sys/dev/irdma/irdma_cm.c

   1 /*-
   2  * SPDX-License-Identifier: GPL-2.0 or Linux-OpenIB
   3  *
   4  * Copyright (c) 2015 - 2022 Intel Corporation
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenFabrics.org BSD license below:
  11  *
  12  *   Redistribution and use in source and binary forms, with or
  13  *   without modification, are permitted provided that the following
  14  *   conditions are met:
  15  *
  16  *    - Redistributions of source code must retain the above
  17  *      copyright notice, this list of conditions and the following
  18  *      disclaimer.
  19  *
  20  *    - Redistributions in binary form must reproduce the above
  21  *      copyright notice, this list of conditions and the following
  22  *      disclaimer in the documentation and/or other materials
  23  *      provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34 /*$FreeBSD$*/
  35
  36 #include "irdma_main.h"
  37
  38 static void irdma_cm_post_event(struct irdma_cm_event *event);
  39 static void irdma_disconnect_worker(struct work_struct *work);
  40
  41 /**
  42  * irdma_free_sqbuf - put back puda buffer if refcount is 0
  43  * @vsi: The VSI structure of the device
  44  * @bufp: puda buffer to free
  45  */
  46 void
  47 irdma_free_sqbuf(struct irdma_sc_vsi *vsi, void *bufp)
  48 {
  49         struct irdma_puda_buf *buf = bufp;
  50         struct irdma_puda_rsrc *ilq = vsi->ilq;
  51
  52         if (atomic_dec_and_test(&buf->refcount))
  53                 irdma_puda_ret_bufpool(ilq, buf);
  54 }
  55
  56 /**
  57  * irdma_record_ird_ord - Record IRD/ORD passed in
  58  * @cm_node: connection's node
  59  * @conn_ird: connection IRD
  60  * @conn_ord: connection ORD
  61  */
  62 static void
  63 irdma_record_ird_ord(struct irdma_cm_node *cm_node, u32 conn_ird,
  64                      u32 conn_ord)
  65 {
  66         if (conn_ird > cm_node->dev->hw_attrs.max_hw_ird)
  67                 conn_ird = cm_node->dev->hw_attrs.max_hw_ird;
  68
  69         if (conn_ord > cm_node->dev->hw_attrs.max_hw_ord)
  70                 conn_ord = cm_node->dev->hw_attrs.max_hw_ord;
  71         else if (!conn_ord && cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO)
  72                 conn_ord = 1;
  73         cm_node->ird_size = conn_ird;
  74         cm_node->ord_size = conn_ord;
  75 }
  76
  77 /**
  78  * irdma_copy_ip_ntohl - copy IP address from  network to host
  79  * @dst: IP address in host order
  80  * @src: IP address in network order (big endian)
  81  */
  82 void
  83 irdma_copy_ip_ntohl(u32 *dst, __be32 *src)
  84 {
  85         *dst++ = ntohl(*src++);
  86         *dst++ = ntohl(*src++);
  87         *dst++ = ntohl(*src++);
  88         *dst = ntohl(*src);
  89 }
  90
  91 /**
  92  * irdma_copy_ip_htonl - copy IP address from host to network order
  93  * @dst: IP address in network order (big endian)
  94  * @src: IP address in host order
  95  */
  96 void
  97 irdma_copy_ip_htonl(__be32 *dst, u32 *src)
  98 {
  99         *dst++ = htonl(*src++);
 100         *dst++ = htonl(*src++);
 101         *dst++ = htonl(*src++);
 102         *dst = htonl(*src);
 103 }
 104
 105 /**
 106  * irdma_get_addr_info
 107  * @cm_node: contains ip/tcp info
 108  * @cm_info: to get a copy of the cm_node ip/tcp info
 109  */
 110 static void
 111 irdma_get_addr_info(struct irdma_cm_node *cm_node,
 112                     struct irdma_cm_info *cm_info)
 113 {
 114         memset(cm_info, 0, sizeof(*cm_info));
 115         cm_info->ipv4 = cm_node->ipv4;
 116         cm_info->vlan_id = cm_node->vlan_id;
 117         memcpy(cm_info->loc_addr, cm_node->loc_addr, sizeof(cm_info->loc_addr));
 118         memcpy(cm_info->rem_addr, cm_node->rem_addr, sizeof(cm_info->rem_addr));
 119         cm_info->loc_port = cm_node->loc_port;
 120         cm_info->rem_port = cm_node->rem_port;
 121 }
 122
 123 /**
 124  * irdma_fill_sockaddr4 - fill in addr info for IPv4 connection
 125  * @cm_node: connection's node
 126  * @event: upper layer's cm event
 127  */
 128 static inline void
 129 irdma_fill_sockaddr4(struct irdma_cm_node *cm_node,
 130                      struct iw_cm_event *event)
 131 {
 132         struct sockaddr_in *laddr = (struct sockaddr_in *)&event->local_addr;
 133         struct sockaddr_in *raddr = (struct sockaddr_in *)&event->remote_addr;
 134
 135         laddr->sin_family = AF_INET;
 136         raddr->sin_family = AF_INET;
 137
 138         laddr->sin_port = htons(cm_node->loc_port);
 139         raddr->sin_port = htons(cm_node->rem_port);
 140
 141         laddr->sin_addr.s_addr = htonl(cm_node->loc_addr[0]);
 142         raddr->sin_addr.s_addr = htonl(cm_node->rem_addr[0]);
 143 }
 144
 145 /**
 146  * irdma_fill_sockaddr6 - fill in addr info for IPv6 connection
 147  * @cm_node: connection's node
 148  * @event: upper layer's cm event
 149  */
 150 static inline void
 151 irdma_fill_sockaddr6(struct irdma_cm_node *cm_node,
 152                      struct iw_cm_event *event)
 153 {
 154         struct sockaddr_in6 *laddr6 = (struct sockaddr_in6 *)&event->local_addr;
 155         struct sockaddr_in6 *raddr6 = (struct sockaddr_in6 *)&event->remote_addr;
 156
 157         laddr6->sin6_family = AF_INET6;
 158         raddr6->sin6_family = AF_INET6;
 159
 160         laddr6->sin6_port = htons(cm_node->loc_port);
 161         raddr6->sin6_port = htons(cm_node->rem_port);
 162
 163         irdma_copy_ip_htonl(laddr6->sin6_addr.__u6_addr.__u6_addr32,
 164                             cm_node->loc_addr);
 165         irdma_copy_ip_htonl(raddr6->sin6_addr.__u6_addr.__u6_addr32,
 166                             cm_node->rem_addr);
 167 }
 168
 169 /**
 170  * irdma_get_cmevent_info - for cm event upcall
 171  * @cm_node: connection's node
 172  * @cm_id: upper layers cm struct for the event
 173  * @event: upper layer's cm event
 174  */
 175 static inline void
 176 irdma_get_cmevent_info(struct irdma_cm_node *cm_node,
 177                        struct iw_cm_id *cm_id,
 178                        struct iw_cm_event *event)
 179 {
 180         memcpy(&event->local_addr, &cm_id->m_local_addr,
 181                sizeof(event->local_addr));
 182         memcpy(&event->remote_addr, &cm_id->m_remote_addr,
 183                sizeof(event->remote_addr));
 184         if (cm_node) {
 185                 event->private_data = cm_node->pdata_buf;
 186                 event->private_data_len = (u8)cm_node->pdata.size;
 187                 event->ird = cm_node->ird_size;
 188                 event->ord = cm_node->ord_size;
 189         }
 190 }
 191
 192 /**
 193  * irdma_send_cm_event - upcall cm's event handler
 194  * @cm_node: connection's node
 195  * @cm_id: upper layer's cm info struct
 196  * @type: Event type to indicate
 197  * @status: status for the event type
 198  */
 199 static int
 200 irdma_send_cm_event(struct irdma_cm_node *cm_node,
 201                     struct iw_cm_id *cm_id,
 202                     enum iw_cm_event_type type, int status)
 203 {
 204         struct iw_cm_event event = {0};
 205
 206         event.event = type;
 207         event.status = status;
 208
 209         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 210                     "cm_node %p cm_id=%p state=%d accel=%d event_type=%d status=%d\n",
 211                     cm_node, cm_id, cm_node->accelerated, cm_node->state, type, status);
 212
 213         switch (type) {
 214         case IW_CM_EVENT_CONNECT_REQUEST:
 215                 if (cm_node->ipv4)
 216                         irdma_fill_sockaddr4(cm_node, &event);
 217                 else
 218                         irdma_fill_sockaddr6(cm_node, &event);
 219                 event.provider_data = cm_node;
 220                 event.private_data = cm_node->pdata_buf;
 221                 event.private_data_len = (u8)cm_node->pdata.size;
 222                 event.ird = cm_node->ird_size;
 223                 break;
 224         case IW_CM_EVENT_CONNECT_REPLY:
 225                 irdma_get_cmevent_info(cm_node, cm_id, &event);
 226                 break;
 227         case IW_CM_EVENT_ESTABLISHED:
 228                 event.ird = cm_node->ird_size;
 229                 event.ord = cm_node->ord_size;
 230                 break;
 231         case IW_CM_EVENT_DISCONNECT:
 232         case IW_CM_EVENT_CLOSE:
 233                 /* Wait if we are in RTS but havent issued the iwcm event upcall */
 234                 if (!cm_node->accelerated)
 235                         wait_for_completion(&cm_node->establish_comp);
 236                 break;
 237         default:
 238                 return -EINVAL;
 239         }
 240
 241         return cm_id->event_handler(cm_id, &event);
 242 }
 243
 244 /**
 245  * irdma_timer_list_prep - add connection nodes to a list to perform timer tasks
 246  * @cm_core: cm's core
 247  * @timer_list: a timer list to which cm_node will be selected
 248  */
 249 static void
 250 irdma_timer_list_prep(struct irdma_cm_core *cm_core,
 251                       struct list_head *timer_list)
 252 {
 253         struct irdma_cm_node *cm_node;
 254         int bkt;
 255
 256         HASH_FOR_EACH_RCU(cm_core->cm_hash_tbl, bkt, cm_node, list) {
 257                 if ((cm_node->close_entry || cm_node->send_entry) &&
 258                     atomic_inc_not_zero(&cm_node->refcnt))
 259                         list_add(&cm_node->timer_entry, timer_list);
 260         }
 261 }
 262
 263 /**
 264  * irdma_create_event - create cm event
 265  * @cm_node: connection's node
 266  * @type: Event type to generate
 267  */
 268 static struct irdma_cm_event *
 269 irdma_create_event(struct irdma_cm_node *cm_node,
 270                    enum irdma_cm_event_type type)
 271 {
 272         struct irdma_cm_event *event;
 273
 274         if (!cm_node->cm_id)
 275                 return NULL;
 276
 277         event = kzalloc(sizeof(*event), GFP_ATOMIC);
 278
 279         if (!event)
 280                 return NULL;
 281
 282         event->type = type;
 283         event->cm_node = cm_node;
 284         memcpy(event->cm_info.rem_addr, cm_node->rem_addr,
 285                sizeof(event->cm_info.rem_addr));
 286         memcpy(event->cm_info.loc_addr, cm_node->loc_addr,
 287                sizeof(event->cm_info.loc_addr));
 288         event->cm_info.rem_port = cm_node->rem_port;
 289         event->cm_info.loc_port = cm_node->loc_port;
 290         event->cm_info.cm_id = cm_node->cm_id;
 291         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 292                     "node=%p event=%p type=%u dst=%pI4 src=%pI4\n", cm_node,
 293                     event, type, event->cm_info.loc_addr,
 294                     event->cm_info.rem_addr);
 295         irdma_cm_post_event(event);
 296
 297         return event;
 298 }
 299
 300 /**
 301  * irdma_free_retrans_entry - free send entry
 302  * @cm_node: connection's node
 303  */
 304 static void
 305 irdma_free_retrans_entry(struct irdma_cm_node *cm_node)
 306 {
 307         struct irdma_device *iwdev = cm_node->iwdev;
 308         struct irdma_timer_entry *send_entry;
 309
 310         send_entry = cm_node->send_entry;
 311         if (!send_entry)
 312                 return;
 313
 314         cm_node->send_entry = NULL;
 315         irdma_free_sqbuf(&iwdev->vsi, send_entry->sqbuf);
 316         kfree(send_entry);
 317         atomic_dec(&cm_node->refcnt);
 318 }
 319
 320 /**
 321  * irdma_cleanup_retrans_entry - free send entry with lock
 322  * @cm_node: connection's node
 323  */
 324 static void
 325 irdma_cleanup_retrans_entry(struct irdma_cm_node *cm_node)
 326 {
 327         unsigned long flags;
 328
 329         spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
 330         irdma_free_retrans_entry(cm_node);
 331         spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
 332 }
 333
 334 /**
 335  * irdma_form_ah_cm_frame - get a free packet and build frame with address handle
 336  * @cm_node: connection's node ionfo to use in frame
 337  * @options: pointer to options info
 338  * @hdr: pointer mpa header
 339  * @pdata: pointer to private data
 340  * @flags:  indicates FIN or ACK
 341  */
 342 static struct irdma_puda_buf *
 343 irdma_form_ah_cm_frame(struct irdma_cm_node *cm_node,
 344                        struct irdma_kmem_info *options,
 345                        struct irdma_kmem_info *hdr,
 346                        struct irdma_mpa_priv_info *pdata,
 347                        u8 flags)
 348 {
 349         struct irdma_puda_buf *sqbuf;
 350         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
 351         u8 *buf;
 352         struct tcphdr *tcph;
 353         u16 pktsize;
 354         u32 opts_len = 0;
 355         u32 pd_len = 0;
 356         u32 hdr_len = 0;
 357
 358         if (!cm_node->ah || !cm_node->ah->ah_info.ah_valid) {
 359                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 360                             "AH invalid\n");
 361                 return NULL;
 362         }
 363
 364         sqbuf = irdma_puda_get_bufpool(vsi->ilq);
 365         if (!sqbuf) {
 366                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 367                             "SQ buf NULL\n");
 368                 return NULL;
 369         }
 370
 371         sqbuf->ah_id = cm_node->ah->ah_info.ah_idx;
 372         buf = sqbuf->mem.va;
 373         if (options)
 374                 opts_len = (u32)options->size;
 375
 376         if (hdr)
 377                 hdr_len = hdr->size;
 378
 379         if (pdata)
 380                 pd_len = pdata->size;
 381
 382         pktsize = sizeof(*tcph) + opts_len + hdr_len + pd_len;
 383
 384         memset(buf, 0, pktsize);
 385
 386         sqbuf->totallen = pktsize;
 387         sqbuf->tcphlen = sizeof(*tcph) + opts_len;
 388         sqbuf->scratch = cm_node;
 389
 390         tcph = (struct tcphdr *)buf;
 391         buf += sizeof(*tcph);
 392
 393         tcph->th_sport = htons(cm_node->loc_port);
 394         tcph->th_dport = htons(cm_node->rem_port);
 395         tcph->th_seq = htonl(cm_node->tcp_cntxt.loc_seq_num);
 396
 397         if (flags & SET_ACK) {
 398                 cm_node->tcp_cntxt.loc_ack_num = cm_node->tcp_cntxt.rcv_nxt;
 399                 tcph->th_ack = htonl(cm_node->tcp_cntxt.loc_ack_num);
 400                 tcph->th_flags |= TH_ACK;
 401         } else {
 402                 tcph->th_ack = 0;
 403         }
 404
 405         if (flags & SET_SYN) {
 406                 cm_node->tcp_cntxt.loc_seq_num++;
 407                 tcph->th_flags |= TH_SYN;
 408         } else {
 409                 cm_node->tcp_cntxt.loc_seq_num += hdr_len + pd_len;
 410         }
 411
 412         if (flags & SET_FIN) {
 413                 cm_node->tcp_cntxt.loc_seq_num++;
 414                 tcph->th_flags |= TH_FIN;
 415         }
 416
 417         if (flags & SET_RST)
 418                 tcph->th_flags |= TH_RST;
 419
 420         tcph->th_off = (u16)((sizeof(*tcph) + opts_len + 3) >> 2);
 421         sqbuf->tcphlen = tcph->th_off << 2;
 422         tcph->th_win = htons(cm_node->tcp_cntxt.rcv_wnd);
 423         tcph->th_urp = 0;
 424
 425         if (opts_len) {
 426                 memcpy(buf, options->addr, opts_len);
 427                 buf += opts_len;
 428         }
 429
 430         if (hdr_len) {
 431                 memcpy(buf, hdr->addr, hdr_len);
 432                 buf += hdr_len;
 433         }
 434
 435         if (pdata && pdata->addr)
 436                 memcpy(buf, pdata->addr, pdata->size);
 437
 438         atomic_set(&sqbuf->refcount, 1);
 439
 440         irdma_debug_buf(vsi->dev, IRDMA_DEBUG_ILQ, "TRANSMIT ILQ BUFFER",
 441                         sqbuf->mem.va, sqbuf->totallen);
 442
 443         return sqbuf;
 444 }
 445
 446 /**
 447  * irdma_form_uda_cm_frame - get a free packet and build frame full tcpip packet
 448  * @cm_node: connection's node ionfo to use in frame
 449  * @options: pointer to options info
 450  * @hdr: pointer mpa header
 451  * @pdata: pointer to private data
 452  * @flags:  indicates FIN or ACK
 453  */
 454 static struct irdma_puda_buf *
 455 irdma_form_uda_cm_frame(struct irdma_cm_node *cm_node,
 456                         struct irdma_kmem_info *options,
 457                         struct irdma_kmem_info *hdr,
 458                         struct irdma_mpa_priv_info *pdata,
 459                         u8 flags)
 460 {
 461         struct irdma_puda_buf *sqbuf;
 462         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
 463         u8 *buf;
 464
 465         struct tcphdr *tcph;
 466         struct ip *iph;
 467         struct ip6_hdr *ip6h;
 468         struct ether_header *ethh;
 469         u16 pktsize;
 470         u16 eth_hlen = ETH_HLEN;
 471         u32 opts_len = 0;
 472         u32 pd_len = 0;
 473         u32 hdr_len = 0;
 474
 475         u16 vtag;
 476
 477         sqbuf = irdma_puda_get_bufpool(vsi->ilq);
 478         if (!sqbuf)
 479                 return NULL;
 480
 481         buf = sqbuf->mem.va;
 482
 483         if (options)
 484                 opts_len = (u32)options->size;
 485
 486         if (hdr)
 487                 hdr_len = hdr->size;
 488
 489         if (pdata)
 490                 pd_len = pdata->size;
 491
 492         if (cm_node->vlan_id < VLAN_N_VID)
 493                 eth_hlen += 4;
 494
 495         if (cm_node->ipv4)
 496                 pktsize = sizeof(*iph) + sizeof(*tcph);
 497         else
 498                 pktsize = sizeof(*ip6h) + sizeof(*tcph);
 499         pktsize += opts_len + hdr_len + pd_len;
 500
 501         memset(buf, 0, eth_hlen + pktsize);
 502
 503         sqbuf->totallen = pktsize + eth_hlen;
 504         sqbuf->maclen = eth_hlen;
 505         sqbuf->tcphlen = sizeof(*tcph) + opts_len;
 506         sqbuf->scratch = cm_node;
 507
 508         ethh = (struct ether_header *)buf;
 509         buf += eth_hlen;
 510
 511         if (cm_node->do_lpb)
 512                 sqbuf->do_lpb = true;
 513
 514         if (cm_node->ipv4) {
 515                 sqbuf->ipv4 = true;
 516
 517                 iph = (struct ip *)buf;
 518                 buf += sizeof(*iph);
 519                 tcph = (struct tcphdr *)buf;
 520                 buf += sizeof(*tcph);
 521
 522                 ether_addr_copy(ethh->ether_dhost, cm_node->rem_mac);
 523                 ether_addr_copy(ethh->ether_shost, cm_node->loc_mac);
 524                 if (cm_node->vlan_id < VLAN_N_VID) {
 525                         ((struct ether_vlan_header *)ethh)->evl_proto =
 526                             htons(ETH_P_8021Q);
 527                         vtag = (cm_node->user_pri << VLAN_PRIO_SHIFT) |
 528                             cm_node->vlan_id;
 529                         ((struct ether_vlan_header *)ethh)->evl_tag = htons(vtag);
 530
 531                         ((struct ether_vlan_header *)ethh)->evl_encap_proto =
 532                             htons(ETH_P_IP);
 533                 } else {
 534                         ethh->ether_type = htons(ETH_P_IP);
 535                 }
 536
 537                 iph->ip_v = IPVERSION;
 538                 iph->ip_hl = 5; /* 5 * 4Byte words, IP headr len */
 539                 iph->ip_tos = cm_node->tos;
 540                 iph->ip_len = htons(pktsize);
 541                 iph->ip_id = htons(++cm_node->tcp_cntxt.loc_id);
 542
 543                 iph->ip_off = htons(0x4000);
 544                 iph->ip_ttl = 0x40;
 545                 iph->ip_p = IPPROTO_TCP;
 546                 iph->ip_src.s_addr = htonl(cm_node->loc_addr[0]);
 547                 iph->ip_dst.s_addr = htonl(cm_node->rem_addr[0]);
 548         } else {
 549                 sqbuf->ipv4 = false;
 550                 ip6h = (struct ip6_hdr *)buf;
 551                 buf += sizeof(*ip6h);
 552                 tcph = (struct tcphdr *)buf;
 553                 buf += sizeof(*tcph);
 554
 555                 ether_addr_copy(ethh->ether_dhost, cm_node->rem_mac);
 556                 ether_addr_copy(ethh->ether_shost, cm_node->loc_mac);
 557                 if (cm_node->vlan_id < VLAN_N_VID) {
 558                         ((struct ether_vlan_header *)ethh)->evl_proto =
 559                             htons(ETH_P_8021Q);
 560                         vtag = (cm_node->user_pri << VLAN_PRIO_SHIFT) |
 561                             cm_node->vlan_id;
 562                         ((struct ether_vlan_header *)ethh)->evl_tag = htons(vtag);
 563                         ((struct ether_vlan_header *)ethh)->evl_encap_proto =
 564                             htons(ETH_P_IPV6);
 565                 } else {
 566                         ethh->ether_type = htons(ETH_P_IPV6);
 567                 }
 568                 ip6h->ip6_vfc = 6 << 4;
 569                 ip6h->ip6_vfc |= cm_node->tos >> 4;
 570                 ip6h->ip6_flow = cm_node->tos << 20;
 571                 ip6h->ip6_plen = htons(pktsize - sizeof(*ip6h));
 572                 ip6h->ip6_nxt = 6;
 573                 ip6h->ip6_hops = 128;
 574                 irdma_copy_ip_htonl(ip6h->ip6_src.__u6_addr.__u6_addr32,
 575                                     cm_node->loc_addr);
 576                 irdma_copy_ip_htonl(ip6h->ip6_dst.__u6_addr.__u6_addr32,
 577                                     cm_node->rem_addr);
 578         }
 579
 580         tcph->th_sport = htons(cm_node->loc_port);
 581         tcph->th_dport = htons(cm_node->rem_port);
 582         tcph->th_seq = htonl(cm_node->tcp_cntxt.loc_seq_num);
 583
 584         if (flags & SET_ACK) {
 585                 cm_node->tcp_cntxt.loc_ack_num = cm_node->tcp_cntxt.rcv_nxt;
 586                 tcph->th_ack = htonl(cm_node->tcp_cntxt.loc_ack_num);
 587                 tcph->th_flags |= TH_ACK;
 588         } else {
 589                 tcph->th_ack = 0;
 590         }
 591
 592         if (flags & SET_SYN) {
 593                 cm_node->tcp_cntxt.loc_seq_num++;
 594                 tcph->th_flags |= TH_SYN;
 595         } else {
 596                 cm_node->tcp_cntxt.loc_seq_num += hdr_len + pd_len;
 597         }
 598
 599         if (flags & SET_FIN) {
 600                 cm_node->tcp_cntxt.loc_seq_num++;
 601                 tcph->th_flags |= TH_FIN;
 602         }
 603
 604         if (flags & SET_RST)
 605                 tcph->th_flags |= TH_RST;
 606
 607         tcph->th_off = (u16)((sizeof(*tcph) + opts_len + 3) >> 2);
 608         sqbuf->tcphlen = tcph->th_off << 2;
 609         tcph->th_win = htons(cm_node->tcp_cntxt.rcv_wnd);
 610         tcph->th_urp = 0;
 611
 612         if (opts_len) {
 613                 memcpy(buf, options->addr, opts_len);
 614                 buf += opts_len;
 615         }
 616
 617         if (hdr_len) {
 618                 memcpy(buf, hdr->addr, hdr_len);
 619                 buf += hdr_len;
 620         }
 621
 622         if (pdata && pdata->addr)
 623                 memcpy(buf, pdata->addr, pdata->size);
 624
 625         atomic_set(&sqbuf->refcount, 1);
 626
 627         irdma_debug_buf(vsi->dev, IRDMA_DEBUG_ILQ, "TRANSMIT ILQ BUFFER",
 628                         sqbuf->mem.va, sqbuf->totallen);
 629
 630         return sqbuf;
 631 }
 632
 633 /**
 634  * irdma_send_reset - Send RST packet
 635  * @cm_node: connection's node
 636  */
 637 int
 638 irdma_send_reset(struct irdma_cm_node *cm_node)
 639 {
 640         struct irdma_puda_buf *sqbuf;
 641         int flags = SET_RST | SET_ACK;
 642
 643         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
 644                                                 flags);
 645         if (!sqbuf)
 646                 return -ENOMEM;
 647
 648         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 649                     "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",
 650                     __builtin_return_address(0), cm_node, cm_node->cm_id,
 651                     cm_node->accelerated, cm_node->state, cm_node->rem_port,
 652                     cm_node->loc_port, cm_node->rem_addr, cm_node->loc_addr);
 653
 654         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 0,
 655                                        1);
 656 }
 657
 658 /**
 659  * irdma_active_open_err - send event for active side cm error
 660  * @cm_node: connection's node
 661  * @reset: Flag to send reset or not
 662  */
 663 static void
 664 irdma_active_open_err(struct irdma_cm_node *cm_node, bool reset)
 665 {
 666         irdma_cleanup_retrans_entry(cm_node);
 667         cm_node->cm_core->stats_connect_errs++;
 668         if (reset) {
 669                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 670                             "cm_node=%p state=%d\n", cm_node, cm_node->state);
 671                 atomic_inc(&cm_node->refcnt);
 672                 irdma_send_reset(cm_node);
 673         }
 674
 675         cm_node->state = IRDMA_CM_STATE_CLOSED;
 676         irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
 677 }
 678
 679 /**
 680  * irdma_passive_open_err - handle passive side cm error
 681  * @cm_node: connection's node
 682  * @reset: send reset or just free cm_node
 683  */
 684 static void
 685 irdma_passive_open_err(struct irdma_cm_node *cm_node, bool reset)
 686 {
 687         irdma_cleanup_retrans_entry(cm_node);
 688         cm_node->cm_core->stats_passive_errs++;
 689         cm_node->state = IRDMA_CM_STATE_CLOSED;
 690         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 691                     "cm_node=%p state =%d\n", cm_node, cm_node->state);
 692         if (reset)
 693                 irdma_send_reset(cm_node);
 694         else
 695                 irdma_rem_ref_cm_node(cm_node);
 696 }
 697
 698 /**
 699  * irdma_event_connect_error - to create connect error event
 700  * @event: cm information for connect event
 701  */
 702 static void
 703 irdma_event_connect_error(struct irdma_cm_event *event)
 704 {
 705         struct irdma_qp *iwqp;
 706         struct iw_cm_id *cm_id;
 707
 708         cm_id = event->cm_node->cm_id;
 709         if (!cm_id)
 710                 return;
 711
 712         iwqp = cm_id->provider_data;
 713
 714         if (!iwqp || !iwqp->iwdev)
 715                 return;
 716
 717         iwqp->cm_id = NULL;
 718         cm_id->provider_data = NULL;
 719         irdma_send_cm_event(event->cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY,
 720                             -ECONNRESET);
 721         irdma_rem_ref_cm_node(event->cm_node);
 722 }
 723
 724 /**
 725  * irdma_process_options - process options from TCP header
 726  * @cm_node: connection's node
 727  * @optionsloc: point to start of options
 728  * @optionsize: size of all options
 729  * @syn_pkt: flag if syn packet
 730  */
 731 static int
 732 irdma_process_options(struct irdma_cm_node *cm_node, u8 *optionsloc,
 733                       u32 optionsize, u32 syn_pkt)
 734 {
 735         u32 tmp;
 736         u32 offset = 0;
 737         union all_known_options *all_options;
 738         char got_mss_option = 0;
 739
 740         while (offset < optionsize) {
 741                 all_options = (union all_known_options *)(optionsloc + offset);
 742                 switch (all_options->base.optionnum) {
 743                 case OPTION_NUM_EOL:
 744                         offset = optionsize;
 745                         break;
 746                 case OPTION_NUM_NONE:
 747                         offset += 1;
 748                         continue;
 749                 case OPTION_NUM_MSS:
 750                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
 751                                     IRDMA_DEBUG_CM,
 752                                     "MSS Length: %d Offset: %d Size: %d\n",
 753                                     all_options->mss.len, offset, optionsize);
 754                         got_mss_option = 1;
 755                         if (all_options->mss.len != 4)
 756                                 return -EINVAL;
 757                         tmp = ntohs(all_options->mss.mss);
 758                         if ((cm_node->ipv4 &&
 759                              (tmp + IRDMA_MTU_TO_MSS_IPV4) < IRDMA_MIN_MTU_IPV4) ||
 760                             (!cm_node->ipv4 &&
 761                              (tmp + IRDMA_MTU_TO_MSS_IPV6) < IRDMA_MIN_MTU_IPV6))
 762                                 return -EINVAL;
 763                         if (tmp < cm_node->tcp_cntxt.mss)
 764                                 cm_node->tcp_cntxt.mss = tmp;
 765                         break;
 766                 case OPTION_NUM_WINDOW_SCALE:
 767                         cm_node->tcp_cntxt.snd_wscale =
 768                             all_options->windowscale.shiftcount;
 769                         break;
 770                 default:
 771                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
 772                                     IRDMA_DEBUG_CM,
 773                                     "Unsupported TCP Option: %x\n",
 774                                     all_options->base.optionnum);
 775                         break;
 776                 }
 777                 offset += all_options->base.len;
 778         }
 779         if (!got_mss_option && syn_pkt)
 780                 cm_node->tcp_cntxt.mss = IRDMA_CM_DEFAULT_MSS;
 781
 782         return 0;
 783 }
 784
 785 /**
 786  * irdma_handle_tcp_options - setup TCP context info after parsing TCP options
 787  * @cm_node: connection's node
 788  * @tcph: pointer tcp header
 789  * @optionsize: size of options rcvd
 790  * @passive: active or passive flag
 791  */
 792 static int
 793 irdma_handle_tcp_options(struct irdma_cm_node *cm_node,
 794                          struct tcphdr *tcph, int optionsize,
 795                          int passive)
 796 {
 797         u8 *optionsloc = (u8 *)&tcph[1];
 798         int ret;
 799
 800         if (optionsize) {
 801                 ret = irdma_process_options(cm_node, optionsloc, optionsize,
 802                                             (u32)tcph->th_flags & TH_SYN);
 803                 if (ret) {
 804                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
 805                                     IRDMA_DEBUG_CM,
 806                                     "Node %p, Sending Reset\n", cm_node);
 807                         if (passive)
 808                                 irdma_passive_open_err(cm_node, true);
 809                         else
 810                                 irdma_active_open_err(cm_node, true);
 811                         return ret;
 812                 }
 813         }
 814
 815         cm_node->tcp_cntxt.snd_wnd = ntohs(tcph->th_win)
 816             << cm_node->tcp_cntxt.snd_wscale;
 817
 818         if (cm_node->tcp_cntxt.snd_wnd > cm_node->tcp_cntxt.max_snd_wnd)
 819                 cm_node->tcp_cntxt.max_snd_wnd = cm_node->tcp_cntxt.snd_wnd;
 820
 821         return 0;
 822 }
 823
 824 /**
 825  * irdma_build_mpa_v1 - build a MPA V1 frame
 826  * @cm_node: connection's node
 827  * @start_addr: address where to build frame
 828  * @mpa_key: to do read0 or write0
 829  */
 830 static void
 831 irdma_build_mpa_v1(struct irdma_cm_node *cm_node, void *start_addr,
 832                    u8 mpa_key)
 833 {
 834         struct ietf_mpa_v1 *mpa_frame = start_addr;
 835
 836         switch (mpa_key) {
 837         case MPA_KEY_REQUEST:
 838                 memcpy(mpa_frame->key, IEFT_MPA_KEY_REQ, IETF_MPA_KEY_SIZE);
 839                 break;
 840         case MPA_KEY_REPLY:
 841                 memcpy(mpa_frame->key, IEFT_MPA_KEY_REP, IETF_MPA_KEY_SIZE);
 842                 break;
 843         default:
 844                 break;
 845         }
 846         mpa_frame->flags = IETF_MPA_FLAGS_CRC;
 847         mpa_frame->rev = cm_node->mpa_frame_rev;
 848         mpa_frame->priv_data_len = htons(cm_node->pdata.size);
 849 }
 850
 851 /**
 852  * irdma_build_mpa_v2 - build a MPA V2 frame
 853  * @cm_node: connection's node
 854  * @start_addr: buffer start address
 855  * @mpa_key: to do read0 or write0
 856  */
 857 static void
 858 irdma_build_mpa_v2(struct irdma_cm_node *cm_node, void *start_addr,
 859                    u8 mpa_key)
 860 {
 861         struct ietf_mpa_v2 *mpa_frame = start_addr;
 862         struct ietf_rtr_msg *rtr_msg = &mpa_frame->rtr_msg;
 863         u16 ctrl_ird, ctrl_ord;
 864
 865         /* initialize the upper 5 bytes of the frame */
 866         irdma_build_mpa_v1(cm_node, start_addr, mpa_key);
 867         mpa_frame->flags |= IETF_MPA_V2_FLAG;
 868         if (cm_node->iwdev->iw_ooo) {
 869                 mpa_frame->flags |= IETF_MPA_FLAGS_MARKERS;
 870                 cm_node->rcv_mark_en = true;
 871         }
 872         mpa_frame->priv_data_len = cpu_to_be16(be16_to_cpu(mpa_frame->priv_data_len) +
 873                                                IETF_RTR_MSG_SIZE);
 874
 875         /* initialize RTR msg */
 876         if (cm_node->mpav2_ird_ord == IETF_NO_IRD_ORD) {
 877                 ctrl_ird = IETF_NO_IRD_ORD;
 878                 ctrl_ord = IETF_NO_IRD_ORD;
 879         } else {
 880                 ctrl_ird = (cm_node->ird_size > IETF_NO_IRD_ORD) ?
 881                     IETF_NO_IRD_ORD :
 882                     cm_node->ird_size;
 883                 ctrl_ord = (cm_node->ord_size > IETF_NO_IRD_ORD) ?
 884                     IETF_NO_IRD_ORD :
 885                     cm_node->ord_size;
 886         }
 887         ctrl_ird |= IETF_PEER_TO_PEER;
 888
 889         switch (mpa_key) {
 890         case MPA_KEY_REQUEST:
 891                 ctrl_ord |= IETF_RDMA0_WRITE;
 892                 ctrl_ord |= IETF_RDMA0_READ;
 893                 break;
 894         case MPA_KEY_REPLY:
 895                 switch (cm_node->send_rdma0_op) {
 896                 case SEND_RDMA_WRITE_ZERO:
 897                         ctrl_ord |= IETF_RDMA0_WRITE;
 898                         break;
 899                 case SEND_RDMA_READ_ZERO:
 900                         ctrl_ord |= IETF_RDMA0_READ;
 901                         break;
 902                 }
 903                 break;
 904         default:
 905                 break;
 906         }
 907         rtr_msg->ctrl_ird = htons(ctrl_ird);
 908         rtr_msg->ctrl_ord = htons(ctrl_ord);
 909 }
 910
 911 /**
 912  * irdma_cm_build_mpa_frame - build mpa frame for mpa version 1 or version 2
 913  * @cm_node: connection's node
 914  * @mpa: mpa: data buffer
 915  * @mpa_key: to do read0 or write0
 916  */
 917 static int
 918 irdma_cm_build_mpa_frame(struct irdma_cm_node *cm_node,
 919                          struct irdma_kmem_info *mpa, u8 mpa_key)
 920 {
 921         int hdr_len = 0;
 922
 923         switch (cm_node->mpa_frame_rev) {
 924         case IETF_MPA_V1:
 925                 hdr_len = sizeof(struct ietf_mpa_v1);
 926                 irdma_build_mpa_v1(cm_node, mpa->addr, mpa_key);
 927                 break;
 928         case IETF_MPA_V2:
 929                 hdr_len = sizeof(struct ietf_mpa_v2);
 930                 irdma_build_mpa_v2(cm_node, mpa->addr, mpa_key);
 931                 break;
 932         default:
 933                 break;
 934         }
 935
 936         return hdr_len;
 937 }
 938
 939 /**
 940  * irdma_send_mpa_request - active node send mpa request to passive node
 941  * @cm_node: connection's node
 942  */
 943 static int
 944 irdma_send_mpa_request(struct irdma_cm_node *cm_node)
 945 {
 946         struct irdma_puda_buf *sqbuf;
 947
 948         cm_node->mpa_hdr.addr = &cm_node->mpa_v2_frame;
 949         cm_node->mpa_hdr.size = irdma_cm_build_mpa_frame(cm_node,
 950                                                          &cm_node->mpa_hdr,
 951                                                          MPA_KEY_REQUEST);
 952         if (!cm_node->mpa_hdr.size) {
 953                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
 954                             "mpa size = %d\n", cm_node->mpa_hdr.size);
 955                 return -EINVAL;
 956         }
 957
 958         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL,
 959                                                 &cm_node->mpa_hdr,
 960                                                 &cm_node->pdata, SET_ACK);
 961         if (!sqbuf)
 962                 return -ENOMEM;
 963
 964         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
 965                                        0);
 966 }
 967
 968 /**
 969  * irdma_send_mpa_reject -
 970  * @cm_node: connection's node
 971  * @pdata: reject data for connection
 972  * @plen: length of reject data
 973  */
 974 static int
 975 irdma_send_mpa_reject(struct irdma_cm_node *cm_node,
 976                       const void *pdata, u8 plen)
 977 {
 978         struct irdma_puda_buf *sqbuf;
 979         struct irdma_mpa_priv_info priv_info;
 980
 981         cm_node->mpa_hdr.addr = &cm_node->mpa_v2_frame;
 982         cm_node->mpa_hdr.size = irdma_cm_build_mpa_frame(cm_node,
 983                                                          &cm_node->mpa_hdr,
 984                                                          MPA_KEY_REPLY);
 985
 986         cm_node->mpa_v2_frame.flags |= IETF_MPA_FLAGS_REJECT;
 987         priv_info.addr = pdata;
 988         priv_info.size = plen;
 989
 990         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL,
 991                                                 &cm_node->mpa_hdr, &priv_info,
 992                                                 SET_ACK | SET_FIN);
 993         if (!sqbuf)
 994                 return -ENOMEM;
 995
 996         cm_node->state = IRDMA_CM_STATE_FIN_WAIT1;
 997
 998         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
 999                                        0);
1000 }
1001
1002 /**
1003  * irdma_negotiate_mpa_v2_ird_ord - negotiate MPAv2 IRD/ORD
1004  * @cm_node: connection's node
1005  * @buf: Data pointer
1006  */
1007 static int
1008 irdma_negotiate_mpa_v2_ird_ord(struct irdma_cm_node *cm_node,
1009                                u8 *buf)
1010 {
1011         struct ietf_mpa_v2 *mpa_v2_frame;
1012         struct ietf_rtr_msg *rtr_msg;
1013         u16 ird_size;
1014         u16 ord_size;
1015         u16 ctrl_ord;
1016         u16 ctrl_ird;
1017
1018         mpa_v2_frame = (struct ietf_mpa_v2 *)buf;
1019         rtr_msg = &mpa_v2_frame->rtr_msg;
1020
1021         /* parse rtr message */
1022         ctrl_ord = ntohs(rtr_msg->ctrl_ord);
1023         ctrl_ird = ntohs(rtr_msg->ctrl_ird);
1024         ird_size = ctrl_ird & IETF_NO_IRD_ORD;
1025         ord_size = ctrl_ord & IETF_NO_IRD_ORD;
1026
1027         if (!(ctrl_ird & IETF_PEER_TO_PEER))
1028                 return -EOPNOTSUPP;
1029
1030         if (ird_size == IETF_NO_IRD_ORD || ord_size == IETF_NO_IRD_ORD) {
1031                 cm_node->mpav2_ird_ord = IETF_NO_IRD_ORD;
1032                 goto negotiate_done;
1033         }
1034
1035         if (cm_node->state != IRDMA_CM_STATE_MPAREQ_SENT) {
1036                 /* responder */
1037                 if (!ord_size && (ctrl_ord & IETF_RDMA0_READ))
1038                         cm_node->ird_size = 1;
1039                 if (cm_node->ord_size > ird_size)
1040                         cm_node->ord_size = ird_size;
1041         } else {
1042                 /* initiator */
1043                 if (!ird_size && (ctrl_ord & IETF_RDMA0_READ))
1044                         /* Remote peer doesn't support RDMA0_READ */
1045                         return -EOPNOTSUPP;
1046
1047                 if (cm_node->ord_size > ird_size)
1048                         cm_node->ord_size = ird_size;
1049
1050                 if (cm_node->ird_size < ord_size)
1051                         /* no resources available */
1052                         return -EINVAL;
1053         }
1054
1055 negotiate_done:
1056         if (ctrl_ord & IETF_RDMA0_READ)
1057                 cm_node->send_rdma0_op = SEND_RDMA_READ_ZERO;
1058         else if (ctrl_ord & IETF_RDMA0_WRITE)
1059                 cm_node->send_rdma0_op = SEND_RDMA_WRITE_ZERO;
1060         else
1061                 /* Not supported RDMA0 operation */
1062                 return -EOPNOTSUPP;
1063
1064         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1065                     "MPAV2 Negotiated ORD: %d, IRD: %d\n", cm_node->ord_size,
1066                     cm_node->ird_size);
1067         return 0;
1068 }
1069
1070 /**
1071  * irdma_parse_mpa - process an IETF MPA frame
1072  * @cm_node: connection's node
1073  * @buf: Data pointer
1074  * @type: to return accept or reject
1075  * @len: Len of mpa buffer
1076  */
1077 static int
1078 irdma_parse_mpa(struct irdma_cm_node *cm_node, u8 *buf, u32 *type,
1079                 u32 len)
1080 {
1081         struct ietf_mpa_v1 *mpa_frame;
1082         int mpa_hdr_len, priv_data_len, ret;
1083
1084         *type = IRDMA_MPA_REQUEST_ACCEPT;
1085
1086         if (len < sizeof(struct ietf_mpa_v1)) {
1087                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1088                             "ietf buffer small (%x)\n", len);
1089                 return -EINVAL;
1090         }
1091
1092         mpa_frame = (struct ietf_mpa_v1 *)buf;
1093         mpa_hdr_len = sizeof(struct ietf_mpa_v1);
1094         priv_data_len = ntohs(mpa_frame->priv_data_len);
1095
1096         if (priv_data_len > IETF_MAX_PRIV_DATA_LEN) {
1097                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1098                             "private_data too big %d\n", priv_data_len);
1099                 return -EOVERFLOW;
1100         }
1101
1102         if (mpa_frame->rev != IETF_MPA_V1 && mpa_frame->rev != IETF_MPA_V2) {
1103                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1104                             "unsupported mpa rev = %d\n", mpa_frame->rev);
1105                 return -EINVAL;
1106         }
1107
1108         if (mpa_frame->rev > cm_node->mpa_frame_rev) {
1109                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1110                             "rev %d\n", mpa_frame->rev);
1111                 return -EINVAL;
1112         }
1113
1114         cm_node->mpa_frame_rev = mpa_frame->rev;
1115         if (cm_node->state != IRDMA_CM_STATE_MPAREQ_SENT) {
1116                 if (memcmp(mpa_frame->key, IEFT_MPA_KEY_REQ,
1117                            IETF_MPA_KEY_SIZE)) {
1118                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
1119                                     IRDMA_DEBUG_CM,
1120                                     "Unexpected MPA Key received\n");
1121                         return -EINVAL;
1122                 }
1123         } else {
1124                 if (memcmp(mpa_frame->key, IEFT_MPA_KEY_REP,
1125                            IETF_MPA_KEY_SIZE)) {
1126                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
1127                                     IRDMA_DEBUG_CM,
1128                                     "Unexpected MPA Key received\n");
1129                         return -EINVAL;
1130                 }
1131         }
1132
1133         if (priv_data_len + mpa_hdr_len > len) {
1134                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1135                             "ietf buffer len(%x + %x != %x)\n", priv_data_len,
1136                             mpa_hdr_len, len);
1137                 return -EOVERFLOW;
1138         }
1139
1140         if (len > IRDMA_MAX_CM_BUF) {
1141                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
1142                             "ietf buffer large len = %d\n", len);
1143                 return -EOVERFLOW;
1144         }
1145
1146         switch (mpa_frame->rev) {
1147         case IETF_MPA_V2:
1148                 mpa_hdr_len += IETF_RTR_MSG_SIZE;
1149                 ret = irdma_negotiate_mpa_v2_ird_ord(cm_node, buf);
1150                 if (ret)
1151                         return ret;
1152                 break;
1153         case IETF_MPA_V1:
1154         default:
1155                 break;
1156         }
1157
1158         memcpy(cm_node->pdata_buf, buf + mpa_hdr_len, priv_data_len);
1159         cm_node->pdata.size = priv_data_len;
1160
1161         if (mpa_frame->flags & IETF_MPA_FLAGS_REJECT)
1162                 *type = IRDMA_MPA_REQUEST_REJECT;
1163
1164         if (mpa_frame->flags & IETF_MPA_FLAGS_MARKERS)
1165                 cm_node->snd_mark_en = true;
1166
1167         return 0;
1168 }
1169
1170 /**
1171  * irdma_schedule_cm_timer
1172  * @cm_node: connection's node
1173  * @sqbuf: buffer to send
1174  * @type: if it is send or close
1175  * @send_retrans: if rexmits to be done
1176  * @close_when_complete: is cm_node to be removed
1177  *
1178  * note - cm_node needs to be protected before calling this. Encase in:
1179  *              irdma_rem_ref_cm_node(cm_core, cm_node);
1180  *              irdma_schedule_cm_timer(...)
1181  *              atomic_inc(&cm_node->refcnt);
1182  */
1183 int
1184 irdma_schedule_cm_timer(struct irdma_cm_node *cm_node,
1185                         struct irdma_puda_buf *sqbuf,
1186                         enum irdma_timer_type type, int send_retrans,
1187                         int close_when_complete)
1188 {
1189         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
1190         struct irdma_cm_core *cm_core = cm_node->cm_core;
1191         struct irdma_timer_entry *new_send;
1192         u32 was_timer_set;
1193         unsigned long flags;
1194
1195         new_send = kzalloc(sizeof(*new_send), GFP_ATOMIC);
1196         if (!new_send) {
1197                 if (type != IRDMA_TIMER_TYPE_CLOSE)
1198                         irdma_free_sqbuf(vsi, sqbuf);
1199                 return -ENOMEM;
1200         }
1201
1202         new_send->retrycount = IRDMA_DEFAULT_RETRYS;
1203         new_send->retranscount = IRDMA_DEFAULT_RETRANS;
1204         new_send->sqbuf = sqbuf;
1205         new_send->timetosend = jiffies;
1206         new_send->type = type;
1207         new_send->send_retrans = send_retrans;
1208         new_send->close_when_complete = close_when_complete;
1209
1210         if (type == IRDMA_TIMER_TYPE_CLOSE) {
1211                 new_send->timetosend += (HZ / 10);
1212                 if (cm_node->close_entry) {
1213                         kfree(new_send);
1214                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
1215                                     IRDMA_DEBUG_CM, "already close entry\n");
1216                         return -EINVAL;
1217                 }
1218
1219                 cm_node->close_entry = new_send;
1220         } else {                /* type == IRDMA_TIMER_TYPE_SEND */
1221                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1222                 cm_node->send_entry = new_send;
1223                 atomic_inc(&cm_node->refcnt);
1224                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1225                 new_send->timetosend = jiffies + IRDMA_RETRY_TIMEOUT;
1226
1227                 atomic_inc(&sqbuf->refcount);
1228                 irdma_puda_send_buf(vsi->ilq, sqbuf);
1229                 if (!send_retrans) {
1230                         irdma_cleanup_retrans_entry(cm_node);
1231                         if (close_when_complete)
1232                                 irdma_rem_ref_cm_node(cm_node);
1233                         return 0;
1234                 }
1235         }
1236
1237         spin_lock_irqsave(&cm_core->ht_lock, flags);
1238         was_timer_set = timer_pending(&cm_core->tcp_timer);
1239
1240         if (!was_timer_set) {
1241                 cm_core->tcp_timer.expires = new_send->timetosend;
1242                 add_timer(&cm_core->tcp_timer);
1243         }
1244         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
1245
1246         return 0;
1247 }
1248
1249 /**
1250  * irdma_retrans_expired - Could not rexmit the packet
1251  * @cm_node: connection's node
1252  */
1253 static void
1254 irdma_retrans_expired(struct irdma_cm_node *cm_node)
1255 {
1256         enum irdma_cm_node_state state = cm_node->state;
1257
1258         cm_node->state = IRDMA_CM_STATE_CLOSED;
1259         switch (state) {
1260         case IRDMA_CM_STATE_SYN_RCVD:
1261         case IRDMA_CM_STATE_CLOSING:
1262                 irdma_rem_ref_cm_node(cm_node);
1263                 break;
1264         case IRDMA_CM_STATE_FIN_WAIT1:
1265         case IRDMA_CM_STATE_LAST_ACK:
1266                 irdma_send_reset(cm_node);
1267                 break;
1268         default:
1269                 atomic_inc(&cm_node->refcnt);
1270                 irdma_send_reset(cm_node);
1271                 irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
1272                 break;
1273         }
1274 }
1275
1276 /**
1277  * irdma_handle_close_entry - for handling retry/timeouts
1278  * @cm_node: connection's node
1279  * @rem_node: flag for remove cm_node
1280  */
1281 static void
1282 irdma_handle_close_entry(struct irdma_cm_node *cm_node,
1283                          u32 rem_node)
1284 {
1285         struct irdma_timer_entry *close_entry = cm_node->close_entry;
1286         struct irdma_qp *iwqp;
1287         unsigned long flags;
1288
1289         if (!close_entry)
1290                 return;
1291         iwqp = (struct irdma_qp *)close_entry->sqbuf;
1292         if (iwqp) {
1293                 spin_lock_irqsave(&iwqp->lock, flags);
1294                 if (iwqp->cm_id) {
1295                         iwqp->hw_tcp_state = IRDMA_TCP_STATE_CLOSED;
1296                         iwqp->hw_iwarp_state = IRDMA_QP_STATE_ERROR;
1297                         iwqp->last_aeq = IRDMA_AE_RESET_SENT;
1298                         iwqp->ibqp_state = IB_QPS_ERR;
1299                         spin_unlock_irqrestore(&iwqp->lock, flags);
1300                         irdma_cm_disconn(iwqp);
1301                 } else {
1302                         spin_unlock_irqrestore(&iwqp->lock, flags);
1303                 }
1304         } else if (rem_node) {
1305                 /* TIME_WAIT state */
1306                 irdma_rem_ref_cm_node(cm_node);
1307         }
1308
1309         kfree(close_entry);
1310         cm_node->close_entry = NULL;
1311 }
1312
1313 /**
1314  * irdma_cm_timer_tick - system's timer expired callback
1315  * @t: Pointer to timer_list
1316  */
1317 static void
1318 irdma_cm_timer_tick(struct timer_list *t)
1319 {
1320         unsigned long nexttimeout = jiffies + IRDMA_LONG_TIME;
1321         struct irdma_cm_node *cm_node;
1322         struct irdma_timer_entry *send_entry, *close_entry;
1323         struct list_head *list_core_temp;
1324         struct list_head *list_node;
1325         struct irdma_cm_core *cm_core = from_timer(cm_core, t, tcp_timer);
1326         struct irdma_sc_vsi *vsi;
1327         u32 settimer = 0;
1328         unsigned long timetosend;
1329         unsigned long flags;
1330         struct list_head timer_list;
1331
1332         INIT_LIST_HEAD(&timer_list);
1333
1334         rcu_read_lock();
1335         irdma_timer_list_prep(cm_core, &timer_list);
1336         rcu_read_unlock();
1337
1338         list_for_each_safe(list_node, list_core_temp, &timer_list) {
1339                 cm_node = container_of(list_node, struct irdma_cm_node,
1340                                        timer_entry);
1341                 close_entry = cm_node->close_entry;
1342
1343                 if (close_entry) {
1344                         if (time_after(close_entry->timetosend, jiffies)) {
1345                                 if (nexttimeout > close_entry->timetosend ||
1346                                     !settimer) {
1347                                         nexttimeout = close_entry->timetosend;
1348                                         settimer = 1;
1349                                 }
1350                         } else {
1351                                 irdma_handle_close_entry(cm_node, 1);
1352                         }
1353                 }
1354
1355                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1356
1357                 send_entry = cm_node->send_entry;
1358                 if (!send_entry)
1359                         goto done;
1360                 if (time_after(send_entry->timetosend, jiffies)) {
1361                         if (cm_node->state != IRDMA_CM_STATE_OFFLOADED) {
1362                                 if (nexttimeout > send_entry->timetosend ||
1363                                     !settimer) {
1364                                         nexttimeout = send_entry->timetosend;
1365                                         settimer = 1;
1366                                 }
1367                         } else {
1368                                 irdma_free_retrans_entry(cm_node);
1369                         }
1370                         goto done;
1371                 }
1372
1373                 if (cm_node->state == IRDMA_CM_STATE_OFFLOADED ||
1374                     cm_node->state == IRDMA_CM_STATE_CLOSED) {
1375                         irdma_free_retrans_entry(cm_node);
1376                         goto done;
1377                 }
1378
1379                 if (!send_entry->retranscount || !send_entry->retrycount) {
1380                         irdma_free_retrans_entry(cm_node);
1381
1382                         spin_unlock_irqrestore(&cm_node->retrans_list_lock,
1383                                                flags);
1384                         irdma_retrans_expired(cm_node);
1385                         cm_node->state = IRDMA_CM_STATE_CLOSED;
1386                         spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1387                         goto done;
1388                 }
1389                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1390
1391                 vsi = &cm_node->iwdev->vsi;
1392                 if (!cm_node->ack_rcvd) {
1393                         atomic_inc(&send_entry->sqbuf->refcount);
1394                         irdma_puda_send_buf(vsi->ilq, send_entry->sqbuf);
1395                         cm_node->cm_core->stats_pkt_retrans++;
1396                 }
1397
1398                 spin_lock_irqsave(&cm_node->retrans_list_lock, flags);
1399                 if (send_entry->send_retrans) {
1400                         send_entry->retranscount--;
1401                         timetosend = (IRDMA_RETRY_TIMEOUT <<
1402                                       (IRDMA_DEFAULT_RETRANS -
1403                                        send_entry->retranscount));
1404
1405                         send_entry->timetosend = jiffies +
1406                             min(timetosend, IRDMA_MAX_TIMEOUT);
1407                         if (nexttimeout > send_entry->timetosend || !settimer) {
1408                                 nexttimeout = send_entry->timetosend;
1409                                 settimer = 1;
1410                         }
1411                 } else {
1412                         int close_when_complete;
1413
1414                         close_when_complete = send_entry->close_when_complete;
1415                         irdma_free_retrans_entry(cm_node);
1416                         if (close_when_complete)
1417                                 irdma_rem_ref_cm_node(cm_node);
1418                 }
1419 done:
1420                 spin_unlock_irqrestore(&cm_node->retrans_list_lock, flags);
1421                 irdma_rem_ref_cm_node(cm_node);
1422         }
1423
1424         if (settimer) {
1425                 spin_lock_irqsave(&cm_core->ht_lock, flags);
1426                 if (!timer_pending(&cm_core->tcp_timer)) {
1427                         cm_core->tcp_timer.expires = nexttimeout;
1428                         add_timer(&cm_core->tcp_timer);
1429                 }
1430                 spin_unlock_irqrestore(&cm_core->ht_lock, flags);
1431         }
1432 }
1433
1434 /**
1435  * irdma_send_syn - send SYN packet
1436  * @cm_node: connection's node
1437  * @sendack: flag to set ACK bit or not
1438  */
1439 int
1440 irdma_send_syn(struct irdma_cm_node *cm_node, u32 sendack)
1441 {
1442         struct irdma_puda_buf *sqbuf;
1443         int flags = SET_SYN;
1444         char optionsbuf[sizeof(struct option_mss) +
1445                         sizeof(struct option_windowscale) +
1446                         sizeof(struct option_base) + TCP_OPTIONS_PADDING];
1447         struct irdma_kmem_info opts;
1448         int optionssize = 0;
1449         /* Sending MSS option */
1450         union all_known_options *options;
1451
1452         opts.addr = optionsbuf;
1453         if (!cm_node)
1454                 return -EINVAL;
1455
1456         options = (union all_known_options *)&optionsbuf[optionssize];
1457         options->mss.optionnum = OPTION_NUM_MSS;
1458         options->mss.len = sizeof(struct option_mss);
1459         options->mss.mss = htons(cm_node->tcp_cntxt.mss);
1460         optionssize += sizeof(struct option_mss);
1461
1462         options = (union all_known_options *)&optionsbuf[optionssize];
1463         options->windowscale.optionnum = OPTION_NUM_WINDOW_SCALE;
1464         options->windowscale.len = sizeof(struct option_windowscale);
1465         options->windowscale.shiftcount = cm_node->tcp_cntxt.rcv_wscale;
1466         optionssize += sizeof(struct option_windowscale);
1467         options = (union all_known_options *)&optionsbuf[optionssize];
1468         options->eol = OPTION_NUM_EOL;
1469         optionssize += 1;
1470
1471         if (sendack)
1472                 flags |= SET_ACK;
1473
1474         opts.size = optionssize;
1475
1476         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, &opts, NULL, NULL,
1477                                                 flags);
1478         if (!sqbuf)
1479                 return -ENOMEM;
1480
1481         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
1482                                        0);
1483 }
1484
1485 /**
1486  * irdma_send_ack - Send ACK packet
1487  * @cm_node: connection's node
1488  */
1489 void
1490 irdma_send_ack(struct irdma_cm_node *cm_node)
1491 {
1492         struct irdma_puda_buf *sqbuf;
1493         struct irdma_sc_vsi *vsi = &cm_node->iwdev->vsi;
1494
1495         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
1496                                                 SET_ACK);
1497         if (sqbuf)
1498                 irdma_puda_send_buf(vsi->ilq, sqbuf);
1499 }
1500
1501 /**
1502  * irdma_send_fin - Send FIN pkt
1503  * @cm_node: connection's node
1504  */
1505 static int
1506 irdma_send_fin(struct irdma_cm_node *cm_node)
1507 {
1508         struct irdma_puda_buf *sqbuf;
1509
1510         sqbuf = cm_node->cm_core->form_cm_frame(cm_node, NULL, NULL, NULL,
1511                                                 SET_ACK | SET_FIN);
1512         if (!sqbuf)
1513                 return -ENOMEM;
1514
1515         return irdma_schedule_cm_timer(cm_node, sqbuf, IRDMA_TIMER_TYPE_SEND, 1,
1516                                        0);
1517 }
1518
1519 /**
1520  * irdma_find_listener - find a cm node listening on this addr-port pair
1521  * @cm_core: cm's core
1522  * @dst_addr: listener ip addr
1523  * @dst_port: listener tcp port num
1524  * @vlan_id: virtual LAN ID
1525  * @listener_state: state to match with listen node's
1526  */
1527 static struct irdma_cm_listener *
1528 irdma_find_listener(struct irdma_cm_core *cm_core, u32 *dst_addr, u16 dst_port,
1529                     u16 vlan_id, enum irdma_cm_listener_state listener_state)
1530 {
1531         struct irdma_cm_listener *listen_node;
1532         static const u32 ip_zero[4] = {0, 0, 0, 0};
1533         u32 listen_addr[4];
1534         u16 listen_port;
1535         unsigned long flags;
1536
1537         /* walk list and find cm_node associated with this session ID */
1538         spin_lock_irqsave(&cm_core->listen_list_lock, flags);
1539         list_for_each_entry(listen_node, &cm_core->listen_list, list) {
1540                 memcpy(listen_addr, listen_node->loc_addr, sizeof(listen_addr));
1541                 listen_port = listen_node->loc_port;
1542                 if (listen_port != dst_port ||
1543                     !(listener_state & listen_node->listener_state))
1544                         continue;
1545                 /* compare node pair, return node handle if a match */
1546                 if (!memcmp(listen_addr, ip_zero, sizeof(listen_addr)) ||
1547                     (!memcmp(listen_addr, dst_addr, sizeof(listen_addr)) &&
1548                      vlan_id == listen_node->vlan_id)) {
1549                         atomic_inc(&listen_node->refcnt);
1550                         spin_unlock_irqrestore(&cm_core->listen_list_lock,
1551                                                flags);
1552                         return listen_node;
1553                 }
1554         }
1555         spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
1556
1557         return NULL;
1558 }
1559
1560 /**
1561  * irdma_del_multiple_qhash - Remove qhash and child listens
1562  * @iwdev: iWarp device
1563  * @cm_info: CM info for parent listen node
1564  * @cm_parent_listen_node: The parent listen node
1565  */
1566 static int
1567 irdma_del_multiple_qhash(struct irdma_device *iwdev,
1568                          struct irdma_cm_info *cm_info,
1569                          struct irdma_cm_listener *cm_parent_listen_node)
1570 {
1571         struct irdma_cm_listener *child_listen_node;
1572         struct list_head *pos, *tpos;
1573         unsigned long flags;
1574         int ret = -EINVAL;
1575
1576         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock, flags);
1577         list_for_each_safe(pos, tpos,
1578                            &cm_parent_listen_node->child_listen_list) {
1579                 child_listen_node = list_entry(pos, struct irdma_cm_listener,
1580                                                child_listen_list);
1581                 if (child_listen_node->ipv4)
1582                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1583                                     "removing child listen for IP=%pI4, port=%d, vlan=%d\n",
1584                                     child_listen_node->loc_addr,
1585                                     child_listen_node->loc_port,
1586                                     child_listen_node->vlan_id);
1587                 else
1588                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1589                                     "removing child listen for IP=%pI6, port=%d, vlan=%d\n",
1590                                     child_listen_node->loc_addr,
1591                                     child_listen_node->loc_port,
1592                                     child_listen_node->vlan_id);
1593                 list_del(pos);
1594                 memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1595                        sizeof(cm_info->loc_addr));
1596                 cm_info->vlan_id = child_listen_node->vlan_id;
1597                 if (child_listen_node->qhash_set) {
1598                         ret = irdma_manage_qhash(iwdev, cm_info,
1599                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1600                                                  IRDMA_QHASH_MANAGE_TYPE_DELETE,
1601                                                  NULL, false);
1602                         child_listen_node->qhash_set = false;
1603                 } else {
1604                         ret = 0;
1605                 }
1606                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1607                             "Child listen node freed = %p\n",
1608                             child_listen_node);
1609                 kfree(child_listen_node);
1610                 cm_parent_listen_node->cm_core->stats_listen_nodes_destroyed++;
1611         }
1612         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1613
1614         return ret;
1615 }
1616
1617 /**
1618  * irdma_netdev_vlan_ipv6 - Gets the netdev and mac
1619  * @addr: local IPv6 address
1620  * @vlan_id: vlan id for the given IPv6 address
1621  * @mac: mac address for the given IPv6 address
1622  *
1623  * Returns the net_device of the IPv6 address and also sets the
1624  * vlan id and mac for that address.
1625  */
1626 struct ifnet *
1627 irdma_netdev_vlan_ipv6(u32 *addr, u16 *vlan_id, u8 *mac)
1628 {
1629         struct ifnet *ip_dev = NULL;
1630         struct in6_addr laddr6;
1631
1632         irdma_copy_ip_htonl(laddr6.__u6_addr.__u6_addr32, addr);
1633         if (vlan_id)
1634                 *vlan_id = 0xFFFF;      /* Match rdma_vlan_dev_vlan_id() */
1635         if (mac)
1636                 eth_zero_addr(mac);
1637
1638         ip_dev = ip6_ifp_find(&init_net, laddr6, 0);
1639         if (ip_dev) {
1640                 if (vlan_id)
1641                         *vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1642                 if (ip_dev->if_addr && ip_dev->if_addr->ifa_addr && mac)
1643                         ether_addr_copy(mac, IF_LLADDR(ip_dev));
1644         }
1645
1646         return ip_dev;
1647 }
1648
1649 /**
1650  * irdma_get_vlan_ipv4 - Returns the vlan_id for IPv4 address
1651  * @addr: local IPv4 address
1652  */
1653 u16
1654 irdma_get_vlan_ipv4(u32 *addr)
1655 {
1656         struct ifnet *netdev;
1657         u16 vlan_id = 0xFFFF;
1658
1659         netdev = ip_ifp_find(&init_net, htonl(addr[0]));
1660         if (netdev) {
1661                 vlan_id = rdma_vlan_dev_vlan_id(netdev);
1662                 dev_put(netdev);
1663         }
1664
1665         return vlan_id;
1666 }
1667
1668 /**
1669  * irdma_add_mqh_6 - Adds multiple qhashes for IPv6
1670  * @iwdev: iWarp device
1671  * @cm_info: CM info for parent listen node
1672  * @cm_parent_listen_node: The parent listen node
1673  *
1674  * Adds a qhash and a child listen node for every IPv6 address
1675  * on the adapter and adds the associated qhash filter
1676  */
1677 static int
1678 irdma_add_mqh_6(struct irdma_device *iwdev,
1679                 struct irdma_cm_info *cm_info,
1680                 struct irdma_cm_listener *cm_parent_listen_node)
1681 {
1682         struct ifnet *ip_dev;
1683         struct ifaddr *ifp;
1684         struct irdma_cm_listener *child_listen_node;
1685         unsigned long flags;
1686         int ret = 0;
1687
1688         IFNET_RLOCK();
1689         IRDMA_TAILQ_FOREACH((ip_dev), &V_ifnet, if_link) {
1690                 if (!(ip_dev->if_flags & IFF_UP))
1691                         continue;
1692
1693                 if (((rdma_vlan_dev_vlan_id(ip_dev) >= VLAN_N_VID) ||
1694                      (rdma_vlan_dev_real_dev(ip_dev) != iwdev->netdev)) &&
1695                     ip_dev != iwdev->netdev)
1696                         continue;
1697
1698                 if_addr_rlock(ip_dev);
1699                 IRDMA_TAILQ_FOREACH(ifp, &ip_dev->if_addrhead, ifa_link) {
1700                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1701                                     "IP=%pI6, vlan_id=%d, MAC=%pM\n",
1702                                     &((struct sockaddr_in6 *)ifp->ifa_addr)->sin6_addr, rdma_vlan_dev_vlan_id(ip_dev),
1703                                     IF_LLADDR(ip_dev));
1704                         if (((struct sockaddr_in6 *)ifp->ifa_addr)->sin6_family != AF_INET6)
1705                                 continue;
1706                         child_listen_node = kzalloc(sizeof(*child_listen_node), GFP_KERNEL);
1707                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1708                                     "Allocating child listener %p\n",
1709                                     child_listen_node);
1710                         if (!child_listen_node) {
1711                                 irdma_debug(iwdev_to_idev(iwdev),
1712                                             IRDMA_DEBUG_CM,
1713                                             "listener memory allocation\n");
1714                                 ret = -ENOMEM;
1715                                 if_addr_runlock(ip_dev);
1716                                 goto exit;
1717                         }
1718
1719                         memcpy(child_listen_node, cm_parent_listen_node,
1720                                sizeof(*child_listen_node));
1721                         cm_info->vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1722                         child_listen_node->vlan_id = cm_info->vlan_id;
1723                         irdma_copy_ip_ntohl(child_listen_node->loc_addr,
1724                                             ((struct sockaddr_in6 *)ifp->ifa_addr)->sin6_addr.__u6_addr.__u6_addr32);
1725                         memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1726                                sizeof(cm_info->loc_addr));
1727                         ret = irdma_manage_qhash(iwdev, cm_info,
1728                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1729                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
1730                                                  NULL, true);
1731                         if (ret) {
1732                                 kfree(child_listen_node);
1733                                 continue;
1734                         }
1735
1736                         child_listen_node->qhash_set = true;
1737                         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock, flags);
1738                         list_add(&child_listen_node->child_listen_list,
1739                                  &cm_parent_listen_node->child_listen_list);
1740                         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1741                         cm_parent_listen_node->cm_core->stats_listen_nodes_created++;
1742                 }
1743                 if_addr_runlock(ip_dev);
1744         }
1745 exit:
1746         IFNET_RUNLOCK();
1747
1748         return ret;
1749 }
1750
1751 /**
1752  * irdma_add_mqh_4 - Adds multiple qhashes for IPv4
1753  * @iwdev: iWarp device
1754  * @cm_info: CM info for parent listen node
1755  * @cm_parent_listen_node: The parent listen node
1756  *
1757  * Adds a qhash and a child listen node for every IPv4 address
1758  * on the adapter and adds the associated qhash filter
1759  */
1760 static int
1761 irdma_add_mqh_4(struct irdma_device *iwdev,
1762                 struct irdma_cm_info *cm_info,
1763                 struct irdma_cm_listener *cm_parent_listen_node)
1764 {
1765         struct ifnet *ip_dev;
1766         struct irdma_cm_listener *child_listen_node;
1767         unsigned long flags;
1768         struct ifaddr *ifa;
1769         int ret = 0;
1770
1771         IFNET_RLOCK();
1772         IRDMA_TAILQ_FOREACH((ip_dev), &V_ifnet, if_link) {
1773                 if (!(ip_dev->if_flags & IFF_UP))
1774                         continue;
1775
1776                 if (((rdma_vlan_dev_vlan_id(ip_dev) >= VLAN_N_VID) ||
1777                      (rdma_vlan_dev_real_dev(ip_dev) != iwdev->netdev)) &&
1778                     ip_dev != iwdev->netdev)
1779                         continue;
1780
1781                 if_addr_rlock(ip_dev);
1782                 IRDMA_TAILQ_FOREACH(ifa, &ip_dev->if_addrhead, ifa_link) {
1783                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1784                                     "Allocating child CM Listener forIP=%pI4, vlan_id=%d, MAC=%pM\n",
1785                                     &ifa->ifa_addr,
1786                                     rdma_vlan_dev_vlan_id(ip_dev), IF_LLADDR(ip_dev));
1787                         if (((struct sockaddr_in *)ifa->ifa_addr)->sin_family != AF_INET)
1788                                 continue;
1789                         child_listen_node = kzalloc(sizeof(*child_listen_node), GFP_KERNEL);
1790                         cm_parent_listen_node->cm_core->stats_listen_nodes_created++;
1791                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1792                                     "Allocating child listener %p\n",
1793                                     child_listen_node);
1794                         if (!child_listen_node) {
1795                                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
1796                                             "listener memory allocation\n");
1797                                 if_addr_runlock(ip_dev);
1798                                 ret = -ENOMEM;
1799                                 goto exit;
1800                         }
1801
1802                         memcpy(child_listen_node, cm_parent_listen_node,
1803                                sizeof(*child_listen_node));
1804                         child_listen_node->vlan_id = rdma_vlan_dev_vlan_id(ip_dev);
1805                         cm_info->vlan_id = child_listen_node->vlan_id;
1806                         child_listen_node->loc_addr[0] =
1807                             ntohl(((struct sockaddr_in *)ifa->ifa_addr)->sin_addr.s_addr);
1808                         memcpy(cm_info->loc_addr, child_listen_node->loc_addr,
1809                                sizeof(cm_info->loc_addr));
1810                         ret = irdma_manage_qhash(iwdev, cm_info,
1811                                                  IRDMA_QHASH_TYPE_TCP_SYN,
1812                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
1813                                                  NULL, true);
1814                         if (ret) {
1815                                 kfree(child_listen_node);
1816                                 cm_parent_listen_node->cm_core
1817                                     ->stats_listen_nodes_created--;
1818                                 continue;
1819                         }
1820
1821                         child_listen_node->qhash_set = true;
1822                         spin_lock_irqsave(&iwdev->cm_core.listen_list_lock,
1823                                           flags);
1824                         list_add(&child_listen_node->child_listen_list,
1825                                  &cm_parent_listen_node->child_listen_list);
1826                         spin_unlock_irqrestore(&iwdev->cm_core.listen_list_lock, flags);
1827                 }
1828                 if_addr_runlock(ip_dev);
1829         }
1830 exit:
1831         IFNET_RUNLOCK();
1832
1833         return ret;
1834 }
1835
1836 /**
1837  * irdma_add_mqh - Adds multiple qhashes
1838  * @iwdev: iWarp device
1839  * @cm_info: CM info for parent listen node
1840  * @cm_listen_node: The parent listen node
1841  */
1842 static int
1843 irdma_add_mqh(struct irdma_device *iwdev,
1844               struct irdma_cm_info *cm_info,
1845               struct irdma_cm_listener *cm_listen_node)
1846 {
1847         int err;
1848         VNET_ITERATOR_DECL(vnet_iter);
1849
1850         VNET_LIST_RLOCK();
1851         VNET_FOREACH(vnet_iter) {
1852                 IFNET_RLOCK();
1853                 CURVNET_SET_QUIET(vnet_iter);
1854
1855                 if (cm_info->ipv4)
1856                         err = irdma_add_mqh_4(iwdev, cm_info, cm_listen_node);
1857                 else
1858                         err = irdma_add_mqh_6(iwdev, cm_info, cm_listen_node);
1859                 CURVNET_RESTORE();
1860                 IFNET_RUNLOCK();
1861         }
1862         VNET_LIST_RUNLOCK();
1863
1864         return err;
1865 }
1866
1867 /**
1868  * irdma_reset_list_prep - add connection nodes slated for reset to list
1869  * @cm_core: cm's core
1870  * @listener: pointer to listener node
1871  * @reset_list: a list to which cm_node will be selected
1872  */
1873 static void
1874 irdma_reset_list_prep(struct irdma_cm_core *cm_core,
1875                       struct irdma_cm_listener *listener,
1876                       struct list_head *reset_list)
1877 {
1878         struct irdma_cm_node *cm_node;
1879         int bkt;
1880
1881         HASH_FOR_EACH_RCU(cm_core->cm_hash_tbl, bkt, cm_node, list) {
1882                 if (cm_node->listener == listener &&
1883                     !cm_node->accelerated &&
1884                     atomic_inc_not_zero(&cm_node->refcnt))
1885                         list_add(&cm_node->reset_entry, reset_list);
1886         }
1887 }
1888
1889 /**
1890  * irdma_dec_refcnt_listen - delete listener and associated cm nodes
1891  * @cm_core: cm's core
1892  * @listener: pointer to listener node
1893  * @free_hanging_nodes: to free associated cm_nodes
1894  * @apbvt_del: flag to delete the apbvt
1895  */
1896 static int
1897 irdma_dec_refcnt_listen(struct irdma_cm_core *cm_core,
1898                         struct irdma_cm_listener *listener,
1899                         int free_hanging_nodes, bool apbvt_del)
1900 {
1901         struct list_head *list_pos;
1902         struct list_head *list_temp;
1903         struct irdma_cm_node *cm_node;
1904         struct list_head reset_list;
1905         struct irdma_cm_info nfo;
1906         enum irdma_cm_node_state old_state;
1907         unsigned long flags;
1908         int err;
1909
1910         /* free non-accelerated child nodes for this listener */
1911         INIT_LIST_HEAD(&reset_list);
1912         if (free_hanging_nodes) {
1913                 rcu_read_lock();
1914                 irdma_reset_list_prep(cm_core, listener, &reset_list);
1915                 rcu_read_unlock();
1916         }
1917
1918         list_for_each_safe(list_pos, list_temp, &reset_list) {
1919                 cm_node = container_of(list_pos, struct irdma_cm_node,
1920                                        reset_entry);
1921                 if (cm_node->state >= IRDMA_CM_STATE_FIN_WAIT1) {
1922                         irdma_rem_ref_cm_node(cm_node);
1923                         continue;
1924                 }
1925
1926                 irdma_cleanup_retrans_entry(cm_node);
1927                 err = irdma_send_reset(cm_node);
1928                 if (err) {
1929                         cm_node->state = IRDMA_CM_STATE_CLOSED;
1930                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
1931                                     IRDMA_DEBUG_CM, "send reset failed\n");
1932                 } else {
1933                         old_state = cm_node->state;
1934                         cm_node->state = IRDMA_CM_STATE_LISTENER_DESTROYED;
1935                         if (old_state != IRDMA_CM_STATE_MPAREQ_RCVD)
1936                                 irdma_rem_ref_cm_node(cm_node);
1937                 }
1938         }
1939
1940         if (atomic_dec_and_test(&listener->refcnt)) {
1941                 spin_lock_irqsave(&cm_core->listen_list_lock, flags);
1942                 list_del(&listener->list);
1943                 spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
1944
1945                 if (apbvt_del)
1946                         irdma_del_apbvt(listener->iwdev,
1947                                         listener->apbvt_entry);
1948                 memcpy(nfo.loc_addr, listener->loc_addr, sizeof(nfo.loc_addr));
1949                 nfo.loc_port = listener->loc_port;
1950                 nfo.ipv4 = listener->ipv4;
1951                 nfo.vlan_id = listener->vlan_id;
1952                 nfo.user_pri = listener->user_pri;
1953                 nfo.qh_qpid = listener->iwdev->vsi.ilq->qp_id;
1954
1955                 if (!list_empty(&listener->child_listen_list)) {
1956                         irdma_del_multiple_qhash(listener->iwdev, &nfo,
1957                                                  listener);
1958                 } else {
1959                         if (listener->qhash_set)
1960                                 irdma_manage_qhash(listener->iwdev,
1961                                                    &nfo,
1962                                                    IRDMA_QHASH_TYPE_TCP_SYN,
1963                                                    IRDMA_QHASH_MANAGE_TYPE_DELETE,
1964                                                    NULL, false);
1965                 }
1966
1967                 cm_core->stats_listen_destroyed++;
1968                 cm_core->stats_listen_nodes_destroyed++;
1969                 irdma_debug(iwdev_to_idev(listener->iwdev), IRDMA_DEBUG_CM,
1970                     "loc_port=0x%04x loc_addr=%pI4 cm_listen_node=%p cm_id=%p qhash_set=%d vlan_id=%d apbvt_del=%d\n",
1971                             listener->loc_port, listener->loc_addr, listener, listener->cm_id, listener->qhash_set,
1972                             listener->vlan_id, apbvt_del);
1973                 kfree(listener);
1974                 listener = NULL;
1975                 return 0;
1976         }
1977
1978         return -EINVAL;
1979 }
1980
1981 /**
1982  * irdma_cm_del_listen - delete a listener
1983  * @cm_core: cm's core
1984  * @listener: passive connection's listener
1985  * @apbvt_del: flag to delete apbvt
1986  */
1987 static int
1988 irdma_cm_del_listen(struct irdma_cm_core *cm_core,
1989                     struct irdma_cm_listener *listener,
1990                     bool apbvt_del)
1991 {
1992         listener->listener_state = IRDMA_CM_LISTENER_PASSIVE_STATE;
1993         listener->cm_id = NULL;
1994
1995         return irdma_dec_refcnt_listen(cm_core, listener, 1, apbvt_del);
1996 }
1997
1998 /**
1999  * irdma_find_node - find a cm node that matches the reference cm node
2000  * @cm_core: cm's core
2001  * @rem_port: remote tcp port num
2002  * @rem_addr: remote ip addr
2003  * @loc_port: local tcp port num
2004  * @loc_addr: local ip addr
2005  * @vlan_id: local VLAN ID
2006  */
2007 struct irdma_cm_node *
2008 irdma_find_node(struct irdma_cm_core *cm_core,
2009                 u16 rem_port, u32 *rem_addr, u16 loc_port,
2010                 u32 *loc_addr, u16 vlan_id)
2011 {
2012         struct irdma_cm_node *cm_node;
2013         u32 key = (rem_port << 16) | loc_port;
2014
2015         rcu_read_lock();
2016         HASH_FOR_EACH_POSSIBLE_RCU(cm_core->cm_hash_tbl, cm_node, list, key) {
2017                 if (cm_node->vlan_id == vlan_id &&
2018                     cm_node->loc_port == loc_port && cm_node->rem_port == rem_port &&
2019                     !memcmp(cm_node->loc_addr, loc_addr, sizeof(cm_node->loc_addr)) &&
2020                     !memcmp(cm_node->rem_addr, rem_addr, sizeof(cm_node->rem_addr))) {
2021                         if (!atomic_inc_not_zero(&cm_node->refcnt))
2022                                 goto exit;
2023                         rcu_read_unlock();
2024                         return cm_node;
2025                 }
2026         }
2027
2028 exit:
2029         rcu_read_unlock();
2030
2031         /* no owner node */
2032         return NULL;
2033 }
2034
2035 /**
2036  * irdma_add_hte_node - add a cm node to the hash table
2037  * @cm_core: cm's core
2038  * @cm_node: connection's node
2039  */
2040 static void
2041 irdma_add_hte_node(struct irdma_cm_core *cm_core,
2042                    struct irdma_cm_node *cm_node)
2043 {
2044         unsigned long flags;
2045         u32 key = (cm_node->rem_port << 16) | cm_node->loc_port;
2046
2047         spin_lock_irqsave(&cm_core->ht_lock, flags);
2048         HASH_ADD_RCU(cm_core->cm_hash_tbl, &cm_node->list, key);
2049         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2050 }
2051
2052 /**
2053  * irdma_ipv4_is_lpb - check if loopback
2054  * @loc_addr: local addr to compare
2055  * @rem_addr: remote address
2056  */
2057 bool
2058 irdma_ipv4_is_lpb(u32 loc_addr, u32 rem_addr)
2059 {
2060         return ipv4_is_loopback(htonl(rem_addr)) || (loc_addr == rem_addr);
2061 }
2062
2063 /**
2064  * irdma_ipv6_is_lpb - check if loopback
2065  * @loc_addr: local addr to compare
2066  * @rem_addr: remote address
2067  */
2068 bool
2069 irdma_ipv6_is_lpb(u32 *loc_addr, u32 *rem_addr)
2070 {
2071         struct in6_addr raddr6;
2072
2073         irdma_copy_ip_htonl(raddr6.__u6_addr.__u6_addr32, rem_addr);
2074
2075         return !memcmp(loc_addr, rem_addr, 16) || ipv6_addr_loopback(&raddr6);
2076 }
2077
2078 /**
2079  * irdma_cm_create_ah - create a cm address handle
2080  * @cm_node: The connection manager node to create AH for
2081  * @wait: Provides option to wait for ah creation or not
2082  */
2083 static int
2084 irdma_cm_create_ah(struct irdma_cm_node *cm_node, bool wait)
2085 {
2086         struct irdma_ah_info ah_info = {0};
2087         struct irdma_device *iwdev = cm_node->iwdev;
2088
2089         ether_addr_copy(ah_info.mac_addr, IF_LLADDR(iwdev->netdev));
2090
2091         ah_info.hop_ttl = 0x40;
2092         ah_info.tc_tos = cm_node->tos;
2093         ah_info.vsi = &iwdev->vsi;
2094
2095         if (cm_node->ipv4) {
2096                 ah_info.ipv4_valid = true;
2097                 ah_info.dest_ip_addr[0] = cm_node->rem_addr[0];
2098                 ah_info.src_ip_addr[0] = cm_node->loc_addr[0];
2099                 ah_info.do_lpbk = irdma_ipv4_is_lpb(ah_info.src_ip_addr[0],
2100                                                     ah_info.dest_ip_addr[0]);
2101         } else {
2102                 memcpy(ah_info.dest_ip_addr, cm_node->rem_addr,
2103                        sizeof(ah_info.dest_ip_addr));
2104                 memcpy(ah_info.src_ip_addr, cm_node->loc_addr,
2105                        sizeof(ah_info.src_ip_addr));
2106                 ah_info.do_lpbk = irdma_ipv6_is_lpb(ah_info.src_ip_addr,
2107                                                     ah_info.dest_ip_addr);
2108         }
2109
2110         ah_info.vlan_tag = cm_node->vlan_id;
2111         if (cm_node->vlan_id < VLAN_N_VID) {
2112                 ah_info.insert_vlan_tag = 1;
2113                 ah_info.vlan_tag |= cm_node->user_pri << VLAN_PRIO_SHIFT;
2114         }
2115
2116         ah_info.dst_arpindex =
2117             irdma_arp_table(iwdev->rf, ah_info.dest_ip_addr,
2118                             NULL, IRDMA_ARP_RESOLVE);
2119
2120         if (irdma_puda_create_ah(&iwdev->rf->sc_dev, &ah_info, wait,
2121                                  IRDMA_PUDA_RSRC_TYPE_ILQ, cm_node,
2122                                  &cm_node->ah))
2123                 return -ENOMEM;
2124
2125         return 0;
2126 }
2127
2128 /**
2129  * irdma_cm_free_ah - free a cm address handle
2130  * @cm_node: The connection manager node to create AH for
2131  */
2132 static void
2133 irdma_cm_free_ah(struct irdma_cm_node *cm_node)
2134 {
2135         struct irdma_device *iwdev = cm_node->iwdev;
2136
2137         irdma_puda_free_ah(&iwdev->rf->sc_dev, cm_node->ah);
2138         cm_node->ah = NULL;
2139 }
2140
2141 /**
2142  * irdma_make_cm_node - create a new instance of a cm node
2143  * @cm_core: cm's core
2144  * @iwdev: iwarp device structure
2145  * @cm_info: quad info for connection
2146  * @listener: passive connection's listener
2147  */
2148 static struct irdma_cm_node *
2149 irdma_make_cm_node(struct irdma_cm_core *cm_core, struct irdma_device *iwdev,
2150                    struct irdma_cm_info *cm_info,
2151                    struct irdma_cm_listener *listener)
2152 {
2153         struct irdma_cm_node *cm_node;
2154         int arpindex;
2155         struct ifnet *netdev = iwdev->netdev;
2156
2157         /* create an hte and cm_node for this instance */
2158         cm_node = kzalloc(sizeof(*cm_node), GFP_ATOMIC);
2159         if (!cm_node)
2160                 return NULL;
2161
2162         /* set our node specific transport info */
2163         cm_node->ipv4 = cm_info->ipv4;
2164         cm_node->vlan_id = cm_info->vlan_id;
2165         if (cm_node->vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
2166                 cm_node->vlan_id = 0;
2167         cm_node->tos = cm_info->tos;
2168         cm_node->user_pri = cm_info->user_pri;
2169         if (listener) {
2170                 if (listener->tos != cm_info->tos)
2171                         irdma_dev_warn(
2172                                        &iwdev->rf->sc_dev,
2173                                        "application TOS[%d] and remote client TOS[%d] mismatch\n",
2174                                        listener->tos, cm_info->tos);
2175                 if (iwdev->vsi.dscp_mode) {
2176                         cm_node->user_pri = listener->user_pri;
2177                 } else {
2178                         cm_node->tos = max(listener->tos, cm_info->tos);
2179                         cm_node->user_pri = rt_tos2priority(cm_node->tos);
2180                 }
2181                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_DCB,
2182                             "listener: TOS:[%d] UP:[%d]\n", cm_node->tos,
2183                             cm_node->user_pri);
2184         }
2185         memcpy(cm_node->loc_addr, cm_info->loc_addr, sizeof(cm_node->loc_addr));
2186         memcpy(cm_node->rem_addr, cm_info->rem_addr, sizeof(cm_node->rem_addr));
2187         cm_node->loc_port = cm_info->loc_port;
2188         cm_node->rem_port = cm_info->rem_port;
2189
2190         cm_node->mpa_frame_rev = IRDMA_CM_DEFAULT_MPA_VER;
2191         cm_node->send_rdma0_op = SEND_RDMA_READ_ZERO;
2192         cm_node->iwdev = iwdev;
2193         cm_node->dev = &iwdev->rf->sc_dev;
2194
2195         cm_node->ird_size = cm_node->dev->hw_attrs.max_hw_ird;
2196         cm_node->ord_size = cm_node->dev->hw_attrs.max_hw_ord;
2197
2198         cm_node->listener = listener;
2199         cm_node->cm_id = cm_info->cm_id;
2200         ether_addr_copy(cm_node->loc_mac, IF_LLADDR(netdev));
2201         spin_lock_init(&cm_node->retrans_list_lock);
2202         cm_node->ack_rcvd = false;
2203
2204         init_completion(&cm_node->establish_comp);
2205         atomic_set(&cm_node->refcnt, 1);
2206         /* associate our parent CM core */
2207         cm_node->cm_core = cm_core;
2208         cm_node->tcp_cntxt.loc_id = IRDMA_CM_DEFAULT_LOCAL_ID;
2209         cm_node->tcp_cntxt.rcv_wscale = iwdev->rcv_wscale;
2210         cm_node->tcp_cntxt.rcv_wnd = iwdev->rcv_wnd >> cm_node->tcp_cntxt.rcv_wscale;
2211         kc_set_loc_seq_num_mss(cm_node);
2212
2213         arpindex = irdma_resolve_neigh_lpb_chk(iwdev, cm_node, cm_info);
2214         if (arpindex < 0)
2215                 goto err;
2216
2217         ether_addr_copy(cm_node->rem_mac, iwdev->rf->arp_table[arpindex].mac_addr);
2218         irdma_add_hte_node(cm_core, cm_node);
2219         cm_core->stats_nodes_created++;
2220         return cm_node;
2221
2222 err:
2223         kfree(cm_node);
2224
2225         return NULL;
2226 }
2227
2228 static void
2229 irdma_cm_node_free_cb(struct rcu_head *rcu_head)
2230 {
2231         struct irdma_cm_node *cm_node =
2232         container_of(rcu_head, struct irdma_cm_node, rcu_head);
2233         struct irdma_cm_core *cm_core = cm_node->cm_core;
2234         struct irdma_qp *iwqp;
2235         struct irdma_cm_info nfo;
2236
2237         /* if the node is destroyed before connection was accelerated */
2238         if (!cm_node->accelerated && cm_node->accept_pend) {
2239                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2240                             "node destroyed before established\n");
2241                 atomic_dec(&cm_node->listener->pend_accepts_cnt);
2242         }
2243         if (cm_node->close_entry)
2244                 irdma_handle_close_entry(cm_node, 0);
2245         if (cm_node->listener) {
2246                 irdma_dec_refcnt_listen(cm_core, cm_node->listener, 0, true);
2247         } else {
2248                 if (cm_node->apbvt_set) {
2249                         irdma_del_apbvt(cm_node->iwdev, cm_node->apbvt_entry);
2250                         cm_node->apbvt_set = 0;
2251                 }
2252                 irdma_get_addr_info(cm_node, &nfo);
2253                 if (cm_node->qhash_set) {
2254                         nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2255                         irdma_manage_qhash(cm_node->iwdev, &nfo,
2256                                            IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2257                                            IRDMA_QHASH_MANAGE_TYPE_DELETE, NULL,
2258                                            false);
2259                         cm_node->qhash_set = 0;
2260                 }
2261         }
2262
2263         iwqp = cm_node->iwqp;
2264         if (iwqp) {
2265                 cm_node->cm_id->rem_ref(cm_node->cm_id);
2266                 cm_node->cm_id = NULL;
2267                 iwqp->cm_id = NULL;
2268                 irdma_qp_rem_ref(&iwqp->ibqp);
2269                 cm_node->iwqp = NULL;
2270         } else if (cm_node->qhash_set) {
2271                 irdma_get_addr_info(cm_node, &nfo);
2272                 nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2273                 irdma_manage_qhash(cm_node->iwdev, &nfo,
2274                                    IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2275                                    IRDMA_QHASH_MANAGE_TYPE_DELETE, NULL, false);
2276                 cm_node->qhash_set = 0;
2277         }
2278
2279         cm_core->cm_free_ah(cm_node);
2280         kfree(cm_node);
2281 }
2282
2283 /**
2284  * irdma_rem_ref_cm_node - destroy an instance of a cm node
2285  * @cm_node: connection's node
2286  */
2287 void
2288 irdma_rem_ref_cm_node(struct irdma_cm_node *cm_node)
2289 {
2290         struct irdma_cm_core *cm_core = cm_node->cm_core;
2291         unsigned long flags;
2292
2293         spin_lock_irqsave(&cm_core->ht_lock, flags);
2294
2295         if (!atomic_dec_and_test(&cm_node->refcnt)) {
2296                 spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2297                 return;
2298         }
2299         if (cm_node->iwqp) {
2300                 cm_node->iwqp->cm_node = NULL;
2301                 cm_node->iwqp->cm_id = NULL;
2302         }
2303         HASH_DEL_RCU(cm_core->cm_hash_tbl, &cm_node->list);
2304         cm_node->cm_core->stats_nodes_destroyed++;
2305
2306         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
2307
2308         /* wait for all list walkers to exit their grace period */
2309         call_rcu(&cm_node->rcu_head, irdma_cm_node_free_cb);
2310 }
2311
2312 /**
2313  * irdma_handle_fin_pkt - FIN packet received
2314  * @cm_node: connection's node
2315  */
2316 static void
2317 irdma_handle_fin_pkt(struct irdma_cm_node *cm_node)
2318 {
2319         switch (cm_node->state) {
2320         case IRDMA_CM_STATE_SYN_RCVD:
2321         case IRDMA_CM_STATE_SYN_SENT:
2322         case IRDMA_CM_STATE_ESTABLISHED:
2323         case IRDMA_CM_STATE_MPAREJ_RCVD:
2324                 cm_node->tcp_cntxt.rcv_nxt++;
2325                 irdma_cleanup_retrans_entry(cm_node);
2326                 cm_node->state = IRDMA_CM_STATE_LAST_ACK;
2327                 irdma_send_fin(cm_node);
2328                 break;
2329         case IRDMA_CM_STATE_MPAREQ_SENT:
2330                 irdma_create_event(cm_node, IRDMA_CM_EVENT_ABORTED);
2331                 cm_node->tcp_cntxt.rcv_nxt++;
2332                 irdma_cleanup_retrans_entry(cm_node);
2333                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2334                 atomic_inc(&cm_node->refcnt);
2335                 irdma_send_reset(cm_node);
2336                 break;
2337         case IRDMA_CM_STATE_FIN_WAIT1:
2338                 cm_node->tcp_cntxt.rcv_nxt++;
2339                 irdma_cleanup_retrans_entry(cm_node);
2340                 cm_node->state = IRDMA_CM_STATE_CLOSING;
2341                 irdma_send_ack(cm_node);
2342                 /*
2343                  * Wait for ACK as this is simultaneous close. After we receive ACK, do not send anything. Just rm the
2344                  * node.
2345                  */
2346                 break;
2347         case IRDMA_CM_STATE_FIN_WAIT2:
2348                 cm_node->tcp_cntxt.rcv_nxt++;
2349                 irdma_cleanup_retrans_entry(cm_node);
2350                 cm_node->state = IRDMA_CM_STATE_TIME_WAIT;
2351                 irdma_send_ack(cm_node);
2352                 irdma_schedule_cm_timer(cm_node, NULL, IRDMA_TIMER_TYPE_CLOSE,
2353                                         1, 0);
2354                 break;
2355         case IRDMA_CM_STATE_TIME_WAIT:
2356                 cm_node->tcp_cntxt.rcv_nxt++;
2357                 irdma_cleanup_retrans_entry(cm_node);
2358                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2359                 irdma_rem_ref_cm_node(cm_node);
2360                 break;
2361         case IRDMA_CM_STATE_OFFLOADED:
2362         default:
2363                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2364                             "bad state node state = %d\n", cm_node->state);
2365                 break;
2366         }
2367 }
2368
2369 /**
2370  * irdma_handle_rst_pkt - process received RST packet
2371  * @cm_node: connection's node
2372  * @rbuf: receive buffer
2373  */
2374 static void
2375 irdma_handle_rst_pkt(struct irdma_cm_node *cm_node,
2376                      struct irdma_puda_buf *rbuf)
2377 {
2378         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2379                     "caller: %pS cm_node=%p state=%d rem_port=0x%04x loc_port=0x%04x rem_addr=%pI4 loc_addr=%pI4\n",
2380                     __builtin_return_address(0), cm_node, cm_node->state,
2381                     cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr, cm_node->loc_addr);
2382
2383         irdma_cleanup_retrans_entry(cm_node);
2384         switch (cm_node->state) {
2385         case IRDMA_CM_STATE_SYN_SENT:
2386         case IRDMA_CM_STATE_MPAREQ_SENT:
2387                 switch (cm_node->mpa_frame_rev) {
2388                 case IETF_MPA_V2:
2389                         /* Drop down to MPA_V1 */
2390                         cm_node->mpa_frame_rev = IETF_MPA_V1;
2391                         /* send a syn and goto syn sent state */
2392                         cm_node->state = IRDMA_CM_STATE_SYN_SENT;
2393                         if (irdma_send_syn(cm_node, 0))
2394                                 irdma_active_open_err(cm_node, false);
2395                         break;
2396                 case IETF_MPA_V1:
2397                 default:
2398                         irdma_active_open_err(cm_node, false);
2399                         break;
2400                 }
2401                 break;
2402         case IRDMA_CM_STATE_MPAREQ_RCVD:
2403                 atomic_inc(&cm_node->passive_state);
2404                 break;
2405         case IRDMA_CM_STATE_ESTABLISHED:
2406         case IRDMA_CM_STATE_SYN_RCVD:
2407         case IRDMA_CM_STATE_LISTENING:
2408                 irdma_passive_open_err(cm_node, false);
2409                 break;
2410         case IRDMA_CM_STATE_OFFLOADED:
2411                 irdma_active_open_err(cm_node, false);
2412                 break;
2413         case IRDMA_CM_STATE_CLOSED:
2414                 break;
2415         case IRDMA_CM_STATE_FIN_WAIT2:
2416         case IRDMA_CM_STATE_FIN_WAIT1:
2417         case IRDMA_CM_STATE_LAST_ACK:
2418         case IRDMA_CM_STATE_TIME_WAIT:
2419                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2420                 irdma_rem_ref_cm_node(cm_node);
2421                 break;
2422         default:
2423                 break;
2424         }
2425 }
2426
2427 /**
2428  * irdma_handle_rcv_mpa - Process a recv'd mpa buffer
2429  * @cm_node: connection's node
2430  * @rbuf: receive buffer
2431  */
2432 static void
2433 irdma_handle_rcv_mpa(struct irdma_cm_node *cm_node,
2434                      struct irdma_puda_buf *rbuf)
2435 {
2436         int err;
2437         int datasize = rbuf->datalen;
2438         u8 *dataloc = rbuf->data;
2439
2440         enum irdma_cm_event_type type = IRDMA_CM_EVENT_UNKNOWN;
2441         u32 res_type;
2442
2443         err = irdma_parse_mpa(cm_node, dataloc, &res_type, datasize);
2444         if (err) {
2445                 if (cm_node->state == IRDMA_CM_STATE_MPAREQ_SENT)
2446                         irdma_active_open_err(cm_node, true);
2447                 else
2448                         irdma_passive_open_err(cm_node, true);
2449                 return;
2450         }
2451
2452         switch (cm_node->state) {
2453         case IRDMA_CM_STATE_ESTABLISHED:
2454                 if (res_type == IRDMA_MPA_REQUEST_REJECT)
2455                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
2456                                     IRDMA_DEBUG_CM, "state for reject\n");
2457                 cm_node->state = IRDMA_CM_STATE_MPAREQ_RCVD;
2458                 type = IRDMA_CM_EVENT_MPA_REQ;
2459                 irdma_send_ack(cm_node);        /* ACK received MPA request */
2460                 atomic_set(&cm_node->passive_state,
2461                            IRDMA_PASSIVE_STATE_INDICATED);
2462                 break;
2463         case IRDMA_CM_STATE_MPAREQ_SENT:
2464                 irdma_cleanup_retrans_entry(cm_node);
2465                 if (res_type == IRDMA_MPA_REQUEST_REJECT) {
2466                         type = IRDMA_CM_EVENT_MPA_REJECT;
2467                         cm_node->state = IRDMA_CM_STATE_MPAREJ_RCVD;
2468                 } else {
2469                         type = IRDMA_CM_EVENT_CONNECTED;
2470                         cm_node->state = IRDMA_CM_STATE_OFFLOADED;
2471                 }
2472                 irdma_send_ack(cm_node);
2473                 break;
2474         default:
2475                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2476                             "wrong cm_node state =%d\n", cm_node->state);
2477                 break;
2478         }
2479         irdma_create_event(cm_node, type);
2480 }
2481
2482 /**
2483  * irdma_check_syn - Check for error on received syn ack
2484  * @cm_node: connection's node
2485  * @tcph: pointer tcp header
2486  */
2487 static int
2488 irdma_check_syn(struct irdma_cm_node *cm_node, struct tcphdr *tcph)
2489 {
2490         if (ntohl(tcph->th_ack) != cm_node->tcp_cntxt.loc_seq_num) {
2491                 irdma_active_open_err(cm_node, true);
2492                 return 1;
2493         }
2494
2495         return 0;
2496 }
2497
2498 /**
2499  * irdma_check_seq - check seq numbers if OK
2500  * @cm_node: connection's node
2501  * @tcph: pointer tcp header
2502  */
2503 static int
2504 irdma_check_seq(struct irdma_cm_node *cm_node, struct tcphdr *tcph)
2505 {
2506         u32 seq;
2507         u32 ack_seq;
2508         u32 loc_seq_num = cm_node->tcp_cntxt.loc_seq_num;
2509         u32 rcv_nxt = cm_node->tcp_cntxt.rcv_nxt;
2510         u32 rcv_wnd;
2511         int err = 0;
2512
2513         seq = ntohl(tcph->th_seq);
2514         ack_seq = ntohl(tcph->th_ack);
2515         rcv_wnd = cm_node->tcp_cntxt.rcv_wnd;
2516         if (ack_seq != loc_seq_num ||
2517             !between(seq, rcv_nxt, (rcv_nxt + rcv_wnd)))
2518                 err = -1;
2519         if (err)
2520                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2521                             "seq number err\n");
2522
2523         return err;
2524 }
2525
2526 void
2527 irdma_add_conn_est_qh(struct irdma_cm_node *cm_node)
2528 {
2529         struct irdma_cm_info nfo;
2530
2531         irdma_get_addr_info(cm_node, &nfo);
2532         nfo.qh_qpid = cm_node->iwdev->vsi.ilq->qp_id;
2533         irdma_manage_qhash(cm_node->iwdev, &nfo,
2534                            IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
2535                            IRDMA_QHASH_MANAGE_TYPE_ADD,
2536                            cm_node, false);
2537         cm_node->qhash_set = true;
2538 }
2539
2540 /**
2541  * irdma_handle_syn_pkt - is for Passive node
2542  * @cm_node: connection's node
2543  * @rbuf: receive buffer
2544  */
2545 static void
2546 irdma_handle_syn_pkt(struct irdma_cm_node *cm_node,
2547                      struct irdma_puda_buf *rbuf)
2548 {
2549         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2550         int err;
2551         u32 inc_sequence;
2552         int optionsize;
2553
2554         optionsize = (tcph->th_off << 2) - sizeof(struct tcphdr);
2555         inc_sequence = ntohl(tcph->th_seq);
2556
2557         switch (cm_node->state) {
2558         case IRDMA_CM_STATE_SYN_SENT:
2559         case IRDMA_CM_STATE_MPAREQ_SENT:
2560                 /* Rcvd syn on active open connection */
2561                 irdma_active_open_err(cm_node, 1);
2562                 break;
2563         case IRDMA_CM_STATE_LISTENING:
2564                 /* Passive OPEN */
2565                 if (atomic_read(&cm_node->listener->pend_accepts_cnt) >
2566                     cm_node->listener->backlog) {
2567                         cm_node->cm_core->stats_backlog_drops++;
2568                         irdma_passive_open_err(cm_node, false);
2569                         break;
2570                 }
2571                 err = irdma_handle_tcp_options(cm_node, tcph, optionsize, 1);
2572                 if (err) {
2573                         irdma_passive_open_err(cm_node, false);
2574                         /* drop pkt */
2575                         break;
2576                 }
2577                 err = cm_node->cm_core->cm_create_ah(cm_node, false);
2578                 if (err) {
2579                         irdma_passive_open_err(cm_node, false);
2580                         /* drop pkt */
2581                         break;
2582                 }
2583                 cm_node->tcp_cntxt.rcv_nxt = inc_sequence + 1;
2584                 cm_node->accept_pend = 1;
2585                 atomic_inc(&cm_node->listener->pend_accepts_cnt);
2586
2587                 cm_node->state = IRDMA_CM_STATE_SYN_RCVD;
2588                 break;
2589         case IRDMA_CM_STATE_CLOSED:
2590                 irdma_cleanup_retrans_entry(cm_node);
2591                 atomic_inc(&cm_node->refcnt);
2592                 irdma_send_reset(cm_node);
2593                 break;
2594         case IRDMA_CM_STATE_OFFLOADED:
2595         case IRDMA_CM_STATE_ESTABLISHED:
2596         case IRDMA_CM_STATE_FIN_WAIT1:
2597         case IRDMA_CM_STATE_FIN_WAIT2:
2598         case IRDMA_CM_STATE_MPAREQ_RCVD:
2599         case IRDMA_CM_STATE_LAST_ACK:
2600         case IRDMA_CM_STATE_CLOSING:
2601         case IRDMA_CM_STATE_UNKNOWN:
2602         default:
2603                 break;
2604         }
2605 }
2606
2607 /**
2608  * irdma_handle_synack_pkt - Process SYN+ACK packet (active side)
2609  * @cm_node: connection's node
2610  * @rbuf: receive buffer
2611  */
2612 static void
2613 irdma_handle_synack_pkt(struct irdma_cm_node *cm_node,
2614                         struct irdma_puda_buf *rbuf)
2615 {
2616         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2617         int err;
2618         u32 inc_sequence;
2619         int optionsize;
2620
2621         optionsize = (tcph->th_off << 2) - sizeof(struct tcphdr);
2622         inc_sequence = ntohl(tcph->th_seq);
2623         switch (cm_node->state) {
2624         case IRDMA_CM_STATE_SYN_SENT:
2625                 irdma_cleanup_retrans_entry(cm_node);
2626                 /* active open */
2627                 if (irdma_check_syn(cm_node, tcph)) {
2628                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
2629                                     IRDMA_DEBUG_CM, "check syn fail\n");
2630                         return;
2631                 }
2632                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->th_ack);
2633                 /* setup options */
2634                 err = irdma_handle_tcp_options(cm_node, tcph, optionsize, 0);
2635                 if (err) {
2636                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
2637                                     IRDMA_DEBUG_CM,
2638                                     "cm_node=%p tcp_options failed\n",
2639                                     cm_node);
2640                         break;
2641                 }
2642                 irdma_cleanup_retrans_entry(cm_node);
2643                 cm_node->tcp_cntxt.rcv_nxt = inc_sequence + 1;
2644                 irdma_send_ack(cm_node);        /* ACK  for the syn_ack */
2645                 err = irdma_send_mpa_request(cm_node);
2646                 if (err) {
2647                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
2648                                     IRDMA_DEBUG_CM,
2649                                     "cm_node=%p irdma_send_mpa_request failed\n",
2650                                     cm_node);
2651                         break;
2652                 }
2653                 cm_node->state = IRDMA_CM_STATE_MPAREQ_SENT;
2654                 break;
2655         case IRDMA_CM_STATE_MPAREQ_RCVD:
2656                 irdma_passive_open_err(cm_node, true);
2657                 break;
2658         case IRDMA_CM_STATE_LISTENING:
2659                 cm_node->tcp_cntxt.loc_seq_num = ntohl(tcph->th_ack);
2660                 irdma_cleanup_retrans_entry(cm_node);
2661                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2662                 irdma_send_reset(cm_node);
2663                 break;
2664         case IRDMA_CM_STATE_CLOSED:
2665                 cm_node->tcp_cntxt.loc_seq_num = ntohl(tcph->th_ack);
2666                 irdma_cleanup_retrans_entry(cm_node);
2667                 atomic_inc(&cm_node->refcnt);
2668                 irdma_send_reset(cm_node);
2669                 break;
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_LAST_ACK:
2674         case IRDMA_CM_STATE_OFFLOADED:
2675         case IRDMA_CM_STATE_CLOSING:
2676         case IRDMA_CM_STATE_UNKNOWN:
2677         case IRDMA_CM_STATE_MPAREQ_SENT:
2678         default:
2679                 break;
2680         }
2681 }
2682
2683 /**
2684  * irdma_handle_ack_pkt - process packet with ACK
2685  * @cm_node: connection's node
2686  * @rbuf: receive buffer
2687  */
2688 static int
2689 irdma_handle_ack_pkt(struct irdma_cm_node *cm_node,
2690                      struct irdma_puda_buf *rbuf)
2691 {
2692         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2693         u32 inc_sequence;
2694         int ret;
2695         int optionsize;
2696         u32 datasize = rbuf->datalen;
2697
2698         optionsize = (tcph->th_off << 2) - sizeof(struct tcphdr);
2699
2700         if (irdma_check_seq(cm_node, tcph))
2701                 return -EINVAL;
2702
2703         inc_sequence = ntohl(tcph->th_seq);
2704         switch (cm_node->state) {
2705         case IRDMA_CM_STATE_SYN_RCVD:
2706                 irdma_cleanup_retrans_entry(cm_node);
2707                 ret = irdma_handle_tcp_options(cm_node, tcph, optionsize, 1);
2708                 if (ret)
2709                         return ret;
2710                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->th_ack);
2711                 cm_node->state = IRDMA_CM_STATE_ESTABLISHED;
2712                 if (datasize) {
2713                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2714                         irdma_handle_rcv_mpa(cm_node, rbuf);
2715                 }
2716                 break;
2717         case IRDMA_CM_STATE_ESTABLISHED:
2718                 irdma_cleanup_retrans_entry(cm_node);
2719                 if (datasize) {
2720                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2721                         irdma_handle_rcv_mpa(cm_node, rbuf);
2722                 }
2723                 break;
2724         case IRDMA_CM_STATE_MPAREQ_SENT:
2725                 cm_node->tcp_cntxt.rem_ack_num = ntohl(tcph->th_ack);
2726                 if (datasize) {
2727                         cm_node->tcp_cntxt.rcv_nxt = inc_sequence + datasize;
2728                         cm_node->ack_rcvd = false;
2729                         irdma_handle_rcv_mpa(cm_node, rbuf);
2730                 } else {
2731                         cm_node->ack_rcvd = true;
2732                 }
2733                 break;
2734         case IRDMA_CM_STATE_LISTENING:
2735                 irdma_cleanup_retrans_entry(cm_node);
2736                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2737                 irdma_send_reset(cm_node);
2738                 break;
2739         case IRDMA_CM_STATE_CLOSED:
2740                 irdma_cleanup_retrans_entry(cm_node);
2741                 atomic_inc(&cm_node->refcnt);
2742                 irdma_send_reset(cm_node);
2743                 break;
2744         case IRDMA_CM_STATE_LAST_ACK:
2745         case IRDMA_CM_STATE_CLOSING:
2746                 irdma_cleanup_retrans_entry(cm_node);
2747                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2748                 irdma_rem_ref_cm_node(cm_node);
2749                 break;
2750         case IRDMA_CM_STATE_FIN_WAIT1:
2751                 irdma_cleanup_retrans_entry(cm_node);
2752                 cm_node->state = IRDMA_CM_STATE_FIN_WAIT2;
2753                 break;
2754         case IRDMA_CM_STATE_SYN_SENT:
2755         case IRDMA_CM_STATE_FIN_WAIT2:
2756         case IRDMA_CM_STATE_OFFLOADED:
2757         case IRDMA_CM_STATE_MPAREQ_RCVD:
2758         case IRDMA_CM_STATE_UNKNOWN:
2759         default:
2760                 irdma_cleanup_retrans_entry(cm_node);
2761                 break;
2762         }
2763
2764         return 0;
2765 }
2766
2767 /**
2768  * irdma_process_pkt - process cm packet
2769  * @cm_node: connection's node
2770  * @rbuf: receive buffer
2771  */
2772 static void
2773 irdma_process_pkt(struct irdma_cm_node *cm_node,
2774                   struct irdma_puda_buf *rbuf)
2775 {
2776         enum irdma_tcpip_pkt_type pkt_type = IRDMA_PKT_TYPE_UNKNOWN;
2777         struct tcphdr *tcph = (struct tcphdr *)rbuf->tcph;
2778         u32 fin_set = 0;
2779         int err;
2780
2781         if (tcph->th_flags & TH_RST) {
2782                 pkt_type = IRDMA_PKT_TYPE_RST;
2783         } else if (tcph->th_flags & TH_SYN) {
2784                 pkt_type = IRDMA_PKT_TYPE_SYN;
2785                 if (tcph->th_flags & TH_ACK)
2786                         pkt_type = IRDMA_PKT_TYPE_SYNACK;
2787         } else if (tcph->th_flags & TH_ACK) {
2788                 pkt_type = IRDMA_PKT_TYPE_ACK;
2789         }
2790         if (tcph->th_flags & TH_FIN)
2791                 fin_set = 1;
2792
2793         switch (pkt_type) {
2794         case IRDMA_PKT_TYPE_SYN:
2795                 irdma_handle_syn_pkt(cm_node, rbuf);
2796                 break;
2797         case IRDMA_PKT_TYPE_SYNACK:
2798                 irdma_handle_synack_pkt(cm_node, rbuf);
2799                 break;
2800         case IRDMA_PKT_TYPE_ACK:
2801                 err = irdma_handle_ack_pkt(cm_node, rbuf);
2802                 if (fin_set && !err)
2803                         irdma_handle_fin_pkt(cm_node);
2804                 break;
2805         case IRDMA_PKT_TYPE_RST:
2806                 irdma_handle_rst_pkt(cm_node, rbuf);
2807                 break;
2808         default:
2809                 if (fin_set &&
2810                     (!irdma_check_seq(cm_node, (struct tcphdr *)rbuf->tcph)))
2811                         irdma_handle_fin_pkt(cm_node);
2812                 break;
2813         }
2814 }
2815
2816 /**
2817  * irdma_make_listen_node - create a listen node with params
2818  * @cm_core: cm's core
2819  * @iwdev: iwarp device structure
2820  * @cm_info: quad info for connection
2821  */
2822 static struct irdma_cm_listener *
2823 irdma_make_listen_node(struct irdma_cm_core *cm_core,
2824                        struct irdma_device *iwdev,
2825                        struct irdma_cm_info *cm_info)
2826 {
2827         struct irdma_cm_listener *listener;
2828         unsigned long flags;
2829
2830         /* cannot have multiple matching listeners */
2831         listener = irdma_find_listener(cm_core, cm_info->loc_addr,
2832                                        cm_info->loc_port, cm_info->vlan_id,
2833                                        IRDMA_CM_LISTENER_EITHER_STATE);
2834         if (listener &&
2835             listener->listener_state == IRDMA_CM_LISTENER_ACTIVE_STATE) {
2836                 atomic_dec(&listener->refcnt);
2837                 return NULL;
2838         }
2839
2840         if (!listener) {
2841                 /*
2842                  * create a CM listen node 1/2 node to compare incoming traffic to
2843                  */
2844                 listener = kzalloc(sizeof(*listener), GFP_KERNEL);
2845                 if (!listener)
2846                         return NULL;
2847                 cm_core->stats_listen_nodes_created++;
2848                 memcpy(listener->loc_addr, cm_info->loc_addr,
2849                        sizeof(listener->loc_addr));
2850                 listener->loc_port = cm_info->loc_port;
2851
2852                 INIT_LIST_HEAD(&listener->child_listen_list);
2853
2854                 atomic_set(&listener->refcnt, 1);
2855         } else {
2856                 listener->reused_node = 1;
2857         }
2858
2859         listener->cm_id = cm_info->cm_id;
2860         listener->ipv4 = cm_info->ipv4;
2861         listener->vlan_id = cm_info->vlan_id;
2862         atomic_set(&listener->pend_accepts_cnt, 0);
2863         listener->cm_core = cm_core;
2864         listener->iwdev = iwdev;
2865
2866         listener->backlog = cm_info->backlog;
2867         listener->listener_state = IRDMA_CM_LISTENER_ACTIVE_STATE;
2868
2869         if (!listener->reused_node) {
2870                 spin_lock_irqsave(&cm_core->listen_list_lock, flags);
2871                 list_add(&listener->list, &cm_core->listen_list);
2872                 spin_unlock_irqrestore(&cm_core->listen_list_lock, flags);
2873         }
2874
2875         return listener;
2876 }
2877
2878 /**
2879  * irdma_create_cm_node - make a connection node with params
2880  * @cm_core: cm's core
2881  * @iwdev: iwarp device structure
2882  * @conn_param: connection parameters
2883  * @cm_info: quad info for connection
2884  * @caller_cm_node: pointer to cm_node structure to return
2885  */
2886 static int
2887 irdma_create_cm_node(struct irdma_cm_core *cm_core,
2888                      struct irdma_device *iwdev,
2889                      struct iw_cm_conn_param *conn_param,
2890                      struct irdma_cm_info *cm_info,
2891                      struct irdma_cm_node **caller_cm_node)
2892 {
2893         struct irdma_cm_node *cm_node;
2894         u16 private_data_len = conn_param->private_data_len;
2895         const void *private_data = conn_param->private_data;
2896
2897         /* create a CM connection node */
2898         cm_node = irdma_make_cm_node(cm_core, iwdev, cm_info, NULL);
2899         if (!cm_node)
2900                 return -ENOMEM;
2901
2902         /* set our node side to client (active) side */
2903         cm_node->tcp_cntxt.client = 1;
2904         cm_node->tcp_cntxt.rcv_wscale = IRDMA_CM_DEFAULT_RCV_WND_SCALE;
2905
2906         irdma_record_ird_ord(cm_node, conn_param->ird, conn_param->ord);
2907
2908         cm_node->pdata.size = private_data_len;
2909         cm_node->pdata.addr = cm_node->pdata_buf;
2910
2911         memcpy(cm_node->pdata_buf, private_data, private_data_len);
2912         *caller_cm_node = cm_node;
2913
2914         return 0;
2915 }
2916
2917 /**
2918  * irdma_cm_reject - reject and teardown a connection
2919  * @cm_node: connection's node
2920  * @pdata: ptr to private data for reject
2921  * @plen: size of private data
2922  */
2923 static int
2924 irdma_cm_reject(struct irdma_cm_node *cm_node, const void *pdata,
2925                 u8 plen)
2926 {
2927         int ret;
2928         int passive_state;
2929
2930         if (cm_node->tcp_cntxt.client)
2931                 return 0;
2932
2933         irdma_cleanup_retrans_entry(cm_node);
2934
2935         passive_state = atomic_add_return(1, &cm_node->passive_state);
2936         if (passive_state == IRDMA_SEND_RESET_EVENT) {
2937                 cm_node->state = IRDMA_CM_STATE_CLOSED;
2938                 irdma_rem_ref_cm_node(cm_node);
2939                 return 0;
2940         }
2941
2942         if (cm_node->state == IRDMA_CM_STATE_LISTENER_DESTROYED) {
2943                 irdma_rem_ref_cm_node(cm_node);
2944                 return 0;
2945         }
2946
2947         ret = irdma_send_mpa_reject(cm_node, pdata, plen);
2948         if (!ret)
2949                 return 0;
2950
2951         cm_node->state = IRDMA_CM_STATE_CLOSED;
2952         if (irdma_send_reset(cm_node))
2953                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
2954                             "send reset failed\n");
2955
2956         return ret;
2957 }
2958
2959 /**
2960  * irdma_cm_close - close of cm connection
2961  * @cm_node: connection's node
2962  */
2963 static int
2964 irdma_cm_close(struct irdma_cm_node *cm_node)
2965 {
2966         switch (cm_node->state) {
2967         case IRDMA_CM_STATE_SYN_RCVD:
2968         case IRDMA_CM_STATE_SYN_SENT:
2969         case IRDMA_CM_STATE_ONE_SIDE_ESTABLISHED:
2970         case IRDMA_CM_STATE_ESTABLISHED:
2971         case IRDMA_CM_STATE_ACCEPTING:
2972         case IRDMA_CM_STATE_MPAREQ_SENT:
2973         case IRDMA_CM_STATE_MPAREQ_RCVD:
2974                 irdma_cleanup_retrans_entry(cm_node);
2975                 irdma_send_reset(cm_node);
2976                 break;
2977         case IRDMA_CM_STATE_CLOSE_WAIT:
2978                 cm_node->state = IRDMA_CM_STATE_LAST_ACK;
2979                 irdma_send_fin(cm_node);
2980                 break;
2981         case IRDMA_CM_STATE_FIN_WAIT1:
2982         case IRDMA_CM_STATE_FIN_WAIT2:
2983         case IRDMA_CM_STATE_LAST_ACK:
2984         case IRDMA_CM_STATE_TIME_WAIT:
2985         case IRDMA_CM_STATE_CLOSING:
2986                 return -EINVAL;
2987         case IRDMA_CM_STATE_LISTENING:
2988                 irdma_cleanup_retrans_entry(cm_node);
2989                 irdma_send_reset(cm_node);
2990                 break;
2991         case IRDMA_CM_STATE_MPAREJ_RCVD:
2992         case IRDMA_CM_STATE_UNKNOWN:
2993         case IRDMA_CM_STATE_INITED:
2994         case IRDMA_CM_STATE_CLOSED:
2995         case IRDMA_CM_STATE_LISTENER_DESTROYED:
2996                 irdma_rem_ref_cm_node(cm_node);
2997                 break;
2998         case IRDMA_CM_STATE_OFFLOADED:
2999                 if (cm_node->send_entry)
3000                         irdma_debug(iwdev_to_idev(cm_node->iwdev),
3001                                     IRDMA_DEBUG_CM,
3002                                     "CM send_entry in OFFLOADED state\n");
3003                 irdma_rem_ref_cm_node(cm_node);
3004                 break;
3005         }
3006
3007         return 0;
3008 }
3009
3010 /**
3011  * irdma_receive_ilq - recv an ETHERNET packet, and process it
3012  * through CM
3013  * @vsi: VSI structure of dev
3014  * @rbuf: receive buffer
3015  */
3016 void
3017 irdma_receive_ilq(struct irdma_sc_vsi *vsi, struct irdma_puda_buf *rbuf)
3018 {
3019         struct irdma_cm_node *cm_node;
3020         struct irdma_cm_listener *listener;
3021         struct ip *iph;
3022         struct ip6_hdr *ip6h;
3023         struct tcphdr *tcph;
3024         struct irdma_cm_info cm_info = {0};
3025         struct irdma_device *iwdev = vsi->back_vsi;
3026         struct irdma_cm_core *cm_core = &iwdev->cm_core;
3027         struct ether_vlan_header *ethh;
3028         u16 vtag;
3029
3030         /* if vlan, then maclen = 18 else 14 */
3031         iph = (struct ip *)rbuf->iph;
3032         irdma_debug_buf(vsi->dev, IRDMA_DEBUG_ILQ, "RECEIVE ILQ BUFFER",
3033                         rbuf->mem.va, rbuf->totallen);
3034         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev >= IRDMA_GEN_2) {
3035                 if (rbuf->vlan_valid) {
3036                         vtag = rbuf->vlan_id;
3037                         cm_info.user_pri = (vtag & EVL_PRI_MASK) >>
3038                             VLAN_PRIO_SHIFT;
3039                         cm_info.vlan_id = vtag & EVL_VLID_MASK;
3040                 } else {
3041                         cm_info.vlan_id = 0xFFFF;
3042                 }
3043         } else {
3044                 ethh = rbuf->mem.va;
3045
3046                 if (ethh->evl_proto == htons(ETH_P_8021Q)) {
3047                         vtag = ntohs(ethh->evl_tag);
3048                         cm_info.user_pri = (vtag & EVL_PRI_MASK) >>
3049                             VLAN_PRIO_SHIFT;
3050                         cm_info.vlan_id = vtag & EVL_VLID_MASK;
3051                         irdma_debug(iwdev_to_idev(cm_core->iwdev),
3052                                     IRDMA_DEBUG_CM, "vlan_id=%d\n",
3053                                     cm_info.vlan_id);
3054                 } else {
3055                         cm_info.vlan_id = 0xFFFF;
3056                 }
3057         }
3058         tcph = (struct tcphdr *)rbuf->tcph;
3059
3060         if (rbuf->ipv4) {
3061                 cm_info.loc_addr[0] = ntohl(iph->ip_dst.s_addr);
3062                 cm_info.rem_addr[0] = ntohl(iph->ip_src.s_addr);
3063                 cm_info.ipv4 = true;
3064                 cm_info.tos = iph->ip_tos;
3065         } else {
3066                 ip6h = (struct ip6_hdr *)rbuf->iph;
3067                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3068                                     ip6h->ip6_dst.__u6_addr.__u6_addr32);
3069                 irdma_copy_ip_ntohl(cm_info.rem_addr,
3070                                     ip6h->ip6_src.__u6_addr.__u6_addr32);
3071                 cm_info.ipv4 = false;
3072                 cm_info.tos = (ip6h->ip6_vfc << 4) | ip6h->ip6_flow;
3073         }
3074         cm_info.loc_port = ntohs(tcph->th_dport);
3075         cm_info.rem_port = ntohs(tcph->th_sport);
3076         cm_node = irdma_find_node(cm_core, cm_info.rem_port, cm_info.rem_addr,
3077                                   cm_info.loc_port, cm_info.loc_addr, cm_info.vlan_id);
3078
3079         if (!cm_node) {
3080                 /*
3081                  * Only type of packet accepted are for the PASSIVE open (syn only)
3082                  */
3083                 if (!(tcph->th_flags & TH_SYN) || tcph->th_flags & TH_ACK)
3084                         return;
3085
3086                 listener = irdma_find_listener(cm_core,
3087                                                cm_info.loc_addr,
3088                                                cm_info.loc_port,
3089                                                cm_info.vlan_id,
3090                                                IRDMA_CM_LISTENER_ACTIVE_STATE);
3091                 if (!listener) {
3092                         cm_info.cm_id = NULL;
3093                         irdma_debug(iwdev_to_idev(cm_core->iwdev),
3094                                     IRDMA_DEBUG_CM, "no listener found\n");
3095                         return;
3096                 }
3097
3098                 cm_info.cm_id = listener->cm_id;
3099                 cm_node = irdma_make_cm_node(cm_core, iwdev, &cm_info,
3100                                              listener);
3101                 if (!cm_node) {
3102                         irdma_debug(iwdev_to_idev(cm_core->iwdev),
3103                                     IRDMA_DEBUG_CM, "allocate node failed\n");
3104                         atomic_dec(&listener->refcnt);
3105                         return;
3106                 }
3107
3108                 if (!(tcph->th_flags & (TH_RST | TH_FIN))) {
3109                         cm_node->state = IRDMA_CM_STATE_LISTENING;
3110                 } else {
3111                         irdma_rem_ref_cm_node(cm_node);
3112                         return;
3113                 }
3114
3115                 atomic_inc(&cm_node->refcnt);
3116         } else if (cm_node->state == IRDMA_CM_STATE_OFFLOADED) {
3117                 irdma_rem_ref_cm_node(cm_node);
3118                 return;
3119         }
3120
3121         irdma_process_pkt(cm_node, rbuf);
3122         irdma_rem_ref_cm_node(cm_node);
3123 }
3124
3125 static int
3126 irdma_add_qh(struct irdma_cm_node *cm_node, bool active)
3127 {
3128         if (!active)
3129                 irdma_add_conn_est_qh(cm_node);
3130         return 0;
3131 }
3132
3133 static void
3134 irdma_cm_free_ah_nop(struct irdma_cm_node *cm_node)
3135 {
3136 }
3137
3138 /**
3139  * irdma_setup_cm_core - setup top level instance of a cm core
3140  * @iwdev: iwarp device structure
3141  * @rdma_ver: HW version
3142  */
3143 int
3144 irdma_setup_cm_core(struct irdma_device *iwdev, u8 rdma_ver)
3145 {
3146         struct irdma_cm_core *cm_core = &iwdev->cm_core;
3147
3148         cm_core->iwdev = iwdev;
3149         cm_core->dev = &iwdev->rf->sc_dev;
3150
3151         /* Handles CM event work items send to Iwarp core */
3152         cm_core->event_wq = alloc_ordered_workqueue("iwarp-event-wq", 0);
3153         if (!cm_core->event_wq)
3154                 return -ENOMEM;
3155
3156         INIT_LIST_HEAD(&cm_core->listen_list);
3157
3158         timer_setup(&cm_core->tcp_timer, irdma_cm_timer_tick, 0);
3159
3160         spin_lock_init(&cm_core->ht_lock);
3161         spin_lock_init(&cm_core->listen_list_lock);
3162         spin_lock_init(&cm_core->apbvt_lock);
3163         switch (rdma_ver) {
3164         case IRDMA_GEN_1:
3165                 cm_core->form_cm_frame = irdma_form_uda_cm_frame;
3166                 cm_core->cm_create_ah = irdma_add_qh;
3167                 cm_core->cm_free_ah = irdma_cm_free_ah_nop;
3168                 break;
3169         case IRDMA_GEN_2:
3170         default:
3171                 cm_core->form_cm_frame = irdma_form_ah_cm_frame;
3172                 cm_core->cm_create_ah = irdma_cm_create_ah;
3173                 cm_core->cm_free_ah = irdma_cm_free_ah;
3174         }
3175
3176         return 0;
3177 }
3178
3179 /**
3180  * irdma_cleanup_cm_core - deallocate a top level instance of a
3181  * cm core
3182  * @cm_core: cm's core
3183  */
3184 void
3185 irdma_cleanup_cm_core(struct irdma_cm_core *cm_core)
3186 {
3187         unsigned long flags;
3188
3189         if (!cm_core)
3190                 return;
3191
3192         spin_lock_irqsave(&cm_core->ht_lock, flags);
3193         if (timer_pending(&cm_core->tcp_timer))
3194                 del_timer_sync(&cm_core->tcp_timer);
3195         spin_unlock_irqrestore(&cm_core->ht_lock, flags);
3196
3197         destroy_workqueue(cm_core->event_wq);
3198         cm_core->dev->ws_reset(&cm_core->iwdev->vsi);
3199 }
3200
3201 /**
3202  * irdma_init_tcp_ctx - setup qp context
3203  * @cm_node: connection's node
3204  * @tcp_info: offload info for tcp
3205  * @iwqp: associate qp for the connection
3206  */
3207 static void
3208 irdma_init_tcp_ctx(struct irdma_cm_node *cm_node,
3209                    struct irdma_tcp_offload_info *tcp_info,
3210                    struct irdma_qp *iwqp)
3211 {
3212         tcp_info->ipv4 = cm_node->ipv4;
3213         tcp_info->drop_ooo_seg = !iwqp->iwdev->iw_ooo;
3214         tcp_info->wscale = true;
3215         tcp_info->ignore_tcp_opt = true;
3216         tcp_info->ignore_tcp_uns_opt = true;
3217         tcp_info->no_nagle = false;
3218
3219         tcp_info->ttl = IRDMA_DEFAULT_TTL;
3220         tcp_info->rtt_var = IRDMA_DEFAULT_RTT_VAR;
3221         tcp_info->ss_thresh = IRDMA_DEFAULT_SS_THRESH;
3222         tcp_info->rexmit_thresh = IRDMA_DEFAULT_REXMIT_THRESH;
3223
3224         tcp_info->tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
3225         tcp_info->snd_wscale = cm_node->tcp_cntxt.snd_wscale;
3226         tcp_info->rcv_wscale = cm_node->tcp_cntxt.rcv_wscale;
3227
3228         tcp_info->snd_nxt = cm_node->tcp_cntxt.loc_seq_num;
3229         tcp_info->snd_wnd = cm_node->tcp_cntxt.snd_wnd;
3230         tcp_info->rcv_nxt = cm_node->tcp_cntxt.rcv_nxt;
3231         tcp_info->snd_max = cm_node->tcp_cntxt.loc_seq_num;
3232
3233         tcp_info->snd_una = cm_node->tcp_cntxt.loc_seq_num;
3234         tcp_info->cwnd = 2 * cm_node->tcp_cntxt.mss;
3235         tcp_info->snd_wl1 = cm_node->tcp_cntxt.rcv_nxt;
3236         tcp_info->snd_wl2 = cm_node->tcp_cntxt.loc_seq_num;
3237         tcp_info->max_snd_window = cm_node->tcp_cntxt.max_snd_wnd;
3238         tcp_info->rcv_wnd = cm_node->tcp_cntxt.rcv_wnd
3239             << cm_node->tcp_cntxt.rcv_wscale;
3240
3241         tcp_info->flow_label = 0;
3242         tcp_info->snd_mss = (u32)cm_node->tcp_cntxt.mss;
3243         tcp_info->tos = cm_node->tos;
3244         if (cm_node->vlan_id < VLAN_N_VID) {
3245                 tcp_info->insert_vlan_tag = true;
3246                 tcp_info->vlan_tag = cm_node->vlan_id;
3247                 tcp_info->vlan_tag |= cm_node->user_pri << VLAN_PRIO_SHIFT;
3248         }
3249         tcp_info->src_port = cm_node->loc_port;
3250         tcp_info->dst_port = cm_node->rem_port;
3251         tcp_info->arp_idx = (u16)irdma_arp_table(iwqp->iwdev->rf,
3252                                                   cm_node->rem_addr, NULL,
3253                                                   IRDMA_ARP_RESOLVE);
3254         if (cm_node->ipv4) {
3255                 tcp_info->dest_ip_addr[3] = cm_node->rem_addr[0];
3256                 tcp_info->local_ipaddr[3] = cm_node->loc_addr[0];
3257         } else {
3258                 memcpy(tcp_info->dest_ip_addr, cm_node->rem_addr,
3259                        sizeof(tcp_info->dest_ip_addr));
3260                 memcpy(tcp_info->local_ipaddr, cm_node->loc_addr,
3261                        sizeof(tcp_info->local_ipaddr));
3262         }
3263 }
3264
3265 /**
3266  * irdma_cm_init_tsa_conn - setup qp for RTS
3267  * @iwqp: associate qp for the connection
3268  * @cm_node: connection's node
3269  */
3270 static void
3271 irdma_cm_init_tsa_conn(struct irdma_qp *iwqp,
3272                        struct irdma_cm_node *cm_node)
3273 {
3274         struct irdma_iwarp_offload_info *iwarp_info;
3275         struct irdma_qp_host_ctx_info *ctx_info;
3276
3277         iwarp_info = &iwqp->iwarp_info;
3278         ctx_info = &iwqp->ctx_info;
3279
3280         ctx_info->tcp_info = &iwqp->tcp_info;
3281         ctx_info->send_cq_num = iwqp->iwscq->sc_cq.cq_uk.cq_id;
3282         ctx_info->rcv_cq_num = iwqp->iwrcq->sc_cq.cq_uk.cq_id;
3283
3284         iwarp_info->ord_size = cm_node->ord_size;
3285         iwarp_info->ird_size = cm_node->ird_size;
3286         iwarp_info->rd_en = true;
3287         iwarp_info->rdmap_ver = 1;
3288         iwarp_info->ddp_ver = 1;
3289         iwarp_info->pd_id = iwqp->iwpd->sc_pd.pd_id;
3290
3291         ctx_info->tcp_info_valid = true;
3292         ctx_info->iwarp_info_valid = true;
3293         ctx_info->user_pri = cm_node->user_pri;
3294
3295         irdma_init_tcp_ctx(cm_node, &iwqp->tcp_info, iwqp);
3296         if (cm_node->snd_mark_en) {
3297                 iwarp_info->snd_mark_en = true;
3298                 iwarp_info->snd_mark_offset = (iwqp->tcp_info.snd_nxt & SNDMARKER_SEQNMASK) +
3299                     cm_node->lsmm_size;
3300         }
3301
3302         cm_node->state = IRDMA_CM_STATE_OFFLOADED;
3303         iwqp->tcp_info.tcp_state = IRDMA_TCP_STATE_ESTABLISHED;
3304         iwqp->tcp_info.src_mac_addr_idx = iwqp->iwdev->mac_ip_table_idx;
3305
3306         if (cm_node->rcv_mark_en) {
3307                 iwarp_info->rcv_mark_en = true;
3308                 iwarp_info->align_hdrs = true;
3309         }
3310
3311         irdma_sc_qp_setctx(&iwqp->sc_qp, iwqp->host_ctx.va, ctx_info);
3312
3313         /* once tcp_info is set, no need to do it again */
3314         ctx_info->tcp_info_valid = false;
3315         ctx_info->iwarp_info_valid = false;
3316 }
3317
3318 /**
3319  * irdma_cm_disconn - when a connection is being closed
3320  * @iwqp: associated qp for the connection
3321  */
3322 void
3323 irdma_cm_disconn(struct irdma_qp *iwqp)
3324 {
3325         struct irdma_device *iwdev = iwqp->iwdev;
3326         struct disconn_work *work;
3327         unsigned long flags;
3328
3329         work = kzalloc(sizeof(*work), GFP_ATOMIC);
3330         if (!work)
3331                 return;
3332
3333         spin_lock_irqsave(&iwdev->rf->qptable_lock, flags);
3334         if (!iwdev->rf->qp_table[iwqp->ibqp.qp_num]) {
3335                 spin_unlock_irqrestore(&iwdev->rf->qptable_lock, flags);
3336                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3337                             "qp_id %d is already freed\n", iwqp->ibqp.qp_num);
3338                 kfree(work);
3339                 return;
3340         }
3341         irdma_qp_add_ref(&iwqp->ibqp);
3342         spin_unlock_irqrestore(&iwdev->rf->qptable_lock, flags);
3343
3344         work->iwqp = iwqp;
3345         INIT_WORK(&work->work, irdma_disconnect_worker);
3346         queue_work(iwdev->cleanup_wq, &work->work);
3347 }
3348
3349 /**
3350  * irdma_qp_disconnect - free qp and close cm
3351  * @iwqp: associate qp for the connection
3352  */
3353 static void
3354 irdma_qp_disconnect(struct irdma_qp *iwqp)
3355 {
3356         struct irdma_device *iwdev = iwqp->iwdev;
3357
3358         iwqp->active_conn = 0;
3359         /* close the CM node down if it is still active */
3360         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3361                     "Call close API\n");
3362         irdma_cm_close(iwqp->cm_node);
3363 }
3364
3365 /**
3366  * irdma_cm_disconn_true - called by worker thread to disconnect qp
3367  * @iwqp: associate qp for the connection
3368  */
3369 static void
3370 irdma_cm_disconn_true(struct irdma_qp *iwqp)
3371 {
3372         struct iw_cm_id *cm_id;
3373         struct irdma_device *iwdev;
3374         struct irdma_sc_qp *qp = &iwqp->sc_qp;
3375         u16 last_ae;
3376         u8 original_hw_tcp_state;
3377         u8 original_ibqp_state;
3378         int disconn_status = 0;
3379         int issue_disconn = 0;
3380         int issue_close = 0;
3381         int issue_flush = 0;
3382         unsigned long flags;
3383         int err;
3384
3385         iwdev = iwqp->iwdev;
3386         spin_lock_irqsave(&iwqp->lock, flags);
3387         if (rdma_protocol_roce(&iwdev->ibdev, 1)) {
3388                 struct ib_qp_attr attr;
3389
3390                 if (iwqp->flush_issued || iwqp->sc_qp.qp_uk.destroy_pending) {
3391                         spin_unlock_irqrestore(&iwqp->lock, flags);
3392                         return;
3393                 }
3394
3395                 spin_unlock_irqrestore(&iwqp->lock, flags);
3396
3397                 attr.qp_state = IB_QPS_ERR;
3398                 irdma_modify_qp_roce(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
3399                 irdma_ib_qp_event(iwqp, qp->event_type);
3400                 return;
3401         }
3402
3403         cm_id = iwqp->cm_id;
3404         /* make sure we havent already closed this connection */
3405         if (!cm_id) {
3406                 spin_unlock_irqrestore(&iwqp->lock, flags);
3407                 return;
3408         }
3409
3410         original_hw_tcp_state = iwqp->hw_tcp_state;
3411         original_ibqp_state = iwqp->ibqp_state;
3412         last_ae = iwqp->last_aeq;
3413
3414         if (qp->term_flags) {
3415                 issue_disconn = 1;
3416                 issue_close = 1;
3417                 iwqp->cm_id = NULL;
3418                 irdma_terminate_del_timer(qp);
3419                 if (!iwqp->flush_issued) {
3420                         iwqp->flush_issued = 1;
3421                         issue_flush = 1;
3422                 }
3423         } else if ((original_hw_tcp_state == IRDMA_TCP_STATE_CLOSE_WAIT) ||
3424                    ((original_ibqp_state == IB_QPS_RTS) &&
3425                     (last_ae == IRDMA_AE_LLP_CONNECTION_RESET))) {
3426                 issue_disconn = 1;
3427                 if (last_ae == IRDMA_AE_LLP_CONNECTION_RESET)
3428                         disconn_status = -ECONNRESET;
3429         }
3430
3431         if ((original_hw_tcp_state == IRDMA_TCP_STATE_CLOSED ||
3432              original_hw_tcp_state == IRDMA_TCP_STATE_TIME_WAIT ||
3433              last_ae == IRDMA_AE_RDMAP_ROE_BAD_LLP_CLOSE ||
3434              last_ae == IRDMA_AE_BAD_CLOSE ||
3435              last_ae == IRDMA_AE_LLP_CONNECTION_RESET || iwdev->rf->reset)) {
3436                 issue_close = 1;
3437                 iwqp->cm_id = NULL;
3438                 qp->term_flags = 0;
3439                 if (!iwqp->flush_issued) {
3440                         iwqp->flush_issued = 1;
3441                         issue_flush = 1;
3442                 }
3443         }
3444
3445         spin_unlock_irqrestore(&iwqp->lock, flags);
3446         if (issue_flush && !iwqp->sc_qp.qp_uk.destroy_pending) {
3447                 if (!iwqp->user_mode)
3448                         queue_delayed_work(iwqp->iwdev->cleanup_wq,
3449                                            &iwqp->dwork_flush,
3450                                            msecs_to_jiffies(IRDMA_FLUSH_DELAY_MS));
3451                 irdma_flush_wqes(iwqp, IRDMA_FLUSH_SQ | IRDMA_FLUSH_RQ |
3452                                  IRDMA_FLUSH_WAIT);
3453
3454                 if (qp->term_flags)
3455                         irdma_ib_qp_event(iwqp, qp->event_type);
3456         }
3457
3458         if (!cm_id || !cm_id->event_handler)
3459                 return;
3460
3461         spin_lock_irqsave(&iwdev->cm_core.ht_lock, flags);
3462         if (!iwqp->cm_node) {
3463                 spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
3464                 return;
3465         }
3466         atomic_inc(&iwqp->cm_node->refcnt);
3467
3468         spin_unlock_irqrestore(&iwdev->cm_core.ht_lock, flags);
3469
3470         if (issue_disconn) {
3471                 err = irdma_send_cm_event(iwqp->cm_node, cm_id,
3472                                           IW_CM_EVENT_DISCONNECT,
3473                                           disconn_status);
3474                 if (err)
3475                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3476                                     "disconnect event failed: - cm_id = %p\n",
3477                                     cm_id);
3478         }
3479         if (issue_close) {
3480                 cm_id->provider_data = iwqp;
3481                 err = irdma_send_cm_event(iwqp->cm_node, cm_id,
3482                                           IW_CM_EVENT_CLOSE, 0);
3483                 if (err)
3484                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3485                                     "close event failed: - cm_id = %p\n",
3486                                     cm_id);
3487                 irdma_qp_disconnect(iwqp);
3488         }
3489         irdma_rem_ref_cm_node(iwqp->cm_node);
3490 }
3491
3492 /**
3493  * irdma_disconnect_worker - worker for connection close
3494  * @work: points or disconn structure
3495  */
3496 static void
3497 irdma_disconnect_worker(struct work_struct *work)
3498 {
3499         struct disconn_work *dwork = container_of(work, struct disconn_work, work);
3500         struct irdma_qp *iwqp = dwork->iwqp;
3501
3502         kfree(dwork);
3503         irdma_cm_disconn_true(iwqp);
3504         irdma_qp_rem_ref(&iwqp->ibqp);
3505 }
3506
3507 /**
3508  * irdma_free_lsmm_rsrc - free lsmm memory and deregister
3509  * @iwqp: associate qp for the connection
3510  */
3511 void
3512 irdma_free_lsmm_rsrc(struct irdma_qp *iwqp)
3513 {
3514         struct irdma_device *iwdev;
3515
3516         iwdev = iwqp->iwdev;
3517
3518         if (iwqp->ietf_mem.va) {
3519                 if (iwqp->lsmm_mr)
3520                         kc_free_lsmm_dereg_mr(iwdev, iwqp);
3521                 irdma_free_dma_mem(iwdev->rf->sc_dev.hw,
3522                                    &iwqp->ietf_mem);
3523                 iwqp->ietf_mem.va = NULL;
3524         }
3525 }
3526
3527 /**
3528  * irdma_accept - registered call for connection to be accepted
3529  * @cm_id: cm information for passive connection
3530  * @conn_param: accpet parameters
3531  */
3532 int
3533 irdma_accept(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
3534 {
3535         struct ib_qp *ibqp;
3536         struct irdma_qp *iwqp;
3537         struct irdma_device *iwdev;
3538         struct irdma_sc_dev *dev;
3539         struct irdma_cm_node *cm_node;
3540         struct ib_qp_attr attr = {0};
3541         int passive_state;
3542         struct ib_mr *ibmr;
3543         struct irdma_pd *iwpd;
3544         u16 buf_len = 0;
3545         struct irdma_kmem_info accept;
3546         u64 tagged_offset;
3547         int wait_ret;
3548         int ret = 0;
3549
3550         ibqp = irdma_get_qp(cm_id->device, conn_param->qpn);
3551         if (!ibqp)
3552                 return -EINVAL;
3553
3554         iwqp = to_iwqp(ibqp);
3555         iwdev = iwqp->iwdev;
3556         dev = &iwdev->rf->sc_dev;
3557         cm_node = cm_id->provider_data;
3558
3559         if (((struct sockaddr_in *)&cm_id->local_addr)->sin_family == AF_INET) {
3560                 cm_node->ipv4 = true;
3561                 cm_node->vlan_id = irdma_get_vlan_ipv4(cm_node->loc_addr);
3562         } else {
3563                 cm_node->ipv4 = false;
3564                 irdma_netdev_vlan_ipv6(cm_node->loc_addr, &cm_node->vlan_id,
3565                                        NULL);
3566         }
3567         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3568                     "Accept vlan_id=%d\n", cm_node->vlan_id);
3569
3570         if (cm_node->state == IRDMA_CM_STATE_LISTENER_DESTROYED) {
3571                 ret = -EINVAL;
3572                 goto error;
3573         }
3574
3575         passive_state = atomic_add_return(1, &cm_node->passive_state);
3576         if (passive_state == IRDMA_SEND_RESET_EVENT) {
3577                 ret = -ECONNRESET;
3578                 goto error;
3579         }
3580
3581         buf_len = conn_param->private_data_len + IRDMA_MAX_IETF_SIZE;
3582         iwqp->ietf_mem.size = buf_len;
3583         iwqp->ietf_mem.va = irdma_allocate_dma_mem(dev->hw, &iwqp->ietf_mem,
3584                                                    iwqp->ietf_mem.size, 1);
3585         if (!iwqp->ietf_mem.va) {
3586                 ret = -ENOMEM;
3587                 goto error;
3588         }
3589
3590         cm_node->pdata.size = conn_param->private_data_len;
3591         accept.addr = iwqp->ietf_mem.va;
3592         accept.size = irdma_cm_build_mpa_frame(cm_node, &accept, MPA_KEY_REPLY);
3593         memcpy((u8 *)accept.addr + accept.size, conn_param->private_data,
3594                conn_param->private_data_len);
3595
3596         if (cm_node->dev->ws_add(iwqp->sc_qp.vsi, cm_node->user_pri)) {
3597                 ret = -ENOMEM;
3598                 goto error;
3599         }
3600         iwqp->sc_qp.user_pri = cm_node->user_pri;
3601         irdma_qp_add_qos(&iwqp->sc_qp);
3602         if (cm_node->dev->hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2)
3603                 iwdev->rf->check_fc(&iwdev->vsi, &iwqp->sc_qp);
3604         /* setup our first outgoing iWarp send WQE (the IETF frame response) */
3605         iwpd = iwqp->iwpd;
3606         tagged_offset = (uintptr_t)iwqp->ietf_mem.va;
3607         ibmr = irdma_reg_phys_mr(&iwpd->ibpd, iwqp->ietf_mem.pa, buf_len,
3608                                  IB_ACCESS_LOCAL_WRITE, &tagged_offset);
3609         if (IS_ERR(ibmr)) {
3610                 ret = -ENOMEM;
3611                 goto error;
3612         }
3613
3614         ibmr->pd = &iwpd->ibpd;
3615         ibmr->device = iwpd->ibpd.device;
3616         iwqp->lsmm_mr = ibmr;
3617         if (iwqp->page)
3618                 iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
3619
3620         cm_node->lsmm_size = accept.size + conn_param->private_data_len;
3621         irdma_sc_send_lsmm(&iwqp->sc_qp, iwqp->ietf_mem.va, cm_node->lsmm_size,
3622                            ibmr->lkey);
3623
3624         if (iwqp->page)
3625                 kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
3626
3627         iwqp->cm_id = cm_id;
3628         cm_node->cm_id = cm_id;
3629
3630         cm_id->provider_data = iwqp;
3631         iwqp->active_conn = 0;
3632         iwqp->cm_node = cm_node;
3633         cm_node->iwqp = iwqp;
3634         irdma_cm_init_tsa_conn(iwqp, cm_node);
3635         irdma_qp_add_ref(&iwqp->ibqp);
3636         cm_id->add_ref(cm_id);
3637
3638         attr.qp_state = IB_QPS_RTS;
3639         cm_node->qhash_set = false;
3640         cm_node->cm_core->cm_free_ah(cm_node);
3641
3642         irdma_modify_qp(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
3643         if (dev->hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_RTS_AE) {
3644                 wait_ret = wait_event_interruptible_timeout(iwqp->waitq,
3645                                                             iwqp->rts_ae_rcvd,
3646                                                             IRDMA_MAX_TIMEOUT);
3647                 if (!wait_ret) {
3648                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3649                                     "Slow Connection: cm_node=%p, loc_port=%d, rem_port=%d, cm_id=%p\n",
3650                                     cm_node, cm_node->loc_port,
3651                                     cm_node->rem_port, cm_node->cm_id);
3652                         ret = -ECONNRESET;
3653                         goto error;
3654                 }
3655         }
3656
3657         irdma_send_cm_event(cm_node, cm_id, IW_CM_EVENT_ESTABLISHED, 0);
3658         cm_node->accelerated = true;
3659         complete(&cm_node->establish_comp);
3660
3661         if (cm_node->accept_pend) {
3662                 atomic_dec(&cm_node->listener->pend_accepts_cnt);
3663                 cm_node->accept_pend = 0;
3664         }
3665
3666         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3667                     "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",
3668                     cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr,
3669                     cm_node->loc_addr, cm_node, cm_id, ibqp->qp_num);
3670         cm_node->cm_core->stats_accepts++;
3671
3672         return 0;
3673 error:
3674         irdma_free_lsmm_rsrc(iwqp);
3675         irdma_rem_ref_cm_node(cm_node);
3676
3677         return ret;
3678 }
3679
3680 /**
3681  * irdma_reject - registered call for connection to be rejected
3682  * @cm_id: cm information for passive connection
3683  * @pdata: private data to be sent
3684  * @pdata_len: private data length
3685  */
3686 int
3687 irdma_reject(struct iw_cm_id *cm_id, const void *pdata, u8 pdata_len)
3688 {
3689         struct irdma_device *iwdev;
3690         struct irdma_cm_node *cm_node;
3691
3692         cm_node = cm_id->provider_data;
3693         cm_node->pdata.size = pdata_len;
3694
3695         iwdev = to_iwdev(cm_id->device);
3696         if (!iwdev)
3697                 return -EINVAL;
3698
3699         cm_node->cm_core->stats_rejects++;
3700
3701         if (pdata_len + sizeof(struct ietf_mpa_v2) > IRDMA_MAX_CM_BUF)
3702                 return -EINVAL;
3703
3704         return irdma_cm_reject(cm_node, pdata, pdata_len);
3705 }
3706
3707 /**
3708  * irdma_connect - registered call for connection to be established
3709  * @cm_id: cm information for passive connection
3710  * @conn_param: Information about the connection
3711  */
3712 int
3713 irdma_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *conn_param)
3714 {
3715         struct ib_qp *ibqp;
3716         struct irdma_qp *iwqp;
3717         struct irdma_device *iwdev;
3718         struct irdma_cm_node *cm_node;
3719         struct irdma_cm_info cm_info;
3720         struct sockaddr_in *laddr;
3721         struct sockaddr_in *raddr;
3722         struct sockaddr_in6 *laddr6;
3723         struct sockaddr_in6 *raddr6;
3724         int ret = 0;
3725
3726         ibqp = irdma_get_qp(cm_id->device, conn_param->qpn);
3727         if (!ibqp)
3728                 return -EINVAL;
3729         iwqp = to_iwqp(ibqp);
3730         if (!iwqp)
3731                 return -EINVAL;
3732         iwdev = iwqp->iwdev;
3733         if (!iwdev)
3734                 return -EINVAL;
3735
3736         laddr = (struct sockaddr_in *)&cm_id->m_local_addr;
3737         raddr = (struct sockaddr_in *)&cm_id->m_remote_addr;
3738         laddr6 = (struct sockaddr_in6 *)&cm_id->m_local_addr;
3739         raddr6 = (struct sockaddr_in6 *)&cm_id->m_remote_addr;
3740
3741         if (!(laddr->sin_port) || !(raddr->sin_port))
3742                 return -EINVAL;
3743
3744         iwqp->active_conn = 1;
3745         iwqp->cm_id = NULL;
3746         cm_id->provider_data = iwqp;
3747
3748         /* set up the connection params for the node */
3749         if (cm_id->remote_addr.ss_family == AF_INET) {
3750                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV4)
3751                         return -EINVAL;
3752
3753                 cm_info.ipv4 = true;
3754                 memset(cm_info.loc_addr, 0, sizeof(cm_info.loc_addr));
3755                 memset(cm_info.rem_addr, 0, sizeof(cm_info.rem_addr));
3756                 cm_info.loc_addr[0] = ntohl(laddr->sin_addr.s_addr);
3757                 cm_info.rem_addr[0] = ntohl(raddr->sin_addr.s_addr);
3758                 cm_info.loc_port = ntohs(laddr->sin_port);
3759                 cm_info.rem_port = ntohs(raddr->sin_port);
3760                 cm_info.vlan_id = irdma_get_vlan_ipv4(cm_info.loc_addr);
3761         } else {
3762                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV6)
3763                         return -EINVAL;
3764
3765                 cm_info.ipv4 = false;
3766                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3767                                     laddr6->sin6_addr.__u6_addr.__u6_addr32);
3768                 irdma_copy_ip_ntohl(cm_info.rem_addr,
3769                                     raddr6->sin6_addr.__u6_addr.__u6_addr32);
3770                 cm_info.loc_port = ntohs(laddr6->sin6_port);
3771                 cm_info.rem_port = ntohs(raddr6->sin6_port);
3772                 irdma_netdev_vlan_ipv6(cm_info.loc_addr, &cm_info.vlan_id,
3773                                        NULL);
3774         }
3775         cm_info.cm_id = cm_id;
3776         cm_info.qh_qpid = iwdev->vsi.ilq->qp_id;
3777         cm_info.tos = cm_id->tos;
3778         if (iwdev->vsi.dscp_mode)
3779                 cm_info.user_pri =
3780                     iwqp->sc_qp.vsi->dscp_map[irdma_tos2dscp(cm_info.tos)];
3781         else
3782                 cm_info.user_pri = rt_tos2priority(cm_id->tos);
3783
3784         if (iwqp->sc_qp.dev->ws_add(iwqp->sc_qp.vsi, cm_info.user_pri))
3785                 return -ENOMEM;
3786         iwqp->sc_qp.user_pri = cm_info.user_pri;
3787         irdma_qp_add_qos(&iwqp->sc_qp);
3788         if (iwdev->rf->sc_dev.hw_attrs.uk_attrs.hw_rev == IRDMA_GEN_2)
3789                 iwdev->rf->check_fc(&iwdev->vsi, &iwqp->sc_qp);
3790         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_DCB,
3791                     "TOS:[%d] UP:[%d]\n", cm_id->tos, cm_info.user_pri);
3792
3793         ret = irdma_create_cm_node(&iwdev->cm_core, iwdev, conn_param, &cm_info,
3794                                    &cm_node);
3795         if (ret)
3796                 return ret;
3797         ret = cm_node->cm_core->cm_create_ah(cm_node, true);
3798         if (ret)
3799                 goto err;
3800         if (irdma_manage_qhash(iwdev, &cm_info,
3801                                IRDMA_QHASH_TYPE_TCP_ESTABLISHED,
3802                                IRDMA_QHASH_MANAGE_TYPE_ADD, NULL, true)) {
3803                 ret = -EINVAL;
3804                 goto err;
3805         }
3806         cm_node->qhash_set = true;
3807
3808         cm_node->apbvt_entry = irdma_add_apbvt(iwdev, cm_info.loc_port);
3809         if (!cm_node->apbvt_entry) {
3810                 ret = -EINVAL;
3811                 goto err;
3812         }
3813
3814         cm_node->apbvt_set = true;
3815         iwqp->cm_node = cm_node;
3816         cm_node->iwqp = iwqp;
3817         iwqp->cm_id = cm_id;
3818         irdma_qp_add_ref(&iwqp->ibqp);
3819         cm_id->add_ref(cm_id);
3820
3821         if (cm_node->state != IRDMA_CM_STATE_OFFLOADED) {
3822                 cm_node->state = IRDMA_CM_STATE_SYN_SENT;
3823                 ret = irdma_send_syn(cm_node, 0);
3824                 if (ret)
3825                         goto err;
3826         }
3827
3828         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3829                     "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",
3830                     cm_node->rem_port, cm_node->loc_port, cm_node->rem_addr,
3831                     cm_node->loc_addr, cm_node, cm_id, ibqp->qp_num);
3832
3833         return 0;
3834
3835 err:
3836         if (cm_info.ipv4)
3837                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3838                             "connect() FAILED: dest addr=%pI4",
3839                             cm_info.rem_addr);
3840         else
3841                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3842                             "connect() FAILED: dest addr=%pI6",
3843                             cm_info.rem_addr);
3844         irdma_rem_ref_cm_node(cm_node);
3845         iwdev->cm_core.stats_connect_errs++;
3846
3847         return ret;
3848 }
3849
3850 /**
3851  * irdma_create_listen - registered call creating listener
3852  * @cm_id: cm information for passive connection
3853  * @backlog: to max accept pending count
3854  */
3855 int
3856 irdma_create_listen(struct iw_cm_id *cm_id, int backlog)
3857 {
3858         struct irdma_device *iwdev;
3859         struct irdma_cm_listener *cm_listen_node;
3860         struct irdma_cm_info cm_info = {0};
3861         struct sockaddr_in *laddr;
3862         struct sockaddr_in6 *laddr6;
3863         bool wildcard = false;
3864         int err;
3865
3866         iwdev = to_iwdev(cm_id->device);
3867         if (!iwdev)
3868                 return -EINVAL;
3869
3870         laddr = (struct sockaddr_in *)&cm_id->m_local_addr;
3871         laddr6 = (struct sockaddr_in6 *)&cm_id->m_local_addr;
3872         cm_info.qh_qpid = iwdev->vsi.ilq->qp_id;
3873
3874         if (laddr->sin_family == AF_INET) {
3875                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV4)
3876                         return -EINVAL;
3877
3878                 cm_info.ipv4 = true;
3879                 cm_info.loc_addr[0] = ntohl(laddr->sin_addr.s_addr);
3880                 cm_info.loc_port = ntohs(laddr->sin_port);
3881
3882                 if (laddr->sin_addr.s_addr != htonl(INADDR_ANY)) {
3883                         cm_info.vlan_id = irdma_get_vlan_ipv4(cm_info.loc_addr);
3884                 } else {
3885                         cm_info.vlan_id = 0xFFFF;
3886                         wildcard = true;
3887                 }
3888         } else {
3889                 if (iwdev->vsi.mtu < IRDMA_MIN_MTU_IPV6)
3890                         return -EINVAL;
3891
3892                 cm_info.ipv4 = false;
3893                 irdma_copy_ip_ntohl(cm_info.loc_addr,
3894                                     laddr6->sin6_addr.__u6_addr.__u6_addr32);
3895                 cm_info.loc_port = ntohs(laddr6->sin6_port);
3896                 if (!IN6_IS_ADDR_UNSPECIFIED(&laddr6->sin6_addr)) {
3897                         irdma_netdev_vlan_ipv6(cm_info.loc_addr,
3898                                                &cm_info.vlan_id, NULL);
3899                 } else {
3900                         cm_info.vlan_id = 0xFFFF;
3901                         wildcard = true;
3902                 }
3903         }
3904
3905         if (cm_info.vlan_id >= VLAN_N_VID && iwdev->dcb_vlan_mode)
3906                 cm_info.vlan_id = 0;
3907         cm_info.backlog = backlog;
3908         cm_info.cm_id = cm_id;
3909
3910         cm_listen_node = irdma_make_listen_node(&iwdev->cm_core, iwdev,
3911                                                 &cm_info);
3912         if (!cm_listen_node) {
3913                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3914                             "cm_listen_node == NULL\n");
3915                 return -ENOMEM;
3916         }
3917
3918         cm_id->provider_data = cm_listen_node;
3919
3920         cm_listen_node->tos = cm_id->tos;
3921         if (iwdev->vsi.dscp_mode)
3922                 cm_listen_node->user_pri =
3923                     iwdev->vsi.dscp_map[irdma_tos2dscp(cm_id->tos)];
3924         else
3925                 cm_listen_node->user_pri = rt_tos2priority(cm_id->tos);
3926         cm_info.user_pri = cm_listen_node->user_pri;
3927         if (!cm_listen_node->reused_node) {
3928                 if (wildcard) {
3929                         err = irdma_add_mqh(iwdev, &cm_info, cm_listen_node);
3930                         if (err)
3931                                 goto error;
3932                 } else {
3933                         err = irdma_manage_qhash(iwdev, &cm_info,
3934                                                  IRDMA_QHASH_TYPE_TCP_SYN,
3935                                                  IRDMA_QHASH_MANAGE_TYPE_ADD,
3936                                                  NULL, true);
3937                         if (err)
3938                                 goto error;
3939
3940                         cm_listen_node->qhash_set = true;
3941                 }
3942
3943                 cm_listen_node->apbvt_entry = irdma_add_apbvt(iwdev,
3944                                                               cm_info.loc_port);
3945                 if (!cm_listen_node->apbvt_entry)
3946                         goto error;
3947         }
3948         cm_id->add_ref(cm_id);
3949         cm_listen_node->cm_core->stats_listen_created++;
3950         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3951                     "loc_port=0x%04x loc_addr=%pI4 cm_listen_node=%p cm_id=%p qhash_set=%d vlan_id=%d\n",
3952                     cm_listen_node->loc_port, cm_listen_node->loc_addr, cm_listen_node, cm_listen_node->cm_id,
3953                     cm_listen_node->qhash_set, cm_listen_node->vlan_id);
3954
3955         return 0;
3956
3957 error:
3958
3959         irdma_cm_del_listen(&iwdev->cm_core, cm_listen_node, false);
3960
3961         return -EINVAL;
3962 }
3963
3964 /**
3965  * irdma_destroy_listen - registered call to destroy listener
3966  * @cm_id: cm information for passive connection
3967  */
3968 int
3969 irdma_destroy_listen(struct iw_cm_id *cm_id)
3970 {
3971         struct irdma_device *iwdev;
3972
3973         iwdev = to_iwdev(cm_id->device);
3974         if (cm_id->provider_data)
3975                 irdma_cm_del_listen(&iwdev->cm_core, cm_id->provider_data,
3976                                     true);
3977         else
3978                 irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
3979                             "cm_id->provider_data was NULL\n");
3980
3981         cm_id->rem_ref(cm_id);
3982
3983         return 0;
3984 }
3985
3986 /**
3987  * irdma_teardown_list_prep - add conn nodes slated for tear down to list
3988  * @cm_core: cm's core
3989  * @teardown_list: a list to which cm_node will be selected
3990  * @ipaddr: pointer to ip address
3991  * @nfo: pointer to cm_info structure instance
3992  * @disconnect_all: flag indicating disconnect all QPs
3993  */
3994 static void
3995 irdma_teardown_list_prep(struct irdma_cm_core *cm_core,
3996                          struct list_head *teardown_list,
3997                          u32 *ipaddr,
3998                          struct irdma_cm_info *nfo,
3999                          bool disconnect_all)
4000 {
4001         struct irdma_cm_node *cm_node;
4002         int bkt;
4003
4004         HASH_FOR_EACH_RCU(cm_core->cm_hash_tbl, bkt, cm_node, list) {
4005                 if ((disconnect_all ||
4006                      (nfo->vlan_id == cm_node->vlan_id &&
4007                       !memcmp(cm_node->loc_addr, ipaddr, nfo->ipv4 ? 4 : 16))) &&
4008                     atomic_inc_not_zero(&cm_node->refcnt))
4009                         list_add(&cm_node->teardown_entry, teardown_list);
4010         }
4011 }
4012
4013 /**
4014  * irdma_cm_event_connected - handle connected active node
4015  * @event: the info for cm_node of connection
4016  */
4017 static void
4018 irdma_cm_event_connected(struct irdma_cm_event *event)
4019 {
4020         struct irdma_qp *iwqp;
4021         struct irdma_device *iwdev;
4022         struct irdma_cm_node *cm_node;
4023         struct irdma_sc_dev *dev;
4024         struct ib_qp_attr attr = {0};
4025         struct iw_cm_id *cm_id;
4026         int status;
4027         bool read0;
4028         int wait_ret = 0;
4029
4030         cm_node = event->cm_node;
4031         cm_id = cm_node->cm_id;
4032         iwqp = cm_id->provider_data;
4033         iwdev = iwqp->iwdev;
4034         dev = &iwdev->rf->sc_dev;
4035         if (iwqp->sc_qp.qp_uk.destroy_pending) {
4036                 status = -ETIMEDOUT;
4037                 goto error;
4038         }
4039
4040         irdma_cm_init_tsa_conn(iwqp, cm_node);
4041         read0 = (cm_node->send_rdma0_op == SEND_RDMA_READ_ZERO);
4042         if (iwqp->page)
4043                 iwqp->sc_qp.qp_uk.sq_base = kmap_local_page(iwqp->page);
4044         irdma_sc_send_rtt(&iwqp->sc_qp, read0);
4045         if (iwqp->page)
4046                 kunmap_local(iwqp->sc_qp.qp_uk.sq_base);
4047
4048         attr.qp_state = IB_QPS_RTS;
4049         cm_node->qhash_set = false;
4050         irdma_modify_qp(&iwqp->ibqp, &attr, IB_QP_STATE, NULL);
4051         if (dev->hw_attrs.uk_attrs.feature_flags & IRDMA_FEATURE_RTS_AE) {
4052                 wait_ret = wait_event_interruptible_timeout(iwqp->waitq,
4053                                                             iwqp->rts_ae_rcvd,
4054                                                             IRDMA_MAX_TIMEOUT);
4055                 if (!wait_ret)
4056                         irdma_debug(iwdev_to_idev(iwdev), IRDMA_DEBUG_CM,
4057                                     "Slow Connection: cm_node=%p, loc_port=%d, rem_port=%d, cm_id=%p\n",
4058                                     cm_node, cm_node->loc_port,
4059                                     cm_node->rem_port, cm_node->cm_id);
4060         }
4061
4062         irdma_send_cm_event(cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY, 0);
4063         cm_node->accelerated = true;
4064         complete(&cm_node->establish_comp);
4065         cm_node->cm_core->cm_free_ah(cm_node);
4066         return;
4067
4068 error:
4069         iwqp->cm_id = NULL;
4070         cm_id->provider_data = NULL;
4071         irdma_send_cm_event(event->cm_node, cm_id, IW_CM_EVENT_CONNECT_REPLY,
4072                             status);
4073         irdma_rem_ref_cm_node(event->cm_node);
4074 }
4075
4076 /**
4077  * irdma_cm_event_reset - handle reset
4078  * @event: the info for cm_node of connection
4079  */
4080 static void
4081 irdma_cm_event_reset(struct irdma_cm_event *event)
4082 {
4083         struct irdma_cm_node *cm_node = event->cm_node;
4084         struct iw_cm_id *cm_id = cm_node->cm_id;
4085         struct irdma_qp *iwqp;
4086
4087         if (!cm_id)
4088                 return;
4089
4090         iwqp = cm_id->provider_data;
4091         if (!iwqp)
4092                 return;
4093
4094         irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
4095                     "reset event %p - cm_id = %p\n", event->cm_node, cm_id);
4096         iwqp->cm_id = NULL;
4097
4098         irdma_send_cm_event(cm_node, cm_node->cm_id, IW_CM_EVENT_DISCONNECT,
4099                             -ECONNRESET);
4100         irdma_send_cm_event(cm_node, cm_node->cm_id, IW_CM_EVENT_CLOSE, 0);
4101 }
4102
4103 /**
4104  * irdma_cm_event_handler - send event to cm upper layer
4105  * @work: pointer of cm event info.
4106  */
4107 static void
4108 irdma_cm_event_handler(struct work_struct *work)
4109 {
4110         struct irdma_cm_event *event = container_of(work, struct irdma_cm_event, event_work);
4111         struct irdma_cm_node *cm_node;
4112
4113         if (!event || !event->cm_node || !event->cm_node->cm_core)
4114                 return;
4115
4116         cm_node = event->cm_node;
4117
4118         switch (event->type) {
4119         case IRDMA_CM_EVENT_MPA_REQ:
4120                 irdma_send_cm_event(cm_node, cm_node->cm_id,
4121                                     IW_CM_EVENT_CONNECT_REQUEST, 0);
4122                 break;
4123         case IRDMA_CM_EVENT_RESET:
4124                 irdma_cm_event_reset(event);
4125                 break;
4126         case IRDMA_CM_EVENT_CONNECTED:
4127                 if (!event->cm_node->cm_id ||
4128                     event->cm_node->state != IRDMA_CM_STATE_OFFLOADED)
4129                         break;
4130                 irdma_cm_event_connected(event);
4131                 break;
4132         case IRDMA_CM_EVENT_MPA_REJECT:
4133                 if (!event->cm_node->cm_id ||
4134                     cm_node->state == IRDMA_CM_STATE_OFFLOADED)
4135                         break;
4136                 irdma_send_cm_event(cm_node, cm_node->cm_id,
4137                                     IW_CM_EVENT_CONNECT_REPLY, -ECONNREFUSED);
4138                 break;
4139         case IRDMA_CM_EVENT_ABORTED:
4140                 if (!event->cm_node->cm_id ||
4141                     event->cm_node->state == IRDMA_CM_STATE_OFFLOADED)
4142                         break;
4143                 irdma_event_connect_error(event);
4144                 break;
4145         default:
4146                 irdma_debug(iwdev_to_idev(cm_node->iwdev), IRDMA_DEBUG_CM,
4147                             "bad event type = %d\n", event->type);
4148                 break;
4149         }
4150
4151         irdma_rem_ref_cm_node(event->cm_node);
4152         kfree(event);
4153 }
4154
4155 /**
4156  * irdma_cm_post_event - queue event request for worker thread
4157  * @event: cm node's info for up event call
4158  */
4159 static void
4160 irdma_cm_post_event(struct irdma_cm_event *event)
4161 {
4162         atomic_inc(&event->cm_node->refcnt);
4163         INIT_WORK(&event->event_work, irdma_cm_event_handler);
4164         queue_work(event->cm_node->cm_core->event_wq, &event->event_work);
4165 }
4166
4167 /**
4168  * irdma_cm_teardown_connections - teardown QPs
4169  * @iwdev: device pointer
4170  * @ipaddr: Pointer to IPv4 or IPv6 address
4171  * @nfo: Connection info
4172  * @disconnect_all: flag indicating disconnect all QPs
4173  *
4174  * teardown QPs where source or destination addr matches ip addr
4175  */
4176 void
4177 irdma_cm_teardown_connections(struct irdma_device *iwdev, u32 *ipaddr,
4178                               struct irdma_cm_info *nfo,
4179                               bool disconnect_all)
4180 {
4181         struct irdma_cm_core *cm_core = &iwdev->cm_core;
4182         struct list_head *list_core_temp;
4183         struct list_head *list_node;
4184         struct irdma_cm_node *cm_node;
4185         struct list_head teardown_list;
4186         struct ib_qp_attr attr;
4187         struct irdma_sc_vsi *vsi = &iwdev->vsi;
4188         struct irdma_sc_qp *sc_qp;
4189         struct irdma_qp *qp;
4190         int i;
4191
4192         INIT_LIST_HEAD(&teardown_list);
4193
4194         rcu_read_lock();
4195         irdma_teardown_list_prep(cm_core, &teardown_list, ipaddr, nfo, disconnect_all);
4196         rcu_read_unlock();
4197
4198         list_for_each_safe(list_node, list_core_temp, &teardown_list) {
4199                 cm_node = container_of(list_node, struct irdma_cm_node,
4200                                        teardown_entry);
4201                 attr.qp_state = IB_QPS_ERR;
4202                 irdma_modify_qp(&cm_node->iwqp->ibqp, &attr, IB_QP_STATE, NULL);
4203                 if (iwdev->rf->reset)
4204                         irdma_cm_disconn(cm_node->iwqp);
4205                 irdma_rem_ref_cm_node(cm_node);
4206         }
4207         if (!iwdev->roce_mode)
4208                 return;
4209
4210         INIT_LIST_HEAD(&teardown_list);
4211         for (i = 0; i < IRDMA_MAX_USER_PRIORITY; i++) {
4212                 mutex_lock(&vsi->qos[i].qos_mutex);
4213                 list_for_each_safe(list_node, list_core_temp,
4214                                    &vsi->qos[i].qplist) {
4215                         u32 qp_ip[4];
4216
4217                         sc_qp = container_of(list_node, struct irdma_sc_qp,
4218                                              list);
4219                         if (sc_qp->qp_uk.qp_type != IRDMA_QP_TYPE_ROCE_RC)
4220                                 continue;
4221
4222                         qp = sc_qp->qp_uk.back_qp;
4223                         if (!disconnect_all) {
4224                                 if (nfo->ipv4)
4225                                         qp_ip[0] = qp->udp_info.local_ipaddr[3];
4226                                 else
4227                                         memcpy(qp_ip,
4228                                                &qp->udp_info.local_ipaddr[0],
4229                                                sizeof(qp_ip));
4230                         }
4231
4232                         if (disconnect_all ||
4233                             (nfo->vlan_id == (qp->udp_info.vlan_tag & EVL_VLID_MASK) &&
4234                              !memcmp(qp_ip, ipaddr, nfo->ipv4 ? 4 : 16))) {
4235                                 spin_lock(&iwdev->rf->qptable_lock);
4236                                 if (iwdev->rf->qp_table[sc_qp->qp_uk.qp_id]) {
4237                                         irdma_qp_add_ref(&qp->ibqp);
4238                                         list_add(&qp->teardown_entry,
4239                                                  &teardown_list);
4240                                 }
4241                                 spin_unlock(&iwdev->rf->qptable_lock);
4242                         }
4243                 }
4244                 mutex_unlock(&vsi->qos[i].qos_mutex);
4245         }
4246
4247         list_for_each_safe(list_node, list_core_temp, &teardown_list) {
4248                 qp = container_of(list_node, struct irdma_qp, teardown_entry);
4249                 attr.qp_state = IB_QPS_ERR;
4250                 irdma_modify_qp_roce(&qp->ibqp, &attr, IB_QP_STATE, NULL);
4251                 irdma_qp_rem_ref(&qp->ibqp);
4252         }
4253 }