         }
1435
1436         if (!list_empty(&priv->cm.passive_ids))
1437                 queue_delayed_work(ipoib_workqueue,
1438                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1439         spin_unlock_irq(&priv->lock);
1440 }
1441
1442
1443 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1444                          char *buf)
1445 {
1446         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1447
1448         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1449                 return sprintf(buf, "connected\n");
1450         else
1451                 return sprintf(buf, "datagram\n");
1452 }
1453
1454 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1455                         const char *buf, size_t count)
1456 {
1457         struct net_device *dev = to_net_dev(d);
1458         struct ipoib_dev_priv *priv = netdev_priv(dev);
1459
1460         if (!rtnl_trylock())
1461                 return restart_syscall();
1462
1463         /* flush paths if we switch modes so that connections are restarted */
1464         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1465                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1466                 ipoib_warn(priv, "enabling connected mode "
1467                            "will cause multicast packet drops\n");
1468
1469                 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1470                 rtnl_unlock();
1471                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1472
1473                 ipoib_flush_paths(dev);
1474                 return count;
1475         }
1476
1477         if (!strcmp(buf, "datagram\n")) {
1478                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1479
1480                 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1481                         dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1482                         if (priv->hca_caps & IB_DEVICE_UD_TSO)
1483                                 dev->features |= NETIF_F_TSO;
1484                 }
1485                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1486                 rtnl_unlock();
1487                 ipoib_flush_paths(dev);
1488
1489                 return count;
1490         }
1491         rtnl_unlock();
1492
1493         return -EINVAL;
1494 }
1495
1496 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1497
1498 int ipoib_cm_add_mode_attr(struct net_device *dev)
1499 {
1500         return device_create_file(&dev->dev, &dev_attr_mode);
1501 }
1502
1503 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1504 {
1505         struct ipoib_dev_priv *priv = netdev_priv(dev);
1506         struct ib_srq_init_attr srq_init_attr = {
1507                 .attr = {
1508                         .max_wr  = ipoib_recvq_size,
1509                         .max_sge = max_sge
1510                 }
1511         };
1512
1513         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1514         if (IS_ERR(priv->cm.srq)) {
1515                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1516                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1517                                priv->ca->name, PTR_ERR(priv->cm.srq));
1518                 priv->cm.srq = NULL;
1519                 return;
1520         }
1521
1522         priv->cm.srq_ring = vmalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1523         if (!priv->cm.srq_ring) {
1524                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1525                        priv->ca->name, ipoib_recvq_size);
1526                 ib_destroy_srq(priv->cm.srq);
1527                 priv->cm.srq = NULL;
1528                 return;
1529         }
1530
1531         memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1532 }
1533
1534 int ipoib_cm_dev_init(struct net_device *dev)
1535 {
1536         struct ipoib_dev_priv *priv = netdev_priv(dev);
1537         int i, ret;
1538         struct ib_device_attr attr;
1539
1540         INIT_LIST_HEAD(&priv->cm.passive_ids);
1541         INIT_LIST_HEAD(&priv->cm.reap_list);
1542         INIT_LIST_HEAD(&priv->cm.start_list);
1543         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1544         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1545         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1546         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1547         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1548         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1549         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1550         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1551         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1552
1553         skb_queue_head_init(&priv->cm.skb_queue);
1554
1555         ret = ib_query_device(priv->ca, &attr);
1556         if (ret) {
1557                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1558                 return ret;
1559         }
1560
1561         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1562
1563         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1564         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1565         if (ipoib_cm_has_srq(dev)) {
1566                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1567                 priv->cm.num_frags  = attr.max_srq_sge;
1568                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1569                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1570         } else {
1571                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1572                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1573         }
1574
1575         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1576
1577         if (ipoib_cm_has_srq(dev)) {
1578                 for (i = 0; i < ipoib_recvq_size; ++i) {
1579                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1580                                                    priv->cm.num_frags - 1,
1581                                                    priv->cm.srq_ring[i].mapping)) {
1582                                 ipoib_warn(priv, "failed to allocate "
1583                                            "receive buffer %d\n", i);
1584                                 ipoib_cm_dev_cleanup(dev);
1585                                 return -ENOMEM;
1586                         }
1587
1588                         if (ipoib_cm_post_receive_srq(dev, i)) {
1589                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1590                                            "failed for buf %d\n", i);
1591                                 ipoib_cm_dev_cleanup(dev);
1592                                 return -EIO;
1593                         }
1594                 }
1595         }
1596
1597         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1598         return 0;
1599 }
1600
1601 void ipoib_cm_dev_cleanup(struct net_device *dev)
1602 {
1603         struct ipoib_dev_priv *priv = netdev_priv(dev);
1604         int ret;
1605
1606         if (!priv->cm.srq)
1607                 return;
1608
1609         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1610
1611         ret = ib_destroy_srq(priv->cm.srq);
1612         if (ret)
1613                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1614
1615         priv->cm.srq = NULL;
1616         if (!priv->cm.srq_ring)
1617                 return;
1618
1619         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1620         priv->cm.srq_ring = NULL;
1621 }