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         .opcode = IB_WR_SEND,
  76 };
  77
  78 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  79                                struct ib_cm_event *event);
  80
  81 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  82                                   u64 mapping[IPOIB_CM_RX_SG])
  83 {
  84         int i;
  85
  86         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  87
  88         for (i = 0; i < frags; ++i)
  89                 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  90 }
  91
  92 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  93 {
  94         struct ipoib_dev_priv *priv = netdev_priv(dev);
  95         struct ib_recv_wr *bad_wr;
  96         int i, ret;
  97
  98         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
  99
 100         for (i = 0; i < priv->cm.num_frags; ++i)
 101                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 102
 103         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 104         if (unlikely(ret)) {
 105                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 106                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 107                                       priv->cm.srq_ring[id].mapping);
 108                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 109                 priv->cm.srq_ring[id].skb = NULL;
 110         }
 111
 112         return ret;
 113 }
 114
 115 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 116                                         struct ipoib_cm_rx *rx,
 117                                         struct ib_recv_wr *wr,
 118                                         struct ib_sge *sge, int id)
 119 {
 120         struct ipoib_dev_priv *priv = netdev_priv(dev);
 121         struct ib_recv_wr *bad_wr;
 122         int i, ret;
 123
 124         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 125
 126         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 127                 sge[i].addr = rx->rx_ring[id].mapping[i];
 128
 129         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 130         if (unlikely(ret)) {
 131                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 132                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 133                                       rx->rx_ring[id].mapping);
 134                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 135                 rx->rx_ring[id].skb = NULL;
 136         }
 137
 138         return ret;
 139 }
 140
 141 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 142                                              struct ipoib_cm_rx_buf *rx_ring,
 143                                              int id, int frags,
 144                                              u64 mapping[IPOIB_CM_RX_SG],
 145                                              gfp_t gfp)
 146 {
 147         struct ipoib_dev_priv *priv = netdev_priv(dev);
 148         struct sk_buff *skb;
 149         int i;
 150
 151         skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
 152         if (unlikely(!skb))
 153                 return NULL;
 154
 155         /*
 156          * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
 157          * IP header to a multiple of 16.
 158          */
 159         skb_reserve(skb, IPOIB_CM_RX_RESERVE);
 160
 161         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 162                                        DMA_FROM_DEVICE);
 163         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 164                 dev_kfree_skb_any(skb);
 165                 return NULL;
 166         }
 167
 168         for (i = 0; i < frags; i++) {
 169                 struct page *page = alloc_page(gfp);
 170
 171                 if (!page)
 172                         goto partial_error;
 173                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 174
 175                 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
 176                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 177                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 178                         goto partial_error;
 179         }
 180
 181         rx_ring[id].skb = skb;
 182         return skb;
 183
 184 partial_error:
 185
 186         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 187
 188         for (; i > 0; --i)
 189                 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 190
 191         dev_kfree_skb_any(skb);
 192         return NULL;
 193 }
 194
 195 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 196                                   struct ipoib_cm_rx_buf *rx_ring)
 197 {
 198         struct ipoib_dev_priv *priv = netdev_priv(dev);
 199         int i;
 200
 201         for (i = 0; i < ipoib_recvq_size; ++i)
 202                 if (rx_ring[i].skb) {
 203                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 204                                               rx_ring[i].mapping);
 205                         dev_kfree_skb_any(rx_ring[i].skb);
 206                 }
 207
 208         vfree(rx_ring);
 209 }
 210
 211 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 212 {
 213         struct ib_send_wr *bad_wr;
 214         struct ipoib_cm_rx *p;
 215
 216         /* We only reserved 1 extra slot in CQ for drain WRs, so
 217          * make sure we have at most 1 outstanding WR. */
 218         if (list_empty(&priv->cm.rx_flush_list) ||
 219             !list_empty(&priv->cm.rx_drain_list))
 220                 return;
 221
 222         /*
 223          * QPs on flush list are error state.  This way, a "flush
 224          * error" WC will be immediately generated for each WR we post.
 225          */
 226         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 227         ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
 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                 return -ENOMEM;
 362
 363         t = kmalloc(sizeof *t, GFP_KERNEL);
 364         if (!t) {
 365                 ret = -ENOMEM;
 366                 goto err_free;
 367         }
 368
 369         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 370
 371         spin_lock_irq(&priv->lock);
 372
 373         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 374                 spin_unlock_irq(&priv->lock);
 375                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 376                 ret = -EINVAL;
 377                 goto err_free;
 378         } else
 379                 ++priv->cm.nonsrq_conn_qp;
 380
 381         spin_unlock_irq(&priv->lock);
 382
 383         for (i = 0; i < ipoib_recvq_size; ++i) {
 384                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 385                                            rx->rx_ring[i].mapping,
 386                                            GFP_KERNEL)) {
 387                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 388                         ret = -ENOMEM;
 389                         goto err_count;
 390                 }
 391                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 392                 if (ret) {
 393                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 394                                    "failed for buf %d\n", i);
 395                         ret = -EIO;
 396                         goto err_count;
 397                 }
 398         }
 399
 400         rx->recv_count = ipoib_recvq_size;
 401
 402         kfree(t);
 403
 404         return 0;
 405
 406 err_count:
 407         spin_lock_irq(&priv->lock);
 408         --priv->cm.nonsrq_conn_qp;
 409         spin_unlock_irq(&priv->lock);
 410
 411 err_free:
 412         kfree(t);
 413         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 414
 415         return ret;
 416 }
 417
 418 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 419                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 420                              unsigned psn)
 421 {
 422         struct ipoib_dev_priv *priv = netdev_priv(dev);
 423         struct ipoib_cm_data data = {};
 424         struct ib_cm_rep_param rep = {};
 425
 426         data.qpn = cpu_to_be32(priv->qp->qp_num);
 427         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 428
 429         rep.private_data = &data;
 430         rep.private_data_len = sizeof data;
 431         rep.flow_control = 0;
 432         rep.rnr_retry_count = req->rnr_retry_count;
 433         rep.srq = ipoib_cm_has_srq(dev);
 434         rep.qp_num = qp->qp_num;
 435         rep.starting_psn = psn;
 436         return ib_send_cm_rep(cm_id, &rep);
 437 }
 438
 439 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 440 {
 441         struct net_device *dev = cm_id->context;
 442         struct ipoib_dev_priv *priv = netdev_priv(dev);
 443         struct ipoib_cm_rx *p;
 444         unsigned psn;
 445         int ret;
 446
 447         ipoib_dbg(priv, "REQ arrived\n");
 448         p = kzalloc(sizeof *p, GFP_KERNEL);
 449         if (!p)
 450                 return -ENOMEM;
 451         p->dev = dev;
 452         p->id = cm_id;
 453         cm_id->context = p;
 454         p->state = IPOIB_CM_RX_LIVE;
 455         p->jiffies = jiffies;
 456         INIT_LIST_HEAD(&p->list);
 457
 458         p->qp = ipoib_cm_create_rx_qp(dev, p);
 459         if (IS_ERR(p->qp)) {
 460                 ret = PTR_ERR(p->qp);
 461                 goto err_qp;
 462         }
 463
 464         psn = prandom_u32() & 0xffffff;
 465         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 466         if (ret)
 467                 goto err_modify;
 468
 469         if (!ipoib_cm_has_srq(dev)) {
 470                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 471                 if (ret)
 472                         goto err_modify;
 473         }
 474
 475         spin_lock_irq(&priv->lock);
 476         queue_delayed_work(priv->wq,
 477                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 478         /* Add this entry to passive ids list head, but do not re-add it
 479          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 480         p->jiffies = jiffies;
 481         if (p->state == IPOIB_CM_RX_LIVE)
 482                 list_move(&p->list, &priv->cm.passive_ids);
 483         spin_unlock_irq(&priv->lock);
 484
 485         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 486         if (ret) {
 487                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 488                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 489                         ipoib_warn(priv, "unable to move qp to error state\n");
 490         }
 491         return 0;
 492
 493 err_modify:
 494         ib_destroy_qp(p->qp);
 495 err_qp:
 496         kfree(p);
 497         return ret;
 498 }
 499
 500 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 501                                struct ib_cm_event *event)
 502 {
 503         struct ipoib_cm_rx *p;
 504         struct ipoib_dev_priv *priv;
 505
 506         switch (event->event) {
 507         case IB_CM_REQ_RECEIVED:
 508                 return ipoib_cm_req_handler(cm_id, event);
 509         case IB_CM_DREQ_RECEIVED:
 510                 p = cm_id->context;
 511                 ib_send_cm_drep(cm_id, NULL, 0);
 512                 /* Fall through */
 513         case IB_CM_REJ_RECEIVED:
 514                 p = cm_id->context;
 515                 priv = netdev_priv(p->dev);
 516                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 517                         ipoib_warn(priv, "unable to move qp to error state\n");
 518                 /* Fall through */
 519         default:
 520                 return 0;
 521         }
 522 }
 523 /* Adjust length of skb with fragments to match received data */
 524 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 525                           unsigned int length, struct sk_buff *toskb)
 526 {
 527         int i, num_frags;
 528         unsigned int size;
 529
 530         /* put header into skb */
 531         size = min(length, hdr_space);
 532         skb->tail += size;
 533         skb->len += size;
 534         length -= size;
 535
 536         num_frags = skb_shinfo(skb)->nr_frags;
 537         for (i = 0; i < num_frags; i++) {
 538                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 539
 540                 if (length == 0) {
 541                         /* don't need this page */
 542                         skb_fill_page_desc(toskb, i, skb_frag_page(frag),
 543                                            0, PAGE_SIZE);
 544                         --skb_shinfo(skb)->nr_frags;
 545                 } else {
 546                         size = min(length, (unsigned) PAGE_SIZE);
 547
 548                         skb_frag_size_set(frag, size);
 549                         skb->data_len += size;
 550                         skb->truesize += size;
 551                         skb->len += size;
 552                         length -= size;
 553                 }
 554         }
 555 }
 556
 557 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 558 {
 559         struct ipoib_dev_priv *priv = netdev_priv(dev);
 560         struct ipoib_cm_rx_buf *rx_ring;
 561         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 562         struct sk_buff *skb, *newskb;
 563         struct ipoib_cm_rx *p;
 564         unsigned long flags;
 565         u64 mapping[IPOIB_CM_RX_SG];
 566         int frags;
 567         int has_srq;
 568         struct sk_buff *small_skb;
 569
 570         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 571                        wr_id, wc->status);
 572
 573         if (unlikely(wr_id >= ipoib_recvq_size)) {
 574                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 575                         spin_lock_irqsave(&priv->lock, flags);
 576                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 577                         ipoib_cm_start_rx_drain(priv);
 578                         queue_work(priv->wq, &priv->cm.rx_reap_task);
 579                         spin_unlock_irqrestore(&priv->lock, flags);
 580                 } else
 581                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 582                                    wr_id, ipoib_recvq_size);
 583                 return;
 584         }
 585
 586         p = wc->qp->qp_context;
 587
 588         has_srq = ipoib_cm_has_srq(dev);
 589         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 590
 591         skb = rx_ring[wr_id].skb;
 592
 593         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 594                 ipoib_dbg(priv, "cm recv error "
 595                            "(status=%d, wrid=%d vend_err %x)\n",
 596                            wc->status, wr_id, wc->vendor_err);
 597                 ++dev->stats.rx_dropped;
 598                 if (has_srq)
 599                         goto repost;
 600                 else {
 601                         if (!--p->recv_count) {
 602                                 spin_lock_irqsave(&priv->lock, flags);
 603                                 list_move(&p->list, &priv->cm.rx_reap_list);
 604                                 spin_unlock_irqrestore(&priv->lock, flags);
 605                                 queue_work(priv->wq, &priv->cm.rx_reap_task);
 606                         }
 607                         return;
 608                 }
 609         }
 610
 611         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 612                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 613                         spin_lock_irqsave(&priv->lock, flags);
 614                         p->jiffies = jiffies;
 615                         /* Move this entry to list head, but do not re-add it
 616                          * if it has been moved out of list. */
 617                         if (p->state == IPOIB_CM_RX_LIVE)
 618                                 list_move(&p->list, &priv->cm.passive_ids);
 619                         spin_unlock_irqrestore(&priv->lock, flags);
 620                 }
 621         }
 622
 623         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 624                 int dlen = wc->byte_len;
 625
 626                 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
 627                 if (small_skb) {
 628                         skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
 629                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 630                                                    dlen, DMA_FROM_DEVICE);
 631                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 632                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 633                                                       dlen, DMA_FROM_DEVICE);
 634                         skb_put(small_skb, dlen);
 635                         skb = small_skb;
 636                         goto copied;
 637                 }
 638         }
 639
 640         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 641                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 642
 643         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
 644                                        mapping, GFP_ATOMIC);
 645         if (unlikely(!newskb)) {
 646                 /*
 647                  * If we can't allocate a new RX buffer, dump
 648                  * this packet and reuse the old buffer.
 649                  */
 650                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 651                 ++dev->stats.rx_dropped;
 652                 goto repost;
 653         }
 654
 655         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 656         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 657
 658         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 659                        wc->byte_len, wc->slid);
 660
 661         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 662
 663 copied:
 664         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 665         skb_add_pseudo_hdr(skb);
 666
 667         ++dev->stats.rx_packets;
 668         dev->stats.rx_bytes += skb->len;
 669
 670         skb->dev = dev;
 671         /* XXX get correct PACKET_ type here */
 672         skb->pkt_type = PACKET_HOST;
 673         netif_receive_skb(skb);
 674
 675 repost:
 676         if (has_srq) {
 677                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 678                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 679                                    "for buf %d\n", wr_id);
 680         } else {
 681                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 682                                                           &priv->cm.rx_wr,
 683                                                           priv->cm.rx_sge,
 684                                                           wr_id))) {
 685                         --p->recv_count;
 686                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 687                                    "for buf %d\n", wr_id);
 688                 }
 689         }
 690 }
 691
 692 static inline int post_send(struct ipoib_dev_priv *priv,
 693                             struct ipoib_cm_tx *tx,
 694                             unsigned int wr_id,
 695                             struct ipoib_tx_buf *tx_req)
 696 {
 697         struct ib_send_wr *bad_wr;
 698
 699         ipoib_build_sge(priv, tx_req);
 700
 701         priv->tx_wr.wr.wr_id    = wr_id | IPOIB_OP_CM;
 702
 703         return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
 704 }
 705
 706 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 707 {
 708         struct ipoib_dev_priv *priv = netdev_priv(dev);
 709         struct ipoib_tx_buf *tx_req;
 710         int rc;
 711         unsigned usable_sge = tx->max_send_sge - !!skb_headlen(skb);
 712
 713         if (unlikely(skb->len > tx->mtu)) {
 714                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 715                            skb->len, tx->mtu);
 716                 ++dev->stats.tx_dropped;
 717                 ++dev->stats.tx_errors;
 718                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 719                 return;
 720         }
 721         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 722                 if (skb_linearize(skb) < 0) {
 723                         ipoib_warn(priv, "skb could not be linearized\n");
 724                         ++dev->stats.tx_dropped;
 725                         ++dev->stats.tx_errors;
 726                         dev_kfree_skb_any(skb);
 727                         return;
 728                 }
 729                 /* Does skb_linearize return ok without reducing nr_frags? */
 730                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 731                         ipoib_warn(priv, "too many frags after skb linearize\n");
 732                         ++dev->stats.tx_dropped;
 733                         ++dev->stats.tx_errors;
 734                         dev_kfree_skb_any(skb);
 735                         return;
 736                 }
 737         }
 738         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 739                        tx->tx_head, skb->len, tx->qp->qp_num);
 740
 741         /*
 742          * We put the skb into the tx_ring _before_ we call post_send()
 743          * because it's entirely possible that the completion handler will
 744          * run before we execute anything after the post_send().  That
 745          * means we have to make sure everything is properly recorded and
 746          * our state is consistent before we call post_send().
 747          */
 748         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 749         tx_req->skb = skb;
 750
 751         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 752                 ++dev->stats.tx_errors;
 753                 dev_kfree_skb_any(skb);
 754                 return;
 755         }
 756
 757         skb_orphan(skb);
 758         skb_dst_drop(skb);
 759
 760         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
 761         if (unlikely(rc)) {
 762                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 763                 ++dev->stats.tx_errors;
 764                 ipoib_dma_unmap_tx(priv, tx_req);
 765                 dev_kfree_skb_any(skb);
 766         } else {
 767                 netif_trans_update(dev);
 768                 ++tx->tx_head;
 769
 770                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 771                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 772                                   tx->qp->qp_num);
 773                         netif_stop_queue(dev);
 774                         rc = ib_req_notify_cq(priv->send_cq,
 775                                 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
 776                         if (rc < 0)
 777                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 778                         else if (rc)
 779                                 ipoib_send_comp_handler(priv->send_cq, dev);
 780                 }
 781         }
 782 }
 783
 784 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 785 {
 786         struct ipoib_dev_priv *priv = netdev_priv(dev);
 787         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 788         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 789         struct ipoib_tx_buf *tx_req;
 790         unsigned long flags;
 791
 792         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 793                        wr_id, wc->status);
 794
 795         if (unlikely(wr_id >= ipoib_sendq_size)) {
 796                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 797                            wr_id, ipoib_sendq_size);
 798                 return;
 799         }
 800
 801         tx_req = &tx->tx_ring[wr_id];
 802
 803         ipoib_dma_unmap_tx(priv, tx_req);
 804
 805         /* FIXME: is this right? Shouldn't we only increment on success? */
 806         ++dev->stats.tx_packets;
 807         dev->stats.tx_bytes += tx_req->skb->len;
 808
 809         dev_kfree_skb_any(tx_req->skb);
 810
 811         netif_tx_lock(dev);
 812
 813         ++tx->tx_tail;
 814         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 815             netif_queue_stopped(dev) &&
 816             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 817                 netif_wake_queue(dev);
 818
 819         if (wc->status != IB_WC_SUCCESS &&
 820             wc->status != IB_WC_WR_FLUSH_ERR) {
 821                 struct ipoib_neigh *neigh;
 822
 823                 ipoib_dbg(priv, "failed cm send event "
 824                            "(status=%d, wrid=%d vend_err %x)\n",
 825                            wc->status, wr_id, wc->vendor_err);
 826
 827                 spin_lock_irqsave(&priv->lock, flags);
 828                 neigh = tx->neigh;
 829
 830                 if (neigh) {
 831                         neigh->cm = NULL;
 832                         ipoib_neigh_free(neigh);
 833
 834                         tx->neigh = NULL;
 835                 }
 836
 837                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 838                         list_move(&tx->list, &priv->cm.reap_list);
 839                         queue_work(priv->wq, &priv->cm.reap_task);
 840                 }
 841
 842                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 843
 844                 spin_unlock_irqrestore(&priv->lock, flags);
 845         }
 846
 847         netif_tx_unlock(dev);
 848 }
 849
 850 int ipoib_cm_dev_open(struct net_device *dev)
 851 {
 852         struct ipoib_dev_priv *priv = netdev_priv(dev);
 853         int ret;
 854
 855         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 856                 return 0;
 857
 858         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 859         if (IS_ERR(priv->cm.id)) {
 860                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 861                 ret = PTR_ERR(priv->cm.id);
 862                 goto err_cm;
 863         }
 864
 865         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 866                            0);
 867         if (ret) {
 868                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 869                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 870                 goto err_listen;
 871         }
 872
 873         return 0;
 874
 875 err_listen:
 876         ib_destroy_cm_id(priv->cm.id);
 877 err_cm:
 878         priv->cm.id = NULL;
 879         return ret;
 880 }
 881
 882 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 883 {
 884         struct ipoib_dev_priv *priv = netdev_priv(dev);
 885         struct ipoib_cm_rx *rx, *n;
 886         LIST_HEAD(list);
 887
 888         spin_lock_irq(&priv->lock);
 889         list_splice_init(&priv->cm.rx_reap_list, &list);
 890         spin_unlock_irq(&priv->lock);
 891
 892         list_for_each_entry_safe(rx, n, &list, list) {
 893                 ib_destroy_cm_id(rx->id);
 894                 ib_destroy_qp(rx->qp);
 895                 if (!ipoib_cm_has_srq(dev)) {
 896                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 897                         spin_lock_irq(&priv->lock);
 898                         --priv->cm.nonsrq_conn_qp;
 899                         spin_unlock_irq(&priv->lock);
 900                 }
 901                 kfree(rx);
 902         }
 903 }
 904
 905 void ipoib_cm_dev_stop(struct net_device *dev)
 906 {
 907         struct ipoib_dev_priv *priv = netdev_priv(dev);
 908         struct ipoib_cm_rx *p;
 909         unsigned long begin;
 910         int ret;
 911
 912         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 913                 return;
 914
 915         ib_destroy_cm_id(priv->cm.id);
 916         priv->cm.id = NULL;
 917
 918         spin_lock_irq(&priv->lock);
 919         while (!list_empty(&priv->cm.passive_ids)) {
 920                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 921                 list_move(&p->list, &priv->cm.rx_error_list);
 922                 p->state = IPOIB_CM_RX_ERROR;
 923                 spin_unlock_irq(&priv->lock);
 924                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 925                 if (ret)
 926                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 927                 spin_lock_irq(&priv->lock);
 928         }
 929
 930         /* Wait for all RX to be drained */
 931         begin = jiffies;
 932
 933         while (!list_empty(&priv->cm.rx_error_list) ||
 934                !list_empty(&priv->cm.rx_flush_list) ||
 935                !list_empty(&priv->cm.rx_drain_list)) {
 936                 if (time_after(jiffies, begin + 5 * HZ)) {
 937                         ipoib_warn(priv, "RX drain timing out\n");
 938
 939                         /*
 940                          * assume the HW is wedged and just free up everything.
 941                          */
 942                         list_splice_init(&priv->cm.rx_flush_list,
 943                                          &priv->cm.rx_reap_list);
 944                         list_splice_init(&priv->cm.rx_error_list,
 945                                          &priv->cm.rx_reap_list);
 946                         list_splice_init(&priv->cm.rx_drain_list,
 947                                          &priv->cm.rx_reap_list);
 948                         break;
 949                 }
 950                 spin_unlock_irq(&priv->lock);
 951                 msleep(1);
 952                 ipoib_drain_cq(dev);
 953                 spin_lock_irq(&priv->lock);
 954         }
 955
 956         spin_unlock_irq(&priv->lock);
 957
 958         ipoib_cm_free_rx_reap_list(dev);
 959
 960         cancel_delayed_work(&priv->cm.stale_task);
 961 }
 962
 963 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 964 {
 965         struct ipoib_cm_tx *p = cm_id->context;
 966         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
 967         struct ipoib_cm_data *data = event->private_data;
 968         struct sk_buff_head skqueue;
 969         struct ib_qp_attr qp_attr;
 970         int qp_attr_mask, ret;
 971         struct sk_buff *skb;
 972
 973         p->mtu = be32_to_cpu(data->mtu);
 974
 975         if (p->mtu <= IPOIB_ENCAP_LEN) {
 976                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 977                            p->mtu, IPOIB_ENCAP_LEN);
 978                 return -EINVAL;
 979         }
 980
 981         qp_attr.qp_state = IB_QPS_RTR;
 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 RTR: %d\n", ret);
 985                 return ret;
 986         }
 987
 988         qp_attr.rq_psn = 0 /* FIXME */;
 989         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 990         if (ret) {
 991                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 992                 return ret;
 993         }
 994
 995         qp_attr.qp_state = IB_QPS_RTS;
 996         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 997         if (ret) {
 998                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 999                 return ret;
1000         }
1001         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1002         if (ret) {
1003                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1004                 return ret;
1005         }
1006
1007         skb_queue_head_init(&skqueue);
1008
1009         spin_lock_irq(&priv->lock);
1010         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1011         if (p->neigh)
1012                 while ((skb = __skb_dequeue(&p->neigh->queue)))
1013                         __skb_queue_tail(&skqueue, skb);
1014         spin_unlock_irq(&priv->lock);
1015
1016         while ((skb = __skb_dequeue(&skqueue))) {
1017                 skb->dev = p->dev;
1018                 ret = dev_queue_xmit(skb);
1019                 if (ret)
1020                         ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1021                                    __func__, ret);
1022         }
1023
1024         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1025         if (ret) {
1026                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1027                 return ret;
1028         }
1029         return 0;
1030 }
1031
1032 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1033 {
1034         struct ipoib_dev_priv *priv = netdev_priv(dev);
1035         struct ib_qp_init_attr attr = {
1036                 .send_cq                = priv->recv_cq,
1037                 .recv_cq                = priv->recv_cq,
1038                 .srq                    = priv->cm.srq,
1039                 .cap.max_send_wr        = ipoib_sendq_size,
1040                 .cap.max_send_sge       = 1,
1041                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1042                 .qp_type                = IB_QPT_RC,
1043                 .qp_context             = tx,
1044                 .create_flags           = IB_QP_CREATE_USE_GFP_NOIO
1045         };
1046
1047         struct ib_qp *tx_qp;
1048
1049         if (dev->features & NETIF_F_SG)
1050                 attr.cap.max_send_sge =
1051                         min_t(u32, priv->ca->attrs.max_sge, MAX_SKB_FRAGS + 1);
1052
1053         tx_qp = ib_create_qp(priv->pd, &attr);
1054         if (PTR_ERR(tx_qp) == -EINVAL) {
1055                 attr.create_flags &= ~IB_QP_CREATE_USE_GFP_NOIO;
1056                 tx_qp = ib_create_qp(priv->pd, &attr);
1057         }
1058         tx->max_send_sge = attr.cap.max_send_sge;
1059         return tx_qp;
1060 }
1061
1062 static int ipoib_cm_send_req(struct net_device *dev,
1063                              struct ib_cm_id *id, struct ib_qp *qp,
1064                              u32 qpn,
1065                              struct ib_sa_path_rec *pathrec)
1066 {
1067         struct ipoib_dev_priv *priv = netdev_priv(dev);
1068         struct ipoib_cm_data data = {};
1069         struct ib_cm_req_param req = {};
1070
1071         data.qpn = cpu_to_be32(priv->qp->qp_num);
1072         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1073
1074         req.primary_path                = pathrec;
1075         req.alternate_path              = NULL;
1076         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1077         req.qp_num                      = qp->qp_num;
1078         req.qp_type                     = qp->qp_type;
1079         req.private_data                = &data;
1080         req.private_data_len            = sizeof data;
1081         req.flow_control                = 0;
1082
1083         req.starting_psn                = 0; /* FIXME */
1084
1085         /*
1086          * Pick some arbitrary defaults here; we could make these
1087          * module parameters if anyone cared about setting them.
1088          */
1089         req.responder_resources         = 4;
1090         req.remote_cm_response_timeout  = 20;
1091         req.local_cm_response_timeout   = 20;
1092         req.retry_count                 = 0; /* RFC draft warns against retries */
1093         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1094         req.max_cm_retries              = 15;
1095         req.srq                         = ipoib_cm_has_srq(dev);
1096         return ib_send_cm_req(id, &req);
1097 }
1098
1099 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1100                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1101 {
1102         struct ipoib_dev_priv *priv = netdev_priv(dev);
1103         struct ib_qp_attr qp_attr;
1104         int qp_attr_mask, ret;
1105         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1106         if (ret) {
1107                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1108                 return ret;
1109         }
1110
1111         qp_attr.qp_state = IB_QPS_INIT;
1112         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1113         qp_attr.port_num = priv->port;
1114         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1115
1116         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1117         if (ret) {
1118                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1119                 return ret;
1120         }
1121         return 0;
1122 }
1123
1124 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1125                             struct ib_sa_path_rec *pathrec)
1126 {
1127         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1128         int ret;
1129
1130         p->tx_ring = __vmalloc(ipoib_sendq_size * sizeof *p->tx_ring,
1131                                GFP_NOIO, PAGE_KERNEL);
1132         if (!p->tx_ring) {
1133                 ret = -ENOMEM;
1134                 goto err_tx;
1135         }
1136         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1137
1138         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1139         if (IS_ERR(p->qp)) {
1140                 ret = PTR_ERR(p->qp);
1141                 ipoib_warn(priv, "failed to allocate tx qp: %d\n", ret);
1142                 goto err_qp;
1143         }
1144
1145         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1146         if (IS_ERR(p->id)) {
1147                 ret = PTR_ERR(p->id);
1148                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1149                 goto err_id;
1150         }
1151
1152         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1153         if (ret) {
1154                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1155                 goto err_modify;
1156         }
1157
1158         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1159         if (ret) {
1160                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1161                 goto err_send_cm;
1162         }
1163
1164         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1165                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1166
1167         return 0;
1168
1169 err_send_cm:
1170 err_modify:
1171         ib_destroy_cm_id(p->id);
1172 err_id:
1173         p->id = NULL;
1174         ib_destroy_qp(p->qp);
1175 err_qp:
1176         p->qp = NULL;
1177         vfree(p->tx_ring);
1178 err_tx:
1179         return ret;
1180 }
1181
1182 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1183 {
1184         struct ipoib_dev_priv *priv = netdev_priv(p->dev);
1185         struct ipoib_tx_buf *tx_req;
1186         unsigned long begin;
1187
1188         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1189                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1190
1191         if (p->id)
1192                 ib_destroy_cm_id(p->id);
1193
1194         if (p->tx_ring) {
1195                 /* Wait for all sends to complete */
1196                 begin = jiffies;
1197                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1198                         if (time_after(jiffies, begin + 5 * HZ)) {
1199                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1200                                            p->tx_head - p->tx_tail);
1201                                 goto timeout;
1202                         }
1203
1204                         msleep(1);
1205                 }
1206         }
1207
1208 timeout:
1209
1210         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1211                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1212                 ipoib_dma_unmap_tx(priv, tx_req);
1213                 dev_kfree_skb_any(tx_req->skb);
1214                 ++p->tx_tail;
1215                 netif_tx_lock_bh(p->dev);
1216                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1217                     netif_queue_stopped(p->dev) &&
1218                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1219                         netif_wake_queue(p->dev);
1220                 netif_tx_unlock_bh(p->dev);
1221         }
1222
1223         if (p->qp)
1224                 ib_destroy_qp(p->qp);
1225
1226         vfree(p->tx_ring);
1227         kfree(p);
1228 }
1229
1230 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1231                                struct ib_cm_event *event)
1232 {
1233         struct ipoib_cm_tx *tx = cm_id->context;
1234         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1235         struct net_device *dev = priv->dev;
1236         struct ipoib_neigh *neigh;
1237         unsigned long flags;
1238         int ret;
1239
1240         switch (event->event) {
1241         case IB_CM_DREQ_RECEIVED:
1242                 ipoib_dbg(priv, "DREQ received.\n");
1243                 ib_send_cm_drep(cm_id, NULL, 0);
1244                 break;
1245         case IB_CM_REP_RECEIVED:
1246                 ipoib_dbg(priv, "REP received.\n");
1247                 ret = ipoib_cm_rep_handler(cm_id, event);
1248                 if (ret)
1249                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1250                                        NULL, 0, NULL, 0);
1251                 break;
1252         case IB_CM_REQ_ERROR:
1253         case IB_CM_REJ_RECEIVED:
1254         case IB_CM_TIMEWAIT_EXIT:
1255                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1256                 netif_tx_lock_bh(dev);
1257                 spin_lock_irqsave(&priv->lock, flags);
1258                 neigh = tx->neigh;
1259
1260                 if (neigh) {
1261                         neigh->cm = NULL;
1262                         ipoib_neigh_free(neigh);
1263
1264                         tx->neigh = NULL;
1265                 }
1266
1267                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1268                         list_move(&tx->list, &priv->cm.reap_list);
1269                         queue_work(priv->wq, &priv->cm.reap_task);
1270                 }
1271
1272                 spin_unlock_irqrestore(&priv->lock, flags);
1273                 netif_tx_unlock_bh(dev);
1274                 break;
1275         default:
1276                 break;
1277         }
1278
1279         return 0;
1280 }
1281
1282 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1283                                        struct ipoib_neigh *neigh)
1284 {
1285         struct ipoib_dev_priv *priv = netdev_priv(dev);
1286         struct ipoib_cm_tx *tx;
1287
1288         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1289         if (!tx)
1290                 return NULL;
1291
1292         neigh->cm = tx;
1293         tx->neigh = neigh;
1294         tx->path = path;
1295         tx->dev = dev;
1296         list_add(&tx->list, &priv->cm.start_list);
1297         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1298         queue_work(priv->wq, &priv->cm.start_task);
1299         return tx;
1300 }
1301
1302 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1303 {
1304         struct ipoib_dev_priv *priv = netdev_priv(tx->dev);
1305         unsigned long flags;
1306         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1307                 spin_lock_irqsave(&priv->lock, flags);
1308                 list_move(&tx->list, &priv->cm.reap_list);
1309                 queue_work(priv->wq, &priv->cm.reap_task);
1310                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1311                           tx->neigh->daddr + 4);
1312                 tx->neigh = NULL;
1313                 spin_unlock_irqrestore(&priv->lock, flags);
1314         }
1315 }
1316
1317 #define QPN_AND_OPTIONS_OFFSET  4
1318
1319 static void ipoib_cm_tx_start(struct work_struct *work)
1320 {
1321         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1322                                                    cm.start_task);
1323         struct net_device *dev = priv->dev;
1324         struct ipoib_neigh *neigh;
1325         struct ipoib_cm_tx *p;
1326         unsigned long flags;
1327         struct ipoib_path *path;
1328         int ret;
1329
1330         struct ib_sa_path_rec pathrec;
1331         u32 qpn;
1332
1333         netif_tx_lock_bh(dev);
1334         spin_lock_irqsave(&priv->lock, flags);
1335
1336         while (!list_empty(&priv->cm.start_list)) {
1337                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1338                 list_del_init(&p->list);
1339                 neigh = p->neigh;
1340
1341                 qpn = IPOIB_QPN(neigh->daddr);
1342                 /*
1343                  * As long as the search is with these 2 locks,
1344                  * path existence indicates its validity.
1345                  */
1346                 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1347                 if (!path) {
1348                         pr_info("%s ignore not valid path %pI6\n",
1349                                 __func__,
1350                                 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1351                         goto free_neigh;
1352                 }
1353                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1354
1355                 spin_unlock_irqrestore(&priv->lock, flags);
1356                 netif_tx_unlock_bh(dev);
1357
1358                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1359
1360                 netif_tx_lock_bh(dev);
1361                 spin_lock_irqsave(&priv->lock, flags);
1362
1363                 if (ret) {
1364 free_neigh:
1365                         neigh = p->neigh;
1366                         if (neigh) {
1367                                 neigh->cm = NULL;
1368                                 ipoib_neigh_free(neigh);
1369                         }
1370                         list_del(&p->list);
1371                         kfree(p);
1372                 }
1373         }
1374
1375         spin_unlock_irqrestore(&priv->lock, flags);
1376         netif_tx_unlock_bh(dev);
1377 }
1378
1379 static void ipoib_cm_tx_reap(struct work_struct *work)
1380 {
1381         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1382                                                    cm.reap_task);
1383         struct net_device *dev = priv->dev;
1384         struct ipoib_cm_tx *p;
1385         unsigned long flags;
1386
1387         netif_tx_lock_bh(dev);
1388         spin_lock_irqsave(&priv->lock, flags);
1389
1390         while (!list_empty(&priv->cm.reap_list)) {
1391                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1392                 list_del(&p->list);
1393                 spin_unlock_irqrestore(&priv->lock, flags);
1394                 netif_tx_unlock_bh(dev);
1395                 ipoib_cm_tx_destroy(p);
1396                 netif_tx_lock_bh(dev);
1397                 spin_lock_irqsave(&priv->lock, flags);
1398         }
1399
1400         spin_unlock_irqrestore(&priv->lock, flags);
1401         netif_tx_unlock_bh(dev);
1402 }
1403
1404 static void ipoib_cm_skb_reap(struct work_struct *work)
1405 {
1406         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1407                                                    cm.skb_task);
1408         struct net_device *dev = priv->dev;
1409         struct sk_buff *skb;
1410         unsigned long flags;
1411         unsigned mtu = priv->mcast_mtu;
1412
1413         netif_tx_lock_bh(dev);
1414         spin_lock_irqsave(&priv->lock, flags);
1415
1416         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1417                 spin_unlock_irqrestore(&priv->lock, flags);
1418                 netif_tx_unlock_bh(dev);
1419
1420                 if (skb->protocol == htons(ETH_P_IP))
1421                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1422 #if IS_ENABLED(CONFIG_IPV6)
1423                 else if (skb->protocol == htons(ETH_P_IPV6))
1424                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1425 #endif
1426                 dev_kfree_skb_any(skb);
1427
1428                 netif_tx_lock_bh(dev);
1429                 spin_lock_irqsave(&priv->lock, flags);
1430         }
1431
1432         spin_unlock_irqrestore(&priv->lock, flags);
1433         netif_tx_unlock_bh(dev);
1434 }
1435
1436 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1437                            unsigned int mtu)
1438 {
1439         struct ipoib_dev_priv *priv = netdev_priv(dev);
1440         int e = skb_queue_empty(&priv->cm.skb_queue);
1441
1442         if (skb_dst(skb))
1443                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1444
1445         skb_queue_tail(&priv->cm.skb_queue, skb);
1446         if (e)
1447                 queue_work(priv->wq, &priv->cm.skb_task);
1448 }
1449
1450 static void ipoib_cm_rx_reap(struct work_struct *work)
1451 {
1452         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1453                                                 cm.rx_reap_task)->dev);
1454 }
1455
1456 static void ipoib_cm_stale_task(struct work_struct *work)
1457 {
1458         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1459                                                    cm.stale_task.work);
1460         struct ipoib_cm_rx *p;
1461         int ret;
1462
1463         spin_lock_irq(&priv->lock);
1464         while (!list_empty(&priv->cm.passive_ids)) {
1465                 /* List is sorted by LRU, start from tail,
1466                  * stop when we see a recently used entry */
1467                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1468                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1469                         break;
1470                 list_move(&p->list, &priv->cm.rx_error_list);
1471                 p->state = IPOIB_CM_RX_ERROR;
1472                 spin_unlock_irq(&priv->lock);
1473                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1474                 if (ret)
1475                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1476                 spin_lock_irq(&priv->lock);
1477         }
1478
1479         if (!list_empty(&priv->cm.passive_ids))
1480                 queue_delayed_work(priv->wq,
1481                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1482         spin_unlock_irq(&priv->lock);
1483 }
1484
1485 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1486                          char *buf)
1487 {
1488         struct ipoib_dev_priv *priv = netdev_priv(to_net_dev(d));
1489
1490         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1491                 return sprintf(buf, "connected\n");
1492         else
1493                 return sprintf(buf, "datagram\n");
1494 }
1495
1496 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1497                         const char *buf, size_t count)
1498 {
1499         struct net_device *dev = to_net_dev(d);
1500         int ret;
1501         struct ipoib_dev_priv *priv = netdev_priv(dev);
1502
1503         if (test_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags))
1504                 return -EPERM;
1505
1506         if (!rtnl_trylock())
1507                 return restart_syscall();
1508
1509         ret = ipoib_set_mode(dev, buf);
1510
1511         /* The assumption is that the function ipoib_set_mode returned
1512          * with the rtnl held by it, if not the value -EBUSY returned,
1513          * then no need to rtnl_unlock
1514          */
1515         if (ret != -EBUSY)
1516                 rtnl_unlock();
1517
1518         return (!ret || ret == -EBUSY) ? count : ret;
1519 }
1520
1521 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1522
1523 int ipoib_cm_add_mode_attr(struct net_device *dev)
1524 {
1525         return device_create_file(&dev->dev, &dev_attr_mode);
1526 }
1527
1528 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1529 {
1530         struct ipoib_dev_priv *priv = netdev_priv(dev);
1531         struct ib_srq_init_attr srq_init_attr = {
1532                 .srq_type = IB_SRQT_BASIC,
1533                 .attr = {
1534                         .max_wr  = ipoib_recvq_size,
1535                         .max_sge = max_sge
1536                 }
1537         };
1538
1539         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1540         if (IS_ERR(priv->cm.srq)) {
1541                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1542                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1543                                priv->ca->name, PTR_ERR(priv->cm.srq));
1544                 priv->cm.srq = NULL;
1545                 return;
1546         }
1547
1548         priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1549         if (!priv->cm.srq_ring) {
1550                 ib_destroy_srq(priv->cm.srq);
1551                 priv->cm.srq = NULL;
1552                 return;
1553         }
1554
1555 }
1556
1557 int ipoib_cm_dev_init(struct net_device *dev)
1558 {
1559         struct ipoib_dev_priv *priv = netdev_priv(dev);
1560         int max_srq_sge, i;
1561
1562         INIT_LIST_HEAD(&priv->cm.passive_ids);
1563         INIT_LIST_HEAD(&priv->cm.reap_list);
1564         INIT_LIST_HEAD(&priv->cm.start_list);
1565         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1566         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1567         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1568         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1569         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1570         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1571         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1572         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1573         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1574
1575         skb_queue_head_init(&priv->cm.skb_queue);
1576
1577         ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1578
1579         max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1580         ipoib_cm_create_srq(dev, max_srq_sge);
1581         if (ipoib_cm_has_srq(dev)) {
1582                 priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1583                 priv->cm.num_frags  = max_srq_sge;
1584                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1585                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1586         } else {
1587                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1588                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1589         }
1590
1591         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1592
1593         if (ipoib_cm_has_srq(dev)) {
1594                 for (i = 0; i < ipoib_recvq_size; ++i) {
1595                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1596                                                    priv->cm.num_frags - 1,
1597                                                    priv->cm.srq_ring[i].mapping,
1598                                                    GFP_KERNEL)) {
1599                                 ipoib_warn(priv, "failed to allocate "
1600                                            "receive buffer %d\n", i);
1601                                 ipoib_cm_dev_cleanup(dev);
1602                                 return -ENOMEM;
1603                         }
1604
1605                         if (ipoib_cm_post_receive_srq(dev, i)) {
1606                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1607                                            "failed for buf %d\n", i);
1608                                 ipoib_cm_dev_cleanup(dev);
1609                                 return -EIO;
1610                         }
1611                 }
1612         }
1613
1614         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1615         return 0;
1616 }
1617
1618 void ipoib_cm_dev_cleanup(struct net_device *dev)
1619 {
1620         struct ipoib_dev_priv *priv = netdev_priv(dev);
1621         int ret;
1622
1623         if (!priv->cm.srq)
1624                 return;
1625
1626         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1627
1628         ret = ib_destroy_srq(priv->cm.srq);
1629         if (ret)
1630                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1631
1632         priv->cm.srq = NULL;
1633         if (!priv->cm.srq_ring)
1634                 return;
1635
1636         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1637         priv->cm.srq_ring = NULL;
1638 }