sys/ofed/drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
   3  *
   4  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <sys/cdefs.h>
  36 __FBSDID("$FreeBSD$");
  37
  38 #include "ipoib.h"
  39
  40 #ifdef CONFIG_INFINIBAND_IPOIB_CM
  41
  42 #include <netinet/ip.h>
  43 #include <netinet/ip_icmp.h>
  44 #include <netinet/icmp6.h>
  45
  46 #include <rdma/ib_cm.h>
  47 #include <rdma/ib_cache.h>
  48 #include <linux/delay.h>
  49
  50 int ipoib_max_conn_qp = 128;
  51
  52 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  53 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  54                  "Max number of connected-mode QPs per interface "
  55                  "(applied only if shared receive queue is not available)");
  56
  57 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  58 static int data_debug_level;
  59
  60 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  61 MODULE_PARM_DESC(cm_data_debug_level,
  62                  "Enable data path debug tracing for connected mode if > 0");
  63 #endif
  64
  65 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  66
  67 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  68 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  69 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  70 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  71
  72 static struct ib_qp_attr ipoib_cm_err_attr = {
  73         .qp_state = IB_QPS_ERR
  74 };
  75
  76 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  77
  78 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  79         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  80         .opcode = IB_WR_SEND,
  81 };
  82
  83 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  84                                struct ib_cm_event *event);
  85
  86 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
  87 {
  88
  89         ipoib_dma_unmap_rx(priv, (struct ipoib_rx_buf *)rx_req);
  90
  91 }
  92
  93 static int ipoib_cm_post_receive_srq(struct ipoib_dev_priv *priv, int id)
  94 {
  95         struct ib_recv_wr *bad_wr;
  96         struct ipoib_rx_buf *rx_req;
  97         struct mbuf *m;
  98         int ret;
  99         int i;
 100
 101         rx_req = (struct ipoib_rx_buf *)&priv->cm.srq_ring[id];
 102         for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
 103                 priv->cm.rx_sge[i].addr = rx_req->mapping[i];
 104                 priv->cm.rx_sge[i].length = m->m_len;
 105         }
 106
 107         priv->cm.rx_wr.num_sge = i;
 108         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 109
 110         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 111         if (unlikely(ret)) {
 112                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 113                 ipoib_dma_unmap_rx(priv, rx_req);
 114                 m_freem(priv->cm.srq_ring[id].mb);
 115                 priv->cm.srq_ring[id].mb = NULL;
 116         }
 117
 118         return ret;
 119 }
 120
 121 static int ipoib_cm_post_receive_nonsrq(struct ipoib_dev_priv *priv,
 122                                         struct ipoib_cm_rx *rx,
 123                                         struct ib_recv_wr *wr,
 124                                         struct ib_sge *sge, int id)
 125 {
 126         struct ipoib_rx_buf *rx_req;
 127         struct ib_recv_wr *bad_wr;
 128         struct mbuf *m;
 129         int ret;
 130         int i;
 131
 132         rx_req = (struct ipoib_rx_buf *)&rx->rx_ring[id];
 133         for (m = rx_req->mb, i = 0; m != NULL; m = m->m_next, i++) {
 134                 sge[i].addr = rx_req->mapping[i];
 135                 sge[i].length = m->m_len;
 136         }
 137
 138         wr->num_sge = i;
 139         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 140
 141         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 142         if (unlikely(ret)) {
 143                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 144                 ipoib_dma_unmap_rx(priv, rx_req);
 145                 m_freem(rx->rx_ring[id].mb);
 146                 rx->rx_ring[id].mb = NULL;
 147         }
 148
 149         return ret;
 150 }
 151
 152 static struct mbuf *
 153 ipoib_cm_alloc_rx_mb(struct ipoib_dev_priv *priv, struct ipoib_cm_rx_buf *rx_req)
 154 {
 155         return ipoib_alloc_map_mb(priv, (struct ipoib_rx_buf *)rx_req,
 156             priv->cm.max_cm_mtu);
 157 }
 158
 159 static void ipoib_cm_free_rx_ring(struct ipoib_dev_priv *priv,
 160                                   struct ipoib_cm_rx_buf *rx_ring)
 161 {
 162         int i;
 163
 164         for (i = 0; i < ipoib_recvq_size; ++i)
 165                 if (rx_ring[i].mb) {
 166                         ipoib_cm_dma_unmap_rx(priv, &rx_ring[i]);
 167                         m_freem(rx_ring[i].mb);
 168                 }
 169
 170         kfree(rx_ring);
 171 }
 172
 173 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 174 {
 175         struct ib_send_wr *bad_wr;
 176         struct ipoib_cm_rx *p;
 177
 178         /* We only reserved 1 extra slot in CQ for drain WRs, so
 179          * make sure we have at most 1 outstanding WR. */
 180         if (list_empty(&priv->cm.rx_flush_list) ||
 181             !list_empty(&priv->cm.rx_drain_list))
 182                 return;
 183
 184         /*
 185          * QPs on flush list are error state.  This way, a "flush
 186          * error" WC will be immediately generated for each WR we post.
 187          */
 188         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 189         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 190                 ipoib_warn(priv, "failed to post drain wr\n");
 191
 192         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 193 }
 194
 195 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 196 {
 197         struct ipoib_cm_rx *p = ctx;
 198         struct ipoib_dev_priv *priv = p->priv;
 199         unsigned long flags;
 200
 201         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 202                 return;
 203
 204         spin_lock_irqsave(&priv->lock, flags);
 205         list_move(&p->list, &priv->cm.rx_flush_list);
 206         p->state = IPOIB_CM_RX_FLUSH;
 207         ipoib_cm_start_rx_drain(priv);
 208         spin_unlock_irqrestore(&priv->lock, flags);
 209 }
 210
 211 static struct ib_qp *ipoib_cm_create_rx_qp(struct ipoib_dev_priv *priv,
 212                                            struct ipoib_cm_rx *p)
 213 {
 214         struct ib_qp_init_attr attr = {
 215                 .event_handler = ipoib_cm_rx_event_handler,
 216                 .send_cq = priv->recv_cq, /* For drain WR */
 217                 .recv_cq = priv->recv_cq,
 218                 .srq = priv->cm.srq,
 219                 .cap.max_send_wr = 1, /* For drain WR */
 220                 .cap.max_send_sge = 1,
 221                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 222                 .qp_type = IB_QPT_RC,
 223                 .qp_context = p,
 224         };
 225
 226         if (!ipoib_cm_has_srq(priv)) {
 227                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 228                 attr.cap.max_recv_sge = priv->cm.num_frags;
 229         }
 230
 231         return ib_create_qp(priv->pd, &attr);
 232 }
 233
 234 static int ipoib_cm_modify_rx_qp(struct ipoib_dev_priv *priv,
 235                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 236                                  unsigned psn)
 237 {
 238         struct ib_qp_attr qp_attr;
 239         int qp_attr_mask, ret;
 240
 241         qp_attr.qp_state = IB_QPS_INIT;
 242         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 243         if (ret) {
 244                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 245                 return ret;
 246         }
 247         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 248         if (ret) {
 249                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 250                 return ret;
 251         }
 252         qp_attr.qp_state = IB_QPS_RTR;
 253         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 254         if (ret) {
 255                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 256                 return ret;
 257         }
 258         qp_attr.rq_psn = psn;
 259         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 260         if (ret) {
 261                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 262                 return ret;
 263         }
 264
 265         /*
 266          * Current Mellanox HCA firmware won't generate completions
 267          * with error for drain WRs unless the QP has been moved to
 268          * RTS first. This work-around leaves a window where a QP has
 269          * moved to error asynchronously, but this will eventually get
 270          * fixed in firmware, so let's not error out if modify QP
 271          * fails.
 272          */
 273         qp_attr.qp_state = IB_QPS_RTS;
 274         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 275         if (ret) {
 276                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 277                 return 0;
 278         }
 279         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 280         if (ret) {
 281                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 282                 return 0;
 283         }
 284
 285         return 0;
 286 }
 287
 288 static void ipoib_cm_init_rx_wr(struct ipoib_dev_priv *priv,
 289                                 struct ib_recv_wr *wr,
 290                                 struct ib_sge *sge)
 291 {
 292         int i;
 293
 294         for (i = 0; i < IPOIB_CM_RX_SG; i++)
 295                 sge[i].lkey = priv->pd->local_dma_lkey;
 296
 297         wr->next    = NULL;
 298         wr->sg_list = sge;
 299         wr->num_sge = 1;
 300 }
 301
 302 static int ipoib_cm_nonsrq_init_rx(struct ipoib_dev_priv *priv,
 303     struct ib_cm_id *cm_id, struct ipoib_cm_rx *rx)
 304 {
 305         struct {
 306                 struct ib_recv_wr wr;
 307                 struct ib_sge sge[IPOIB_CM_RX_SG];
 308         } *t;
 309         int ret;
 310         int i;
 311
 312         rx->rx_ring = kzalloc(ipoib_recvq_size * sizeof *rx->rx_ring, GFP_KERNEL);
 313         if (!rx->rx_ring) {
 314                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 315                        priv->ca->name, ipoib_recvq_size);
 316                 return -ENOMEM;
 317         }
 318
 319         memset(rx->rx_ring, 0, ipoib_recvq_size * sizeof *rx->rx_ring);
 320
 321         t = kmalloc(sizeof *t, GFP_KERNEL);
 322         if (!t) {
 323                 ret = -ENOMEM;
 324                 goto err_free;
 325         }
 326
 327         ipoib_cm_init_rx_wr(priv, &t->wr, t->sge);
 328
 329         spin_lock_irq(&priv->lock);
 330
 331         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 332                 spin_unlock_irq(&priv->lock);
 333                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 334                 ret = -EINVAL;
 335                 goto err_free;
 336         } else
 337                 ++priv->cm.nonsrq_conn_qp;
 338
 339         spin_unlock_irq(&priv->lock);
 340
 341         for (i = 0; i < ipoib_recvq_size; ++i) {
 342                 if (!ipoib_cm_alloc_rx_mb(priv, &rx->rx_ring[i])) {
 343                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 344                                 ret = -ENOMEM;
 345                                 goto err_count;
 346                 }
 347                 ret = ipoib_cm_post_receive_nonsrq(priv, rx, &t->wr, t->sge, i);
 348                 if (ret) {
 349                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 350                                    "failed for buf %d\n", i);
 351                         ret = -EIO;
 352                         goto err_count;
 353                 }
 354         }
 355
 356         rx->recv_count = ipoib_recvq_size;
 357
 358         kfree(t);
 359
 360         return 0;
 361
 362 err_count:
 363         spin_lock_irq(&priv->lock);
 364         --priv->cm.nonsrq_conn_qp;
 365         spin_unlock_irq(&priv->lock);
 366
 367 err_free:
 368         kfree(t);
 369         ipoib_cm_free_rx_ring(priv, rx->rx_ring);
 370
 371         return ret;
 372 }
 373
 374 static int ipoib_cm_send_rep(struct ipoib_dev_priv *priv, struct ib_cm_id *cm_id,
 375                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 376                              unsigned psn)
 377 {
 378         struct ipoib_cm_data data = {};
 379         struct ib_cm_rep_param rep = {};
 380
 381         data.qpn = cpu_to_be32(priv->qp->qp_num);
 382         data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
 383
 384         rep.private_data = &data;
 385         rep.private_data_len = sizeof data;
 386         rep.flow_control = 0;
 387         rep.rnr_retry_count = req->rnr_retry_count;
 388         rep.srq = ipoib_cm_has_srq(priv);
 389         rep.qp_num = qp->qp_num;
 390         rep.starting_psn = psn;
 391         return ib_send_cm_rep(cm_id, &rep);
 392 }
 393
 394 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 395 {
 396         struct ipoib_dev_priv *priv = cm_id->context;
 397         struct ipoib_cm_rx *p;
 398         unsigned psn;
 399         int ret;
 400
 401         ipoib_dbg(priv, "REQ arrived\n");
 402         p = kzalloc(sizeof *p, GFP_KERNEL);
 403         if (!p)
 404                 return -ENOMEM;
 405         p->priv = priv;
 406         p->id = cm_id;
 407         cm_id->context = p;
 408         p->state = IPOIB_CM_RX_LIVE;
 409         p->jiffies = jiffies;
 410         INIT_LIST_HEAD(&p->list);
 411
 412         p->qp = ipoib_cm_create_rx_qp(priv, p);
 413         if (IS_ERR(p->qp)) {
 414                 ret = PTR_ERR(p->qp);
 415                 goto err_qp;
 416         }
 417
 418         psn = random() & 0xffffff;
 419         ret = ipoib_cm_modify_rx_qp(priv, cm_id, p->qp, psn);
 420         if (ret)
 421                 goto err_modify;
 422
 423         if (!ipoib_cm_has_srq(priv)) {
 424                 ret = ipoib_cm_nonsrq_init_rx(priv, cm_id, p);
 425                 if (ret)
 426                         goto err_modify;
 427         }
 428
 429         spin_lock_irq(&priv->lock);
 430         queue_delayed_work(ipoib_workqueue,
 431                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 432         /* Add this entry to passive ids list head, but do not re-add it
 433          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 434         p->jiffies = jiffies;
 435         if (p->state == IPOIB_CM_RX_LIVE)
 436                 list_move(&p->list, &priv->cm.passive_ids);
 437         spin_unlock_irq(&priv->lock);
 438
 439         ret = ipoib_cm_send_rep(priv, cm_id, p->qp, &event->param.req_rcvd, psn);
 440         if (ret) {
 441                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 442                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 443                         ipoib_warn(priv, "unable to move qp to error state\n");
 444         }
 445         return 0;
 446
 447 err_modify:
 448         ib_destroy_qp(p->qp);
 449 err_qp:
 450         kfree(p);
 451         return ret;
 452 }
 453
 454 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 455                                struct ib_cm_event *event)
 456 {
 457         struct ipoib_cm_rx *p;
 458         struct ipoib_dev_priv *priv;
 459
 460         switch (event->event) {
 461         case IB_CM_REQ_RECEIVED:
 462                 return ipoib_cm_req_handler(cm_id, event);
 463         case IB_CM_DREQ_RECEIVED:
 464                 p = cm_id->context;
 465                 ib_send_cm_drep(cm_id, NULL, 0);
 466                 /* Fall through */
 467         case IB_CM_REJ_RECEIVED:
 468                 p = cm_id->context;
 469                 priv = p->priv;
 470                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 471                         ipoib_warn(priv, "unable to move qp to error state\n");
 472                 /* Fall through */
 473         default:
 474                 return 0;
 475         }
 476 }
 477
 478 void ipoib_cm_handle_rx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
 479 {
 480         struct ipoib_cm_rx_buf saverx;
 481         struct ipoib_cm_rx_buf *rx_ring;
 482         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 483         struct ifnet *dev = priv->dev;
 484         struct mbuf *mb, *newmb;
 485         struct ipoib_cm_rx *p;
 486         int has_srq;
 487         u_short proto;
 488
 489         CURVNET_SET_QUIET(dev->if_vnet);
 490
 491         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 492                        wr_id, wc->status);
 493
 494         if (unlikely(wr_id >= ipoib_recvq_size)) {
 495                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 496                         spin_lock(&priv->lock);
 497                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 498                         ipoib_cm_start_rx_drain(priv);
 499                         if (priv->cm.id != NULL)
 500                                 queue_work(ipoib_workqueue,
 501                                     &priv->cm.rx_reap_task);
 502                         spin_unlock(&priv->lock);
 503                 } else
 504                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 505                                    wr_id, ipoib_recvq_size);
 506                 goto done;
 507         }
 508
 509         p = wc->qp->qp_context;
 510
 511         has_srq = ipoib_cm_has_srq(priv);
 512         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 513
 514         mb = rx_ring[wr_id].mb;
 515
 516         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 517                 ipoib_dbg(priv, "cm recv error "
 518                            "(status=%d, wrid=%d vend_err %x)\n",
 519                            wc->status, wr_id, wc->vendor_err);
 520                 if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
 521                 if (has_srq)
 522                         goto repost;
 523                 else {
 524                         if (!--p->recv_count) {
 525                                 spin_lock(&priv->lock);
 526                                 list_move(&p->list, &priv->cm.rx_reap_list);
 527                                 queue_work(ipoib_workqueue, &priv->cm.rx_reap_task);
 528                                 spin_unlock(&priv->lock);
 529                         }
 530                         goto done;
 531                 }
 532         }
 533
 534         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 535                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 536                         p->jiffies = jiffies;
 537                         /* Move this entry to list head, but do not re-add it
 538                          * if it has been moved out of list. */
 539                         if (p->state == IPOIB_CM_RX_LIVE)
 540                                 list_move(&p->list, &priv->cm.passive_ids);
 541                 }
 542         }
 543
 544         memcpy(&saverx, &rx_ring[wr_id], sizeof(saverx));
 545         newmb = ipoib_cm_alloc_rx_mb(priv, &rx_ring[wr_id]);
 546         if (unlikely(!newmb)) {
 547                 /*
 548                  * If we can't allocate a new RX buffer, dump
 549                  * this packet and reuse the old buffer.
 550                  */
 551                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 552                 if_inc_counter(dev, IFCOUNTER_IERRORS, 1);
 553                 memcpy(&rx_ring[wr_id], &saverx, sizeof(saverx));
 554                 goto repost;
 555         }
 556
 557         ipoib_cm_dma_unmap_rx(priv, &saverx);
 558
 559         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 560                        wc->byte_len, wc->slid);
 561
 562         ipoib_dma_mb(priv, mb, wc->byte_len);
 563
 564         if_inc_counter(dev, IFCOUNTER_IPACKETS, 1);
 565         if_inc_counter(dev, IFCOUNTER_IBYTES, mb->m_pkthdr.len);
 566
 567         mb->m_pkthdr.rcvif = dev;
 568         proto = *mtod(mb, uint16_t *);
 569         m_adj(mb, IPOIB_ENCAP_LEN);
 570
 571         IPOIB_MTAP_PROTO(dev, mb, proto);
 572         ipoib_demux(dev, mb, ntohs(proto));
 573
 574 repost:
 575         if (has_srq) {
 576                 if (unlikely(ipoib_cm_post_receive_srq(priv, wr_id)))
 577                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 578                                    "for buf %d\n", wr_id);
 579         } else {
 580                 if (unlikely(ipoib_cm_post_receive_nonsrq(priv, p,
 581                                                           &priv->cm.rx_wr,
 582                                                           priv->cm.rx_sge,
 583                                                           wr_id))) {
 584                         --p->recv_count;
 585                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 586                                    "for buf %d\n", wr_id);
 587                 }
 588         }
 589 done:
 590         CURVNET_RESTORE();
 591         return;
 592 }
 593
 594 static inline int post_send(struct ipoib_dev_priv *priv,
 595                             struct ipoib_cm_tx *tx,
 596                             struct ipoib_cm_tx_buf *tx_req,
 597                             unsigned int wr_id)
 598 {
 599         struct ib_send_wr *bad_wr;
 600         struct mbuf *mb = tx_req->mb;
 601         u64 *mapping = tx_req->mapping;
 602         struct mbuf *m;
 603         int i;
 604
 605         for (m = mb, i = 0; m != NULL; m = m->m_next, i++) {
 606                 priv->tx_sge[i].addr = mapping[i];
 607                 priv->tx_sge[i].length = m->m_len;
 608         }
 609         priv->tx_wr.wr.num_sge = i;
 610         priv->tx_wr.wr.wr_id = wr_id | IPOIB_OP_CM;
 611         priv->tx_wr.wr.opcode = IB_WR_SEND;
 612
 613         return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
 614 }
 615
 616 void ipoib_cm_send(struct ipoib_dev_priv *priv, struct mbuf *mb, struct ipoib_cm_tx *tx)
 617 {
 618         struct ipoib_cm_tx_buf *tx_req;
 619         struct ifnet *dev = priv->dev;
 620
 621         if (unlikely(priv->tx_outstanding > MAX_SEND_CQE)) {
 622                 while (ipoib_poll_tx(priv, false))
 623                         ;       /* nothing */
 624         }
 625
 626         m_adj(mb, sizeof(struct ipoib_pseudoheader));
 627         if (unlikely(mb->m_pkthdr.len > tx->mtu)) {
 628                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 629                            mb->m_pkthdr.len, tx->mtu);
 630                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 631                 ipoib_cm_mb_too_long(priv, mb, IPOIB_CM_MTU(tx->mtu));
 632                 return;
 633         }
 634
 635         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 636                        tx->tx_head, mb->m_pkthdr.len, tx->qp->qp_num);
 637
 638
 639         /*
 640          * We put the mb into the tx_ring _before_ we call post_send()
 641          * because it's entirely possible that the completion handler will
 642          * run before we execute anything after the post_send().  That
 643          * means we have to make sure everything is properly recorded and
 644          * our state is consistent before we call post_send().
 645          */
 646         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 647         tx_req->mb = mb;
 648         if (unlikely(ipoib_dma_map_tx(priv->ca, (struct ipoib_tx_buf *)tx_req,
 649             priv->cm.num_frags))) {
 650                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 651                 if (tx_req->mb)
 652                         m_freem(tx_req->mb);
 653                 return;
 654         }
 655
 656         if (unlikely(post_send(priv, tx, tx_req, tx->tx_head & (ipoib_sendq_size - 1)))) {
 657                 ipoib_warn(priv, "post_send failed\n");
 658                 if_inc_counter(dev, IFCOUNTER_OERRORS, 1);
 659                 ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
 660                 m_freem(mb);
 661         } else {
 662                 ++tx->tx_head;
 663
 664                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 665                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 666                                   tx->qp->qp_num);
 667                         if (ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP))
 668                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 669                         dev->if_drv_flags |= IFF_DRV_OACTIVE;
 670                 }
 671         }
 672
 673 }
 674
 675 void ipoib_cm_handle_tx_wc(struct ipoib_dev_priv *priv, struct ib_wc *wc)
 676 {
 677         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 678         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 679         struct ifnet *dev = priv->dev;
 680         struct ipoib_cm_tx_buf *tx_req;
 681
 682         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 683                        wr_id, wc->status);
 684
 685         if (unlikely(wr_id >= ipoib_sendq_size)) {
 686                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 687                            wr_id, ipoib_sendq_size);
 688                 return;
 689         }
 690
 691         tx_req = &tx->tx_ring[wr_id];
 692
 693         ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
 694
 695         /* FIXME: is this right? Shouldn't we only increment on success? */
 696         if_inc_counter(dev, IFCOUNTER_OPACKETS, 1);
 697
 698         m_freem(tx_req->mb);
 699
 700         ++tx->tx_tail;
 701         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 702             (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
 703             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 704                 dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
 705
 706         if (wc->status != IB_WC_SUCCESS &&
 707             wc->status != IB_WC_WR_FLUSH_ERR) {
 708                 struct ipoib_path *path;
 709
 710                 ipoib_dbg(priv, "failed cm send event "
 711                            "(status=%d, wrid=%d vend_err %x)\n",
 712                            wc->status, wr_id, wc->vendor_err);
 713
 714                 path = tx->path;
 715
 716                 if (path) {
 717                         path->cm = NULL;
 718                         rb_erase(&path->rb_node, &priv->path_tree);
 719                         list_del(&path->list);
 720                 }
 721
 722                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 723                         list_move(&tx->list, &priv->cm.reap_list);
 724                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
 725                 }
 726
 727                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 728         }
 729
 730 }
 731
 732 int ipoib_cm_dev_open(struct ipoib_dev_priv *priv)
 733 {
 734         int ret;
 735
 736         if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)))
 737                 return 0;
 738
 739         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, priv);
 740         if (IS_ERR(priv->cm.id)) {
 741                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 742                 ret = PTR_ERR(priv->cm.id);
 743                 goto err_cm;
 744         }
 745
 746         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num), 0);
 747         if (ret) {
 748                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 749                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 750                 goto err_listen;
 751         }
 752
 753         return 0;
 754
 755 err_listen:
 756         ib_destroy_cm_id(priv->cm.id);
 757 err_cm:
 758         priv->cm.id = NULL;
 759         return ret;
 760 }
 761
 762 static void ipoib_cm_free_rx_reap_list(struct ipoib_dev_priv *priv)
 763 {
 764         struct ipoib_cm_rx *rx, *n;
 765         LIST_HEAD(list);
 766
 767         spin_lock_irq(&priv->lock);
 768         list_splice_init(&priv->cm.rx_reap_list, &list);
 769         spin_unlock_irq(&priv->lock);
 770
 771         list_for_each_entry_safe(rx, n, &list, list) {
 772                 ib_destroy_cm_id(rx->id);
 773                 ib_destroy_qp(rx->qp);
 774                 if (!ipoib_cm_has_srq(priv)) {
 775                         ipoib_cm_free_rx_ring(priv, rx->rx_ring);
 776                         spin_lock_irq(&priv->lock);
 777                         --priv->cm.nonsrq_conn_qp;
 778                         spin_unlock_irq(&priv->lock);
 779                 }
 780                 kfree(rx);
 781         }
 782 }
 783
 784 void ipoib_cm_dev_stop(struct ipoib_dev_priv *priv)
 785 {
 786         struct ipoib_cm_rx *p;
 787         unsigned long begin;
 788         int ret;
 789
 790         if (!IPOIB_CM_SUPPORTED(IF_LLADDR(priv->dev)) || !priv->cm.id)
 791                 return;
 792
 793         ib_destroy_cm_id(priv->cm.id);
 794         priv->cm.id = NULL;
 795
 796         cancel_work_sync(&priv->cm.rx_reap_task);
 797
 798         spin_lock_irq(&priv->lock);
 799         while (!list_empty(&priv->cm.passive_ids)) {
 800                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 801                 list_move(&p->list, &priv->cm.rx_error_list);
 802                 p->state = IPOIB_CM_RX_ERROR;
 803                 spin_unlock_irq(&priv->lock);
 804                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 805                 if (ret)
 806                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 807                 spin_lock_irq(&priv->lock);
 808         }
 809
 810         /* Wait for all RX to be drained */
 811         begin = jiffies;
 812
 813         while (!list_empty(&priv->cm.rx_error_list) ||
 814                !list_empty(&priv->cm.rx_flush_list) ||
 815                !list_empty(&priv->cm.rx_drain_list)) {
 816                 if (time_after(jiffies, begin + 5 * HZ)) {
 817                         ipoib_warn(priv, "RX drain timing out\n");
 818
 819                         /*
 820                          * assume the HW is wedged and just free up everything.
 821                          */
 822                         list_splice_init(&priv->cm.rx_flush_list,
 823                                          &priv->cm.rx_reap_list);
 824                         list_splice_init(&priv->cm.rx_error_list,
 825                                          &priv->cm.rx_reap_list);
 826                         list_splice_init(&priv->cm.rx_drain_list,
 827                                          &priv->cm.rx_reap_list);
 828                         break;
 829                 }
 830                 spin_unlock_irq(&priv->lock);
 831                 msleep(1);
 832                 ipoib_drain_cq(priv);
 833                 spin_lock_irq(&priv->lock);
 834         }
 835
 836         spin_unlock_irq(&priv->lock);
 837
 838         ipoib_cm_free_rx_reap_list(priv);
 839
 840         cancel_delayed_work_sync(&priv->cm.stale_task);
 841 }
 842
 843 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 844 {
 845         struct ipoib_cm_tx *p = cm_id->context;
 846         struct ipoib_dev_priv *priv = p->priv;
 847         struct ipoib_cm_data *data = event->private_data;
 848         struct ifqueue mbqueue;
 849         struct ib_qp_attr qp_attr;
 850         int qp_attr_mask, ret;
 851         struct mbuf *mb;
 852
 853         ipoib_dbg(priv, "cm rep handler\n");
 854         p->mtu = be32_to_cpu(data->mtu);
 855
 856         if (p->mtu <= IPOIB_ENCAP_LEN) {
 857                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 858                            p->mtu, IPOIB_ENCAP_LEN);
 859                 return -EINVAL;
 860         }
 861
 862         qp_attr.qp_state = IB_QPS_RTR;
 863         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 864         if (ret) {
 865                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 866                 return ret;
 867         }
 868
 869         qp_attr.rq_psn = 0 /* FIXME */;
 870         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 871         if (ret) {
 872                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 873                 return ret;
 874         }
 875
 876         qp_attr.qp_state = IB_QPS_RTS;
 877         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 878         if (ret) {
 879                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 880                 return ret;
 881         }
 882         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 883         if (ret) {
 884                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 885                 return ret;
 886         }
 887
 888         bzero(&mbqueue, sizeof(mbqueue));
 889
 890         spin_lock_irq(&priv->lock);
 891         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 892         if (p->path)
 893                 for (;;) {
 894                         _IF_DEQUEUE(&p->path->queue, mb);
 895                         if (mb == NULL)
 896                                 break;
 897                         _IF_ENQUEUE(&mbqueue, mb);
 898                 }
 899         spin_unlock_irq(&priv->lock);
 900
 901         for (;;) {
 902                 struct ifnet *dev = p->priv->dev;
 903                 _IF_DEQUEUE(&mbqueue, mb);
 904                 if (mb == NULL)
 905                         break;
 906                 mb->m_pkthdr.rcvif = dev;
 907                 if (dev->if_transmit(dev, mb))
 908                         ipoib_warn(priv, "dev_queue_xmit failed "
 909                                    "to requeue packet\n");
 910         }
 911
 912         ret = ib_send_cm_rtu(cm_id, NULL, 0);
 913         if (ret) {
 914                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
 915                 return ret;
 916         }
 917         return 0;
 918 }
 919
 920 static struct ib_qp *ipoib_cm_create_tx_qp(struct ipoib_dev_priv *priv,
 921     struct ipoib_cm_tx *tx)
 922 {
 923         struct ib_qp_init_attr attr = {
 924                 .send_cq                = priv->send_cq,
 925                 .recv_cq                = priv->recv_cq,
 926                 .srq                    = priv->cm.srq,
 927                 .cap.max_send_wr        = ipoib_sendq_size,
 928                 .cap.max_send_sge       = priv->cm.num_frags,
 929                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
 930                 .qp_type                = IB_QPT_RC,
 931                 .qp_context             = tx
 932         };
 933
 934         return ib_create_qp(priv->pd, &attr);
 935 }
 936
 937 static int ipoib_cm_send_req(struct ipoib_dev_priv *priv,
 938                              struct ib_cm_id *id, struct ib_qp *qp,
 939                              u32 qpn,
 940                              struct ib_sa_path_rec *pathrec)
 941 {
 942         struct ipoib_cm_data data = {};
 943         struct ib_cm_req_param req = {};
 944
 945         ipoib_dbg(priv, "cm send req\n");
 946
 947         data.qpn = cpu_to_be32(priv->qp->qp_num);
 948         data.mtu = cpu_to_be32(priv->cm.max_cm_mtu);
 949
 950         req.primary_path                = pathrec;
 951         req.alternate_path              = NULL;
 952         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
 953         req.qp_num                      = qp->qp_num;
 954         req.qp_type                     = qp->qp_type;
 955         req.private_data                = &data;
 956         req.private_data_len            = sizeof data;
 957         req.flow_control                = 0;
 958
 959         req.starting_psn                = 0; /* FIXME */
 960
 961         /*
 962          * Pick some arbitrary defaults here; we could make these
 963          * module parameters if anyone cared about setting them.
 964          */
 965         req.responder_resources         = 4;
 966         req.remote_cm_response_timeout  = 20;
 967         req.local_cm_response_timeout   = 20;
 968         req.retry_count                 = 0; /* RFC draft warns against retries */
 969         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
 970         req.max_cm_retries              = 15;
 971         req.srq                         = ipoib_cm_has_srq(priv);
 972         return ib_send_cm_req(id, &req);
 973 }
 974
 975 static int ipoib_cm_modify_tx_init(struct ipoib_dev_priv *priv,
 976                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
 977 {
 978         struct ib_qp_attr qp_attr;
 979         int qp_attr_mask, ret;
 980         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
 981         if (ret) {
 982                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
 983                 return ret;
 984         }
 985
 986         qp_attr.qp_state = IB_QPS_INIT;
 987         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
 988         qp_attr.port_num = priv->port;
 989         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
 990
 991         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 992         if (ret) {
 993                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
 994                 return ret;
 995         }
 996         return 0;
 997 }
 998
 999 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1000                             struct ib_sa_path_rec *pathrec)
1001 {
1002         struct ipoib_dev_priv *priv = p->priv;
1003         int ret;
1004
1005         p->tx_ring = kzalloc(ipoib_sendq_size * sizeof *p->tx_ring, GFP_KERNEL);
1006         if (!p->tx_ring) {
1007                 ipoib_warn(priv, "failed to allocate tx ring\n");
1008                 ret = -ENOMEM;
1009                 goto err_tx;
1010         }
1011         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1012
1013         p->qp = ipoib_cm_create_tx_qp(p->priv, p);
1014         if (IS_ERR(p->qp)) {
1015                 ret = PTR_ERR(p->qp);
1016                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1017                 goto err_qp;
1018         }
1019
1020         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1021         if (IS_ERR(p->id)) {
1022                 ret = PTR_ERR(p->id);
1023                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1024                 goto err_id;
1025         }
1026
1027         ret = ipoib_cm_modify_tx_init(p->priv, p->id,  p->qp);
1028         if (ret) {
1029                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1030                 goto err_modify;
1031         }
1032
1033         ret = ipoib_cm_send_req(p->priv, p->id, p->qp, qpn, pathrec);
1034         if (ret) {
1035                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1036                 goto err_send_cm;
1037         }
1038
1039         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1040                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1041
1042         return 0;
1043
1044 err_send_cm:
1045 err_modify:
1046         ib_destroy_cm_id(p->id);
1047 err_id:
1048         p->id = NULL;
1049         ib_destroy_qp(p->qp);
1050 err_qp:
1051         p->qp = NULL;
1052         kfree(p->tx_ring);
1053 err_tx:
1054         return ret;
1055 }
1056
1057 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1058 {
1059         struct ipoib_dev_priv *priv = p->priv;
1060         struct ifnet *dev = priv->dev;
1061         struct ipoib_cm_tx_buf *tx_req;
1062         unsigned long begin;
1063
1064         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1065                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1066
1067         if (p->path)
1068                 ipoib_path_free(priv, p->path);
1069
1070         if (p->id)
1071                 ib_destroy_cm_id(p->id);
1072
1073         if (p->tx_ring) {
1074                 /* Wait for all sends to complete */
1075                 begin = jiffies;
1076                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1077                         if (time_after(jiffies, begin + 5 * HZ)) {
1078                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1079                                            p->tx_head - p->tx_tail);
1080                                 goto timeout;
1081                         }
1082
1083                         msleep(1);
1084                 }
1085         }
1086
1087 timeout:
1088
1089         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1090                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1091                 ipoib_dma_unmap_tx(priv->ca, (struct ipoib_tx_buf *)tx_req);
1092                 m_freem(tx_req->mb);
1093                 ++p->tx_tail;
1094                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1095                     (dev->if_drv_flags & IFF_DRV_OACTIVE) != 0 &&
1096                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1097                         dev->if_drv_flags &= ~IFF_DRV_OACTIVE;
1098         }
1099
1100         if (p->qp)
1101                 ib_destroy_qp(p->qp);
1102
1103         kfree(p->tx_ring);
1104         kfree(p);
1105 }
1106
1107 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1108                                struct ib_cm_event *event)
1109 {
1110         struct ipoib_cm_tx *tx = cm_id->context;
1111         struct ipoib_dev_priv *priv = tx->priv;
1112         struct ipoib_path *path;
1113         unsigned long flags;
1114         int ret;
1115
1116         switch (event->event) {
1117         case IB_CM_DREQ_RECEIVED:
1118                 ipoib_dbg(priv, "DREQ received.\n");
1119                 ib_send_cm_drep(cm_id, NULL, 0);
1120                 break;
1121         case IB_CM_REP_RECEIVED:
1122                 ipoib_dbg(priv, "REP received.\n");
1123                 ret = ipoib_cm_rep_handler(cm_id, event);
1124                 if (ret)
1125                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1126                                        NULL, 0, NULL, 0);
1127                 break;
1128         case IB_CM_REQ_ERROR:
1129         case IB_CM_REJ_RECEIVED:
1130         case IB_CM_TIMEWAIT_EXIT:
1131                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1132                 spin_lock_irqsave(&priv->lock, flags);
1133                 path = tx->path;
1134
1135                 if (path) {
1136                         path->cm = NULL;
1137                         tx->path = NULL;
1138                         rb_erase(&path->rb_node, &priv->path_tree);
1139                         list_del(&path->list);
1140                 }
1141
1142                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1143                         list_move(&tx->list, &priv->cm.reap_list);
1144                         queue_work(ipoib_workqueue, &priv->cm.reap_task);
1145                 }
1146
1147                 spin_unlock_irqrestore(&priv->lock, flags);
1148                 if (path)
1149                         ipoib_path_free(tx->priv, path);
1150                 break;
1151         default:
1152                 break;
1153         }
1154
1155         return 0;
1156 }
1157
1158 struct ipoib_cm_tx *ipoib_cm_create_tx(struct ipoib_dev_priv *priv,
1159     struct ipoib_path *path)
1160 {
1161         struct ipoib_cm_tx *tx;
1162
1163         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1164         if (!tx)
1165                 return NULL;
1166
1167         ipoib_dbg(priv, "Creating cm tx\n");
1168         path->cm = tx;
1169         tx->path = path;
1170         tx->priv = priv;
1171         list_add(&tx->list, &priv->cm.start_list);
1172         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1173         queue_work(ipoib_workqueue, &priv->cm.start_task);
1174         return tx;
1175 }
1176
1177 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1178 {
1179         struct ipoib_dev_priv *priv = tx->priv;
1180         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1181                 spin_lock(&priv->lock);
1182                 list_move(&tx->list, &priv->cm.reap_list);
1183                 spin_unlock(&priv->lock);
1184                 queue_work(ipoib_workqueue, &priv->cm.reap_task);
1185                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1186                           tx->path->pathrec.dgid.raw);
1187                 tx->path = NULL;
1188         }
1189 }
1190
1191 static void ipoib_cm_tx_start(struct work_struct *work)
1192 {
1193         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1194                                                    cm.start_task);
1195         struct ipoib_path *path;
1196         struct ipoib_cm_tx *p;
1197         unsigned long flags;
1198         int ret;
1199
1200         struct ib_sa_path_rec pathrec;
1201         u32 qpn;
1202
1203         ipoib_dbg(priv, "cm start task\n");
1204         spin_lock_irqsave(&priv->lock, flags);
1205
1206         while (!list_empty(&priv->cm.start_list)) {
1207                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1208                 list_del_init(&p->list);
1209                 path = p->path;
1210                 qpn = IPOIB_QPN(path->hwaddr);
1211                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1212
1213                 spin_unlock_irqrestore(&priv->lock, flags);
1214
1215                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1216
1217                 spin_lock_irqsave(&priv->lock, flags);
1218
1219                 if (ret) {
1220                         path = p->path;
1221                         if (path) {
1222                                 path->cm = NULL;
1223                                 rb_erase(&path->rb_node, &priv->path_tree);
1224                                 list_del(&path->list);
1225                                 ipoib_path_free(priv, path);
1226                         }
1227                         list_del(&p->list);
1228                         kfree(p);
1229                 }
1230         }
1231
1232         spin_unlock_irqrestore(&priv->lock, flags);
1233 }
1234
1235 static void ipoib_cm_tx_reap(struct work_struct *work)
1236 {
1237         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1238                                                    cm.reap_task);
1239         struct ipoib_cm_tx *p;
1240         unsigned long flags;
1241
1242         spin_lock_irqsave(&priv->lock, flags);
1243
1244         while (!list_empty(&priv->cm.reap_list)) {
1245                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1246                 list_del(&p->list);
1247                 spin_unlock_irqrestore(&priv->lock, flags);
1248                 ipoib_cm_tx_destroy(p);
1249                 spin_lock_irqsave(&priv->lock, flags);
1250         }
1251
1252         spin_unlock_irqrestore(&priv->lock, flags);
1253 }
1254
1255 static void ipoib_cm_mb_reap(struct work_struct *work)
1256 {
1257         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1258                                                    cm.mb_task);
1259         struct mbuf *mb;
1260         unsigned long flags;
1261 #if defined(INET) || defined(INET6)
1262         unsigned mtu = priv->mcast_mtu;
1263 #endif
1264         uint16_t proto;
1265
1266         spin_lock_irqsave(&priv->lock, flags);
1267
1268         for (;;) {
1269                 IF_DEQUEUE(&priv->cm.mb_queue, mb);
1270                 if (mb == NULL)
1271                         break;
1272                 spin_unlock_irqrestore(&priv->lock, flags);
1273
1274                 proto = htons(*mtod(mb, uint16_t *));
1275                 m_adj(mb, IPOIB_ENCAP_LEN);
1276                 switch (proto) {
1277 #if defined(INET)
1278                 case ETHERTYPE_IP:
1279                         icmp_error(mb, ICMP_UNREACH, ICMP_UNREACH_NEEDFRAG, 0, mtu);
1280                         break;
1281 #endif
1282 #if defined(INET6)
1283                 case ETHERTYPE_IPV6:
1284                         icmp6_error(mb, ICMP6_PACKET_TOO_BIG, 0, mtu);
1285                         break;
1286 #endif
1287                 default:
1288                         m_freem(mb);
1289                 }
1290
1291                 spin_lock_irqsave(&priv->lock, flags);
1292         }
1293
1294         spin_unlock_irqrestore(&priv->lock, flags);
1295 }
1296
1297 void
1298 ipoib_cm_mb_too_long(struct ipoib_dev_priv *priv, struct mbuf *mb, unsigned int mtu)
1299 {
1300         int e = priv->cm.mb_queue.ifq_len;
1301
1302         IF_ENQUEUE(&priv->cm.mb_queue, mb);
1303         if (e == 0)
1304                 queue_work(ipoib_workqueue, &priv->cm.mb_task);
1305 }
1306
1307 static void ipoib_cm_rx_reap(struct work_struct *work)
1308 {
1309         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1310                                                 cm.rx_reap_task));
1311 }
1312
1313 static void ipoib_cm_stale_task(struct work_struct *work)
1314 {
1315         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1316                                                    cm.stale_task.work);
1317         struct ipoib_cm_rx *p;
1318         int ret;
1319
1320         spin_lock_irq(&priv->lock);
1321         while (!list_empty(&priv->cm.passive_ids)) {
1322                 /* List is sorted by LRU, start from tail,
1323                  * stop when we see a recently used entry */
1324                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1325                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1326                         break;
1327                 list_move(&p->list, &priv->cm.rx_error_list);
1328                 p->state = IPOIB_CM_RX_ERROR;
1329                 spin_unlock_irq(&priv->lock);
1330                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1331                 if (ret)
1332                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1333                 spin_lock_irq(&priv->lock);
1334         }
1335
1336         if (!list_empty(&priv->cm.passive_ids))
1337                 queue_delayed_work(ipoib_workqueue,
1338                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1339         spin_unlock_irq(&priv->lock);
1340 }
1341
1342
1343 static void ipoib_cm_create_srq(struct ipoib_dev_priv *priv, int max_sge)
1344 {
1345         struct ib_srq_init_attr srq_init_attr = {
1346                 .attr = {
1347                         .max_wr  = ipoib_recvq_size,
1348                         .max_sge = max_sge
1349                 }
1350         };
1351
1352         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1353         if (IS_ERR(priv->cm.srq)) {
1354                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1355                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1356                                priv->ca->name, PTR_ERR(priv->cm.srq));
1357                 priv->cm.srq = NULL;
1358                 return;
1359         }
1360
1361         priv->cm.srq_ring = kzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring, GFP_KERNEL);
1362         if (!priv->cm.srq_ring) {
1363                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1364                        priv->ca->name, ipoib_recvq_size);
1365                 ib_destroy_srq(priv->cm.srq);
1366                 priv->cm.srq = NULL;
1367                 return;
1368         }
1369
1370         memset(priv->cm.srq_ring, 0, ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1371 }
1372
1373 int ipoib_cm_dev_init(struct ipoib_dev_priv *priv)
1374 {
1375         struct ifnet *dev = priv->dev;
1376         int i;
1377         int max_srq_sge;
1378
1379         INIT_LIST_HEAD(&priv->cm.passive_ids);
1380         INIT_LIST_HEAD(&priv->cm.reap_list);
1381         INIT_LIST_HEAD(&priv->cm.start_list);
1382         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1383         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1384         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1385         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1386         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1387         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1388         INIT_WORK(&priv->cm.mb_task, ipoib_cm_mb_reap);
1389         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1390         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1391
1392         bzero(&priv->cm.mb_queue, sizeof(priv->cm.mb_queue));
1393         mtx_init(&priv->cm.mb_queue.ifq_mtx,
1394             dev->if_xname, "if send queue", MTX_DEF);
1395
1396         max_srq_sge = priv->ca->attrs.max_srq_sge;
1397
1398         ipoib_dbg(priv, "max_srq_sge=%d\n", max_srq_sge);
1399
1400         max_srq_sge = min_t(int, IPOIB_CM_RX_SG, max_srq_sge);
1401         ipoib_cm_create_srq(priv, max_srq_sge);
1402         if (ipoib_cm_has_srq(priv)) {
1403                 priv->cm.max_cm_mtu = max_srq_sge * MJUMPAGESIZE;
1404                 priv->cm.num_frags  = max_srq_sge;
1405                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1406                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1407         } else {
1408                 priv->cm.max_cm_mtu = IPOIB_CM_MAX_MTU;
1409                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1410         }
1411
1412         ipoib_cm_init_rx_wr(priv, &priv->cm.rx_wr, priv->cm.rx_sge);
1413
1414         if (ipoib_cm_has_srq(priv)) {
1415                 for (i = 0; i < ipoib_recvq_size; ++i) {
1416                         if (!ipoib_cm_alloc_rx_mb(priv, &priv->cm.srq_ring[i])) {
1417                                 ipoib_warn(priv, "failed to allocate "
1418                                            "receive buffer %d\n", i);
1419                                 ipoib_cm_dev_cleanup(priv);
1420                                 return -ENOMEM;
1421                         }
1422
1423                         if (ipoib_cm_post_receive_srq(priv, i)) {
1424                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1425                                            "failed for buf %d\n", i);
1426                                 ipoib_cm_dev_cleanup(priv);
1427                                 return -EIO;
1428                         }
1429                 }
1430         }
1431
1432         IF_LLADDR(priv->dev)[0] = IPOIB_FLAGS_RC;
1433         return 0;
1434 }
1435
1436 void ipoib_cm_dev_cleanup(struct ipoib_dev_priv *priv)
1437 {
1438         int ret;
1439
1440         if (!priv->cm.srq)
1441                 return;
1442
1443         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1444
1445         ret = ib_destroy_srq(priv->cm.srq);
1446         if (ret)
1447                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1448
1449         priv->cm.srq = NULL;
1450         if (!priv->cm.srq_ring)
1451                 return;
1452
1453         ipoib_cm_free_rx_ring(priv, priv->cm.srq_ring);
1454         priv->cm.srq_ring = NULL;
1455
1456         mtx_destroy(&priv->cm.mb_queue.ifq_mtx);
1457 }
1458
1459 #endif /* CONFIG_INFINIBAND_IPOIB_CM */