kernel/drivers/infiniband/ulp/ipoib/ipoib_cm.c

   1 /*
   2  * Copyright (c) 2006 Mellanox Technologies. All rights reserved
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32
  33 #include <rdma/ib_cm.h>
  34 #include <net/dst.h>
  35 #include <net/icmp.h>
  36 #include <linux/icmpv6.h>
  37 #include <linux/delay.h>
  38 #include <linux/slab.h>
  39 #include <linux/vmalloc.h>
  40 #include <linux/moduleparam.h>
  41
  42 #include "ipoib.h"
  43
  44 int ipoib_max_conn_qp = 128;
  45
  46 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  47 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  48                  "Max number of connected-mode QPs per interface "
  49                  "(applied only if shared receive queue is not available)");
  50
  51 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  52 static int data_debug_level;
  53
  54 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  55 MODULE_PARM_DESC(cm_data_debug_level,
  56                  "Enable data path debug tracing for connected mode if > 0");
  57 #endif
  58
  59 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  60
  61 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  62 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  63 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  64 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  65
  66 #define IPOIB_CM_RX_RESERVE     (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
  67
  68 static struct ib_qp_attr ipoib_cm_err_attr = {
  69         .qp_state = IB_QPS_ERR
  70 };
  71
  72 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  73
  74 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  75         .wr_id = IPOIB_CM_RX_DRAIN_WRID,
  76         .opcode = IB_WR_SEND,
  77 };
  78
  79 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  80                                struct ib_cm_event *event);
  81
  82 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  83                                   u64 mapping[IPOIB_CM_RX_SG])
  84 {
  85         int i;
  86
  87         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  88
  89         for (i = 0; i < frags; ++i)
  90                 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  91 }
  92
  93 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  94 {
  95         struct ipoib_dev_priv *priv = netdev_priv(dev);
  96         struct ib_recv_wr *bad_wr;
  97         int i, ret;
  98
  99         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 100
 101         for (i = 0; i < priv->cm.num_frags; ++i)
 102                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 103
 104         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 105         if (unlikely(ret)) {
 106                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 107                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 108                                       priv->cm.srq_ring[id].mapping);
 109                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 110                 priv->cm.srq_ring[id].skb = NULL;
 111         }
 112
 113         return ret;
 114 }
 115
 116 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 117                                         struct ipoib_cm_rx *rx,
 118                                         struct ib_recv_wr *wr,
 119                                         struct ib_sge *sge, int id)
 120 {
 121         struct ipoib_dev_priv *priv = netdev_priv(dev);
 122         struct ib_recv_wr *bad_wr;
 123         int i, ret;
 124
 125         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 126
 127         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 128                 sge[i].addr = rx->rx_ring[id].mapping[i];
 129
 130         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 131         if (unlikely(ret)) {
 132                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 133                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 134                                       rx->rx_ring[id].mapping);
 135                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 136                 rx->rx_ring[id].skb = NULL;
 137         }
 138
 139         return ret;
 140 }
 141
 142 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 143                                              struct ipoib_cm_rx_buf *rx_ring,
 144                                              int id, int frags,
 145                                              u64 mapping[IPOIB_CM_RX_SG],
 146                                              gfp_t gfp)
 147 {
 148         struct ipoib_dev_priv *priv = netdev_priv(dev);
 149         struct sk_buff *skb;
 150         int i;
 151
 152         skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
 153         if (unlikely(!skb))
 154                 return NULL;
 155
 156         /*
 157          * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
 158          * IP header to a multiple of 16.
 159          */
 160         skb_reserve(skb, IPOIB_CM_RX_RESERVE);
 161
 162         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 163                                        DMA_FROM_DEVICE);
 164         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 165                 dev_kfree_skb_any(skb);
 166                 return NULL;
 167         }
 168
 169         for (i = 0; i < frags; i++) {
 170                 struct page *page = alloc_page(gfp);
 171
 172                 if (!page)
 173                         goto partial_error;
 174                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 175
 176                 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
 177                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 178                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 179                         goto partial_error;
 180         }
 181
 182         rx_ring[id].skb = skb;
 183         return skb;
 184
 185 partial_error:
 186
 187         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 188
 189         for (; i > 0; --i)
 190                 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 191
 192         dev_kfree_skb_any(skb);
 193         return NULL;
 194 }
 195
 196 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 197                                   struct ipoib_cm_rx_buf *rx_ring)
 198 {
 199         struct ipoib_dev_priv *priv = netdev_priv(dev);
 200         int i;
 201
 202         for (i = 0; i < ipoib_recvq_size; ++i)
 203                 if (rx_ring[i].skb) {
 204                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 205                                               rx_ring[i].mapping);
 206                         dev_kfree_skb_any(rx_ring[i].skb);
 207                 }
 208
 209         vfree(rx_ring);
 210 }
 211
 212 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 213 {
 214         struct ib_send_wr *bad_wr;
 215         struct ipoib_cm_rx *p;
 216
 217         /* We only reserved 1 extra slot in CQ for drain WRs, so
 218          * make sure we have at most 1 outstanding WR. */
 219         if (list_empty(&priv->cm.rx_flush_list) ||
 220             !list_empty(&priv->cm.rx_drain_list))
 221                 return;
 222
 223         /*
 224          * QPs on flush list are error state.  This way, a "flush
 225          * error" WC will be immediately generated for each WR we post.
 226          */
 227         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 228         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 229                 ipoib_warn(priv, "failed to post drain wr\n");
 230
 231         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 232 }
 233
 234 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 235 {
 236         struct ipoib_cm_rx *p = ctx;
 237         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 238         unsigned long flags;
 239
 240         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 241                 return;
 242
 243         spin_lock_irqsave(&priv->lock, flags);
 244         list_move(&p->list, &priv->cm.rx_flush_list);
 245         p->state = IPOIB_CM_RX_FLUSH;
 246         ipoib_cm_start_rx_drain(priv);
 247         spin_unlock_irqrestore(&priv->lock, flags);
 248 }
 249
 250 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 251                                            struct ipoib_cm_rx *p)
 252 {
 253         struct ipoib_dev_priv *priv = netdev_priv(dev);
 254         struct ib_qp_init_attr attr = {
 255                 .event_handler = ipoib_cm_rx_event_handler,
 256                 .send_cq = priv->recv_cq, /* For drain WR */
 257                 .recv_cq = priv->recv_cq,
 258                 .srq = priv->cm.srq,
 259                 .cap.max_send_wr = 1, /* For drain WR */
 260                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 261                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 262                 .qp_type = IB_QPT_RC,
 263                 .qp_context = p,
 264         };
 265
 266         if (!ipoib_cm_has_srq(dev)) {
 267                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 268                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 269         }
 270
 271         return ib_create_qp(priv->pd, &attr);
 272 }
 273
 274 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 275                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 276                                  unsigned psn)
 277 {
 278         struct ipoib_dev_priv *priv = netdev_priv(dev);
 279         struct ib_qp_attr qp_attr;
 280         int qp_attr_mask, ret;
 281
 282         qp_attr.qp_state = IB_QPS_INIT;
 283         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 284         if (ret) {
 285                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 286                 return ret;
 287         }
 288         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 289         if (ret) {
 290                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 291                 return ret;
 292         }
 293         qp_attr.qp_state = IB_QPS_RTR;
 294         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 295         if (ret) {
 296                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 297                 return ret;
 298         }
 299         qp_attr.rq_psn = psn;
 300         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 301         if (ret) {
 302                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 303                 return ret;
 304         }
 305
 306         /*
 307          * Current Mellanox HCA firmware won't generate completions
 308          * with error for drain WRs unless the QP has been moved to
 309          * RTS first. This work-around leaves a window where a QP has
 310          * moved to error asynchronously, but this will eventually get
 311          * fixed in firmware, so let's not error out if modify QP
 312          * fails.
 313          */
 314         qp_attr.qp_state = IB_QPS_RTS;
 315         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 316         if (ret) {
 317                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 318                 return 0;
 319         }
 320         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 321         if (ret) {
 322                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 323                 return 0;
 324         }
 325
 326         return 0;
 327 }
 328
 329 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 330                                 struct ib_recv_wr *wr,
 331                                 struct ib_sge *sge)
 332 {
 333         struct ipoib_dev_priv *priv = netdev_priv(dev);
 334         int i;
 335
 336         for (i = 0; i < priv->cm.num_frags; ++i)
 337                 sge[i].lkey = priv->pd->local_dma_lkey;
 338
 339         sge[0].length = IPOIB_CM_HEAD_SIZE;
 340         for (i = 1; i < priv->cm.num_frags; ++i)
 341                 sge[i].length = PAGE_SIZE;
 342
 343         wr->next    = NULL;
 344         wr->sg_list = sge;
 345         wr->num_sge = priv->cm.num_frags;
 346 }
 347
 348 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 349                                    struct ipoib_cm_rx *rx)
 350 {
 351         struct ipoib_dev_priv *priv = netdev_priv(dev);
 352         struct {
 353                 struct ib_recv_wr wr;
 354                 struct ib_sge sge[IPOIB_CM_RX_SG];
 355         } *t;
 356         int ret;
 357         int i;
 358
 359         rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 360         if (!rx->rx_ring) {
 361                 printk(KERN_WARNING "%s: failed to allocate CM non-SRQ ring (%d entries)\n",
 362                        priv->ca->name, ipoib_recvq_size);
 363                 return -ENOMEM;
 364         }
 365
 366         t = kmalloc(sizeof *t, GFP_KERNEL);
 367         if (!t) {
 368                 ret = -ENOMEM;
 369                 goto err_free;
 370         }
 371
 372         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 373
 374         spin_lock_irq(&priv->lock);
 375
 376         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 377                 spin_unlock_irq(&priv->lock);
 378                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 379                 ret = -EINVAL;
 380                 goto err_free;
 381         } else
 382                 ++priv->cm.nonsrq_conn_qp;
 383
 384         spin_unlock_irq(&priv->lock);
 385
 386         for (i = 0; i < ipoib_recvq_size; ++i) {
 387                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 388                                            rx->rx_ring[i].mapping,
 389                                            GFP_KERNEL)) {
 390                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 391                         ret = -ENOMEM;
 392                         goto err_count;
 393                 }
 394                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 395                 if (ret) {
 396                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 397                                    "failed for buf %d\n", i);
 398                         ret = -EIO;
 399                         goto err_count;
 400                 }
 401         }
 402
 403         rx->recv_count = ipoib_recvq_size;
 404
 405         kfree(t);
 406
 407         return 0;
 408
 409 err_count:
 410         spin_lock_irq(&priv->lock);
 411         --priv->cm.nonsrq_conn_qp;
 412         spin_unlock_irq(&priv->lock);
 413
 414 err_free:
 415         kfree(t);
 416         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 417
 418         return ret;
 419 }
 420
 421 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 422                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 423                              unsigned psn)
 424 {
 425         struct ipoib_dev_priv *priv = netdev_priv(dev);
 426         struct ipoib_cm_data data = {};
 427         struct ib_cm_rep_param rep = {};
 428
 429         data.qpn = cpu_to_be32(priv->qp->qp_num);
 430         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 431
 432         rep.private_data = &data;
 433         rep.private_data_len = sizeof data;
 434         rep.flow_control = 0;
 435         rep.rnr_retry_count = req->rnr_retry_count;
 436         rep.srq = ipoib_cm_has_srq(dev);
 437         rep.qp_num = qp->qp_num;
 438         rep.starting_psn = psn;
 439         return ib_send_cm_rep(cm_id, &rep);
 440 }
 441
 442 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 443 {
 444         struct net_device *dev = cm_id->context;
 445         struct ipoib_dev_priv *priv = netdev_priv(dev);
 446         struct ipoib_cm_rx *p;
 447         unsigned psn;
 448         int ret;
 449
 450         ipoib_dbg(priv, "REQ arrived\n");
 451         p = kzalloc(sizeof *p, GFP_KERNEL);
 452         if (!p)
 453                 return -ENOMEM;
 454         p->dev = dev;
 455         p->id = cm_id;
 456         cm_id->context = p;
 457         p->state = IPOIB_CM_RX_LIVE;
 458         p->jiffies = jiffies;
 459         INIT_LIST_HEAD(&p->list);
 460
 461         p->qp = ipoib_cm_create_rx_qp(dev, p);
 462         if (IS_ERR(p->qp)) {
 463                 ret = PTR_ERR(p->qp);
 464                 goto err_qp;
 465         }
 466
 467         psn = prandom_u32() & 0xffffff;
 468         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 469         if (ret)
 470                 goto err_modify;
 471
 472         if (!ipoib_cm_has_srq(dev)) {
 473                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 474                 if (ret)
 475                         goto err_modify;
 476         }
 477
 478         spin_lock_irq(&priv->lock);
 479         queue_delayed_work(priv->wq,
 480                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 481         /* Add this entry to passive ids list head, but do not re-add it
 482          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 483         p->jiffies = jiffies;
 484         if (p->state == IPOIB_CM_RX_LIVE)
 485                 list_move(&p->list, &priv->cm.passive_ids);
 486         spin_unlock_irq(&priv->lock);
 487
 488         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 489         if (ret) {
 490                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 491                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 492                         ipoib_warn(priv, "unable to move qp to error state\n");
 493         }
 494         return 0;
 495
 496 err_modify:
 497         ib_destroy_qp(p->qp);
 498 err_qp:
 499         kfree(p);
 500         return ret;
 501 }
 502
 503 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 504                                struct ib_cm_event *event)
 505 {
 506         struct ipoib_cm_rx *p;
 507         struct ipoib_dev_priv *priv;
 508
 509         switch (event->event) {
 510         case IB_CM_REQ_RECEIVED:
 511                 return ipoib_cm_req_handler(cm_id, event);
 512         case IB_CM_DREQ_RECEIVED:
 513                 p = cm_id->context;
 514                 ib_send_cm_drep(cm_id, NULL, 0);
 515                 /* Fall through */
 516         case IB_CM_REJ_RECEIVED:
 517                 p = cm_id->context;
 518                 priv = netdev_priv(p->dev);
 519                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 520                         ipoib_warn(priv, "unable to move qp to error state\n");
 521                 /* Fall through */
 522         default:
 523                 return 0;
 524         }
 525 }
 526 /* Adjust length of skb with fragments to match received data */
 527 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 528                           unsigned int length, struct sk_buff *toskb)
 529 {
 530         int i, num_frags;
 531         unsigned int size;
 532
 533         /* put header into skb */
 534         size = min(length, hdr_space);
 535         skb->tail += size;
 536         skb->len += size;
 537         length -= size;
 538
 539         num_frags = skb_shinfo(skb)->nr_frags;
 540         for (i = 0; i < num_frags; i++) {
 541                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 542
 543                 if (length == 0) {
 544                         /* don't need this page */
 545                         skb_fill_page_desc(toskb, i, skb_frag_page(frag),
 546                                            0, PAGE_SIZE);
 547                         --skb_shinfo(skb)->nr_frags;
 548                 } else {
 549                         size = min(length, (unsigned) PAGE_SIZE);
 550
 551                         skb_frag_size_set(frag, size);
 552                         skb->data_len += size;
 553                         skb->truesize += size;
 554                         skb->len += size;
 555                         length -= size;
 556                 }
 557         }
 558 }
 559
 560 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 561 {
 562         struct ipoib_dev_priv *priv = netdev_priv(dev);
 563         struct ipoib_cm_rx_buf *rx_ring;
 564         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 565         struct sk_buff *skb, *newskb;
 566         struct ipoib_cm_rx *p;
 567         unsigned long flags;
 568         u64 mapping[IPOIB_CM_RX_SG];
 569         int frags;
 570         int has_srq;
 571         struct sk_buff *small_skb;
 572
 573         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 574                        wr_id, wc->status);
 575
 576         if (unlikely(wr_id >= ipoib_recvq_size)) {
 577                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 578                         spin_lock_irqsave(&priv->lock, flags);
 579                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 580                         ipoib_cm_start_rx_drain(priv);
 581                         queue_work(priv->wq, &priv->cm.rx_reap_task);
 582                         spin_unlock_irqrestore(&priv->lock, flags);
 583                 } else
 584                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 585                                    wr_id, ipoib_recvq_size);
 586                 return;
 587         }
 588
 589         p = wc->qp->qp_context;
 590
 591         has_srq = ipoib_cm_has_srq(dev);
 592         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 593
 594         skb = rx_ring[wr_id].skb;
 595
 596         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 597                 ipoib_dbg(priv, "cm recv error "
 598                            "(status=%d, wrid=%d vend_err %x)\n",
 599                            wc->status, wr_id, wc->vendor_err);
 600                 ++dev->stats.rx_dropped;
 601                 if (has_srq)
 602                         goto repost;
 603                 else {
 604                         if (!--p->recv_count) {
 605                                 spin_lock_irqsave(&priv->lock, flags);
 606                                 list_move(&p->list, &priv->cm.rx_reap_list);
 607                                 spin_unlock_irqrestore(&priv->lock, flags);
 608                                 queue_work(priv->wq, &priv->cm.rx_reap_task);
 609                         }
 610                         return;
 611                 }
 612         }
 613
 614         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 615                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 616                         spin_lock_irqsave(&priv->lock, flags);
 617                         p->jiffies = jiffies;
 618                         /* Move this entry to list head, but do not re-add it
 619                          * if it has been moved out of list. */
 620                         if (p->state == IPOIB_CM_RX_LIVE)
 621                                 list_move(&p->list, &priv->cm.passive_ids);
 622                         spin_unlock_irqrestore(&priv->lock, flags);
 623                 }
 624         }
 625
 626         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 627                 int dlen = wc->byte_len;
 628
 629                 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
 630                 if (small_skb) {
 631                         skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
 632                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 633                                                    dlen, DMA_FROM_DEVICE);
 634                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 635                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 636                                                       dlen, DMA_FROM_DEVICE);
 637                         skb_put(small_skb, dlen);
 638                         skb = small_skb;
 639                         goto copied;
 640                 }
 641         }
 642
 643         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 644                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 645
 646         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
 647                                        mapping, GFP_ATOMIC);
 648         if (unlikely(!newskb)) {
 649                 /*
 650                  * If we can't allocate a new RX buffer, dump
 651                  * this packet and reuse the old buffer.
 652                  */
 653                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 654                 ++dev->stats.rx_dropped;
 655                 goto repost;
 656         }
 657
 658         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 659         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 660
 661         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 662                        wc->byte_len, wc->slid);
 663
 664         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 665
 666 copied:
 667         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 668         skb_add_pseudo_hdr(skb);
 669
 670         ++dev->stats.rx_packets;
 671         dev->stats.rx_bytes += skb->len;
 672
 673         skb->dev = dev;
 674         /* XXX get correct PACKET_ type here */
 675         skb->pkt_type = PACKET_HOST;
 676         netif_receive_skb(skb);
 677
 678 repost:
 679         if (has_srq) {
 680                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 681                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 682                                    "for buf %d\n", wr_id);
 683         } else {
 684                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 685                                                           &priv->cm.rx_wr,
 686                                                           priv->cm.rx_sge,
 687                                                           wr_id))) {
 688                         --p->recv_count;
 689                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 690                                    "for buf %d\n", wr_id);
 691                 }
 692         }
 693 }
 694
 695 static inline int post_send(struct ipoib_dev_priv *priv,
 696                             struct ipoib_cm_tx *tx,
 697                             unsigned int wr_id,
 698                             struct ipoib_tx_buf *tx_req)
 699 {
 700         struct ib_send_wr *bad_wr;
 701
 702         ipoib_build_sge(priv, tx_req);
 703
 704         priv->tx_wr.wr.wr_id    = wr_id | IPOIB_OP_CM;
 705
 706         return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
 707 }
 708
 709 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 710 {
 711         struct ipoib_dev_priv *priv = netdev_priv(dev);
 712         struct ipoib_tx_buf *tx_req;
 713         int rc;
 714
 715         if (unlikely(skb->len > tx->mtu)) {
 716                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 717                            skb->len, tx->mtu);
 718                 ++dev->stats.tx_dropped;
 719                 ++dev->stats.tx_errors;
 720                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 721                 return;
 722         }
 723
 724         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 725                        tx->tx_head, skb->len, tx->qp->qp_num);
 726
 727         /*
 728          * We put the skb into the tx_ring _before_ we call post_send()
 729          * because it's entirely possible that the completion handler will
 730          * run before we execute anything after the post_send().  That
 731          * means we have to make sure everything is properly recorded and
 732          * our state is consistent before we call post_send().
 733          */
 734         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 735         tx_req->skb = skb;
 736
 737         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 738                 ++dev->stats.tx_errors;
 739                 dev_kfree_skb_any(skb);
 740                 return;
 741         }
 742
 743         skb_orphan(skb);
 744         skb_dst_drop(skb);
 745
 746         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
 747         if (unlikely(rc)) {
 748                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 749                 ++dev->stats.tx_errors;
 750                 ipoib_dma_unmap_tx(priv, tx_req);
 751                 dev_kfree_skb_any(skb);
 752         } else {
 753                 dev->trans_start = jiffies;
 754                 ++tx->tx_head;
 755
 756                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 757                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 758                                   tx->qp->qp_num);
 759                         netif_stop_queue(dev);
 760                         rc = ib_req_notify_cq(priv->send_cq,
 761                                 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
 762                         if (rc < 0)
 763                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 764                         else if (rc)
 765                                 ipoib_send_comp_handler(priv->send_cq, dev);
 766                 }
 767         }
 768 }
 769
 770 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 771 {
 772         struct ipoib_dev_priv *priv = netdev_priv(dev);
 773         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 774         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 775         struct ipoib_tx_buf *tx_req;
 776         unsigned long flags;
 777
 778         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 779                        wr_id, wc->status);
 780
 781         if (unlikely(wr_id >= ipoib_sendq_size)) {
 782                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 783                            wr_id, ipoib_sendq_size);
 784                 return;
 785         }
 786
 787         tx_req = &tx->tx_ring[wr_id];
 788
 789         ipoib_dma_unmap_tx(priv, tx_req);
 790
 791         /* FIXME: is this right? Shouldn't we only increment on success? */
 792         ++dev->stats.tx_packets;
 793         dev->stats.tx_bytes += tx_req->skb->len;
 794
 795         dev_kfree_skb_any(tx_req->skb);
 796
 797         netif_tx_lock(dev);
 798
 799         ++tx->tx_tail;
 800         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 801             netif_queue_stopped(dev) &&
 802             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 803                 netif_wake_queue(dev);
 804
 805         if (wc->status != IB_WC_SUCCESS &&
 806             wc->status != IB_WC_WR_FLUSH_ERR) {
 807                 struct ipoib_neigh *neigh;
 808
 809                 ipoib_dbg(priv, "failed cm send event "
 810                            "(status=%d, wrid=%d vend_err %x)\n",
 811                            wc->status, wr_id, wc->vendor_err);
 812
 813                 spin_lock_irqsave(&priv->lock, flags);
 814                 neigh = tx->neigh;
 815
 816                 if (neigh) {
 817                         neigh->cm = NULL;
 818                         ipoib_neigh_free(neigh);
 819
 820                         tx->neigh = NULL;
 821                 }
 822
 823                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 824                         list_move(&tx->list, &priv->cm.reap_list);
 825                         queue_work(priv->wq, &priv->cm.reap_task);
 826                 }
 827
 828                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 829
 830                 spin_unlock_irqrestore(&priv->lock, flags);
 831         }
 832
 833         netif_tx_unlock(dev);
 834 }
 835
 836 int ipoib_cm_dev_open(struct net_device *dev)
 837 {
 838         struct ipoib_dev_priv *priv = netdev_priv(dev);
 839         int ret;
 840
 841         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 842                 return 0;
 843
 844         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 845         if (IS_ERR(priv->cm.id)) {
 846                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 847                 ret = PTR_ERR(priv->cm.id);
 848                 goto err_cm;
 849         }
 850
 851         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 852                            0);
 853         if (ret) {
 854                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 855                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 856                 goto err_listen;
 857         }
 858
 859         return 0;
 860
 861 err_listen:
 862         ib_destroy_cm_id(priv->cm.id);
 863 err_cm:
 864         priv->cm.id = NULL;
 865         return ret;
 866 }
 867
 868 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 869 {
 870         struct ipoib_dev_priv *priv = netdev_priv(dev);
 871         struct ipoib_cm_rx *rx, *n;
 872         LIST_HEAD(list);
 873
 874         spin_lock_irq(&priv->lock);
 875         list_splice_init(&priv->cm.rx_reap_list, &list);
 876         spin_unlock_irq(&priv->lock);
 877
 878         list_for_each_entry_safe(rx, n, &list, list) {
 879                 ib_destroy_cm_id(rx->id);
 880                 ib_destroy_qp(rx->qp);
 881                 if (!ipoib_cm_has_srq(dev)) {
 882                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 883                         spin_lock_irq(&priv->lock);
 884                         --priv->cm.nonsrq_conn_qp;
 885                         spin_unlock_irq(&priv->lock);
 886                 }
 887                 kfree(rx);
 888         }
 889 }
 890
 891 void ipoib_cm_dev_stop(struct net_device *dev)
 892 {
 893         struct ipoib_dev_priv *priv = netdev_priv(dev);
 894         struct ipoib_cm_rx *p;
 895         unsigned long begin;
 896         int ret;
 897
 898         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 899                 return;
 900
 901         ib_destroy_cm_id(priv->cm.id);
 902         priv->cm.id = NULL;
 903
 904         spin_lock_irq(&priv->lock);
 905         while (!list_empty(&priv->cm.passive_ids)) {
 906                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 907                 list_move(&p->list, &priv->cm.rx_error_list);
 908                 p->state = IPOIB_CM_RX_ERROR;
 909                 spin_unlock_irq(&priv->lock);
 910                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 911                 if (ret)
 912                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 913                 spin_lock_irq(&priv->lock);
 914         }
 915
 916         /* Wait for all RX to be drained */
 917         begin = jiffies;
 918
 919         while (!list_empty(&priv->cm.rx_error_list) ||
 920                !list_empty(&priv->cm.rx_flush_list) ||
 921                !list_empty(&priv->cm.rx_drain_list)) {
 922                 if (time_after(jiffies, begin + 5 * HZ)) {
 923                         ipoib_warn(priv, "RX drain timing out\n");
 924
 925                         /*
 926                          * assume the HW is wedged and just free up everything.
 927                          */
 928                         list_splice_init(&priv->cm.rx_flush_list,
 929                                          &priv->cm.rx_reap_list);
 930                         list_splice_init(&priv->cm.rx_error_list,
 931                                          &priv->cm.rx_reap_list);
 932                         list_splice_init(&priv->cm.rx_drain_list,
 933                                          &priv->cm.rx_reap_list);
 934                         break;
 935                 }
 936                 spin_unlock_irq(&priv->lock);
 937                 msleep(1);
 938                 ipoib_drain_cq(dev);
 939                 spin_lock_irq(&priv->lock);
 940         }
 941
 942         spin_unlock_irq(&priv->lock);
 943
 944         ipoib_cm_free_rx_reap_list(dev);
 945
 946         cancel_delayed_work(&priv->cm.stale_task);
 947 }
 948
 949 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 950 {
 951         struct ipoib_cm_tx *p = cm_id->context;
 952         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 953         struct ipoib_cm_data *data = event->private_data;
 954         struct sk_buff_head skqueue;
 955         struct ib_qp_attr qp_attr;
 956         int qp_attr_mask, ret;
 957         struct sk_buff *skb;
 958
 959         p->mtu = be32_to_cpu(data->mtu);
 960
 961         if (p->mtu <= IPOIB_ENCAP_LEN) {
 962                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 963                            p->mtu, IPOIB_ENCAP_LEN);
 964                 return -EINVAL;
 965         }
 966
 967         qp_attr.qp_state = IB_QPS_RTR;
 968         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 969         if (ret) {
 970                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 971                 return ret;
 972         }
 973
 974         qp_attr.rq_psn = 0 /* FIXME */;
 975         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 976         if (ret) {
 977                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 978                 return ret;
 979         }
 980
 981         qp_attr.qp_state = IB_QPS_RTS;
 982         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 983         if (ret) {
 984                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 985                 return ret;
 986         }
 987         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 988         if (ret) {
 989                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 990                 return ret;
 991         }
 992
 993         skb_queue_head_init(&skqueue);
 994
 995         spin_lock_irq(&priv->lock);
 996         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
 997         if (p->neigh)
 998                 while ((skb = __skb_dequeue(&p->neigh->queue)))
 999                         __skb_queue_tail(&skqueue, skb);
1000         spin_unlock_irq(&priv->lock);
1001
1002         while ((skb = __skb_dequeue(&skqueue))) {
1003                 skb->dev = p->dev;
1004                 if (dev_queue_xmit(skb))
1005                         ipoib_warn(priv, "dev_queue_xmit failed "
1006                                    "to requeue packet\n");
1007         }
1008
1009         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1010         if (ret) {
1011                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1012                 return ret;
1013         }
1014         return 0;
1015 }
1016
1017 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1018 {
1019         struct ipoib_dev_priv *priv = netdev_priv(dev);
1020         struct ib_qp_init_attr attr = {
1021                 .send_cq                = priv->recv_cq,
1022                 .recv_cq                = priv->recv_cq,
1023                 .srq                    = priv->cm.srq,
1024                 .cap.max_send_wr        = ipoib_sendq_size,
1025                 .cap.max_send_sge       = 1,
1026                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1027                 .qp_type                = IB_QPT_RC,
1028                 .qp_context             = tx,
1029                 .create_flags           = IB_QP_CREATE_USE_GFP_NOIO
1030         };
1031
1032         struct ib_qp *tx_qp;
1033
1034         if (dev->features & NETIF_F_SG)
1035                 attr.cap.max_send_sge = MAX_SKB_FRAGS + 1;
1036
1037         tx_qp = ib_create_qp(priv->pd, &attr);
1038         if (PTR_ERR(tx_qp) == -EINVAL) {
1039                 attr.create_flags &= ~IB_QP_CREATE_USE_GFP_NOIO;
1040                 tx_qp = ib_create_qp(priv->pd, &attr);
1041         }
1042         return tx_qp;
1043 }
1044
1045 static int ipoib_cm_send_req(struct net_device *dev,
1046                              struct ib_cm_id *id, struct ib_qp *qp,
1047                              u32 qpn,
1048                              struct ib_sa_path_rec *pathrec)
1049 {
1050         struct ipoib_dev_priv *priv = netdev_priv(dev);
1051         struct ipoib_cm_data data = {};
1052         struct ib_cm_req_param req = {};
1053
1054         data.qpn = cpu_to_be32(priv->qp->qp_num);
1055         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1056
1057         req.primary_path                = pathrec;
1058         req.alternate_path              = NULL;
1059         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1060         req.qp_num                      = qp->qp_num;
1061         req.qp_type                     = qp->qp_type;
1062         req.private_data                = &data;
1063         req.private_data_len            = sizeof data;
1064         req.flow_control                = 0;
1065
1066         req.starting_psn                = 0; /* FIXME */
1067
1068         /*
1069          * Pick some arbitrary defaults here; we could make these
1070          * module parameters if anyone cared about setting them.
1071          */
1072         req.responder_resources         = 4;
1073         req.remote_cm_response_timeout  = 20;
1074         req.local_cm_response_timeout   = 20;
1075         req.retry_count                 = 0; /* RFC draft warns against retries */
1076         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1077         req.max_cm_retries              = 15;
1078         req.srq                         = ipoib_cm_has_srq(dev);
1079         return ib_send_cm_req(id, &req);
1080 }
1081
1082 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1083                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1084 {
1085         struct ipoib_dev_priv *priv = netdev_priv(dev);
1086         struct ib_qp_attr qp_attr;
1087         int qp_attr_mask, ret;
1088         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1089         if (ret) {
1090                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1091                 return ret;
1092         }
1093
1094         qp_attr.qp_state = IB_QPS_INIT;
1095         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1096         qp_attr.port_num = priv->port;
1097         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1098
1099         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1100         if (ret) {
1101                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1102                 return ret;
1103         }
1104         return 0;
1105 }
1106
1107 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1108                             struct ib_sa_path_rec *pathrec)
1109 {
1110         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1111         int ret;
1112
1113         p->tx_ring = __vmalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1114                                GFP_NOIO, PAGE_KERNEL);
1115         if (!p->tx_ring) {
1116                 ipoib_warn(priv, "failed to allocate tx ring\n");
1117                 ret = -ENOMEM;
1118                 goto err_tx;
1119         }
1120         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1121
1122         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1123         if (IS_ERR(p->qp)) {
1124                 ret = PTR_ERR(p->qp);
1125                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1126                 goto err_qp;
1127         }
1128
1129         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1130         if (IS_ERR(p->id)) {
1131                 ret = PTR_ERR(p->id);
1132                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1133                 goto err_id;
1134         }
1135
1136         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1137         if (ret) {
1138                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1139                 goto err_modify;
1140         }
1141
1142         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1143         if (ret) {
1144                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1145                 goto err_send_cm;
1146         }
1147
1148         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1149                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1150
1151         return 0;
1152
1153 err_send_cm:
1154 err_modify:
1155         ib_destroy_cm_id(p->id);
1156 err_id:
1157         p->id = NULL;
1158         ib_destroy_qp(p->qp);
1159 err_qp:
1160         p->qp = NULL;
1161         vfree(p->tx_ring);
1162 err_tx:
1163         return ret;
1164 }
1165
1166 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1167 {
1168         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1169         struct ipoib_tx_buf *tx_req;
1170         unsigned long begin;
1171
1172         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1173                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1174
1175         if (p->id)
1176                 ib_destroy_cm_id(p->id);
1177
1178         if (p->tx_ring) {
1179                 /* Wait for all sends to complete */
1180                 begin = jiffies;
1181                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1182                         if (time_after(jiffies, begin + 5 * HZ)) {
1183                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1184                                            p->tx_head - p->tx_tail);
1185                                 goto timeout;
1186                         }
1187
1188                         msleep(1);
1189                 }
1190         }
1191
1192 timeout:
1193
1194         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1195                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1196                 ipoib_dma_unmap_tx(priv, tx_req);
1197                 dev_kfree_skb_any(tx_req->skb);
1198                 ++p->tx_tail;
1199                 netif_tx_lock_bh(p->dev);
1200                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1201                     netif_queue_stopped(p->dev) &&
1202                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1203                         netif_wake_queue(p->dev);
1204                 netif_tx_unlock_bh(p->dev);
1205         }
1206
1207         if (p->qp)
1208                 ib_destroy_qp(p->qp);
1209
1210         vfree(p->tx_ring);
1211         kfree(p);
1212 }
1213
1214 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1215                                struct ib_cm_event *event)
1216 {
1217         struct ipoib_cm_tx *tx = cm_id->context;
1218         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1219         struct net_device *dev = priv->dev;
1220         struct ipoib_neigh *neigh;
1221         unsigned long flags;
1222         int ret;
1223
1224         switch (event->event) {
1225         case IB_CM_DREQ_RECEIVED:
1226                 ipoib_dbg(priv, "DREQ received.\n");
1227                 ib_send_cm_drep(cm_id, NULL, 0);
1228                 break;
1229         case IB_CM_REP_RECEIVED:
1230                 ipoib_dbg(priv, "REP received.\n");
1231                 ret = ipoib_cm_rep_handler(cm_id, event);
1232                 if (ret)
1233                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1234                                        NULL, 0, NULL, 0);
1235                 break;
1236         case IB_CM_REQ_ERROR:
1237         case IB_CM_REJ_RECEIVED:
1238         case IB_CM_TIMEWAIT_EXIT:
1239                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1240                 netif_tx_lock_bh(dev);
1241                 spin_lock_irqsave(&priv->lock, flags);
1242                 neigh = tx->neigh;
1243
1244                 if (neigh) {
1245                         neigh->cm = NULL;
1246                         ipoib_neigh_free(neigh);
1247
1248                         tx->neigh = NULL;
1249                 }
1250
1251                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1252                         list_move(&tx->list, &priv->cm.reap_list);
1253                         queue_work(priv->wq, &priv->cm.reap_task);
1254                 }
1255
1256                 spin_unlock_irqrestore(&priv->lock, flags);
1257                 netif_tx_unlock_bh(dev);
1258                 break;
1259         default:
1260                 break;
1261         }
1262
1263         return 0;
1264 }
1265
1266 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1267                                        struct ipoib_neigh *neigh)
1268 {
1269         struct ipoib_dev_priv *priv = netdev_priv(dev);
1270         struct ipoib_cm_tx *tx;
1271
1272         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1273         if (!tx)
1274                 return NULL;
1275
1276         neigh->cm = tx;
1277         tx->neigh = neigh;
1278         tx->path = path;
1279         tx->dev = dev;
1280         list_add(&tx->list, &priv->cm.start_list);
1281         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1282         queue_work(priv->wq, &priv->cm.start_task);
1283         return tx;
1284 }
1285
1286 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1287 {
1288         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1289         unsigned long flags;
1290         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1291                 spin_lock_irqsave(&priv->lock, flags);
1292                 list_move(&tx->list, &priv->cm.reap_list);
1293                 queue_work(priv->wq, &priv->cm.reap_task);
1294                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1295                           tx->neigh->daddr + 4);
1296                 tx->neigh = NULL;
1297                 spin_unlock_irqrestore(&priv->lock, flags);
1298         }
1299 }
1300
1301 #define QPN_AND_OPTIONS_OFFSET  4
1302
1303 static void ipoib_cm_tx_start(struct work_struct *work)
1304 {
1305         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1306                                                    cm.start_task);
1307         struct net_device *dev = priv->dev;
1308         struct ipoib_neigh *neigh;
1309         struct ipoib_cm_tx *p;
1310         unsigned long flags;
1311         struct ipoib_path *path;
1312         int ret;
1313
1314         struct ib_sa_path_rec pathrec;
1315         u32 qpn;
1316
1317         netif_tx_lock_bh(dev);
1318         spin_lock_irqsave(&priv->lock, flags);
1319
1320         while (!list_empty(&priv->cm.start_list)) {
1321                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1322                 list_del_init(&p->list);
1323                 neigh = p->neigh;
1324
1325                 qpn = IPOIB_QPN(neigh->daddr);
1326                 /*
1327                  * As long as the search is with these 2 locks,
1328                  * path existence indicates its validity.
1329                  */
1330                 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1331                 if (!path) {
1332                         pr_info("%s ignore not valid path %pI6\n",
1333                                 __func__,
1334                                 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1335                         goto free_neigh;
1336                 }
1337                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1338
1339                 spin_unlock_irqrestore(&priv->lock, flags);
1340                 netif_tx_unlock_bh(dev);
1341
1342                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1343
1344                 netif_tx_lock_bh(dev);
1345                 spin_lock_irqsave(&priv->lock, flags);
1346
1347                 if (ret) {
1348 free_neigh:
1349                         neigh = p->neigh;
1350                         if (neigh) {
1351                                 neigh->cm = NULL;
1352                                 ipoib_neigh_free(neigh);
1353                         }
1354                         list_del(&p->list);
1355                         kfree(p);
1356                 }
1357         }
1358
1359         spin_unlock_irqrestore(&priv->lock, flags);
1360         netif_tx_unlock_bh(dev);
1361 }
1362
1363 static void ipoib_cm_tx_reap(struct work_struct *work)
1364 {
1365         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1366                                                    cm.reap_task);
1367         struct net_device *dev = priv->dev;
1368         struct ipoib_cm_tx *p;
1369         unsigned long flags;
1370
1371         netif_tx_lock_bh(dev);
1372         spin_lock_irqsave(&priv->lock, flags);
1373
1374         while (!list_empty(&priv->cm.reap_list)) {
1375                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1376                 list_del(&p->list);
1377                 spin_unlock_irqrestore(&priv->lock, flags);
1378                 netif_tx_unlock_bh(dev);
1379                 ipoib_cm_tx_destroy(p);
1380                 netif_tx_lock_bh(dev);
1381                 spin_lock_irqsave(&priv->lock, flags);
1382         }
1383
1384         spin_unlock_irqrestore(&priv->lock, flags);
1385         netif_tx_unlock_bh(dev);
1386 }
1387
1388 static void ipoib_cm_skb_reap(struct work_struct *work)
1389 {
1390         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1391                                                    cm.skb_task);
1392         struct net_device *dev = priv->dev;
1393         struct sk_buff *skb;
1394         unsigned long flags;
1395         unsigned mtu = priv->mcast_mtu;
1396
1397         netif_tx_lock_bh(dev);
1398         spin_lock_irqsave(&priv->lock, flags);
1399
1400         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1401                 spin_unlock_irqrestore(&priv->lock, flags);
1402                 netif_tx_unlock_bh(dev);
1403
1404                 if (skb->protocol == htons(ETH_P_IP))
1405                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1406 #if IS_ENABLED(CONFIG_IPV6)
1407                 else if (skb->protocol == htons(ETH_P_IPV6))
1408                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1409 #endif
1410                 dev_kfree_skb_any(skb);
1411
1412                 netif_tx_lock_bh(dev);
1413                 spin_lock_irqsave(&priv->lock, flags);
1414         }
1415
1416         spin_unlock_irqrestore(&priv->lock, flags);
1417         netif_tx_unlock_bh(dev);
1418 }
1419
1420 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1421                            unsigned int mtu)
1422 {
1423         struct ipoib_dev_priv *priv = netdev_priv(dev);
1424         int e = skb_queue_empty(&priv->cm.skb_queue);
1425
1426         if (skb_dst(skb))
1427                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1428
1429         skb_queue_tail(&priv->cm.skb_queue, skb);
1430         if (e)
1431                 queue_work(priv->wq, &priv->cm.skb_task);
1432 }
1433
1434 static void ipoib_cm_rx_reap(struct work_struct *work)
1435 {
1436         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1437                                                 cm.rx_reap_task)->dev);
1438 }
1439
1440 static void ipoib_cm_stale_task(struct work_struct *work)
1441 {
1442         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1443                                                    cm.stale_task.work);
1444         struct ipoib_cm_rx *p;
1445         int ret;
1446
1447         spin_lock_irq(&priv->lock);
1448         while (!list_empty(&priv->cm.passive_ids)) {
1449                 /* List is sorted by LRU, start from tail,
1450                  * stop when we see a recently used entry */
1451                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1452                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1453                         break;
1454                 list_move(&p->list, &priv->cm.rx_error_list);
1455                 p->state = IPOIB_CM_RX_ERROR;
1456                 spin_unlock_irq(&priv->lock);
1457                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1458                 if (ret)
1459                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1460                 spin_lock_irq(&priv->lock);
1461         }
1462
1463         if (!list_empty(&priv->cm.passive_ids))
1464                 queue_delayed_work(priv->wq,
1465                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1466         spin_unlock_irq(&priv->lock);
1467 }
1468
1469 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1470                          char *buf)
1471 {
1472         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1473
1474         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1475                 return sprintf(buf, "connected\n");
1476         else
1477                 return sprintf(buf, "datagram\n");
1478 }
1479
1480 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1481                         const char *buf, size_t count)
1482 {
1483         struct net_device *dev = to_net_dev(d);
1484         int ret;
1485
1486         if (!rtnl_trylock())
1487                 return restart_syscall();
1488
1489         ret = ipoib_set_mode(dev, buf);
1490
1491         rtnl_unlock();
1492
1493         if (!ret)
1494                 return count;
1495
1496         return ret;
1497 }
1498
1499 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1500
1501 int ipoib_cm_add_mode_attr(struct net_device *dev)
1502 {
1503         return device_create_file(&dev->dev, &dev_attr_mode);
1504 }
1505
1506 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1507 {
1508         struct ipoib_dev_priv *priv = netdev_priv(dev);
1509         struct ib_srq_init_attr srq_init_attr = {
1510                 .srq_type = IB_SRQT_BASIC,
1511                 .attr = {
1512                         .max_wr  = ipoib_recvq_size,
1513                         .max_sge = max_sge
1514                 }
1515         };
1516
1517         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1518         if (IS_ERR(priv->cm.srq)) {
1519                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1520                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1521                                priv->ca->name, PTR_ERR(priv->cm.srq));
1522                 priv->cm.srq = NULL;
1523                 return;
1524         }
1525
1526         priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1527         if (!priv->cm.srq_ring) {
1528                 printk(KERN_WARNING "%s: failed to allocate CM SRQ ring (%d entries)\n",
1529                        priv->ca->name, ipoib_recvq_size);
1530                 ib_destroy_srq(priv->cm.srq);
1531                 priv->cm.srq = NULL;
1532                 return;
1533         }
1534
1535 }
1536
1537 int ipoib_cm_dev_init(struct net_device *dev)
1538 {
1539         struct ipoib_dev_priv *priv = netdev_priv(dev);
1540         int i, ret;
1541         struct ib_device_attr attr;
1542
1543         INIT_LIST_HEAD(&priv->cm.passive_ids);
1544         INIT_LIST_HEAD(&priv->cm.reap_list);
1545         INIT_LIST_HEAD(&priv->cm.start_list);
1546         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1547         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1548         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1549         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1550         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1551         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1552         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1553         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1554         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1555
1556         skb_queue_head_init(&priv->cm.skb_queue);
1557
1558         ret = ib_query_device(priv->ca, &attr);
1559         if (ret) {
1560                 printk(KERN_WARNING "ib_query_device() failed with %d\n", ret);
1561                 return ret;
1562         }
1563
1564         ipoib_dbg(priv, "max_srq_sge=%d\n", attr.max_srq_sge);
1565
1566         attr.max_srq_sge = min_t(int, IPOIB_CM_RX_SG, attr.max_srq_sge);
1567         ipoib_cm_create_srq(dev, attr.max_srq_sge);
1568         if (ipoib_cm_has_srq(dev)) {
1569                 priv->cm.max_cm_mtu = attr.max_srq_sge * PAGE_SIZE - 0x10;
1570                 priv->cm.num_frags  = attr.max_srq_sge;
1571                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1572                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1573         } else {
1574                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1575                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1576         }
1577
1578         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1579
1580         if (ipoib_cm_has_srq(dev)) {
1581                 for (i = 0; i < ipoib_recvq_size; ++i) {
1582                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1583                                                    priv->cm.num_frags - 1,
1584                                                    priv->cm.srq_ring[i].mapping,
1585                                                    GFP_KERNEL)) {
1586                                 ipoib_warn(priv, "failed to allocate "
1587                                            "receive buffer %d\n", i);
1588                                 ipoib_cm_dev_cleanup(dev);
1589                                 return -ENOMEM;
1590                         }
1591
1592                         if (ipoib_cm_post_receive_srq(dev, i)) {
1593                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1594                                            "failed for buf %d\n", i);
1595                                 ipoib_cm_dev_cleanup(dev);
1596                                 return -EIO;
1597                         }
1598                 }
1599         }
1600
1601         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1602         return 0;
1603 }
1604
1605 void ipoib_cm_dev_cleanup(struct net_device *dev)
1606 {
1607         struct ipoib_dev_priv *priv = netdev_priv(dev);
1608         int ret;
1609
1610         if (!priv->cm.srq)
1611                 return;
1612
1613         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1614
1615         ret = ib_destroy_srq(priv->cm.srq);
1616         if (ret)
1617                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1618
1619         priv->cm.srq = NULL;
1620         if (!priv->cm.srq_ring)
1621                 return;
1622
1623         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1624         priv->cm.srq_ring = NULL;
1625 }