drivers/infiniband/hw/irdma/cm.c

   1 // SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   2 /* Copyright (c) 2015 - 2021 Intel Corporation */
   3 #include "main.h"
   4 #include "trace.h"
   5
   6 static void irdma_cm_post_event(struct irdma_cm_event *event);
   7 static void irdma_disconnect_worker(struct work_struct *work);
   8
   9 /**
  10  * irdma_free_sqbuf - put back puda buffer if refcount is 0
  11  * @vsi: The VSI structure of the device
  12  * @bufp: puda buffer to free
  13  */
  14 void irdma_free_sqbuf(struct irdma_sc_vsi *vsi, void *bufp)
  15 {
  16         struct irdma_puda_buf *buf = bufp;
  17         struct irdma_puda_rsrc *ilq = vsi->ilq;
  18
  19         if (refcount_dec_and_test(&buf->refcount))
  20                 irdma_puda_ret_bufpool(ilq, buf);
  21 }
  22
  23 /**
  24  * irdma_record_ird_ord - Record IRD/ORD passed in
  25  * @cm_node: connection's node
  26  * @conn_ird: connection IRD
  27  * @conn_ord: connection ORD
  28  */
  29 static void irdma_record_ird_ord(struct irdma_cm_node *cm_node, u32 conn_ird,
  30                                  u32 conn_ord)
  31 {
  32         if (conn_ird > cm_node->dev->hw_attrs.max_hw_ird)
  33                 conn_ird = cm_node->dev->hw_attrs.max_hw_ird;
  34
  35         if (conn_ord > cm_node->dev->hw_attrs.max_hw_ord)
  36                 conn_ord = cm_node->dev->hw_attrs.max_hw_ord;
  37         else if (!conn_ord && cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO)
  38                 conn_ord = 1;
  39         cm_node->ird_size = conn_ird;
  40         cm_node->ord_size = conn_ord;
  41 }
  42
  43 /**
  44  * irdma_copy_ip_ntohl - copy IP address from  network to host
  45  * @dst: IP address in host order
  46  * @src: IP address in network order (big endian)
  47  */
  48 void irdma_copy_ip_ntohl(u32 *dst, __be32 *src)
  49 {
  50         *dst++ = ntohl(*src++);
  51         *dst++ = ntohl(*src++);
  52         *dst++ = ntohl(*src++);
  53         *dst = ntohl(*src);
  54 }
  55
  56 /**
  57  * irdma_copy_ip_htonl - copy IP address from host to network order
  58  * @dst: IP address in network order (big endian)
  59  * @src: IP address in host order
  60  */
  61 void irdma_copy_ip_htonl(__be32 *dst, u32 *src)
  62 {
  63         *dst++ = htonl(*src++);
  64         *dst++ = htonl(*src++);
  65         *dst++ = htonl(*src++);
  66         *dst = htonl(*src);
  67 }
  68
  69 /**
  70  * irdma_get_addr_info
  71  * @cm_node: contains ip/tcp info
  72  * @cm_info: to get a copy of the cm_node ip/tcp info
  73  */
  74 static void irdma_get_addr_info(struct irdma_cm_node *cm_node,
  75                                 struct irdma_cm_info *cm_info)
  76 {
  77         memset(cm_info, 0, sizeof(*cm_info));
  78         cm_info->ipv4 = cm_node->ipv4;
  79         cm_info->vlan_id = cm_node->vlan_id;
  80         memcpy(cm_info->loc_addr, cm_node->loc_addr, sizeof(cm_info->loc_addr));
  81         memcpy(cm_info->rem_addr, cm_node->rem_addr, sizeof(cm_info->rem_addr));
  82         cm_info->loc_port = cm_node->loc_port;
  83         cm_info->rem_port = cm_node->rem_port;
  84 }
  85
  86 /**
  87  * irdma_fill_sockaddr4 - fill in addr info for IPv4 connection
  88  * @cm_node: connection's node
  89  * @event: upper layer's cm event
  90  */
  91 static inline void irdma_fill_sockaddr4(struct irdma_cm_node *cm_node,
  92                                         struct iw_cm_event *event)
  93 {
  94         struct sockaddr_in *laddr = (struct sockaddr_in *)&event->local_addr;
  95         struct sockaddr_in *raddr = (struct sockaddr_in *)&event->remote_addr;
  96
  97         laddr->sin_family = AF_INET;
  98         raddr->sin_family = AF_INET;
  99
 100         laddr->sin_port = htons(cm_node->loc_port);
 101         raddr->sin_port = htons(cm_node->rem_port);
 102
 103         laddr->sin_addr.s_addr = htonl(cm_node->loc_addr[0]);
 104         raddr->sin_addr.s_addr = htonl(cm_node->rem_addr[0]);
 105 }
 106
 107 /**
 108  * irdma_fill_sockaddr6 - fill in addr info for IPv6 connection
 109  * @cm_node: connection's node
 110  * @event: upper layer's cm event
 111  */
 112 static inline void irdma_fill_sockaddr6(struct irdma_cm_node *cm_node,
 113                                         struct iw_cm_event *event)
 114 {
 115         struct sockaddr_in6 *laddr6 = (struct sockaddr_in6 *)&event->local_addr;
 116         struct sockaddr_in6 *raddr6 = (struct sockaddr_in6 *)&event->remote_addr;
 117
 118         laddr6->sin6_family = AF_INET6;
 119         raddr6->sin6_family = AF_INET6;
 120
 121         laddr6->sin6_port = htons(cm_node->loc_port);
 122         raddr6->sin6_port = htons(cm_node->rem_port);
 123
 124         irdma_copy_ip_htonl(laddr6->sin6_addr.in6_u.u6_addr32,
 125                             cm_node->loc_addr);
 126         irdma_copy_ip_htonl(raddr6->sin6_addr.in6_u.u6_addr32,
 127                             cm_node->rem_addr);
 128 }
 129
 130 /**
 131  * irdma_get_cmevent_info - for cm event upcall
 132  * @cm_node: connection's node
 133  * @cm_id: upper layers cm struct for the event
 134  * @event: upper layer's cm event
 135  */
 136 static inline void irdma_get_cmevent_info(struct irdma_cm_node *cm_node,
 137                                           struct iw_cm_id *cm_id,
 138                                           struct iw_cm_event *event)
 139 {
 140         memcpy(&event->local_addr, &cm_id->m_local_addr,
 141                sizeof(event->local_addr));
 142         memcpy(&event->remote_addr, &cm_id->m_remote_addr,
 143                sizeof(event->remote_addr));
 144         if (cm_node) {
 145                 event->private_data = cm_node->pdata_buf;
 146                 event->private_data_len = (u8)cm_node->pdata.size;
 147                 event->ird = cm_node->ird_size;
 148                 event->ord = cm_node->ord_size;
 149         }
 150 }
 151
 152 /**
 153  * irdma_send_cm_event - upcall cm's event handler
 154  * @cm_node: connection's node
 155  * @cm_id: upper layer's cm info struct
 156  * @type: Event type to indicate
 157  * @status: status for the event type
 158  */
 159 static int irdma_send_cm_event(struct irdma_cm_node *cm_node,
 160                                struct iw_cm_id *cm_id,
 161                                enum iw_cm_event_type type, int status)
 162 {
 163         struct iw_cm_event event = {};
 164
 165         event.event = type;
 166         event.status = status;
 167         trace_irdma_send_cm_event(cm_node, cm_id, type, status,
 168                                   __builtin_return_address(0));
 169
 170         ibdev_dbg(&cm_node->iwdev->ibdev,
 171                   "CM: cm_node %p cm_id=%p state=%d accel=%d event_type=%d status=%d\n",
 172                   cm_node, cm_id, cm_node->accelerated, cm_node->state, type,
 173                   status);
 174
 175         switch (type) {
 176         case IW_CM_EVENT_CONNECT_REQUEST:
 177                 if (cm_node->ipv4)
 178                         irdma_fill_sockaddr4(cm_node, &event);
 179                 else
 180                         irdma_fill_sockaddr6(cm_node, &event);
 181                 event.provider_data = cm_node;
 182                 event.private_data = cm_node->pdata_buf;
 183                 event.private_data_len = (u8)cm_node->pdata.size;
 184                 event.ird = cm_node->ird_size;
 185                 break;
 186         case IW_CM_EVENT_CONNECT_REPLY:
 187                 irdma_get_cmevent_info(cm_node, cm_id, &event);
 188                 break;
 189         case IW_CM_EVENT_ESTABLISHED:
 190                 event.ird = cm_node->ird_size;
 191                 event.ord = cm_node->ord_size;
 192                 break;
 193         case IW_CM_EVENT_DISCONNECT:
 194         case IW_CM_EVENT_CLOSE:
 195                 /* Wait if we are in RTS but havent issued the iwcm event upcall */
 196                 if (!cm_node->accelerated)
 197                         wait_for_completion(&cm_node->establish_comp);
 198                 break;
 199         default:
 200                 return -EINVAL;
 201         }
 202
 203         return cm_id->event_handler(cm_id, &event);
 204 }
 205
 206 /**
 207  * irdma_timer_list_prep - add connection nodes to a list to perform timer tasks
 208  * @cm_core: cm's core
 209  * @timer_list: a timer list to which cm_node will be selected
 210  */
 211 static void irdma_timer_list_prep(struct irdma_cm_core *cm_core,
 212                                   struct list_head *timer_list)
 213 {
 214         struct irdma_cm_node *cm_node;
 215         int bkt;
 216
 217         hash_for_each_rcu(cm_core->cm_hash_tbl, bkt, cm_node, list) {
 218                 if ((cm_node->close_entry || cm_node->send_entry) &&
 219                     refcount_inc_not_zero(&cm_node->refcnt))
 220                         list_add(&cm_node->timer_entry, timer_list);
 221         }
 222 }
 223
 224 /**
 225  * irdma_create_event - create cm event
 226  * @cm_node: connection's node
 227  * @type: Event type to generate
 228  */
 229 static struct irdma_cm_event *irdma_create_event(struct irdma_cm_node *cm_node,
 230                                                  enum irdma_cm_event_type type)
 231 {
 232         struct irdma_cm_event *event;
 233
 234         if (!cm_node->cm_id)
 235                 return NULL;
 236
 237         event = kzalloc(sizeof(*event), GFP_ATOMIC);
 238
 239         if (!event)
 240                 return NULL;
 241
 242         event->type = type;
 243         event->cm_node = cm_node;
 244         memcpy(event->cm_info.rem_addr, cm_node->rem_addr,
 245                sizeof(event->cm_info.rem_addr));
 246         memcpy(event->cm_info.loc_addr, cm_node->loc_addr,
 247                sizeof(event->cm_info.loc_addr));
 248         event->cm_info.rem_port = cm_node->rem_port;
 249         event->cm_info.loc_port = cm_node->loc_port;
 250         event->cm_info.cm_id = cm_node->cm_id;
 251         ibdev_dbg(&cm_node->iwdev->ibdev,
 252                   "CM: node=%p event=%p type=%u dst=%pI4 src=%pI4\n", cm_node,
 253                   event, type, event->cm_info.loc_addr,
 254                   event->cm_info.rem_addr);
 255         trace_irdma_create_event(cm_node, type, __builtin_return_address(0));
 256         irdma_cm_post_event(event);
 257
 258         return event;
 259 }
 260
 261 /**
 262  * irdma_free_retrans_entry - free send entry
 263  * @cm_node: connection's node
 264  */
 265 static void irdma_free_retrans_entry(struct irdma_cm_node *cm_node)
 266 {
 267         struct irdma_device *iwdev = cm_node->iwdev;
 268         struct irdma_timer_entry *send_entry;
 269
 270         send_entry = cm_node->send_entry;
 271         if (!send_entry)
 272                 return;
 273
 274         cm_node->send_entry = NULL;
 275         irdma_free_sqbuf(&iwdev->vsi, send_entry->sqbuf);
 276         kfree(send_entry);
 277         refcount_dec(&cm_node->refcnt);
 278 }
 279
 280 /**
 281  * irdma_cleanup_retrans_entry - free send entry with lock
 282  * @cm_node: connection's node
 283  */
 284 static void irdma_cleanup_retrans_entry(struct irdma_cm_node *cm_node)
 285 {
 286         unsigned long flags;
 287
 288         spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
 289         irdma_free_retrans_entry(cm_node);
 290         spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
 291 }
 292
 293 /**
 294  * irdma_form_ah_cm_frame - get a free packet and build frame with address handle
 295  * @cm_node: connection's node ionfo to use in frame
 296  * @options: pointer to options info
 297  * @hdr: pointer mpa header
 298  * @pdata: pointer to private data
 299  * @flags:  indicates FIN or ACK
 300  */
 301 static struct irdma_puda_buf *irdma_form_ah_cm_frame(struct irdma_cm_node *cm_node,
 302                                                      struct irdma_kmem_info *options,
 303                                                      struct irdma_kmem_info *hdr,
 304                                                      struct irdma_mpa_priv_info *pdata,
 305                                                      u8 flags)
 306 {
 307         struct irdma_puda_buf *sqbuf;
 308         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
 309         u8 *buf;
 310         struct tcphdr *tcph;
 311         u16 pktsize;
 312         u32 opts_len = 0;
 313         u32 pd_len = 0;
 314         u32 hdr_len = 0;
 315
 316         if (!cm_node->ah || !cm_node->ah->ah_info.ah_valid) {
 317                 ibdev_dbg(&cm_node->iwdev->ibdev, "CM: AH invalid\n");
 318                 return NULL;
 319         }
 320
 321         sqbuf = irdma_puda_get_bufpool(vsi->ilq);
 322         if (!sqbuf) {
 323                 ibdev_dbg(&cm_node->iwdev->ibdev, "CM: SQ buf NULL\n");
 324                 return NULL;
 325         }
 326
 327         sqbuf->ah_id = cm_node->ah->ah_info.ah_idx;
 328         buf = sqbuf->mem.va;
 329         if (options)
 330                 opts_len = (u32)options->size;
 331
 332         if (hdr)
 333                 hdr_len = hdr->size;
 334
 335         if (pdata)
 336                 pd_len = pdata->size;
 337
 338         pktsize = sizeof(*tcph) + opts_len + hdr_len + pd_len;
 339
 340         memset(buf, 0, pktsize);
 341
 342         sqbuf->totallen = pktsize;
 343         sqbuf->tcphlen = sizeof(*tcph) + opts_len;
 344         sqbuf->scratch = cm_node;
 345
 346         tcph = (struct tcphdr *)buf;
 347         buf += sizeof(*tcph);
 348
 349         tcph->source = htons(cm_node->loc_port);
 350         tcph->dest = htons(cm_node->rem_port);
 351         tcph->seq = htonl(cm_node->tcp_cntxt.loc_seq_num);
 352
 353         if (flags & SET_ACK) {
 354                 cm_node->tcp_cntxt.loc_ack_num = cm_node->tcp_cntxt.rcv_nxt;
 355                 tcph->ack_seq = htonl(cm_node->tcp_cntxt.loc_ack_num);
 356                 tcph->ack = 1;
 357         } else {
 358                 tcph->ack_seq = 0;
 359         }
 360
 361         if (flags & SET_SYN) {
 362                 cm_node->tcp_cntxt.loc_seq_num++;
 363                 tcph->syn = 1;
 364         } else {
 365                 cm_node->tcp_cntxt.loc_seq_num += hdr_len + pd_len;
 366         }
 367
 368         if (flags & SET_FIN) {
 369                 cm_node->tcp_cntxt.loc_seq_num++;
 370                 tcph->fin = 1;
 371         }
 372
 373         if (flags & SET_RST)
 374                 tcph->rst = 1;
 375
 376         tcph->doff = (u16)((sizeof(*tcph) + opts_len + 3) >> 2);
 377         sqbuf->tcphlen = tcph->doff << 2;
 378         tcph->window = htons(cm_node->tcp_cntxt.rcv_wnd);
 379         tcph->urg_ptr = 0;
 380
 381         if (opts_len) {
 382                 memcpy(buf, options->addr, opts_len);
 383                 buf += opts_len;
 384         }
 385
 386         if (hdr_len) {
 387                 memcpy(buf, hdr->addr, hdr_len);
 388                 buf += hdr_len;
 389         }
 390
 391         if (pdata && pdata->addr)
 392                 memcpy(buf, pdata->addr, pdata->size);
 393
 394         refcount_set(&sqbuf->refcount, 1);
 395
 396         print_hex_dump_debug("ILQ: TRANSMIT ILQ BUFFER", DUMP_PREFIX_OFFSET,
 397                              16, 8, sqbuf->mem.va, sqbuf->totallen, false);
 398
 399         return sqbuf;
 400 }
 401
 402 /**
 403  * irdma_form_uda_cm_frame - get a free packet and build frame full tcpip packet
 404  * @cm_node: connection's node ionfo to use in frame
 405  * @options: pointer to options info
 406  * @hdr: pointer mpa header
 407  * @pdata: pointer to private data
 408  * @flags:  indicates FIN or ACK
 409  */
 410 static struct irdma_puda_buf *irdma_form_uda_cm_frame(struct irdma_cm_node *cm_node,
 411                                                       struct irdma_kmem_info *options,
 412                                                       struct irdma_kmem_info *hdr,
 413                                                       struct irdma_mpa_priv_info *pdata,
 414                                                       u8 flags)
 415 {
 416         struct irdma_puda_buf *sqbuf;
 417         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
 418         u8 *buf;
 419
 420         struct tcphdr *tcph;
 421         struct iphdr *iph;
 422         struct ipv6hdr *ip6h;
 423         struct ethhdr *ethh;
 424         u16 pktsize;
 425         u16 eth_hlen = ETH_HLEN;
 426         u32 opts_len = 0;
 427         u32 pd_len = 0;
 428         u32 hdr_len = 0;
 429
 430         u16 vtag;
 431
 432         sqbuf = irdma_puda_get_bufpool(vsi->ilq);
 433         if (!sqbuf)
 434                 return NULL;
 435
 436         buf = sqbuf->mem.va;
 437
 438         if (options)
 439                 opts_len = (u32)options->size;
 440
 441         if (hdr)
 442                 hdr_len = hdr->size;
 443
 444         if (pdata)
 445                 pd_len = pdata->size;
 446
 447         if (cm_node->vlan_id < VLAN_N_VID)
 448                 eth_hlen += 4;
 449
 450         if (cm_node->ipv4)
 451                 pktsize = sizeof(*iph) + sizeof(*tcph);
 452         else
 453                 pktsize = sizeof(*ip6h) + sizeof(*tcph);
 454         pktsize += opts_len + hdr_len + pd_len;
 455
 456         memset(buf, 0, eth_hlen + pktsize);
 457
 458         sqbuf->totallen = pktsize + eth_hlen;
 459         sqbuf->maclen = eth_hlen;
 460         sqbuf->tcphlen = sizeof(*tcph) + opts_len;
 461         sqbuf->scratch = cm_node;
 462
 463         ethh = (struct ethhdr *)buf;
 464         buf += eth_hlen;
 465
 466         if (cm_node->do_lpb)
 467                 sqbuf->do_lpb = true;
 468
 469         if (cm_node->ipv4) {
 470                 sqbuf->ipv4 = true;
 471
 472                 iph = (struct iphdr *)buf;
 473                 buf += sizeof(*iph);
 474                 tcph = (struct tcphdr *)buf;
 475                 buf += sizeof(*tcph);
 476
 477                 ether_addr_copy(ethh->h_dest, cm_node->rem_mac);
 478                 ether_addr_copy(ethh->h_source, cm_node->loc_mac);
 479                 if (cm_node->vlan_id < VLAN_N_VID) {
 480                         ((struct vlan_ethhdr *)ethh)->h_vlan_proto =
 481                                 htons(ETH_P_8021Q);
 482                         vtag = (cm_node->user_pri << VLAN_PRIO_SHIFT) |
 483                                cm_node->vlan_id;
 484                         ((struct vlan_ethhdr *)ethh)->h_vlan_TCI = htons(vtag);
 485
 486                         ((struct vlan_ethhdr *)ethh)->h_vlan_encapsulated_proto =
 487                                 htons(ETH_P_IP);
 488                 } else {
 489                         ethh->h_proto = htons(ETH_P_IP);
 490                 }
 491
 492                 iph->version = IPVERSION;
 493                 iph->ihl = 5; /* 5 * 4Byte words, IP headr len */
 494                 iph->tos = cm_node->tos;
 495                 iph->tot_len = htons(pktsize);
 496                 iph->id = htons(++cm_node->tcp_cntxt.loc_id);
 497
 498                 iph->frag_off = htons(0x4000);
 499                 iph->ttl = 0x40;
 500                 iph->protocol = IPPROTO_TCP;
 501                 iph->saddr = htonl(cm_node->loc_addr[0]);
 502                 iph->daddr = htonl(cm_node->rem_addr[0]);
 503         } else {
 504                 sqbuf->ipv4 = false;
 505                 ip6h = (struct ipv6hdr *)buf;
 506                 buf += sizeof(*ip6h);
 507                 tcph = (struct tcphdr *)buf;
 508                 buf += sizeof(*tcph);
 509
 510                 ether_addr_copy(ethh->h_dest, cm_node->rem_mac);
 511                 ether_addr_copy(ethh->h_source, cm_node->loc_mac);
 512                 if (cm_node->vlan_id < VLAN_N_VID) {
 513                         ((struct vlan_ethhdr *)ethh)->h_vlan_proto =
 514                                 htons(ETH_P_8021Q);
 515                         vtag = (cm_node->user_pri << VLAN_PRIO_SHIFT) |
 516                                cm_node->vlan_id;
 517                         ((struct vlan_ethhdr *)ethh)->h_vlan_TCI = htons(vtag);
 518                         ((struct vlan_ethhdr *)ethh)->h_vlan_encapsulated_proto =
 519                                 htons(ETH_P_IPV6);
 520                 } else {
 521                         ethh->h_proto = htons(ETH_P_IPV6);
 522                 }
 523                 ip6h->version = 6;
 524                 ip6h->priority = cm_node->tos >> 4;
 525                 ip6h->flow_lbl[0] = cm_node->tos << 4;
 526                 ip6h->flow_lbl[1] = 0;
 527                 ip6h->flow_lbl[2] = 0;
 528                 ip6h->payload_len = htons(pktsize - sizeof(*ip6h));
 529                 ip6h->nexthdr = 6;
 530                 ip6h->hop_limit = 128;
 531                 irdma_copy_ip_htonl(ip6h->saddr.in6_u.u6_addr32,
 532                                     cm_node->loc_addr);
 533                 irdma_copy_ip_htonl(ip6h->daddr.in6_u.u6_addr32,
 534                                     cm_node->rem_addr);
 535         }
 536
 537         tcph->source = htons(cm_node->loc_port);
 538         tcph->dest = htons(cm_node->rem_port);
 539         tcph->seq = htonl(cm_node->tcp_cntxt.loc_seq_num);
 540
 541         if (flags & SET_ACK) {
 542                 cm_node->tcp_cntxt.loc_ack_num = cm_node->tcp_cntxt.rcv_nxt;
 543                 tcph->ack_seq = htonl(cm_node->tcp_cntxt.loc_ack_num);
 544                 tcph->ack = 1;
 545         } else {
 546                 tcph->ack_seq = 0;
 547         }
 548
 549         if (flags & SET_SYN) {
 550                 cm_node->tcp_cntxt.loc_seq_num++;
 551                 tcph->syn = 1;
 552         } else {
 553                 cm_node->tcp_cntxt.loc_seq_num += hdr_len + pd_len;
 554         }
 555
 556         if (flags & SET_FIN) {
 557                 cm_node->tcp_cntxt.loc_seq_num++;
 558                 tcph->fin = 1;
 559         }
 560
 561         if (flags & SET_RST)
 562                 tcph->rst = 1;
 563
 564         tcph->doff = (u16)((sizeof(*tcph) + opts_len + 3) >> 2);
 565         sqbuf->tcphlen = tcph->doff << 2;
 566         tcph->window = htons(cm_node->tcp_cntxt.rcv_wnd);
 567         tcph->urg_ptr = 0;
 568
 569         if (opts_len) {
 570                 memcpy(buf, options->addr, opts_len);
 571                 buf += opts_len;
 572         }
 573
 574         if (hdr_len) {
 575                 memcpy(buf, hdr->addr, hdr_len);
 576                 buf += hdr_len;
 577         }
 578
 579         if (pdata && pdata->addr)
 580                 memcpy(buf, pdata->addr, pdata->size);
 581
 582         refcount_set(&sqbuf->refcount, 1);
 583
 584         print_hex_dump_debug("ILQ: TRANSMIT ILQ BUFFER", DUMP_PREFIX_OFFSET,
 585                              16, 8, sqbuf->mem.va, sqbuf->totallen, false);
 586         return sqbuf;
 587 }
 588
 589 /**
 590  * irdma_send_reset - Send RST packet
 591  * @cm_node: connection's node
 592  */
 593 int irdma_send_reset(struct irdma_cm_node *cm_node)
 594 {
 595         struct irdma_puda_buf *sqbuf;
 596         int flags = SET_RST | SET_ACK;
 597
 598         trace_irdma_send_reset(cm_node, 0, __builtin_return_address(0));
 599         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
 600                                                 flags);
 601         if (!sqbuf)
 602                 return -ENOMEM;
 603
 604         ibdev_dbg(&cm_node->iwdev->ibdev,
 605                   "CM: caller: %pS cm_node %p cm_id=%p accel=%d state=%d rem_port=0x%04x, loc_port=0x%04x rem_addr=%pI4 loc_addr=%pI4\n",
 606                   __builtin_return_address(0), cm_node, cm_node->cm_id,
 607                   cm_node->accelerated, cm_node->state, cm_node->rem_port,
 608                   cm_node->loc_port, cm_node->rem_addr, cm_node->loc_addr);
 609
 610         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 0,
 611                                        1);
 612 }
 613
 614 /**
 615  * irdma_active_open_err - send event for active side cm error
 616  * @cm_node: connection's node
 617  * @reset: Flag to send reset or not
 618  */
 619 static void irdma_active_open_err(struct irdma_cm_node *cm_node, bool reset)
 620 {
 621         trace_irdma_active_open_err(cm_node, reset,
 622                                     __builtin_return_address(0));
 623         irdma_cleanup_retrans_entry(cm_node);
 624         cm_node->cm_core->stats_connect_errs++;
 625         if (reset) {
 626                 ibdev_dbg(&cm_node->iwdev->ibdev,
 627                           "CM: cm_node=%p state=%d\n", cm_node,
 628                           cm_node->state);
 629                 refcount_inc(&cm_node->refcnt);
 630                 irdma_send_reset(cm_node);
 631         }
 632
 633         cm_node->state = IRDMA_CM_STATE_CLOSED;
 634         irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
 635 }
 636
 637 /**
 638  * irdma_passive_open_err - handle passive side cm error
 639  * @cm_node: connection's node
 640  * @reset: send reset or just free cm_node
 641  */
 642 static void irdma_passive_open_err(struct irdma_cm_node *cm_node, bool reset)
 643 {
 644         irdma_cleanup_retrans_entry(cm_node);
 645         cm_node->cm_core->stats_passive_errs++;
 646         cm_node->state = IRDMA_CM_STATE_CLOSED;
 647         ibdev_dbg(&cm_node->iwdev->ibdev, "CM: cm_node=%p state =%d\n",
 648                   cm_node, cm_node->state);
 649         trace_irdma_passive_open_err(cm_node, reset,
 650                                      __builtin_return_address(0));
 651         if (reset)
 652                 irdma_send_reset(cm_node);
 653         else
 654                 irdma_rem_ref_cm_node(cm_node);
 655 }
 656
 657 /**
 658  * irdma_event_connect_error - to create connect error event
 659  * @event: cm information for connect event
 660  */
 661 static void irdma_event_connect_error(struct irdma_cm_event *event)
 662 {
 663         struct irdma_qp *iwqp;
 664         struct iw_cm_id *cm_id;
 665
 666         cm_id = event->cm_node->cm_id;
 667         if (!cm_id)
 668                 return;
 669
 670         iwqp = cm_id->provider_data;
 671
 672         if (!iwqp || !iwqp->iwdev)
 673                 return;
 674
 675         iwqp->cm_id = NULL;
 676         cm_id->provider_data = NULL;
 677         irdma_send_cm_event(event->cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY,
 678                             -ECONNRESET);
 679         irdma_rem_ref_cm_node(event->cm_node);
 680 }
 681
 682 /**
 683  * irdma_process_options - process options from TCP header
 684  * @cm_node: connection's node
 685  * @optionsloc: point to start of options
 686  * @optionsize: size of all options
 687  * @syn_pkt: flag if syn packet
 688  */
 689 static int irdma_process_options(struct irdma_cm_node *cm_node, u8 *optionsloc,
 690                                  u32 optionsize, u32 syn_pkt)
 691 {
 692         u32 tmp;
 693         u32 offset = 0;
 694         union all_known_options *all_options;
 695         char got_mss_option = 0;
 696
 697         while (offset < optionsize) {
 698                 all_options = (union all_known_options *)(optionsloc + offset);
 699                 switch (all_options->base.optionnum) {
 700                 case OPTION_NUM_EOL:
 701                         offset = optionsize;
 702                         break;
 703                 case OPTION_NUM_NONE:
 704                         offset += 1;
 705                         continue;
 706                 case OPTION_NUM_MSS:
 707                         ibdev_dbg(&cm_node->iwdev->ibdev,
 708                                   "CM: MSS Length: %d Offset: %d Size: %d\n",
 709                                   all_options->mss.len, offset, optionsize);
 710                         got_mss_option = 1;
 711                         if (all_options->mss.len != 4)
 712                                 return -EINVAL;
 713                         tmp = ntohs(all_options->mss.mss);
 714                         if ((cm_node->ipv4 &&
 715                              (tmp + IRDMA_MTU_TO_MSS_IPV4) < IRDMA_MIN_MTU_IPV4) ||
 716                             (!cm_node->ipv4 &&
 717                              (tmp + IRDMA_MTU_TO_MSS_IPV6) < IRDMA_MIN_MTU_IPV6))
 718                                 return -EINVAL;
 719                         if (tmp < cm_node->tcp_cntxt.mss)
 720                                 cm_node->tcp_cntxt.mss = tmp;
 721                         break;
 722                 case OPTION_NUM_WINDOW_SCALE:
 723                         cm_node->tcp_cntxt.snd_wscale =
 724                                 all_options->windowscale.shiftcount;
 725                         break;
 726                 default:
 727                         ibdev_dbg(&cm_node->iwdev->ibdev,
 728                                   "CM: Unsupported TCP Option: %x\n",
 729                                   all_options->base.optionnum);
 730                         break;
 731                 }
 732                 offset += all_options->base.len;
 733         }
 734         if (!got_mss_option && syn_pkt)
 735                 cm_node->tcp_cntxt.mss = IRDMA_CM_DEFAULT_MSS;
 736
 737         return 0;
 738 }
 739
 740 /**
 741  * irdma_handle_tcp_options - setup TCP context info after parsing TCP options
 742  * @cm_node: connection's node
 743  * @tcph: pointer tcp header
 744  * @optionsize: size of options rcvd
 745  * @passive: active or passive flag
 746  */
 747 static int irdma_handle_tcp_options(struct irdma_cm_node *cm_node,
 748                                     struct tcphdr *tcph, int optionsize,
 749                                     int passive)
 750 {
 751         u8 *optionsloc = (u8 *)&tcph[1];
 752         int ret;
 753
 754         if (optionsize) {
 755                 ret = irdma_process_options(cm_node, optionsloc, optionsize,
 756                                             (u32)tcph->syn);
 757                 if (ret) {
 758                         ibdev_dbg(&cm_node->iwdev->ibdev,
 759                                   "CM: Node %p, Sending Reset\n", cm_node);
 760                         if (passive)
 761                                 irdma_passive_open_err(cm_node, true);
 762                         else
 763                                 irdma_active_open_err(cm_node, true);
 764                         return ret;
 765                 }
 766         }
 767
 768         cm_node->tcp_cntxt.snd_wnd = ntohs(tcph->window)
 769                                      << cm_node->tcp_cntxt.snd_wscale;
 770
 771         if (cm_node->tcp_cntxt.snd_wnd > cm_node->tcp_cntxt.max_snd_wnd)
 772                 cm_node->tcp_cntxt.max_snd_wnd = cm_node->tcp_cntxt.snd_wnd;
 773
 774         return 0;
 775 }
 776
 777 /**
 778  * irdma_build_mpa_v1 - build a MPA V1 frame
 779  * @cm_node: connection's node
 780  * @start_addr: address where to build frame
 781  * @mpa_key: to do read0 or write0
 782  */
 783 static void irdma_build_mpa_v1(struct irdma_cm_node *cm_node, void *start_addr,
 784                                u8 mpa_key)
 785 {
 786         struct ietf_mpa_v1 *mpa_frame = start_addr;
 787
 788         switch (mpa_key) {
 789         case MPA_KEY_REQUEST:
 790                 memcpy(mpa_frame->key, IEFT_MPA_KEY_REQ, IETF_MPA_KEY_SIZE);
 791                 break;
 792         case MPA_KEY_REPLY:
 793                 memcpy(mpa_frame->key, IEFT_MPA_KEY_REP, IETF_MPA_KEY_SIZE);
 794                 break;
 795         default:
 796                 break;
 797         }
 798         mpa_frame->flags = IETF_MPA_FLAGS_CRC;
 799         mpa_frame->rev = cm_node->mpa_frame_rev;
 800         mpa_frame->priv_data_len = htons(cm_node->pdata.size);
 801 }
 802
 803 /**
 804  * irdma_build_mpa_v2 - build a MPA V2 frame
 805  * @cm_node: connection's node
 806  * @start_addr: buffer start address
 807  * @mpa_key: to do read0 or write0
 808  */
 809 static void irdma_build_mpa_v2(struct irdma_cm_node *cm_node, void *start_addr,
 810                                u8 mpa_key)
 811 {
 812         struct ietf_mpa_v2 *mpa_frame = start_addr;
 813         struct ietf_rtr_msg *rtr_msg = &mpa_frame->rtr_msg;
 814         u16 ctrl_ird, ctrl_ord;
 815
 816         /* initialize the upper 5 bytes of the frame */
 817         irdma_build_mpa_v1(cm_node, start_addr, mpa_key);
 818         mpa_frame->flags |= IETF_MPA_V2_FLAG;
 819         if (cm_node->iwdev->iw_ooo) {
 820                 mpa_frame->flags |= IETF_MPA_FLAGS_MARKERS;
 821                 cm_node->rcv_mark_en = true;
 822         }
 823         mpa_frame->priv_data_len = cpu_to_be16(be16_to_cpu(mpa_frame->priv_data_len) +
 824                                                IETF_RTR_MSG_SIZE);
 825
 826         /* initialize RTR msg */
 827         if (cm_node->mpav2_ird_ord == IETF_NO_IRD_ORD) {
 828                 ctrl_ird = IETF_NO_IRD_ORD;
 829                 ctrl_ord = IETF_NO_IRD_ORD;
 830         } else {
 831                 ctrl_ird = (cm_node->ird_size > IETF_NO_IRD_ORD) ?
 832                                    IETF_NO_IRD_ORD :
 833                                    cm_node->ird_size;
 834                 ctrl_ord = (cm_node->ord_size > IETF_NO_IRD_ORD) ?
 835                                    IETF_NO_IRD_ORD :
 836                                    cm_node->ord_size;
 837         }
 838         ctrl_ird |= IETF_PEER_TO_PEER;
 839
 840         switch (mpa_key) {
 841         case MPA_KEY_REQUEST:
 842                 ctrl_ord |= IETF_RDMA0_WRITE;
 843                 ctrl_ord |= IETF_RDMA0_READ;
 844                 break;
 845         case MPA_KEY_REPLY:
 846                 switch (cm_node->send_rdma0_op) {
 847                 case SEND_RDMA_WRITE_ZERO:
 848                         ctrl_ord |= IETF_RDMA0_WRITE;
 849                         break;
 850                 case SEND_RDMA_READ_ZERO:
 851                         ctrl_ord |= IETF_RDMA0_READ;
 852                         break;
 853                 }
 854                 break;
 855         default:
 856                 break;
 857         }
 858         rtr_msg->ctrl_ird = htons(ctrl_ird);
 859         rtr_msg->ctrl_ord = htons(ctrl_ord);
 860 }
 861
 862 /**
 863  * irdma_cm_build_mpa_frame - build mpa frame for mpa version 1 or version 2
 864  * @cm_node: connection's node
 865  * @mpa: mpa: data buffer
 866  * @mpa_key: to do read0 or write0
 867  */
 868 static int irdma_cm_build_mpa_frame(struct irdma_cm_node *cm_node,
 869                                     struct irdma_kmem_info *mpa, u8 mpa_key)
 870 {
 871         int hdr_len = 0;
 872
 873         switch (cm_node->mpa_frame_rev) {
 874         case IETF_MPA_V1:
 875                 hdr_len = sizeof(struct ietf_mpa_v1);
 876                 irdma_build_mpa_v1(cm_node, mpa->addr, mpa_key);
 877                 break;
 878         case IETF_MPA_V2:
 879                 hdr_len = sizeof(struct ietf_mpa_v2);
 880                 irdma_build_mpa_v2(cm_node, mpa->addr, mpa_key);
 881                 break;
 882         default:
 883                 break;
 884         }
 885
 886         return hdr_len;
 887 }
 888
 889 /**
 890  * irdma_send_mpa_request - active node send mpa request to passive node
 891  * @cm_node: connection's node
 892  */
 893 static int irdma_send_mpa_request(struct irdma_cm_node *cm_node)
 894 {
 895         struct irdma_puda_buf *sqbuf;
 896
 897         cm_node->mpa_hdr.addr = &cm_node->mpa_v2_frame;
 898         cm_node->mpa_hdr.size = irdma_cm_build_mpa_frame(cm_node,
 899                                                          &cm_node->mpa_hdr,
 900                                                          MPA_KEY_REQUEST);
 901         if (!cm_node->mpa_hdr.size) {
 902                 ibdev_dbg(&cm_node->iwdev->ibdev,
 903                           "CM: mpa size = %d\n", cm_node->mpa_hdr.size);
 904                 return -EINVAL;
 905         }
 906
 907         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL,
 908                                                 &cm_node->mpa_hdr,
 909                                                 &cm_node->pdata, SET_ACK);
 910         if (!sqbuf)
 911                 return -ENOMEM;
 912
 913         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
 914                                        0);
 915 }
 916
 917 /**
 918  * irdma_send_mpa_reject -
 919  * @cm_node: connection's node
 920  * @pdata: reject data for connection
 921  * @plen: length of reject data
 922  */
 923 static int irdma_send_mpa_reject(struct irdma_cm_node *cm_node,
 924                                  const void *pdata, u8 plen)
 925 {
 926         struct irdma_puda_buf *sqbuf;
 927         struct irdma_mpa_priv_info priv_info;
 928
 929         cm_node->mpa_hdr.addr = &cm_node->mpa_v2_frame;
 930         cm_node->mpa_hdr.size = irdma_cm_build_mpa_frame(cm_node,
 931                                                          &cm_node->mpa_hdr,
 932                                                          MPA_KEY_REPLY);
 933
 934         cm_node->mpa_frame.flags |= IETF_MPA_FLAGS_REJECT;
 935         priv_info.addr = pdata;
 936         priv_info.size = plen;
 937
 938         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL,
 939                                                 &cm_node->mpa_hdr, &priv_info,
 940                                                 SET_ACK | SET_FIN);
 941         if (!sqbuf)
 942                 return -ENOMEM;
 943
 944         cm_node->state = IRDMA_CM_STATE_FIN_WAIT1;
 945
 946         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
 947                                        0);
 948 }
 949
 950 /**
 951  * irdma_negotiate_mpa_v2_ird_ord - negotiate MPAv2 IRD/ORD
 952  * @cm_node: connection's node
 953  * @buf: Data pointer
 954  */
 955 static int irdma_negotiate_mpa_v2_ird_ord(struct irdma_cm_node *cm_node,
 956                                           u8 *buf)
 957 {
 958         struct ietf_mpa_v2 *mpa_v2_frame;
 959         struct ietf_rtr_msg *rtr_msg;
 960         u16 ird_size;
 961         u16 ord_size;
 962         u16 ctrl_ord;
 963         u16 ctrl_ird;
 964
 965         mpa_v2_frame = (struct ietf_mpa_v2 *)buf;
 966         rtr_msg = &mpa_v2_frame->rtr_msg;
 967
 968         /* parse rtr message */
 969         ctrl_ord = ntohs(rtr_msg->ctrl_ord);
 970         ctrl_ird = ntohs(rtr_msg->ctrl_ird);
 971         ird_size = ctrl_ird & IETF_NO_IRD_ORD;
 972         ord_size = ctrl_ord & IETF_NO_IRD_ORD;
 973
 974         if (!(ctrl_ird & IETF_PEER_TO_PEER))
 975                 return -EOPNOTSUPP;
 976
 977         if (ird_size == IETF_NO_IRD_ORD || ord_size == IETF_NO_IRD_ORD) {
 978                 cm_node->mpav2_ird_ord = IETF_NO_IRD_ORD;
 979                 goto negotiate_done;
 980         }
 981
 982         if (cm_node->state != IRDMA_CM_STATE_MPAREQ_SENT) {
 983                 /* responder */
 984                 if (!ord_size && (ctrl_ord & IETF_RDMA0_READ))
 985                         cm_node->ird_size = 1;
 986                 if (cm_node->ord_size > ird_size)
 987                         cm_node->ord_size = ird_size;
 988         } else {
 989                 /* initiator */
 990                 if (!ird_size && (ctrl_ord & IETF_RDMA0_READ))
 991                         /* Remote peer doesn't support RDMA0_READ */
 992                         return -EOPNOTSUPP;
 993
 994                 if (cm_node->ord_size > ird_size)
 995                         cm_node->ord_size = ird_size;
 996
 997                 if (cm_node->ird_size < ord_size)
 998                 /* no resources available */
 999                         return -EINVAL;
1000         }
1001
1002 negotiate_done:
1003         if (ctrl_ord & IETF_RDMA0_READ)
1004                 cm_node->send_rdma0_op = SEND_RDMA_READ_ZERO;
1005         else if (ctrl_ord & IETF_RDMA0_WRITE)
1006                 cm_node->send_rdma0_op = SEND_RDMA_WRITE_ZERO;
1007         else
1008                 /* Not supported RDMA0 operation */
1009                 return -EOPNOTSUPP;
1010
1011         ibdev_dbg(&cm_node->iwdev->ibdev,
1012                   "CM: MPAV2 Negotiated ORD: %d, IRD: %d\n",
1013                   cm_node->ord_size, cm_node->ird_size);
1014         trace_irdma_negotiate_mpa_v2(cm_node);
1015         return 0;
1016 }
1017
1018 /**
1019  * irdma_parse_mpa - process an IETF MPA frame
1020  * @cm_node: connection's node
1021  * @buf: Data pointer
1022  * @type: to return accept or reject
1023  * @len: Len of mpa buffer
1024  */
1025 static int irdma_parse_mpa(struct irdma_cm_node *cm_node, u8 *buf, u32 *type,
1026                            u32 len)
1027 {
1028         struct ietf_mpa_v1 *mpa_frame;
1029         int mpa_hdr_len, priv_data_len, ret;
1030
1031         *type = IRDMA_MPA_REQUEST_ACCEPT;
1032
1033         if (len < sizeof(struct ietf_mpa_v1)) {
1034                 ibdev_dbg(&cm_node->iwdev->ibdev,
1035                           "CM: ietf buffer small (%x)\n", len);
1036                 return -EINVAL;
1037         }
1038
1039         mpa_frame = (struct ietf_mpa_v1 *)buf;
1040         mpa_hdr_len = sizeof(struct ietf_mpa_v1);
1041         priv_data_len = ntohs(mpa_frame->priv_data_len);
1042
1043         if (priv_data_len > IETF_MAX_PRIV_DATA_LEN) {
1044                 ibdev_dbg(&cm_node->iwdev->ibdev,
1045                           "CM: private_data too big %d\n", priv_data_len);
1046                 return -EOVERFLOW;
1047         }
1048
1049         if (mpa_frame->rev != IETF_MPA_V1 && mpa_frame->rev != IETF_MPA_V2) {
1050                 ibdev_dbg(&cm_node->iwdev->ibdev,
1051                           "CM: unsupported mpa rev = %d\n", mpa_frame->rev);
1052                 return -EINVAL;
1053         }
1054
1055         if (mpa_frame->rev > cm_node->mpa_frame_rev) {
1056                 ibdev_dbg(&cm_node->iwdev->ibdev, "CM: rev %d\n",
1057                           mpa_frame->rev);
1058                 return -EINVAL;
1059         }
1060
1061         cm_node->mpa_frame_rev = mpa_frame->rev;
1062         if (cm_node->state != IRDMA_CM_STATE_MPAREQ_SENT) {
1063                 if (memcmp(mpa_frame->key, IEFT_MPA_KEY_REQ,
1064                            IETF_MPA_KEY_SIZE)) {
1065                         ibdev_dbg(&cm_node->iwdev->ibdev,
1066                                   "CM: Unexpected MPA Key received\n");
1067                         return -EINVAL;
1068                 }
1069         } else {
1070                 if (memcmp(mpa_frame->key, IEFT_MPA_KEY_REP,
1071                            IETF_MPA_KEY_SIZE)) {
1072                         ibdev_dbg(&cm_node->iwdev->ibdev,
1073                                   "CM: Unexpected MPA Key received\n");
1074                         return -EINVAL;
1075                 }
1076         }
1077
1078         if (priv_data_len + mpa_hdr_len > len) {
1079                 ibdev_dbg(&cm_node->iwdev->ibdev,
1080                           "CM: ietf buffer len(%x + %x != %x)\n",
1081                           priv_data_len, mpa_hdr_len, len);
1082                 return -EOVERFLOW;
1083         }
1084
1085         if (len > IRDMA_MAX_CM_BUF) {
1086                 ibdev_dbg(&cm_node->iwdev->ibdev,
1087                           "CM: ietf buffer large len = %d\n", len);
1088                 return -EOVERFLOW;
1089         }
1090
1091         switch (mpa_frame->rev) {
1092         case IETF_MPA_V2:
1093                 mpa_hdr_len += IETF_RTR_MSG_SIZE;
1094                 ret = irdma_negotiate_mpa_v2_ird_ord(cm_node, buf);
1095                 if (ret)
1096                         return ret;
1097                 break;
1098         case IETF_MPA_V1:
1099         default:
1100                 break;
1101         }
1102
1103         memcpy(cm_node->pdata_buf, buf + mpa_hdr_len, priv_data_len);
1104         cm_node->pdata.size = priv_data_len;
1105
1106         if (mpa_frame->flags & IETF_MPA_FLAGS_REJECT)
1107                 *type = IRDMA_MPA_REQUEST_REJECT;
1108
1109         if (mpa_frame->flags & IETF_MPA_FLAGS_MARKERS)
1110                 cm_node->snd_mark_en = true;
1111
1112         return 0;
1113 }
1114
1115 /**
1116  * irdma_schedule_cm_timer
1117  * @cm_node: connection's node
1118  * @sqbuf: buffer to send
1119  * @type: if it is send or close
1120  * @send_retrans: if rexmits to be done
1121  * @close_when_complete: is cm_node to be removed
1122  *
1123  * note - cm_node needs to be protected before calling this. Encase in:
1124  *              irdma_rem_ref_cm_node(cm_core, cm_node);
1125  *              irdma_schedule_cm_timer(...)
1126  *              refcount_inc(&cm_node->refcnt);
1127  */
1128 int irdma_schedule_cm_timer(struct irdma_cm_node *cm_node,
1129                             struct irdma_puda_buf *sqbuf,
1130                             enum irdma_timer_type type, int send_retrans,
1131                             int close_when_complete)
1132 {
1133         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
1134         struct irdma_cm_core *cm_core = cm_node->cm_core;
1135         struct irdma_timer_entry *new_send;
1136         u32 was_timer_set;
1137         unsigned long flags;
1138
1139         new_send = kzalloc(sizeof(*new_send), GFP_ATOMIC);
1140         if (!new_send) {
1141                 if (type != IRDMA_TIMER_TYPE_CLOSE)
1142                         irdma_free_sqbuf(vsi, sqbuf);
1143                 return -ENOMEM;
1144         }
1145
1146         new_send->retrycount = IRDMA_DEFAULT_RETRYS;
1147         new_send->retranscount = IRDMA_DEFAULT_RETRANS;
1148         new_send->sqbuf = sqbuf;
1149         new_send->timetosend = jiffies;
1150         new_send->type = type;
1151         new_send->send_retrans = send_retrans;
1152         new_send->close_when_complete = close_when_complete;
1153
1154         if (type == IRDMA_TIMER_TYPE_CLOSE) {
1155                 new_send->timetosend += (HZ / 10);
1156                 if (cm_node->close_entry) {
1157                         kfree(new_send);
1158                         ibdev_dbg(&cm_node->iwdev->ibdev,
1159                                   "CM: already close entry\n");
1160                         return -EINVAL;
1161                 }
1162
1163                 cm_node->close_entry = new_send;
1164         } else { /* type == IRDMA_TIMER_TYPE_SEND */
1165                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1166                 cm_node->send_entry = new_send;
1167                 refcount_inc(&cm_node->refcnt);
1168                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1169                 new_send->timetosend = jiffies + IRDMA_RETRY_TIMEOUT;
1170
1171                 refcount_inc(&sqbuf->refcount);
1172                 irdma_puda_send_buf(vsi->ilq, sqbuf);
1173                 if (!send_retrans) {
1174                         irdma_cleanup_retrans_entry(cm_node);
1175                         if (close_when_complete)
1176                                 irdma_rem_ref_cm_node(cm_node);
1177                         return 0;
1178                 }
1179         }
1180
1181         spin_lock_irqsave(&cm_core->ht_lock, flags);
1182         was_timer_set = timer_pending(&cm_core->tcp_timer);
1183
1184         if (!was_timer_set) {
1185                 cm_core->tcp_timer.expires = new_send->timetosend;
1186                 add_timer(&cm_core->tcp_timer);
1187         }
1188         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
1189
1190         return 0;
1191 }
1192
1193 /**
1194  * irdma_retrans_expired - Could not rexmit the packet
1195  * @cm_node: connection's node
1196  */
1197 static void irdma_retrans_expired(struct irdma_cm_node *cm_node)
1198 {
1199         enum irdma_cm_node_state state = cm_node->state;
1200
1201         cm_node->state = IRDMA_CM_STATE_CLOSED;
1202         switch (state) {
1203         case IRDMA_CM_STATE_SYN_RCVD:
1204         case IRDMA_CM_STATE_CLOSING:
1205                 irdma_rem_ref_cm_node(cm_node);
1206                 break;
1207         case IRDMA_CM_STATE_FIN_WAIT1:
1208         case IRDMA_CM_STATE_LAST_ACK:
1209                 irdma_send_reset(cm_node);
1210                 break;
1211         default:
1212                 refcount_inc(&cm_node->refcnt);
1213                 irdma_send_reset(cm_node);
1214                 irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
1215                 break;
1216         }
1217 }
1218
1219 /**
1220  * irdma_handle_close_entry - for handling retry/timeouts
1221  * @cm_node: connection's node
1222  * @rem_node: flag for remove cm_node
1223  */
1224 static void irdma_handle_close_entry(struct irdma_cm_node *cm_node,
1225                                      u32 rem_node)
1226 {
1227         struct irdma_timer_entry *close_entry = cm_node->close_entry;
1228         struct irdma_qp *iwqp;
1229         unsigned long flags;
1230
1231         if (!close_entry)
1232                 return;
1233         iwqp = (struct irdma_qp *)close_entry->sqbuf;
1234         if (iwqp) {
1235                 spin_lock_irqsave(&iwqp->lock, flags);
1236                 if (iwqp->cm_id) {
1237                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1238                         iwqp->hw_iwarp_state = IRDMA_QP_STATE_ERROR;
1239                         iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1240                         iwqp->ibqp_state = IB_QPS_ERR;
1241                         spin_unlock_irqrestore(&iwqp->lock, flags);
1242                         irdma_cm_disconn(iwqp);
1243                 } else {
1244                         spin_unlock_irqrestore(&iwqp->lock, flags);
1245                 }
1246         } else if (rem_node) {
1247                 /* TIME_WAIT state */
1248                 irdma_rem_ref_cm_node(cm_node);
1249         }
1250
1251         kfree(close_entry);
1252         cm_node->close_entry = NULL;
1253 }
1254
1255 /**
1256  * irdma_cm_timer_tick - system's timer expired callback
1257  * @t: Pointer to timer_list
1258  */
1259 static void irdma_cm_timer_tick(struct timer_list *t)
1260 {
1261         unsigned long nexttimeout = jiffies + IRDMA_LONG_TIME;
1262         struct irdma_cm_node *cm_node;
1263         struct irdma_timer_entry *send_entry, *close_entry;
1264         struct list_head *list_core_temp;
1265         struct list_head *list_node;
1266         struct irdma_cm_core *cm_core = from_timer(cm_core, t, tcp_timer);
1267         struct irdma_sc_vsi *vsi;
1268         u32 settimer = 0;
1269         unsigned long timetosend;
1270         unsigned long flags;
1271         struct list_head timer_list;
1272
1273         INIT_LIST_HEAD(&timer_list);
1274
1275         rcu_read_lock();
1276         irdma_timer_list_prep(cm_core, &timer_list);
1277         rcu_read_unlock();
1278
1279         list_for_each_safe (list_node, list_core_temp, &timer_list) {
1280                 cm_node = container_of(list_node, struct irdma_cm_node,
1281                                        timer_entry);
1282                 close_entry = cm_node->close_entry;
1283
1284                 if (close_entry) {
1285                         if (time_after(close_entry->timetosend, jiffies)) {
1286                                 if (nexttimeout > close_entry->timetosend ||
1287                                     !settimer) {
1288                                         nexttimeout = close_entry->timetosend;
1289                                         settimer = 1;
1290                                 }
1291                         } else {
1292                                 irdma_handle_close_entry(cm_node, 1);
1293                         }
1294                 }
1295
1296                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1297
1298                 send_entry = cm_node->send_entry;
1299                 if (!send_entry)
1300                         goto done;
1301                 if (time_after(send_entry->timetosend, jiffies)) {
1302                         if (cm_node->state != IRDMA_CM_STATE_OFFLOADED) {
1303                                 if (nexttimeout > send_entry->timetosend ||
1304                                     !settimer) {
1305                                         nexttimeout = send_entry->timetosend;
1306                                         settimer = 1;
1307                                 }
1308                         } else {
1309                                 irdma_free_retrans_entry(cm_node);
1310                         }
1311                         goto done;
1312                 }
1313
1314                 if (cm_node->state == IRDMA_CM_STATE_OFFLOADED ||
1315                     cm_node->state == IRDMA_CM_STATE_CLOSED) {
1316                         irdma_free_retrans_entry(cm_node);
1317                         goto done;
1318                 }
1319
1320                 if (!send_entry->retranscount || !send_entry->retrycount) {
1321                         irdma_free_retrans_entry(cm_node);
1322
1323                         spin_unlock_irqrestore(&cm_node->retrans_list_lock,
1324                                                flags);
1325                         irdma_retrans_expired(cm_node);
1326                         cm_node->state = IRDMA_CM_STATE_CLOSED;
1327                         spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1328                         goto done;
1329                 }
1330                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1331
1332                 vsi = &cm_node->iwdev->vsi;
1333                 if (!cm_node->ack_rcvd) {
1334                         refcount_inc(&send_entry->sqbuf->refcount);
1335                         irdma_puda_send_buf(vsi->ilq, send_entry->sqbuf);
1336                         cm_node->cm_core->stats_pkt_retrans++;
1337                 }
1338
1339                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1340                 if (send_entry->send_retrans) {
1341                         send_entry->retranscount--;
1342                         timetosend = (IRDMA_RETRY_TIMEOUT <<
1343                                       (IRDMA_DEFAULT_RETRANS -
1344                                        send_entry->retranscount));
1345
1346                         send_entry->timetosend = jiffies +
1347                             min(timetosend, IRDMA_MAX_TIMEOUT);
1348                         if (nexttimeout > send_entry->timetosend || !settimer) {
1349                                 nexttimeout = send_entry->timetosend;
1350                                 settimer = 1;
1351                         }
1352                 } else {
1353                         int close_when_complete;
1354
1355                         close_when_complete = send_entry->close_when_complete;
1356                         irdma_free_retrans_entry(cm_node);
1357                         if (close_when_complete)
1358                                 irdma_rem_ref_cm_node(cm_node);
1359                 }
1360 done:
1361                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1362                 irdma_rem_ref_cm_node(cm_node);
1363         }
1364
1365         if (settimer) {
1366                 spin_lock_irqsave(&cm_core->ht_lock, flags);
1367                 if (!timer_pending(&cm_core->tcp_timer)) {
1368                         cm_core->tcp_timer.expires = nexttimeout;
1369                         add_timer(&cm_core->tcp_timer);
1370                 }
1371                 spin_unlock_irqrestore(&cm_core->ht_lock, flags);
1372         }
1373 }
1374
1375 /**
1376  * irdma_send_syn - send SYN packet
1377  * @cm_node: connection's node
1378  * @sendack: flag to set ACK bit or not
1379  */
1380 int irdma_send_syn(struct irdma_cm_node *cm_node, u32 sendack)
1381 {
1382         struct irdma_puda_buf *sqbuf;
1383         int flags = SET_SYN;
1384         char optionsbuf[sizeof(struct option_mss) +
1385                         sizeof(struct option_windowscale) +
1386                         sizeof(struct option_base) + TCP_OPTIONS_PADDING];
1387         struct irdma_kmem_info opts;
1388         int optionssize = 0;
1389         /* Sending MSS option */
1390         union all_known_options *options;
1391
1392         opts.addr = optionsbuf;
1393         if (!cm_node)
1394                 return -EINVAL;
1395
1396         options = (union all_known_options *)&optionsbuf[optionssize];
1397         options->mss.optionnum = OPTION_NUM_MSS;
1398         options->mss.len = sizeof(struct option_mss);
1399         options->mss.mss = htons(cm_node->tcp_cntxt.mss);
1400         optionssize += sizeof(struct option_mss);
1401
1402         options = (union all_known_options *)&optionsbuf[optionssize];
1403         options->windowscale.optionnum = OPTION_NUM_WINDOW_SCALE;
1404         options->windowscale.len = sizeof(struct option_windowscale);
1405         options->windowscale.shiftcount = cm_node->tcp_cntxt.rcv_wscale;
1406         optionssize += sizeof(struct option_windowscale);
1407         options = (union all_known_options *)&optionsbuf[optionssize];
1408         options->eol = OPTION_NUM_EOL;
1409         optionssize += 1;
1410
1411         if (sendack)
1412                 flags |= SET_ACK;
1413
1414         opts.size = optionssize;
1415
1416         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, &opts, NULL, NULL,
1417                                                 flags);
1418         if (!sqbuf)
1419                 return -ENOMEM;
1420
1421         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
1422                                        0);
1423 }
1424
1425 /**
1426  * irdma_send_ack - Send ACK packet
1427  * @cm_node: connection's node
1428  */
1429 void irdma_send_ack(struct irdma_cm_node *cm_node)
1430 {
1431         struct irdma_puda_buf *sqbuf;
1432         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
1433
1434         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
1435                                                 SET_ACK);
1436         if (sqbuf)
1437                 irdma_puda_send_buf(vsi->ilq, sqbuf);
1438 }
1439
1440 /**
1441  * irdma_send_fin - Send FIN pkt
1442  * @cm_node: connection's node
1443  */
1444 static int irdma_send_fin(struct irdma_cm_node *cm_node)
1445 {
1446         struct irdma_puda_buf *sqbuf;
1447
1448         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
1449                                                 SET_ACK | SET_FIN);
1450         if (!sqbuf)
1451                 return -ENOMEM;
1452
1453         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
1454                                        0);
1455 }
1456
1457 /**
1458  * irdma_find_listener - find a cm node listening on this addr-port pair
1459  * @cm_core: cm's core
1460  * @dst_addr: listener ip addr
1461  * @dst_port: listener tcp port num
1462  * @vlan_id: virtual LAN ID
1463  * @listener_state: state to match with listen node's
1464  */
1465 static struct irdma_cm_listener *
1466 irdma_find_listener(struct irdma_cm_core *cm_core, u32 *dst_addr, u16 dst_port,
1467                     u16 vlan_id, enum irdma_cm_listener_state listener_state)
1468 {
1469         struct irdma_cm_listener *listen_node;
1470         static const u32 ip_zero[4] = { 0, 0, 0, 0 };
1471         u32 listen_addr[4];
1472         u16 listen_port;
1473         unsigned long flags;
1474
1475         /* walk list and find cm_node associated with this session ID */
1476         spin_lock_irqsave(&cm_core->listen_list_lock, flags);
1477         list_for_each_entry (listen_node, &cm_core->listen_list, list) {
1478                 memcpy(listen_addr, listen_node->loc_addr, sizeof(listen_addr));
1479                 listen_port = listen_node->loc_port;
1480                 /* compare node pair, return node handle if a match */
1481                 if ((!memcmp(listen_addr, dst_addr, sizeof(listen_addr)) ||
1482                      !memcmp(listen_addr, ip_zero, sizeof(listen_addr))) &&
1483                     listen_port == dst_port &&
1484                     vlan_id == listen_node->vlan_id &&
1485                     (listener_state & listen_node->listener_state)) {
1486                         refcount_inc(&listen_node->refcnt);
1487                         spin_unlock_irqrestore(&cm_core->listen_list_lock,
1488                                                flags);
1489                         trace_irdma_find_listener(listen_node);
1490                         return listen_node;
1491                 }
1492         }
1493         spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
1494
1495         return NULL;
1496 }
1497
1498 /**
1499  * irdma_del_multiple_qhash - Remove qhash and child listens
1500  * @iwdev: iWarp device
1501  * @cm_info: CM info for parent listen node
1502  * @cm_parent_listen_node: The parent listen node
1503  */
1504 static int irdma_del_multiple_qhash(struct irdma_device *iwdev,
1505                                     struct irdma_cm_info *cm_info,
1506                                     struct irdma_cm_listener *cm_parent_listen_node)
1507 {
1508         struct irdma_cm_listener *child_listen_node;
1509         struct list_head *pos, *tpos;
1510         unsigned long flags;
1511         int ret = -EINVAL;
1512
1513         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock, flags);
1514         list_for_each_safe (pos, tpos,
1515                             &cm_parent_listen_node->child_listen_list) {
1516                 child_listen_node = list_entry(pos, struct irdma_cm_listener,
1517                                                child_listen_list);
1518                 if (child_listen_node->ipv4)
1519                         ibdev_dbg(&iwdev->ibdev,
1520                                   "CM: removing child listen for IP=%pI4, port=%d, vlan=%d\n",
1521                                   child_listen_node->loc_addr,
1522                                   child_listen_node->loc_port,
1523                                   child_listen_node->vlan_id);
1524                 else
1525                         ibdev_dbg(&iwdev->ibdev,
1526                                   "CM: removing child listen for IP=%pI6, port=%d, vlan=%d\n",
1527                                   child_listen_node->loc_addr,
1528                                   child_listen_node->loc_port,
1529                                   child_listen_node->vlan_id);
1530                 trace_irdma_del_multiple_qhash(child_listen_node);
1531                 list_del(pos);
1532                 memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1533                        sizeof(cm_info->loc_addr));
1534                 cm_info->vlan_id = child_listen_node->vlan_id;
1535                 if (child_listen_node->qhash_set) {
1536                         ret = irdma_manage_qhash(iwdev, cm_info,
1537                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1538                                                  IRDMA_QHASH_MANAGE_TYPE_DELETE,
1539                                                  NULL, false);
1540                         child_listen_node->qhash_set = false;
1541                 } else {
1542                         ret = 0;
1543                 }
1544                 ibdev_dbg(&iwdev->ibdev,
1545                           "CM: Child listen node freed = %p\n",
1546                           child_listen_node);
1547                 kfree(child_listen_node);
1548                 cm_parent_listen_node->cm_core->stats_listen_nodes_destroyed++;
1549         }
1550         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1551
1552         return ret;
1553 }
1554
1555 /**
1556  * irdma_netdev_vlan_ipv6 - Gets the netdev and mac
1557  * @addr: local IPv6 address
1558  * @vlan_id: vlan id for the given IPv6 address
1559  * @mac: mac address for the given IPv6 address
1560  *
1561  * Returns the net_device of the IPv6 address and also sets the
1562  * vlan id and mac for that address.
1563  */
1564 struct net_device *irdma_netdev_vlan_ipv6(u32 *addr, u16 *vlan_id, u8 *mac)
1565 {
1566         struct net_device *ip_dev = NULL;
1567         struct in6_addr laddr6;
1568
1569         if (!IS_ENABLED(CONFIG_IPV6))
1570                 return NULL;
1571
1572         irdma_copy_ip_htonl(laddr6.in6_u.u6_addr32, addr);
1573         if (vlan_id)
1574                 *vlan_id = 0xFFFF;      /* Match rdma_vlan_dev_vlan_id() */
1575         if (mac)
1576                 eth_zero_addr(mac);
1577
1578         rcu_read_lock();
1579         for_each_netdev_rcu (&init_net, ip_dev) {
1580                 if (ipv6_chk_addr(&init_net, &laddr6, ip_dev, 1)) {
1581                         if (vlan_id)
1582                                 *vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1583                         if (ip_dev->dev_addr && mac)
1584                                 ether_addr_copy(mac, ip_dev->dev_addr);
1585                         break;
1586                 }
1587         }
1588         rcu_read_unlock();
1589
1590         return ip_dev;
1591 }
1592
1593 /**
1594  * irdma_get_vlan_ipv4 - Returns the vlan_id for IPv4 address
1595  * @addr: local IPv4 address
1596  */
1597 u16 irdma_get_vlan_ipv4(u32 *addr)
1598 {
1599         struct net_device *netdev;
1600         u16 vlan_id = 0xFFFF;
1601
1602         netdev = ip_dev_find(&init_net, htonl(addr[0]));
1603         if (netdev) {
1604                 vlan_id = rdma_vlan_dev_vlan_id(netdev);
1605                 dev_put(netdev);
1606         }
1607
1608         return vlan_id;
1609 }
1610
1611 /**
1612  * irdma_add_mqh_6 - Adds multiple qhashes for IPv6
1613  * @iwdev: iWarp device
1614  * @cm_info: CM info for parent listen node
1615  * @cm_parent_listen_node: The parent listen node
1616  *
1617  * Adds a qhash and a child listen node for every IPv6 address
1618  * on the adapter and adds the associated qhash filter
1619  */
1620 static int irdma_add_mqh_6(struct irdma_device *iwdev,
1621                            struct irdma_cm_info *cm_info,
1622                            struct irdma_cm_listener *cm_parent_listen_node)
1623 {
1624         struct net_device *ip_dev;
1625         struct inet6_dev *idev;
1626         struct inet6_ifaddr *ifp, *tmp;
1627         struct irdma_cm_listener *child_listen_node;
1628         unsigned long flags;
1629         int ret = 0;
1630
1631         rtnl_lock();
1632         for_each_netdev(&init_net, ip_dev) {
1633                 if (!(ip_dev->flags & IFF_UP))
1634                         continue;
1635
1636                 if (((rdma_vlan_dev_vlan_id(ip_dev) >= VLAN_N_VID) ||
1637                      (rdma_vlan_dev_real_dev(ip_dev) != iwdev->netdev)) &&
1638                     ip_dev != iwdev->netdev)
1639                         continue;
1640
1641                 idev = __in6_dev_get(ip_dev);
1642                 if (!idev) {
1643                         ibdev_dbg(&iwdev->ibdev, "CM: idev == NULL\n");
1644                         break;
1645                 }
1646                 list_for_each_entry_safe (ifp, tmp, &idev->addr_list, if_list) {
1647                         ibdev_dbg(&iwdev->ibdev, "CM: IP=%pI6, vlan_id=%d, MAC=%pM\n",
1648                                   &ifp->addr, rdma_vlan_dev_vlan_id(ip_dev),
1649                                   ip_dev->dev_addr);
1650                         child_listen_node = kzalloc(sizeof(*child_listen_node), GFP_KERNEL);
1651                         ibdev_dbg(&iwdev->ibdev, "CM: Allocating child listener %p\n",
1652                                   child_listen_node);
1653                         if (!child_listen_node) {
1654                                 ibdev_dbg(&iwdev->ibdev, "CM: listener memory allocation\n");
1655                                 ret = -ENOMEM;
1656                                 goto exit;
1657                         }
1658
1659                         cm_info->vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1660                         cm_parent_listen_node->vlan_id = cm_info->vlan_id;
1661                         memcpy(child_listen_node, cm_parent_listen_node,
1662                                sizeof(*child_listen_node));
1663                         irdma_copy_ip_ntohl(child_listen_node->loc_addr,
1664                                             ifp->addr.in6_u.u6_addr32);
1665                         memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1666                                sizeof(cm_info->loc_addr));
1667                         ret = irdma_manage_qhash(iwdev, cm_info,
1668                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1669                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
1670                                                  NULL, true);
1671                         if (ret) {
1672                                 kfree(child_listen_node);
1673                                 continue;
1674                         }
1675
1676                         trace_irdma_add_mqh_6(iwdev, child_listen_node,
1677                                               ip_dev->dev_addr);
1678
1679                         child_listen_node->qhash_set = true;
1680                         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock, flags);
1681                         list_add(&child_listen_node->child_listen_list,
1682                                  &cm_parent_listen_node->child_listen_list);
1683                         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1684                         cm_parent_listen_node->cm_core->stats_listen_nodes_created++;
1685                 }
1686         }
1687 exit:
1688         rtnl_unlock();
1689
1690         return ret;
1691 }
1692
1693 /**
1694  * irdma_add_mqh_4 - Adds multiple qhashes for IPv4
1695  * @iwdev: iWarp device
1696  * @cm_info: CM info for parent listen node
1697  * @cm_parent_listen_node: The parent listen node
1698  *
1699  * Adds a qhash and a child listen node for every IPv4 address
1700  * on the adapter and adds the associated qhash filter
1701  */
1702 static int irdma_add_mqh_4(struct irdma_device *iwdev,
1703                            struct irdma_cm_info *cm_info,
1704                            struct irdma_cm_listener *cm_parent_listen_node)
1705 {
1706         struct net_device *ip_dev;
1707         struct in_device *idev;
1708         struct irdma_cm_listener *child_listen_node;
1709         unsigned long flags;
1710         const struct in_ifaddr *ifa;
1711         int ret = 0;
1712
1713         rtnl_lock();
1714         for_each_netdev(&init_net, ip_dev) {
1715                 if (!(ip_dev->flags & IFF_UP))
1716                         continue;
1717
1718                 if (((rdma_vlan_dev_vlan_id(ip_dev) >= VLAN_N_VID) ||
1719                      (rdma_vlan_dev_real_dev(ip_dev) != iwdev->netdev)) &&
1720                     ip_dev != iwdev->netdev)
1721                         continue;
1722
1723                 idev = in_dev_get(ip_dev);
1724                 in_dev_for_each_ifa_rtnl(ifa, idev) {
1725                         ibdev_dbg(&iwdev->ibdev,
1726                                   "CM: Allocating child CM Listener forIP=%pI4, vlan_id=%d, MAC=%pM\n",
1727                                   &ifa->ifa_address, rdma_vlan_dev_vlan_id(ip_dev),
1728                                   ip_dev->dev_addr);
1729                         child_listen_node = kzalloc(sizeof(*child_listen_node), GFP_KERNEL);
1730                         cm_parent_listen_node->cm_core->stats_listen_nodes_created++;
1731                         ibdev_dbg(&iwdev->ibdev, "CM: Allocating child listener %p\n",
1732                                   child_listen_node);
1733                         if (!child_listen_node) {
1734                                 ibdev_dbg(&iwdev->ibdev, "CM: listener memory allocation\n");
1735                                 in_dev_put(idev);
1736                                 ret = -ENOMEM;
1737                                 goto exit;
1738                         }
1739
1740                         cm_info->vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1741                         cm_parent_listen_node->vlan_id = cm_info->vlan_id;
1742                         memcpy(child_listen_node, cm_parent_listen_node,
1743                                sizeof(*child_listen_node));
1744                         child_listen_node->loc_addr[0] =
1745                                 ntohl(ifa->ifa_address);
1746                         memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1747                                sizeof(cm_info->loc_addr));
1748                         ret = irdma_manage_qhash(iwdev, cm_info,
1749                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1750                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
1751                                                  NULL, true);
1752                         if (ret) {
1753                                 kfree(child_listen_node);
1754                                 cm_parent_listen_node->cm_core
1755                                         ->stats_listen_nodes_created--;
1756                                 continue;
1757                         }
1758
1759                         trace_irdma_add_mqh_4(iwdev, child_listen_node,
1760                                               ip_dev->dev_addr);
1761
1762                         child_listen_node->qhash_set = true;
1763                         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock,
1764                                           flags);
1765                         list_add(&child_listen_node->child_listen_list,
1766                                  &cm_parent_listen_node->child_listen_list);
1767                         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1768                 }
1769                 in_dev_put(idev);
1770         }
1771 exit:
1772         rtnl_unlock();
1773
1774         return ret;
1775 }
1776
1777 /**
1778  * irdma_add_mqh - Adds multiple qhashes
1779  * @iwdev: iWarp device
1780  * @cm_info: CM info for parent listen node
1781  * @cm_listen_node: The parent listen node
1782  */
1783 static int irdma_add_mqh(struct irdma_device *iwdev,
1784                          struct irdma_cm_info *cm_info,
1785                          struct irdma_cm_listener *cm_listen_node)
1786 {
1787         if (cm_info->ipv4)
1788                 return irdma_add_mqh_4(iwdev, cm_info, cm_listen_node);
1789         else
1790                 return irdma_add_mqh_6(iwdev, cm_info, cm_listen_node);
1791 }
1792
1793 /**
1794  * irdma_reset_list_prep - add connection nodes slated for reset to list
1795  * @cm_core: cm's core
1796  * @listener: pointer to listener node
1797  * @reset_list: a list to which cm_node will be selected
1798  */
1799 static void irdma_reset_list_prep(struct irdma_cm_core *cm_core,
1800                                   struct irdma_cm_listener *listener,
1801                                   struct list_head *reset_list)
1802 {
1803         struct irdma_cm_node *cm_node;
1804         int bkt;
1805
1806         hash_for_each_rcu(cm_core->cm_hash_tbl, bkt, cm_node, list) {
1807                 if (cm_node->listener == listener &&
1808                     !cm_node->accelerated &&
1809                     refcount_inc_not_zero(&cm_node->refcnt))
1810                         list_add(&cm_node->reset_entry, reset_list);
1811         }
1812 }
1813
1814 /**
1815  * irdma_dec_refcnt_listen - delete listener and associated cm nodes
1816  * @cm_core: cm's core
1817  * @listener: pointer to listener node
1818  * @free_hanging_nodes: to free associated cm_nodes
1819  * @apbvt_del: flag to delete the apbvt
1820  */
1821 static int irdma_dec_refcnt_listen(struct irdma_cm_core *cm_core,
1822                                    struct irdma_cm_listener *listener,
1823                                    int free_hanging_nodes, bool apbvt_del)
1824 {
1825         int err;
1826         struct list_head *list_pos;
1827         struct list_head *list_temp;
1828         struct irdma_cm_node *cm_node;
1829         struct list_head reset_list;
1830         struct irdma_cm_info nfo;
1831         enum irdma_cm_node_state old_state;
1832         unsigned long flags;
1833
1834         trace_irdma_dec_refcnt_listen(listener, __builtin_return_address(0));
1835         /* free non-accelerated child nodes for this listener */
1836         INIT_LIST_HEAD(&reset_list);
1837         if (free_hanging_nodes) {
1838                 rcu_read_lock();
1839                 irdma_reset_list_prep(cm_core, listener, &reset_list);
1840                 rcu_read_unlock();
1841         }
1842
1843         list_for_each_safe (list_pos, list_temp, &reset_list) {
1844                 cm_node = container_of(list_pos, struct irdma_cm_node,
1845                                        reset_entry);
1846                 if (cm_node->state >= IRDMA_CM_STATE_FIN_WAIT1) {
1847                         irdma_rem_ref_cm_node(cm_node);
1848                         continue;
1849                 }
1850
1851                 irdma_cleanup_retrans_entry(cm_node);
1852                 err = irdma_send_reset(cm_node);
1853                 if (err) {
1854                         cm_node->state = IRDMA_CM_STATE_CLOSED;
1855                         ibdev_dbg(&cm_node->iwdev->ibdev,
1856                                   "CM: send reset failed\n");
1857                 } else {
1858                         old_state = cm_node->state;
1859                         cm_node->state = IRDMA_CM_STATE_LISTENER_DESTROYED;
1860                         if (old_state != IRDMA_CM_STATE_MPAREQ_RCVD)
1861                                 irdma_rem_ref_cm_node(cm_node);
1862                 }
1863         }
1864
1865         if (refcount_dec_and_test(&listener->refcnt)) {
1866                 spin_lock_irqsave(&cm_core->listen_list_lock, flags);
1867                 list_del(&listener->list);
1868                 spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
1869
1870                 if (apbvt_del)
1871                         irdma_del_apbvt(listener->iwdev,
1872                                         listener->apbvt_entry);
1873                 memcpy(nfo.loc_addr, listener->loc_addr, sizeof(nfo.loc_addr));
1874                 nfo.loc_port = listener->loc_port;
1875                 nfo.ipv4 = listener->ipv4;
1876                 nfo.vlan_id = listener->vlan_id;
1877                 nfo.user_pri = listener->user_pri;
1878                 nfo.qh_qpid = listener->iwdev->vsi.ilq->qp_id;
1879
1880                 if (!list_empty(&listener->child_listen_list)) {
1881                         irdma_del_multiple_qhash(listener->iwdev, &nfo,
1882                                                  listener);
1883                 } else {
1884                         if (listener->qhash_set)
1885                                 irdma_manage_qhash(listener->iwdev,
1886                                                    &nfo,
1887                                                    IRDMA_QHASH_TYPE_TCP_SYN,
1888                                                    IRDMA_QHASH_MANAGE_TYPE_DELETE,
1889                                                    NULL, false);
1890                 }
1891
1892                 cm_core->stats_listen_destroyed++;
1893                 cm_core->stats_listen_nodes_destroyed++;
1894                 ibdev_dbg(&listener->iwdev->ibdev,
1895                           "CM: loc_port=0x%04x loc_addr=%pI4 cm_listen_node=%p cm_id=%p qhash_set=%d vlan_id=%d apbvt_del=%d\n",
1896                           listener->loc_port, listener->loc_addr, listener,
1897                           listener->cm_id, listener->qhash_set,
1898                           listener->vlan_id, apbvt_del);
1899                 kfree(listener);
1900                 listener = NULL;
1901                 return 0;
1902         }
1903
1904         return -EINVAL;
1905 }
1906
1907 /**
1908  * irdma_cm_del_listen - delete a listener
1909  * @cm_core: cm's core
1910  * @listener: passive connection's listener
1911  * @apbvt_del: flag to delete apbvt
1912  */
1913 static int irdma_cm_del_listen(struct irdma_cm_core *cm_core,
1914                                struct irdma_cm_listener *listener,
1915                                bool apbvt_del)
1916 {
1917         listener->listener_state = IRDMA_CM_LISTENER_PASSIVE_STATE;
1918         listener->cm_id = NULL;
1919
1920         return irdma_dec_refcnt_listen(cm_core, listener, 1, apbvt_del);
1921 }
1922
1923 /**
1924  * irdma_addr_resolve_neigh - resolve neighbor address
1925  * @iwdev: iwarp device structure
1926  * @src_ip: local ip address
1927  * @dst_ip: remote ip address
1928  * @arpindex: if there is an arp entry
1929  */
1930 static int irdma_addr_resolve_neigh(struct irdma_device *iwdev, u32 src_ip,
1931                                     u32 dst_ip, int arpindex)
1932 {
1933         struct rtable *rt;
1934         struct neighbour *neigh;
1935         int rc = arpindex;
1936         __be32 dst_ipaddr = htonl(dst_ip);
1937         __be32 src_ipaddr = htonl(src_ip);
1938
1939         rt = ip_route_output(&init_net, dst_ipaddr, src_ipaddr, 0, 0);
1940         if (IS_ERR(rt)) {
1941                 ibdev_dbg(&iwdev->ibdev, "CM: ip_route_output fail\n");
1942                 return -EINVAL;
1943         }
1944
1945         neigh = dst_neigh_lookup(&rt->dst, &dst_ipaddr);
1946         if (!neigh)
1947                 goto exit;
1948
1949         if (neigh->nud_state & NUD_VALID)
1950                 rc = irdma_add_arp(iwdev->rf, &dst_ip, true, neigh->ha);
1951         else
1952                 neigh_event_send(neigh, NULL);
1953         if (neigh)
1954                 neigh_release(neigh);
1955 exit:
1956         ip_rt_put(rt);
1957
1958         return rc;
1959 }
1960
1961 /**
1962  * irdma_get_dst_ipv6 - get destination cache entry via ipv6 lookup
1963  * @src_addr: local ipv6 sock address
1964  * @dst_addr: destination ipv6 sock address
1965  */
1966 static struct dst_entry *irdma_get_dst_ipv6(struct sockaddr_in6 *src_addr,
1967                                             struct sockaddr_in6 *dst_addr)
1968 {
1969         struct dst_entry *dst = NULL;
1970
1971         if ((IS_ENABLED(CONFIG_IPV6))) {
1972                 struct flowi6 fl6 = {};
1973
1974                 fl6.daddr = dst_addr->sin6_addr;
1975                 fl6.saddr = src_addr->sin6_addr;
1976                 if (ipv6_addr_type(&fl6.daddr) & IPV6_ADDR_LINKLOCAL)
1977                         fl6.flowi6_oif = dst_addr->sin6_scope_id;
1978
1979                 dst = ip6_route_output(&init_net, NULL, &fl6);
1980         }
1981
1982         return dst;
1983 }
1984
1985 /**
1986  * irdma_addr_resolve_neigh_ipv6 - resolve neighbor ipv6 address
1987  * @iwdev: iwarp device structure
1988  * @src: local ip address
1989  * @dest: remote ip address
1990  * @arpindex: if there is an arp entry
1991  */
1992 static int irdma_addr_resolve_neigh_ipv6(struct irdma_device *iwdev, u32 *src,
1993                                          u32 *dest, int arpindex)
1994 {
1995         struct neighbour *neigh;
1996         int rc = arpindex;
1997         struct dst_entry *dst;
1998         struct sockaddr_in6 dst_addr = {};
1999         struct sockaddr_in6 src_addr = {};
2000
2001         dst_addr.sin6_family = AF_INET6;
2002         irdma_copy_ip_htonl(dst_addr.sin6_addr.in6_u.u6_addr32, dest);
2003         src_addr.sin6_family = AF_INET6;
2004         irdma_copy_ip_htonl(src_addr.sin6_addr.in6_u.u6_addr32, src);
2005         dst = irdma_get_dst_ipv6(&src_addr, &dst_addr);
2006         if (!dst || dst->error) {
2007                 if (dst) {
2008                         dst_release(dst);
2009                         ibdev_dbg(&iwdev->ibdev,
2010                                   "CM: ip6_route_output returned dst->error = %d\n",
2011                                   dst->error);
2012                 }
2013                 return -EINVAL;
2014         }
2015
2016         neigh = dst_neigh_lookup(dst, dst_addr.sin6_addr.in6_u.u6_addr32);
2017         if (!neigh)
2018                 goto exit;
2019
2020         ibdev_dbg(&iwdev->ibdev, "CM: dst_neigh_lookup MAC=%pM\n",
2021                   neigh->ha);
2022
2023         trace_irdma_addr_resolve(iwdev, neigh->ha);
2024
2025         if (neigh->nud_state & NUD_VALID)
2026                 rc = irdma_add_arp(iwdev->rf, dest, false, neigh->ha);
2027         else
2028                 neigh_event_send(neigh, NULL);
2029         if (neigh)
2030                 neigh_release(neigh);
2031 exit:
2032         dst_release(dst);
2033
2034         return rc;
2035 }
2036
2037 /**
2038  * irdma_find_node - find a cm node that matches the reference cm node
2039  * @cm_core: cm's core
2040  * @rem_port: remote tcp port num
2041  * @rem_addr: remote ip addr
2042  * @loc_port: local tcp port num
2043  * @loc_addr: local ip addr
2044  * @vlan_id: local VLAN ID
2045  */
2046 struct irdma_cm_node *irdma_find_node(struct irdma_cm_core *cm_core,
2047                                       u16 rem_port, u32 *rem_addr, u16 loc_port,
2048                                       u32 *loc_addr, u16 vlan_id)
2049 {
2050         struct irdma_cm_node *cm_node;
2051         u32 key = (rem_port << 16) | loc_port;
2052
2053         rcu_read_lock();
2054         hash_for_each_possible_rcu(cm_core->cm_hash_tbl, cm_node, list, key) {
2055                 if (cm_node->vlan_id == vlan_id &&
2056                     cm_node->loc_port == loc_port && cm_node->rem_port == rem_port &&
2057                     !memcmp(cm_node->loc_addr, loc_addr, sizeof(cm_node->loc_addr)) &&
2058                     !memcmp(cm_node->rem_addr, rem_addr, sizeof(cm_node->rem_addr))) {
2059                         if (!refcount_inc_not_zero(&cm_node->refcnt))
2060                                 goto exit;
2061                         rcu_read_unlock();
2062                         trace_irdma_find_node(cm_node, 0, NULL);
2063                         return cm_node;
2064                 }
2065         }
2066
2067 exit:
2068         rcu_read_unlock();
2069
2070         /* no owner node */
2071         return NULL;
2072 }
2073
2074 /**
2075  * irdma_add_hte_node - add a cm node to the hash table
2076  * @cm_core: cm's core
2077  * @cm_node: connection's node
2078  */
2079 static void irdma_add_hte_node(struct irdma_cm_core *cm_core,
2080                                struct irdma_cm_node *cm_node)
2081 {
2082         unsigned long flags;
2083         u32 key = (cm_node->rem_port << 16) | cm_node->loc_port;
2084
2085         spin_lock_irqsave(&cm_core->ht_lock, flags);
2086         hash_add_rcu(cm_core->cm_hash_tbl, &cm_node->list, key);
2087         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2088 }
2089
2090 /**
2091  * irdma_ipv4_is_lpb - check if loopback
2092  * @loc_addr: local addr to compare
2093  * @rem_addr: remote address
2094  */
2095 bool irdma_ipv4_is_lpb(u32 loc_addr, u32 rem_addr)
2096 {
2097         return ipv4_is_loopback(htonl(rem_addr)) || (loc_addr == rem_addr);
2098 }
2099
2100 /**
2101  * irdma_ipv6_is_lpb - check if loopback
2102  * @loc_addr: local addr to compare
2103  * @rem_addr: remote address
2104  */
2105 bool irdma_ipv6_is_lpb(u32 *loc_addr, u32 *rem_addr)
2106 {
2107         struct in6_addr raddr6;
2108
2109         irdma_copy_ip_htonl(raddr6.in6_u.u6_addr32, rem_addr);
2110
2111         return !memcmp(loc_addr, rem_addr, 16) || ipv6_addr_loopback(&raddr6);
2112 }
2113
2114 /**
2115  * irdma_cm_create_ah - create a cm address handle
2116  * @cm_node: The connection manager node to create AH for
2117  * @wait: Provides option to wait for ah creation or not
2118  */
2119 static int irdma_cm_create_ah(struct irdma_cm_node *cm_node, bool wait)
2120 {
2121         struct irdma_ah_info ah_info = {};
2122         struct irdma_device *iwdev = cm_node->iwdev;
2123
2124         ether_addr_copy(ah_info.mac_addr, iwdev->netdev->dev_addr);
2125
2126         ah_info.hop_ttl = 0x40;
2127         ah_info.tc_tos = cm_node->tos;
2128         ah_info.vsi = &iwdev->vsi;
2129
2130         if (cm_node->ipv4) {
2131                 ah_info.ipv4_valid = true;
2132                 ah_info.dest_ip_addr[0] = cm_node->rem_addr[0];
2133                 ah_info.src_ip_addr[0] = cm_node->loc_addr[0];
2134                 ah_info.do_lpbk = irdma_ipv4_is_lpb(ah_info.src_ip_addr[0],
2135                                                     ah_info.dest_ip_addr[0]);
2136         } else {
2137                 memcpy(ah_info.dest_ip_addr, cm_node->rem_addr,
2138                        sizeof(ah_info.dest_ip_addr));
2139                 memcpy(ah_info.src_ip_addr, cm_node->loc_addr,
2140                        sizeof(ah_info.src_ip_addr));
2141                 ah_info.do_lpbk = irdma_ipv6_is_lpb(ah_info.src_ip_addr,
2142                                                     ah_info.dest_ip_addr);
2143         }
2144
2145         ah_info.vlan_tag = cm_node->vlan_id;
2146         if (cm_node->vlan_id < VLAN_N_VID) {
2147                 ah_info.insert_vlan_tag = 1;
2148                 ah_info.vlan_tag |= cm_node->user_pri << VLAN_PRIO_SHIFT;
2149         }
2150
2151         ah_info.dst_arpindex =
2152                 irdma_arp_table(iwdev->rf, ah_info.dest_ip_addr,
2153                                 ah_info.ipv4_valid, NULL, IRDMA_ARP_RESOLVE);
2154
2155         if (irdma_puda_create_ah(&iwdev->rf->sc_dev, &ah_info, wait,
2156                                  IRDMA_PUDA_RSRC_TYPE_ILQ, cm_node,
2157                                  &cm_node->ah))
2158                 return -ENOMEM;
2159
2160         trace_irdma_create_ah(cm_node);
2161         return 0;
2162 }
2163
2164 /**
2165  * irdma_cm_free_ah - free a cm address handle
2166  * @cm_node: The connection manager node to create AH for
2167  */
2168 static void irdma_cm_free_ah(struct irdma_cm_node *cm_node)
2169 {
2170         struct irdma_device *iwdev = cm_node->iwdev;
2171
2172         trace_irdma_cm_free_ah(cm_node);
2173         irdma_puda_free_ah(&iwdev->rf->sc_dev, cm_node->ah);
2174         cm_node->ah = NULL;
2175 }
2176
2177 /**
2178  * irdma_make_cm_node - create a new instance of a cm node
2179  * @cm_core: cm's core
2180  * @iwdev: iwarp device structure
2181  * @cm_info: quad info for connection
2182  * @listener: passive connection's listener
2183  */
2184 static struct irdma_cm_node *
2185 irdma_make_cm_node(struct irdma_cm_core *cm_core, struct irdma_device *iwdev,
2186                    struct irdma_cm_info *cm_info,
2187                    struct irdma_cm_listener *listener)
2188 {
2189         struct irdma_cm_node *cm_node;
2190         int oldarpindex;
2191         int arpindex;
2192         struct net_device *netdev = iwdev->netdev;
2193
2194         /* create an hte and cm_node for this instance */
2195         cm_node = kzalloc(sizeof(*cm_node), GFP_ATOMIC);
2196         if (!cm_node)
2197                 return NULL;
2198
2199         /* set our node specific transport info */
2200         cm_node->ipv4 = cm_info->ipv4;
2201         cm_node->vlan_id = cm_info->vlan_id;
2202         if (cm_node->vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
2203                 cm_node->vlan_id = 0;
2204         cm_node->tos = cm_info->tos;
2205         cm_node->user_pri = cm_info->user_pri;
2206         if (listener) {
2207                 if (listener->tos != cm_info->tos)
2208                         ibdev_warn(&iwdev->ibdev,
2209                                    "application TOS[%d] and remote client TOS[%d] mismatch\n",
2210                                    listener->tos, cm_info->tos);
2211                 if (iwdev->vsi.dscp_mode) {
2212                         cm_node->user_pri = listener->user_pri;
2213                 } else {
2214                         cm_node->tos = max(listener->tos, cm_info->tos);
2215                         cm_node->user_pri = rt_tos2priority(cm_node->tos);
2216                 }
2217                 ibdev_dbg(&iwdev->ibdev,
2218                           "DCB: listener: TOS:[%d] UP:[%d]\n", cm_node->tos,
2219                           cm_node->user_pri);
2220                 trace_irdma_listener_tos(iwdev, cm_node->tos,
2221                                          cm_node->user_pri);
2222         }
2223         memcpy(cm_node->loc_addr, cm_info->loc_addr, sizeof(cm_node->loc_addr));
2224         memcpy(cm_node->rem_addr, cm_info->rem_addr, sizeof(cm_node->rem_addr));
2225         cm_node->loc_port = cm_info->loc_port;
2226         cm_node->rem_port = cm_info->rem_port;
2227
2228         cm_node->mpa_frame_rev = IRDMA_CM_DEFAULT_MPA_VER;
2229         cm_node->send_rdma0_op = SEND_RDMA_READ_ZERO;
2230         cm_node->iwdev = iwdev;
2231         cm_node->dev = &iwdev->rf->sc_dev;
2232
2233         cm_node->ird_size = cm_node->dev->hw_attrs.max_hw_ird;
2234         cm_node->ord_size = cm_node->dev->hw_attrs.max_hw_ord;
2235
2236         cm_node->listener = listener;
2237         cm_node->cm_id = cm_info->cm_id;
2238         ether_addr_copy(cm_node->loc_mac, netdev->dev_addr);
2239         spin_lock_init(&cm_node->retrans_list_lock);
2240         cm_node->ack_rcvd = false;
2241
2242         init_completion(&cm_node->establish_comp);
2243         refcount_set(&cm_node->refcnt, 1);
2244         /* associate our parent CM core */
2245         cm_node->cm_core = cm_core;
2246         cm_node->tcp_cntxt.loc_id = IRDMA_CM_DEFAULT_LOCAL_ID;
2247         cm_node->tcp_cntxt.rcv_wscale = iwdev->rcv_wscale;
2248         cm_node->tcp_cntxt.rcv_wnd = iwdev->rcv_wnd >> cm_node->tcp_cntxt.rcv_wscale;
2249         if (cm_node->ipv4) {
2250                 cm_node->tcp_cntxt.loc_seq_num = secure_tcp_seq(htonl(cm_node->loc_addr[0]),
2251                                                                 htonl(cm_node->rem_addr[0]),
2252                                                                 htons(cm_node->loc_port),
2253                                                                 htons(cm_node->rem_port));
2254                 cm_node->tcp_cntxt.mss = iwdev->vsi.mtu - IRDMA_MTU_TO_MSS_IPV4;
2255         } else if (IS_ENABLED(CONFIG_IPV6)) {
2256                 __be32 loc[4] = {
2257                         htonl(cm_node->loc_addr[0]), htonl(cm_node->loc_addr[1]),
2258                         htonl(cm_node->loc_addr[2]), htonl(cm_node->loc_addr[3])
2259                 };
2260                 __be32 rem[4] = {
2261                         htonl(cm_node->rem_addr[0]), htonl(cm_node->rem_addr[1]),
2262                         htonl(cm_node->rem_addr[2]), htonl(cm_node->rem_addr[3])
2263                 };
2264                 cm_node->tcp_cntxt.loc_seq_num = secure_tcpv6_seq(loc, rem,
2265                                                                   htons(cm_node->loc_port),
2266                                                                   htons(cm_node->rem_port));
2267                 cm_node->tcp_cntxt.mss = iwdev->vsi.mtu - IRDMA_MTU_TO_MSS_IPV6;
2268         }
2269
2270         if ((cm_node->ipv4 &&
2271              irdma_ipv4_is_lpb(cm_node->loc_addr[0], cm_node->rem_addr[0])) ||
2272             (!cm_node->ipv4 &&
2273              irdma_ipv6_is_lpb(cm_node->loc_addr, cm_node->rem_addr))) {
2274                 cm_node->do_lpb = true;
2275                 arpindex = irdma_arp_table(iwdev->rf, cm_node->rem_addr,
2276                                            cm_node->ipv4, NULL,
2277                                            IRDMA_ARP_RESOLVE);
2278         } else {
2279                 oldarpindex = irdma_arp_table(iwdev->rf, cm_node->rem_addr,
2280                                               cm_node->ipv4, NULL,
2281                                               IRDMA_ARP_RESOLVE);
2282                 if (cm_node->ipv4)
2283                         arpindex = irdma_addr_resolve_neigh(iwdev,
2284                                                             cm_info->loc_addr[0],
2285                                                             cm_info->rem_addr[0],
2286                                                             oldarpindex);
2287                 else if (IS_ENABLED(CONFIG_IPV6))
2288                         arpindex = irdma_addr_resolve_neigh_ipv6(iwdev,
2289                                                                  cm_info->loc_addr,
2290                                                                  cm_info->rem_addr,
2291                                                                  oldarpindex);
2292                 else
2293                         arpindex = -EINVAL;
2294         }
2295
2296         if (arpindex < 0)
2297                 goto err;
2298
2299         ether_addr_copy(cm_node->rem_mac,
2300                         iwdev->rf->arp_table[arpindex].mac_addr);
2301         irdma_add_hte_node(cm_core, cm_node);
2302         cm_core->stats_nodes_created++;
2303         return cm_node;
2304
2305 err:
2306         kfree(cm_node);
2307
2308         return NULL;
2309 }
2310
2311 static void irdma_destroy_connection(struct irdma_cm_node *cm_node)
2312 {
2313         struct irdma_cm_core *cm_core = cm_node->cm_core;
2314         struct irdma_qp *iwqp;
2315         struct irdma_cm_info nfo;
2316
2317         /* if the node is destroyed before connection was accelerated */
2318         if (!cm_node->accelerated && cm_node->accept_pend) {
2319                 ibdev_dbg(&cm_node->iwdev->ibdev,
2320                           "CM: node destroyed before established\n");
2321                 atomic_dec(&cm_node->listener->pend_accepts_cnt);
2322         }
2323         if (cm_node->close_entry)
2324                 irdma_handle_close_entry(cm_node, 0);
2325         if (cm_node->listener) {
2326                 irdma_dec_refcnt_listen(cm_core, cm_node->listener, 0, true);
2327         } else {
2328                 if (cm_node->apbvt_set) {
2329                         irdma_del_apbvt(cm_node->iwdev, cm_node->apbvt_entry);
2330                         cm_node->apbvt_set = 0;
2331                 }
2332                 irdma_get_addr_info(cm_node, &nfo);
2333                 if (cm_node->qhash_set) {
2334                         nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2335                         irdma_manage_qhash(cm_node->iwdev, &nfo,
2336                                            IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2337                                            IRDMA_QHASH_MANAGE_TYPE_DELETE, NULL,
2338                                            false);
2339                         cm_node->qhash_set = 0;
2340                 }
2341         }
2342
2343         iwqp = cm_node->iwqp;
2344         if (iwqp) {
2345                 cm_node->cm_id->rem_ref(cm_node->cm_id);
2346                 cm_node->cm_id = NULL;
2347                 iwqp->cm_id = NULL;
2348                 irdma_qp_rem_ref(&iwqp->ibqp);
2349                 cm_node->iwqp = NULL;
2350         } else if (cm_node->qhash_set) {
2351                 irdma_get_addr_info(cm_node, &nfo);
2352                 nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2353                 irdma_manage_qhash(cm_node->iwdev, &nfo,
2354                                    IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2355                                    IRDMA_QHASH_MANAGE_TYPE_DELETE, NULL, false);
2356                 cm_node->qhash_set = 0;
2357         }
2358
2359         cm_core->cm_free_ah(cm_node);
2360 }
2361
2362 /**
2363  * irdma_rem_ref_cm_node - destroy an instance of a cm node
2364  * @cm_node: connection's node
2365  */
2366 void irdma_rem_ref_cm_node(struct irdma_cm_node *cm_node)
2367 {
2368         struct irdma_cm_core *cm_core = cm_node->cm_core;
2369         unsigned long flags;
2370
2371         trace_irdma_rem_ref_cm_node(cm_node, 0, __builtin_return_address(0));
2372         spin_lock_irqsave(&cm_core->ht_lock, flags);
2373
2374         if (!refcount_dec_and_test(&cm_node->refcnt)) {
2375                 spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2376                 return;
2377         }
2378         if (cm_node->iwqp) {
2379                 cm_node->iwqp->cm_node = NULL;
2380                 cm_node->iwqp->cm_id = NULL;
2381         }
2382         hash_del_rcu(&cm_node->list);
2383         cm_node->cm_core->stats_nodes_destroyed++;
2384
2385         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2386
2387         irdma_destroy_connection(cm_node);
2388
2389         kfree_rcu(cm_node, rcu_head);
2390 }
2391
2392 /**
2393  * irdma_handle_fin_pkt - FIN packet received
2394  * @cm_node: connection's node
2395  */
2396 static void irdma_handle_fin_pkt(struct irdma_cm_node *cm_node)
2397 {
2398         switch (cm_node->state) {
2399         case IRDMA_CM_STATE_SYN_RCVD:
2400         case IRDMA_CM_STATE_SYN_SENT:
2401         case IRDMA_CM_STATE_ESTABLISHED:
2402         case IRDMA_CM_STATE_MPAREJ_RCVD:
2403                 cm_node->tcp_cntxt.rcv_nxt++;
2404                 irdma_cleanup_retrans_entry(cm_node);
2405                 cm_node->state = IRDMA_CM_STATE_LAST_ACK;
2406                 irdma_send_fin(cm_node);
2407                 break;
2408         case IRDMA_CM_STATE_MPAREQ_SENT:
2409                 irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
2410                 cm_node->tcp_cntxt.rcv_nxt++;
2411                 irdma_cleanup_retrans_entry(cm_node);
2412                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2413                 refcount_inc(&cm_node->refcnt);
2414                 irdma_send_reset(cm_node);
2415                 break;
2416         case IRDMA_CM_STATE_FIN_WAIT1:
2417                 cm_node->tcp_cntxt.rcv_nxt++;
2418                 irdma_cleanup_retrans_entry(cm_node);
2419                 cm_node->state = IRDMA_CM_STATE_CLOSING;
2420                 irdma_send_ack(cm_node);
2421                 /*
2422                  * Wait for ACK as this is simultaneous close.
2423                  * After we receive ACK, do not send anything.
2424                  * Just rm the node.
2425                  */
2426                 break;
2427         case IRDMA_CM_STATE_FIN_WAIT2:
2428                 cm_node->tcp_cntxt.rcv_nxt++;
2429                 irdma_cleanup_retrans_entry(cm_node);
2430                 cm_node->state = IRDMA_CM_STATE_TIME_WAIT;
2431                 irdma_send_ack(cm_node);
2432                 irdma_schedule_cm_timer(cm_node, NULL, IRDMA_TIMER_TYPE_CLOSE,
2433                                         1, 0);
2434                 break;
2435         case IRDMA_CM_STATE_TIME_WAIT:
2436                 cm_node->tcp_cntxt.rcv_nxt++;
2437                 irdma_cleanup_retrans_entry(cm_node);
2438                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2439                 irdma_rem_ref_cm_node(cm_node);
2440                 break;
2441         case IRDMA_CM_STATE_OFFLOADED:
2442         default:
2443                 ibdev_dbg(&cm_node->iwdev->ibdev,
2444                           "CM: bad state node state = %d\n", cm_node->state);
2445                 break;
2446         }
2447 }
2448
2449 /**
2450  * irdma_handle_rst_pkt - process received RST packet
2451  * @cm_node: connection's node
2452  * @rbuf: receive buffer
2453  */
2454 static void irdma_handle_rst_pkt(struct irdma_cm_node *cm_node,
2455                                  struct irdma_puda_buf *rbuf)
2456 {
2457         ibdev_dbg(&cm_node->iwdev->ibdev,
2458                   "CM: caller: %pS cm_node=%p state=%d rem_port=0x%04x loc_port=0x%04x rem_addr=%pI4 loc_addr=%pI4\n",
2459                   __builtin_return_address(0), cm_node, cm_node->state,
2460                   cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr,
2461                   cm_node->loc_addr);
2462
2463         irdma_cleanup_retrans_entry(cm_node);
2464         switch (cm_node->state) {
2465         case IRDMA_CM_STATE_SYN_SENT:
2466         case IRDMA_CM_STATE_MPAREQ_SENT:
2467                 switch (cm_node->mpa_frame_rev) {
2468                 case IETF_MPA_V2:
2469                         /* Drop down to MPA_V1*/
2470                         cm_node->mpa_frame_rev = IETF_MPA_V1;
2471                         /* send a syn and goto syn sent state */
2472                         cm_node->state = IRDMA_CM_STATE_SYN_SENT;
2473                         if (irdma_send_syn(cm_node, 0))
2474                                 irdma_active_open_err(cm_node, false);
2475                         break;
2476                 case IETF_MPA_V1:
2477                 default:
2478                         irdma_active_open_err(cm_node, false);
2479                         break;
2480                 }
2481                 break;
2482         case IRDMA_CM_STATE_MPAREQ_RCVD:
2483                 atomic_inc(&cm_node->passive_state);
2484                 break;
2485         case IRDMA_CM_STATE_ESTABLISHED:
2486         case IRDMA_CM_STATE_SYN_RCVD:
2487         case IRDMA_CM_STATE_LISTENING:
2488                 irdma_passive_open_err(cm_node, false);
2489                 break;
2490         case IRDMA_CM_STATE_OFFLOADED:
2491                 irdma_active_open_err(cm_node, false);
2492                 break;
2493         case IRDMA_CM_STATE_CLOSED:
2494                 break;
2495         case IRDMA_CM_STATE_FIN_WAIT2:
2496         case IRDMA_CM_STATE_FIN_WAIT1:
2497         case IRDMA_CM_STATE_LAST_ACK:
2498         case IRDMA_CM_STATE_TIME_WAIT:
2499                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2500                 irdma_rem_ref_cm_node(cm_node);
2501                 break;
2502         default:
2503                 break;
2504         }
2505 }
2506
2507 /**
2508  * irdma_handle_rcv_mpa - Process a recv'd mpa buffer
2509  * @cm_node: connection's node
2510  * @rbuf: receive buffer
2511  */
2512 static void irdma_handle_rcv_mpa(struct irdma_cm_node *cm_node,
2513                                  struct irdma_puda_buf *rbuf)
2514 {
2515         int err;
2516         int datasize = rbuf->datalen;
2517         u8 *dataloc = rbuf->data;
2518
2519         enum irdma_cm_event_type type = IRDMA_CM_EVENT_UNKNOWN;
2520         u32 res_type;
2521
2522         err = irdma_parse_mpa(cm_node, dataloc, &res_type, datasize);
2523         if (err) {
2524                 if (cm_node->state == IRDMA_CM_STATE_MPAREQ_SENT)
2525                         irdma_active_open_err(cm_node, true);
2526                 else
2527                         irdma_passive_open_err(cm_node, true);
2528                 return;
2529         }
2530
2531         switch (cm_node->state) {
2532         case IRDMA_CM_STATE_ESTABLISHED:
2533                 if (res_type == IRDMA_MPA_REQUEST_REJECT)
2534                         ibdev_dbg(&cm_node->iwdev->ibdev,
2535                                   "CM: state for reject\n");
2536                 cm_node->state = IRDMA_CM_STATE_MPAREQ_RCVD;
2537                 type = IRDMA_CM_EVENT_MPA_REQ;
2538                 irdma_send_ack(cm_node); /* ACK received MPA request */
2539                 atomic_set(&cm_node->passive_state,
2540                            IRDMA_PASSIVE_STATE_INDICATED);
2541                 break;
2542         case IRDMA_CM_STATE_MPAREQ_SENT:
2543                 irdma_cleanup_retrans_entry(cm_node);
2544                 if (res_type == IRDMA_MPA_REQUEST_REJECT) {
2545                         type = IRDMA_CM_EVENT_MPA_REJECT;
2546                         cm_node->state = IRDMA_CM_STATE_MPAREJ_RCVD;
2547                 } else {
2548                         type = IRDMA_CM_EVENT_CONNECTED;
2549                         cm_node->state = IRDMA_CM_STATE_OFFLOADED;
2550                 }
2551                 irdma_send_ack(cm_node);
2552                 break;
2553         default:
2554                 ibdev_dbg(&cm_node->iwdev->ibdev,
2555                           "CM: wrong cm_node state =%d\n", cm_node->state);
2556                 break;
2557         }
2558         irdma_create_event(cm_node, type);
2559 }
2560
2561 /**
2562  * irdma_check_syn - Check for error on received syn ack
2563  * @cm_node: connection's node
2564  * @tcph: pointer tcp header
2565  */
2566 static int irdma_check_syn(struct irdma_cm_node *cm_node, struct tcphdr *tcph)
2567 {
2568         if (ntohl(tcph->ack_seq) != cm_node->tcp_cntxt.loc_seq_num) {
2569                 irdma_active_open_err(cm_node, true);
2570                 return 1;
2571         }
2572
2573         return 0;
2574 }
2575
2576 /**
2577  * irdma_check_seq - check seq numbers if OK
2578  * @cm_node: connection's node
2579  * @tcph: pointer tcp header
2580  */
2581 static int irdma_check_seq(struct irdma_cm_node *cm_node, struct tcphdr *tcph)
2582 {
2583         u32 seq;
2584         u32 ack_seq;
2585         u32 loc_seq_num = cm_node->tcp_cntxt.loc_seq_num;
2586         u32 rcv_nxt = cm_node->tcp_cntxt.rcv_nxt;
2587         u32 rcv_wnd;
2588         int err = 0;
2589
2590         seq = ntohl(tcph->seq);
2591         ack_seq = ntohl(tcph->ack_seq);
2592         rcv_wnd = cm_node->tcp_cntxt.rcv_wnd;
2593         if (ack_seq != loc_seq_num ||
2594             !between(seq, rcv_nxt, (rcv_nxt + rcv_wnd)))
2595                 err = -1;
2596         if (err)
2597                 ibdev_dbg(&cm_node->iwdev->ibdev,
2598                           "CM: seq number err\n");
2599
2600         return err;
2601 }
2602
2603 void irdma_add_conn_est_qh(struct irdma_cm_node *cm_node)
2604 {
2605         struct irdma_cm_info nfo;
2606
2607         irdma_get_addr_info(cm_node, &nfo);
2608         nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2609         irdma_manage_qhash(cm_node->iwdev, &nfo,
2610                            IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2611                            IRDMA_QHASH_MANAGE_TYPE_ADD,
2612                            cm_node, false);
2613         cm_node->qhash_set = true;
2614 }
2615
2616 /**
2617  * irdma_handle_syn_pkt - is for Passive node
2618  * @cm_node: connection's node
2619  * @rbuf: receive buffer
2620  */
2621 static void irdma_handle_syn_pkt(struct irdma_cm_node *cm_node,
2622                                  struct irdma_puda_buf *rbuf)
2623 {
2624         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2625         int err;
2626         u32 inc_sequence;
2627         int optionsize;
2628
2629         optionsize = (tcph->doff << 2) - sizeof(struct tcphdr);
2630         inc_sequence = ntohl(tcph->seq);
2631
2632         switch (cm_node->state) {
2633         case IRDMA_CM_STATE_SYN_SENT:
2634         case IRDMA_CM_STATE_MPAREQ_SENT:
2635                 /* Rcvd syn on active open connection */
2636                 irdma_active_open_err(cm_node, 1);
2637                 break;
2638         case IRDMA_CM_STATE_LISTENING:
2639                 /* Passive OPEN */
2640                 if (atomic_read(&cm_node->listener->pend_accepts_cnt) >
2641                     cm_node->listener->backlog) {
2642                         cm_node->cm_core->stats_backlog_drops++;
2643                         irdma_passive_open_err(cm_node, false);
2644                         break;
2645                 }
2646                 err = irdma_handle_tcp_options(cm_node, tcph, optionsize, 1);
2647                 if (err) {
2648                         irdma_passive_open_err(cm_node, false);
2649                         /* drop pkt */
2650                         break;
2651                 }
2652                 err = cm_node->cm_core->cm_create_ah(cm_node, false);
2653                 if (err) {
2654                         irdma_passive_open_err(cm_node, false);
2655                         /* drop pkt */
2656                         break;
2657                 }
2658                 cm_node->tcp_cntxt.rcv_nxt = inc_sequence + 1;
2659                 cm_node->accept_pend = 1;
2660                 atomic_inc(&cm_node->listener->pend_accepts_cnt);
2661
2662                 cm_node->state = IRDMA_CM_STATE_SYN_RCVD;
2663                 break;
2664         case IRDMA_CM_STATE_CLOSED:
2665                 irdma_cleanup_retrans_entry(cm_node);
2666                 refcount_inc(&cm_node->refcnt);
2667                 irdma_send_reset(cm_node);
2668                 break;
2669         case IRDMA_CM_STATE_OFFLOADED:
2670         case IRDMA_CM_STATE_ESTABLISHED:
2671         case IRDMA_CM_STATE_FIN_WAIT1:
2672         case IRDMA_CM_STATE_FIN_WAIT2:
2673         case IRDMA_CM_STATE_MPAREQ_RCVD:
2674         case IRDMA_CM_STATE_LAST_ACK:
2675         case IRDMA_CM_STATE_CLOSING:
2676         case IRDMA_CM_STATE_UNKNOWN:
2677         default:
2678                 break;
2679         }
2680 }
2681
2682 /**
2683  * irdma_handle_synack_pkt - Process SYN+ACK packet (active side)
2684  * @cm_node: connection's node
2685  * @rbuf: receive buffer
2686  */
2687 static void irdma_handle_synack_pkt(struct irdma_cm_node *cm_node,
2688                                     struct irdma_puda_buf *rbuf)
2689 {
2690         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2691         int err;
2692         u32 inc_sequence;
2693         int optionsize;
2694
2695         optionsize = (tcph->doff << 2) - sizeof(struct tcphdr);
2696         inc_sequence = ntohl(tcph->seq);
2697         switch (cm_node->state) {
2698         case IRDMA_CM_STATE_SYN_SENT:
2699                 irdma_cleanup_retrans_entry(cm_node);
2700                 /* active open */
2701                 if (irdma_check_syn(cm_node, tcph)) {
2702                         ibdev_dbg(&cm_node->iwdev->ibdev,
2703                                   "CM: check syn fail\n");
2704                         return;
2705                 }
2706                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->ack_seq);
2707                 /* setup options */
2708                 err = irdma_handle_tcp_options(cm_node, tcph, optionsize, 0);
2709                 if (err) {
2710                         ibdev_dbg(&cm_node->iwdev->ibdev,
2711                                   "CM: cm_node=%p tcp_options failed\n",
2712                                   cm_node);
2713                         break;
2714                 }
2715                 irdma_cleanup_retrans_entry(cm_node);
2716                 cm_node->tcp_cntxt.rcv_nxt = inc_sequence + 1;
2717                 irdma_send_ack(cm_node); /* ACK  for the syn_ack */
2718                 err = irdma_send_mpa_request(cm_node);
2719                 if (err) {
2720                         ibdev_dbg(&cm_node->iwdev->ibdev,
2721                                   "CM: cm_node=%p irdma_send_mpa_request failed\n",
2722                                   cm_node);
2723                         break;
2724                 }
2725                 cm_node->state = IRDMA_CM_STATE_MPAREQ_SENT;
2726                 break;
2727         case IRDMA_CM_STATE_MPAREQ_RCVD:
2728                 irdma_passive_open_err(cm_node, true);
2729                 break;
2730         case IRDMA_CM_STATE_LISTENING:
2731                 cm_node->tcp_cntxt.loc_seq_num = ntohl(tcph->ack_seq);
2732                 irdma_cleanup_retrans_entry(cm_node);
2733                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2734                 irdma_send_reset(cm_node);
2735                 break;
2736         case IRDMA_CM_STATE_CLOSED:
2737                 cm_node->tcp_cntxt.loc_seq_num = ntohl(tcph->ack_seq);
2738                 irdma_cleanup_retrans_entry(cm_node);
2739                 refcount_inc(&cm_node->refcnt);
2740                 irdma_send_reset(cm_node);
2741                 break;
2742         case IRDMA_CM_STATE_ESTABLISHED:
2743         case IRDMA_CM_STATE_FIN_WAIT1:
2744         case IRDMA_CM_STATE_FIN_WAIT2:
2745         case IRDMA_CM_STATE_LAST_ACK:
2746         case IRDMA_CM_STATE_OFFLOADED:
2747         case IRDMA_CM_STATE_CLOSING:
2748         case IRDMA_CM_STATE_UNKNOWN:
2749         case IRDMA_CM_STATE_MPAREQ_SENT:
2750         default:
2751                 break;
2752         }
2753 }
2754
2755 /**
2756  * irdma_handle_ack_pkt - process packet with ACK
2757  * @cm_node: connection's node
2758  * @rbuf: receive buffer
2759  */
2760 static int irdma_handle_ack_pkt(struct irdma_cm_node *cm_node,
2761                                 struct irdma_puda_buf *rbuf)
2762 {
2763         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2764         u32 inc_sequence;
2765         int ret;
2766         int optionsize;
2767         u32 datasize = rbuf->datalen;
2768
2769         optionsize = (tcph->doff << 2) - sizeof(struct tcphdr);
2770
2771         if (irdma_check_seq(cm_node, tcph))
2772                 return -EINVAL;
2773
2774         inc_sequence = ntohl(tcph->seq);
2775         switch (cm_node->state) {
2776         case IRDMA_CM_STATE_SYN_RCVD:
2777                 irdma_cleanup_retrans_entry(cm_node);
2778                 ret = irdma_handle_tcp_options(cm_node, tcph, optionsize, 1);
2779                 if (ret)
2780                         return ret;
2781                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->ack_seq);
2782                 cm_node->state = IRDMA_CM_STATE_ESTABLISHED;
2783                 if (datasize) {
2784                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2785                         irdma_handle_rcv_mpa(cm_node, rbuf);
2786                 }
2787                 break;
2788         case IRDMA_CM_STATE_ESTABLISHED:
2789                 irdma_cleanup_retrans_entry(cm_node);
2790                 if (datasize) {
2791                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2792                         irdma_handle_rcv_mpa(cm_node, rbuf);
2793                 }
2794                 break;
2795         case IRDMA_CM_STATE_MPAREQ_SENT:
2796                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->ack_seq);
2797                 if (datasize) {
2798                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2799                         cm_node->ack_rcvd = false;
2800                         irdma_handle_rcv_mpa(cm_node, rbuf);
2801                 } else {
2802                         cm_node->ack_rcvd = true;
2803                 }
2804                 break;
2805         case IRDMA_CM_STATE_LISTENING:
2806                 irdma_cleanup_retrans_entry(cm_node);
2807                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2808                 irdma_send_reset(cm_node);
2809                 break;
2810         case IRDMA_CM_STATE_CLOSED:
2811                 irdma_cleanup_retrans_entry(cm_node);
2812                 refcount_inc(&cm_node->refcnt);
2813                 irdma_send_reset(cm_node);
2814                 break;
2815         case IRDMA_CM_STATE_LAST_ACK:
2816         case IRDMA_CM_STATE_CLOSING:
2817                 irdma_cleanup_retrans_entry(cm_node);
2818                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2819                 irdma_rem_ref_cm_node(cm_node);
2820                 break;
2821         case IRDMA_CM_STATE_FIN_WAIT1:
2822                 irdma_cleanup_retrans_entry(cm_node);
2823                 cm_node->state = IRDMA_CM_STATE_FIN_WAIT2;
2824                 break;
2825         case IRDMA_CM_STATE_SYN_SENT:
2826         case IRDMA_CM_STATE_FIN_WAIT2:
2827         case IRDMA_CM_STATE_OFFLOADED:
2828         case IRDMA_CM_STATE_MPAREQ_RCVD:
2829         case IRDMA_CM_STATE_UNKNOWN:
2830         default:
2831                 irdma_cleanup_retrans_entry(cm_node);
2832                 break;
2833         }
2834
2835         return 0;
2836 }
2837
2838 /**
2839  * irdma_process_pkt - process cm packet
2840  * @cm_node: connection's node
2841  * @rbuf: receive buffer
2842  */
2843 static void irdma_process_pkt(struct irdma_cm_node *cm_node,
2844                               struct irdma_puda_buf *rbuf)
2845 {
2846         enum irdma_tcpip_pkt_type pkt_type = IRDMA_PKT_TYPE_UNKNOWN;
2847         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2848         u32 fin_set = 0;
2849         int err;
2850
2851         if (tcph->rst) {
2852                 pkt_type = IRDMA_PKT_TYPE_RST;
2853         } else if (tcph->syn) {
2854                 pkt_type = IRDMA_PKT_TYPE_SYN;
2855                 if (tcph->ack)
2856                         pkt_type = IRDMA_PKT_TYPE_SYNACK;
2857         } else if (tcph->ack) {
2858                 pkt_type = IRDMA_PKT_TYPE_ACK;
2859         }
2860         if (tcph->fin)
2861                 fin_set = 1;
2862
2863         switch (pkt_type) {
2864         case IRDMA_PKT_TYPE_SYN:
2865                 irdma_handle_syn_pkt(cm_node, rbuf);
2866                 break;
2867         case IRDMA_PKT_TYPE_SYNACK:
2868                 irdma_handle_synack_pkt(cm_node, rbuf);
2869                 break;
2870         case IRDMA_PKT_TYPE_ACK:
2871                 err = irdma_handle_ack_pkt(cm_node, rbuf);
2872                 if (fin_set && !err)
2873                         irdma_handle_fin_pkt(cm_node);
2874                 break;
2875         case IRDMA_PKT_TYPE_RST:
2876                 irdma_handle_rst_pkt(cm_node, rbuf);
2877                 break;
2878         default:
2879                 if (fin_set &&
2880                     (!irdma_check_seq(cm_node, (struct tcphdr *)rbuf->tcph)))
2881                         irdma_handle_fin_pkt(cm_node);
2882                 break;
2883         }
2884 }
2885
2886 /**
2887  * irdma_make_listen_node - create a listen node with params
2888  * @cm_core: cm's core
2889  * @iwdev: iwarp device structure
2890  * @cm_info: quad info for connection
2891  */
2892 static struct irdma_cm_listener *
2893 irdma_make_listen_node(struct irdma_cm_core *cm_core,
2894                        struct irdma_device *iwdev,
2895                        struct irdma_cm_info *cm_info)
2896 {
2897         struct irdma_cm_listener *listener;
2898         unsigned long flags;
2899
2900         /* cannot have multiple matching listeners */
2901         listener = irdma_find_listener(cm_core, cm_info->loc_addr,
2902                                        cm_info->loc_port, cm_info->vlan_id,
2903                                        IRDMA_CM_LISTENER_EITHER_STATE);
2904         if (listener &&
2905             listener->listener_state == IRDMA_CM_LISTENER_ACTIVE_STATE) {
2906                 refcount_dec(&listener->refcnt);
2907                 return NULL;
2908         }
2909
2910         if (!listener) {
2911                 /* create a CM listen node
2912                  * 1/2 node to compare incoming traffic to
2913                  */
2914                 listener = kzalloc(sizeof(*listener), GFP_KERNEL);
2915                 if (!listener)
2916                         return NULL;
2917                 cm_core->stats_listen_nodes_created++;
2918                 memcpy(listener->loc_addr, cm_info->loc_addr,
2919                        sizeof(listener->loc_addr));
2920                 listener->loc_port = cm_info->loc_port;
2921
2922                 INIT_LIST_HEAD(&listener->child_listen_list);
2923
2924                 refcount_set(&listener->refcnt, 1);
2925         } else {
2926                 listener->reused_node = 1;
2927         }
2928
2929         listener->cm_id = cm_info->cm_id;
2930         listener->ipv4 = cm_info->ipv4;
2931         listener->vlan_id = cm_info->vlan_id;
2932         atomic_set(&listener->pend_accepts_cnt, 0);
2933         listener->cm_core = cm_core;
2934         listener->iwdev = iwdev;
2935
2936         listener->backlog = cm_info->backlog;
2937         listener->listener_state = IRDMA_CM_LISTENER_ACTIVE_STATE;
2938
2939         if (!listener->reused_node) {
2940                 spin_lock_irqsave(&cm_core->listen_list_lock, flags);
2941                 list_add(&listener->list, &cm_core->listen_list);
2942                 spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
2943         }
2944
2945         return listener;
2946 }
2947
2948 /**
2949  * irdma_create_cm_node - make a connection node with params
2950  * @cm_core: cm's core
2951  * @iwdev: iwarp device structure
2952  * @conn_param: connection parameters
2953  * @cm_info: quad info for connection
2954  * @caller_cm_node: pointer to cm_node structure to return
2955  */
2956 static int irdma_create_cm_node(struct irdma_cm_core *cm_core,
2957                                 struct irdma_device *iwdev,
2958                                 struct iw_cm_conn_param *conn_param,
2959                                 struct irdma_cm_info *cm_info,
2960                                 struct irdma_cm_node **caller_cm_node)
2961 {
2962         struct irdma_cm_node *cm_node;
2963         u16 private_data_len = conn_param->private_data_len;
2964         const void *private_data = conn_param->private_data;
2965
2966         /* create a CM connection node */
2967         cm_node = irdma_make_cm_node(cm_core, iwdev, cm_info, NULL);
2968         if (!cm_node)
2969                 return -ENOMEM;
2970
2971         /* set our node side to client (active) side */
2972         cm_node->tcp_cntxt.client = 1;
2973         cm_node->tcp_cntxt.rcv_wscale = IRDMA_CM_DEFAULT_RCV_WND_SCALE;
2974
2975         irdma_record_ird_ord(cm_node, conn_param->ird, conn_param->ord);
2976
2977         cm_node->pdata.size = private_data_len;
2978         cm_node->pdata.addr = cm_node->pdata_buf;
2979
2980         memcpy(cm_node->pdata_buf, private_data, private_data_len);
2981         *caller_cm_node = cm_node;
2982
2983         return 0;
2984 }
2985
2986 /**
2987  * irdma_cm_reject - reject and teardown a connection
2988  * @cm_node: connection's node
2989  * @pdata: ptr to private data for reject
2990  * @plen: size of private data
2991  */
2992 static int irdma_cm_reject(struct irdma_cm_node *cm_node, const void *pdata,
2993                            u8 plen)
2994 {
2995         int ret;
2996         int passive_state;
2997
2998         if (cm_node->tcp_cntxt.client)
2999                 return 0;
3000
3001         irdma_cleanup_retrans_entry(cm_node);
3002
3003         passive_state = atomic_add_return(1, &cm_node->passive_state);
3004         if (passive_state == IRDMA_SEND_RESET_EVENT) {
3005                 cm_node->state = IRDMA_CM_STATE_CLOSED;
3006                 irdma_rem_ref_cm_node(cm_node);
3007                 return 0;
3008         }
3009
3010         if (cm_node->state == IRDMA_CM_STATE_LISTENER_DESTROYED) {
3011                 irdma_rem_ref_cm_node(cm_node);
3012                 return 0;
3013         }
3014
3015         ret = irdma_send_mpa_reject(cm_node, pdata, plen);
3016         if (!ret)
3017                 return 0;
3018
3019         cm_node->state = IRDMA_CM_STATE_CLOSED;
3020         if (irdma_send_reset(cm_node))
3021                 ibdev_dbg(&cm_node->iwdev->ibdev,
3022                           "CM: send reset failed\n");
3023
3024         return ret;
3025 }
3026
3027 /**
3028  * irdma_cm_close - close of cm connection
3029  * @cm_node: connection's node
3030  */
3031 static int irdma_cm_close(struct irdma_cm_node *cm_node)
3032 {
3033         switch (cm_node->state) {
3034         case IRDMA_CM_STATE_SYN_RCVD:
3035         case IRDMA_CM_STATE_SYN_SENT:
3036         case IRDMA_CM_STATE_ONE_SIDE_ESTABLISHED:
3037         case IRDMA_CM_STATE_ESTABLISHED:
3038         case IRDMA_CM_STATE_ACCEPTING:
3039         case IRDMA_CM_STATE_MPAREQ_SENT:
3040         case IRDMA_CM_STATE_MPAREQ_RCVD:
3041                 irdma_cleanup_retrans_entry(cm_node);
3042                 irdma_send_reset(cm_node);
3043                 break;
3044         case IRDMA_CM_STATE_CLOSE_WAIT:
3045                 cm_node->state = IRDMA_CM_STATE_LAST_ACK;
3046                 irdma_send_fin(cm_node);
3047                 break;
3048         case IRDMA_CM_STATE_FIN_WAIT1:
3049         case IRDMA_CM_STATE_FIN_WAIT2:
3050         case IRDMA_CM_STATE_LAST_ACK:
3051         case IRDMA_CM_STATE_TIME_WAIT:
3052         case IRDMA_CM_STATE_CLOSING:
3053                 return -EINVAL;
3054         case IRDMA_CM_STATE_LISTENING:
3055                 irdma_cleanup_retrans_entry(cm_node);
3056                 irdma_send_reset(cm_node);
3057                 break;
3058         case IRDMA_CM_STATE_MPAREJ_RCVD:
3059         case IRDMA_CM_STATE_UNKNOWN:
3060         case IRDMA_CM_STATE_INITED:
3061         case IRDMA_CM_STATE_CLOSED:
3062         case IRDMA_CM_STATE_LISTENER_DESTROYED:
3063                 irdma_rem_ref_cm_node(cm_node);
3064                 break;
3065         case IRDMA_CM_STATE_OFFLOADED:
3066                 if (cm_node->send_entry)
3067                         ibdev_dbg(&cm_node->iwdev->ibdev,
3068                                   "CM: CM send_entry in OFFLOADED state\n");
3069                 irdma_rem_ref_cm_node(cm_node);
3070                 break;
3071         }
3072
3073         return 0;
3074 }
3075
3076 /**
3077  * irdma_receive_ilq - recv an ETHERNET packet, and process it
3078  * through CM
3079  * @vsi: VSI structure of dev
3080  * @rbuf: receive buffer
3081  */
3082 void irdma_receive_ilq(struct irdma_sc_vsi *vsi, struct irdma_puda_buf *rbuf)
3083 {
3084         struct irdma_cm_node *cm_node;
3085         struct irdma_cm_listener *listener;
3086         struct iphdr *iph;
3087         struct ipv6hdr *ip6h;
3088         struct tcphdr *tcph;
3089         struct irdma_cm_info cm_info = {};
3090         struct irdma_device *iwdev = vsi->back_vsi;
3091         struct irdma_cm_core *cm_core = &iwdev->cm_core;
3092         struct vlan_ethhdr *ethh;
3093         u16 vtag;
3094
3095         /* if vlan, then maclen = 18 else 14 */
3096         iph = (struct iphdr *)rbuf->iph;
3097         print_hex_dump_debug("ILQ: RECEIVE ILQ BUFFER", DUMP_PREFIX_OFFSET,
3098                              16, 8, rbuf->mem.va, rbuf->totallen, false);
3099         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
3100                 if (rbuf->vlan_valid) {
3101                         vtag = rbuf->vlan_id;
3102                         cm_info.user_pri = (vtag & VLAN_PRIO_MASK) >>
3103                                            VLAN_PRIO_SHIFT;
3104                         cm_info.vlan_id = vtag & VLAN_VID_MASK;
3105                 } else {
3106                         cm_info.vlan_id = 0xFFFF;
3107                 }
3108         } else {
3109                 ethh = rbuf->mem.va;
3110
3111                 if (ethh->h_vlan_proto == htons(ETH_P_8021Q)) {
3112                         vtag = ntohs(ethh->h_vlan_TCI);
3113                         cm_info.user_pri = (vtag & VLAN_PRIO_MASK) >>
3114                                            VLAN_PRIO_SHIFT;
3115                         cm_info.vlan_id = vtag & VLAN_VID_MASK;
3116                         ibdev_dbg(&cm_core->iwdev->ibdev,
3117                                   "CM: vlan_id=%d\n", cm_info.vlan_id);
3118                 } else {
3119                         cm_info.vlan_id = 0xFFFF;
3120                 }
3121         }
3122         tcph = (struct tcphdr *)rbuf->tcph;
3123
3124         if (rbuf->ipv4) {
3125                 cm_info.loc_addr[0] = ntohl(iph->daddr);
3126                 cm_info.rem_addr[0] = ntohl(iph->saddr);
3127                 cm_info.ipv4 = true;
3128                 cm_info.tos = iph->tos;
3129         } else {
3130                 ip6h = (struct ipv6hdr *)rbuf->iph;
3131                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3132                                     ip6h->daddr.in6_u.u6_addr32);
3133                 irdma_copy_ip_ntohl(cm_info.rem_addr,
3134                                     ip6h->saddr.in6_u.u6_addr32);
3135                 cm_info.ipv4 = false;
3136                 cm_info.tos = (ip6h->priority << 4) | (ip6h->flow_lbl[0] >> 4);
3137         }
3138         cm_info.loc_port = ntohs(tcph->dest);
3139         cm_info.rem_port = ntohs(tcph->source);
3140         cm_node = irdma_find_node(cm_core, cm_info.rem_port, cm_info.rem_addr,
3141                                   cm_info.loc_port, cm_info.loc_addr, cm_info.vlan_id);
3142
3143         if (!cm_node) {
3144                 /* Only type of packet accepted are for the
3145                  * PASSIVE open (syn only)
3146                  */
3147                 if (!tcph->syn || tcph->ack)
3148                         return;
3149
3150                 listener = irdma_find_listener(cm_core,
3151                                                cm_info.loc_addr,
3152                                                cm_info.loc_port,
3153                                                cm_info.vlan_id,
3154                                                IRDMA_CM_LISTENER_ACTIVE_STATE);
3155                 if (!listener) {
3156                         cm_info.cm_id = NULL;
3157                         ibdev_dbg(&cm_core->iwdev->ibdev,
3158                                   "CM: no listener found\n");
3159                         return;
3160                 }
3161
3162                 cm_info.cm_id = listener->cm_id;
3163                 cm_node = irdma_make_cm_node(cm_core, iwdev, &cm_info,
3164                                              listener);
3165                 if (!cm_node) {
3166                         ibdev_dbg(&cm_core->iwdev->ibdev,
3167                                   "CM: allocate node failed\n");
3168                         refcount_dec(&listener->refcnt);
3169                         return;
3170                 }
3171
3172                 if (!tcph->rst && !tcph->fin) {
3173                         cm_node->state = IRDMA_CM_STATE_LISTENING;
3174                 } else {
3175                         irdma_rem_ref_cm_node(cm_node);
3176                         return;
3177                 }
3178
3179                 refcount_inc(&cm_node->refcnt);
3180         } else if (cm_node->state == IRDMA_CM_STATE_OFFLOADED) {
3181                 irdma_rem_ref_cm_node(cm_node);
3182                 return;
3183         }
3184
3185         irdma_process_pkt(cm_node, rbuf);
3186         irdma_rem_ref_cm_node(cm_node);
3187 }
3188
3189 static int irdma_add_qh(struct irdma_cm_node *cm_node, bool active)
3190 {
3191         if (!active)
3192                 irdma_add_conn_est_qh(cm_node);
3193         return 0;
3194 }
3195
3196 static void irdma_cm_free_ah_nop(struct irdma_cm_node *cm_node)
3197 {
3198 }
3199
3200 /**
3201  * irdma_setup_cm_core - setup top level instance of a cm core
3202  * @iwdev: iwarp device structure
3203  * @rdma_ver: HW version
3204  */
3205 int irdma_setup_cm_core(struct irdma_device *iwdev, u8 rdma_ver)
3206 {
3207         struct irdma_cm_core *cm_core = &iwdev->cm_core;
3208
3209         cm_core->iwdev = iwdev;
3210         cm_core->dev = &iwdev->rf->sc_dev;
3211
3212         /* Handles CM event work items send to Iwarp core */
3213         cm_core->event_wq = alloc_ordered_workqueue("iwarp-event-wq", 0);
3214         if (!cm_core->event_wq)
3215                 return -ENOMEM;
3216
3217         INIT_LIST_HEAD(&cm_core->listen_list);
3218
3219         timer_setup(&cm_core->tcp_timer, irdma_cm_timer_tick, 0);
3220
3221         spin_lock_init(&cm_core->ht_lock);
3222         spin_lock_init(&cm_core->listen_list_lock);
3223         spin_lock_init(&cm_core->apbvt_lock);
3224         switch (rdma_ver) {
3225         case IRDMA_GEN_1:
3226                 cm_core->form_cm_frame = irdma_form_uda_cm_frame;
3227                 cm_core->cm_create_ah = irdma_add_qh;
3228                 cm_core->cm_free_ah = irdma_cm_free_ah_nop;
3229                 break;
3230         case IRDMA_GEN_2:
3231         default:
3232                 cm_core->form_cm_frame = irdma_form_ah_cm_frame;
3233                 cm_core->cm_create_ah = irdma_cm_create_ah;
3234                 cm_core->cm_free_ah = irdma_cm_free_ah;
3235         }
3236
3237         return 0;
3238 }
3239
3240 /**
3241  * irdma_cleanup_cm_core - deallocate a top level instance of a
3242  * cm core
3243  * @cm_core: cm's core
3244  */
3245 void irdma_cleanup_cm_core(struct irdma_cm_core *cm_core)
3246 {
3247         if (!cm_core)
3248                 return;
3249
3250         del_timer_sync(&cm_core->tcp_timer);
3251
3252         destroy_workqueue(cm_core->event_wq);
3253         cm_core->dev->ws_reset(&cm_core->iwdev->vsi);
3254 }
3255
3256 /**
3257  * irdma_init_tcp_ctx - setup qp context
3258  * @cm_node: connection's node
3259  * @tcp_info: offload info for tcp
3260  * @iwqp: associate qp for the connection
3261  */
3262 static void irdma_init_tcp_ctx(struct irdma_cm_node *cm_node,
3263                                struct irdma_tcp_offload_info *tcp_info,
3264                                struct irdma_qp *iwqp)
3265 {
3266         tcp_info->ipv4 = cm_node->ipv4;
3267         tcp_info->drop_ooo_seg = !iwqp->iwdev->iw_ooo;
3268         tcp_info->wscale = true;
3269         tcp_info->ignore_tcp_opt = true;
3270         tcp_info->ignore_tcp_uns_opt = true;
3271         tcp_info->no_nagle = false;
3272
3273         tcp_info->ttl = IRDMA_DEFAULT_TTL;
3274         tcp_info->rtt_var = IRDMA_DEFAULT_RTT_VAR;
3275         tcp_info->ss_thresh = IRDMA_DEFAULT_SS_THRESH;
3276         tcp_info->rexmit_thresh = IRDMA_DEFAULT_REXMIT_THRESH;
3277
3278         tcp_info->tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
3279         tcp_info->snd_wscale = cm_node->tcp_cntxt.snd_wscale;
3280         tcp_info->rcv_wscale = cm_node->tcp_cntxt.rcv_wscale;
3281
3282         tcp_info->snd_nxt = cm_node->tcp_cntxt.loc_seq_num;
3283         tcp_info->snd_wnd = cm_node->tcp_cntxt.snd_wnd;
3284         tcp_info->rcv_nxt = cm_node->tcp_cntxt.rcv_nxt;
3285         tcp_info->snd_max = cm_node->tcp_cntxt.loc_seq_num;
3286
3287         tcp_info->snd_una = cm_node->tcp_cntxt.loc_seq_num;
3288         tcp_info->cwnd = 2 * cm_node->tcp_cntxt.mss;
3289         tcp_info->snd_wl1 = cm_node->tcp_cntxt.rcv_nxt;
3290         tcp_info->snd_wl2 = cm_node->tcp_cntxt.loc_seq_num;
3291         tcp_info->max_snd_window = cm_node->tcp_cntxt.max_snd_wnd;
3292         tcp_info->rcv_wnd = cm_node->tcp_cntxt.rcv_wnd
3293                             << cm_node->tcp_cntxt.rcv_wscale;
3294
3295         tcp_info->flow_label = 0;
3296         tcp_info->snd_mss = (u32)cm_node->tcp_cntxt.mss;
3297         tcp_info->tos = cm_node->tos;
3298         if (cm_node->vlan_id < VLAN_N_VID) {
3299                 tcp_info->insert_vlan_tag = true;
3300                 tcp_info->vlan_tag = cm_node->vlan_id;
3301                 tcp_info->vlan_tag |= cm_node->user_pri << VLAN_PRIO_SHIFT;
3302         }
3303         if (cm_node->ipv4) {
3304                 tcp_info->src_port = cm_node->loc_port;
3305                 tcp_info->dst_port = cm_node->rem_port;
3306
3307                 tcp_info->dest_ip_addr[3] = cm_node->rem_addr[0];
3308                 tcp_info->local_ipaddr[3] = cm_node->loc_addr[0];
3309                 tcp_info->arp_idx = (u16)irdma_arp_table(iwqp->iwdev->rf,
3310                                                          &tcp_info->dest_ip_addr[3],
3311                                                          true, NULL,
3312                                                          IRDMA_ARP_RESOLVE);
3313         } else {
3314                 tcp_info->src_port = cm_node->loc_port;
3315                 tcp_info->dst_port = cm_node->rem_port;
3316                 memcpy(tcp_info->dest_ip_addr, cm_node->rem_addr,
3317                        sizeof(tcp_info->dest_ip_addr));
3318                 memcpy(tcp_info->local_ipaddr, cm_node->loc_addr,
3319                        sizeof(tcp_info->local_ipaddr));
3320
3321                 tcp_info->arp_idx = (u16)irdma_arp_table(iwqp->iwdev->rf,
3322                                                          &tcp_info->dest_ip_addr[0],
3323                                                          false, NULL,
3324                                                          IRDMA_ARP_RESOLVE);
3325         }
3326 }
3327
3328 /**
3329  * irdma_cm_init_tsa_conn - setup qp for RTS
3330  * @iwqp: associate qp for the connection
3331  * @cm_node: connection's node
3332  */
3333 static void irdma_cm_init_tsa_conn(struct irdma_qp *iwqp,
3334                                    struct irdma_cm_node *cm_node)
3335 {
3336         struct irdma_iwarp_offload_info *iwarp_info;
3337         struct irdma_qp_host_ctx_info *ctx_info;
3338
3339         iwarp_info = &iwqp->iwarp_info;
3340         ctx_info = &iwqp->ctx_info;
3341
3342         ctx_info->tcp_info = &iwqp->tcp_info;
3343         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
3344         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
3345
3346         iwarp_info->ord_size = cm_node->ord_size;
3347         iwarp_info->ird_size = cm_node->ird_size;
3348         iwarp_info->rd_en = true;
3349         iwarp_info->rdmap_ver = 1;
3350         iwarp_info->ddp_ver = 1;
3351         iwarp_info->pd_id = iwqp->iwpd->sc_pd.pd_id;
3352
3353         ctx_info->tcp_info_valid = true;
3354         ctx_info->iwarp_info_valid = true;
3355         ctx_info->user_pri = cm_node->user_pri;
3356
3357         irdma_init_tcp_ctx(cm_node, &iwqp->tcp_info, iwqp);
3358         if (cm_node->snd_mark_en) {
3359                 iwarp_info->snd_mark_en = true;
3360                 iwarp_info->snd_mark_offset = (iwqp->tcp_info.snd_nxt & SNDMARKER_SEQNMASK) +
3361                                                cm_node->lsmm_size;
3362         }
3363
3364         cm_node->state = IRDMA_CM_STATE_OFFLOADED;
3365         iwqp->tcp_info.tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
3366         iwqp->tcp_info.src_mac_addr_idx = iwqp->iwdev->mac_ip_table_idx;
3367
3368         if (cm_node->rcv_mark_en) {
3369                 iwarp_info->rcv_mark_en = true;
3370                 iwarp_info->align_hdrs = true;
3371         }
3372
3373         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
3374
3375         /* once tcp_info is set, no need to do it again */
3376         ctx_info->tcp_info_valid = false;
3377         ctx_info->iwarp_info_valid = false;
3378 }
3379
3380 /**
3381  * irdma_cm_disconn - when a connection is being closed
3382  * @iwqp: associated qp for the connection
3383  */
3384 void irdma_cm_disconn(struct irdma_qp *iwqp)
3385 {
3386         struct irdma_device *iwdev = iwqp->iwdev;
3387         struct disconn_work *work;
3388         unsigned long flags;
3389
3390         work = kzalloc(sizeof(*work), GFP_ATOMIC);
3391         if (!work)
3392                 return;
3393
3394         spin_lock_irqsave(&iwdev->rf->qptable_lock, flags);
3395         if (!iwdev->rf->qp_table[iwqp->ibqp.qp_num]) {
3396                 spin_unlock_irqrestore(&iwdev->rf->qptable_lock, flags);
3397                 ibdev_dbg(&iwdev->ibdev,
3398                           "CM: qp_id %d is already freed\n",
3399                           iwqp->ibqp.qp_num);
3400                 kfree(work);
3401                 return;
3402         }
3403         irdma_qp_add_ref(&iwqp->ibqp);
3404         spin_unlock_irqrestore(&iwdev->rf->qptable_lock, flags);
3405
3406         work->iwqp = iwqp;
3407         INIT_WORK(&work->work, irdma_disconnect_worker);
3408         queue_work(iwdev->cleanup_wq, &work->work);
3409 }
3410
3411 /**
3412  * irdma_qp_disconnect - free qp and close cm
3413  * @iwqp: associate qp for the connection
3414  */
3415 static void irdma_qp_disconnect(struct irdma_qp *iwqp)
3416 {
3417         struct irdma_device *iwdev = iwqp->iwdev;
3418
3419         iwqp->active_conn = 0;
3420         /* close the CM node down if it is still active */
3421         ibdev_dbg(&iwdev->ibdev, "CM: Call close API\n");
3422         irdma_cm_close(iwqp->cm_node);
3423 }
3424
3425 /**
3426  * irdma_cm_disconn_true - called by worker thread to disconnect qp
3427  * @iwqp: associate qp for the connection
3428  */
3429 static void irdma_cm_disconn_true(struct irdma_qp *iwqp)
3430 {
3431         struct iw_cm_id *cm_id;
3432         struct irdma_device *iwdev;
3433         struct irdma_sc_qp *qp = &iwqp->sc_qp;
3434         u16 last_ae;
3435         u8 original_hw_tcp_state;
3436         u8 original_ibqp_state;
3437         int disconn_status = 0;
3438         int issue_disconn = 0;
3439         int issue_close = 0;
3440         int issue_flush = 0;
3441         unsigned long flags;
3442         int err;
3443
3444         iwdev = iwqp->iwdev;
3445         spin_lock_irqsave(&iwqp->lock, flags);
3446         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
3447                 struct ib_qp_attr attr;
3448
3449                 if (iwqp->flush_issued || iwqp->sc_qp.qp_uk.destroy_pending) {
3450                         spin_unlock_irqrestore(&iwqp->lock, flags);
3451                         return;
3452                 }
3453
3454                 spin_unlock_irqrestore(&iwqp->lock, flags);
3455
3456                 attr.qp_state = IB_QPS_ERR;
3457                 irdma_modify_qp_roce(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
3458                 irdma_ib_qp_event(iwqp, qp->event_type);
3459                 return;
3460         }
3461
3462         cm_id = iwqp->cm_id;
3463         original_hw_tcp_state = iwqp->hw_tcp_state;
3464         original_ibqp_state = iwqp->ibqp_state;
3465         last_ae = iwqp->last_aeq;
3466
3467         if (qp->term_flags) {
3468                 issue_disconn = 1;
3469                 issue_close = 1;
3470                 iwqp->cm_id = NULL;
3471                 irdma_terminate_del_timer(qp);
3472                 if (!iwqp->flush_issued) {
3473                         iwqp->flush_issued = 1;
3474                         issue_flush = 1;
3475                 }
3476         } else if ((original_hw_tcp_state == IRDMA_TCP_STATE_CLOSE_WAIT) ||
3477                    ((original_ibqp_state == IB_QPS_RTS) &&
3478                     (last_ae == IRDMA_AE_LLP_CONNECTION_RESET))) {
3479                 issue_disconn = 1;
3480                 if (last_ae == IRDMA_AE_LLP_CONNECTION_RESET)
3481                         disconn_status = -ECONNRESET;
3482         }
3483
3484         if (original_hw_tcp_state == IRDMA_TCP_STATE_CLOSED ||
3485             original_hw_tcp_state == IRDMA_TCP_STATE_TIME_WAIT ||
3486             last_ae == IRDMA_AE_RDMAP_ROE_BAD_LLP_CLOSE ||
3487             last_ae == IRDMA_AE_BAD_CLOSE ||
3488             last_ae == IRDMA_AE_LLP_CONNECTION_RESET || iwdev->rf->reset || !cm_id) {
3489                 issue_close = 1;
3490                 iwqp->cm_id = NULL;
3491                 qp->term_flags = 0;
3492                 if (!iwqp->flush_issued) {
3493                         iwqp->flush_issued = 1;
3494                         issue_flush = 1;
3495                 }
3496         }
3497
3498         spin_unlock_irqrestore(&iwqp->lock, flags);
3499         if (issue_flush && !iwqp->sc_qp.qp_uk.destroy_pending) {
3500                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_FLUSH_RQ |
3501                                  IRDMA_FLUSH_WAIT);
3502
3503                 if (qp->term_flags)
3504                         irdma_ib_qp_event(iwqp, qp->event_type);
3505         }
3506
3507         if (!cm_id || !cm_id->event_handler)
3508                 return;
3509
3510         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
3511         if (!iwqp->cm_node) {
3512                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
3513                 return;
3514         }
3515         refcount_inc(&iwqp->cm_node->refcnt);
3516
3517         spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
3518
3519         if (issue_disconn) {
3520                 err = irdma_send_cm_event(iwqp->cm_node, cm_id,
3521                                           IW_CM_EVENT_DISCONNECT,
3522                                           disconn_status);
3523                 if (err)
3524                         ibdev_dbg(&iwdev->ibdev,
3525                                   "CM: disconnect event failed: - cm_id = %p\n",
3526                                   cm_id);
3527         }
3528         if (issue_close) {
3529                 cm_id->provider_data = iwqp;
3530                 err = irdma_send_cm_event(iwqp->cm_node, cm_id,
3531                                           IW_CM_EVENT_CLOSE, 0);
3532                 if (err)
3533                         ibdev_dbg(&iwdev->ibdev,
3534                                   "CM: close event failed: - cm_id = %p\n",
3535                                   cm_id);
3536                 irdma_qp_disconnect(iwqp);
3537         }
3538         irdma_rem_ref_cm_node(iwqp->cm_node);
3539 }
3540
3541 /**
3542  * irdma_disconnect_worker - worker for connection close
3543  * @work: points or disconn structure
3544  */
3545 static void irdma_disconnect_worker(struct work_struct *work)
3546 {
3547         struct disconn_work *dwork = container_of(work, struct disconn_work, work);
3548         struct irdma_qp *iwqp = dwork->iwqp;
3549
3550         kfree(dwork);
3551         irdma_cm_disconn_true(iwqp);
3552         irdma_qp_rem_ref(&iwqp->ibqp);
3553 }
3554
3555 /**
3556  * irdma_free_lsmm_rsrc - free lsmm memory and deregister
3557  * @iwqp: associate qp for the connection
3558  */
3559 void irdma_free_lsmm_rsrc(struct irdma_qp *iwqp)
3560 {
3561         struct irdma_device *iwdev;
3562
3563         iwdev = iwqp->iwdev;
3564
3565         if (iwqp->ietf_mem.va) {
3566                 if (iwqp->lsmm_mr)
3567                         iwdev->ibdev.ops.dereg_mr(iwqp->lsmm_mr, NULL);
3568                 dma_free_coherent(iwdev->rf->sc_dev.hw->device,
3569                                   iwqp->ietf_mem.size, iwqp->ietf_mem.va,
3570                                   iwqp->ietf_mem.pa);
3571                 iwqp->ietf_mem.va = NULL;
3572                 iwqp->ietf_mem.va = NULL;
3573         }
3574 }
3575
3576 /**
3577  * irdma_accept - registered call for connection to be accepted
3578  * @cm_id: cm information for passive connection
3579  * @conn_param: accpet parameters
3580  */
3581 int irdma_accept(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
3582 {
3583         struct ib_qp *ibqp;
3584         struct irdma_qp *iwqp;
3585         struct irdma_device *iwdev;
3586         struct irdma_sc_dev *dev;
3587         struct irdma_cm_node *cm_node;
3588         struct ib_qp_attr attr = {};
3589         int passive_state;
3590         struct ib_mr *ibmr;
3591         struct irdma_pd *iwpd;
3592         u16 buf_len = 0;
3593         struct irdma_kmem_info accept;
3594         u64 tagged_offset;
3595         int wait_ret;
3596         int ret = 0;
3597
3598         ibqp = irdma_get_qp(cm_id->device, conn_param->qpn);
3599         if (!ibqp)
3600                 return -EINVAL;
3601
3602         iwqp = to_iwqp(ibqp);
3603         iwdev = iwqp->iwdev;
3604         dev = &iwdev->rf->sc_dev;
3605         cm_node = cm_id->provider_data;
3606
3607         if (((struct sockaddr_in *)&cm_id->local_addr)->sin_family == AF_INET) {
3608                 cm_node->ipv4 = true;
3609                 cm_node->vlan_id = irdma_get_vlan_ipv4(cm_node->loc_addr);
3610         } else {
3611                 cm_node->ipv4 = false;
3612                 irdma_netdev_vlan_ipv6(cm_node->loc_addr, &cm_node->vlan_id,
3613                                        NULL);
3614         }
3615         ibdev_dbg(&iwdev->ibdev, "CM: Accept vlan_id=%d\n",
3616                   cm_node->vlan_id);
3617
3618         trace_irdma_accept(cm_node, 0, NULL);
3619
3620         if (cm_node->state == IRDMA_CM_STATE_LISTENER_DESTROYED) {
3621                 ret = -EINVAL;
3622                 goto error;
3623         }
3624
3625         passive_state = atomic_add_return(1, &cm_node->passive_state);
3626         if (passive_state == IRDMA_SEND_RESET_EVENT) {
3627                 ret = -ECONNRESET;
3628                 goto error;
3629         }
3630
3631         buf_len = conn_param->private_data_len + IRDMA_MAX_IETF_SIZE;
3632         iwqp->ietf_mem.size = ALIGN(buf_len, 1);
3633         iwqp->ietf_mem.va = dma_alloc_coherent(dev->hw->device,
3634                                                iwqp->ietf_mem.size,
3635                                                &iwqp->ietf_mem.pa, GFP_KERNEL);
3636         if (!iwqp->ietf_mem.va) {
3637                 ret = -ENOMEM;
3638                 goto error;
3639         }
3640
3641         cm_node->pdata.size = conn_param->private_data_len;
3642         accept.addr = iwqp->ietf_mem.va;
3643         accept.size = irdma_cm_build_mpa_frame(cm_node, &accept, MPA_KEY_REPLY);
3644         memcpy((u8 *)accept.addr + accept.size, conn_param->private_data,
3645                conn_param->private_data_len);
3646
3647         if (cm_node->dev->ws_add(iwqp->sc_qp.vsi, cm_node->user_pri)) {
3648                 ret = -ENOMEM;
3649                 goto error;
3650         }
3651         iwqp->sc_qp.user_pri = cm_node->user_pri;
3652         irdma_qp_add_qos(&iwqp->sc_qp);
3653         /* setup our first outgoing iWarp send WQE (the IETF frame response) */
3654         iwpd = iwqp->iwpd;
3655         tagged_offset = (uintptr_t)iwqp->ietf_mem.va;
3656         ibmr = irdma_reg_phys_mr(&iwpd->ibpd, iwqp->ietf_mem.pa, buf_len,
3657                                  IB_ACCESS_LOCAL_WRITE, &tagged_offset);
3658         if (IS_ERR(ibmr)) {
3659                 ret = -ENOMEM;
3660                 goto error;
3661         }
3662
3663         ibmr->pd = &iwpd->ibpd;
3664         ibmr->device = iwpd->ibpd.device;
3665         iwqp->lsmm_mr = ibmr;
3666         if (iwqp->page)
3667                 iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
3668
3669         cm_node->lsmm_size = accept.size + conn_param->private_data_len;
3670         irdma_sc_send_lsmm(&iwqp->sc_qp, iwqp->ietf_mem.va, cm_node->lsmm_size,
3671                            ibmr->lkey);
3672
3673         if (iwqp->page)
3674                 kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
3675
3676         iwqp->cm_id = cm_id;
3677         cm_node->cm_id = cm_id;
3678
3679         cm_id->provider_data = iwqp;
3680         iwqp->active_conn = 0;
3681         iwqp->cm_node = cm_node;
3682         cm_node->iwqp = iwqp;
3683         irdma_cm_init_tsa_conn(iwqp, cm_node);
3684         irdma_qp_add_ref(&iwqp->ibqp);
3685         cm_id->add_ref(cm_id);
3686
3687         attr.qp_state = IB_QPS_RTS;
3688         cm_node->qhash_set = false;
3689         cm_node->cm_core->cm_free_ah(cm_node);
3690
3691         irdma_modify_qp(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
3692         if (dev->hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_RTS_AE) {
3693                 wait_ret = wait_event_interruptible_timeout(iwqp->waitq,
3694                                                             iwqp->rts_ae_rcvd,
3695                                                             IRDMA_MAX_TIMEOUT);
3696                 if (!wait_ret) {
3697                         ibdev_dbg(&iwdev->ibdev,
3698                                   "CM: Slow Connection: cm_node=%p, loc_port=%d, rem_port=%d, cm_id=%p\n",
3699                                   cm_node, cm_node->loc_port,
3700                                   cm_node->rem_port, cm_node->cm_id);
3701                         ret = -ECONNRESET;
3702                         goto error;
3703                 }
3704         }
3705
3706         irdma_send_cm_event(cm_node, cm_id, IW_CM_EVENT_ESTABLISHED, 0);
3707         cm_node->accelerated = true;
3708         complete(&cm_node->establish_comp);
3709
3710         if (cm_node->accept_pend) {
3711                 atomic_dec(&cm_node->listener->pend_accepts_cnt);
3712                 cm_node->accept_pend = 0;
3713         }
3714
3715         ibdev_dbg(&iwdev->ibdev,
3716                   "CM: rem_port=0x%04x, loc_port=0x%04x rem_addr=%pI4 loc_addr=%pI4 cm_node=%p cm_id=%p qp_id = %d\n\n",
3717                   cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr,
3718                   cm_node->loc_addr, cm_node, cm_id, ibqp->qp_num);
3719         cm_node->cm_core->stats_accepts++;
3720
3721         return 0;
3722 error:
3723         irdma_free_lsmm_rsrc(iwqp);
3724         irdma_rem_ref_cm_node(cm_node);
3725
3726         return ret;
3727 }
3728
3729 /**
3730  * irdma_reject - registered call for connection to be rejected
3731  * @cm_id: cm information for passive connection
3732  * @pdata: private data to be sent
3733  * @pdata_len: private data length
3734  */
3735 int irdma_reject(struct iw_cm_id *cm_id, const void *pdata, u8 pdata_len)
3736 {
3737         struct irdma_device *iwdev;
3738         struct irdma_cm_node *cm_node;
3739
3740         cm_node = cm_id->provider_data;
3741         cm_node->pdata.size = pdata_len;
3742
3743         trace_irdma_reject(cm_node, 0, NULL);
3744
3745         iwdev = to_iwdev(cm_id->device);
3746         if (!iwdev)
3747                 return -EINVAL;
3748
3749         cm_node->cm_core->stats_rejects++;
3750
3751         if (pdata_len + sizeof(struct ietf_mpa_v2) > IRDMA_MAX_CM_BUF)
3752                 return -EINVAL;
3753
3754         return irdma_cm_reject(cm_node, pdata, pdata_len);
3755 }
3756
3757 /**
3758  * irdma_connect - registered call for connection to be established
3759  * @cm_id: cm information for passive connection
3760  * @conn_param: Information about the connection
3761  */
3762 int irdma_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
3763 {
3764         struct ib_qp *ibqp;
3765         struct irdma_qp *iwqp;
3766         struct irdma_device *iwdev;
3767         struct irdma_cm_node *cm_node;
3768         struct irdma_cm_info cm_info;
3769         struct sockaddr_in *laddr;
3770         struct sockaddr_in *raddr;
3771         struct sockaddr_in6 *laddr6;
3772         struct sockaddr_in6 *raddr6;
3773         int ret = 0;
3774
3775         ibqp = irdma_get_qp(cm_id->device, conn_param->qpn);
3776         if (!ibqp)
3777                 return -EINVAL;
3778         iwqp = to_iwqp(ibqp);
3779         if (!iwqp)
3780                 return -EINVAL;
3781         iwdev = iwqp->iwdev;
3782         if (!iwdev)
3783                 return -EINVAL;
3784
3785         laddr = (struct sockaddr_in *)&cm_id->m_local_addr;
3786         raddr = (struct sockaddr_in *)&cm_id->m_remote_addr;
3787         laddr6 = (struct sockaddr_in6 *)&cm_id->m_local_addr;
3788         raddr6 = (struct sockaddr_in6 *)&cm_id->m_remote_addr;
3789
3790         if (!(laddr->sin_port) || !(raddr->sin_port))
3791                 return -EINVAL;
3792
3793         iwqp->active_conn = 1;
3794         iwqp->cm_id = NULL;
3795         cm_id->provider_data = iwqp;
3796
3797         /* set up the connection params for the node */
3798         if (cm_id->remote_addr.ss_family == AF_INET) {
3799                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV4)
3800                         return -EINVAL;
3801
3802                 cm_info.ipv4 = true;
3803                 memset(cm_info.loc_addr, 0, sizeof(cm_info.loc_addr));
3804                 memset(cm_info.rem_addr, 0, sizeof(cm_info.rem_addr));
3805                 cm_info.loc_addr[0] = ntohl(laddr->sin_addr.s_addr);
3806                 cm_info.rem_addr[0] = ntohl(raddr->sin_addr.s_addr);
3807                 cm_info.loc_port = ntohs(laddr->sin_port);
3808                 cm_info.rem_port = ntohs(raddr->sin_port);
3809                 cm_info.vlan_id = irdma_get_vlan_ipv4(cm_info.loc_addr);
3810         } else {
3811                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV6)
3812                         return -EINVAL;
3813
3814                 cm_info.ipv4 = false;
3815                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3816                                     laddr6->sin6_addr.in6_u.u6_addr32);
3817                 irdma_copy_ip_ntohl(cm_info.rem_addr,
3818                                     raddr6->sin6_addr.in6_u.u6_addr32);
3819                 cm_info.loc_port = ntohs(laddr6->sin6_port);
3820                 cm_info.rem_port = ntohs(raddr6->sin6_port);
3821                 irdma_netdev_vlan_ipv6(cm_info.loc_addr, &cm_info.vlan_id,
3822                                        NULL);
3823         }
3824         cm_info.cm_id = cm_id;
3825         cm_info.qh_qpid = iwdev->vsi.ilq->qp_id;
3826         cm_info.tos = cm_id->tos;
3827         if (iwdev->vsi.dscp_mode)
3828                 cm_info.user_pri =
3829                         iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(cm_info.tos)];
3830         else
3831                 cm_info.user_pri = rt_tos2priority(cm_id->tos);
3832
3833         if (iwqp->sc_qp.dev->ws_add(iwqp->sc_qp.vsi, cm_info.user_pri))
3834                 return -ENOMEM;
3835         iwqp->sc_qp.user_pri = cm_info.user_pri;
3836         irdma_qp_add_qos(&iwqp->sc_qp);
3837         ibdev_dbg(&iwdev->ibdev, "DCB: TOS:[%d] UP:[%d]\n", cm_id->tos,
3838                   cm_info.user_pri);
3839
3840         trace_irdma_dcb_tos(iwdev, cm_id->tos, cm_info.user_pri);
3841
3842         ret = irdma_create_cm_node(&iwdev->cm_core, iwdev, conn_param, &cm_info,
3843                                    &cm_node);
3844         if (ret)
3845                 return ret;
3846         ret = cm_node->cm_core->cm_create_ah(cm_node, true);
3847         if (ret)
3848                 goto err;
3849         if (irdma_manage_qhash(iwdev, &cm_info,
3850                                IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
3851                                IRDMA_QHASH_MANAGE_TYPE_ADD, NULL, true)) {
3852                 ret = -EINVAL;
3853                 goto err;
3854         }
3855         cm_node->qhash_set = true;
3856
3857         cm_node->apbvt_entry = irdma_add_apbvt(iwdev, cm_info.loc_port);
3858         if (!cm_node->apbvt_entry) {
3859                 ret = -EINVAL;
3860                 goto err;
3861         }
3862
3863         cm_node->apbvt_set = true;
3864         iwqp->cm_node = cm_node;
3865         cm_node->iwqp = iwqp;
3866         iwqp->cm_id = cm_id;
3867         irdma_qp_add_ref(&iwqp->ibqp);
3868         cm_id->add_ref(cm_id);
3869
3870         if (cm_node->state != IRDMA_CM_STATE_OFFLOADED) {
3871                 cm_node->state = IRDMA_CM_STATE_SYN_SENT;
3872                 ret = irdma_send_syn(cm_node, 0);
3873                 if (ret)
3874                         goto err;
3875         }
3876
3877         ibdev_dbg(&iwdev->ibdev,
3878                   "CM: rem_port=0x%04x, loc_port=0x%04x rem_addr=%pI4 loc_addr=%pI4 cm_node=%p cm_id=%p qp_id = %d\n\n",
3879                   cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr,
3880                   cm_node->loc_addr, cm_node, cm_id, ibqp->qp_num);
3881
3882         trace_irdma_connect(cm_node, 0, NULL);
3883
3884         return 0;
3885
3886 err:
3887         if (cm_info.ipv4)
3888                 ibdev_dbg(&iwdev->ibdev,
3889                           "CM: connect() FAILED: dest addr=%pI4",
3890                           cm_info.rem_addr);
3891         else
3892                 ibdev_dbg(&iwdev->ibdev,
3893                           "CM: connect() FAILED: dest addr=%pI6",
3894                           cm_info.rem_addr);
3895         irdma_rem_ref_cm_node(cm_node);
3896         iwdev->cm_core.stats_connect_errs++;
3897
3898         return ret;
3899 }
3900
3901 /**
3902  * irdma_create_listen - registered call creating listener
3903  * @cm_id: cm information for passive connection
3904  * @backlog: to max accept pending count
3905  */
3906 int irdma_create_listen(struct iw_cm_id *cm_id, int backlog)
3907 {
3908         struct irdma_device *iwdev;
3909         struct irdma_cm_listener *cm_listen_node;
3910         struct irdma_cm_info cm_info = {};
3911         struct sockaddr_in *laddr;
3912         struct sockaddr_in6 *laddr6;
3913         bool wildcard = false;
3914         int err;
3915
3916         iwdev = to_iwdev(cm_id->device);
3917         if (!iwdev)
3918                 return -EINVAL;
3919
3920         laddr = (struct sockaddr_in *)&cm_id->m_local_addr;
3921         laddr6 = (struct sockaddr_in6 *)&cm_id->m_local_addr;
3922         cm_info.qh_qpid = iwdev->vsi.ilq->qp_id;
3923
3924         if (laddr->sin_family == AF_INET) {
3925                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV4)
3926                         return -EINVAL;
3927
3928                 cm_info.ipv4 = true;
3929                 cm_info.loc_addr[0] = ntohl(laddr->sin_addr.s_addr);
3930                 cm_info.loc_port = ntohs(laddr->sin_port);
3931
3932                 if (laddr->sin_addr.s_addr != htonl(INADDR_ANY)) {
3933                         cm_info.vlan_id = irdma_get_vlan_ipv4(cm_info.loc_addr);
3934                 } else {
3935                         cm_info.vlan_id = 0xFFFF;
3936                         wildcard = true;
3937                 }
3938         } else {
3939                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV6)
3940                         return -EINVAL;
3941
3942                 cm_info.ipv4 = false;
3943                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3944                                     laddr6->sin6_addr.in6_u.u6_addr32);
3945                 cm_info.loc_port = ntohs(laddr6->sin6_port);
3946                 if (ipv6_addr_type(&laddr6->sin6_addr) != IPV6_ADDR_ANY) {
3947                         irdma_netdev_vlan_ipv6(cm_info.loc_addr,
3948                                                &cm_info.vlan_id, NULL);
3949                 } else {
3950                         cm_info.vlan_id = 0xFFFF;
3951                         wildcard = true;
3952                 }
3953         }
3954
3955         if (cm_info.vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
3956                 cm_info.vlan_id = 0;
3957         cm_info.backlog = backlog;
3958         cm_info.cm_id = cm_id;
3959
3960         trace_irdma_create_listen(iwdev, &cm_info);
3961
3962         cm_listen_node = irdma_make_listen_node(&iwdev->cm_core, iwdev,
3963                                                 &cm_info);
3964         if (!cm_listen_node) {
3965                 ibdev_dbg(&iwdev->ibdev,
3966                           "CM: cm_listen_node == NULL\n");
3967                 return -ENOMEM;
3968         }
3969
3970         cm_id->provider_data = cm_listen_node;
3971
3972         cm_listen_node->tos = cm_id->tos;
3973         if (iwdev->vsi.dscp_mode)
3974                 cm_listen_node->user_pri =
3975                         iwdev->vsi.dscp_map[irdma_tos2dscp(cm_id->tos)];
3976         else
3977                 cm_listen_node->user_pri = rt_tos2priority(cm_id->tos);
3978         cm_info.user_pri = cm_listen_node->user_pri;
3979         if (!cm_listen_node->reused_node) {
3980                 if (wildcard) {
3981                         err = irdma_add_mqh(iwdev, &cm_info, cm_listen_node);
3982                         if (err)
3983                                 goto error;
3984                 } else {
3985                         err = irdma_manage_qhash(iwdev, &cm_info,
3986                                                  IRDMA_QHASH_TYPE_TCP_SYN,
3987                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
3988                                                  NULL, true);
3989                         if (err)
3990                                 goto error;
3991
3992                         cm_listen_node->qhash_set = true;
3993                 }
3994
3995                 cm_listen_node->apbvt_entry = irdma_add_apbvt(iwdev,
3996                                                               cm_info.loc_port);
3997                 if (!cm_listen_node->apbvt_entry)
3998                         goto error;
3999         }
4000         cm_id->add_ref(cm_id);
4001         cm_listen_node->cm_core->stats_listen_created++;
4002         ibdev_dbg(&iwdev->ibdev,
4003                   "CM: loc_port=0x%04x loc_addr=%pI4 cm_listen_node=%p cm_id=%p qhash_set=%d vlan_id=%d\n",
4004                   cm_listen_node->loc_port, cm_listen_node->loc_addr,
4005                   cm_listen_node, cm_listen_node->cm_id,
4006                   cm_listen_node->qhash_set, cm_listen_node->vlan_id);
4007
4008         return 0;
4009
4010 error:
4011
4012         irdma_cm_del_listen(&iwdev->cm_core, cm_listen_node, false);
4013
4014         return -EINVAL;
4015 }
4016
4017 /**
4018  * irdma_destroy_listen - registered call to destroy listener
4019  * @cm_id: cm information for passive connection
4020  */
4021 int irdma_destroy_listen(struct iw_cm_id *cm_id)
4022 {
4023         struct irdma_device *iwdev;
4024
4025         iwdev = to_iwdev(cm_id->device);
4026         if (cm_id->provider_data)
4027                 irdma_cm_del_listen(&iwdev->cm_core, cm_id->provider_data,
4028                                     true);
4029         else
4030                 ibdev_dbg(&iwdev->ibdev,
4031                           "CM: cm_id->provider_data was NULL\n");
4032
4033         cm_id->rem_ref(cm_id);
4034
4035         return 0;
4036 }
4037
4038 /**
4039  * irdma_teardown_list_prep - add conn nodes slated for tear down to list
4040  * @cm_core: cm's core
4041  * @teardown_list: a list to which cm_node will be selected
4042  * @ipaddr: pointer to ip address
4043  * @nfo: pointer to cm_info structure instance
4044  * @disconnect_all: flag indicating disconnect all QPs
4045  */
4046 static void irdma_teardown_list_prep(struct irdma_cm_core *cm_core,
4047                                      struct list_head *teardown_list,
4048                                      u32 *ipaddr,
4049                                      struct irdma_cm_info *nfo,
4050                                      bool disconnect_all)
4051 {
4052         struct irdma_cm_node *cm_node;
4053         int bkt;
4054
4055         hash_for_each_rcu(cm_core->cm_hash_tbl, bkt, cm_node, list) {
4056                 if ((disconnect_all ||
4057                      (nfo->vlan_id == cm_node->vlan_id &&
4058                       !memcmp(cm_node->loc_addr, ipaddr, nfo->ipv4 ? 4 : 16))) &&
4059                     refcount_inc_not_zero(&cm_node->refcnt))
4060                         list_add(&cm_node->teardown_entry, teardown_list);
4061         }
4062 }
4063
4064 /**
4065  * irdma_cm_event_connected - handle connected active node
4066  * @event: the info for cm_node of connection
4067  */
4068 static void irdma_cm_event_connected(struct irdma_cm_event *event)
4069 {
4070         struct irdma_qp *iwqp;
4071         struct irdma_device *iwdev;
4072         struct irdma_cm_node *cm_node;
4073         struct irdma_sc_dev *dev;
4074         struct ib_qp_attr attr = {};
4075         struct iw_cm_id *cm_id;
4076         int status;
4077         bool read0;
4078         int wait_ret = 0;
4079
4080         cm_node = event->cm_node;
4081         cm_id = cm_node->cm_id;
4082         iwqp = cm_id->provider_data;
4083         iwdev = iwqp->iwdev;
4084         dev = &iwdev->rf->sc_dev;
4085         if (iwqp->sc_qp.qp_uk.destroy_pending) {
4086                 status = -ETIMEDOUT;
4087                 goto error;
4088         }
4089
4090         irdma_cm_init_tsa_conn(iwqp, cm_node);
4091         read0 = (cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO);
4092         if (iwqp->page)
4093                 iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
4094         irdma_sc_send_rtt(&iwqp->sc_qp, read0);
4095         if (iwqp->page)
4096                 kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
4097
4098         attr.qp_state = IB_QPS_RTS;
4099         cm_node->qhash_set = false;
4100         irdma_modify_qp(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
4101         if (dev->hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_RTS_AE) {
4102                 wait_ret = wait_event_interruptible_timeout(iwqp->waitq,
4103                                                             iwqp->rts_ae_rcvd,
4104                                                             IRDMA_MAX_TIMEOUT);
4105                 if (!wait_ret)
4106                         ibdev_dbg(&iwdev->ibdev,
4107                                   "CM: Slow Connection: cm_node=%p, loc_port=%d, rem_port=%d, cm_id=%p\n",
4108                                   cm_node, cm_node->loc_port,
4109                                   cm_node->rem_port, cm_node->cm_id);
4110         }
4111
4112         irdma_send_cm_event(cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY, 0);
4113         cm_node->accelerated = true;
4114         complete(&cm_node->establish_comp);
4115         cm_node->cm_core->cm_free_ah(cm_node);
4116         return;
4117
4118 error:
4119         iwqp->cm_id = NULL;
4120         cm_id->provider_data = NULL;
4121         irdma_send_cm_event(event->cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY,
4122                             status);
4123         irdma_rem_ref_cm_node(event->cm_node);
4124 }
4125
4126 /**
4127  * irdma_cm_event_reset - handle reset
4128  * @event: the info for cm_node of connection
4129  */
4130 static void irdma_cm_event_reset(struct irdma_cm_event *event)
4131 {
4132         struct irdma_cm_node *cm_node = event->cm_node;
4133         struct iw_cm_id *cm_id = cm_node->cm_id;
4134         struct irdma_qp *iwqp;
4135
4136         if (!cm_id)
4137                 return;
4138
4139         iwqp = cm_id->provider_data;
4140         if (!iwqp)
4141                 return;
4142
4143         ibdev_dbg(&cm_node->iwdev->ibdev,
4144                   "CM: reset event %p - cm_id = %p\n", event->cm_node, cm_id);
4145         iwqp->cm_id = NULL;
4146
4147         irdma_send_cm_event(cm_node, cm_node->cm_id, IW_CM_EVENT_DISCONNECT,
4148                             -ECONNRESET);
4149         irdma_send_cm_event(cm_node, cm_node->cm_id, IW_CM_EVENT_CLOSE, 0);
4150 }
4151
4152 /**
4153  * irdma_cm_event_handler - send event to cm upper layer
4154  * @work: pointer of cm event info.
4155  */
4156 static void irdma_cm_event_handler(struct work_struct *work)
4157 {
4158         struct irdma_cm_event *event = container_of(work, struct irdma_cm_event, event_work);
4159         struct irdma_cm_node *cm_node;
4160
4161         if (!event || !event->cm_node || !event->cm_node->cm_core)
4162                 return;
4163
4164         cm_node = event->cm_node;
4165         trace_irdma_cm_event_handler(cm_node, event->type, NULL);
4166
4167         switch (event->type) {
4168         case IRDMA_CM_EVENT_MPA_REQ:
4169                 irdma_send_cm_event(cm_node, cm_node->cm_id,
4170                                     IW_CM_EVENT_CONNECT_REQUEST, 0);
4171                 break;
4172         case IRDMA_CM_EVENT_RESET:
4173                 irdma_cm_event_reset(event);
4174                 break;
4175         case IRDMA_CM_EVENT_CONNECTED:
4176                 if (!event->cm_node->cm_id ||
4177                     event->cm_node->state != IRDMA_CM_STATE_OFFLOADED)
4178                         break;
4179                 irdma_cm_event_connected(event);
4180                 break;
4181         case IRDMA_CM_EVENT_MPA_REJECT:
4182                 if (!event->cm_node->cm_id ||
4183                     cm_node->state == IRDMA_CM_STATE_OFFLOADED)
4184                         break;
4185                 irdma_send_cm_event(cm_node, cm_node->cm_id,
4186                                     IW_CM_EVENT_CONNECT_REPLY, -ECONNREFUSED);
4187                 break;
4188         case IRDMA_CM_EVENT_ABORTED:
4189                 if (!event->cm_node->cm_id ||
4190                     event->cm_node->state == IRDMA_CM_STATE_OFFLOADED)
4191                         break;
4192                 irdma_event_connect_error(event);
4193                 break;
4194         default:
4195                 ibdev_dbg(&cm_node->iwdev->ibdev,
4196                           "CM: bad event type = %d\n", event->type);
4197                 break;
4198         }
4199
4200         irdma_rem_ref_cm_node(event->cm_node);
4201         kfree(event);
4202 }
4203
4204 /**
4205  * irdma_cm_post_event - queue event request for worker thread
4206  * @event: cm node's info for up event call
4207  */
4208 static void irdma_cm_post_event(struct irdma_cm_event *event)
4209 {
4210         refcount_inc(&event->cm_node->refcnt);
4211         INIT_WORK(&event->event_work, irdma_cm_event_handler);
4212         queue_work(event->cm_node->cm_core->event_wq, &event->event_work);
4213 }
4214
4215 /**
4216  * irdma_cm_teardown_connections - teardown QPs
4217  * @iwdev: device pointer
4218  * @ipaddr: Pointer to IPv4 or IPv6 address
4219  * @nfo: Connection info
4220  * @disconnect_all: flag indicating disconnect all QPs
4221  *
4222  * teardown QPs where source or destination addr matches ip addr
4223  */
4224 void irdma_cm_teardown_connections(struct irdma_device *iwdev, u32 *ipaddr,
4225                                    struct irdma_cm_info *nfo,
4226                                    bool disconnect_all)
4227 {
4228         struct irdma_cm_core *cm_core = &iwdev->cm_core;
4229         struct list_head *list_core_temp;
4230         struct list_head *list_node;
4231         struct irdma_cm_node *cm_node;
4232         struct list_head teardown_list;
4233         struct ib_qp_attr attr;
4234
4235         INIT_LIST_HEAD(&teardown_list);
4236
4237         rcu_read_lock();
4238         irdma_teardown_list_prep(cm_core, &teardown_list, ipaddr, nfo, disconnect_all);
4239         rcu_read_unlock();
4240
4241         list_for_each_safe (list_node, list_core_temp, &teardown_list) {
4242                 cm_node = container_of(list_node, struct irdma_cm_node,
4243                                        teardown_entry);
4244                 attr.qp_state = IB_QPS_ERR;
4245                 irdma_modify_qp(&cm_node->iwqp->ibqp, &attr, IB_QP_STATE, NULL);
4246                 if (iwdev->rf->reset)
4247                         irdma_cm_disconn(cm_node->iwqp);
4248                 irdma_rem_ref_cm_node(cm_node);
4249         }
4250 }
4251
4252 /**
4253  * irdma_qhash_ctrl - enable/disable qhash for list
4254  * @iwdev: device pointer
4255  * @parent_listen_node: parent listen node
4256  * @nfo: cm info node
4257  * @ipaddr: Pointer to IPv4 or IPv6 address
4258  * @ipv4: flag indicating IPv4 when true
4259  * @ifup: flag indicating interface up when true
4260  *
4261  * Enables or disables the qhash for the node in the child
4262  * listen list that matches ipaddr. If no matching IP was found
4263  * it will allocate and add a new child listen node to the
4264  * parent listen node. The listen_list_lock is assumed to be
4265  * held when called.
4266  */
4267 static void irdma_qhash_ctrl(struct irdma_device *iwdev,
4268                              struct irdma_cm_listener *parent_listen_node,
4269                              struct irdma_cm_info *nfo, u32 *ipaddr, bool ipv4,
4270                              bool ifup)
4271 {
4272         struct list_head *child_listen_list = &parent_listen_node->child_listen_list;
4273         struct irdma_cm_listener *child_listen_node;
4274         struct list_head *pos, *tpos;
4275         bool node_allocated = false;
4276         enum irdma_quad_hash_manage_type op = ifup ?
4277                                               IRDMA_QHASH_MANAGE_TYPE_ADD :
4278                                               IRDMA_QHASH_MANAGE_TYPE_DELETE;
4279         int err;
4280
4281         list_for_each_safe (pos, tpos, child_listen_list) {
4282                 child_listen_node = list_entry(pos, struct irdma_cm_listener,
4283                                                child_listen_list);
4284                 if (!memcmp(child_listen_node->loc_addr, ipaddr, ipv4 ? 4 : 16))
4285                         goto set_qhash;
4286         }
4287
4288         /* if not found then add a child listener if interface is going up */
4289         if (!ifup)
4290                 return;
4291         child_listen_node = kmemdup(parent_listen_node,
4292                                     sizeof(*child_listen_node), GFP_ATOMIC);
4293         if (!child_listen_node)
4294                 return;
4295
4296         node_allocated = true;
4297         memcpy(child_listen_node->loc_addr, ipaddr, ipv4 ? 4 : 16);
4298
4299 set_qhash:
4300         memcpy(nfo->loc_addr, child_listen_node->loc_addr,
4301                sizeof(nfo->loc_addr));
4302         nfo->vlan_id = child_listen_node->vlan_id;
4303         err = irdma_manage_qhash(iwdev, nfo, IRDMA_QHASH_TYPE_TCP_SYN, op, NULL,
4304                                  false);
4305         if (!err) {
4306                 child_listen_node->qhash_set = ifup;
4307                 if (node_allocated)
4308                         list_add(&child_listen_node->child_listen_list,
4309                                  &parent_listen_node->child_listen_list);
4310         } else if (node_allocated) {
4311                 kfree(child_listen_node);
4312         }
4313 }
4314
4315 /**
4316  * irdma_if_notify - process an ifdown on an interface
4317  * @iwdev: device pointer
4318  * @netdev: network device structure
4319  * @ipaddr: Pointer to IPv4 or IPv6 address
4320  * @ipv4: flag indicating IPv4 when true
4321  * @ifup: flag indicating interface up when true
4322  */
4323 void irdma_if_notify(struct irdma_device *iwdev, struct net_device *netdev,
4324                      u32 *ipaddr, bool ipv4, bool ifup)
4325 {
4326         struct irdma_cm_core *cm_core = &iwdev->cm_core;
4327         unsigned long flags;
4328         struct irdma_cm_listener *listen_node;
4329         static const u32 ip_zero[4] = { 0, 0, 0, 0 };
4330         struct irdma_cm_info nfo = {};
4331         u16 vlan_id = rdma_vlan_dev_vlan_id(netdev);
4332         enum irdma_quad_hash_manage_type op = ifup ?
4333                                               IRDMA_QHASH_MANAGE_TYPE_ADD :
4334                                               IRDMA_QHASH_MANAGE_TYPE_DELETE;
4335
4336         nfo.vlan_id = vlan_id;
4337         nfo.ipv4 = ipv4;
4338         nfo.qh_qpid = 1;
4339
4340         /* Disable or enable qhash for listeners */
4341         spin_lock_irqsave(&cm_core->listen_list_lock, flags);
4342         list_for_each_entry (listen_node, &cm_core->listen_list, list) {
4343                 if (vlan_id != listen_node->vlan_id ||
4344                     (memcmp(listen_node->loc_addr, ipaddr, ipv4 ? 4 : 16) &&
4345                      memcmp(listen_node->loc_addr, ip_zero, ipv4 ? 4 : 16)))
4346                         continue;
4347
4348                 memcpy(nfo.loc_addr, listen_node->loc_addr,
4349                        sizeof(nfo.loc_addr));
4350                 nfo.loc_port = listen_node->loc_port;
4351                 nfo.user_pri = listen_node->user_pri;
4352                 if (!list_empty(&listen_node->child_listen_list)) {
4353                         irdma_qhash_ctrl(iwdev, listen_node, &nfo, ipaddr, ipv4,
4354                                          ifup);
4355                 } else if (memcmp(listen_node->loc_addr, ip_zero,
4356                                   ipv4 ? 4 : 16)) {
4357                         if (!irdma_manage_qhash(iwdev, &nfo,
4358                                                 IRDMA_QHASH_TYPE_TCP_SYN, op,
4359                                                 NULL, false))
4360                                 listen_node->qhash_set = ifup;
4361                 }
4362         }
4363         spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
4364
4365         /* disconnect any connected qp's on ifdown */
4366         if (!ifup)
4367                 irdma_cm_teardown_connections(iwdev, ipaddr, &nfo, false);
4368 }