drivers/infiniband/ulp/ipoib/ipoib_cm.c

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