drivers/infiniband/sw/rxe/rxe_recv.c

   1 /*
   2  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
   3  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include <linux/skbuff.h>
  35
  36 #include "rxe.h"
  37 #include "rxe_loc.h"
  38
  39 static int check_type_state(struct rxe_dev *rxe, struct rxe_pkt_info *pkt,
  40                             struct rxe_qp *qp)
  41 {
  42         if (unlikely(!qp->valid))
  43                 goto err1;
  44
  45         switch (qp_type(qp)) {
  46         case IB_QPT_RC:
  47                 if (unlikely((pkt->opcode & IB_OPCODE_RC) != 0)) {
  48                         pr_warn_ratelimited("bad qp type\n");
  49                         goto err1;
  50                 }
  51                 break;
  52         case IB_QPT_UC:
  53                 if (unlikely(!(pkt->opcode & IB_OPCODE_UC))) {
  54                         pr_warn_ratelimited("bad qp type\n");
  55                         goto err1;
  56                 }
  57                 break;
  58         case IB_QPT_UD:
  59         case IB_QPT_SMI:
  60         case IB_QPT_GSI:
  61                 if (unlikely(!(pkt->opcode & IB_OPCODE_UD))) {
  62                         pr_warn_ratelimited("bad qp type\n");
  63                         goto err1;
  64                 }
  65                 break;
  66         default:
  67                 pr_warn_ratelimited("unsupported qp type\n");
  68                 goto err1;
  69         }
  70
  71         if (pkt->mask & RXE_REQ_MASK) {
  72                 if (unlikely(qp->resp.state != QP_STATE_READY))
  73                         goto err1;
  74         } else if (unlikely(qp->req.state < QP_STATE_READY ||
  75                                 qp->req.state > QP_STATE_DRAINED)) {
  76                 goto err1;
  77         }
  78
  79         return 0;
  80
  81 err1:
  82         return -EINVAL;
  83 }
  84
  85 static void set_bad_pkey_cntr(struct rxe_port *port)
  86 {
  87         spin_lock_bh(&port->port_lock);
  88         port->attr.bad_pkey_cntr = min((u32)0xffff,
  89                                        port->attr.bad_pkey_cntr + 1);
  90         spin_unlock_bh(&port->port_lock);
  91 }
  92
  93 static void set_qkey_viol_cntr(struct rxe_port *port)
  94 {
  95         spin_lock_bh(&port->port_lock);
  96         port->attr.qkey_viol_cntr = min((u32)0xffff,
  97                                         port->attr.qkey_viol_cntr + 1);
  98         spin_unlock_bh(&port->port_lock);
  99 }
 100
 101 static int check_keys(struct rxe_dev *rxe, struct rxe_pkt_info *pkt,
 102                       u32 qpn, struct rxe_qp *qp)
 103 {
 104         int i;
 105         int found_pkey = 0;
 106         struct rxe_port *port = &rxe->port;
 107         u16 pkey = bth_pkey(pkt);
 108
 109         pkt->pkey_index = 0;
 110
 111         if (qpn == 1) {
 112                 for (i = 0; i < port->attr.pkey_tbl_len; i++) {
 113                         if (pkey_match(pkey, port->pkey_tbl[i])) {
 114                                 pkt->pkey_index = i;
 115                                 found_pkey = 1;
 116                                 break;
 117                         }
 118                 }
 119
 120                 if (!found_pkey) {
 121                         pr_warn_ratelimited("bad pkey = 0x%x\n", pkey);
 122                         set_bad_pkey_cntr(port);
 123                         goto err1;
 124                 }
 125         } else if (qpn != 0) {
 126                 if (unlikely(!pkey_match(pkey,
 127                                          port->pkey_tbl[qp->attr.pkey_index]
 128                                         ))) {
 129                         pr_warn_ratelimited("bad pkey = 0x%0x\n", pkey);
 130                         set_bad_pkey_cntr(port);
 131                         goto err1;
 132                 }
 133                 pkt->pkey_index = qp->attr.pkey_index;
 134         }
 135
 136         if ((qp_type(qp) == IB_QPT_UD || qp_type(qp) == IB_QPT_GSI) &&
 137             qpn != 0 && pkt->mask) {
 138                 u32 qkey = (qpn == 1) ? GSI_QKEY : qp->attr.qkey;
 139
 140                 if (unlikely(deth_qkey(pkt) != qkey)) {
 141                         pr_warn_ratelimited("bad qkey, got 0x%x expected 0x%x for qpn 0x%x\n",
 142                                             deth_qkey(pkt), qkey, qpn);
 143                         set_qkey_viol_cntr(port);
 144                         goto err1;
 145                 }
 146         }
 147
 148         return 0;
 149
 150 err1:
 151         return -EINVAL;
 152 }
 153
 154 static int check_addr(struct rxe_dev *rxe, struct rxe_pkt_info *pkt,
 155                       struct rxe_qp *qp)
 156 {
 157         struct sk_buff *skb = PKT_TO_SKB(pkt);
 158
 159         if (qp_type(qp) != IB_QPT_RC && qp_type(qp) != IB_QPT_UC)
 160                 goto done;
 161
 162         if (unlikely(pkt->port_num != qp->attr.port_num)) {
 163                 pr_warn_ratelimited("port %d != qp port %d\n",
 164                                     pkt->port_num, qp->attr.port_num);
 165                 goto err1;
 166         }
 167
 168         if (skb->protocol == htons(ETH_P_IP)) {
 169                 struct in_addr *saddr =
 170                         &qp->pri_av.sgid_addr._sockaddr_in.sin_addr;
 171                 struct in_addr *daddr =
 172                         &qp->pri_av.dgid_addr._sockaddr_in.sin_addr;
 173
 174                 if (ip_hdr(skb)->daddr != saddr->s_addr) {
 175                         pr_warn_ratelimited("dst addr %pI4 != qp source addr %pI4\n",
 176                                             &ip_hdr(skb)->daddr,
 177                                             &saddr->s_addr);
 178                         goto err1;
 179                 }
 180
 181                 if (ip_hdr(skb)->saddr != daddr->s_addr) {
 182                         pr_warn_ratelimited("source addr %pI4 != qp dst addr %pI4\n",
 183                                             &ip_hdr(skb)->saddr,
 184                                             &daddr->s_addr);
 185                         goto err1;
 186                 }
 187
 188         } else if (skb->protocol == htons(ETH_P_IPV6)) {
 189                 struct in6_addr *saddr =
 190                         &qp->pri_av.sgid_addr._sockaddr_in6.sin6_addr;
 191                 struct in6_addr *daddr =
 192                         &qp->pri_av.dgid_addr._sockaddr_in6.sin6_addr;
 193
 194                 if (memcmp(&ipv6_hdr(skb)->daddr, saddr, sizeof(*saddr))) {
 195                         pr_warn_ratelimited("dst addr %pI6 != qp source addr %pI6\n",
 196                                             &ipv6_hdr(skb)->daddr, saddr);
 197                         goto err1;
 198                 }
 199
 200                 if (memcmp(&ipv6_hdr(skb)->saddr, daddr, sizeof(*daddr))) {
 201                         pr_warn_ratelimited("source addr %pI6 != qp dst addr %pI6\n",
 202                                             &ipv6_hdr(skb)->saddr, daddr);
 203                         goto err1;
 204                 }
 205         }
 206
 207 done:
 208         return 0;
 209
 210 err1:
 211         return -EINVAL;
 212 }
 213
 214 static int hdr_check(struct rxe_pkt_info *pkt)
 215 {
 216         struct rxe_dev *rxe = pkt->rxe;
 217         struct rxe_port *port = &rxe->port;
 218         struct rxe_qp *qp = NULL;
 219         u32 qpn = bth_qpn(pkt);
 220         int index;
 221         int err;
 222
 223         if (unlikely(bth_tver(pkt) != BTH_TVER)) {
 224                 pr_warn_ratelimited("bad tver\n");
 225                 goto err1;
 226         }
 227
 228         if (qpn != IB_MULTICAST_QPN) {
 229                 index = (qpn == 0) ? port->qp_smi_index :
 230                         ((qpn == 1) ? port->qp_gsi_index : qpn);
 231                 qp = rxe_pool_get_index(&rxe->qp_pool, index);
 232                 if (unlikely(!qp)) {
 233                         pr_warn_ratelimited("no qp matches qpn 0x%x\n", qpn);
 234                         goto err1;
 235                 }
 236
 237                 err = check_type_state(rxe, pkt, qp);
 238                 if (unlikely(err))
 239                         goto err2;
 240
 241                 err = check_addr(rxe, pkt, qp);
 242                 if (unlikely(err))
 243                         goto err2;
 244
 245                 err = check_keys(rxe, pkt, qpn, qp);
 246                 if (unlikely(err))
 247                         goto err2;
 248         } else {
 249                 if (unlikely((pkt->mask & RXE_GRH_MASK) == 0)) {
 250                         pr_warn_ratelimited("no grh for mcast qpn\n");
 251                         goto err1;
 252                 }
 253         }
 254
 255         pkt->qp = qp;
 256         return 0;
 257
 258 err2:
 259         if (qp)
 260                 rxe_drop_ref(qp);
 261 err1:
 262         return -EINVAL;
 263 }
 264
 265 static inline void rxe_rcv_pkt(struct rxe_dev *rxe,
 266                                struct rxe_pkt_info *pkt,
 267                                struct sk_buff *skb)
 268 {
 269         if (pkt->mask & RXE_REQ_MASK)
 270                 rxe_resp_queue_pkt(rxe, pkt->qp, skb);
 271         else
 272                 rxe_comp_queue_pkt(rxe, pkt->qp, skb);
 273 }
 274
 275 static void rxe_rcv_mcast_pkt(struct rxe_dev *rxe, struct sk_buff *skb)
 276 {
 277         struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
 278         struct rxe_mc_grp *mcg;
 279         struct sk_buff *skb_copy;
 280         struct rxe_mc_elem *mce;
 281         struct rxe_qp *qp;
 282         union ib_gid dgid;
 283         int err;
 284
 285         if (skb->protocol == htons(ETH_P_IP))
 286                 ipv6_addr_set_v4mapped(ip_hdr(skb)->daddr,
 287                                        (struct in6_addr *)&dgid);
 288         else if (skb->protocol == htons(ETH_P_IPV6))
 289                 memcpy(&dgid, &ipv6_hdr(skb)->daddr, sizeof(dgid));
 290
 291         /* lookup mcast group corresponding to mgid, takes a ref */
 292         mcg = rxe_pool_get_key(&rxe->mc_grp_pool, &dgid);
 293         if (!mcg)
 294                 goto err1;      /* mcast group not registered */
 295
 296         spin_lock_bh(&mcg->mcg_lock);
 297
 298         list_for_each_entry(mce, &mcg->qp_list, qp_list) {
 299                 qp = mce->qp;
 300                 pkt = SKB_TO_PKT(skb);
 301
 302                 /* validate qp for incoming packet */
 303                 err = check_type_state(rxe, pkt, qp);
 304                 if (err)
 305                         continue;
 306
 307                 err = check_keys(rxe, pkt, bth_qpn(pkt), qp);
 308                 if (err)
 309                         continue;
 310
 311                 /* if *not* the last qp in the list
 312                  * make a copy of the skb to post to the next qp
 313                  */
 314                 skb_copy = (mce->qp_list.next != &mcg->qp_list) ?
 315                                 skb_clone(skb, GFP_ATOMIC) : NULL;
 316
 317                 pkt->qp = qp;
 318                 rxe_add_ref(qp);
 319                 rxe_rcv_pkt(rxe, pkt, skb);
 320
 321                 skb = skb_copy;
 322                 if (!skb)
 323                         break;
 324         }
 325
 326         spin_unlock_bh(&mcg->mcg_lock);
 327
 328         rxe_drop_ref(mcg);      /* drop ref from rxe_pool_get_key. */
 329
 330 err1:
 331         if (skb)
 332                 kfree_skb(skb);
 333 }
 334
 335 static int rxe_match_dgid(struct rxe_dev *rxe, struct sk_buff *skb)
 336 {
 337         union ib_gid dgid;
 338         union ib_gid *pdgid;
 339         u16 index;
 340
 341         if (skb->protocol == htons(ETH_P_IP)) {
 342                 ipv6_addr_set_v4mapped(ip_hdr(skb)->daddr,
 343                                        (struct in6_addr *)&dgid);
 344                 pdgid = &dgid;
 345         } else {
 346                 pdgid = (union ib_gid *)&ipv6_hdr(skb)->daddr;
 347         }
 348
 349         return ib_find_cached_gid_by_port(&rxe->ib_dev, pdgid,
 350                                           IB_GID_TYPE_ROCE_UDP_ENCAP,
 351                                           1, rxe->ndev, &index);
 352 }
 353
 354 /* rxe_rcv is called from the interface driver */
 355 int rxe_rcv(struct sk_buff *skb)
 356 {
 357         int err;
 358         struct rxe_pkt_info *pkt = SKB_TO_PKT(skb);
 359         struct rxe_dev *rxe = pkt->rxe;
 360         __be32 *icrcp;
 361         u32 calc_icrc, pack_icrc;
 362
 363         pkt->offset = 0;
 364
 365         if (unlikely(skb->len < pkt->offset + RXE_BTH_BYTES))
 366                 goto drop;
 367
 368         if (unlikely(rxe_match_dgid(rxe, skb) < 0)) {
 369                 pr_warn_ratelimited("failed matching dgid\n");
 370                 goto drop;
 371         }
 372
 373         pkt->opcode = bth_opcode(pkt);
 374         pkt->psn = bth_psn(pkt);
 375         pkt->qp = NULL;
 376         pkt->mask |= rxe_opcode[pkt->opcode].mask;
 377
 378         if (unlikely(skb->len < header_size(pkt)))
 379                 goto drop;
 380
 381         err = hdr_check(pkt);
 382         if (unlikely(err))
 383                 goto drop;
 384
 385         /* Verify ICRC */
 386         icrcp = (__be32 *)(pkt->hdr + pkt->paylen - RXE_ICRC_SIZE);
 387         pack_icrc = be32_to_cpu(*icrcp);
 388
 389         calc_icrc = rxe_icrc_hdr(pkt, skb);
 390         calc_icrc = crc32_le(calc_icrc, (u8 *)payload_addr(pkt), payload_size(pkt));
 391         calc_icrc = cpu_to_be32(~calc_icrc);
 392         if (unlikely(calc_icrc != pack_icrc)) {
 393                 char saddr[sizeof(struct in6_addr)];
 394
 395                 if (skb->protocol == htons(ETH_P_IPV6))
 396                         sprintf(saddr, "%pI6", &ipv6_hdr(skb)->saddr);
 397                 else if (skb->protocol == htons(ETH_P_IP))
 398                         sprintf(saddr, "%pI4", &ip_hdr(skb)->saddr);
 399                 else
 400                         sprintf(saddr, "unknown");
 401
 402                 pr_warn_ratelimited("bad ICRC from %s\n", saddr);
 403                 goto drop;
 404         }
 405
 406         if (unlikely(bth_qpn(pkt) == IB_MULTICAST_QPN))
 407                 rxe_rcv_mcast_pkt(rxe, skb);
 408         else
 409                 rxe_rcv_pkt(rxe, pkt, skb);
 410
 411         return 0;
 412
 413 drop:
 414         if (pkt->qp)
 415                 rxe_drop_ref(pkt->qp);
 416
 417         kfree_skb(skb);
 418         return 0;
 419 }
 420 EXPORT_SYMBOL(rxe_rcv);