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 #include <linux/sched/signal.h>
  42 #include <linux/sched/mm.h>
  43
  44 #include "ipoib.h"
  45
  46 int ipoib_max_conn_qp = 128;
  47
  48 module_param_named(max_nonsrq_conn_qp, ipoib_max_conn_qp, int, 0444);
  49 MODULE_PARM_DESC(max_nonsrq_conn_qp,
  50                  "Max number of connected-mode QPs per interface "
  51                  "(applied only if shared receive queue is not available)");
  52
  53 #ifdef CONFIG_INFINIBAND_IPOIB_DEBUG_DATA
  54 static int data_debug_level;
  55
  56 module_param_named(cm_data_debug_level, data_debug_level, int, 0644);
  57 MODULE_PARM_DESC(cm_data_debug_level,
  58                  "Enable data path debug tracing for connected mode if > 0");
  59 #endif
  60
  61 #define IPOIB_CM_IETF_ID 0x1000000000000000ULL
  62
  63 #define IPOIB_CM_RX_UPDATE_TIME (256 * HZ)
  64 #define IPOIB_CM_RX_TIMEOUT     (2 * 256 * HZ)
  65 #define IPOIB_CM_RX_DELAY       (3 * 256 * HZ)
  66 #define IPOIB_CM_RX_UPDATE_MASK (0x3)
  67
  68 #define IPOIB_CM_RX_RESERVE     (ALIGN(IPOIB_HARD_LEN, 16) - IPOIB_ENCAP_LEN)
  69
  70 static struct ib_qp_attr ipoib_cm_err_attr = {
  71         .qp_state = IB_QPS_ERR
  72 };
  73
  74 #define IPOIB_CM_RX_DRAIN_WRID 0xffffffff
  75
  76 static struct ib_send_wr ipoib_cm_rx_drain_wr = {
  77         .opcode = IB_WR_SEND,
  78 };
  79
  80 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
  81                                struct ib_cm_event *event);
  82
  83 static void ipoib_cm_dma_unmap_rx(struct ipoib_dev_priv *priv, int frags,
  84                                   u64 mapping[IPOIB_CM_RX_SG])
  85 {
  86         int i;
  87
  88         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
  89
  90         for (i = 0; i < frags; ++i)
  91                 ib_dma_unmap_page(priv->ca, mapping[i + 1], PAGE_SIZE, DMA_FROM_DEVICE);
  92 }
  93
  94 static int ipoib_cm_post_receive_srq(struct net_device *dev, int id)
  95 {
  96         struct ipoib_dev_priv *priv = ipoib_priv(dev);
  97         struct ib_recv_wr *bad_wr;
  98         int i, ret;
  99
 100         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 101
 102         for (i = 0; i < priv->cm.num_frags; ++i)
 103                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 104
 105         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, &bad_wr);
 106         if (unlikely(ret)) {
 107                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 108                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 109                                       priv->cm.srq_ring[id].mapping);
 110                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 111                 priv->cm.srq_ring[id].skb = NULL;
 112         }
 113
 114         return ret;
 115 }
 116
 117 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 118                                         struct ipoib_cm_rx *rx,
 119                                         struct ib_recv_wr *wr,
 120                                         struct ib_sge *sge, int id)
 121 {
 122         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 123         struct ib_recv_wr *bad_wr;
 124         int i, ret;
 125
 126         wr->wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 127
 128         for (i = 0; i < IPOIB_CM_RX_SG; ++i)
 129                 sge[i].addr = rx->rx_ring[id].mapping[i];
 130
 131         ret = ib_post_recv(rx->qp, wr, &bad_wr);
 132         if (unlikely(ret)) {
 133                 ipoib_warn(priv, "post recv failed for buf %d (%d)\n", id, ret);
 134                 ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 135                                       rx->rx_ring[id].mapping);
 136                 dev_kfree_skb_any(rx->rx_ring[id].skb);
 137                 rx->rx_ring[id].skb = NULL;
 138         }
 139
 140         return ret;
 141 }
 142
 143 static struct sk_buff *ipoib_cm_alloc_rx_skb(struct net_device *dev,
 144                                              struct ipoib_cm_rx_buf *rx_ring,
 145                                              int id, int frags,
 146                                              u64 mapping[IPOIB_CM_RX_SG],
 147                                              gfp_t gfp)
 148 {
 149         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 150         struct sk_buff *skb;
 151         int i;
 152
 153         skb = dev_alloc_skb(ALIGN(IPOIB_CM_HEAD_SIZE + IPOIB_PSEUDO_LEN, 16));
 154         if (unlikely(!skb))
 155                 return NULL;
 156
 157         /*
 158          * IPoIB adds a IPOIB_ENCAP_LEN byte header, this will align the
 159          * IP header to a multiple of 16.
 160          */
 161         skb_reserve(skb, IPOIB_CM_RX_RESERVE);
 162
 163         mapping[0] = ib_dma_map_single(priv->ca, skb->data, IPOIB_CM_HEAD_SIZE,
 164                                        DMA_FROM_DEVICE);
 165         if (unlikely(ib_dma_mapping_error(priv->ca, mapping[0]))) {
 166                 dev_kfree_skb_any(skb);
 167                 return NULL;
 168         }
 169
 170         for (i = 0; i < frags; i++) {
 171                 struct page *page = alloc_page(gfp);
 172
 173                 if (!page)
 174                         goto partial_error;
 175                 skb_fill_page_desc(skb, i, page, 0, PAGE_SIZE);
 176
 177                 mapping[i + 1] = ib_dma_map_page(priv->ca, page,
 178                                                  0, PAGE_SIZE, DMA_FROM_DEVICE);
 179                 if (unlikely(ib_dma_mapping_error(priv->ca, mapping[i + 1])))
 180                         goto partial_error;
 181         }
 182
 183         rx_ring[id].skb = skb;
 184         return skb;
 185
 186 partial_error:
 187
 188         ib_dma_unmap_single(priv->ca, mapping[0], IPOIB_CM_HEAD_SIZE, DMA_FROM_DEVICE);
 189
 190         for (; i > 0; --i)
 191                 ib_dma_unmap_page(priv->ca, mapping[i], PAGE_SIZE, DMA_FROM_DEVICE);
 192
 193         dev_kfree_skb_any(skb);
 194         return NULL;
 195 }
 196
 197 static void ipoib_cm_free_rx_ring(struct net_device *dev,
 198                                   struct ipoib_cm_rx_buf *rx_ring)
 199 {
 200         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 201         int i;
 202
 203         for (i = 0; i < ipoib_recvq_size; ++i)
 204                 if (rx_ring[i].skb) {
 205                         ipoib_cm_dma_unmap_rx(priv, IPOIB_CM_RX_SG - 1,
 206                                               rx_ring[i].mapping);
 207                         dev_kfree_skb_any(rx_ring[i].skb);
 208                 }
 209
 210         vfree(rx_ring);
 211 }
 212
 213 static void ipoib_cm_start_rx_drain(struct ipoib_dev_priv *priv)
 214 {
 215         struct ib_send_wr *bad_wr;
 216         struct ipoib_cm_rx *p;
 217
 218         /* We only reserved 1 extra slot in CQ for drain WRs, so
 219          * make sure we have at most 1 outstanding WR. */
 220         if (list_empty(&priv->cm.rx_flush_list) ||
 221             !list_empty(&priv->cm.rx_drain_list))
 222                 return;
 223
 224         /*
 225          * QPs on flush list are error state.  This way, a "flush
 226          * error" WC will be immediately generated for each WR we post.
 227          */
 228         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 229         ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
 230         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, &bad_wr))
 231                 ipoib_warn(priv, "failed to post drain wr\n");
 232
 233         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 234 }
 235
 236 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 237 {
 238         struct ipoib_cm_rx *p = ctx;
 239         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
 240         unsigned long flags;
 241
 242         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 243                 return;
 244
 245         spin_lock_irqsave(&priv->lock, flags);
 246         list_move(&p->list, &priv->cm.rx_flush_list);
 247         p->state = IPOIB_CM_RX_FLUSH;
 248         ipoib_cm_start_rx_drain(priv);
 249         spin_unlock_irqrestore(&priv->lock, flags);
 250 }
 251
 252 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 253                                            struct ipoib_cm_rx *p)
 254 {
 255         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 256         struct ib_qp_init_attr attr = {
 257                 .event_handler = ipoib_cm_rx_event_handler,
 258                 .send_cq = priv->recv_cq, /* For drain WR */
 259                 .recv_cq = priv->recv_cq,
 260                 .srq = priv->cm.srq,
 261                 .cap.max_send_wr = 1, /* For drain WR */
 262                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 263                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 264                 .qp_type = IB_QPT_RC,
 265                 .qp_context = p,
 266         };
 267
 268         if (!ipoib_cm_has_srq(dev)) {
 269                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 270                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 271         }
 272
 273         return ib_create_qp(priv->pd, &attr);
 274 }
 275
 276 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 277                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 278                                  unsigned psn)
 279 {
 280         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 281         struct ib_qp_attr qp_attr;
 282         int qp_attr_mask, ret;
 283
 284         qp_attr.qp_state = IB_QPS_INIT;
 285         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 286         if (ret) {
 287                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 288                 return ret;
 289         }
 290         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 291         if (ret) {
 292                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 293                 return ret;
 294         }
 295         qp_attr.qp_state = IB_QPS_RTR;
 296         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 297         if (ret) {
 298                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 299                 return ret;
 300         }
 301         qp_attr.rq_psn = psn;
 302         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 303         if (ret) {
 304                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 305                 return ret;
 306         }
 307
 308         /*
 309          * Current Mellanox HCA firmware won't generate completions
 310          * with error for drain WRs unless the QP has been moved to
 311          * RTS first. This work-around leaves a window where a QP has
 312          * moved to error asynchronously, but this will eventually get
 313          * fixed in firmware, so let's not error out if modify QP
 314          * fails.
 315          */
 316         qp_attr.qp_state = IB_QPS_RTS;
 317         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 318         if (ret) {
 319                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 320                 return 0;
 321         }
 322         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 323         if (ret) {
 324                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 325                 return 0;
 326         }
 327
 328         return 0;
 329 }
 330
 331 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 332                                 struct ib_recv_wr *wr,
 333                                 struct ib_sge *sge)
 334 {
 335         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 336         int i;
 337
 338         for (i = 0; i < priv->cm.num_frags; ++i)
 339                 sge[i].lkey = priv->pd->local_dma_lkey;
 340
 341         sge[0].length = IPOIB_CM_HEAD_SIZE;
 342         for (i = 1; i < priv->cm.num_frags; ++i)
 343                 sge[i].length = PAGE_SIZE;
 344
 345         wr->next    = NULL;
 346         wr->sg_list = sge;
 347         wr->num_sge = priv->cm.num_frags;
 348 }
 349
 350 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 351                                    struct ipoib_cm_rx *rx)
 352 {
 353         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 354         struct {
 355                 struct ib_recv_wr wr;
 356                 struct ib_sge sge[IPOIB_CM_RX_SG];
 357         } *t;
 358         int ret;
 359         int i;
 360
 361         rx->rx_ring = vzalloc(ipoib_recvq_size * sizeof *rx->rx_ring);
 362         if (!rx->rx_ring)
 363                 return -ENOMEM;
 364
 365         t = kmalloc(sizeof *t, GFP_KERNEL);
 366         if (!t) {
 367                 ret = -ENOMEM;
 368                 goto err_free_1;
 369         }
 370
 371         ipoib_cm_init_rx_wr(dev, &t->wr, t->sge);
 372
 373         spin_lock_irq(&priv->lock);
 374
 375         if (priv->cm.nonsrq_conn_qp >= ipoib_max_conn_qp) {
 376                 spin_unlock_irq(&priv->lock);
 377                 ib_send_cm_rej(cm_id, IB_CM_REJ_NO_QP, NULL, 0, NULL, 0);
 378                 ret = -EINVAL;
 379                 goto err_free;
 380         } else
 381                 ++priv->cm.nonsrq_conn_qp;
 382
 383         spin_unlock_irq(&priv->lock);
 384
 385         for (i = 0; i < ipoib_recvq_size; ++i) {
 386                 if (!ipoib_cm_alloc_rx_skb(dev, rx->rx_ring, i, IPOIB_CM_RX_SG - 1,
 387                                            rx->rx_ring[i].mapping,
 388                                            GFP_KERNEL)) {
 389                         ipoib_warn(priv, "failed to allocate receive buffer %d\n", i);
 390                         ret = -ENOMEM;
 391                         goto err_count;
 392                 }
 393                 ret = ipoib_cm_post_receive_nonsrq(dev, rx, &t->wr, t->sge, i);
 394                 if (ret) {
 395                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq "
 396                                    "failed for buf %d\n", i);
 397                         ret = -EIO;
 398                         goto err_count;
 399                 }
 400         }
 401
 402         rx->recv_count = ipoib_recvq_size;
 403
 404         kfree(t);
 405
 406         return 0;
 407
 408 err_count:
 409         spin_lock_irq(&priv->lock);
 410         --priv->cm.nonsrq_conn_qp;
 411         spin_unlock_irq(&priv->lock);
 412
 413 err_free:
 414         kfree(t);
 415
 416 err_free_1:
 417         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 418
 419         return ret;
 420 }
 421
 422 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 423                              struct ib_qp *qp, struct ib_cm_req_event_param *req,
 424                              unsigned psn)
 425 {
 426         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 427         struct ipoib_cm_data data = {};
 428         struct ib_cm_rep_param rep = {};
 429
 430         data.qpn = cpu_to_be32(priv->qp->qp_num);
 431         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 432
 433         rep.private_data = &data;
 434         rep.private_data_len = sizeof data;
 435         rep.flow_control = 0;
 436         rep.rnr_retry_count = req->rnr_retry_count;
 437         rep.srq = ipoib_cm_has_srq(dev);
 438         rep.qp_num = qp->qp_num;
 439         rep.starting_psn = psn;
 440         return ib_send_cm_rep(cm_id, &rep);
 441 }
 442
 443 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 444 {
 445         struct net_device *dev = cm_id->context;
 446         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 447         struct ipoib_cm_rx *p;
 448         unsigned psn;
 449         int ret;
 450
 451         ipoib_dbg(priv, "REQ arrived\n");
 452         p = kzalloc(sizeof *p, GFP_KERNEL);
 453         if (!p)
 454                 return -ENOMEM;
 455         p->dev = dev;
 456         p->id = cm_id;
 457         cm_id->context = p;
 458         p->state = IPOIB_CM_RX_LIVE;
 459         p->jiffies = jiffies;
 460         INIT_LIST_HEAD(&p->list);
 461
 462         p->qp = ipoib_cm_create_rx_qp(dev, p);
 463         if (IS_ERR(p->qp)) {
 464                 ret = PTR_ERR(p->qp);
 465                 goto err_qp;
 466         }
 467
 468         psn = prandom_u32() & 0xffffff;
 469         ret = ipoib_cm_modify_rx_qp(dev, cm_id, p->qp, psn);
 470         if (ret)
 471                 goto err_modify;
 472
 473         if (!ipoib_cm_has_srq(dev)) {
 474                 ret = ipoib_cm_nonsrq_init_rx(dev, cm_id, p);
 475                 if (ret)
 476                         goto err_modify;
 477         }
 478
 479         spin_lock_irq(&priv->lock);
 480         queue_delayed_work(priv->wq,
 481                            &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
 482         /* Add this entry to passive ids list head, but do not re-add it
 483          * if IB_EVENT_QP_LAST_WQE_REACHED has moved it to flush list. */
 484         p->jiffies = jiffies;
 485         if (p->state == IPOIB_CM_RX_LIVE)
 486                 list_move(&p->list, &priv->cm.passive_ids);
 487         spin_unlock_irq(&priv->lock);
 488
 489         ret = ipoib_cm_send_rep(dev, cm_id, p->qp, &event->param.req_rcvd, psn);
 490         if (ret) {
 491                 ipoib_warn(priv, "failed to send REP: %d\n", ret);
 492                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 493                         ipoib_warn(priv, "unable to move qp to error state\n");
 494         }
 495         return 0;
 496
 497 err_modify:
 498         ib_destroy_qp(p->qp);
 499 err_qp:
 500         kfree(p);
 501         return ret;
 502 }
 503
 504 static int ipoib_cm_rx_handler(struct ib_cm_id *cm_id,
 505                                struct ib_cm_event *event)
 506 {
 507         struct ipoib_cm_rx *p;
 508         struct ipoib_dev_priv *priv;
 509
 510         switch (event->event) {
 511         case IB_CM_REQ_RECEIVED:
 512                 return ipoib_cm_req_handler(cm_id, event);
 513         case IB_CM_DREQ_RECEIVED:
 514                 p = cm_id->context;
 515                 ib_send_cm_drep(cm_id, NULL, 0);
 516                 /* Fall through */
 517         case IB_CM_REJ_RECEIVED:
 518                 p = cm_id->context;
 519                 priv = ipoib_priv(p->dev);
 520                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 521                         ipoib_warn(priv, "unable to move qp to error state\n");
 522                 /* Fall through */
 523         default:
 524                 return 0;
 525         }
 526 }
 527 /* Adjust length of skb with fragments to match received data */
 528 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 529                           unsigned int length, struct sk_buff *toskb)
 530 {
 531         int i, num_frags;
 532         unsigned int size;
 533
 534         /* put header into skb */
 535         size = min(length, hdr_space);
 536         skb->tail += size;
 537         skb->len += size;
 538         length -= size;
 539
 540         num_frags = skb_shinfo(skb)->nr_frags;
 541         for (i = 0; i < num_frags; i++) {
 542                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 543
 544                 if (length == 0) {
 545                         /* don't need this page */
 546                         skb_fill_page_desc(toskb, i, skb_frag_page(frag),
 547                                            0, PAGE_SIZE);
 548                         --skb_shinfo(skb)->nr_frags;
 549                 } else {
 550                         size = min(length, (unsigned) PAGE_SIZE);
 551
 552                         skb_frag_size_set(frag, size);
 553                         skb->data_len += size;
 554                         skb->truesize += size;
 555                         skb->len += size;
 556                         length -= size;
 557                 }
 558         }
 559 }
 560
 561 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 562 {
 563         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 564         struct ipoib_cm_rx_buf *rx_ring;
 565         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 566         struct sk_buff *skb, *newskb;
 567         struct ipoib_cm_rx *p;
 568         unsigned long flags;
 569         u64 mapping[IPOIB_CM_RX_SG];
 570         int frags;
 571         int has_srq;
 572         struct sk_buff *small_skb;
 573
 574         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 575                        wr_id, wc->status);
 576
 577         if (unlikely(wr_id >= ipoib_recvq_size)) {
 578                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 579                         spin_lock_irqsave(&priv->lock, flags);
 580                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 581                         ipoib_cm_start_rx_drain(priv);
 582                         queue_work(priv->wq, &priv->cm.rx_reap_task);
 583                         spin_unlock_irqrestore(&priv->lock, flags);
 584                 } else
 585                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 586                                    wr_id, ipoib_recvq_size);
 587                 return;
 588         }
 589
 590         p = wc->qp->qp_context;
 591
 592         has_srq = ipoib_cm_has_srq(dev);
 593         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 594
 595         skb = rx_ring[wr_id].skb;
 596
 597         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 598                 ipoib_dbg(priv, "cm recv error "
 599                            "(status=%d, wrid=%d vend_err %x)\n",
 600                            wc->status, wr_id, wc->vendor_err);
 601                 ++dev->stats.rx_dropped;
 602                 if (has_srq)
 603                         goto repost;
 604                 else {
 605                         if (!--p->recv_count) {
 606                                 spin_lock_irqsave(&priv->lock, flags);
 607                                 list_move(&p->list, &priv->cm.rx_reap_list);
 608                                 spin_unlock_irqrestore(&priv->lock, flags);
 609                                 queue_work(priv->wq, &priv->cm.rx_reap_task);
 610                         }
 611                         return;
 612                 }
 613         }
 614
 615         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 616                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 617                         spin_lock_irqsave(&priv->lock, flags);
 618                         p->jiffies = jiffies;
 619                         /* Move this entry to list head, but do not re-add it
 620                          * if it has been moved out of list. */
 621                         if (p->state == IPOIB_CM_RX_LIVE)
 622                                 list_move(&p->list, &priv->cm.passive_ids);
 623                         spin_unlock_irqrestore(&priv->lock, flags);
 624                 }
 625         }
 626
 627         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 628                 int dlen = wc->byte_len;
 629
 630                 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
 631                 if (small_skb) {
 632                         skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
 633                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 634                                                    dlen, DMA_FROM_DEVICE);
 635                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 636                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 637                                                       dlen, DMA_FROM_DEVICE);
 638                         skb_put(small_skb, dlen);
 639                         skb = small_skb;
 640                         goto copied;
 641                 }
 642         }
 643
 644         frags = PAGE_ALIGN(wc->byte_len - min(wc->byte_len,
 645                                               (unsigned)IPOIB_CM_HEAD_SIZE)) / PAGE_SIZE;
 646
 647         newskb = ipoib_cm_alloc_rx_skb(dev, rx_ring, wr_id, frags,
 648                                        mapping, GFP_ATOMIC);
 649         if (unlikely(!newskb)) {
 650                 /*
 651                  * If we can't allocate a new RX buffer, dump
 652                  * this packet and reuse the old buffer.
 653                  */
 654                 ipoib_dbg(priv, "failed to allocate receive buffer %d\n", wr_id);
 655                 ++dev->stats.rx_dropped;
 656                 goto repost;
 657         }
 658
 659         ipoib_cm_dma_unmap_rx(priv, frags, rx_ring[wr_id].mapping);
 660         memcpy(rx_ring[wr_id].mapping, mapping, (frags + 1) * sizeof *mapping);
 661
 662         ipoib_dbg_data(priv, "received %d bytes, SLID 0x%04x\n",
 663                        wc->byte_len, wc->slid);
 664
 665         skb_put_frags(skb, IPOIB_CM_HEAD_SIZE, wc->byte_len, newskb);
 666
 667 copied:
 668         skb->protocol = ((struct ipoib_header *) skb->data)->proto;
 669         skb_add_pseudo_hdr(skb);
 670
 671         ++dev->stats.rx_packets;
 672         dev->stats.rx_bytes += skb->len;
 673
 674         skb->dev = dev;
 675         /* XXX get correct PACKET_ type here */
 676         skb->pkt_type = PACKET_HOST;
 677         netif_receive_skb(skb);
 678
 679 repost:
 680         if (has_srq) {
 681                 if (unlikely(ipoib_cm_post_receive_srq(dev, wr_id)))
 682                         ipoib_warn(priv, "ipoib_cm_post_receive_srq failed "
 683                                    "for buf %d\n", wr_id);
 684         } else {
 685                 if (unlikely(ipoib_cm_post_receive_nonsrq(dev, p,
 686                                                           &priv->cm.rx_wr,
 687                                                           priv->cm.rx_sge,
 688                                                           wr_id))) {
 689                         --p->recv_count;
 690                         ipoib_warn(priv, "ipoib_cm_post_receive_nonsrq failed "
 691                                    "for buf %d\n", wr_id);
 692                 }
 693         }
 694 }
 695
 696 static inline int post_send(struct ipoib_dev_priv *priv,
 697                             struct ipoib_cm_tx *tx,
 698                             unsigned int wr_id,
 699                             struct ipoib_tx_buf *tx_req)
 700 {
 701         struct ib_send_wr *bad_wr;
 702
 703         ipoib_build_sge(priv, tx_req);
 704
 705         priv->tx_wr.wr.wr_id    = wr_id | IPOIB_OP_CM;
 706
 707         return ib_post_send(tx->qp, &priv->tx_wr.wr, &bad_wr);
 708 }
 709
 710 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 711 {
 712         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 713         struct ipoib_tx_buf *tx_req;
 714         int rc;
 715         unsigned usable_sge = tx->max_send_sge - !!skb_headlen(skb);
 716
 717         if (unlikely(skb->len > tx->mtu)) {
 718                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 719                            skb->len, tx->mtu);
 720                 ++dev->stats.tx_dropped;
 721                 ++dev->stats.tx_errors;
 722                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 723                 return;
 724         }
 725         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 726                 if (skb_linearize(skb) < 0) {
 727                         ipoib_warn(priv, "skb could not be linearized\n");
 728                         ++dev->stats.tx_dropped;
 729                         ++dev->stats.tx_errors;
 730                         dev_kfree_skb_any(skb);
 731                         return;
 732                 }
 733                 /* Does skb_linearize return ok without reducing nr_frags? */
 734                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 735                         ipoib_warn(priv, "too many frags after skb linearize\n");
 736                         ++dev->stats.tx_dropped;
 737                         ++dev->stats.tx_errors;
 738                         dev_kfree_skb_any(skb);
 739                         return;
 740                 }
 741         }
 742         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 743                        tx->tx_head, skb->len, tx->qp->qp_num);
 744
 745         /*
 746          * We put the skb into the tx_ring _before_ we call post_send()
 747          * because it's entirely possible that the completion handler will
 748          * run before we execute anything after the post_send().  That
 749          * means we have to make sure everything is properly recorded and
 750          * our state is consistent before we call post_send().
 751          */
 752         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 753         tx_req->skb = skb;
 754
 755         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 756                 ++dev->stats.tx_errors;
 757                 dev_kfree_skb_any(skb);
 758                 return;
 759         }
 760
 761         skb_orphan(skb);
 762         skb_dst_drop(skb);
 763
 764         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
 765         if (unlikely(rc)) {
 766                 ipoib_warn(priv, "post_send failed, error %d\n", rc);
 767                 ++dev->stats.tx_errors;
 768                 ipoib_dma_unmap_tx(priv, tx_req);
 769                 dev_kfree_skb_any(skb);
 770         } else {
 771                 netif_trans_update(dev);
 772                 ++tx->tx_head;
 773
 774                 if (++priv->tx_outstanding == ipoib_sendq_size) {
 775                         ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 776                                   tx->qp->qp_num);
 777                         netif_stop_queue(dev);
 778                         rc = ib_req_notify_cq(priv->send_cq,
 779                                 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
 780                         if (rc < 0)
 781                                 ipoib_warn(priv, "request notify on send CQ failed\n");
 782                         else if (rc)
 783                                 ipoib_send_comp_handler(priv->send_cq, dev);
 784                 }
 785         }
 786 }
 787
 788 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 789 {
 790         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 791         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 792         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 793         struct ipoib_tx_buf *tx_req;
 794         unsigned long flags;
 795
 796         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 797                        wr_id, wc->status);
 798
 799         if (unlikely(wr_id >= ipoib_sendq_size)) {
 800                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 801                            wr_id, ipoib_sendq_size);
 802                 return;
 803         }
 804
 805         tx_req = &tx->tx_ring[wr_id];
 806
 807         ipoib_dma_unmap_tx(priv, tx_req);
 808
 809         /* FIXME: is this right? Shouldn't we only increment on success? */
 810         ++dev->stats.tx_packets;
 811         dev->stats.tx_bytes += tx_req->skb->len;
 812
 813         dev_kfree_skb_any(tx_req->skb);
 814
 815         netif_tx_lock(dev);
 816
 817         ++tx->tx_tail;
 818         if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
 819             netif_queue_stopped(dev) &&
 820             test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
 821                 netif_wake_queue(dev);
 822
 823         if (wc->status != IB_WC_SUCCESS &&
 824             wc->status != IB_WC_WR_FLUSH_ERR) {
 825                 struct ipoib_neigh *neigh;
 826
 827                 if (wc->status != IB_WC_RNR_RETRY_EXC_ERR)
 828                         ipoib_warn(priv, "failed cm send event (status=%d, wrid=%d vend_err %x)\n",
 829                                    wc->status, wr_id, wc->vendor_err);
 830                 else
 831                         ipoib_dbg(priv, "failed cm send event (status=%d, wrid=%d vend_err %x)\n",
 832                                   wc->status, wr_id, wc->vendor_err);
 833
 834                 spin_lock_irqsave(&priv->lock, flags);
 835                 neigh = tx->neigh;
 836
 837                 if (neigh) {
 838                         neigh->cm = NULL;
 839                         ipoib_neigh_free(neigh);
 840
 841                         tx->neigh = NULL;
 842                 }
 843
 844                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 845                         list_move(&tx->list, &priv->cm.reap_list);
 846                         queue_work(priv->wq, &priv->cm.reap_task);
 847                 }
 848
 849                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 850
 851                 spin_unlock_irqrestore(&priv->lock, flags);
 852         }
 853
 854         netif_tx_unlock(dev);
 855 }
 856
 857 int ipoib_cm_dev_open(struct net_device *dev)
 858 {
 859         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 860         int ret;
 861
 862         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 863                 return 0;
 864
 865         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 866         if (IS_ERR(priv->cm.id)) {
 867                 printk(KERN_WARNING "%s: failed to create CM ID\n", priv->ca->name);
 868                 ret = PTR_ERR(priv->cm.id);
 869                 goto err_cm;
 870         }
 871
 872         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 873                            0);
 874         if (ret) {
 875                 printk(KERN_WARNING "%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 876                        IPOIB_CM_IETF_ID | priv->qp->qp_num);
 877                 goto err_listen;
 878         }
 879
 880         return 0;
 881
 882 err_listen:
 883         ib_destroy_cm_id(priv->cm.id);
 884 err_cm:
 885         priv->cm.id = NULL;
 886         return ret;
 887 }
 888
 889 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 890 {
 891         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 892         struct ipoib_cm_rx *rx, *n;
 893         LIST_HEAD(list);
 894
 895         spin_lock_irq(&priv->lock);
 896         list_splice_init(&priv->cm.rx_reap_list, &list);
 897         spin_unlock_irq(&priv->lock);
 898
 899         list_for_each_entry_safe(rx, n, &list, list) {
 900                 ib_destroy_cm_id(rx->id);
 901                 ib_destroy_qp(rx->qp);
 902                 if (!ipoib_cm_has_srq(dev)) {
 903                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 904                         spin_lock_irq(&priv->lock);
 905                         --priv->cm.nonsrq_conn_qp;
 906                         spin_unlock_irq(&priv->lock);
 907                 }
 908                 kfree(rx);
 909         }
 910 }
 911
 912 void ipoib_cm_dev_stop(struct net_device *dev)
 913 {
 914         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 915         struct ipoib_cm_rx *p;
 916         unsigned long begin;
 917         int ret;
 918
 919         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 920                 return;
 921
 922         ib_destroy_cm_id(priv->cm.id);
 923         priv->cm.id = NULL;
 924
 925         spin_lock_irq(&priv->lock);
 926         while (!list_empty(&priv->cm.passive_ids)) {
 927                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 928                 list_move(&p->list, &priv->cm.rx_error_list);
 929                 p->state = IPOIB_CM_RX_ERROR;
 930                 spin_unlock_irq(&priv->lock);
 931                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 932                 if (ret)
 933                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 934                 spin_lock_irq(&priv->lock);
 935         }
 936
 937         /* Wait for all RX to be drained */
 938         begin = jiffies;
 939
 940         while (!list_empty(&priv->cm.rx_error_list) ||
 941                !list_empty(&priv->cm.rx_flush_list) ||
 942                !list_empty(&priv->cm.rx_drain_list)) {
 943                 if (time_after(jiffies, begin + 5 * HZ)) {
 944                         ipoib_warn(priv, "RX drain timing out\n");
 945
 946                         /*
 947                          * assume the HW is wedged and just free up everything.
 948                          */
 949                         list_splice_init(&priv->cm.rx_flush_list,
 950                                          &priv->cm.rx_reap_list);
 951                         list_splice_init(&priv->cm.rx_error_list,
 952                                          &priv->cm.rx_reap_list);
 953                         list_splice_init(&priv->cm.rx_drain_list,
 954                                          &priv->cm.rx_reap_list);
 955                         break;
 956                 }
 957                 spin_unlock_irq(&priv->lock);
 958                 usleep_range(1000, 2000);
 959                 ipoib_drain_cq(dev);
 960                 spin_lock_irq(&priv->lock);
 961         }
 962
 963         spin_unlock_irq(&priv->lock);
 964
 965         ipoib_cm_free_rx_reap_list(dev);
 966
 967         cancel_delayed_work(&priv->cm.stale_task);
 968 }
 969
 970 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
 971 {
 972         struct ipoib_cm_tx *p = cm_id->context;
 973         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
 974         struct ipoib_cm_data *data = event->private_data;
 975         struct sk_buff_head skqueue;
 976         struct ib_qp_attr qp_attr;
 977         int qp_attr_mask, ret;
 978         struct sk_buff *skb;
 979
 980         p->mtu = be32_to_cpu(data->mtu);
 981
 982         if (p->mtu <= IPOIB_ENCAP_LEN) {
 983                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 984                            p->mtu, IPOIB_ENCAP_LEN);
 985                 return -EINVAL;
 986         }
 987
 988         qp_attr.qp_state = IB_QPS_RTR;
 989         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 990         if (ret) {
 991                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 992                 return ret;
 993         }
 994
 995         qp_attr.rq_psn = 0 /* FIXME */;
 996         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
 997         if (ret) {
 998                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 999                 return ret;
1000         }
1001
1002         qp_attr.qp_state = IB_QPS_RTS;
1003         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1004         if (ret) {
1005                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1006                 return ret;
1007         }
1008         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1009         if (ret) {
1010                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1011                 return ret;
1012         }
1013
1014         skb_queue_head_init(&skqueue);
1015
1016         spin_lock_irq(&priv->lock);
1017         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1018         if (p->neigh)
1019                 while ((skb = __skb_dequeue(&p->neigh->queue)))
1020                         __skb_queue_tail(&skqueue, skb);
1021         spin_unlock_irq(&priv->lock);
1022
1023         while ((skb = __skb_dequeue(&skqueue))) {
1024                 skb->dev = p->dev;
1025                 ret = dev_queue_xmit(skb);
1026                 if (ret)
1027                         ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1028                                    __func__, ret);
1029         }
1030
1031         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1032         if (ret) {
1033                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1034                 return ret;
1035         }
1036         return 0;
1037 }
1038
1039 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1040 {
1041         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1042         struct ib_qp_init_attr attr = {
1043                 .send_cq                = priv->recv_cq,
1044                 .recv_cq                = priv->recv_cq,
1045                 .srq                    = priv->cm.srq,
1046                 .cap.max_send_wr        = ipoib_sendq_size,
1047                 .cap.max_send_sge       = 1,
1048                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1049                 .qp_type                = IB_QPT_RC,
1050                 .qp_context             = tx,
1051                 .create_flags           = 0
1052         };
1053         struct ib_qp *tx_qp;
1054
1055         if (dev->features & NETIF_F_SG)
1056                 attr.cap.max_send_sge =
1057                         min_t(u32, priv->ca->attrs.max_sge, MAX_SKB_FRAGS + 1);
1058
1059         tx_qp = ib_create_qp(priv->pd, &attr);
1060         tx->max_send_sge = attr.cap.max_send_sge;
1061         return tx_qp;
1062 }
1063
1064 static int ipoib_cm_send_req(struct net_device *dev,
1065                              struct ib_cm_id *id, struct ib_qp *qp,
1066                              u32 qpn,
1067                              struct sa_path_rec *pathrec)
1068 {
1069         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1070         struct ipoib_cm_data data = {};
1071         struct ib_cm_req_param req = {};
1072
1073         data.qpn = cpu_to_be32(priv->qp->qp_num);
1074         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1075
1076         req.primary_path                = pathrec;
1077         req.alternate_path              = NULL;
1078         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1079         req.qp_num                      = qp->qp_num;
1080         req.qp_type                     = qp->qp_type;
1081         req.private_data                = &data;
1082         req.private_data_len            = sizeof data;
1083         req.flow_control                = 0;
1084
1085         req.starting_psn                = 0; /* FIXME */
1086
1087         /*
1088          * Pick some arbitrary defaults here; we could make these
1089          * module parameters if anyone cared about setting them.
1090          */
1091         req.responder_resources         = 4;
1092         req.remote_cm_response_timeout  = 20;
1093         req.local_cm_response_timeout   = 20;
1094         req.retry_count                 = 0; /* RFC draft warns against retries */
1095         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1096         req.max_cm_retries              = 15;
1097         req.srq                         = ipoib_cm_has_srq(dev);
1098         return ib_send_cm_req(id, &req);
1099 }
1100
1101 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1102                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1103 {
1104         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1105         struct ib_qp_attr qp_attr;
1106         int qp_attr_mask, ret;
1107         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1108         if (ret) {
1109                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1110                 return ret;
1111         }
1112
1113         qp_attr.qp_state = IB_QPS_INIT;
1114         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1115         qp_attr.port_num = priv->port;
1116         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1117
1118         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1119         if (ret) {
1120                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1121                 return ret;
1122         }
1123         return 0;
1124 }
1125
1126 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1127                             struct sa_path_rec *pathrec)
1128 {
1129         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1130         unsigned int noio_flag;
1131         int ret;
1132
1133         noio_flag = memalloc_noio_save();
1134         p->tx_ring = vzalloc(ipoib_sendq_size * sizeof(*p->tx_ring));
1135         if (!p->tx_ring) {
1136                 ret = -ENOMEM;
1137                 goto err_tx;
1138         }
1139         memset(p->tx_ring, 0, ipoib_sendq_size * sizeof *p->tx_ring);
1140
1141         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1142         memalloc_noio_restore(noio_flag);
1143         if (IS_ERR(p->qp)) {
1144                 ret = PTR_ERR(p->qp);
1145                 ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1146                 goto err_qp;
1147         }
1148
1149         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1150         if (IS_ERR(p->id)) {
1151                 ret = PTR_ERR(p->id);
1152                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1153                 goto err_id;
1154         }
1155
1156         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1157         if (ret) {
1158                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1159                 goto err_modify_send;
1160         }
1161
1162         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1163         if (ret) {
1164                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1165                 goto err_modify_send;
1166         }
1167
1168         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1169                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1170
1171         return 0;
1172
1173 err_modify_send:
1174         ib_destroy_cm_id(p->id);
1175 err_id:
1176         p->id = NULL;
1177         ib_destroy_qp(p->qp);
1178 err_qp:
1179         p->qp = NULL;
1180         vfree(p->tx_ring);
1181 err_tx:
1182         return ret;
1183 }
1184
1185 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1186 {
1187         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1188         struct ipoib_tx_buf *tx_req;
1189         unsigned long begin;
1190
1191         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1192                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1193
1194         if (p->id)
1195                 ib_destroy_cm_id(p->id);
1196
1197         if (p->tx_ring) {
1198                 /* Wait for all sends to complete */
1199                 begin = jiffies;
1200                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1201                         if (time_after(jiffies, begin + 5 * HZ)) {
1202                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1203                                            p->tx_head - p->tx_tail);
1204                                 goto timeout;
1205                         }
1206
1207                         usleep_range(1000, 2000);
1208                 }
1209         }
1210
1211 timeout:
1212
1213         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1214                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1215                 ipoib_dma_unmap_tx(priv, tx_req);
1216                 dev_kfree_skb_any(tx_req->skb);
1217                 ++p->tx_tail;
1218                 netif_tx_lock_bh(p->dev);
1219                 if (unlikely(--priv->tx_outstanding == ipoib_sendq_size >> 1) &&
1220                     netif_queue_stopped(p->dev) &&
1221                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1222                         netif_wake_queue(p->dev);
1223                 netif_tx_unlock_bh(p->dev);
1224         }
1225
1226         if (p->qp)
1227                 ib_destroy_qp(p->qp);
1228
1229         vfree(p->tx_ring);
1230         kfree(p);
1231 }
1232
1233 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1234                                struct ib_cm_event *event)
1235 {
1236         struct ipoib_cm_tx *tx = cm_id->context;
1237         struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1238         struct net_device *dev = priv->dev;
1239         struct ipoib_neigh *neigh;
1240         unsigned long flags;
1241         int ret;
1242
1243         switch (event->event) {
1244         case IB_CM_DREQ_RECEIVED:
1245                 ipoib_dbg(priv, "DREQ received.\n");
1246                 ib_send_cm_drep(cm_id, NULL, 0);
1247                 break;
1248         case IB_CM_REP_RECEIVED:
1249                 ipoib_dbg(priv, "REP received.\n");
1250                 ret = ipoib_cm_rep_handler(cm_id, event);
1251                 if (ret)
1252                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1253                                        NULL, 0, NULL, 0);
1254                 break;
1255         case IB_CM_REQ_ERROR:
1256         case IB_CM_REJ_RECEIVED:
1257         case IB_CM_TIMEWAIT_EXIT:
1258                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1259                 netif_tx_lock_bh(dev);
1260                 spin_lock_irqsave(&priv->lock, flags);
1261                 neigh = tx->neigh;
1262
1263                 if (neigh) {
1264                         neigh->cm = NULL;
1265                         ipoib_neigh_free(neigh);
1266
1267                         tx->neigh = NULL;
1268                 }
1269
1270                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1271                         list_move(&tx->list, &priv->cm.reap_list);
1272                         queue_work(priv->wq, &priv->cm.reap_task);
1273                 }
1274
1275                 spin_unlock_irqrestore(&priv->lock, flags);
1276                 netif_tx_unlock_bh(dev);
1277                 break;
1278         default:
1279                 break;
1280         }
1281
1282         return 0;
1283 }
1284
1285 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1286                                        struct ipoib_neigh *neigh)
1287 {
1288         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1289         struct ipoib_cm_tx *tx;
1290
1291         tx = kzalloc(sizeof *tx, GFP_ATOMIC);
1292         if (!tx)
1293                 return NULL;
1294
1295         neigh->cm = tx;
1296         tx->neigh = neigh;
1297         tx->path = path;
1298         tx->dev = dev;
1299         list_add(&tx->list, &priv->cm.start_list);
1300         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1301         queue_work(priv->wq, &priv->cm.start_task);
1302         return tx;
1303 }
1304
1305 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1306 {
1307         struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1308         unsigned long flags;
1309         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1310                 spin_lock_irqsave(&priv->lock, flags);
1311                 list_move(&tx->list, &priv->cm.reap_list);
1312                 queue_work(priv->wq, &priv->cm.reap_task);
1313                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1314                           tx->neigh->daddr + 4);
1315                 tx->neigh = NULL;
1316                 spin_unlock_irqrestore(&priv->lock, flags);
1317         }
1318 }
1319
1320 #define QPN_AND_OPTIONS_OFFSET  4
1321
1322 static void ipoib_cm_tx_start(struct work_struct *work)
1323 {
1324         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1325                                                    cm.start_task);
1326         struct net_device *dev = priv->dev;
1327         struct ipoib_neigh *neigh;
1328         struct ipoib_cm_tx *p;
1329         unsigned long flags;
1330         struct ipoib_path *path;
1331         int ret;
1332
1333         struct sa_path_rec pathrec;
1334         u32 qpn;
1335
1336         netif_tx_lock_bh(dev);
1337         spin_lock_irqsave(&priv->lock, flags);
1338
1339         while (!list_empty(&priv->cm.start_list)) {
1340                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1341                 list_del_init(&p->list);
1342                 neigh = p->neigh;
1343
1344                 qpn = IPOIB_QPN(neigh->daddr);
1345                 /*
1346                  * As long as the search is with these 2 locks,
1347                  * path existence indicates its validity.
1348                  */
1349                 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1350                 if (!path) {
1351                         pr_info("%s ignore not valid path %pI6\n",
1352                                 __func__,
1353                                 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1354                         goto free_neigh;
1355                 }
1356                 memcpy(&pathrec, &p->path->pathrec, sizeof pathrec);
1357
1358                 spin_unlock_irqrestore(&priv->lock, flags);
1359                 netif_tx_unlock_bh(dev);
1360
1361                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1362
1363                 netif_tx_lock_bh(dev);
1364                 spin_lock_irqsave(&priv->lock, flags);
1365
1366                 if (ret) {
1367 free_neigh:
1368                         neigh = p->neigh;
1369                         if (neigh) {
1370                                 neigh->cm = NULL;
1371                                 ipoib_neigh_free(neigh);
1372                         }
1373                         list_del(&p->list);
1374                         kfree(p);
1375                 }
1376         }
1377
1378         spin_unlock_irqrestore(&priv->lock, flags);
1379         netif_tx_unlock_bh(dev);
1380 }
1381
1382 static void ipoib_cm_tx_reap(struct work_struct *work)
1383 {
1384         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1385                                                    cm.reap_task);
1386         struct net_device *dev = priv->dev;
1387         struct ipoib_cm_tx *p;
1388         unsigned long flags;
1389
1390         netif_tx_lock_bh(dev);
1391         spin_lock_irqsave(&priv->lock, flags);
1392
1393         while (!list_empty(&priv->cm.reap_list)) {
1394                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1395                 list_del_init(&p->list);
1396                 spin_unlock_irqrestore(&priv->lock, flags);
1397                 netif_tx_unlock_bh(dev);
1398                 ipoib_cm_tx_destroy(p);
1399                 netif_tx_lock_bh(dev);
1400                 spin_lock_irqsave(&priv->lock, flags);
1401         }
1402
1403         spin_unlock_irqrestore(&priv->lock, flags);
1404         netif_tx_unlock_bh(dev);
1405 }
1406
1407 static void ipoib_cm_skb_reap(struct work_struct *work)
1408 {
1409         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1410                                                    cm.skb_task);
1411         struct net_device *dev = priv->dev;
1412         struct sk_buff *skb;
1413         unsigned long flags;
1414         unsigned mtu = priv->mcast_mtu;
1415
1416         netif_tx_lock_bh(dev);
1417         spin_lock_irqsave(&priv->lock, flags);
1418
1419         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1420                 spin_unlock_irqrestore(&priv->lock, flags);
1421                 netif_tx_unlock_bh(dev);
1422
1423                 if (skb->protocol == htons(ETH_P_IP))
1424                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1425 #if IS_ENABLED(CONFIG_IPV6)
1426                 else if (skb->protocol == htons(ETH_P_IPV6))
1427                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1428 #endif
1429                 dev_kfree_skb_any(skb);
1430
1431                 netif_tx_lock_bh(dev);
1432                 spin_lock_irqsave(&priv->lock, flags);
1433         }
1434
1435         spin_unlock_irqrestore(&priv->lock, flags);
1436         netif_tx_unlock_bh(dev);
1437 }
1438
1439 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1440                            unsigned int mtu)
1441 {
1442         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1443         int e = skb_queue_empty(&priv->cm.skb_queue);
1444
1445         if (skb_dst(skb))
1446                 skb_dst(skb)->ops->update_pmtu(skb_dst(skb), NULL, skb, mtu);
1447
1448         skb_queue_tail(&priv->cm.skb_queue, skb);
1449         if (e)
1450                 queue_work(priv->wq, &priv->cm.skb_task);
1451 }
1452
1453 static void ipoib_cm_rx_reap(struct work_struct *work)
1454 {
1455         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1456                                                 cm.rx_reap_task)->dev);
1457 }
1458
1459 static void ipoib_cm_stale_task(struct work_struct *work)
1460 {
1461         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1462                                                    cm.stale_task.work);
1463         struct ipoib_cm_rx *p;
1464         int ret;
1465
1466         spin_lock_irq(&priv->lock);
1467         while (!list_empty(&priv->cm.passive_ids)) {
1468                 /* List is sorted by LRU, start from tail,
1469                  * stop when we see a recently used entry */
1470                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1471                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1472                         break;
1473                 list_move(&p->list, &priv->cm.rx_error_list);
1474                 p->state = IPOIB_CM_RX_ERROR;
1475                 spin_unlock_irq(&priv->lock);
1476                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1477                 if (ret)
1478                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1479                 spin_lock_irq(&priv->lock);
1480         }
1481
1482         if (!list_empty(&priv->cm.passive_ids))
1483                 queue_delayed_work(priv->wq,
1484                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1485         spin_unlock_irq(&priv->lock);
1486 }
1487
1488 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1489                          char *buf)
1490 {
1491         struct net_device *dev = to_net_dev(d);
1492         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1493
1494         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1495                 return sprintf(buf, "connected\n");
1496         else
1497                 return sprintf(buf, "datagram\n");
1498 }
1499
1500 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1501                         const char *buf, size_t count)
1502 {
1503         struct net_device *dev = to_net_dev(d);
1504         int ret;
1505         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1506
1507         if (test_bit(IPOIB_FLAG_GOING_DOWN, &priv->flags))
1508                 return -EPERM;
1509
1510         if (!rtnl_trylock())
1511                 return restart_syscall();
1512
1513         ret = ipoib_set_mode(dev, buf);
1514
1515         /* The assumption is that the function ipoib_set_mode returned
1516          * with the rtnl held by it, if not the value -EBUSY returned,
1517          * then no need to rtnl_unlock
1518          */
1519         if (ret != -EBUSY)
1520                 rtnl_unlock();
1521
1522         return (!ret || ret == -EBUSY) ? count : ret;
1523 }
1524
1525 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1526
1527 int ipoib_cm_add_mode_attr(struct net_device *dev)
1528 {
1529         return device_create_file(&dev->dev, &dev_attr_mode);
1530 }
1531
1532 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1533 {
1534         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1535         struct ib_srq_init_attr srq_init_attr = {
1536                 .srq_type = IB_SRQT_BASIC,
1537                 .attr = {
1538                         .max_wr  = ipoib_recvq_size,
1539                         .max_sge = max_sge
1540                 }
1541         };
1542
1543         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1544         if (IS_ERR(priv->cm.srq)) {
1545                 if (PTR_ERR(priv->cm.srq) != -ENOSYS)
1546                         printk(KERN_WARNING "%s: failed to allocate SRQ, error %ld\n",
1547                                priv->ca->name, PTR_ERR(priv->cm.srq));
1548                 priv->cm.srq = NULL;
1549                 return;
1550         }
1551
1552         priv->cm.srq_ring = vzalloc(ipoib_recvq_size * sizeof *priv->cm.srq_ring);
1553         if (!priv->cm.srq_ring) {
1554                 ib_destroy_srq(priv->cm.srq);
1555                 priv->cm.srq = NULL;
1556                 return;
1557         }
1558
1559 }
1560
1561 int ipoib_cm_dev_init(struct net_device *dev)
1562 {
1563         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1564         int max_srq_sge, i;
1565
1566         INIT_LIST_HEAD(&priv->cm.passive_ids);
1567         INIT_LIST_HEAD(&priv->cm.reap_list);
1568         INIT_LIST_HEAD(&priv->cm.start_list);
1569         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1570         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1571         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1572         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1573         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1574         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1575         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1576         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1577         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1578
1579         skb_queue_head_init(&priv->cm.skb_queue);
1580
1581         ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1582
1583         max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1584         ipoib_cm_create_srq(dev, max_srq_sge);
1585         if (ipoib_cm_has_srq(dev)) {
1586                 priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1587                 priv->cm.num_frags  = max_srq_sge;
1588                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1589                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1590         } else {
1591                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1592                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1593         }
1594
1595         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1596
1597         if (ipoib_cm_has_srq(dev)) {
1598                 for (i = 0; i < ipoib_recvq_size; ++i) {
1599                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1600                                                    priv->cm.num_frags - 1,
1601                                                    priv->cm.srq_ring[i].mapping,
1602                                                    GFP_KERNEL)) {
1603                                 ipoib_warn(priv, "failed to allocate "
1604                                            "receive buffer %d\n", i);
1605                                 ipoib_cm_dev_cleanup(dev);
1606                                 return -ENOMEM;
1607                         }
1608
1609                         if (ipoib_cm_post_receive_srq(dev, i)) {
1610                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1611                                            "failed for buf %d\n", i);
1612                                 ipoib_cm_dev_cleanup(dev);
1613                                 return -EIO;
1614                         }
1615                 }
1616         }
1617
1618         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1619         return 0;
1620 }
1621
1622 void ipoib_cm_dev_cleanup(struct net_device *dev)
1623 {
1624         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1625         int ret;
1626
1627         if (!priv->cm.srq)
1628                 return;
1629
1630         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1631
1632         ret = ib_destroy_srq(priv->cm.srq);
1633         if (ret)
1634                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1635
1636         priv->cm.srq = NULL;
1637         if (!priv->cm.srq_ring)
1638                 return;
1639
1640         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1641         priv->cm.srq_ring = NULL;
1642 }