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                                const 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         int i, ret;
  98
  99         priv->cm.rx_wr.wr_id = id | IPOIB_OP_CM | IPOIB_OP_RECV;
 100
 101         for (i = 0; i < priv->cm.num_frags; ++i)
 102                 priv->cm.rx_sge[i].addr = priv->cm.srq_ring[id].mapping[i];
 103
 104         ret = ib_post_srq_recv(priv->cm.srq, &priv->cm.rx_wr, NULL);
 105         if (unlikely(ret)) {
 106                 ipoib_warn(priv, "post srq failed for buf %d (%d)\n", id, ret);
 107                 ipoib_cm_dma_unmap_rx(priv, priv->cm.num_frags - 1,
 108                                       priv->cm.srq_ring[id].mapping);
 109                 dev_kfree_skb_any(priv->cm.srq_ring[id].skb);
 110                 priv->cm.srq_ring[id].skb = NULL;
 111         }
 112
 113         return ret;
 114 }
 115
 116 static int ipoib_cm_post_receive_nonsrq(struct net_device *dev,
 117                                         struct ipoib_cm_rx *rx,
 118                                         struct ib_recv_wr *wr,
 119                                         struct ib_sge *sge, int id)
 120 {
 121         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 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, NULL);
 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 = ipoib_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 = ipoib_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 ipoib_cm_rx *p;
 214
 215         /* We only reserved 1 extra slot in CQ for drain WRs, so
 216          * make sure we have at most 1 outstanding WR. */
 217         if (list_empty(&priv->cm.rx_flush_list) ||
 218             !list_empty(&priv->cm.rx_drain_list))
 219                 return;
 220
 221         /*
 222          * QPs on flush list are error state.  This way, a "flush
 223          * error" WC will be immediately generated for each WR we post.
 224          */
 225         p = list_entry(priv->cm.rx_flush_list.next, typeof(*p), list);
 226         ipoib_cm_rx_drain_wr.wr_id = IPOIB_CM_RX_DRAIN_WRID;
 227         if (ib_post_send(p->qp, &ipoib_cm_rx_drain_wr, NULL))
 228                 ipoib_warn(priv, "failed to post drain wr\n");
 229
 230         list_splice_init(&priv->cm.rx_flush_list, &priv->cm.rx_drain_list);
 231 }
 232
 233 static void ipoib_cm_rx_event_handler(struct ib_event *event, void *ctx)
 234 {
 235         struct ipoib_cm_rx *p = ctx;
 236         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
 237         unsigned long flags;
 238
 239         if (event->event != IB_EVENT_QP_LAST_WQE_REACHED)
 240                 return;
 241
 242         spin_lock_irqsave(&priv->lock, flags);
 243         list_move(&p->list, &priv->cm.rx_flush_list);
 244         p->state = IPOIB_CM_RX_FLUSH;
 245         ipoib_cm_start_rx_drain(priv);
 246         spin_unlock_irqrestore(&priv->lock, flags);
 247 }
 248
 249 static struct ib_qp *ipoib_cm_create_rx_qp(struct net_device *dev,
 250                                            struct ipoib_cm_rx *p)
 251 {
 252         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 253         struct ib_qp_init_attr attr = {
 254                 .event_handler = ipoib_cm_rx_event_handler,
 255                 .send_cq = priv->recv_cq, /* For drain WR */
 256                 .recv_cq = priv->recv_cq,
 257                 .srq = priv->cm.srq,
 258                 .cap.max_send_wr = 1, /* For drain WR */
 259                 .cap.max_send_sge = 1, /* FIXME: 0 Seems not to work */
 260                 .sq_sig_type = IB_SIGNAL_ALL_WR,
 261                 .qp_type = IB_QPT_RC,
 262                 .qp_context = p,
 263         };
 264
 265         if (!ipoib_cm_has_srq(dev)) {
 266                 attr.cap.max_recv_wr  = ipoib_recvq_size;
 267                 attr.cap.max_recv_sge = IPOIB_CM_RX_SG;
 268         }
 269
 270         return ib_create_qp(priv->pd, &attr);
 271 }
 272
 273 static int ipoib_cm_modify_rx_qp(struct net_device *dev,
 274                                  struct ib_cm_id *cm_id, struct ib_qp *qp,
 275                                  unsigned int psn)
 276 {
 277         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 278         struct ib_qp_attr qp_attr;
 279         int qp_attr_mask, ret;
 280
 281         qp_attr.qp_state = IB_QPS_INIT;
 282         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 283         if (ret) {
 284                 ipoib_warn(priv, "failed to init QP attr for INIT: %d\n", ret);
 285                 return ret;
 286         }
 287         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 288         if (ret) {
 289                 ipoib_warn(priv, "failed to modify QP to INIT: %d\n", ret);
 290                 return ret;
 291         }
 292         qp_attr.qp_state = IB_QPS_RTR;
 293         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 294         if (ret) {
 295                 ipoib_warn(priv, "failed to init QP attr for RTR: %d\n", ret);
 296                 return ret;
 297         }
 298         qp_attr.rq_psn = psn;
 299         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 300         if (ret) {
 301                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
 302                 return ret;
 303         }
 304
 305         /*
 306          * Current Mellanox HCA firmware won't generate completions
 307          * with error for drain WRs unless the QP has been moved to
 308          * RTS first. This work-around leaves a window where a QP has
 309          * moved to error asynchronously, but this will eventually get
 310          * fixed in firmware, so let's not error out if modify QP
 311          * fails.
 312          */
 313         qp_attr.qp_state = IB_QPS_RTS;
 314         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
 315         if (ret) {
 316                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
 317                 return 0;
 318         }
 319         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
 320         if (ret) {
 321                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
 322                 return 0;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static void ipoib_cm_init_rx_wr(struct net_device *dev,
 329                                 struct ib_recv_wr *wr,
 330                                 struct ib_sge *sge)
 331 {
 332         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 333         int i;
 334
 335         for (i = 0; i < priv->cm.num_frags; ++i)
 336                 sge[i].lkey = priv->pd->local_dma_lkey;
 337
 338         sge[0].length = IPOIB_CM_HEAD_SIZE;
 339         for (i = 1; i < priv->cm.num_frags; ++i)
 340                 sge[i].length = PAGE_SIZE;
 341
 342         wr->next    = NULL;
 343         wr->sg_list = sge;
 344         wr->num_sge = priv->cm.num_frags;
 345 }
 346
 347 static int ipoib_cm_nonsrq_init_rx(struct net_device *dev, struct ib_cm_id *cm_id,
 348                                    struct ipoib_cm_rx *rx)
 349 {
 350         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 351         struct {
 352                 struct ib_recv_wr wr;
 353                 struct ib_sge sge[IPOIB_CM_RX_SG];
 354         } *t;
 355         int ret;
 356         int i;
 357
 358         rx->rx_ring = vzalloc(array_size(ipoib_recvq_size,
 359                                          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_1;
 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
 414 err_free_1:
 415         ipoib_cm_free_rx_ring(dev, rx->rx_ring);
 416
 417         return ret;
 418 }
 419
 420 static int ipoib_cm_send_rep(struct net_device *dev, struct ib_cm_id *cm_id,
 421                              struct ib_qp *qp,
 422                              const struct ib_cm_req_event_param *req,
 423                              unsigned int psn)
 424 {
 425         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 426         struct ipoib_cm_data data = {};
 427         struct ib_cm_rep_param rep = {};
 428
 429         data.qpn = cpu_to_be32(priv->qp->qp_num);
 430         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
 431
 432         rep.private_data = &data;
 433         rep.private_data_len = sizeof(data);
 434         rep.flow_control = 0;
 435         rep.rnr_retry_count = req->rnr_retry_count;
 436         rep.srq = ipoib_cm_has_srq(dev);
 437         rep.qp_num = qp->qp_num;
 438         rep.starting_psn = psn;
 439         return ib_send_cm_rep(cm_id, &rep);
 440 }
 441
 442 static int ipoib_cm_req_handler(struct ib_cm_id *cm_id,
 443                                 const 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 int 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                                const 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                 ib_send_cm_drep(cm_id, NULL, 0);
 515                 /* Fall through */
 516         case IB_CM_REJ_RECEIVED:
 517                 p = cm_id->context;
 518                 priv = ipoib_priv(p->dev);
 519                 if (ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE))
 520                         ipoib_warn(priv, "unable to move qp to error state\n");
 521                 /* Fall through */
 522         default:
 523                 return 0;
 524         }
 525 }
 526 /* Adjust length of skb with fragments to match received data */
 527 static void skb_put_frags(struct sk_buff *skb, unsigned int hdr_space,
 528                           unsigned int length, struct sk_buff *toskb)
 529 {
 530         int i, num_frags;
 531         unsigned int size;
 532
 533         /* put header into skb */
 534         size = min(length, hdr_space);
 535         skb->tail += size;
 536         skb->len += size;
 537         length -= size;
 538
 539         num_frags = skb_shinfo(skb)->nr_frags;
 540         for (i = 0; i < num_frags; i++) {
 541                 skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 542
 543                 if (length == 0) {
 544                         /* don't need this page */
 545                         skb_fill_page_desc(toskb, i, skb_frag_page(frag),
 546                                            0, PAGE_SIZE);
 547                         --skb_shinfo(skb)->nr_frags;
 548                 } else {
 549                         size = min_t(unsigned int, length, PAGE_SIZE);
 550
 551                         skb_frag_size_set(frag, size);
 552                         skb->data_len += size;
 553                         skb->truesize += size;
 554                         skb->len += size;
 555                         length -= size;
 556                 }
 557         }
 558 }
 559
 560 void ipoib_cm_handle_rx_wc(struct net_device *dev, struct ib_wc *wc)
 561 {
 562         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 563         struct ipoib_cm_rx_buf *rx_ring;
 564         unsigned int wr_id = wc->wr_id & ~(IPOIB_OP_CM | IPOIB_OP_RECV);
 565         struct sk_buff *skb, *newskb;
 566         struct ipoib_cm_rx *p;
 567         unsigned long flags;
 568         u64 mapping[IPOIB_CM_RX_SG];
 569         int frags;
 570         int has_srq;
 571         struct sk_buff *small_skb;
 572
 573         ipoib_dbg_data(priv, "cm recv completion: id %d, status: %d\n",
 574                        wr_id, wc->status);
 575
 576         if (unlikely(wr_id >= ipoib_recvq_size)) {
 577                 if (wr_id == (IPOIB_CM_RX_DRAIN_WRID & ~(IPOIB_OP_CM | IPOIB_OP_RECV))) {
 578                         spin_lock_irqsave(&priv->lock, flags);
 579                         list_splice_init(&priv->cm.rx_drain_list, &priv->cm.rx_reap_list);
 580                         ipoib_cm_start_rx_drain(priv);
 581                         queue_work(priv->wq, &priv->cm.rx_reap_task);
 582                         spin_unlock_irqrestore(&priv->lock, flags);
 583                 } else
 584                         ipoib_warn(priv, "cm recv completion event with wrid %d (> %d)\n",
 585                                    wr_id, ipoib_recvq_size);
 586                 return;
 587         }
 588
 589         p = wc->qp->qp_context;
 590
 591         has_srq = ipoib_cm_has_srq(dev);
 592         rx_ring = has_srq ? priv->cm.srq_ring : p->rx_ring;
 593
 594         skb = rx_ring[wr_id].skb;
 595
 596         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 597                 ipoib_dbg(priv,
 598                           "cm recv error (status=%d, wrid=%d vend_err %#x)\n",
 599                           wc->status, wr_id, wc->vendor_err);
 600                 ++dev->stats.rx_dropped;
 601                 if (has_srq)
 602                         goto repost;
 603                 else {
 604                         if (!--p->recv_count) {
 605                                 spin_lock_irqsave(&priv->lock, flags);
 606                                 list_move(&p->list, &priv->cm.rx_reap_list);
 607                                 spin_unlock_irqrestore(&priv->lock, flags);
 608                                 queue_work(priv->wq, &priv->cm.rx_reap_task);
 609                         }
 610                         return;
 611                 }
 612         }
 613
 614         if (unlikely(!(wr_id & IPOIB_CM_RX_UPDATE_MASK))) {
 615                 if (p && time_after_eq(jiffies, p->jiffies + IPOIB_CM_RX_UPDATE_TIME)) {
 616                         spin_lock_irqsave(&priv->lock, flags);
 617                         p->jiffies = jiffies;
 618                         /* Move this entry to list head, but do not re-add it
 619                          * if it has been moved out of list. */
 620                         if (p->state == IPOIB_CM_RX_LIVE)
 621                                 list_move(&p->list, &priv->cm.passive_ids);
 622                         spin_unlock_irqrestore(&priv->lock, flags);
 623                 }
 624         }
 625
 626         if (wc->byte_len < IPOIB_CM_COPYBREAK) {
 627                 int dlen = wc->byte_len;
 628
 629                 small_skb = dev_alloc_skb(dlen + IPOIB_CM_RX_RESERVE);
 630                 if (small_skb) {
 631                         skb_reserve(small_skb, IPOIB_CM_RX_RESERVE);
 632                         ib_dma_sync_single_for_cpu(priv->ca, rx_ring[wr_id].mapping[0],
 633                                                    dlen, DMA_FROM_DEVICE);
 634                         skb_copy_from_linear_data(skb, small_skb->data, dlen);
 635                         ib_dma_sync_single_for_device(priv->ca, rx_ring[wr_id].mapping[0],
 636                                                       dlen, DMA_FROM_DEVICE);
 637                         skb_put(small_skb, dlen);
 638                         skb = small_skb;
 639                         goto copied;
 640                 }
 641         }
 642
 643         frags = PAGE_ALIGN(wc->byte_len -
 644                            min_t(u32, wc->byte_len, IPOIB_CM_HEAD_SIZE)) /
 645                 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         ipoib_build_sge(priv, tx_req);
 702
 703         priv->tx_wr.wr.wr_id    = wr_id | IPOIB_OP_CM;
 704
 705         return ib_post_send(tx->qp, &priv->tx_wr.wr, NULL);
 706 }
 707
 708 void ipoib_cm_send(struct net_device *dev, struct sk_buff *skb, struct ipoib_cm_tx *tx)
 709 {
 710         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 711         struct ipoib_tx_buf *tx_req;
 712         int rc;
 713         unsigned int usable_sge = tx->max_send_sge - !!skb_headlen(skb);
 714
 715         if (unlikely(skb->len > tx->mtu)) {
 716                 ipoib_warn(priv, "packet len %d (> %d) too long to send, dropping\n",
 717                            skb->len, tx->mtu);
 718                 ++dev->stats.tx_dropped;
 719                 ++dev->stats.tx_errors;
 720                 ipoib_cm_skb_too_long(dev, skb, tx->mtu - IPOIB_ENCAP_LEN);
 721                 return;
 722         }
 723         if (skb_shinfo(skb)->nr_frags > usable_sge) {
 724                 if (skb_linearize(skb) < 0) {
 725                         ipoib_warn(priv, "skb could not be linearized\n");
 726                         ++dev->stats.tx_dropped;
 727                         ++dev->stats.tx_errors;
 728                         dev_kfree_skb_any(skb);
 729                         return;
 730                 }
 731                 /* Does skb_linearize return ok without reducing nr_frags? */
 732                 if (skb_shinfo(skb)->nr_frags > usable_sge) {
 733                         ipoib_warn(priv, "too many frags after skb linearize\n");
 734                         ++dev->stats.tx_dropped;
 735                         ++dev->stats.tx_errors;
 736                         dev_kfree_skb_any(skb);
 737                         return;
 738                 }
 739         }
 740         ipoib_dbg_data(priv, "sending packet: head 0x%x length %d connection 0x%x\n",
 741                        tx->tx_head, skb->len, tx->qp->qp_num);
 742
 743         /*
 744          * We put the skb into the tx_ring _before_ we call post_send()
 745          * because it's entirely possible that the completion handler will
 746          * run before we execute anything after the post_send().  That
 747          * means we have to make sure everything is properly recorded and
 748          * our state is consistent before we call post_send().
 749          */
 750         tx_req = &tx->tx_ring[tx->tx_head & (ipoib_sendq_size - 1)];
 751         tx_req->skb = skb;
 752
 753         if (unlikely(ipoib_dma_map_tx(priv->ca, tx_req))) {
 754                 ++dev->stats.tx_errors;
 755                 dev_kfree_skb_any(skb);
 756                 return;
 757         }
 758
 759         if ((priv->tx_head - priv->tx_tail) == ipoib_sendq_size - 1) {
 760                 ipoib_dbg(priv, "TX ring 0x%x full, stopping kernel net queue\n",
 761                           tx->qp->qp_num);
 762                 netif_stop_queue(dev);
 763         }
 764
 765         skb_orphan(skb);
 766         skb_dst_drop(skb);
 767
 768         if (netif_queue_stopped(dev)) {
 769                 rc = ib_req_notify_cq(priv->send_cq, IB_CQ_NEXT_COMP |
 770                                       IB_CQ_REPORT_MISSED_EVENTS);
 771                 if (unlikely(rc < 0))
 772                         ipoib_warn(priv, "IPoIB/CM:request notify on send CQ failed\n");
 773                 else if (rc)
 774                         napi_schedule(&priv->send_napi);
 775         }
 776
 777         rc = post_send(priv, tx, tx->tx_head & (ipoib_sendq_size - 1), tx_req);
 778         if (unlikely(rc)) {
 779                 ipoib_warn(priv, "IPoIB/CM:post_send failed, error %d\n", rc);
 780                 ++dev->stats.tx_errors;
 781                 ipoib_dma_unmap_tx(priv, tx_req);
 782                 dev_kfree_skb_any(skb);
 783
 784                 if (netif_queue_stopped(dev))
 785                         netif_wake_queue(dev);
 786         } else {
 787                 netif_trans_update(dev);
 788                 ++tx->tx_head;
 789                 ++priv->tx_head;
 790         }
 791 }
 792
 793 void ipoib_cm_handle_tx_wc(struct net_device *dev, struct ib_wc *wc)
 794 {
 795         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 796         struct ipoib_cm_tx *tx = wc->qp->qp_context;
 797         unsigned int wr_id = wc->wr_id & ~IPOIB_OP_CM;
 798         struct ipoib_tx_buf *tx_req;
 799         unsigned long flags;
 800
 801         ipoib_dbg_data(priv, "cm send completion: id %d, status: %d\n",
 802                        wr_id, wc->status);
 803
 804         if (unlikely(wr_id >= ipoib_sendq_size)) {
 805                 ipoib_warn(priv, "cm send completion event with wrid %d (> %d)\n",
 806                            wr_id, ipoib_sendq_size);
 807                 return;
 808         }
 809
 810         tx_req = &tx->tx_ring[wr_id];
 811
 812         ipoib_dma_unmap_tx(priv, tx_req);
 813
 814         /* FIXME: is this right? Shouldn't we only increment on success? */
 815         ++dev->stats.tx_packets;
 816         dev->stats.tx_bytes += tx_req->skb->len;
 817
 818         dev_kfree_skb_any(tx_req->skb);
 819
 820         netif_tx_lock(dev);
 821
 822         ++tx->tx_tail;
 823         ++priv->tx_tail;
 824
 825         if (unlikely(netif_queue_stopped(dev) &&
 826                      (priv->tx_head - priv->tx_tail) <= ipoib_sendq_size >> 1 &&
 827                      test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags)))
 828                 netif_wake_queue(dev);
 829
 830         if (wc->status != IB_WC_SUCCESS &&
 831             wc->status != IB_WC_WR_FLUSH_ERR) {
 832                 struct ipoib_neigh *neigh;
 833
 834                 /* IB_WC[_RNR]_RETRY_EXC_ERR error is part of the life cycle,
 835                  * so don't make waves.
 836                  */
 837                 if (wc->status == IB_WC_RNR_RETRY_EXC_ERR ||
 838                     wc->status == IB_WC_RETRY_EXC_ERR)
 839                         ipoib_dbg(priv,
 840                                   "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
 841                                    __func__, wc->status, wr_id, wc->vendor_err);
 842                 else
 843                         ipoib_warn(priv,
 844                                     "%s: failed cm send event (status=%d, wrid=%d vend_err %#x)\n",
 845                                    __func__, wc->status, wr_id, wc->vendor_err);
 846
 847                 spin_lock_irqsave(&priv->lock, flags);
 848                 neigh = tx->neigh;
 849
 850                 if (neigh) {
 851                         neigh->cm = NULL;
 852                         ipoib_neigh_free(neigh);
 853
 854                         tx->neigh = NULL;
 855                 }
 856
 857                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
 858                         list_move(&tx->list, &priv->cm.reap_list);
 859                         queue_work(priv->wq, &priv->cm.reap_task);
 860                 }
 861
 862                 clear_bit(IPOIB_FLAG_OPER_UP, &tx->flags);
 863
 864                 spin_unlock_irqrestore(&priv->lock, flags);
 865         }
 866
 867         netif_tx_unlock(dev);
 868 }
 869
 870 int ipoib_cm_dev_open(struct net_device *dev)
 871 {
 872         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 873         int ret;
 874
 875         if (!IPOIB_CM_SUPPORTED(dev->dev_addr))
 876                 return 0;
 877
 878         priv->cm.id = ib_create_cm_id(priv->ca, ipoib_cm_rx_handler, dev);
 879         if (IS_ERR(priv->cm.id)) {
 880                 pr_warn("%s: failed to create CM ID\n", priv->ca->name);
 881                 ret = PTR_ERR(priv->cm.id);
 882                 goto err_cm;
 883         }
 884
 885         ret = ib_cm_listen(priv->cm.id, cpu_to_be64(IPOIB_CM_IETF_ID | priv->qp->qp_num),
 886                            0);
 887         if (ret) {
 888                 pr_warn("%s: failed to listen on ID 0x%llx\n", priv->ca->name,
 889                         IPOIB_CM_IETF_ID | priv->qp->qp_num);
 890                 goto err_listen;
 891         }
 892
 893         return 0;
 894
 895 err_listen:
 896         ib_destroy_cm_id(priv->cm.id);
 897 err_cm:
 898         priv->cm.id = NULL;
 899         return ret;
 900 }
 901
 902 static void ipoib_cm_free_rx_reap_list(struct net_device *dev)
 903 {
 904         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 905         struct ipoib_cm_rx *rx, *n;
 906         LIST_HEAD(list);
 907
 908         spin_lock_irq(&priv->lock);
 909         list_splice_init(&priv->cm.rx_reap_list, &list);
 910         spin_unlock_irq(&priv->lock);
 911
 912         list_for_each_entry_safe(rx, n, &list, list) {
 913                 ib_destroy_cm_id(rx->id);
 914                 ib_destroy_qp(rx->qp);
 915                 if (!ipoib_cm_has_srq(dev)) {
 916                         ipoib_cm_free_rx_ring(priv->dev, rx->rx_ring);
 917                         spin_lock_irq(&priv->lock);
 918                         --priv->cm.nonsrq_conn_qp;
 919                         spin_unlock_irq(&priv->lock);
 920                 }
 921                 kfree(rx);
 922         }
 923 }
 924
 925 void ipoib_cm_dev_stop(struct net_device *dev)
 926 {
 927         struct ipoib_dev_priv *priv = ipoib_priv(dev);
 928         struct ipoib_cm_rx *p;
 929         unsigned long begin;
 930         int ret;
 931
 932         if (!IPOIB_CM_SUPPORTED(dev->dev_addr) || !priv->cm.id)
 933                 return;
 934
 935         ib_destroy_cm_id(priv->cm.id);
 936         priv->cm.id = NULL;
 937
 938         spin_lock_irq(&priv->lock);
 939         while (!list_empty(&priv->cm.passive_ids)) {
 940                 p = list_entry(priv->cm.passive_ids.next, typeof(*p), list);
 941                 list_move(&p->list, &priv->cm.rx_error_list);
 942                 p->state = IPOIB_CM_RX_ERROR;
 943                 spin_unlock_irq(&priv->lock);
 944                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
 945                 if (ret)
 946                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
 947                 spin_lock_irq(&priv->lock);
 948         }
 949
 950         /* Wait for all RX to be drained */
 951         begin = jiffies;
 952
 953         while (!list_empty(&priv->cm.rx_error_list) ||
 954                !list_empty(&priv->cm.rx_flush_list) ||
 955                !list_empty(&priv->cm.rx_drain_list)) {
 956                 if (time_after(jiffies, begin + 5 * HZ)) {
 957                         ipoib_warn(priv, "RX drain timing out\n");
 958
 959                         /*
 960                          * assume the HW is wedged and just free up everything.
 961                          */
 962                         list_splice_init(&priv->cm.rx_flush_list,
 963                                          &priv->cm.rx_reap_list);
 964                         list_splice_init(&priv->cm.rx_error_list,
 965                                          &priv->cm.rx_reap_list);
 966                         list_splice_init(&priv->cm.rx_drain_list,
 967                                          &priv->cm.rx_reap_list);
 968                         break;
 969                 }
 970                 spin_unlock_irq(&priv->lock);
 971                 usleep_range(1000, 2000);
 972                 ipoib_drain_cq(dev);
 973                 spin_lock_irq(&priv->lock);
 974         }
 975
 976         spin_unlock_irq(&priv->lock);
 977
 978         ipoib_cm_free_rx_reap_list(dev);
 979
 980         cancel_delayed_work(&priv->cm.stale_task);
 981 }
 982
 983 static int ipoib_cm_rep_handler(struct ib_cm_id *cm_id,
 984                                 const struct ib_cm_event *event)
 985 {
 986         struct ipoib_cm_tx *p = cm_id->context;
 987         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
 988         struct ipoib_cm_data *data = event->private_data;
 989         struct sk_buff_head skqueue;
 990         struct ib_qp_attr qp_attr;
 991         int qp_attr_mask, ret;
 992         struct sk_buff *skb;
 993
 994         p->mtu = be32_to_cpu(data->mtu);
 995
 996         if (p->mtu <= IPOIB_ENCAP_LEN) {
 997                 ipoib_warn(priv, "Rejecting connection: mtu %d <= %d\n",
 998                            p->mtu, IPOIB_ENCAP_LEN);
 999                 return -EINVAL;
1000         }
1001
1002         qp_attr.qp_state = IB_QPS_RTR;
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 RTR: %d\n", ret);
1006                 return ret;
1007         }
1008
1009         qp_attr.rq_psn = 0 /* FIXME */;
1010         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1011         if (ret) {
1012                 ipoib_warn(priv, "failed to modify QP to RTR: %d\n", ret);
1013                 return ret;
1014         }
1015
1016         qp_attr.qp_state = IB_QPS_RTS;
1017         ret = ib_cm_init_qp_attr(cm_id, &qp_attr, &qp_attr_mask);
1018         if (ret) {
1019                 ipoib_warn(priv, "failed to init QP attr for RTS: %d\n", ret);
1020                 return ret;
1021         }
1022         ret = ib_modify_qp(p->qp, &qp_attr, qp_attr_mask);
1023         if (ret) {
1024                 ipoib_warn(priv, "failed to modify QP to RTS: %d\n", ret);
1025                 return ret;
1026         }
1027
1028         skb_queue_head_init(&skqueue);
1029
1030         netif_tx_lock_bh(p->dev);
1031         spin_lock_irq(&priv->lock);
1032         set_bit(IPOIB_FLAG_OPER_UP, &p->flags);
1033         if (p->neigh)
1034                 while ((skb = __skb_dequeue(&p->neigh->queue)))
1035                         __skb_queue_tail(&skqueue, skb);
1036         spin_unlock_irq(&priv->lock);
1037         netif_tx_unlock_bh(p->dev);
1038
1039         while ((skb = __skb_dequeue(&skqueue))) {
1040                 skb->dev = p->dev;
1041                 ret = dev_queue_xmit(skb);
1042                 if (ret)
1043                         ipoib_warn(priv, "%s:dev_queue_xmit failed to re-queue packet, ret:%d\n",
1044                                    __func__, ret);
1045         }
1046
1047         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1048         if (ret) {
1049                 ipoib_warn(priv, "failed to send RTU: %d\n", ret);
1050                 return ret;
1051         }
1052         return 0;
1053 }
1054
1055 static struct ib_qp *ipoib_cm_create_tx_qp(struct net_device *dev, struct ipoib_cm_tx *tx)
1056 {
1057         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1058         struct ib_qp_init_attr attr = {
1059                 .send_cq                = priv->send_cq,
1060                 .recv_cq                = priv->recv_cq,
1061                 .srq                    = priv->cm.srq,
1062                 .cap.max_send_wr        = ipoib_sendq_size,
1063                 .cap.max_send_sge       = 1,
1064                 .sq_sig_type            = IB_SIGNAL_ALL_WR,
1065                 .qp_type                = IB_QPT_RC,
1066                 .qp_context             = tx,
1067                 .create_flags           = 0
1068         };
1069         struct ib_qp *tx_qp;
1070
1071         if (dev->features & NETIF_F_SG)
1072                 attr.cap.max_send_sge = min_t(u32, priv->ca->attrs.max_send_sge,
1073                                               MAX_SKB_FRAGS + 1);
1074
1075         tx_qp = ib_create_qp(priv->pd, &attr);
1076         tx->max_send_sge = attr.cap.max_send_sge;
1077         return tx_qp;
1078 }
1079
1080 static int ipoib_cm_send_req(struct net_device *dev,
1081                              struct ib_cm_id *id, struct ib_qp *qp,
1082                              u32 qpn,
1083                              struct sa_path_rec *pathrec)
1084 {
1085         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1086         struct ipoib_cm_data data = {};
1087         struct ib_cm_req_param req = {};
1088
1089         data.qpn = cpu_to_be32(priv->qp->qp_num);
1090         data.mtu = cpu_to_be32(IPOIB_CM_BUF_SIZE);
1091
1092         req.primary_path                = pathrec;
1093         req.alternate_path              = NULL;
1094         req.service_id                  = cpu_to_be64(IPOIB_CM_IETF_ID | qpn);
1095         req.qp_num                      = qp->qp_num;
1096         req.qp_type                     = qp->qp_type;
1097         req.private_data                = &data;
1098         req.private_data_len            = sizeof(data);
1099         req.flow_control                = 0;
1100
1101         req.starting_psn                = 0; /* FIXME */
1102
1103         /*
1104          * Pick some arbitrary defaults here; we could make these
1105          * module parameters if anyone cared about setting them.
1106          */
1107         req.responder_resources         = 4;
1108         req.remote_cm_response_timeout  = 20;
1109         req.local_cm_response_timeout   = 20;
1110         req.retry_count                 = 0; /* RFC draft warns against retries */
1111         req.rnr_retry_count             = 0; /* RFC draft warns against retries */
1112         req.max_cm_retries              = 15;
1113         req.srq                         = ipoib_cm_has_srq(dev);
1114         return ib_send_cm_req(id, &req);
1115 }
1116
1117 static int ipoib_cm_modify_tx_init(struct net_device *dev,
1118                                   struct ib_cm_id *cm_id, struct ib_qp *qp)
1119 {
1120         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1121         struct ib_qp_attr qp_attr;
1122         int qp_attr_mask, ret;
1123         ret = ib_find_pkey(priv->ca, priv->port, priv->pkey, &qp_attr.pkey_index);
1124         if (ret) {
1125                 ipoib_warn(priv, "pkey 0x%x not found: %d\n", priv->pkey, ret);
1126                 return ret;
1127         }
1128
1129         qp_attr.qp_state = IB_QPS_INIT;
1130         qp_attr.qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1131         qp_attr.port_num = priv->port;
1132         qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PKEY_INDEX | IB_QP_PORT;
1133
1134         ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1135         if (ret) {
1136                 ipoib_warn(priv, "failed to modify tx QP to INIT: %d\n", ret);
1137                 return ret;
1138         }
1139         return 0;
1140 }
1141
1142 static int ipoib_cm_tx_init(struct ipoib_cm_tx *p, u32 qpn,
1143                             struct sa_path_rec *pathrec)
1144 {
1145         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1146         unsigned int noio_flag;
1147         int ret;
1148
1149         noio_flag = memalloc_noio_save();
1150         p->tx_ring = vzalloc(array_size(ipoib_sendq_size, sizeof(*p->tx_ring)));
1151         if (!p->tx_ring) {
1152                 memalloc_noio_restore(noio_flag);
1153                 ret = -ENOMEM;
1154                 goto err_tx;
1155         }
1156
1157         p->qp = ipoib_cm_create_tx_qp(p->dev, p);
1158         memalloc_noio_restore(noio_flag);
1159         if (IS_ERR(p->qp)) {
1160                 ret = PTR_ERR(p->qp);
1161                 ipoib_warn(priv, "failed to create tx qp: %d\n", ret);
1162                 goto err_qp;
1163         }
1164
1165         p->id = ib_create_cm_id(priv->ca, ipoib_cm_tx_handler, p);
1166         if (IS_ERR(p->id)) {
1167                 ret = PTR_ERR(p->id);
1168                 ipoib_warn(priv, "failed to create tx cm id: %d\n", ret);
1169                 goto err_id;
1170         }
1171
1172         ret = ipoib_cm_modify_tx_init(p->dev, p->id,  p->qp);
1173         if (ret) {
1174                 ipoib_warn(priv, "failed to modify tx qp to rtr: %d\n", ret);
1175                 goto err_modify_send;
1176         }
1177
1178         ret = ipoib_cm_send_req(p->dev, p->id, p->qp, qpn, pathrec);
1179         if (ret) {
1180                 ipoib_warn(priv, "failed to send cm req: %d\n", ret);
1181                 goto err_modify_send;
1182         }
1183
1184         ipoib_dbg(priv, "Request connection 0x%x for gid %pI6 qpn 0x%x\n",
1185                   p->qp->qp_num, pathrec->dgid.raw, qpn);
1186
1187         return 0;
1188
1189 err_modify_send:
1190         ib_destroy_cm_id(p->id);
1191 err_id:
1192         p->id = NULL;
1193         ib_destroy_qp(p->qp);
1194 err_qp:
1195         p->qp = NULL;
1196         vfree(p->tx_ring);
1197 err_tx:
1198         return ret;
1199 }
1200
1201 static void ipoib_cm_tx_destroy(struct ipoib_cm_tx *p)
1202 {
1203         struct ipoib_dev_priv *priv = ipoib_priv(p->dev);
1204         struct ipoib_tx_buf *tx_req;
1205         unsigned long begin;
1206
1207         ipoib_dbg(priv, "Destroy active connection 0x%x head 0x%x tail 0x%x\n",
1208                   p->qp ? p->qp->qp_num : 0, p->tx_head, p->tx_tail);
1209
1210         if (p->id)
1211                 ib_destroy_cm_id(p->id);
1212
1213         if (p->tx_ring) {
1214                 /* Wait for all sends to complete */
1215                 begin = jiffies;
1216                 while ((int) p->tx_tail - (int) p->tx_head < 0) {
1217                         if (time_after(jiffies, begin + 5 * HZ)) {
1218                                 ipoib_warn(priv, "timing out; %d sends not completed\n",
1219                                            p->tx_head - p->tx_tail);
1220                                 goto timeout;
1221                         }
1222
1223                         usleep_range(1000, 2000);
1224                 }
1225         }
1226
1227 timeout:
1228
1229         while ((int) p->tx_tail - (int) p->tx_head < 0) {
1230                 tx_req = &p->tx_ring[p->tx_tail & (ipoib_sendq_size - 1)];
1231                 ipoib_dma_unmap_tx(priv, tx_req);
1232                 dev_kfree_skb_any(tx_req->skb);
1233                 netif_tx_lock_bh(p->dev);
1234                 ++p->tx_tail;
1235                 ++priv->tx_tail;
1236                 if (unlikely(priv->tx_head - priv->tx_tail == ipoib_sendq_size >> 1) &&
1237                     netif_queue_stopped(p->dev) &&
1238                     test_bit(IPOIB_FLAG_ADMIN_UP, &priv->flags))
1239                         netif_wake_queue(p->dev);
1240                 netif_tx_unlock_bh(p->dev);
1241         }
1242
1243         if (p->qp)
1244                 ib_destroy_qp(p->qp);
1245
1246         vfree(p->tx_ring);
1247         kfree(p);
1248 }
1249
1250 static int ipoib_cm_tx_handler(struct ib_cm_id *cm_id,
1251                                const struct ib_cm_event *event)
1252 {
1253         struct ipoib_cm_tx *tx = cm_id->context;
1254         struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1255         struct net_device *dev = priv->dev;
1256         struct ipoib_neigh *neigh;
1257         unsigned long flags;
1258         int ret;
1259
1260         switch (event->event) {
1261         case IB_CM_DREQ_RECEIVED:
1262                 ipoib_dbg(priv, "DREQ received.\n");
1263                 ib_send_cm_drep(cm_id, NULL, 0);
1264                 break;
1265         case IB_CM_REP_RECEIVED:
1266                 ipoib_dbg(priv, "REP received.\n");
1267                 ret = ipoib_cm_rep_handler(cm_id, event);
1268                 if (ret)
1269                         ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED,
1270                                        NULL, 0, NULL, 0);
1271                 break;
1272         case IB_CM_REQ_ERROR:
1273         case IB_CM_REJ_RECEIVED:
1274         case IB_CM_TIMEWAIT_EXIT:
1275                 ipoib_dbg(priv, "CM error %d.\n", event->event);
1276                 netif_tx_lock_bh(dev);
1277                 spin_lock_irqsave(&priv->lock, flags);
1278                 neigh = tx->neigh;
1279
1280                 if (neigh) {
1281                         neigh->cm = NULL;
1282                         ipoib_neigh_free(neigh);
1283
1284                         tx->neigh = NULL;
1285                 }
1286
1287                 if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1288                         list_move(&tx->list, &priv->cm.reap_list);
1289                         queue_work(priv->wq, &priv->cm.reap_task);
1290                 }
1291
1292                 spin_unlock_irqrestore(&priv->lock, flags);
1293                 netif_tx_unlock_bh(dev);
1294                 break;
1295         default:
1296                 break;
1297         }
1298
1299         return 0;
1300 }
1301
1302 struct ipoib_cm_tx *ipoib_cm_create_tx(struct net_device *dev, struct ipoib_path *path,
1303                                        struct ipoib_neigh *neigh)
1304 {
1305         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1306         struct ipoib_cm_tx *tx;
1307
1308         tx = kzalloc(sizeof(*tx), GFP_ATOMIC);
1309         if (!tx)
1310                 return NULL;
1311
1312         neigh->cm = tx;
1313         tx->neigh = neigh;
1314         tx->dev = dev;
1315         list_add(&tx->list, &priv->cm.start_list);
1316         set_bit(IPOIB_FLAG_INITIALIZED, &tx->flags);
1317         queue_work(priv->wq, &priv->cm.start_task);
1318         return tx;
1319 }
1320
1321 void ipoib_cm_destroy_tx(struct ipoib_cm_tx *tx)
1322 {
1323         struct ipoib_dev_priv *priv = ipoib_priv(tx->dev);
1324         unsigned long flags;
1325         if (test_and_clear_bit(IPOIB_FLAG_INITIALIZED, &tx->flags)) {
1326                 spin_lock_irqsave(&priv->lock, flags);
1327                 list_move(&tx->list, &priv->cm.reap_list);
1328                 queue_work(priv->wq, &priv->cm.reap_task);
1329                 ipoib_dbg(priv, "Reap connection for gid %pI6\n",
1330                           tx->neigh->daddr + 4);
1331                 tx->neigh = NULL;
1332                 spin_unlock_irqrestore(&priv->lock, flags);
1333         }
1334 }
1335
1336 #define QPN_AND_OPTIONS_OFFSET  4
1337
1338 static void ipoib_cm_tx_start(struct work_struct *work)
1339 {
1340         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1341                                                    cm.start_task);
1342         struct net_device *dev = priv->dev;
1343         struct ipoib_neigh *neigh;
1344         struct ipoib_cm_tx *p;
1345         unsigned long flags;
1346         struct ipoib_path *path;
1347         int ret;
1348
1349         struct sa_path_rec pathrec;
1350         u32 qpn;
1351
1352         netif_tx_lock_bh(dev);
1353         spin_lock_irqsave(&priv->lock, flags);
1354
1355         while (!list_empty(&priv->cm.start_list)) {
1356                 p = list_entry(priv->cm.start_list.next, typeof(*p), list);
1357                 list_del_init(&p->list);
1358                 neigh = p->neigh;
1359
1360                 qpn = IPOIB_QPN(neigh->daddr);
1361                 /*
1362                  * As long as the search is with these 2 locks,
1363                  * path existence indicates its validity.
1364                  */
1365                 path = __path_find(dev, neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1366                 if (!path) {
1367                         pr_info("%s ignore not valid path %pI6\n",
1368                                 __func__,
1369                                 neigh->daddr + QPN_AND_OPTIONS_OFFSET);
1370                         goto free_neigh;
1371                 }
1372                 memcpy(&pathrec, &path->pathrec, sizeof(pathrec));
1373
1374                 spin_unlock_irqrestore(&priv->lock, flags);
1375                 netif_tx_unlock_bh(dev);
1376
1377                 ret = ipoib_cm_tx_init(p, qpn, &pathrec);
1378
1379                 netif_tx_lock_bh(dev);
1380                 spin_lock_irqsave(&priv->lock, flags);
1381
1382                 if (ret) {
1383 free_neigh:
1384                         neigh = p->neigh;
1385                         if (neigh) {
1386                                 neigh->cm = NULL;
1387                                 ipoib_neigh_free(neigh);
1388                         }
1389                         list_del(&p->list);
1390                         kfree(p);
1391                 }
1392         }
1393
1394         spin_unlock_irqrestore(&priv->lock, flags);
1395         netif_tx_unlock_bh(dev);
1396 }
1397
1398 static void ipoib_cm_tx_reap(struct work_struct *work)
1399 {
1400         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1401                                                    cm.reap_task);
1402         struct net_device *dev = priv->dev;
1403         struct ipoib_cm_tx *p;
1404         unsigned long flags;
1405
1406         netif_tx_lock_bh(dev);
1407         spin_lock_irqsave(&priv->lock, flags);
1408
1409         while (!list_empty(&priv->cm.reap_list)) {
1410                 p = list_entry(priv->cm.reap_list.next, typeof(*p), list);
1411                 list_del_init(&p->list);
1412                 spin_unlock_irqrestore(&priv->lock, flags);
1413                 netif_tx_unlock_bh(dev);
1414                 ipoib_cm_tx_destroy(p);
1415                 netif_tx_lock_bh(dev);
1416                 spin_lock_irqsave(&priv->lock, flags);
1417         }
1418
1419         spin_unlock_irqrestore(&priv->lock, flags);
1420         netif_tx_unlock_bh(dev);
1421 }
1422
1423 static void ipoib_cm_skb_reap(struct work_struct *work)
1424 {
1425         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1426                                                    cm.skb_task);
1427         struct net_device *dev = priv->dev;
1428         struct sk_buff *skb;
1429         unsigned long flags;
1430         unsigned int mtu = priv->mcast_mtu;
1431
1432         netif_tx_lock_bh(dev);
1433         spin_lock_irqsave(&priv->lock, flags);
1434
1435         while ((skb = skb_dequeue(&priv->cm.skb_queue))) {
1436                 spin_unlock_irqrestore(&priv->lock, flags);
1437                 netif_tx_unlock_bh(dev);
1438
1439                 if (skb->protocol == htons(ETH_P_IP)) {
1440                         memset(IPCB(skb), 0, sizeof(*IPCB(skb)));
1441                         icmp_send(skb, ICMP_DEST_UNREACH, ICMP_FRAG_NEEDED, htonl(mtu));
1442                 }
1443 #if IS_ENABLED(CONFIG_IPV6)
1444                 else if (skb->protocol == htons(ETH_P_IPV6)) {
1445                         memset(IP6CB(skb), 0, sizeof(*IP6CB(skb)));
1446                         icmpv6_send(skb, ICMPV6_PKT_TOOBIG, 0, mtu);
1447                 }
1448 #endif
1449                 dev_kfree_skb_any(skb);
1450
1451                 netif_tx_lock_bh(dev);
1452                 spin_lock_irqsave(&priv->lock, flags);
1453         }
1454
1455         spin_unlock_irqrestore(&priv->lock, flags);
1456         netif_tx_unlock_bh(dev);
1457 }
1458
1459 void ipoib_cm_skb_too_long(struct net_device *dev, struct sk_buff *skb,
1460                            unsigned int mtu)
1461 {
1462         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1463         int e = skb_queue_empty(&priv->cm.skb_queue);
1464
1465         skb_dst_update_pmtu(skb, mtu);
1466
1467         skb_queue_tail(&priv->cm.skb_queue, skb);
1468         if (e)
1469                 queue_work(priv->wq, &priv->cm.skb_task);
1470 }
1471
1472 static void ipoib_cm_rx_reap(struct work_struct *work)
1473 {
1474         ipoib_cm_free_rx_reap_list(container_of(work, struct ipoib_dev_priv,
1475                                                 cm.rx_reap_task)->dev);
1476 }
1477
1478 static void ipoib_cm_stale_task(struct work_struct *work)
1479 {
1480         struct ipoib_dev_priv *priv = container_of(work, struct ipoib_dev_priv,
1481                                                    cm.stale_task.work);
1482         struct ipoib_cm_rx *p;
1483         int ret;
1484
1485         spin_lock_irq(&priv->lock);
1486         while (!list_empty(&priv->cm.passive_ids)) {
1487                 /* List is sorted by LRU, start from tail,
1488                  * stop when we see a recently used entry */
1489                 p = list_entry(priv->cm.passive_ids.prev, typeof(*p), list);
1490                 if (time_before_eq(jiffies, p->jiffies + IPOIB_CM_RX_TIMEOUT))
1491                         break;
1492                 list_move(&p->list, &priv->cm.rx_error_list);
1493                 p->state = IPOIB_CM_RX_ERROR;
1494                 spin_unlock_irq(&priv->lock);
1495                 ret = ib_modify_qp(p->qp, &ipoib_cm_err_attr, IB_QP_STATE);
1496                 if (ret)
1497                         ipoib_warn(priv, "unable to move qp to error state: %d\n", ret);
1498                 spin_lock_irq(&priv->lock);
1499         }
1500
1501         if (!list_empty(&priv->cm.passive_ids))
1502                 queue_delayed_work(priv->wq,
1503                                    &priv->cm.stale_task, IPOIB_CM_RX_DELAY);
1504         spin_unlock_irq(&priv->lock);
1505 }
1506
1507 static ssize_t show_mode(struct device *d, struct device_attribute *attr,
1508                          char *buf)
1509 {
1510         struct net_device *dev = to_net_dev(d);
1511         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1512
1513         if (test_bit(IPOIB_FLAG_ADMIN_CM, &priv->flags))
1514                 return sprintf(buf, "connected\n");
1515         else
1516                 return sprintf(buf, "datagram\n");
1517 }
1518
1519 static ssize_t set_mode(struct device *d, struct device_attribute *attr,
1520                         const char *buf, size_t count)
1521 {
1522         struct net_device *dev = to_net_dev(d);
1523         int ret;
1524
1525         if (!rtnl_trylock()) {
1526                 return restart_syscall();
1527         }
1528
1529         if (dev->reg_state != NETREG_REGISTERED) {
1530                 rtnl_unlock();
1531                 return -EPERM;
1532         }
1533
1534         ret = ipoib_set_mode(dev, buf);
1535
1536         /* The assumption is that the function ipoib_set_mode returned
1537          * with the rtnl held by it, if not the value -EBUSY returned,
1538          * then no need to rtnl_unlock
1539          */
1540         if (ret != -EBUSY)
1541                 rtnl_unlock();
1542
1543         return (!ret || ret == -EBUSY) ? count : ret;
1544 }
1545
1546 static DEVICE_ATTR(mode, S_IWUSR | S_IRUGO, show_mode, set_mode);
1547
1548 int ipoib_cm_add_mode_attr(struct net_device *dev)
1549 {
1550         return device_create_file(&dev->dev, &dev_attr_mode);
1551 }
1552
1553 static void ipoib_cm_create_srq(struct net_device *dev, int max_sge)
1554 {
1555         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1556         struct ib_srq_init_attr srq_init_attr = {
1557                 .srq_type = IB_SRQT_BASIC,
1558                 .attr = {
1559                         .max_wr  = ipoib_recvq_size,
1560                         .max_sge = max_sge
1561                 }
1562         };
1563
1564         priv->cm.srq = ib_create_srq(priv->pd, &srq_init_attr);
1565         if (IS_ERR(priv->cm.srq)) {
1566                 if (PTR_ERR(priv->cm.srq) != -EOPNOTSUPP)
1567                         pr_warn("%s: failed to allocate SRQ, error %ld\n",
1568                                priv->ca->name, PTR_ERR(priv->cm.srq));
1569                 priv->cm.srq = NULL;
1570                 return;
1571         }
1572
1573         priv->cm.srq_ring = vzalloc(array_size(ipoib_recvq_size,
1574                                                sizeof(*priv->cm.srq_ring)));
1575         if (!priv->cm.srq_ring) {
1576                 ib_destroy_srq(priv->cm.srq);
1577                 priv->cm.srq = NULL;
1578                 return;
1579         }
1580
1581 }
1582
1583 int ipoib_cm_dev_init(struct net_device *dev)
1584 {
1585         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1586         int max_srq_sge, i;
1587
1588         INIT_LIST_HEAD(&priv->cm.passive_ids);
1589         INIT_LIST_HEAD(&priv->cm.reap_list);
1590         INIT_LIST_HEAD(&priv->cm.start_list);
1591         INIT_LIST_HEAD(&priv->cm.rx_error_list);
1592         INIT_LIST_HEAD(&priv->cm.rx_flush_list);
1593         INIT_LIST_HEAD(&priv->cm.rx_drain_list);
1594         INIT_LIST_HEAD(&priv->cm.rx_reap_list);
1595         INIT_WORK(&priv->cm.start_task, ipoib_cm_tx_start);
1596         INIT_WORK(&priv->cm.reap_task, ipoib_cm_tx_reap);
1597         INIT_WORK(&priv->cm.skb_task, ipoib_cm_skb_reap);
1598         INIT_WORK(&priv->cm.rx_reap_task, ipoib_cm_rx_reap);
1599         INIT_DELAYED_WORK(&priv->cm.stale_task, ipoib_cm_stale_task);
1600
1601         skb_queue_head_init(&priv->cm.skb_queue);
1602
1603         ipoib_dbg(priv, "max_srq_sge=%d\n", priv->ca->attrs.max_srq_sge);
1604
1605         max_srq_sge = min_t(int, IPOIB_CM_RX_SG, priv->ca->attrs.max_srq_sge);
1606         ipoib_cm_create_srq(dev, max_srq_sge);
1607         if (ipoib_cm_has_srq(dev)) {
1608                 priv->cm.max_cm_mtu = max_srq_sge * PAGE_SIZE - 0x10;
1609                 priv->cm.num_frags  = max_srq_sge;
1610                 ipoib_dbg(priv, "max_cm_mtu = 0x%x, num_frags=%d\n",
1611                           priv->cm.max_cm_mtu, priv->cm.num_frags);
1612         } else {
1613                 priv->cm.max_cm_mtu = IPOIB_CM_MTU;
1614                 priv->cm.num_frags  = IPOIB_CM_RX_SG;
1615         }
1616
1617         ipoib_cm_init_rx_wr(dev, &priv->cm.rx_wr, priv->cm.rx_sge);
1618
1619         if (ipoib_cm_has_srq(dev)) {
1620                 for (i = 0; i < ipoib_recvq_size; ++i) {
1621                         if (!ipoib_cm_alloc_rx_skb(dev, priv->cm.srq_ring, i,
1622                                                    priv->cm.num_frags - 1,
1623                                                    priv->cm.srq_ring[i].mapping,
1624                                                    GFP_KERNEL)) {
1625                                 ipoib_warn(priv, "failed to allocate "
1626                                            "receive buffer %d\n", i);
1627                                 ipoib_cm_dev_cleanup(dev);
1628                                 return -ENOMEM;
1629                         }
1630
1631                         if (ipoib_cm_post_receive_srq(dev, i)) {
1632                                 ipoib_warn(priv, "ipoib_cm_post_receive_srq "
1633                                            "failed for buf %d\n", i);
1634                                 ipoib_cm_dev_cleanup(dev);
1635                                 return -EIO;
1636                         }
1637                 }
1638         }
1639
1640         priv->dev->dev_addr[0] = IPOIB_FLAGS_RC;
1641         return 0;
1642 }
1643
1644 void ipoib_cm_dev_cleanup(struct net_device *dev)
1645 {
1646         struct ipoib_dev_priv *priv = ipoib_priv(dev);
1647         int ret;
1648
1649         if (!priv->cm.srq)
1650                 return;
1651
1652         ipoib_dbg(priv, "Cleanup ipoib connected mode.\n");
1653
1654         ret = ib_destroy_srq(priv->cm.srq);
1655         if (ret)
1656                 ipoib_warn(priv, "ib_destroy_srq failed: %d\n", ret);
1657
1658         priv->cm.srq = NULL;
1659         if (!priv->cm.srq_ring)
1660                 return;
1661
1662         ipoib_cm_free_rx_ring(dev, priv->cm.srq_ring);
1663         priv->cm.srq_ring = NULL;
1664 }