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 #include <linux/moduleparam.h>
  41
  42 #include "ipoib.h"
  43
  44 int ipoib_max_conn_qp = 128;
  45
  46 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  47 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  48                  "Max number of connected-mode QPs per interface "
  49                  "(applied only if shared receive queue is not available)");
  50
  51 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  52 static int data_debug_level;
  53
  54 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  55 MODULE_PARM_DESC(cm_data_debug_level,
  56                  "Enable data path debug tracing for connected mode if > 0");
  57 #endif
  58
  59 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  60
  61 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  62 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  63 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  64 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  65
  66 static struct ib_qp_attr ipoib_cm_err_attr = {
  67         .qp_state = IB_QPS_ERR
  68 };
  69
  70 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  71
  72 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  73         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  74         .opcode = IB_WR_SEND,
  75 };
  76
  77 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  78                                struct ib_cm_event *event);
  79
  80 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  81                                   u64 mapping[IPOIB_CM_RX_SG])
  82 {
  83         int i;
  84
  85         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  86
  87         for (i = 0; i < frags; ++i)
  88                 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  89 }
  90
  91 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  92 {
  93         struct ipoib_dev_priv *priv = netdev_priv(dev);
  94         struct ib_recv_wr *bad_wr;
  95         int i, ret;
  96
  97         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  98
  99         for (i = 0; i < priv->cm.num_frags; ++i)
 100                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 101
 102         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 103         if (unlikely(ret)) {
 104                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 105                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 106                                       priv->cm.srq_ring[id].mapping);
 107                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 108                 priv->cm.srq_ring[id].skb = NULL;
 109         }
 110
 111         return ret;
 112 }
 113
 114 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 115                                         struct ipoib_cm_rx *rx,
 116                                         struct ib_recv_wr *wr,
 117                                         struct ib_sge *sge, int id)
 118 {
 119         struct ipoib_dev_priv *priv = netdev_priv(dev);
 120         struct ib_recv_wr *bad_wr;
 121         int i, ret;
 122
 123         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 124
 125         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 126                 sge[i].addr = rx->rx_ring[id].mapping[i];
 127
 128         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 129         if (unlikely(ret)) {
 130                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 131                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 132                                       rx->rx_ring[id].mapping);
 133                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 134                 rx->rx_ring[id].skb = NULL;
 135         }
 136
 137         return ret;
 138 }
 139
 140 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 141                                              struct ipoib_cm_rx_buf *rx_ring,
 142                                              int id, int frags,
 143                                              u64 mapping[IPOIB_CM_RX_SG])
 144 {
 145         struct ipoib_dev_priv *priv = netdev_priv(dev);
 146         struct sk_buff *skb;
 147         int i;
 148
 149         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 150         if (unlikely(!skb))
 151                 return NULL;
 152
 153         /*
 154          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 155          * IP header to a multiple of 16.
 156          */
 157         skb_reserve(skb, 12);
 158
 159         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 160                                        DMA_FROM_DEVICE);
 161         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 162                 dev_kfree_skb_any(skb);
 163                 return NULL;
 164         }
 165
 166         for (i = 0; i < frags; i++) {
 167                 struct page *page = alloc_page(GFP_ATOMIC);
 168
 169                 if (!page)
 170                         goto partial_error;
 171                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 172
 173                 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
 174                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 175                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 176                         goto partial_error;
 177         }
 178
 179         rx_ring[id].skb = skb;
 180         return skb;
 181
 182 partial_error:
 183
 184         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 185
 186         for (; i > 0; --i)
 187                 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 188
 189         dev_kfree_skb_any(skb);
 190         return NULL;
 191 }
 192
 193 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 194                                   struct ipoib_cm_rx_buf *rx_ring)
 195 {
 196         struct ipoib_dev_priv *priv = netdev_priv(dev);
 197         int i;
 198
 199         for (i = 0; i < ipoib_recvq_size; ++i)
 200                 if (rx_ring[i].skb) {
 201                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 202                                               rx_ring[i].mapping);
 203                         dev_kfree_skb_any(rx_ring[i].skb);
 204                 }
 205
 206         vfree(rx_ring);
 207 }
 208
 209 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 210 {
 211         struct ib_send_wr *bad_wr;
 212         struct ipoib_cm_rx *p;
 213
 214         /* We only reserved 1 extra slot in CQ for drain WRs, so
 215          * make sure we have at most 1 outstanding WR. */
 216         if (list_empty(&priv->cm.rx_flush_list) ||
 217             !list_empty(&priv->cm.rx_drain_list))
 218                 return;
 219
 220         /*
 221          * QPs on flush list are error state.  This way, a "flush
 222          * error" WC will be immediately generated for each WR we post.
 223          */
 224         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 225         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 226                 ipoib_warn(priv, "failed to post drain wr\n");
 227
 228         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 229 }
 230
 231 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 232 {
 233         struct ipoib_cm_rx *p = ctx;
 234         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 235         unsigned long flags;
 236
 237         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 238                 return;
 239
 240         spin_lock_irqsave(&priv->lock, flags);
 241         list_move(&p->list, &priv->cm.rx_flush_list);
 242         p->state = IPOIB_CM_RX_FLUSH;
 243         ipoib_cm_start_rx_drain(priv);
 244         spin_unlock_irqrestore(&priv->lock, flags);
 245 }
 246
 247 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 248                                            struct ipoib_cm_rx *p)
 249 {
 250         struct ipoib_dev_priv *priv = netdev_priv(dev);
 251         struct ib_qp_init_attr attr = {
 252                 .event_handler = ipoib_cm_rx_event_handler,
 253                 .send_cq = priv->recv_cq, /* For drain WR */
 254                 .recv_cq = priv->recv_cq,
 255                 .srq = priv->cm.srq,
 256                 .cap.max_send_wr = 1, /* For drain WR */
 257                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 258                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 259                 .qp_type = IB_QPT_RC,
 260                 .qp_context = p,
 261         };
 262
 263         if (!ipoib_cm_has_srq(dev)) {
 264                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 265                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 266         }
 267
 268         return ib_create_qp(priv->pd, &attr);
 269 }
 270
 271 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 272                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 273                                  unsigned psn)
 274 {
 275         struct ipoib_dev_priv *priv = netdev_priv(dev);
 276         struct ib_qp_attr qp_attr;
 277         int qp_attr_mask, ret;
 278
 279         qp_attr.qp_state = IB_QPS_INIT;
 280         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 281         if (ret) {
 282                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 283                 return ret;
 284         }
 285         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 286         if (ret) {
 287                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 288                 return ret;
 289         }
 290         qp_attr.qp_state = IB_QPS_RTR;
 291         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 292         if (ret) {
 293                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 294                 return ret;
 295         }
 296         qp_attr.rq_psn = psn;
 297         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 298         if (ret) {
 299                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 300                 return ret;
 301         }
 302
 303         /*
 304          * Current Mellanox HCA firmware won't generate completions
 305          * with error for drain WRs unless the QP has been moved to
 306          * RTS first. This work-around leaves a window where a QP has
 307          * moved to error asynchronously, but this will eventually get
 308          * fixed in firmware, so let's not error out if modify QP
 309          * fails.
 310          */
 311         qp_attr.qp_state = IB_QPS_RTS;
 312         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 313         if (ret) {
 314                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 315                 return 0;
 316         }
 317         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 318         if (ret) {
 319                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 320                 return 0;
 321         }
 322
 323         return 0;
 324 }
 325
 326 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 327                                 struct ib_recv_wr *wr,
 328                                 struct ib_sge *sge)
 329 {
 330         struct ipoib_dev_priv *priv = netdev_priv(dev);
 331         int i;
 332
 333         for (i = 0; i < priv->cm.num_frags; ++i)
 334                 sge[i].lkey = priv->mr->lkey;
 335
 336         sge[0].length = IPOIB_CM_HEAD_SIZE;
 337         for (i = 1; i < priv->cm.num_frags; ++i)
 338                 sge[i].length = PAGE_SIZE;
 339
 340         wr->next    = NULL;
 341         wr->sg_list = sge;
 342         wr->num_sge = priv->cm.num_frags;
 343 }
 344
 345 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 346                                    struct ipoib_cm_rx *rx)
 347 {
 348         struct ipoib_dev_priv *priv = netdev_priv(dev);
 349         struct {
 350                 struct ib_recv_wr wr;
 351                 struct ib_sge sge[IPOIB_CM_RX_SG];
 352         } *t;
 353         int ret;
 354         int i;
 355
 356         rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 357         if (!rx->rx_ring) {
 358                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 359                        priv->ca->name, ipoib_recvq_size);
 360                 return -ENOMEM;
 361         }
 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, skb_frag_page(frag),
 542                                            0, PAGE_SIZE);
 543                         --skb_shinfo(skb)->nr_frags;
 544                 } else {
 545                         size = min(length, (unsigned) PAGE_SIZE);
 546
 547                         skb_frag_size_set(frag, size);
 548                         skb->data_len += size;
 549                         skb->truesize += size;
 550                         skb->len += size;
 551                         length -= size;
 552                 }
 553         }
 554 }
 555
 556 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 557 {
 558         struct ipoib_dev_priv *priv = netdev_priv(dev);
 559         struct ipoib_cm_rx_buf *rx_ring;
 560         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 561         struct sk_buff *skb, *newskb;
 562         struct ipoib_cm_rx *p;
 563         unsigned long flags;
 564         u64 mapping[IPOIB_CM_RX_SG];
 565         int frags;
 566         int has_srq;
 567         struct sk_buff *small_skb;
 568
 569         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 570                        wr_id, wc->status);
 571
 572         if (unlikely(wr_id >= ipoib_recvq_size)) {
 573                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 574                         spin_lock_irqsave(&priv->lock, flags);
 575                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 576                         ipoib_cm_start_rx_drain(priv);
 577                         queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 578                         spin_unlock_irqrestore(&priv->lock, flags);
 579                 } else
 580                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 581                                    wr_id, ipoib_recvq_size);
 582                 return;
 583         }
 584
 585         p = wc->qp->qp_context;
 586
 587         has_srq = ipoib_cm_has_srq(dev);
 588         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 589
 590         skb = rx_ring[wr_id].skb;
 591
 592         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 593                 ipoib_dbg(priv, "cm recv error "
 594                            "(status=%d, wrid=%d vend_err %x)\n",
 595                            wc->status, wr_id, wc->vendor_err);
 596                 ++dev->stats.rx_dropped;
 597                 if (has_srq)
 598                         goto repost;
 599                 else {
 600                         if (!--p->recv_count) {
 601                                 spin_lock_irqsave(&priv->lock, flags);
 602                                 list_move(&p->list, &priv->cm.rx_reap_list);
 603                                 spin_unlock_irqrestore(&priv->lock, flags);
 604                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 605                         }
 606                         return;
 607                 }
 608         }
 609
 610         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 611                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 612                         spin_lock_irqsave(&priv->lock, flags);
 613                         p->jiffies = jiffies;
 614                         /* Move this entry to list head, but do not re-add it
 615                          * if it has been moved out of list. */
 616                         if (p->state == IPOIB_CM_RX_LIVE)
 617                                 list_move(&p->list, &priv->cm.passive_ids);
 618                         spin_unlock_irqrestore(&priv->lock, flags);
 619                 }
 620         }
 621
 622         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 623                 int dlen = wc->byte_len;
 624
 625                 small_skb = dev_alloc_skb(dlen + 12);
 626                 if (small_skb) {
 627                         skb_reserve(small_skb, 12);
 628                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 629                                                    dlen, DMA_FROM_DEVICE);
 630                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 631                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 632                                                       dlen, DMA_FROM_DEVICE);
 633                         skb_put(small_skb, dlen);
 634                         skb = small_skb;
 635                         goto copied;
 636                 }
 637         }
 638
 639         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 640                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 641
 642         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags, mapping);
 643         if (unlikely(!newskb)) {
 644                 /*
 645                  * If we can't allocate a new RX buffer, dump
 646                  * this packet and reuse the old buffer.
 647                  */
 648                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 649                 ++dev->stats.rx_dropped;
 650                 goto repost;
 651         }
 652
 653         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 654         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 655
 656         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 657                        wc->byte_len, wc->slid);
 658
 659         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 660
 661 copied:
 662         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 663         skb_reset_mac_header(skb);
 664         skb_pull(skb, IPOIB_ENCAP_LEN);
 665
 666         ++dev->stats.rx_packets;
 667         dev->stats.rx_bytes += skb->len;
 668
 669         skb->dev = dev;
 670         /* XXX get correct PACKET_ type here */
 671         skb->pkt_type = PACKET_HOST;
 672         netif_receive_skb(skb);
 673
 674 repost:
 675         if (has_srq) {
 676                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 677                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 678                                    "for buf %d\n", wr_id);
 679         } else {
 680                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 681                                                           &priv->cm.rx_wr,
 682                                                           priv->cm.rx_sge,
 683                                                           wr_id))) {
 684                         --p->recv_count;
 685                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 686                                    "for buf %d\n", wr_id);
 687                 }
 688         }
 689 }
 690
 691 static inline int post_send(struct ipoib_dev_priv *priv,
 692                             struct ipoib_cm_tx *tx,
 693                             unsigned int wr_id,
 694                             u64 addr, int len)
 695 {
 696         struct ib_send_wr *bad_wr;
 697
 698         priv->tx_sge[0].addr          = addr;
 699         priv->tx_sge[0].length        = len;
 700
 701         priv->tx_wr.num_sge     = 1;
 702         priv->tx_wr.wr_id       = wr_id | IPOIB_OP_CM;
 703
 704         return ib_post_send(tx->qp, &priv->tx_wr, &bad_wr);
 705 }
 706
 707 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 708 {
 709         struct ipoib_dev_priv *priv = netdev_priv(dev);
 710         struct ipoib_cm_tx_buf *tx_req;
 711         u64 addr;
 712         int rc;
 713
 714         if (unlikely(skb->len > tx->mtu)) {
 715                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 716                            skb->len, tx->mtu);
 717                 ++dev->stats.tx_dropped;
 718                 ++dev->stats.tx_errors;
 719                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 720                 return;
 721         }
 722
 723         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 724                        tx->tx_head, skb->len, tx->qp->qp_num);
 725
 726         /*
 727          * We put the skb into the tx_ring _before_ we call post_send()
 728          * because it's entirely possible that the completion handler will
 729          * run before we execute anything after the post_send().  That
 730          * means we have to make sure everything is properly recorded and
 731          * our state is consistent before we call post_send().
 732          */
 733         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 734         tx_req->skb = skb;
 735         addr = ib_dma_map_single(priv->ca, skb->data, skb->len, DMA_TO_DEVICE);
 736         if (unlikely(ib_dma_mapping_error(priv->ca, addr))) {
 737                 ++dev->stats.tx_errors;
 738                 dev_kfree_skb_any(skb);
 739                 return;
 740         }
 741
 742         tx_req->mapping = addr;
 743
 744         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1),
 745                        addr, skb->len);
 746         if (unlikely(rc)) {
 747                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 748                 ++dev->stats.tx_errors;
 749                 ib_dma_unmap_single(priv->ca, addr, skb->len, DMA_TO_DEVICE);
 750                 dev_kfree_skb_any(skb);
 751         } else {
 752                 dev->trans_start = jiffies;
 753                 ++tx->tx_head;
 754
 755                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 756                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 757                                   tx->qp->qp_num);
 758                         netif_stop_queue(dev);
 759                         rc = ib_req_notify_cq(priv->send_cq,
 760                                 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
 761                         if (rc < 0)
 762                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 763                         else if (rc)
 764                                 ipoib_send_comp_handler(priv->send_cq, dev);
 765                 }
 766         }
 767 }
 768
 769 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 770 {
 771         struct ipoib_dev_priv *priv = netdev_priv(dev);
 772         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 773         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 774         struct ipoib_cm_tx_buf *tx_req;
 775         unsigned long flags;
 776
 777         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 778                        wr_id, wc->status);
 779
 780         if (unlikely(wr_id >= ipoib_sendq_size)) {
 781                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 782                            wr_id, ipoib_sendq_size);
 783                 return;
 784         }
 785
 786         tx_req = &tx->tx_ring[wr_id];
 787
 788         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 789
 790         /* FIXME: is this right? Shouldn't we only increment on success? */
 791         ++dev->stats.tx_packets;
 792         dev->stats.tx_bytes += tx_req->skb->len;
 793
 794         dev_kfree_skb_any(tx_req->skb);
 795
 796         netif_tx_lock(dev);
 797
 798         ++tx->tx_tail;
 799         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 800             netif_queue_stopped(dev) &&
 801             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 802                 netif_wake_queue(dev);
 803
 804         if (wc->status != IB_WC_SUCCESS &&
 805             wc->status != IB_WC_WR_FLUSH_ERR) {
 806                 struct ipoib_neigh *neigh;
 807
 808                 ipoib_dbg(priv, "failed cm send event "
 809                            "(status=%d, wrid=%d vend_err %x)\n",
 810                            wc->status, wr_id, wc->vendor_err);
 811
 812                 spin_lock_irqsave(&priv->lock, flags);
 813                 neigh = tx->neigh;
 814
 815                 if (neigh) {
 816                         neigh->cm = NULL;
 817                         list_del(&neigh->list);
 818                         if (neigh->ah)
 819                                 ipoib_put_ah(neigh->ah);
 820                         ipoib_neigh_free(dev, neigh);
 821
 822                         tx->neigh = NULL;
 823                 }
 824
 825                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 826                         list_move(&tx->list, &priv->cm.reap_list);
 827                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 828                 }
 829
 830                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 831
 832                 spin_unlock_irqrestore(&priv->lock, flags);
 833         }
 834
 835         netif_tx_unlock(dev);
 836 }
 837
 838 int ipoib_cm_dev_open(struct net_device *dev)
 839 {
 840         struct ipoib_dev_priv *priv = netdev_priv(dev);
 841         int ret;
 842
 843         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 844                 return 0;
 845
 846         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 847         if (IS_ERR(priv->cm.id)) {
 848                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 849                 ret = PTR_ERR(priv->cm.id);
 850                 goto err_cm;
 851         }
 852
 853         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 854                            0, NULL);
 855         if (ret) {
 856                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 857                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 858                 goto err_listen;
 859         }
 860
 861         return 0;
 862
 863 err_listen:
 864         ib_destroy_cm_id(priv->cm.id);
 865 err_cm:
 866         priv->cm.id = NULL;
 867         return ret;
 868 }
 869
 870 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 871 {
 872         struct ipoib_dev_priv *priv = netdev_priv(dev);
 873         struct ipoib_cm_rx *rx, *n;
 874         LIST_HEAD(list);
 875
 876         spin_lock_irq(&priv->lock);
 877         list_splice_init(&priv->cm.rx_reap_list, &list);
 878         spin_unlock_irq(&priv->lock);
 879
 880         list_for_each_entry_safe(rx, n, &list, list) {
 881                 ib_destroy_cm_id(rx->id);
 882                 ib_destroy_qp(rx->qp);
 883                 if (!ipoib_cm_has_srq(dev)) {
 884                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 885                         spin_lock_irq(&priv->lock);
 886                         --priv->cm.nonsrq_conn_qp;
 887                         spin_unlock_irq(&priv->lock);
 888                 }
 889                 kfree(rx);
 890         }
 891 }
 892
 893 void ipoib_cm_dev_stop(struct net_device *dev)
 894 {
 895         struct ipoib_dev_priv *priv = netdev_priv(dev);
 896         struct ipoib_cm_rx *p;
 897         unsigned long begin;
 898         int ret;
 899
 900         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 901                 return;
 902
 903         ib_destroy_cm_id(priv->cm.id);
 904         priv->cm.id = NULL;
 905
 906         spin_lock_irq(&priv->lock);
 907         while (!list_empty(&priv->cm.passive_ids)) {
 908                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 909                 list_move(&p->list, &priv->cm.rx_error_list);
 910                 p->state = IPOIB_CM_RX_ERROR;
 911                 spin_unlock_irq(&priv->lock);
 912                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 913                 if (ret)
 914                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 915                 spin_lock_irq(&priv->lock);
 916         }
 917
 918         /* Wait for all RX to be drained */
 919         begin = jiffies;
 920
 921         while (!list_empty(&priv->cm.rx_error_list) ||
 922                !list_empty(&priv->cm.rx_flush_list) ||
 923                !list_empty(&priv->cm.rx_drain_list)) {
 924                 if (time_after(jiffies, begin + 5 * HZ)) {
 925                         ipoib_warn(priv, "RX drain timing out\n");
 926
 927                         /*
 928                          * assume the HW is wedged and just free up everything.
 929                          */
 930                         list_splice_init(&priv->cm.rx_flush_list,
 931                                          &priv->cm.rx_reap_list);
 932                         list_splice_init(&priv->cm.rx_error_list,
 933                                          &priv->cm.rx_reap_list);
 934                         list_splice_init(&priv->cm.rx_drain_list,
 935                                          &priv->cm.rx_reap_list);
 936                         break;
 937                 }
 938                 spin_unlock_irq(&priv->lock);
 939                 msleep(1);
 940                 ipoib_drain_cq(dev);
 941                 spin_lock_irq(&priv->lock);
 942         }
 943
 944         spin_unlock_irq(&priv->lock);
 945
 946         ipoib_cm_free_rx_reap_list(dev);
 947
 948         cancel_delayed_work(&priv->cm.stale_task);
 949 }
 950
 951 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 952 {
 953         struct ipoib_cm_tx *p = cm_id->context;
 954         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 955         struct ipoib_cm_data *data = event->private_data;
 956         struct sk_buff_head skqueue;
 957         struct ib_qp_attr qp_attr;
 958         int qp_attr_mask, ret;
 959         struct sk_buff *skb;
 960
 961         p->mtu = be32_to_cpu(data->mtu);
 962
 963         if (p->mtu <= IPOIB_ENCAP_LEN) {
 964                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 965                            p->mtu, IPOIB_ENCAP_LEN);
 966                 return -EINVAL;
 967         }
 968
 969         qp_attr.qp_state = IB_QPS_RTR;
 970         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 971         if (ret) {
 972                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 973                 return ret;
 974         }
 975
 976         qp_attr.rq_psn = 0 /* FIXME */;
 977         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 978         if (ret) {
 979                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 980                 return ret;
 981         }
 982
 983         qp_attr.qp_state = IB_QPS_RTS;
 984         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 985         if (ret) {
 986                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 987                 return ret;
 988         }
 989         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 990         if (ret) {
 991                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 992                 return ret;
 993         }
 994
 995         skb_queue_head_init(&skqueue);
 996
 997         spin_lock_irq(&priv->lock);
 998         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 999         if (p->neigh)
1000                 while ((skb = __skb_dequeue(&p->neigh->queue)))
1001                         __skb_queue_tail(&skqueue, skb);
1002         spin_unlock_irq(&priv->lock);
1003
1004         while ((skb = __skb_dequeue(&skqueue))) {
1005                 skb->dev = p->dev;
1006                 if (dev_queue_xmit(skb))
1007                         ipoib_warn(priv, "dev_queue_xmit failed "
1008                                    "to requeue packet\n");
1009         }
1010
1011         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1012         if (ret) {
1013                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1014                 return ret;
1015         }
1016         return 0;
1017 }
1018
1019 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1020 {
1021         struct ipoib_dev_priv *priv = netdev_priv(dev);
1022         struct ib_qp_init_attr attr = {
1023                 .send_cq                = priv->recv_cq,
1024                 .recv_cq                = priv->recv_cq,
1025                 .srq                    = priv->cm.srq,
1026                 .cap.max_send_wr        = ipoib_sendq_size,
1027                 .cap.max_send_sge       = 1,
1028                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1029                 .qp_type                = IB_QPT_RC,
1030                 .qp_context             = tx
1031         };
1032
1033         return ib_create_qp(priv->pd, &attr);
1034 }
1035
1036 static int ipoib_cm_send_req(struct net_device *dev,
1037                              struct ib_cm_id *id, struct ib_qp *qp,
1038                              u32 qpn,
1039                              struct ib_sa_path_rec *pathrec)
1040 {
1041         struct ipoib_dev_priv *priv = netdev_priv(dev);
1042         struct ipoib_cm_data data = {};
1043         struct ib_cm_req_param req = {};
1044
1045         data.qpn = cpu_to_be32(priv->qp->qp_num);
1046         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1047
1048         req.primary_path                = pathrec;
1049         req.alternate_path              = NULL;
1050         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1051         req.qp_num                      = qp->qp_num;
1052         req.qp_type                     = qp->qp_type;
1053         req.private_data                = &data;
1054         req.private_data_len            = sizeof data;
1055         req.flow_control                = 0;
1056
1057         req.starting_psn                = 0; /* FIXME */
1058
1059         /*
1060          * Pick some arbitrary defaults here; we could make these
1061          * module parameters if anyone cared about setting them.
1062          */
1063         req.responder_resources         = 4;
1064         req.remote_cm_response_timeout  = 20;
1065         req.local_cm_response_timeout   = 20;
1066         req.retry_count                 = 0; /* RFC draft warns against retries */
1067         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1068         req.max_cm_retries              = 15;
1069         req.srq                         = ipoib_cm_has_srq(dev);
1070         return ib_send_cm_req(id, &req);
1071 }
1072
1073 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1074                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1075 {
1076         struct ipoib_dev_priv *priv = netdev_priv(dev);
1077         struct ib_qp_attr qp_attr;
1078         int qp_attr_mask, ret;
1079         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1080         if (ret) {
1081                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1082                 return ret;
1083         }
1084
1085         qp_attr.qp_state = IB_QPS_INIT;
1086         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1087         qp_attr.port_num = priv->port;
1088         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1089
1090         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1091         if (ret) {
1092                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1093                 return ret;
1094         }
1095         return 0;
1096 }
1097
1098 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1099                             struct ib_sa_path_rec *pathrec)
1100 {
1101         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1102         int ret;
1103
1104         p->tx_ring = vzalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1105         if (!p->tx_ring) {
1106                 ipoib_warn(priv, "failed to allocate tx ring\n");
1107                 ret = -ENOMEM;
1108                 goto err_tx;
1109         }
1110
1111         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1112         if (IS_ERR(p->qp)) {
1113                 ret = PTR_ERR(p->qp);
1114                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1115                 goto err_qp;
1116         }
1117
1118         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1119         if (IS_ERR(p->id)) {
1120                 ret = PTR_ERR(p->id);
1121                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1122                 goto err_id;
1123         }
1124
1125         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1126         if (ret) {
1127                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1128                 goto err_modify;
1129         }
1130
1131         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1132         if (ret) {
1133                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1134                 goto err_send_cm;
1135         }
1136
1137         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1138                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1139
1140         return 0;
1141
1142 err_send_cm:
1143 err_modify:
1144         ib_destroy_cm_id(p->id);
1145 err_id:
1146         p->id = NULL;
1147         ib_destroy_qp(p->qp);
1148 err_qp:
1149         p->qp = NULL;
1150         vfree(p->tx_ring);
1151 err_tx:
1152         return ret;
1153 }
1154
1155 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1156 {
1157         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1158         struct ipoib_cm_tx_buf *tx_req;
1159         unsigned long begin;
1160
1161         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1162                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1163
1164         if (p->id)
1165                 ib_destroy_cm_id(p->id);
1166
1167         if (p->tx_ring) {
1168                 /* Wait for all sends to complete */
1169                 begin = jiffies;
1170                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1171                         if (time_after(jiffies, begin + 5 * HZ)) {
1172                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1173                                            p->tx_head - p->tx_tail);
1174                                 goto timeout;
1175                         }
1176
1177                         msleep(1);
1178                 }
1179         }
1180
1181 timeout:
1182
1183         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1184                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1185                 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1186                                     DMA_TO_DEVICE);
1187                 dev_kfree_skb_any(tx_req->skb);
1188                 ++p->tx_tail;
1189                 netif_tx_lock_bh(p->dev);
1190                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1191                     netif_queue_stopped(p->dev) &&
1192                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1193                         netif_wake_queue(p->dev);
1194                 netif_tx_unlock_bh(p->dev);
1195         }
1196
1197         if (p->qp)
1198                 ib_destroy_qp(p->qp);
1199
1200         vfree(p->tx_ring);
1201         kfree(p);
1202 }
1203
1204 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1205                                struct ib_cm_event *event)
1206 {
1207         struct ipoib_cm_tx *tx = cm_id->context;
1208         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1209         struct net_device *dev = priv->dev;
1210         struct ipoib_neigh *neigh;
1211         unsigned long flags;
1212         int ret;
1213
1214         switch (event->event) {
1215         case IB_CM_DREQ_RECEIVED:
1216                 ipoib_dbg(priv, "DREQ received.\n");
1217                 ib_send_cm_drep(cm_id, NULL, 0);
1218                 break;
1219         case IB_CM_REP_RECEIVED:
1220                 ipoib_dbg(priv, "REP received.\n");
1221                 ret = ipoib_cm_rep_handler(cm_id, event);
1222                 if (ret)
1223                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1224                                        NULL, 0, NULL, 0);
1225                 break;
1226         case IB_CM_REQ_ERROR:
1227         case IB_CM_REJ_RECEIVED:
1228         case IB_CM_TIMEWAIT_EXIT:
1229                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1230                 netif_tx_lock_bh(dev);
1231                 spin_lock_irqsave(&priv->lock, flags);
1232                 neigh = tx->neigh;
1233
1234                 if (neigh) {
1235                         neigh->cm = NULL;
1236                         list_del(&neigh->list);
1237                         if (neigh->ah)
1238                                 ipoib_put_ah(neigh->ah);
1239                         ipoib_neigh_free(dev, neigh);
1240
1241                         tx->neigh = NULL;
1242                 }
1243
1244                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1245                         list_move(&tx->list, &priv->cm.reap_list);
1246                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1247                 }
1248
1249                 spin_unlock_irqrestore(&priv->lock, flags);
1250                 netif_tx_unlock_bh(dev);
1251                 break;
1252         default:
1253                 break;
1254         }
1255
1256         return 0;
1257 }
1258
1259 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1260                                        struct ipoib_neigh *neigh)
1261 {
1262         struct ipoib_dev_priv *priv = netdev_priv(dev);
1263         struct ipoib_cm_tx *tx;
1264
1265         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1266         if (!tx)
1267                 return NULL;
1268
1269         neigh->cm = tx;
1270         tx->neigh = neigh;
1271         tx->path = path;
1272         tx->dev = dev;
1273         list_add(&tx->list, &priv->cm.start_list);
1274         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1275         queue_work(ipoib_workqueue, &priv->cm.start_task);
1276         return tx;
1277 }
1278
1279 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1280 {
1281         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1282         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1283                 list_move(&tx->list, &priv->cm.reap_list);
1284                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1285                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1286                           tx->neigh->dgid.raw);
1287                 tx->neigh = NULL;
1288         }
1289 }
1290
1291 #define QPN_AND_OPTIONS_OFFSET  4
1292
1293 static void ipoib_cm_tx_start(struct work_struct *work)
1294 {
1295         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1296                                                    cm.start_task);
1297         struct net_device *dev = priv->dev;
1298         struct ipoib_neigh *neigh;
1299         struct ipoib_cm_tx *p;
1300         unsigned long flags;
1301         struct ipoib_path *path;
1302         int ret;
1303
1304         struct ib_sa_path_rec pathrec;
1305         u32 qpn;
1306
1307         netif_tx_lock_bh(dev);
1308         spin_lock_irqsave(&priv->lock, flags);
1309
1310         while (!list_empty(&priv->cm.start_list)) {
1311                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1312                 list_del_init(&p->list);
1313                 neigh = p->neigh;
1314
1315                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1316                 /*
1317                  * As long as the search is with these 2 locks,
1318                  * path existence indicates its validity.
1319                  */
1320                 path = __path_find(dev, neigh->neighbour->ha + QPN_AND_OPTIONS_OFFSET);
1321                 if (!path) {
1322                         pr_info("%s ignore not valid path %pI6\n",
1323                                 __func__,
1324                                 neigh->neighbour->ha + QPN_AND_OPTIONS_OFFSET);
1325                         goto free_neigh;
1326                 }
1327                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1328
1329                 spin_unlock_irqrestore(&priv->lock, flags);
1330                 netif_tx_unlock_bh(dev);
1331
1332                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1333
1334                 netif_tx_lock_bh(dev);
1335                 spin_lock_irqsave(&priv->lock, flags);
1336
1337                 if (ret) {
1338 free_neigh:
1339                         neigh = p->neigh;
1340                         if (neigh) {
1341                                 neigh->cm = NULL;
1342                                 list_del(&neigh->list);
1343                                 if (neigh->ah)
1344                                         ipoib_put_ah(neigh->ah);
1345                                 ipoib_neigh_free(dev, neigh);
1346                         }
1347                         list_del(&p->list);
1348                         kfree(p);
1349                 }
1350         }
1351
1352         spin_unlock_irqrestore(&priv->lock, flags);
1353         netif_tx_unlock_bh(dev);
1354 }
1355
1356 static void ipoib_cm_tx_reap(struct work_struct *work)
1357 {
1358         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1359                                                    cm.reap_task);
1360         struct net_device *dev = priv->dev;
1361         struct ipoib_cm_tx *p;
1362         unsigned long flags;
1363
1364         netif_tx_lock_bh(dev);
1365         spin_lock_irqsave(&priv->lock, flags);
1366
1367         while (!list_empty(&priv->cm.reap_list)) {
1368                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1369                 list_del(&p->list);
1370                 spin_unlock_irqrestore(&priv->lock, flags);
1371                 netif_tx_unlock_bh(dev);
1372                 ipoib_cm_tx_destroy(p);
1373                 netif_tx_lock_bh(dev);
1374                 spin_lock_irqsave(&priv->lock, flags);
1375         }
1376
1377         spin_unlock_irqrestore(&priv->lock, flags);
1378         netif_tx_unlock_bh(dev);
1379 }
1380
1381 static void ipoib_cm_skb_reap(struct work_struct *work)
1382 {
1383         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1384                                                    cm.skb_task);
1385         struct net_device *dev = priv->dev;
1386         struct sk_buff *skb;
1387         unsigned long flags;
1388         unsigned mtu = priv->mcast_mtu;
1389
1390         netif_tx_lock_bh(dev);
1391         spin_lock_irqsave(&priv->lock, flags);
1392
1393         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1394                 spin_unlock_irqrestore(&priv->lock, flags);
1395                 netif_tx_unlock_bh(dev);
1396
1397                 if (skb->protocol == htons(ETH_P_IP))
1398                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1399 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1400                 else if (skb->protocol == htons(ETH_P_IPV6))
1401                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1402 #endif
1403                 dev_kfree_skb_any(skb);
1404
1405                 netif_tx_lock_bh(dev);
1406                 spin_lock_irqsave(&priv->lock, flags);
1407         }
1408
1409         spin_unlock_irqrestore(&priv->lock, flags);
1410         netif_tx_unlock_bh(dev);
1411 }
1412
1413 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1414                            unsigned int mtu)
1415 {
1416         struct ipoib_dev_priv *priv = netdev_priv(dev);
1417         int e = skb_queue_empty(&priv->cm.skb_queue);
1418
1419         if (skb_dst(skb))
1420                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1421
1422         skb_queue_tail(&priv->cm.skb_queue, skb);
1423         if (e)
1424                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1425 }
1426
1427 static void ipoib_cm_rx_reap(struct work_struct *work)
1428 {
1429         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1430                                                 cm.rx_reap_task)->dev);
1431 }
1432
1433 static void ipoib_cm_stale_task(struct work_struct *work)
1434 {
1435         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1436                                                    cm.stale_task.work);
1437         struct ipoib_cm_rx *p;
1438         int ret;
1439
1440         spin_lock_irq(&priv->lock);
1441         while (!list_empty(&priv->cm.passive_ids)) {
1442                 /* List is sorted by LRU, start from tail,
1443                  * stop when we see a recently used entry */
1444                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1445                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1446                         break;
1447                 list_move(&p->list, &priv->cm.rx_error_list);
1448                 p->state = IPOIB_CM_RX_ERROR;
1449                 spin_unlock_irq(&priv->lock);
1450                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1451                 if (ret)
1452                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1453                 spin_lock_irq(&priv->lock);
1454         }
1455
1456         if (!list_empty(&priv->cm.passive_ids))
1457                 queue_delayed_work(ipoib_workqueue,
1458                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1459         spin_unlock_irq(&priv->lock);
1460 }
1461
1462
1463 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1464                          char *buf)
1465 {
1466         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1467
1468         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1469                 return sprintf(buf, "connected\n");
1470         else
1471                 return sprintf(buf, "datagram\n");
1472 }
1473
1474 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1475                         const char *buf, size_t count)
1476 {
1477         struct net_device *dev = to_net_dev(d);
1478         struct ipoib_dev_priv *priv = netdev_priv(dev);
1479
1480         if (!rtnl_trylock())
1481                 return restart_syscall();
1482
1483         /* flush paths if we switch modes so that connections are restarted */
1484         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1485                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1486                 ipoib_warn(priv, "enabling connected mode "
1487                            "will cause multicast packet drops\n");
1488                 netdev_update_features(dev);
1489                 rtnl_unlock();
1490                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1491
1492                 ipoib_flush_paths(dev);
1493                 return count;
1494         }
1495
1496         if (!strcmp(buf, "datagram\n")) {
1497                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1498                 netdev_update_features(dev);
1499                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1500                 rtnl_unlock();
1501                 ipoib_flush_paths(dev);
1502
1503                 return count;
1504         }
1505         rtnl_unlock();
1506
1507         return -EINVAL;
1508 }
1509
1510 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1511
1512 int ipoib_cm_add_mode_attr(struct net_device *dev)
1513 {
1514         return device_create_file(&dev->dev, &dev_attr_mode);
1515 }
1516
1517 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1518 {
1519         struct ipoib_dev_priv *priv = netdev_priv(dev);
1520         struct ib_srq_init_attr srq_init_attr = {
1521                 .srq_type = IB_SRQT_BASIC,
1522                 .attr = {
1523                         .max_wr  = ipoib_recvq_size,
1524                         .max_sge = max_sge
1525                 }
1526         };
1527
1528         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1529         if (IS_ERR(priv->cm.srq)) {
1530                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1531                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1532                                priv->ca->name, PTR_ERR(priv->cm.srq));
1533                 priv->cm.srq = NULL;
1534                 return;
1535         }
1536
1537         priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1538         if (!priv->cm.srq_ring) {
1539                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1540                        priv->ca->name, ipoib_recvq_size);
1541                 ib_destroy_srq(priv->cm.srq);
1542                 priv->cm.srq = NULL;
1543                 return;
1544         }
1545
1546 }
1547
1548 int ipoib_cm_dev_init(struct net_device *dev)
1549 {
1550         struct ipoib_dev_priv *priv = netdev_priv(dev);
1551         int i, ret;
1552         struct ib_device_attr attr;
1553
1554         INIT_LIST_HEAD(&priv->cm.passive_ids);
1555         INIT_LIST_HEAD(&priv->cm.reap_list);
1556         INIT_LIST_HEAD(&priv->cm.start_list);
1557         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1558         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1559         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1560         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1561         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1562         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1563         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1564         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1565         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1566
1567         skb_queue_head_init(&priv->cm.skb_queue);
1568
1569         ret = ib_query_device(priv->ca, &attr);
1570         if (ret) {
1571                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1572                 return ret;
1573         }
1574
1575         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1576
1577         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1578         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1579         if (ipoib_cm_has_srq(dev)) {
1580                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1581                 priv->cm.num_frags  = attr.max_srq_sge;
1582                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1583                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1584         } else {
1585                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1586                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1587         }
1588
1589         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1590
1591         if (ipoib_cm_has_srq(dev)) {
1592                 for (i = 0; i < ipoib_recvq_size; ++i) {
1593                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1594                                                    priv->cm.num_frags - 1,
1595                                                    priv->cm.srq_ring[i].mapping)) {
1596                                 ipoib_warn(priv, "failed to allocate "
1597                                            "receive buffer %d\n", i);
1598                                 ipoib_cm_dev_cleanup(dev);
1599                                 return -ENOMEM;
1600                         }
1601
1602                         if (ipoib_cm_post_receive_srq(dev, i)) {
1603                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1604                                            "failed for buf %d\n", i);
1605                                 ipoib_cm_dev_cleanup(dev);
1606                                 return -EIO;
1607                         }
1608                 }
1609         }
1610
1611         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1612         return 0;
1613 }
1614
1615 void ipoib_cm_dev_cleanup(struct net_device *dev)
1616 {
1617         struct ipoib_dev_priv *priv = netdev_priv(dev);
1618         int ret;
1619
1620         if (!priv->cm.srq)
1621                 return;
1622
1623         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1624
1625         ret = ib_destroy_srq(priv->cm.srq);
1626         if (ret)
1627                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1628
1629         priv->cm.srq = NULL;
1630         if (!priv->cm.srq_ring)
1631                 return;
1632
1633         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1634         priv->cm.srq_ring = NULL;
1635 }