drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <rdma/ib_cm.h>
  34 #include <net/dst.h>
  35 #include <net/icmp.h>
  36 #include <linux/icmpv6.h>
  37 #include <linux/delay.h>
  38 #include <linux/slab.h>
  39 #include <linux/vmalloc.h>
  40
  41 #include "ipoib.h"
  42
  43 int ipoib_max_conn_qp = 128;
  44
  45 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  46 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  47                  "Max number of connected-mode QPs per interface "
  48                  "(applied only if shared receive queue is not available)");
  49
  50 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  51 static int data_debug_level;
  52
  53 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  54 MODULE_PARM_DESC(cm_data_debug_level,
  55                  "Enable data path debug tracing for connected mode if > 0");
  56 #endif
  57
  58 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  59
  60 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  61 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  62 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  63 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  64
  65 static struct ib_qp_attr ipoib_cm_err_attr = {
  66         .qp_state = IB_QPS_ERR
  67 };
  68
  69 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  70
  71 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  72         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  73         .opcode = IB_WR_SEND,
  74 };
  75
  76 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  77                                struct ib_cm_event *event);
  78
  79 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  80                                   u64 mapping[IPOIB_CM_RX_SG])
  81 {
  82         int i;
  83
  84         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  85
  86         for (i = 0; i < frags; ++i)
  87                 ib_dma_unmap_single(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  88 }
  89
  90 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  91 {
  92         struct ipoib_dev_priv *priv = netdev_priv(dev);
  93         struct ib_recv_wr *bad_wr;
  94         int i, ret;
  95
  96         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  97
  98         for (i = 0; i < priv->cm.num_frags; ++i)
  99                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 100
 101         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 102         if (unlikely(ret)) {
 103                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 104                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 105                                       priv->cm.srq_ring[id].mapping);
 106                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 107                 priv->cm.srq_ring[id].skb = NULL;
 108         }
 109
 110         return ret;
 111 }
 112
 113 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 114                                         struct ipoib_cm_rx *rx,
 115                                         struct ib_recv_wr *wr,
 116                                         struct ib_sge *sge, int id)
 117 {
 118         struct ipoib_dev_priv *priv = netdev_priv(dev);
 119         struct ib_recv_wr *bad_wr;
 120         int i, ret;
 121
 122         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 123
 124         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 125                 sge[i].addr = rx->rx_ring[id].mapping[i];
 126
 127         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 128         if (unlikely(ret)) {
 129                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 130                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 131                                       rx->rx_ring[id].mapping);
 132                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 133                 rx->rx_ring[id].skb = NULL;
 134         }
 135
 136         return ret;
 137 }
 138
 139 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 140                                              struct ipoib_cm_rx_buf *rx_ring,
 141                                              int id, int frags,
 142                                              u64 mapping[IPOIB_CM_RX_SG])
 143 {
 144         struct ipoib_dev_priv *priv = netdev_priv(dev);
 145         struct sk_buff *skb;
 146         int i;
 147
 148         skb = dev_alloc_skb(IPOIB_CM_HEAD_SIZE + 12);
 149         if (unlikely(!skb))
 150                 return NULL;
 151
 152         /*
 153          * IPoIB adds a 4 byte header. So we need 12 more bytes to align the
 154          * IP header to a multiple of 16.
 155          */
 156         skb_reserve(skb, 12);
 157
 158         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 159                                        DMA_FROM_DEVICE);
 160         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 161                 dev_kfree_skb_any(skb);
 162                 return NULL;
 163         }
 164
 165         for (i = 0; i < frags; i++) {
 166                 struct page *page = alloc_page(GFP_ATOMIC);
 167
 168                 if (!page)
 169                         goto partial_error;
 170                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 171
 172                 mapping[i + 1] = ib_dma_map_page(priv->ca, skb_shinfo(skb)->frags[i].page,
 173                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 174                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 175                         goto partial_error;
 176         }
 177
 178         rx_ring[id].skb = skb;
 179         return skb;
 180
 181 partial_error:
 182
 183         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 184
 185         for (; i > 0; --i)
 186                 ib_dma_unmap_single(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 187
 188         dev_kfree_skb_any(skb);
 189         return NULL;
 190 }
 191
 192 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 193                                   struct ipoib_cm_rx_buf *rx_ring)
 194 {
 195         struct ipoib_dev_priv *priv = netdev_priv(dev);
 196         int i;
 197
 198         for (i = 0; i < ipoib_recvq_size; ++i)
 199                 if (rx_ring[i].skb) {
 200                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 201                                               rx_ring[i].mapping);
 202                         dev_kfree_skb_any(rx_ring[i].skb);
 203                 }
 204
 205         vfree(rx_ring);
 206 }
 207
 208 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 209 {
 210         struct ib_send_wr *bad_wr;
 211         struct ipoib_cm_rx *p;
 212
 213         /* We only reserved 1 extra slot in CQ for drain WRs, so
 214          * make sure we have at most 1 outstanding WR. */
 215         if (list_empty(&priv->cm.rx_flush_list) ||
 216             !list_empty(&priv->cm.rx_drain_list))
 217                 return;
 218
 219         /*
 220          * QPs on flush list are error state.  This way, a "flush
 221          * error" WC will be immediately generated for each WR we post.
 222          */
 223         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 224         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 225                 ipoib_warn(priv, "failed to post drain wr\n");
 226
 227         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 228 }
 229
 230 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 231 {
 232         struct ipoib_cm_rx *p = ctx;
 233         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 234         unsigned long flags;
 235
 236         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 237                 return;
 238
 239         spin_lock_irqsave(&priv->lock, flags);
 240         list_move(&p->list, &priv->cm.rx_flush_list);
 241         p->state = IPOIB_CM_RX_FLUSH;
 242         ipoib_cm_start_rx_drain(priv);
 243         spin_unlock_irqrestore(&priv->lock, flags);
 244 }
 245
 246 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 247                                            struct ipoib_cm_rx *p)
 248 {
 249         struct ipoib_dev_priv *priv = netdev_priv(dev);
 250         struct ib_qp_init_attr attr = {
 251                 .event_handler = ipoib_cm_rx_event_handler,
 252                 .send_cq = priv->recv_cq, /* For drain WR */
 253                 .recv_cq = priv->recv_cq,
 254                 .srq = priv->cm.srq,
 255                 .cap.max_send_wr = 1, /* For drain WR */
 256                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 257                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 258                 .qp_type = IB_QPT_RC,
 259                 .qp_context = p,
 260         };
 261
 262         if (!ipoib_cm_has_srq(dev)) {
 263                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 264                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 265         }
 266
 267         return ib_create_qp(priv->pd, &attr);
 268 }
 269
 270 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 271                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 272                                  unsigned psn)
 273 {
 274         struct ipoib_dev_priv *priv = netdev_priv(dev);
 275         struct ib_qp_attr qp_attr;
 276         int qp_attr_mask, ret;
 277
 278         qp_attr.qp_state = IB_QPS_INIT;
 279         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 280         if (ret) {
 281                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 282                 return ret;
 283         }
 284         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 285         if (ret) {
 286                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 287                 return ret;
 288         }
 289         qp_attr.qp_state = IB_QPS_RTR;
 290         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 291         if (ret) {
 292                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 293                 return ret;
 294         }
 295         qp_attr.rq_psn = psn;
 296         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 297         if (ret) {
 298                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 299                 return ret;
 300         }
 301
 302         /*
 303          * Current Mellanox HCA firmware won't generate completions
 304          * with error for drain WRs unless the QP has been moved to
 305          * RTS first. This work-around leaves a window where a QP has
 306          * moved to error asynchronously, but this will eventually get
 307          * fixed in firmware, so let's not error out if modify QP
 308          * fails.
 309          */
 310         qp_attr.qp_state = IB_QPS_RTS;
 311         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 312         if (ret) {
 313                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 314                 return 0;
 315         }
 316         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 317         if (ret) {
 318                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 319                 return 0;
 320         }
 321
 322         return 0;
 323 }
 324
 325 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 326                                 struct ib_recv_wr *wr,
 327                                 struct ib_sge *sge)
 328 {
 329         struct ipoib_dev_priv *priv = netdev_priv(dev);
 330         int i;
 331
 332         for (i = 0; i < priv->cm.num_frags; ++i)
 333                 sge[i].lkey = priv->mr->lkey;
 334
 335         sge[0].length = IPOIB_CM_HEAD_SIZE;
 336         for (i = 1; i < priv->cm.num_frags; ++i)
 337                 sge[i].length = PAGE_SIZE;
 338
 339         wr->next    = NULL;
 340         wr->sg_list = sge;
 341         wr->num_sge = priv->cm.num_frags;
 342 }
 343
 344 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 345                                    struct ipoib_cm_rx *rx)
 346 {
 347         struct ipoib_dev_priv *priv = netdev_priv(dev);
 348         struct {
 349                 struct ib_recv_wr wr;
 350                 struct ib_sge sge[IPOIB_CM_RX_SG];
 351         } *t;
 352         int ret;
 353         int i;
 354
 355         rx->rx_ring = vmalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 356         if (!rx->rx_ring) {
 357                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 358                        priv->ca->name, ipoib_recvq_size);
 359                 return -ENOMEM;
 360         }
 361
 362         memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
 363
 364         t = kmalloc(sizeof *t, GFP_KERNEL);
 365         if (!t) {
 366                 ret = -ENOMEM;
 367                 goto err_free;
 368         }
 369
 370         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 371
 372         spin_lock_irq(&priv->lock);
 373
 374         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 375                 spin_unlock_irq(&priv->lock);
 376                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 377                 ret = -EINVAL;
 378                 goto err_free;
 379         } else
 380                 ++priv->cm.nonsrq_conn_qp;
 381
 382         spin_unlock_irq(&priv->lock);
 383
 384         for (i = 0; i < ipoib_recvq_size; ++i) {
 385                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 386                                            rx->rx_ring[i].mapping)) {
 387                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 388                                 ret = -ENOMEM;
 389                                 goto err_count;
 390                 }
 391                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 392                 if (ret) {
 393                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 394                                    "failed for buf %d\n", i);
 395                         ret = -EIO;
 396                         goto err_count;
 397                 }
 398         }
 399
 400         rx->recv_count = ipoib_recvq_size;
 401
 402         kfree(t);
 403
 404         return 0;
 405
 406 err_count:
 407         spin_lock_irq(&priv->lock);
 408         --priv->cm.nonsrq_conn_qp;
 409         spin_unlock_irq(&priv->lock);
 410
 411 err_free:
 412         kfree(t);
 413         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 414
 415         return ret;
 416 }
 417
 418 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 419                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 420                              unsigned psn)
 421 {
 422         struct ipoib_dev_priv *priv = netdev_priv(dev);
 423         struct ipoib_cm_data data = {};
 424         struct ib_cm_rep_param rep = {};
 425
 426         data.qpn = cpu_to_be32(priv->qp->qp_num);
 427         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 428
 429         rep.private_data = &data;
 430         rep.private_data_len = sizeof data;
 431         rep.flow_control = 0;
 432         rep.rnr_retry_count = req->rnr_retry_count;
 433         rep.srq = ipoib_cm_has_srq(dev);
 434         rep.qp_num = qp->qp_num;
 435         rep.starting_psn = psn;
 436         return ib_send_cm_rep(cm_id, &rep);
 437 }
 438
 439 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 440 {
 441         struct net_device *dev = cm_id->context;
 442         struct ipoib_dev_priv *priv = netdev_priv(dev);
 443         struct ipoib_cm_rx *p;
 444         unsigned psn;
 445         int ret;
 446
 447         ipoib_dbg(priv, "REQ arrived\n");
 448         p = kzalloc(sizeof *p, GFP_KERNEL);
 449         if (!p)
 450                 return -ENOMEM;
 451         p->dev = dev;
 452         p->id = cm_id;
 453         cm_id->context = p;
 454         p->state = IPOIB_CM_RX_LIVE;
 455         p->jiffies = jiffies;
 456         INIT_LIST_HEAD(&p->list);
 457
 458         p->qp = ipoib_cm_create_rx_qp(dev, p);
 459         if (IS_ERR(p->qp)) {
 460                 ret = PTR_ERR(p->qp);
 461                 goto err_qp;
 462         }
 463
 464         psn = random32() & 0xffffff;
 465         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 466         if (ret)
 467                 goto err_modify;
 468
 469         if (!ipoib_cm_has_srq(dev)) {
 470                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 471                 if (ret)
 472                         goto err_modify;
 473         }
 474
 475         spin_lock_irq(&priv->lock);
 476         queue_delayed_work(ipoib_workqueue,
 477                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 478         /* Add this entry to passive ids list head, but do not re-add it
 479          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 480         p->jiffies = jiffies;
 481         if (p->state == IPOIB_CM_RX_LIVE)
 482                 list_move(&p->list, &priv->cm.passive_ids);
 483         spin_unlock_irq(&priv->lock);
 484
 485         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 486         if (ret) {
 487                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 488                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 489                         ipoib_warn(priv, "unable to move qp to error state\n");
 490         }
 491         return 0;
 492
 493 err_modify:
 494         ib_destroy_qp(p->qp);
 495 err_qp:
 496         kfree(p);
 497         return ret;
 498 }
 499
 500 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 501                                struct ib_cm_event *event)
 502 {
 503         struct ipoib_cm_rx *p;
 504         struct ipoib_dev_priv *priv;
 505
 506         switch (event->event) {
 507         case IB_CM_REQ_RECEIVED:
 508                 return ipoib_cm_req_handler(cm_id, event);
 509         case IB_CM_DREQ_RECEIVED:
 510                 p = cm_id->context;
 511                 ib_send_cm_drep(cm_id, NULL, 0);
 512                 /* Fall through */
 513         case IB_CM_REJ_RECEIVED:
 514                 p = cm_id->context;
 515                 priv = netdev_priv(p->dev);
 516                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 517                         ipoib_warn(priv, "unable to move qp to error state\n");
 518                 /* Fall through */
 519         default:
 520                 return 0;
 521         }
 522 }
 523 /* Adjust length of skb with fragments to match received data */
 524 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 525                           unsigned int length, struct sk_buff *toskb)
 526 {
 527         int i, num_frags;
 528         unsigned int size;
 529
 530         /* put header into skb */
 531         size = min(length, hdr_space);
 532         skb->tail += size;
 533         skb->len += size;
 534         length -= size;
 535
 536         num_frags = skb_shinfo(skb)->nr_frags;
 537         for (i = 0; i < num_frags; i++) {
 538                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 539
 540                 if (length == 0) {
 541                         /* don't need this page */
 542                         skb_fill_page_desc(toskb, i, frag->page, 0, PAGE_SIZE);
 543                         --skb_shinfo(skb)->nr_frags;
 544                 } else {
 545                         size = min(length, (unsigned) PAGE_SIZE);
 546
 547                         frag->size = 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                         if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 759                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 760                         netif_stop_queue(dev);
 761                 }
 762         }
 763 }
 764
 765 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 766 {
 767         struct ipoib_dev_priv *priv = netdev_priv(dev);
 768         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 769         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 770         struct ipoib_cm_tx_buf *tx_req;
 771         unsigned long flags;
 772
 773         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 774                        wr_id, wc->status);
 775
 776         if (unlikely(wr_id >= ipoib_sendq_size)) {
 777                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 778                            wr_id, ipoib_sendq_size);
 779                 return;
 780         }
 781
 782         tx_req = &tx->tx_ring[wr_id];
 783
 784         ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len, DMA_TO_DEVICE);
 785
 786         /* FIXME: is this right? Shouldn't we only increment on success? */
 787         ++dev->stats.tx_packets;
 788         dev->stats.tx_bytes += tx_req->skb->len;
 789
 790         dev_kfree_skb_any(tx_req->skb);
 791
 792         netif_tx_lock(dev);
 793
 794         ++tx->tx_tail;
 795         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 796             netif_queue_stopped(dev) &&
 797             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 798                 netif_wake_queue(dev);
 799
 800         if (wc->status != IB_WC_SUCCESS &&
 801             wc->status != IB_WC_WR_FLUSH_ERR) {
 802                 struct ipoib_neigh *neigh;
 803
 804                 ipoib_dbg(priv, "failed cm send event "
 805                            "(status=%d, wrid=%d vend_err %x)\n",
 806                            wc->status, wr_id, wc->vendor_err);
 807
 808                 spin_lock_irqsave(&priv->lock, flags);
 809                 neigh = tx->neigh;
 810
 811                 if (neigh) {
 812                         neigh->cm = NULL;
 813                         list_del(&neigh->list);
 814                         if (neigh->ah)
 815                                 ipoib_put_ah(neigh->ah);
 816                         ipoib_neigh_free(dev, neigh);
 817
 818                         tx->neigh = NULL;
 819                 }
 820
 821                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 822                         list_move(&tx->list, &priv->cm.reap_list);
 823                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 824                 }
 825
 826                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 827
 828                 spin_unlock_irqrestore(&priv->lock, flags);
 829         }
 830
 831         netif_tx_unlock(dev);
 832 }
 833
 834 int ipoib_cm_dev_open(struct net_device *dev)
 835 {
 836         struct ipoib_dev_priv *priv = netdev_priv(dev);
 837         int ret;
 838
 839         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 840                 return 0;
 841
 842         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 843         if (IS_ERR(priv->cm.id)) {
 844                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 845                 ret = PTR_ERR(priv->cm.id);
 846                 goto err_cm;
 847         }
 848
 849         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 850                            0, NULL);
 851         if (ret) {
 852                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 853                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 854                 goto err_listen;
 855         }
 856
 857         return 0;
 858
 859 err_listen:
 860         ib_destroy_cm_id(priv->cm.id);
 861 err_cm:
 862         priv->cm.id = NULL;
 863         return ret;
 864 }
 865
 866 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 867 {
 868         struct ipoib_dev_priv *priv = netdev_priv(dev);
 869         struct ipoib_cm_rx *rx, *n;
 870         LIST_HEAD(list);
 871
 872         spin_lock_irq(&priv->lock);
 873         list_splice_init(&priv->cm.rx_reap_list, &list);
 874         spin_unlock_irq(&priv->lock);
 875
 876         list_for_each_entry_safe(rx, n, &list, list) {
 877                 ib_destroy_cm_id(rx->id);
 878                 ib_destroy_qp(rx->qp);
 879                 if (!ipoib_cm_has_srq(dev)) {
 880                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 881                         spin_lock_irq(&priv->lock);
 882                         --priv->cm.nonsrq_conn_qp;
 883                         spin_unlock_irq(&priv->lock);
 884                 }
 885                 kfree(rx);
 886         }
 887 }
 888
 889 void ipoib_cm_dev_stop(struct net_device *dev)
 890 {
 891         struct ipoib_dev_priv *priv = netdev_priv(dev);
 892         struct ipoib_cm_rx *p;
 893         unsigned long begin;
 894         int ret;
 895
 896         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 897                 return;
 898
 899         ib_destroy_cm_id(priv->cm.id);
 900         priv->cm.id = NULL;
 901
 902         spin_lock_irq(&priv->lock);
 903         while (!list_empty(&priv->cm.passive_ids)) {
 904                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 905                 list_move(&p->list, &priv->cm.rx_error_list);
 906                 p->state = IPOIB_CM_RX_ERROR;
 907                 spin_unlock_irq(&priv->lock);
 908                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 909                 if (ret)
 910                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 911                 spin_lock_irq(&priv->lock);
 912         }
 913
 914         /* Wait for all RX to be drained */
 915         begin = jiffies;
 916
 917         while (!list_empty(&priv->cm.rx_error_list) ||
 918                !list_empty(&priv->cm.rx_flush_list) ||
 919                !list_empty(&priv->cm.rx_drain_list)) {
 920                 if (time_after(jiffies, begin + 5 * HZ)) {
 921                         ipoib_warn(priv, "RX drain timing out\n");
 922
 923                         /*
 924                          * assume the HW is wedged and just free up everything.
 925                          */
 926                         list_splice_init(&priv->cm.rx_flush_list,
 927                                          &priv->cm.rx_reap_list);
 928                         list_splice_init(&priv->cm.rx_error_list,
 929                                          &priv->cm.rx_reap_list);
 930                         list_splice_init(&priv->cm.rx_drain_list,
 931                                          &priv->cm.rx_reap_list);
 932                         break;
 933                 }
 934                 spin_unlock_irq(&priv->lock);
 935                 msleep(1);
 936                 ipoib_drain_cq(dev);
 937                 spin_lock_irq(&priv->lock);
 938         }
 939
 940         spin_unlock_irq(&priv->lock);
 941
 942         ipoib_cm_free_rx_reap_list(dev);
 943
 944         cancel_delayed_work(&priv->cm.stale_task);
 945 }
 946
 947 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 948 {
 949         struct ipoib_cm_tx *p = cm_id->context;
 950         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 951         struct ipoib_cm_data *data = event->private_data;
 952         struct sk_buff_head skqueue;
 953         struct ib_qp_attr qp_attr;
 954         int qp_attr_mask, ret;
 955         struct sk_buff *skb;
 956
 957         p->mtu = be32_to_cpu(data->mtu);
 958
 959         if (p->mtu <= IPOIB_ENCAP_LEN) {
 960                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 961                            p->mtu, IPOIB_ENCAP_LEN);
 962                 return -EINVAL;
 963         }
 964
 965         qp_attr.qp_state = IB_QPS_RTR;
 966         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 967         if (ret) {
 968                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 969                 return ret;
 970         }
 971
 972         qp_attr.rq_psn = 0 /* FIXME */;
 973         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 974         if (ret) {
 975                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 976                 return ret;
 977         }
 978
 979         qp_attr.qp_state = IB_QPS_RTS;
 980         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 981         if (ret) {
 982                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 983                 return ret;
 984         }
 985         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 986         if (ret) {
 987                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 988                 return ret;
 989         }
 990
 991         skb_queue_head_init(&skqueue);
 992
 993         spin_lock_irq(&priv->lock);
 994         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 995         if (p->neigh)
 996                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 997                         __skb_queue_tail(&skqueue, skb);
 998         spin_unlock_irq(&priv->lock);
 999
1000         while ((skb = __skb_dequeue(&skqueue))) {
1001                 skb->dev = p->dev;
1002                 if (dev_queue_xmit(skb))
1003                         ipoib_warn(priv, "dev_queue_xmit failed "
1004                                    "to requeue packet\n");
1005         }
1006
1007         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1008         if (ret) {
1009                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1010                 return ret;
1011         }
1012         return 0;
1013 }
1014
1015 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1016 {
1017         struct ipoib_dev_priv *priv = netdev_priv(dev);
1018         struct ib_qp_init_attr attr = {
1019                 .send_cq                = priv->recv_cq,
1020                 .recv_cq                = priv->recv_cq,
1021                 .srq                    = priv->cm.srq,
1022                 .cap.max_send_wr        = ipoib_sendq_size,
1023                 .cap.max_send_sge       = 1,
1024                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1025                 .qp_type                = IB_QPT_RC,
1026                 .qp_context             = tx
1027         };
1028
1029         return ib_create_qp(priv->pd, &attr);
1030 }
1031
1032 static int ipoib_cm_send_req(struct net_device *dev,
1033                              struct ib_cm_id *id, struct ib_qp *qp,
1034                              u32 qpn,
1035                              struct ib_sa_path_rec *pathrec)
1036 {
1037         struct ipoib_dev_priv *priv = netdev_priv(dev);
1038         struct ipoib_cm_data data = {};
1039         struct ib_cm_req_param req = {};
1040
1041         data.qpn = cpu_to_be32(priv->qp->qp_num);
1042         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1043
1044         req.primary_path                = pathrec;
1045         req.alternate_path              = NULL;
1046         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1047         req.qp_num                      = qp->qp_num;
1048         req.qp_type                     = qp->qp_type;
1049         req.private_data                = &data;
1050         req.private_data_len            = sizeof data;
1051         req.flow_control                = 0;
1052
1053         req.starting_psn                = 0; /* FIXME */
1054
1055         /*
1056          * Pick some arbitrary defaults here; we could make these
1057          * module parameters if anyone cared about setting them.
1058          */
1059         req.responder_resources         = 4;
1060         req.remote_cm_response_timeout  = 20;
1061         req.local_cm_response_timeout   = 20;
1062         req.retry_count                 = 0; /* RFC draft warns against retries */
1063         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1064         req.max_cm_retries              = 15;
1065         req.srq                         = ipoib_cm_has_srq(dev);
1066         return ib_send_cm_req(id, &req);
1067 }
1068
1069 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1070                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1071 {
1072         struct ipoib_dev_priv *priv = netdev_priv(dev);
1073         struct ib_qp_attr qp_attr;
1074         int qp_attr_mask, ret;
1075         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1076         if (ret) {
1077                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1078                 return ret;
1079         }
1080
1081         qp_attr.qp_state = IB_QPS_INIT;
1082         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1083         qp_attr.port_num = priv->port;
1084         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1085
1086         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1087         if (ret) {
1088                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1089                 return ret;
1090         }
1091         return 0;
1092 }
1093
1094 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1095                             struct ib_sa_path_rec *pathrec)
1096 {
1097         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1098         int ret;
1099
1100         p->tx_ring = vmalloc(ipoib_sendq_size * sizeof *p->tx_ring);
1101         if (!p->tx_ring) {
1102                 ipoib_warn(priv, "failed to allocate tx ring\n");
1103                 ret = -ENOMEM;
1104                 goto err_tx;
1105         }
1106         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1107
1108         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1109         if (IS_ERR(p->qp)) {
1110                 ret = PTR_ERR(p->qp);
1111                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1112                 goto err_qp;
1113         }
1114
1115         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1116         if (IS_ERR(p->id)) {
1117                 ret = PTR_ERR(p->id);
1118                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1119                 goto err_id;
1120         }
1121
1122         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1123         if (ret) {
1124                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1125                 goto err_modify;
1126         }
1127
1128         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1129         if (ret) {
1130                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1131                 goto err_send_cm;
1132         }
1133
1134         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1135                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1136
1137         return 0;
1138
1139 err_send_cm:
1140 err_modify:
1141         ib_destroy_cm_id(p->id);
1142 err_id:
1143         p->id = NULL;
1144         ib_destroy_qp(p->qp);
1145 err_qp:
1146         p->qp = NULL;
1147         vfree(p->tx_ring);
1148 err_tx:
1149         return ret;
1150 }
1151
1152 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1153 {
1154         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1155         struct ipoib_cm_tx_buf *tx_req;
1156         unsigned long begin;
1157
1158         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1159                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1160
1161         if (p->id)
1162                 ib_destroy_cm_id(p->id);
1163
1164         if (p->tx_ring) {
1165                 /* Wait for all sends to complete */
1166                 begin = jiffies;
1167                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1168                         if (time_after(jiffies, begin + 5 * HZ)) {
1169                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1170                                            p->tx_head - p->tx_tail);
1171                                 goto timeout;
1172                         }
1173
1174                         msleep(1);
1175                 }
1176         }
1177
1178 timeout:
1179
1180         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1181                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1182                 ib_dma_unmap_single(priv->ca, tx_req->mapping, tx_req->skb->len,
1183                                     DMA_TO_DEVICE);
1184                 dev_kfree_skb_any(tx_req->skb);
1185                 ++p->tx_tail;
1186                 netif_tx_lock_bh(p->dev);
1187                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1188                     netif_queue_stopped(p->dev) &&
1189                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1190                         netif_wake_queue(p->dev);
1191                 netif_tx_unlock_bh(p->dev);
1192         }
1193
1194         if (p->qp)
1195                 ib_destroy_qp(p->qp);
1196
1197         vfree(p->tx_ring);
1198         kfree(p);
1199 }
1200
1201 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1202                                struct ib_cm_event *event)
1203 {
1204         struct ipoib_cm_tx *tx = cm_id->context;
1205         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1206         struct net_device *dev = priv->dev;
1207         struct ipoib_neigh *neigh;
1208         unsigned long flags;
1209         int ret;
1210
1211         switch (event->event) {
1212         case IB_CM_DREQ_RECEIVED:
1213                 ipoib_dbg(priv, "DREQ received.\n");
1214                 ib_send_cm_drep(cm_id, NULL, 0);
1215                 break;
1216         case IB_CM_REP_RECEIVED:
1217                 ipoib_dbg(priv, "REP received.\n");
1218                 ret = ipoib_cm_rep_handler(cm_id, event);
1219                 if (ret)
1220                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1221                                        NULL, 0, NULL, 0);
1222                 break;
1223         case IB_CM_REQ_ERROR:
1224         case IB_CM_REJ_RECEIVED:
1225         case IB_CM_TIMEWAIT_EXIT:
1226                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1227                 netif_tx_lock_bh(dev);
1228                 spin_lock_irqsave(&priv->lock, flags);
1229                 neigh = tx->neigh;
1230
1231                 if (neigh) {
1232                         neigh->cm = NULL;
1233                         list_del(&neigh->list);
1234                         if (neigh->ah)
1235                                 ipoib_put_ah(neigh->ah);
1236                         ipoib_neigh_free(dev, neigh);
1237
1238                         tx->neigh = NULL;
1239                 }
1240
1241                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1242                         list_move(&tx->list, &priv->cm.reap_list);
1243                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1244                 }
1245
1246                 spin_unlock_irqrestore(&priv->lock, flags);
1247                 netif_tx_unlock_bh(dev);
1248                 break;
1249         default:
1250                 break;
1251         }
1252
1253         return 0;
1254 }
1255
1256 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1257                                        struct ipoib_neigh *neigh)
1258 {
1259         struct ipoib_dev_priv *priv = netdev_priv(dev);
1260         struct ipoib_cm_tx *tx;
1261
1262         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1263         if (!tx)
1264                 return NULL;
1265
1266         neigh->cm = tx;
1267         tx->neigh = neigh;
1268         tx->path = path;
1269         tx->dev = dev;
1270         list_add(&tx->list, &priv->cm.start_list);
1271         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1272         queue_work(ipoib_workqueue, &priv->cm.start_task);
1273         return tx;
1274 }
1275
1276 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1277 {
1278         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1279         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1280                 list_move(&tx->list, &priv->cm.reap_list);
1281                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1282                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1283                           tx->neigh->dgid.raw);
1284                 tx->neigh = NULL;
1285         }
1286 }
1287
1288 static void ipoib_cm_tx_start(struct work_struct *work)
1289 {
1290         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1291                                                    cm.start_task);
1292         struct net_device *dev = priv->dev;
1293         struct ipoib_neigh *neigh;
1294         struct ipoib_cm_tx *p;
1295         unsigned long flags;
1296         int ret;
1297
1298         struct ib_sa_path_rec pathrec;
1299         u32 qpn;
1300
1301         netif_tx_lock_bh(dev);
1302         spin_lock_irqsave(&priv->lock, flags);
1303
1304         while (!list_empty(&priv->cm.start_list)) {
1305                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1306                 list_del_init(&p->list);
1307                 neigh = p->neigh;
1308                 qpn = IPOIB_QPN(neigh->neighbour->ha);
1309                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1310
1311                 spin_unlock_irqrestore(&priv->lock, flags);
1312                 netif_tx_unlock_bh(dev);
1313
1314                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1315
1316                 netif_tx_lock_bh(dev);
1317                 spin_lock_irqsave(&priv->lock, flags);
1318
1319                 if (ret) {
1320                         neigh = p->neigh;
1321                         if (neigh) {
1322                                 neigh->cm = NULL;
1323                                 list_del(&neigh->list);
1324                                 if (neigh->ah)
1325                                         ipoib_put_ah(neigh->ah);
1326                                 ipoib_neigh_free(dev, neigh);
1327                         }
1328                         list_del(&p->list);
1329                         kfree(p);
1330                 }
1331         }
1332
1333         spin_unlock_irqrestore(&priv->lock, flags);
1334         netif_tx_unlock_bh(dev);
1335 }
1336
1337 static void ipoib_cm_tx_reap(struct work_struct *work)
1338 {
1339         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1340                                                    cm.reap_task);
1341         struct net_device *dev = priv->dev;
1342         struct ipoib_cm_tx *p;
1343         unsigned long flags;
1344
1345         netif_tx_lock_bh(dev);
1346         spin_lock_irqsave(&priv->lock, flags);
1347
1348         while (!list_empty(&priv->cm.reap_list)) {
1349                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1350                 list_del(&p->list);
1351                 spin_unlock_irqrestore(&priv->lock, flags);
1352                 netif_tx_unlock_bh(dev);
1353                 ipoib_cm_tx_destroy(p);
1354                 netif_tx_lock_bh(dev);
1355                 spin_lock_irqsave(&priv->lock, flags);
1356         }
1357
1358         spin_unlock_irqrestore(&priv->lock, flags);
1359         netif_tx_unlock_bh(dev);
1360 }
1361
1362 static void ipoib_cm_skb_reap(struct work_struct *work)
1363 {
1364         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1365                                                    cm.skb_task);
1366         struct net_device *dev = priv->dev;
1367         struct sk_buff *skb;
1368         unsigned long flags;
1369         unsigned mtu = priv->mcast_mtu;
1370
1371         netif_tx_lock_bh(dev);
1372         spin_lock_irqsave(&priv->lock, flags);
1373
1374         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1375                 spin_unlock_irqrestore(&priv->lock, flags);
1376                 netif_tx_unlock_bh(dev);
1377
1378                 if (skb->protocol == htons(ETH_P_IP))
1379                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1380 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
1381                 else if (skb->protocol == htons(ETH_P_IPV6))
1382                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1383 #endif
1384                 dev_kfree_skb_any(skb);
1385
1386                 netif_tx_lock_bh(dev);
1387                 spin_lock_irqsave(&priv->lock, flags);
1388         }
1389
1390         spin_unlock_irqrestore(&priv->lock, flags);
1391         netif_tx_unlock_bh(dev);
1392 }
1393
1394 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1395                            unsigned int mtu)
1396 {
1397         struct ipoib_dev_priv *priv = netdev_priv(dev);
1398         int e = skb_queue_empty(&priv->cm.skb_queue);
1399
1400         if (skb_dst(skb))
1401                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), mtu);
1402
1403         skb_queue_tail(&priv->cm.skb_queue, skb);
1404         if (e)
1405                 queue_work(ipoib_workqueue, &priv->cm.skb_task);
1406 }
1407
1408 static void ipoib_cm_rx_reap(struct work_struct *work)
1409 {
1410         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1411                                                 cm.rx_reap_task)->dev);
1412 }
1413
1414 static void ipoib_cm_stale_task(struct work_struct *work)
1415 {
1416         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1417                                                    cm.stale_task.work);
1418         struct ipoib_cm_rx *p;
1419         int ret;
1420
1421         spin_lock_irq(&priv->lock);
1422         while (!list_empty(&priv->cm.passive_ids)) {
1423                 /* List is sorted by LRU, start from tail,
1424                  * stop when we see a recently used entry */
1425                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1426                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1427                         break;
1428                 list_move(&p->list, &priv->cm.rx_error_list);
1429                 p->state = IPOIB_CM_RX_ERROR;
1430                 spin_unlock_irq(&priv->lock);
1431                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1432                 if (ret)
1433                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1434                 spin_lock_irq(&priv->lock);
1435         }
1436
1437         if (!list_empty(&priv->cm.passive_ids))
1438                 queue_delayed_work(ipoib_workqueue,
1439                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1440         spin_unlock_irq(&priv->lock);
1441 }
1442
1443
1444 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1445                          char *buf)
1446 {
1447         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1448
1449         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1450                 return sprintf(buf, "connected\n");
1451         else
1452                 return sprintf(buf, "datagram\n");
1453 }
1454
1455 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1456                         const char *buf, size_t count)
1457 {
1458         struct net_device *dev = to_net_dev(d);
1459         struct ipoib_dev_priv *priv = netdev_priv(dev);
1460
1461         if (!rtnl_trylock())
1462                 return restart_syscall();
1463
1464         /* flush paths if we switch modes so that connections are restarted */
1465         if (IPOIB_CM_SUPPORTED(dev->dev_addr) && !strcmp(buf, "connected\n")) {
1466                 set_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1467                 ipoib_warn(priv, "enabling connected mode "
1468                            "will cause multicast packet drops\n");
1469
1470                 dev->features &= ~(NETIF_F_IP_CSUM | NETIF_F_SG | NETIF_F_TSO);
1471                 rtnl_unlock();
1472                 priv->tx_wr.send_flags &= ~IB_SEND_IP_CSUM;
1473
1474                 ipoib_flush_paths(dev);
1475                 return count;
1476         }
1477
1478         if (!strcmp(buf, "datagram\n")) {
1479                 clear_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags);
1480
1481                 if (test_bit(IPOIB_FLAG_CSUM, &priv->flags)) {
1482                         dev->features |= NETIF_F_IP_CSUM | NETIF_F_SG;
1483                         priv->dev->features |= NETIF_F_GRO;
1484                         if (priv->hca_caps & IB_DEVICE_UD_TSO)
1485                                 dev->features |= NETIF_F_TSO;
1486                 }
1487                 dev_set_mtu(dev, min(priv->mcast_mtu, dev->mtu));
1488                 rtnl_unlock();
1489                 ipoib_flush_paths(dev);
1490
1491                 return count;
1492         }
1493         rtnl_unlock();
1494
1495         return -EINVAL;
1496 }
1497
1498 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1499
1500 int ipoib_cm_add_mode_attr(struct net_device *dev)
1501 {
1502         return device_create_file(&dev->dev, &dev_attr_mode);
1503 }
1504
1505 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1506 {
1507         struct ipoib_dev_priv *priv = netdev_priv(dev);
1508         struct ib_srq_init_attr srq_init_attr = {
1509                 .attr = {
1510                         .max_wr  = ipoib_recvq_size,
1511                         .max_sge = max_sge
1512                 }
1513         };
1514
1515         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1516         if (IS_ERR(priv->cm.srq)) {
1517                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1518                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1519                                priv->ca->name, PTR_ERR(priv->cm.srq));
1520                 priv->cm.srq = NULL;
1521                 return;
1522         }
1523
1524         priv->cm.srq_ring = vmalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1525         if (!priv->cm.srq_ring) {
1526                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1527                        priv->ca->name, ipoib_recvq_size);
1528                 ib_destroy_srq(priv->cm.srq);
1529                 priv->cm.srq = NULL;
1530                 return;
1531         }
1532
1533         memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1534 }
1535
1536 int ipoib_cm_dev_init(struct net_device *dev)
1537 {
1538         struct ipoib_dev_priv *priv = netdev_priv(dev);
1539         int i, ret;
1540         struct ib_device_attr attr;
1541
1542         INIT_LIST_HEAD(&priv->cm.passive_ids);
1543         INIT_LIST_HEAD(&priv->cm.reap_list);
1544         INIT_LIST_HEAD(&priv->cm.start_list);
1545         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1546         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1547         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1548         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1549         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1550         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1551         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1552         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1553         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1554
1555         skb_queue_head_init(&priv->cm.skb_queue);
1556
1557         ret = ib_query_device(priv->ca, &attr);
1558         if (ret) {
1559                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1560                 return ret;
1561         }
1562
1563         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1564
1565         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1566         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1567         if (ipoib_cm_has_srq(dev)) {
1568                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1569                 priv->cm.num_frags  = attr.max_srq_sge;
1570                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1571                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1572         } else {
1573                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1574                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1575         }
1576
1577         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1578
1579         if (ipoib_cm_has_srq(dev)) {
1580                 for (i = 0; i < ipoib_recvq_size; ++i) {
1581                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1582                                                    priv->cm.num_frags - 1,
1583                                                    priv->cm.srq_ring[i].mapping)) {
1584                                 ipoib_warn(priv, "failed to allocate "
1585                                            "receive buffer %d\n", i);
1586                                 ipoib_cm_dev_cleanup(dev);
1587                                 return -ENOMEM;
1588                         }
1589
1590                         if (ipoib_cm_post_receive_srq(dev, i)) {
1591                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1592                                            "failed for buf %d\n", i);
1593                                 ipoib_cm_dev_cleanup(dev);
1594                                 return -EIO;
1595                         }
1596                 }
1597         }
1598
1599         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1600         return 0;
1601 }
1602
1603 void ipoib_cm_dev_cleanup(struct net_device *dev)
1604 {
1605         struct ipoib_dev_priv *priv = netdev_priv(dev);
1606         int ret;
1607
1608         if (!priv->cm.srq)
1609                 return;
1610
1611         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1612
1613         ret = ib_destroy_srq(priv->cm.srq);
1614         if (ret)
1615                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1616
1617         priv->cm.srq = NULL;
1618         if (!priv->cm.srq_ring)
1619                 return;
1620
1621         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1622         priv->cm.srq_ring = NULL;
1623 }