drivers/infiniband/hw/mlx4/qp.c

   1 /*
   2  * Copyright (c) 2007 Cisco Systems, Inc. All rights reserved.
   3  * Copyright (c) 2007, 2008 Mellanox Technologies. 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/log2.h>
  35 #include <linux/slab.h>
  36 #include <linux/netdevice.h>
  37
  38 #include <rdma/ib_cache.h>
  39 #include <rdma/ib_pack.h>
  40 #include <rdma/ib_addr.h>
  41 #include <rdma/ib_mad.h>
  42
  43 #include <linux/mlx4/qp.h>
  44
  45 #include "mlx4_ib.h"
  46 #include "user.h"
  47
  48 enum {
  49         MLX4_IB_ACK_REQ_FREQ    = 8,
  50 };
  51
  52 enum {
  53         MLX4_IB_DEFAULT_SCHED_QUEUE     = 0x83,
  54         MLX4_IB_DEFAULT_QP0_SCHED_QUEUE = 0x3f,
  55         MLX4_IB_LINK_TYPE_IB            = 0,
  56         MLX4_IB_LINK_TYPE_ETH           = 1
  57 };
  58
  59 enum {
  60         /*
  61          * Largest possible UD header: send with GRH and immediate
  62          * data plus 18 bytes for an Ethernet header with VLAN/802.1Q
  63          * tag.  (LRH would only use 8 bytes, so Ethernet is the
  64          * biggest case)
  65          */
  66         MLX4_IB_UD_HEADER_SIZE          = 82,
  67         MLX4_IB_LSO_HEADER_SPARE        = 128,
  68 };
  69
  70 enum {
  71         MLX4_IB_IBOE_ETHERTYPE          = 0x8915
  72 };
  73
  74 struct mlx4_ib_sqp {
  75         struct mlx4_ib_qp       qp;
  76         int                     pkey_index;
  77         u32                     qkey;
  78         u32                     send_psn;
  79         struct ib_ud_header     ud_header;
  80         u8                      header_buf[MLX4_IB_UD_HEADER_SIZE];
  81 };
  82
  83 enum {
  84         MLX4_IB_MIN_SQ_STRIDE   = 6,
  85         MLX4_IB_CACHE_LINE_SIZE = 64,
  86 };
  87
  88 enum {
  89         MLX4_RAW_QP_MTU         = 7,
  90         MLX4_RAW_QP_MSGMAX      = 31,
  91 };
  92
  93 #ifndef ETH_ALEN
  94 #define ETH_ALEN        6
  95 #endif
  96 static inline u64 mlx4_mac_to_u64(u8 *addr)
  97 {
  98         u64 mac = 0;
  99         int i;
 100
 101         for (i = 0; i < ETH_ALEN; i++) {
 102                 mac <<= 8;
 103                 mac |= addr[i];
 104         }
 105         return mac;
 106 }
 107
 108 static const __be32 mlx4_ib_opcode[] = {
 109         [IB_WR_SEND]                            = cpu_to_be32(MLX4_OPCODE_SEND),
 110         [IB_WR_LSO]                             = cpu_to_be32(MLX4_OPCODE_LSO),
 111         [IB_WR_SEND_WITH_IMM]                   = cpu_to_be32(MLX4_OPCODE_SEND_IMM),
 112         [IB_WR_RDMA_WRITE]                      = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE),
 113         [IB_WR_RDMA_WRITE_WITH_IMM]             = cpu_to_be32(MLX4_OPCODE_RDMA_WRITE_IMM),
 114         [IB_WR_RDMA_READ]                       = cpu_to_be32(MLX4_OPCODE_RDMA_READ),
 115         [IB_WR_ATOMIC_CMP_AND_SWP]              = cpu_to_be32(MLX4_OPCODE_ATOMIC_CS),
 116         [IB_WR_ATOMIC_FETCH_AND_ADD]            = cpu_to_be32(MLX4_OPCODE_ATOMIC_FA),
 117         [IB_WR_SEND_WITH_INV]                   = cpu_to_be32(MLX4_OPCODE_SEND_INVAL),
 118         [IB_WR_LOCAL_INV]                       = cpu_to_be32(MLX4_OPCODE_LOCAL_INVAL),
 119         [IB_WR_FAST_REG_MR]                     = cpu_to_be32(MLX4_OPCODE_FMR),
 120         [IB_WR_MASKED_ATOMIC_CMP_AND_SWP]       = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_CS),
 121         [IB_WR_MASKED_ATOMIC_FETCH_AND_ADD]     = cpu_to_be32(MLX4_OPCODE_MASKED_ATOMIC_FA),
 122         [IB_WR_BIND_MW]                         = cpu_to_be32(MLX4_OPCODE_BIND_MW),
 123 };
 124
 125 static struct mlx4_ib_sqp *to_msqp(struct mlx4_ib_qp *mqp)
 126 {
 127         return container_of(mqp, struct mlx4_ib_sqp, qp);
 128 }
 129
 130 static int is_tunnel_qp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
 131 {
 132         if (!mlx4_is_master(dev->dev))
 133                 return 0;
 134
 135         return qp->mqp.qpn >= dev->dev->phys_caps.base_tunnel_sqpn &&
 136                qp->mqp.qpn < dev->dev->phys_caps.base_tunnel_sqpn +
 137                 8 * MLX4_MFUNC_MAX;
 138 }
 139
 140 static int is_sqp(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
 141 {
 142         int proxy_sqp = 0;
 143         int real_sqp = 0;
 144         int i;
 145         /* PPF or Native -- real SQP */
 146         real_sqp = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
 147                     qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
 148                     qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 3);
 149         if (real_sqp)
 150                 return 1;
 151         /* VF or PF -- proxy SQP */
 152         if (mlx4_is_mfunc(dev->dev)) {
 153                 for (i = 0; i < dev->dev->caps.num_ports; i++) {
 154                         if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i] ||
 155                             qp->mqp.qpn == dev->dev->caps.qp1_proxy[i]) {
 156                                 proxy_sqp = 1;
 157                                 break;
 158                         }
 159                 }
 160         }
 161         return proxy_sqp;
 162 }
 163
 164 /* used for INIT/CLOSE port logic */
 165 static int is_qp0(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
 166 {
 167         int proxy_qp0 = 0;
 168         int real_qp0 = 0;
 169         int i;
 170         /* PPF or Native -- real QP0 */
 171         real_qp0 = ((mlx4_is_master(dev->dev) || !mlx4_is_mfunc(dev->dev)) &&
 172                     qp->mqp.qpn >= dev->dev->phys_caps.base_sqpn &&
 173                     qp->mqp.qpn <= dev->dev->phys_caps.base_sqpn + 1);
 174         if (real_qp0)
 175                 return 1;
 176         /* VF or PF -- proxy QP0 */
 177         if (mlx4_is_mfunc(dev->dev)) {
 178                 for (i = 0; i < dev->dev->caps.num_ports; i++) {
 179                         if (qp->mqp.qpn == dev->dev->caps.qp0_proxy[i]) {
 180                                 proxy_qp0 = 1;
 181                                 break;
 182                         }
 183                 }
 184         }
 185         return proxy_qp0;
 186 }
 187
 188 static void *get_wqe(struct mlx4_ib_qp *qp, int offset)
 189 {
 190         return mlx4_buf_offset(&qp->buf, offset);
 191 }
 192
 193 static void *get_recv_wqe(struct mlx4_ib_qp *qp, int n)
 194 {
 195         return get_wqe(qp, qp->rq.offset + (n << qp->rq.wqe_shift));
 196 }
 197
 198 static void *get_send_wqe(struct mlx4_ib_qp *qp, int n)
 199 {
 200         return get_wqe(qp, qp->sq.offset + (n << qp->sq.wqe_shift));
 201 }
 202
 203 /*
 204  * Stamp a SQ WQE so that it is invalid if prefetched by marking the
 205  * first four bytes of every 64 byte chunk with
 206  *     0x7FFFFFF | (invalid_ownership_value << 31).
 207  *
 208  * When the max work request size is less than or equal to the WQE
 209  * basic block size, as an optimization, we can stamp all WQEs with
 210  * 0xffffffff, and skip the very first chunk of each WQE.
 211  */
 212 static void stamp_send_wqe(struct mlx4_ib_qp *qp, int n, int size)
 213 {
 214         __be32 *wqe;
 215         int i;
 216         int s;
 217         int ind;
 218         void *buf;
 219         __be32 stamp;
 220         struct mlx4_wqe_ctrl_seg *ctrl;
 221
 222         if (qp->sq_max_wqes_per_wr > 1) {
 223                 s = roundup(size, 1U << qp->sq.wqe_shift);
 224                 for (i = 0; i < s; i += 64) {
 225                         ind = (i >> qp->sq.wqe_shift) + n;
 226                         stamp = ind & qp->sq.wqe_cnt ? cpu_to_be32(0x7fffffff) :
 227                                                        cpu_to_be32(0xffffffff);
 228                         buf = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
 229                         wqe = buf + (i & ((1 << qp->sq.wqe_shift) - 1));
 230                         *wqe = stamp;
 231                 }
 232         } else {
 233                 ctrl = buf = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
 234                 s = (ctrl->fence_size & 0x3f) << 4;
 235                 for (i = 64; i < s; i += 64) {
 236                         wqe = buf + i;
 237                         *wqe = cpu_to_be32(0xffffffff);
 238                 }
 239         }
 240 }
 241
 242 static void post_nop_wqe(struct mlx4_ib_qp *qp, int n, int size)
 243 {
 244         struct mlx4_wqe_ctrl_seg *ctrl;
 245         struct mlx4_wqe_inline_seg *inl;
 246         void *wqe;
 247         int s;
 248
 249         ctrl = wqe = get_send_wqe(qp, n & (qp->sq.wqe_cnt - 1));
 250         s = sizeof(struct mlx4_wqe_ctrl_seg);
 251
 252         if (qp->ibqp.qp_type == IB_QPT_UD) {
 253                 struct mlx4_wqe_datagram_seg *dgram = wqe + sizeof *ctrl;
 254                 struct mlx4_av *av = (struct mlx4_av *)dgram->av;
 255                 memset(dgram, 0, sizeof *dgram);
 256                 av->port_pd = cpu_to_be32((qp->port << 24) | to_mpd(qp->ibqp.pd)->pdn);
 257                 s += sizeof(struct mlx4_wqe_datagram_seg);
 258         }
 259
 260         /* Pad the remainder of the WQE with an inline data segment. */
 261         if (size > s) {
 262                 inl = wqe + s;
 263                 inl->byte_count = cpu_to_be32(1 << 31 | (size - s - sizeof *inl));
 264         }
 265         ctrl->srcrb_flags = 0;
 266         ctrl->fence_size = size / 16;
 267         /*
 268          * Make sure descriptor is fully written before setting ownership bit
 269          * (because HW can start executing as soon as we do).
 270          */
 271         wmb();
 272
 273         ctrl->owner_opcode = cpu_to_be32(MLX4_OPCODE_NOP | MLX4_WQE_CTRL_NEC) |
 274                 (n & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0);
 275
 276         stamp_send_wqe(qp, n + qp->sq_spare_wqes, size);
 277 }
 278
 279 /* Post NOP WQE to prevent wrap-around in the middle of WR */
 280 static inline unsigned pad_wraparound(struct mlx4_ib_qp *qp, int ind)
 281 {
 282         unsigned s = qp->sq.wqe_cnt - (ind & (qp->sq.wqe_cnt - 1));
 283         if (unlikely(s < qp->sq_max_wqes_per_wr)) {
 284                 post_nop_wqe(qp, ind, s << qp->sq.wqe_shift);
 285                 ind += s;
 286         }
 287         return ind;
 288 }
 289
 290 static void mlx4_ib_qp_event(struct mlx4_qp *qp, enum mlx4_event type)
 291 {
 292         struct ib_event event;
 293         struct ib_qp *ibqp = &to_mibqp(qp)->ibqp;
 294
 295         if (type == MLX4_EVENT_TYPE_PATH_MIG)
 296                 to_mibqp(qp)->port = to_mibqp(qp)->alt_port;
 297
 298         if (ibqp->event_handler) {
 299                 event.device     = ibqp->device;
 300                 event.element.qp = ibqp;
 301                 switch (type) {
 302                 case MLX4_EVENT_TYPE_PATH_MIG:
 303                         event.event = IB_EVENT_PATH_MIG;
 304                         break;
 305                 case MLX4_EVENT_TYPE_COMM_EST:
 306                         event.event = IB_EVENT_COMM_EST;
 307                         break;
 308                 case MLX4_EVENT_TYPE_SQ_DRAINED:
 309                         event.event = IB_EVENT_SQ_DRAINED;
 310                         break;
 311                 case MLX4_EVENT_TYPE_SRQ_QP_LAST_WQE:
 312                         event.event = IB_EVENT_QP_LAST_WQE_REACHED;
 313                         break;
 314                 case MLX4_EVENT_TYPE_WQ_CATAS_ERROR:
 315                         event.event = IB_EVENT_QP_FATAL;
 316                         break;
 317                 case MLX4_EVENT_TYPE_PATH_MIG_FAILED:
 318                         event.event = IB_EVENT_PATH_MIG_ERR;
 319                         break;
 320                 case MLX4_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
 321                         event.event = IB_EVENT_QP_REQ_ERR;
 322                         break;
 323                 case MLX4_EVENT_TYPE_WQ_ACCESS_ERROR:
 324                         event.event = IB_EVENT_QP_ACCESS_ERR;
 325                         break;
 326                 default:
 327                         pr_warn("Unexpected event type %d "
 328                                "on QP %06x\n", type, qp->qpn);
 329                         return;
 330                 }
 331
 332                 ibqp->event_handler(&event, ibqp->qp_context);
 333         }
 334 }
 335
 336 static int send_wqe_overhead(enum mlx4_ib_qp_type type, u32 flags)
 337 {
 338         /*
 339          * UD WQEs must have a datagram segment.
 340          * RC and UC WQEs might have a remote address segment.
 341          * MLX WQEs need two extra inline data segments (for the UD
 342          * header and space for the ICRC).
 343          */
 344         switch (type) {
 345         case MLX4_IB_QPT_UD:
 346                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 347                         sizeof (struct mlx4_wqe_datagram_seg) +
 348                         ((flags & MLX4_IB_QP_LSO) ? MLX4_IB_LSO_HEADER_SPARE : 0);
 349         case MLX4_IB_QPT_PROXY_SMI_OWNER:
 350         case MLX4_IB_QPT_PROXY_SMI:
 351         case MLX4_IB_QPT_PROXY_GSI:
 352                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 353                         sizeof (struct mlx4_wqe_datagram_seg) + 64;
 354         case MLX4_IB_QPT_TUN_SMI_OWNER:
 355         case MLX4_IB_QPT_TUN_GSI:
 356                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 357                         sizeof (struct mlx4_wqe_datagram_seg);
 358
 359         case MLX4_IB_QPT_UC:
 360                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 361                         sizeof (struct mlx4_wqe_raddr_seg);
 362         case MLX4_IB_QPT_RC:
 363                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 364                         sizeof (struct mlx4_wqe_atomic_seg) +
 365                         sizeof (struct mlx4_wqe_raddr_seg);
 366         case MLX4_IB_QPT_SMI:
 367         case MLX4_IB_QPT_GSI:
 368                 return sizeof (struct mlx4_wqe_ctrl_seg) +
 369                         ALIGN(MLX4_IB_UD_HEADER_SIZE +
 370                               DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
 371                                            MLX4_INLINE_ALIGN) *
 372                               sizeof (struct mlx4_wqe_inline_seg),
 373                               sizeof (struct mlx4_wqe_data_seg)) +
 374                         ALIGN(4 +
 375                               sizeof (struct mlx4_wqe_inline_seg),
 376                               sizeof (struct mlx4_wqe_data_seg));
 377         default:
 378                 return sizeof (struct mlx4_wqe_ctrl_seg);
 379         }
 380 }
 381
 382 static int set_rq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 383                        int is_user, int has_rq, struct mlx4_ib_qp *qp)
 384 {
 385         /* Sanity check RQ size before proceeding */
 386         if (cap->max_recv_wr > dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE ||
 387             cap->max_recv_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg))
 388                 return -EINVAL;
 389
 390         if (!has_rq) {
 391                 if (cap->max_recv_wr)
 392                         return -EINVAL;
 393
 394                 qp->rq.wqe_cnt = qp->rq.max_gs = 0;
 395         } else {
 396                 /* HW requires >= 1 RQ entry with >= 1 gather entry */
 397                 if (is_user && (!cap->max_recv_wr || !cap->max_recv_sge))
 398                         return -EINVAL;
 399
 400                 qp->rq.wqe_cnt   = roundup_pow_of_two(max(1U, cap->max_recv_wr));
 401                 qp->rq.max_gs    = roundup_pow_of_two(max(1U, cap->max_recv_sge));
 402                 qp->rq.wqe_shift = ilog2(qp->rq.max_gs * sizeof (struct mlx4_wqe_data_seg));
 403         }
 404
 405         /* leave userspace return values as they were, so as not to break ABI */
 406         if (is_user) {
 407                 cap->max_recv_wr  = qp->rq.max_post = qp->rq.wqe_cnt;
 408                 cap->max_recv_sge = qp->rq.max_gs;
 409         } else {
 410                 cap->max_recv_wr  = qp->rq.max_post =
 411                         min(dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE, qp->rq.wqe_cnt);
 412                 cap->max_recv_sge = min(qp->rq.max_gs,
 413                                         min(dev->dev->caps.max_sq_sg,
 414                                             dev->dev->caps.max_rq_sg));
 415         }
 416
 417         return 0;
 418 }
 419
 420 static int set_kernel_sq_size(struct mlx4_ib_dev *dev, struct ib_qp_cap *cap,
 421                               enum mlx4_ib_qp_type type, struct mlx4_ib_qp *qp)
 422 {
 423         int s;
 424
 425         /* Sanity check SQ size before proceeding */
 426         if (cap->max_send_wr  > (dev->dev->caps.max_wqes - MLX4_IB_SQ_MAX_SPARE) ||
 427             cap->max_send_sge > min(dev->dev->caps.max_sq_sg, dev->dev->caps.max_rq_sg) ||
 428             cap->max_inline_data + send_wqe_overhead(type, qp->flags) +
 429             sizeof (struct mlx4_wqe_inline_seg) > dev->dev->caps.max_sq_desc_sz)
 430                 return -EINVAL;
 431
 432         /*
 433          * For MLX transport we need 2 extra S/G entries:
 434          * one for the header and one for the checksum at the end
 435          */
 436         if ((type == MLX4_IB_QPT_SMI || type == MLX4_IB_QPT_GSI ||
 437              type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) &&
 438             cap->max_send_sge + 2 > dev->dev->caps.max_sq_sg)
 439                 return -EINVAL;
 440
 441         s = max(cap->max_send_sge * sizeof (struct mlx4_wqe_data_seg),
 442                 cap->max_inline_data + sizeof (struct mlx4_wqe_inline_seg)) +
 443                 send_wqe_overhead(type, qp->flags);
 444
 445         if (s > dev->dev->caps.max_sq_desc_sz)
 446                 return -EINVAL;
 447
 448         /*
 449          * Hermon supports shrinking WQEs, such that a single work
 450          * request can include multiple units of 1 << wqe_shift.  This
 451          * way, work requests can differ in size, and do not have to
 452          * be a power of 2 in size, saving memory and speeding up send
 453          * WR posting.  Unfortunately, if we do this then the
 454          * wqe_index field in CQEs can't be used to look up the WR ID
 455          * anymore, so we do this only if selective signaling is off.
 456          *
 457          * Further, on 32-bit platforms, we can't use vmap() to make
 458          * the QP buffer virtually contiguous.  Thus we have to use
 459          * constant-sized WRs to make sure a WR is always fully within
 460          * a single page-sized chunk.
 461          *
 462          * Finally, we use NOP work requests to pad the end of the
 463          * work queue, to avoid wrap-around in the middle of WR.  We
 464          * set NEC bit to avoid getting completions with error for
 465          * these NOP WRs, but since NEC is only supported starting
 466          * with firmware 2.2.232, we use constant-sized WRs for older
 467          * firmware.
 468          *
 469          * And, since MLX QPs only support SEND, we use constant-sized
 470          * WRs in this case.
 471          *
 472          * We look for the smallest value of wqe_shift such that the
 473          * resulting number of wqes does not exceed device
 474          * capabilities.
 475          *
 476          * We set WQE size to at least 64 bytes, this way stamping
 477          * invalidates each WQE.
 478          */
 479         if (dev->dev->caps.fw_ver >= MLX4_FW_VER_WQE_CTRL_NEC &&
 480             qp->sq_signal_bits && BITS_PER_LONG == 64 &&
 481             type != MLX4_IB_QPT_SMI && type != MLX4_IB_QPT_GSI &&
 482             !(type & (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_PROXY_SMI |
 483                       MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER)))
 484                 qp->sq.wqe_shift = ilog2(64);
 485         else
 486                 qp->sq.wqe_shift = ilog2(roundup_pow_of_two(s));
 487
 488         for (;;) {
 489                 qp->sq_max_wqes_per_wr = DIV_ROUND_UP(s, 1U << qp->sq.wqe_shift);
 490
 491                 /*
 492                  * We need to leave 2 KB + 1 WR of headroom in the SQ to
 493                  * allow HW to prefetch.
 494                  */
 495                 qp->sq_spare_wqes = (2048 >> qp->sq.wqe_shift) + qp->sq_max_wqes_per_wr;
 496                 qp->sq.wqe_cnt = roundup_pow_of_two(cap->max_send_wr *
 497                                                     qp->sq_max_wqes_per_wr +
 498                                                     qp->sq_spare_wqes);
 499
 500                 if (qp->sq.wqe_cnt <= dev->dev->caps.max_wqes)
 501                         break;
 502
 503                 if (qp->sq_max_wqes_per_wr <= 1)
 504                         return -EINVAL;
 505
 506                 ++qp->sq.wqe_shift;
 507         }
 508
 509         qp->sq.max_gs = (min(dev->dev->caps.max_sq_desc_sz,
 510                              (qp->sq_max_wqes_per_wr << qp->sq.wqe_shift)) -
 511                          send_wqe_overhead(type, qp->flags)) /
 512                 sizeof (struct mlx4_wqe_data_seg);
 513
 514         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
 515                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
 516         if (qp->rq.wqe_shift > qp->sq.wqe_shift) {
 517                 qp->rq.offset = 0;
 518                 qp->sq.offset = qp->rq.wqe_cnt << qp->rq.wqe_shift;
 519         } else {
 520                 qp->rq.offset = qp->sq.wqe_cnt << qp->sq.wqe_shift;
 521                 qp->sq.offset = 0;
 522         }
 523
 524         cap->max_send_wr  = qp->sq.max_post =
 525                 (qp->sq.wqe_cnt - qp->sq_spare_wqes) / qp->sq_max_wqes_per_wr;
 526         cap->max_send_sge = min(qp->sq.max_gs,
 527                                 min(dev->dev->caps.max_sq_sg,
 528                                     dev->dev->caps.max_rq_sg));
 529         /* We don't support inline sends for kernel QPs (yet) */
 530         cap->max_inline_data = 0;
 531
 532         return 0;
 533 }
 534
 535 static int set_user_sq_size(struct mlx4_ib_dev *dev,
 536                             struct mlx4_ib_qp *qp,
 537                             struct mlx4_ib_create_qp *ucmd)
 538 {
 539         /* Sanity check SQ size before proceeding */
 540         if ((1 << ucmd->log_sq_bb_count) > dev->dev->caps.max_wqes       ||
 541             ucmd->log_sq_stride >
 542                 ilog2(roundup_pow_of_two(dev->dev->caps.max_sq_desc_sz)) ||
 543             ucmd->log_sq_stride < MLX4_IB_MIN_SQ_STRIDE)
 544                 return -EINVAL;
 545
 546         qp->sq.wqe_cnt   = 1 << ucmd->log_sq_bb_count;
 547         qp->sq.wqe_shift = ucmd->log_sq_stride;
 548
 549         qp->buf_size = (qp->rq.wqe_cnt << qp->rq.wqe_shift) +
 550                 (qp->sq.wqe_cnt << qp->sq.wqe_shift);
 551
 552         return 0;
 553 }
 554
 555 static int alloc_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
 556 {
 557         int i;
 558
 559         qp->sqp_proxy_rcv =
 560                 kmalloc(sizeof (struct mlx4_ib_buf) * qp->rq.wqe_cnt,
 561                         GFP_KERNEL);
 562         if (!qp->sqp_proxy_rcv)
 563                 return -ENOMEM;
 564         for (i = 0; i < qp->rq.wqe_cnt; i++) {
 565                 qp->sqp_proxy_rcv[i].addr =
 566                         kmalloc(sizeof (struct mlx4_ib_proxy_sqp_hdr),
 567                                 GFP_KERNEL);
 568                 if (!qp->sqp_proxy_rcv[i].addr)
 569                         goto err;
 570                 qp->sqp_proxy_rcv[i].map =
 571                         ib_dma_map_single(dev, qp->sqp_proxy_rcv[i].addr,
 572                                           sizeof (struct mlx4_ib_proxy_sqp_hdr),
 573                                           DMA_FROM_DEVICE);
 574         }
 575         return 0;
 576
 577 err:
 578         while (i > 0) {
 579                 --i;
 580                 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
 581                                     sizeof (struct mlx4_ib_proxy_sqp_hdr),
 582                                     DMA_FROM_DEVICE);
 583                 kfree(qp->sqp_proxy_rcv[i].addr);
 584         }
 585         kfree(qp->sqp_proxy_rcv);
 586         qp->sqp_proxy_rcv = NULL;
 587         return -ENOMEM;
 588 }
 589
 590 static void free_proxy_bufs(struct ib_device *dev, struct mlx4_ib_qp *qp)
 591 {
 592         int i;
 593
 594         for (i = 0; i < qp->rq.wqe_cnt; i++) {
 595                 ib_dma_unmap_single(dev, qp->sqp_proxy_rcv[i].map,
 596                                     sizeof (struct mlx4_ib_proxy_sqp_hdr),
 597                                     DMA_FROM_DEVICE);
 598                 kfree(qp->sqp_proxy_rcv[i].addr);
 599         }
 600         kfree(qp->sqp_proxy_rcv);
 601 }
 602
 603 static int qp_has_rq(struct ib_qp_init_attr *attr)
 604 {
 605         if (attr->qp_type == IB_QPT_XRC_INI || attr->qp_type == IB_QPT_XRC_TGT)
 606                 return 0;
 607
 608         return !attr->srq;
 609 }
 610
 611 static int create_qp_common(struct mlx4_ib_dev *dev, struct ib_pd *pd,
 612                             struct ib_qp_init_attr *init_attr,
 613                             struct ib_udata *udata, int sqpn, struct mlx4_ib_qp **caller_qp,
 614                             gfp_t gfp)
 615 {
 616         int qpn;
 617         int err;
 618         struct mlx4_ib_sqp *sqp;
 619         struct mlx4_ib_qp *qp;
 620         enum mlx4_ib_qp_type qp_type = (enum mlx4_ib_qp_type) init_attr->qp_type;
 621
 622         /* When tunneling special qps, we use a plain UD qp */
 623         if (sqpn) {
 624                 if (mlx4_is_mfunc(dev->dev) &&
 625                     (!mlx4_is_master(dev->dev) ||
 626                      !(init_attr->create_flags & MLX4_IB_SRIOV_SQP))) {
 627                         if (init_attr->qp_type == IB_QPT_GSI)
 628                                 qp_type = MLX4_IB_QPT_PROXY_GSI;
 629                         else if (mlx4_is_master(dev->dev))
 630                                 qp_type = MLX4_IB_QPT_PROXY_SMI_OWNER;
 631                         else
 632                                 qp_type = MLX4_IB_QPT_PROXY_SMI;
 633                 }
 634                 qpn = sqpn;
 635                 /* add extra sg entry for tunneling */
 636                 init_attr->cap.max_recv_sge++;
 637         } else if (init_attr->create_flags & MLX4_IB_SRIOV_TUNNEL_QP) {
 638                 struct mlx4_ib_qp_tunnel_init_attr *tnl_init =
 639                         container_of(init_attr,
 640                                      struct mlx4_ib_qp_tunnel_init_attr, init_attr);
 641                 if ((tnl_init->proxy_qp_type != IB_QPT_SMI &&
 642                      tnl_init->proxy_qp_type != IB_QPT_GSI)   ||
 643                     !mlx4_is_master(dev->dev))
 644                         return -EINVAL;
 645                 if (tnl_init->proxy_qp_type == IB_QPT_GSI)
 646                         qp_type = MLX4_IB_QPT_TUN_GSI;
 647                 else if (tnl_init->slave == mlx4_master_func_num(dev->dev))
 648                         qp_type = MLX4_IB_QPT_TUN_SMI_OWNER;
 649                 else
 650                         qp_type = MLX4_IB_QPT_TUN_SMI;
 651                 /* we are definitely in the PPF here, since we are creating
 652                  * tunnel QPs. base_tunnel_sqpn is therefore valid. */
 653                 qpn = dev->dev->phys_caps.base_tunnel_sqpn + 8 * tnl_init->slave
 654                         + tnl_init->proxy_qp_type * 2 + tnl_init->port - 1;
 655                 sqpn = qpn;
 656         }
 657
 658         if (!*caller_qp) {
 659                 if (qp_type == MLX4_IB_QPT_SMI || qp_type == MLX4_IB_QPT_GSI ||
 660                     (qp_type & (MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_SMI_OWNER |
 661                                 MLX4_IB_QPT_PROXY_GSI | MLX4_IB_QPT_TUN_SMI_OWNER))) {
 662                         sqp = kzalloc(sizeof (struct mlx4_ib_sqp), GFP_KERNEL);
 663                         if (!sqp)
 664                                 return -ENOMEM;
 665                         qp = &sqp->qp;
 666                         qp->pri.vid = 0xFFFF;
 667                         qp->alt.vid = 0xFFFF;
 668                 } else {
 669                         qp = kzalloc(sizeof (struct mlx4_ib_qp), GFP_KERNEL);
 670                         if (!qp)
 671                                 return -ENOMEM;
 672                         qp->pri.vid = 0xFFFF;
 673                         qp->alt.vid = 0xFFFF;
 674                 }
 675         } else
 676                 qp = *caller_qp;
 677
 678         qp->mlx4_ib_qp_type = qp_type;
 679
 680         mutex_init(&qp->mutex);
 681         spin_lock_init(&qp->sq.lock);
 682         spin_lock_init(&qp->rq.lock);
 683         INIT_LIST_HEAD(&qp->gid_list);
 684         INIT_LIST_HEAD(&qp->steering_rules);
 685
 686         qp->state        = IB_QPS_RESET;
 687         if (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR)
 688                 qp->sq_signal_bits = cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
 689
 690         err = set_rq_size(dev, &init_attr->cap, !!pd->uobject, qp_has_rq(init_attr), qp);
 691         if (err)
 692                 goto err;
 693
 694         if (pd->uobject) {
 695                 struct mlx4_ib_create_qp ucmd;
 696
 697                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 698                         err = -EFAULT;
 699                         goto err;
 700                 }
 701
 702                 qp->sq_no_prefetch = ucmd.sq_no_prefetch;
 703
 704                 err = set_user_sq_size(dev, qp, &ucmd);
 705                 if (err)
 706                         goto err;
 707
 708                 qp->umem = ib_umem_get(pd->uobject->context, ucmd.buf_addr,
 709                                        qp->buf_size, 0, 0);
 710                 if (IS_ERR(qp->umem)) {
 711                         err = PTR_ERR(qp->umem);
 712                         goto err;
 713                 }
 714
 715                 err = mlx4_mtt_init(dev->dev, ib_umem_page_count(qp->umem),
 716                                     ilog2(qp->umem->page_size), &qp->mtt);
 717                 if (err)
 718                         goto err_buf;
 719
 720                 err = mlx4_ib_umem_write_mtt(dev, &qp->mtt, qp->umem);
 721                 if (err)
 722                         goto err_mtt;
 723
 724                 if (qp_has_rq(init_attr)) {
 725                         err = mlx4_ib_db_map_user(to_mucontext(pd->uobject->context),
 726                                                   ucmd.db_addr, &qp->db);
 727                         if (err)
 728                                 goto err_mtt;
 729                 }
 730         } else {
 731                 qp->sq_no_prefetch = 0;
 732
 733                 if (init_attr->create_flags & IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK)
 734                         qp->flags |= MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK;
 735
 736                 if (init_attr->create_flags & IB_QP_CREATE_IPOIB_UD_LSO)
 737                         qp->flags |= MLX4_IB_QP_LSO;
 738
 739                 if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
 740                         if (dev->steering_support ==
 741                             MLX4_STEERING_MODE_DEVICE_MANAGED)
 742                                 qp->flags |= MLX4_IB_QP_NETIF;
 743                         else
 744                                 goto err;
 745                 }
 746
 747                 err = set_kernel_sq_size(dev, &init_attr->cap, qp_type, qp);
 748                 if (err)
 749                         goto err;
 750
 751                 if (qp_has_rq(init_attr)) {
 752                         err = mlx4_db_alloc(dev->dev, &qp->db, 0, gfp);
 753                         if (err)
 754                                 goto err;
 755
 756                         *qp->db.db = 0;
 757                 }
 758
 759                 if (mlx4_buf_alloc(dev->dev, qp->buf_size, PAGE_SIZE * 2, &qp->buf, gfp)) {
 760                         err = -ENOMEM;
 761                         goto err_db;
 762                 }
 763
 764                 err = mlx4_mtt_init(dev->dev, qp->buf.npages, qp->buf.page_shift,
 765                                     &qp->mtt);
 766                 if (err)
 767                         goto err_buf;
 768
 769                 err = mlx4_buf_write_mtt(dev->dev, &qp->mtt, &qp->buf, gfp);
 770                 if (err)
 771                         goto err_mtt;
 772
 773                 qp->sq.wrid  = kmalloc(qp->sq.wqe_cnt * sizeof (u64), gfp);
 774                 qp->rq.wrid  = kmalloc(qp->rq.wqe_cnt * sizeof (u64), gfp);
 775                 if (!qp->sq.wrid || !qp->rq.wrid) {
 776                         err = -ENOMEM;
 777                         goto err_wrid;
 778                 }
 779         }
 780
 781         if (sqpn) {
 782                 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
 783                     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
 784                         if (alloc_proxy_bufs(pd->device, qp)) {
 785                                 err = -ENOMEM;
 786                                 goto err_wrid;
 787                         }
 788                 }
 789         } else {
 790                 /* Raw packet QPNs must be aligned to 8 bits. If not, the WQE
 791                  * BlueFlame setup flow wrongly causes VLAN insertion. */
 792                 if (init_attr->qp_type == IB_QPT_RAW_PACKET)
 793                         err = mlx4_qp_reserve_range(dev->dev, 1, 1 << 8, &qpn);
 794                 else
 795                         if (qp->flags & MLX4_IB_QP_NETIF)
 796                                 err = mlx4_ib_steer_qp_alloc(dev, 1, &qpn);
 797                         else
 798                                 err = mlx4_qp_reserve_range(dev->dev, 1, 1,
 799                                                             &qpn);
 800                 if (err)
 801                         goto err_proxy;
 802         }
 803
 804         err = mlx4_qp_alloc(dev->dev, qpn, &qp->mqp, gfp);
 805         if (err)
 806                 goto err_qpn;
 807
 808         if (init_attr->qp_type == IB_QPT_XRC_TGT)
 809                 qp->mqp.qpn |= (1 << 23);
 810
 811         /*
 812          * Hardware wants QPN written in big-endian order (after
 813          * shifting) for send doorbell.  Precompute this value to save
 814          * a little bit when posting sends.
 815          */
 816         qp->doorbell_qpn = swab32(qp->mqp.qpn << 8);
 817
 818         qp->mqp.event = mlx4_ib_qp_event;
 819         if (!*caller_qp)
 820                 *caller_qp = qp;
 821         return 0;
 822
 823 err_qpn:
 824         if (!sqpn) {
 825                 if (qp->flags & MLX4_IB_QP_NETIF)
 826                         mlx4_ib_steer_qp_free(dev, qpn, 1);
 827                 else
 828                         mlx4_qp_release_range(dev->dev, qpn, 1);
 829         }
 830 err_proxy:
 831         if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI)
 832                 free_proxy_bufs(pd->device, qp);
 833 err_wrid:
 834         if (pd->uobject) {
 835                 if (qp_has_rq(init_attr))
 836                         mlx4_ib_db_unmap_user(to_mucontext(pd->uobject->context), &qp->db);
 837         } else {
 838                 kfree(qp->sq.wrid);
 839                 kfree(qp->rq.wrid);
 840         }
 841
 842 err_mtt:
 843         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
 844
 845 err_buf:
 846         if (pd->uobject)
 847                 ib_umem_release(qp->umem);
 848         else
 849                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
 850
 851 err_db:
 852         if (!pd->uobject && qp_has_rq(init_attr))
 853                 mlx4_db_free(dev->dev, &qp->db);
 854
 855 err:
 856         if (!*caller_qp)
 857                 kfree(qp);
 858         return err;
 859 }
 860
 861 static enum mlx4_qp_state to_mlx4_state(enum ib_qp_state state)
 862 {
 863         switch (state) {
 864         case IB_QPS_RESET:      return MLX4_QP_STATE_RST;
 865         case IB_QPS_INIT:       return MLX4_QP_STATE_INIT;
 866         case IB_QPS_RTR:        return MLX4_QP_STATE_RTR;
 867         case IB_QPS_RTS:        return MLX4_QP_STATE_RTS;
 868         case IB_QPS_SQD:        return MLX4_QP_STATE_SQD;
 869         case IB_QPS_SQE:        return MLX4_QP_STATE_SQER;
 870         case IB_QPS_ERR:        return MLX4_QP_STATE_ERR;
 871         default:                return -1;
 872         }
 873 }
 874
 875 static void mlx4_ib_lock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
 876         __acquires(&send_cq->lock) __acquires(&recv_cq->lock)
 877 {
 878         if (send_cq == recv_cq) {
 879                 spin_lock_irq(&send_cq->lock);
 880                 __acquire(&recv_cq->lock);
 881         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
 882                 spin_lock_irq(&send_cq->lock);
 883                 spin_lock_nested(&recv_cq->lock, SINGLE_DEPTH_NESTING);
 884         } else {
 885                 spin_lock_irq(&recv_cq->lock);
 886                 spin_lock_nested(&send_cq->lock, SINGLE_DEPTH_NESTING);
 887         }
 888 }
 889
 890 static void mlx4_ib_unlock_cqs(struct mlx4_ib_cq *send_cq, struct mlx4_ib_cq *recv_cq)
 891         __releases(&send_cq->lock) __releases(&recv_cq->lock)
 892 {
 893         if (send_cq == recv_cq) {
 894                 __release(&recv_cq->lock);
 895                 spin_unlock_irq(&send_cq->lock);
 896         } else if (send_cq->mcq.cqn < recv_cq->mcq.cqn) {
 897                 spin_unlock(&recv_cq->lock);
 898                 spin_unlock_irq(&send_cq->lock);
 899         } else {
 900                 spin_unlock(&send_cq->lock);
 901                 spin_unlock_irq(&recv_cq->lock);
 902         }
 903 }
 904
 905 static void del_gid_entries(struct mlx4_ib_qp *qp)
 906 {
 907         struct mlx4_ib_gid_entry *ge, *tmp;
 908
 909         list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
 910                 list_del(&ge->list);
 911                 kfree(ge);
 912         }
 913 }
 914
 915 static struct mlx4_ib_pd *get_pd(struct mlx4_ib_qp *qp)
 916 {
 917         if (qp->ibqp.qp_type == IB_QPT_XRC_TGT)
 918                 return to_mpd(to_mxrcd(qp->ibqp.xrcd)->pd);
 919         else
 920                 return to_mpd(qp->ibqp.pd);
 921 }
 922
 923 static void get_cqs(struct mlx4_ib_qp *qp,
 924                     struct mlx4_ib_cq **send_cq, struct mlx4_ib_cq **recv_cq)
 925 {
 926         switch (qp->ibqp.qp_type) {
 927         case IB_QPT_XRC_TGT:
 928                 *send_cq = to_mcq(to_mxrcd(qp->ibqp.xrcd)->cq);
 929                 *recv_cq = *send_cq;
 930                 break;
 931         case IB_QPT_XRC_INI:
 932                 *send_cq = to_mcq(qp->ibqp.send_cq);
 933                 *recv_cq = *send_cq;
 934                 break;
 935         default:
 936                 *send_cq = to_mcq(qp->ibqp.send_cq);
 937                 *recv_cq = to_mcq(qp->ibqp.recv_cq);
 938                 break;
 939         }
 940 }
 941
 942 static void destroy_qp_common(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp,
 943                               int is_user)
 944 {
 945         struct mlx4_ib_cq *send_cq, *recv_cq;
 946
 947         if (qp->state != IB_QPS_RESET) {
 948                 if (mlx4_qp_modify(dev->dev, NULL, to_mlx4_state(qp->state),
 949                                    MLX4_QP_STATE_RST, NULL, 0, 0, &qp->mqp))
 950                         pr_warn("modify QP %06x to RESET failed.\n",
 951                                qp->mqp.qpn);
 952                 if (qp->pri.smac) {
 953                         mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
 954                         qp->pri.smac = 0;
 955                 }
 956                 if (qp->alt.smac) {
 957                         mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
 958                         qp->alt.smac = 0;
 959                 }
 960                 if (qp->pri.vid < 0x1000) {
 961                         mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
 962                         qp->pri.vid = 0xFFFF;
 963                         qp->pri.candidate_vid = 0xFFFF;
 964                         qp->pri.update_vid = 0;
 965                 }
 966                 if (qp->alt.vid < 0x1000) {
 967                         mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
 968                         qp->alt.vid = 0xFFFF;
 969                         qp->alt.candidate_vid = 0xFFFF;
 970                         qp->alt.update_vid = 0;
 971                 }
 972         }
 973
 974         get_cqs(qp, &send_cq, &recv_cq);
 975
 976         mlx4_ib_lock_cqs(send_cq, recv_cq);
 977
 978         if (!is_user) {
 979                 __mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
 980                                  qp->ibqp.srq ? to_msrq(qp->ibqp.srq): NULL);
 981                 if (send_cq != recv_cq)
 982                         __mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
 983         }
 984
 985         mlx4_qp_remove(dev->dev, &qp->mqp);
 986
 987         mlx4_ib_unlock_cqs(send_cq, recv_cq);
 988
 989         mlx4_qp_free(dev->dev, &qp->mqp);
 990
 991         if (!is_sqp(dev, qp) && !is_tunnel_qp(dev, qp)) {
 992                 if (qp->flags & MLX4_IB_QP_NETIF)
 993                         mlx4_ib_steer_qp_free(dev, qp->mqp.qpn, 1);
 994                 else
 995                         mlx4_qp_release_range(dev->dev, qp->mqp.qpn, 1);
 996         }
 997
 998         mlx4_mtt_cleanup(dev->dev, &qp->mtt);
 999
1000         if (is_user) {
1001                 if (qp->rq.wqe_cnt)
1002                         mlx4_ib_db_unmap_user(to_mucontext(qp->ibqp.uobject->context),
1003                                               &qp->db);
1004                 ib_umem_release(qp->umem);
1005         } else {
1006                 kfree(qp->sq.wrid);
1007                 kfree(qp->rq.wrid);
1008                 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
1009                     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI))
1010                         free_proxy_bufs(&dev->ib_dev, qp);
1011                 mlx4_buf_free(dev->dev, qp->buf_size, &qp->buf);
1012                 if (qp->rq.wqe_cnt)
1013                         mlx4_db_free(dev->dev, &qp->db);
1014         }
1015
1016         del_gid_entries(qp);
1017 }
1018
1019 static u32 get_sqp_num(struct mlx4_ib_dev *dev, struct ib_qp_init_attr *attr)
1020 {
1021         /* Native or PPF */
1022         if (!mlx4_is_mfunc(dev->dev) ||
1023             (mlx4_is_master(dev->dev) &&
1024              attr->create_flags & MLX4_IB_SRIOV_SQP)) {
1025                 return  dev->dev->phys_caps.base_sqpn +
1026                         (attr->qp_type == IB_QPT_SMI ? 0 : 2) +
1027                         attr->port_num - 1;
1028         }
1029         /* PF or VF -- creating proxies */
1030         if (attr->qp_type == IB_QPT_SMI)
1031                 return dev->dev->caps.qp0_proxy[attr->port_num - 1];
1032         else
1033                 return dev->dev->caps.qp1_proxy[attr->port_num - 1];
1034 }
1035
1036 struct ib_qp *mlx4_ib_create_qp(struct ib_pd *pd,
1037                                 struct ib_qp_init_attr *init_attr,
1038                                 struct ib_udata *udata)
1039 {
1040         struct mlx4_ib_qp *qp = NULL;
1041         int err;
1042         u16 xrcdn = 0;
1043         gfp_t gfp;
1044
1045         gfp = (init_attr->create_flags & MLX4_IB_QP_CREATE_USE_GFP_NOIO) ?
1046                 GFP_NOIO : GFP_KERNEL;
1047         /*
1048          * We only support LSO, vendor flag1, and multicast loopback blocking,
1049          * and only for kernel UD QPs.
1050          */
1051         if (init_attr->create_flags & ~(MLX4_IB_QP_LSO |
1052                                         MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK |
1053                                         MLX4_IB_SRIOV_TUNNEL_QP |
1054                                         MLX4_IB_SRIOV_SQP |
1055                                         MLX4_IB_QP_NETIF |
1056                                         MLX4_IB_QP_CREATE_USE_GFP_NOIO))
1057                 return ERR_PTR(-EINVAL);
1058
1059         if (init_attr->create_flags & IB_QP_CREATE_NETIF_QP) {
1060                 if (init_attr->qp_type != IB_QPT_UD)
1061                         return ERR_PTR(-EINVAL);
1062         }
1063
1064         if (init_attr->create_flags &&
1065             (udata ||
1066              ((init_attr->create_flags & ~(MLX4_IB_SRIOV_SQP | MLX4_IB_QP_CREATE_USE_GFP_NOIO)) &&
1067               init_attr->qp_type != IB_QPT_UD) ||
1068              ((init_attr->create_flags & MLX4_IB_SRIOV_SQP) &&
1069               init_attr->qp_type > IB_QPT_GSI)))
1070                 return ERR_PTR(-EINVAL);
1071
1072         switch (init_attr->qp_type) {
1073         case IB_QPT_XRC_TGT:
1074                 pd = to_mxrcd(init_attr->xrcd)->pd;
1075                 xrcdn = to_mxrcd(init_attr->xrcd)->xrcdn;
1076                 init_attr->send_cq = to_mxrcd(init_attr->xrcd)->cq;
1077                 /* fall through */
1078         case IB_QPT_XRC_INI:
1079                 if (!(to_mdev(pd->device)->dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC))
1080                         return ERR_PTR(-ENOSYS);
1081                 init_attr->recv_cq = init_attr->send_cq;
1082                 /* fall through */
1083         case IB_QPT_RC:
1084         case IB_QPT_UC:
1085         case IB_QPT_RAW_PACKET:
1086                 qp = kzalloc(sizeof *qp, gfp);
1087                 if (!qp)
1088                         return ERR_PTR(-ENOMEM);
1089                 qp->pri.vid = 0xFFFF;
1090                 qp->alt.vid = 0xFFFF;
1091                 /* fall through */
1092         case IB_QPT_UD:
1093         {
1094                 err = create_qp_common(to_mdev(pd->device), pd, init_attr,
1095                                        udata, 0, &qp, gfp);
1096                 if (err)
1097                         return ERR_PTR(err);
1098
1099                 qp->ibqp.qp_num = qp->mqp.qpn;
1100                 qp->xrcdn = xrcdn;
1101
1102                 break;
1103         }
1104         case IB_QPT_SMI:
1105         case IB_QPT_GSI:
1106         {
1107                 /* Userspace is not allowed to create special QPs: */
1108                 if (udata)
1109                         return ERR_PTR(-EINVAL);
1110
1111                 err = create_qp_common(to_mdev(pd->device), pd, init_attr, udata,
1112                                        get_sqp_num(to_mdev(pd->device), init_attr),
1113                                        &qp, gfp);
1114                 if (err)
1115                         return ERR_PTR(err);
1116
1117                 qp->port        = init_attr->port_num;
1118                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
1119
1120                 break;
1121         }
1122         default:
1123                 /* Don't support raw QPs */
1124                 return ERR_PTR(-EINVAL);
1125         }
1126
1127         return &qp->ibqp;
1128 }
1129
1130 int mlx4_ib_destroy_qp(struct ib_qp *qp)
1131 {
1132         struct mlx4_ib_dev *dev = to_mdev(qp->device);
1133         struct mlx4_ib_qp *mqp = to_mqp(qp);
1134         struct mlx4_ib_pd *pd;
1135
1136         if (is_qp0(dev, mqp))
1137                 mlx4_CLOSE_PORT(dev->dev, mqp->port);
1138
1139         pd = get_pd(mqp);
1140         destroy_qp_common(dev, mqp, !!pd->ibpd.uobject);
1141
1142         if (is_sqp(dev, mqp))
1143                 kfree(to_msqp(mqp));
1144         else
1145                 kfree(mqp);
1146
1147         return 0;
1148 }
1149
1150 static int to_mlx4_st(struct mlx4_ib_dev *dev, enum mlx4_ib_qp_type type)
1151 {
1152         switch (type) {
1153         case MLX4_IB_QPT_RC:            return MLX4_QP_ST_RC;
1154         case MLX4_IB_QPT_UC:            return MLX4_QP_ST_UC;
1155         case MLX4_IB_QPT_UD:            return MLX4_QP_ST_UD;
1156         case MLX4_IB_QPT_XRC_INI:
1157         case MLX4_IB_QPT_XRC_TGT:       return MLX4_QP_ST_XRC;
1158         case MLX4_IB_QPT_SMI:
1159         case MLX4_IB_QPT_GSI:
1160         case MLX4_IB_QPT_RAW_PACKET:    return MLX4_QP_ST_MLX;
1161
1162         case MLX4_IB_QPT_PROXY_SMI_OWNER:
1163         case MLX4_IB_QPT_TUN_SMI_OWNER: return (mlx4_is_mfunc(dev->dev) ?
1164                                                 MLX4_QP_ST_MLX : -1);
1165         case MLX4_IB_QPT_PROXY_SMI:
1166         case MLX4_IB_QPT_TUN_SMI:
1167         case MLX4_IB_QPT_PROXY_GSI:
1168         case MLX4_IB_QPT_TUN_GSI:       return (mlx4_is_mfunc(dev->dev) ?
1169                                                 MLX4_QP_ST_UD : -1);
1170         default:                        return -1;
1171         }
1172 }
1173
1174 static __be32 to_mlx4_access_flags(struct mlx4_ib_qp *qp, const struct ib_qp_attr *attr,
1175                                    int attr_mask)
1176 {
1177         u8 dest_rd_atomic;
1178         u32 access_flags;
1179         u32 hw_access_flags = 0;
1180
1181         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1182                 dest_rd_atomic = attr->max_dest_rd_atomic;
1183         else
1184                 dest_rd_atomic = qp->resp_depth;
1185
1186         if (attr_mask & IB_QP_ACCESS_FLAGS)
1187                 access_flags = attr->qp_access_flags;
1188         else
1189                 access_flags = qp->atomic_rd_en;
1190
1191         if (!dest_rd_atomic)
1192                 access_flags &= IB_ACCESS_REMOTE_WRITE;
1193
1194         if (access_flags & IB_ACCESS_REMOTE_READ)
1195                 hw_access_flags |= MLX4_QP_BIT_RRE;
1196         if (access_flags & IB_ACCESS_REMOTE_ATOMIC)
1197                 hw_access_flags |= MLX4_QP_BIT_RAE;
1198         if (access_flags & IB_ACCESS_REMOTE_WRITE)
1199                 hw_access_flags |= MLX4_QP_BIT_RWE;
1200
1201         return cpu_to_be32(hw_access_flags);
1202 }
1203
1204 static void store_sqp_attrs(struct mlx4_ib_sqp *sqp, const struct ib_qp_attr *attr,
1205                             int attr_mask)
1206 {
1207         if (attr_mask & IB_QP_PKEY_INDEX)
1208                 sqp->pkey_index = attr->pkey_index;
1209         if (attr_mask & IB_QP_QKEY)
1210                 sqp->qkey = attr->qkey;
1211         if (attr_mask & IB_QP_SQ_PSN)
1212                 sqp->send_psn = attr->sq_psn;
1213 }
1214
1215 static void mlx4_set_sched(struct mlx4_qp_path *path, u8 port)
1216 {
1217         path->sched_queue = (path->sched_queue & 0xbf) | ((port - 1) << 6);
1218 }
1219
1220 static int _mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_ah_attr *ah,
1221                           u64 smac, u16 vlan_tag, struct mlx4_qp_path *path,
1222                           struct mlx4_roce_smac_vlan_info *smac_info, u8 port)
1223 {
1224         int is_eth = rdma_port_get_link_layer(&dev->ib_dev, port) ==
1225                 IB_LINK_LAYER_ETHERNET;
1226         int vidx;
1227         int smac_index;
1228         int err;
1229
1230
1231         path->grh_mylmc     = ah->src_path_bits & 0x7f;
1232         path->rlid          = cpu_to_be16(ah->dlid);
1233         if (ah->static_rate) {
1234                 path->static_rate = ah->static_rate + MLX4_STAT_RATE_OFFSET;
1235                 while (path->static_rate > IB_RATE_2_5_GBPS + MLX4_STAT_RATE_OFFSET &&
1236                        !(1 << path->static_rate & dev->dev->caps.stat_rate_support))
1237                         --path->static_rate;
1238         } else
1239                 path->static_rate = 0;
1240
1241         if (ah->ah_flags & IB_AH_GRH) {
1242                 if (ah->grh.sgid_index >= dev->dev->caps.gid_table_len[port]) {
1243                         pr_err("sgid_index (%u) too large. max is %d\n",
1244                                ah->grh.sgid_index, dev->dev->caps.gid_table_len[port] - 1);
1245                         return -1;
1246                 }
1247
1248                 path->grh_mylmc |= 1 << 7;
1249                 path->mgid_index = ah->grh.sgid_index;
1250                 path->hop_limit  = ah->grh.hop_limit;
1251                 path->tclass_flowlabel =
1252                         cpu_to_be32((ah->grh.traffic_class << 20) |
1253                                     (ah->grh.flow_label));
1254                 memcpy(path->rgid, ah->grh.dgid.raw, 16);
1255         }
1256
1257         if (is_eth) {
1258                 if (!(ah->ah_flags & IB_AH_GRH))
1259                         return -1;
1260
1261                 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1262                         ((port - 1) << 6) | ((ah->sl & 7) << 3);
1263
1264                 path->feup |= MLX4_FEUP_FORCE_ETH_UP;
1265                 if (vlan_tag < 0x1000) {
1266                         if (smac_info->vid < 0x1000) {
1267                                 /* both valid vlan ids */
1268                                 if (smac_info->vid != vlan_tag) {
1269                                         /* different VIDs.  unreg old and reg new */
1270                                         err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1271                                         if (err)
1272                                                 return err;
1273                                         smac_info->candidate_vid = vlan_tag;
1274                                         smac_info->candidate_vlan_index = vidx;
1275                                         smac_info->candidate_vlan_port = port;
1276                                         smac_info->update_vid = 1;
1277                                         path->vlan_index = vidx;
1278                                 } else {
1279                                         path->vlan_index = smac_info->vlan_index;
1280                                 }
1281                         } else {
1282                                 /* no current vlan tag in qp */
1283                                 err = mlx4_register_vlan(dev->dev, port, vlan_tag, &vidx);
1284                                 if (err)
1285                                         return err;
1286                                 smac_info->candidate_vid = vlan_tag;
1287                                 smac_info->candidate_vlan_index = vidx;
1288                                 smac_info->candidate_vlan_port = port;
1289                                 smac_info->update_vid = 1;
1290                                 path->vlan_index = vidx;
1291                         }
1292                         path->feup |= MLX4_FVL_FORCE_ETH_VLAN;
1293                         path->fl = 1 << 6;
1294                 } else {
1295                         /* have current vlan tag. unregister it at modify-qp success */
1296                         if (smac_info->vid < 0x1000) {
1297                                 smac_info->candidate_vid = 0xFFFF;
1298                                 smac_info->update_vid = 1;
1299                         }
1300                 }
1301
1302                 /* get smac_index for RoCE use.
1303                  * If no smac was yet assigned, register one.
1304                  * If one was already assigned, but the new mac differs,
1305                  * unregister the old one and register the new one.
1306                 */
1307                 if (!smac_info->smac || smac_info->smac != smac) {
1308                         /* register candidate now, unreg if needed, after success */
1309                         smac_index = mlx4_register_mac(dev->dev, port, smac);
1310                         if (smac_index >= 0) {
1311                                 smac_info->candidate_smac_index = smac_index;
1312                                 smac_info->candidate_smac = smac;
1313                                 smac_info->candidate_smac_port = port;
1314                         } else {
1315                                 return -EINVAL;
1316                         }
1317                 } else {
1318                         smac_index = smac_info->smac_index;
1319                 }
1320
1321                 memcpy(path->dmac, ah->dmac, 6);
1322                 path->ackto = MLX4_IB_LINK_TYPE_ETH;
1323                 /* put MAC table smac index for IBoE */
1324                 path->grh_mylmc = (u8) (smac_index) | 0x80;
1325         } else {
1326                 path->sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE |
1327                         ((port - 1) << 6) | ((ah->sl & 0xf) << 2);
1328         }
1329
1330         return 0;
1331 }
1332
1333 static int mlx4_set_path(struct mlx4_ib_dev *dev, const struct ib_qp_attr *qp,
1334                          enum ib_qp_attr_mask qp_attr_mask,
1335                          struct mlx4_ib_qp *mqp,
1336                          struct mlx4_qp_path *path, u8 port)
1337 {
1338         return _mlx4_set_path(dev, &qp->ah_attr,
1339                               mlx4_mac_to_u64((u8 *)qp->smac),
1340                               (qp_attr_mask & IB_QP_VID) ? qp->vlan_id : 0xffff,
1341                               path, &mqp->pri, port);
1342 }
1343
1344 static int mlx4_set_alt_path(struct mlx4_ib_dev *dev,
1345                              const struct ib_qp_attr *qp,
1346                              enum ib_qp_attr_mask qp_attr_mask,
1347                              struct mlx4_ib_qp *mqp,
1348                              struct mlx4_qp_path *path, u8 port)
1349 {
1350         return _mlx4_set_path(dev, &qp->alt_ah_attr,
1351                               mlx4_mac_to_u64((u8 *)qp->alt_smac),
1352                               (qp_attr_mask & IB_QP_ALT_VID) ?
1353                               qp->alt_vlan_id : 0xffff,
1354                               path, &mqp->alt, port);
1355 }
1356
1357 static void update_mcg_macs(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp)
1358 {
1359         struct mlx4_ib_gid_entry *ge, *tmp;
1360
1361         list_for_each_entry_safe(ge, tmp, &qp->gid_list, list) {
1362                 if (!ge->added && mlx4_ib_add_mc(dev, qp, &ge->gid)) {
1363                         ge->added = 1;
1364                         ge->port = qp->port;
1365                 }
1366         }
1367 }
1368
1369 static int handle_eth_ud_smac_index(struct mlx4_ib_dev *dev, struct mlx4_ib_qp *qp, u8 *smac,
1370                                     struct mlx4_qp_context *context)
1371 {
1372         struct net_device *ndev;
1373         u64 u64_mac;
1374         int smac_index;
1375
1376
1377         ndev = dev->iboe.netdevs[qp->port - 1];
1378         if (ndev) {
1379                 smac = ndev->dev_addr;
1380                 u64_mac = mlx4_mac_to_u64(smac);
1381         } else {
1382                 u64_mac = dev->dev->caps.def_mac[qp->port];
1383         }
1384
1385         context->pri_path.sched_queue = MLX4_IB_DEFAULT_SCHED_QUEUE | ((qp->port - 1) << 6);
1386         if (!qp->pri.smac) {
1387                 smac_index = mlx4_register_mac(dev->dev, qp->port, u64_mac);
1388                 if (smac_index >= 0) {
1389                         qp->pri.candidate_smac_index = smac_index;
1390                         qp->pri.candidate_smac = u64_mac;
1391                         qp->pri.candidate_smac_port = qp->port;
1392                         context->pri_path.grh_mylmc = 0x80 | (u8) smac_index;
1393                 } else {
1394                         return -ENOENT;
1395                 }
1396         }
1397         return 0;
1398 }
1399
1400 static int __mlx4_ib_modify_qp(struct ib_qp *ibqp,
1401                                const struct ib_qp_attr *attr, int attr_mask,
1402                                enum ib_qp_state cur_state, enum ib_qp_state new_state)
1403 {
1404         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1405         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1406         struct mlx4_ib_pd *pd;
1407         struct mlx4_ib_cq *send_cq, *recv_cq;
1408         struct mlx4_qp_context *context;
1409         enum mlx4_qp_optpar optpar = 0;
1410         int sqd_event;
1411         int steer_qp = 0;
1412         int err = -EINVAL;
1413
1414         context = kzalloc(sizeof *context, GFP_KERNEL);
1415         if (!context)
1416                 return -ENOMEM;
1417
1418         context->flags = cpu_to_be32((to_mlx4_state(new_state) << 28) |
1419                                      (to_mlx4_st(dev, qp->mlx4_ib_qp_type) << 16));
1420
1421         if (!(attr_mask & IB_QP_PATH_MIG_STATE))
1422                 context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1423         else {
1424                 optpar |= MLX4_QP_OPTPAR_PM_STATE;
1425                 switch (attr->path_mig_state) {
1426                 case IB_MIG_MIGRATED:
1427                         context->flags |= cpu_to_be32(MLX4_QP_PM_MIGRATED << 11);
1428                         break;
1429                 case IB_MIG_REARM:
1430                         context->flags |= cpu_to_be32(MLX4_QP_PM_REARM << 11);
1431                         break;
1432                 case IB_MIG_ARMED:
1433                         context->flags |= cpu_to_be32(MLX4_QP_PM_ARMED << 11);
1434                         break;
1435                 }
1436         }
1437
1438         if (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI)
1439                 context->mtu_msgmax = (IB_MTU_4096 << 5) | 11;
1440         else if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1441                 context->mtu_msgmax = (MLX4_RAW_QP_MTU << 5) | MLX4_RAW_QP_MSGMAX;
1442         else if (ibqp->qp_type == IB_QPT_UD) {
1443                 if (qp->flags & MLX4_IB_QP_LSO)
1444                         context->mtu_msgmax = (IB_MTU_4096 << 5) |
1445                                               ilog2(dev->dev->caps.max_gso_sz);
1446                 else
1447                         context->mtu_msgmax = (IB_MTU_4096 << 5) | 12;
1448         } else if (attr_mask & IB_QP_PATH_MTU) {
1449                 if (attr->path_mtu < IB_MTU_256 || attr->path_mtu > IB_MTU_4096) {
1450                         pr_err("path MTU (%u) is invalid\n",
1451                                attr->path_mtu);
1452                         goto out;
1453                 }
1454                 context->mtu_msgmax = (attr->path_mtu << 5) |
1455                         ilog2(dev->dev->caps.max_msg_sz);
1456         }
1457
1458         if (qp->rq.wqe_cnt)
1459                 context->rq_size_stride = ilog2(qp->rq.wqe_cnt) << 3;
1460         context->rq_size_stride |= qp->rq.wqe_shift - 4;
1461
1462         if (qp->sq.wqe_cnt)
1463                 context->sq_size_stride = ilog2(qp->sq.wqe_cnt) << 3;
1464         context->sq_size_stride |= qp->sq.wqe_shift - 4;
1465
1466         if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1467                 context->sq_size_stride |= !!qp->sq_no_prefetch << 7;
1468                 context->xrcd = cpu_to_be32((u32) qp->xrcdn);
1469                 if (ibqp->qp_type == IB_QPT_RAW_PACKET)
1470                         context->param3 |= cpu_to_be32(1 << 30);
1471         }
1472
1473         if (qp->ibqp.uobject)
1474                 context->usr_page = cpu_to_be32(to_mucontext(ibqp->uobject->context)->uar.index);
1475         else
1476                 context->usr_page = cpu_to_be32(dev->priv_uar.index);
1477
1478         if (attr_mask & IB_QP_DEST_QPN)
1479                 context->remote_qpn = cpu_to_be32(attr->dest_qp_num);
1480
1481         if (attr_mask & IB_QP_PORT) {
1482                 if (cur_state == IB_QPS_SQD && new_state == IB_QPS_SQD &&
1483                     !(attr_mask & IB_QP_AV)) {
1484                         mlx4_set_sched(&context->pri_path, attr->port_num);
1485                         optpar |= MLX4_QP_OPTPAR_SCHED_QUEUE;
1486                 }
1487         }
1488
1489         if (cur_state == IB_QPS_INIT && new_state == IB_QPS_RTR) {
1490                 if (dev->counters[qp->port - 1] != -1) {
1491                         context->pri_path.counter_index =
1492                                                 dev->counters[qp->port - 1];
1493                         optpar |= MLX4_QP_OPTPAR_COUNTER_INDEX;
1494                 } else
1495                         context->pri_path.counter_index = 0xff;
1496
1497                 if (qp->flags & MLX4_IB_QP_NETIF) {
1498                         mlx4_ib_steer_qp_reg(dev, qp, 1);
1499                         steer_qp = 1;
1500                 }
1501         }
1502
1503         if (attr_mask & IB_QP_PKEY_INDEX) {
1504                 if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1505                         context->pri_path.disable_pkey_check = 0x40;
1506                 context->pri_path.pkey_index = attr->pkey_index;
1507                 optpar |= MLX4_QP_OPTPAR_PKEY_INDEX;
1508         }
1509
1510         if (attr_mask & IB_QP_AV) {
1511                 if (mlx4_set_path(dev, attr, attr_mask, qp, &context->pri_path,
1512                                   attr_mask & IB_QP_PORT ?
1513                                   attr->port_num : qp->port))
1514                         goto out;
1515
1516                 optpar |= (MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH |
1517                            MLX4_QP_OPTPAR_SCHED_QUEUE);
1518         }
1519
1520         if (attr_mask & IB_QP_TIMEOUT) {
1521                 context->pri_path.ackto |= attr->timeout << 3;
1522                 optpar |= MLX4_QP_OPTPAR_ACK_TIMEOUT;
1523         }
1524
1525         if (attr_mask & IB_QP_ALT_PATH) {
1526                 if (attr->alt_port_num == 0 ||
1527                     attr->alt_port_num > dev->dev->caps.num_ports)
1528                         goto out;
1529
1530                 if (attr->alt_pkey_index >=
1531                     dev->dev->caps.pkey_table_len[attr->alt_port_num])
1532                         goto out;
1533
1534                 if (mlx4_set_alt_path(dev, attr, attr_mask, qp,
1535                                       &context->alt_path,
1536                                       attr->alt_port_num))
1537                         goto out;
1538
1539                 context->alt_path.pkey_index = attr->alt_pkey_index;
1540                 context->alt_path.ackto = attr->alt_timeout << 3;
1541                 optpar |= MLX4_QP_OPTPAR_ALT_ADDR_PATH;
1542         }
1543
1544         pd = get_pd(qp);
1545         get_cqs(qp, &send_cq, &recv_cq);
1546         context->pd       = cpu_to_be32(pd->pdn);
1547         context->cqn_send = cpu_to_be32(send_cq->mcq.cqn);
1548         context->cqn_recv = cpu_to_be32(recv_cq->mcq.cqn);
1549         context->params1  = cpu_to_be32(MLX4_IB_ACK_REQ_FREQ << 28);
1550
1551         /* Set "fast registration enabled" for all kernel QPs */
1552         if (!qp->ibqp.uobject)
1553                 context->params1 |= cpu_to_be32(1 << 11);
1554
1555         if (attr_mask & IB_QP_RNR_RETRY) {
1556                 context->params1 |= cpu_to_be32(attr->rnr_retry << 13);
1557                 optpar |= MLX4_QP_OPTPAR_RNR_RETRY;
1558         }
1559
1560         if (attr_mask & IB_QP_RETRY_CNT) {
1561                 context->params1 |= cpu_to_be32(attr->retry_cnt << 16);
1562                 optpar |= MLX4_QP_OPTPAR_RETRY_COUNT;
1563         }
1564
1565         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) {
1566                 if (attr->max_rd_atomic)
1567                         context->params1 |=
1568                                 cpu_to_be32(fls(attr->max_rd_atomic - 1) << 21);
1569                 optpar |= MLX4_QP_OPTPAR_SRA_MAX;
1570         }
1571
1572         if (attr_mask & IB_QP_SQ_PSN)
1573                 context->next_send_psn = cpu_to_be32(attr->sq_psn);
1574
1575         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) {
1576                 if (attr->max_dest_rd_atomic)
1577                         context->params2 |=
1578                                 cpu_to_be32(fls(attr->max_dest_rd_atomic - 1) << 21);
1579                 optpar |= MLX4_QP_OPTPAR_RRA_MAX;
1580         }
1581
1582         if (attr_mask & (IB_QP_ACCESS_FLAGS | IB_QP_MAX_DEST_RD_ATOMIC)) {
1583                 context->params2 |= to_mlx4_access_flags(qp, attr, attr_mask);
1584                 optpar |= MLX4_QP_OPTPAR_RWE | MLX4_QP_OPTPAR_RRE | MLX4_QP_OPTPAR_RAE;
1585         }
1586
1587         if (ibqp->srq)
1588                 context->params2 |= cpu_to_be32(MLX4_QP_BIT_RIC);
1589
1590         if (attr_mask & IB_QP_MIN_RNR_TIMER) {
1591                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->min_rnr_timer << 24);
1592                 optpar |= MLX4_QP_OPTPAR_RNR_TIMEOUT;
1593         }
1594         if (attr_mask & IB_QP_RQ_PSN)
1595                 context->rnr_nextrecvpsn |= cpu_to_be32(attr->rq_psn);
1596
1597         /* proxy and tunnel qp qkeys will be changed in modify-qp wrappers */
1598         if (attr_mask & IB_QP_QKEY) {
1599                 if (qp->mlx4_ib_qp_type &
1600                     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))
1601                         context->qkey = cpu_to_be32(IB_QP_SET_QKEY);
1602                 else {
1603                         if (mlx4_is_mfunc(dev->dev) &&
1604                             !(qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV) &&
1605                             (attr->qkey & MLX4_RESERVED_QKEY_MASK) ==
1606                             MLX4_RESERVED_QKEY_BASE) {
1607                                 pr_err("Cannot use reserved QKEY"
1608                                        " 0x%x (range 0xffff0000..0xffffffff"
1609                                        " is reserved)\n", attr->qkey);
1610                                 err = -EINVAL;
1611                                 goto out;
1612                         }
1613                         context->qkey = cpu_to_be32(attr->qkey);
1614                 }
1615                 optpar |= MLX4_QP_OPTPAR_Q_KEY;
1616         }
1617
1618         if (ibqp->srq)
1619                 context->srqn = cpu_to_be32(1 << 24 | to_msrq(ibqp->srq)->msrq.srqn);
1620
1621         if (qp->rq.wqe_cnt && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1622                 context->db_rec_addr = cpu_to_be64(qp->db.dma);
1623
1624         if (cur_state == IB_QPS_INIT &&
1625             new_state == IB_QPS_RTR  &&
1626             (ibqp->qp_type == IB_QPT_GSI || ibqp->qp_type == IB_QPT_SMI ||
1627              ibqp->qp_type == IB_QPT_UD ||
1628              ibqp->qp_type == IB_QPT_RAW_PACKET)) {
1629                 context->pri_path.sched_queue = (qp->port - 1) << 6;
1630                 if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
1631                     qp->mlx4_ib_qp_type &
1632                     (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER)) {
1633                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_QP0_SCHED_QUEUE;
1634                         if (qp->mlx4_ib_qp_type != MLX4_IB_QPT_SMI)
1635                                 context->pri_path.fl = 0x80;
1636                 } else {
1637                         if (qp->mlx4_ib_qp_type & MLX4_IB_QPT_ANY_SRIOV)
1638                                 context->pri_path.fl = 0x80;
1639                         context->pri_path.sched_queue |= MLX4_IB_DEFAULT_SCHED_QUEUE;
1640                 }
1641                 if (rdma_port_get_link_layer(&dev->ib_dev, qp->port) ==
1642                     IB_LINK_LAYER_ETHERNET) {
1643                         if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI ||
1644                             qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI)
1645                                 context->pri_path.feup = 1 << 7; /* don't fsm */
1646                         /* handle smac_index */
1647                         if (qp->mlx4_ib_qp_type == MLX4_IB_QPT_UD ||
1648                             qp->mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_GSI ||
1649                             qp->mlx4_ib_qp_type == MLX4_IB_QPT_TUN_GSI) {
1650                                 err = handle_eth_ud_smac_index(dev, qp, (u8 *)attr->smac, context);
1651                                 if (err)
1652                                         return -EINVAL;
1653                         }
1654                 }
1655         }
1656
1657         if (qp->ibqp.qp_type == IB_QPT_RAW_PACKET)
1658                 context->pri_path.ackto = (context->pri_path.ackto & 0xf8) |
1659                                         MLX4_IB_LINK_TYPE_ETH;
1660
1661         if (ibqp->qp_type == IB_QPT_UD && (new_state == IB_QPS_RTR)) {
1662                 int is_eth = rdma_port_get_link_layer(
1663                                 &dev->ib_dev, qp->port) ==
1664                                 IB_LINK_LAYER_ETHERNET;
1665                 if (is_eth) {
1666                         context->pri_path.ackto = MLX4_IB_LINK_TYPE_ETH;
1667                         optpar |= MLX4_QP_OPTPAR_PRIMARY_ADDR_PATH;
1668                 }
1669         }
1670
1671
1672         if (cur_state == IB_QPS_RTS && new_state == IB_QPS_SQD  &&
1673             attr_mask & IB_QP_EN_SQD_ASYNC_NOTIFY && attr->en_sqd_async_notify)
1674                 sqd_event = 1;
1675         else
1676                 sqd_event = 0;
1677
1678         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1679                 context->rlkey |= (1 << 4);
1680
1681         /*
1682          * Before passing a kernel QP to the HW, make sure that the
1683          * ownership bits of the send queue are set and the SQ
1684          * headroom is stamped so that the hardware doesn't start
1685          * processing stale work requests.
1686          */
1687         if (!ibqp->uobject && cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) {
1688                 struct mlx4_wqe_ctrl_seg *ctrl;
1689                 int i;
1690
1691                 for (i = 0; i < qp->sq.wqe_cnt; ++i) {
1692                         ctrl = get_send_wqe(qp, i);
1693                         ctrl->owner_opcode = cpu_to_be32(1 << 31);
1694                         if (qp->sq_max_wqes_per_wr == 1)
1695                                 ctrl->fence_size = 1 << (qp->sq.wqe_shift - 4);
1696
1697                         stamp_send_wqe(qp, i, 1 << qp->sq.wqe_shift);
1698                 }
1699         }
1700
1701         err = mlx4_qp_modify(dev->dev, &qp->mtt, to_mlx4_state(cur_state),
1702                              to_mlx4_state(new_state), context, optpar,
1703                              sqd_event, &qp->mqp);
1704         if (err)
1705                 goto out;
1706
1707         qp->state = new_state;
1708
1709         if (attr_mask & IB_QP_ACCESS_FLAGS)
1710                 qp->atomic_rd_en = attr->qp_access_flags;
1711         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1712                 qp->resp_depth = attr->max_dest_rd_atomic;
1713         if (attr_mask & IB_QP_PORT) {
1714                 qp->port = attr->port_num;
1715                 update_mcg_macs(dev, qp);
1716         }
1717         if (attr_mask & IB_QP_ALT_PATH)
1718                 qp->alt_port = attr->alt_port_num;
1719
1720         if (is_sqp(dev, qp))
1721                 store_sqp_attrs(to_msqp(qp), attr, attr_mask);
1722
1723         /*
1724          * If we moved QP0 to RTR, bring the IB link up; if we moved
1725          * QP0 to RESET or ERROR, bring the link back down.
1726          */
1727         if (is_qp0(dev, qp)) {
1728                 if (cur_state != IB_QPS_RTR && new_state == IB_QPS_RTR)
1729                         if (mlx4_INIT_PORT(dev->dev, qp->port))
1730                                 pr_warn("INIT_PORT failed for port %d\n",
1731                                        qp->port);
1732
1733                 if (cur_state != IB_QPS_RESET && cur_state != IB_QPS_ERR &&
1734                     (new_state == IB_QPS_RESET || new_state == IB_QPS_ERR))
1735                         mlx4_CLOSE_PORT(dev->dev, qp->port);
1736         }
1737
1738         /*
1739          * If we moved a kernel QP to RESET, clean up all old CQ
1740          * entries and reinitialize the QP.
1741          */
1742         if (new_state == IB_QPS_RESET) {
1743                 if (!ibqp->uobject) {
1744                         mlx4_ib_cq_clean(recv_cq, qp->mqp.qpn,
1745                                          ibqp->srq ? to_msrq(ibqp->srq) : NULL);
1746                         if (send_cq != recv_cq)
1747                                 mlx4_ib_cq_clean(send_cq, qp->mqp.qpn, NULL);
1748
1749                         qp->rq.head = 0;
1750                         qp->rq.tail = 0;
1751                         qp->sq.head = 0;
1752                         qp->sq.tail = 0;
1753                         qp->sq_next_wqe = 0;
1754                         if (qp->rq.wqe_cnt)
1755                                 *qp->db.db  = 0;
1756
1757                         if (qp->flags & MLX4_IB_QP_NETIF)
1758                                 mlx4_ib_steer_qp_reg(dev, qp, 0);
1759                 }
1760                 if (qp->pri.smac) {
1761                         mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1762                         qp->pri.smac = 0;
1763                 }
1764                 if (qp->alt.smac) {
1765                         mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1766                         qp->alt.smac = 0;
1767                 }
1768                 if (qp->pri.vid < 0x1000) {
1769                         mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port, qp->pri.vid);
1770                         qp->pri.vid = 0xFFFF;
1771                         qp->pri.candidate_vid = 0xFFFF;
1772                         qp->pri.update_vid = 0;
1773                 }
1774
1775                 if (qp->alt.vid < 0x1000) {
1776                         mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port, qp->alt.vid);
1777                         qp->alt.vid = 0xFFFF;
1778                         qp->alt.candidate_vid = 0xFFFF;
1779                         qp->alt.update_vid = 0;
1780                 }
1781         }
1782 out:
1783         if (err && steer_qp)
1784                 mlx4_ib_steer_qp_reg(dev, qp, 0);
1785         kfree(context);
1786         if (qp->pri.candidate_smac) {
1787                 if (err) {
1788                         mlx4_unregister_mac(dev->dev, qp->pri.candidate_smac_port, qp->pri.candidate_smac);
1789                 } else {
1790                         if (qp->pri.smac)
1791                                 mlx4_unregister_mac(dev->dev, qp->pri.smac_port, qp->pri.smac);
1792                         qp->pri.smac = qp->pri.candidate_smac;
1793                         qp->pri.smac_index = qp->pri.candidate_smac_index;
1794                         qp->pri.smac_port = qp->pri.candidate_smac_port;
1795                 }
1796                 qp->pri.candidate_smac = 0;
1797                 qp->pri.candidate_smac_index = 0;
1798                 qp->pri.candidate_smac_port = 0;
1799         }
1800         if (qp->alt.candidate_smac) {
1801                 if (err) {
1802                         mlx4_unregister_mac(dev->dev, qp->alt.candidate_smac_port, qp->alt.candidate_smac);
1803                 } else {
1804                         if (qp->alt.smac)
1805                                 mlx4_unregister_mac(dev->dev, qp->alt.smac_port, qp->alt.smac);
1806                         qp->alt.smac = qp->alt.candidate_smac;
1807                         qp->alt.smac_index = qp->alt.candidate_smac_index;
1808                         qp->alt.smac_port = qp->alt.candidate_smac_port;
1809                 }
1810                 qp->alt.candidate_smac = 0;
1811                 qp->alt.candidate_smac_index = 0;
1812                 qp->alt.candidate_smac_port = 0;
1813         }
1814
1815         if (qp->pri.update_vid) {
1816                 if (err) {
1817                         if (qp->pri.candidate_vid < 0x1000)
1818                                 mlx4_unregister_vlan(dev->dev, qp->pri.candidate_vlan_port,
1819                                                      qp->pri.candidate_vid);
1820                 } else {
1821                         if (qp->pri.vid < 0x1000)
1822                                 mlx4_unregister_vlan(dev->dev, qp->pri.vlan_port,
1823                                                      qp->pri.vid);
1824                         qp->pri.vid = qp->pri.candidate_vid;
1825                         qp->pri.vlan_port = qp->pri.candidate_vlan_port;
1826                         qp->pri.vlan_index =  qp->pri.candidate_vlan_index;
1827                 }
1828                 qp->pri.candidate_vid = 0xFFFF;
1829                 qp->pri.update_vid = 0;
1830         }
1831
1832         if (qp->alt.update_vid) {
1833                 if (err) {
1834                         if (qp->alt.candidate_vid < 0x1000)
1835                                 mlx4_unregister_vlan(dev->dev, qp->alt.candidate_vlan_port,
1836                                                      qp->alt.candidate_vid);
1837                 } else {
1838                         if (qp->alt.vid < 0x1000)
1839                                 mlx4_unregister_vlan(dev->dev, qp->alt.vlan_port,
1840                                                      qp->alt.vid);
1841                         qp->alt.vid = qp->alt.candidate_vid;
1842                         qp->alt.vlan_port = qp->alt.candidate_vlan_port;
1843                         qp->alt.vlan_index =  qp->alt.candidate_vlan_index;
1844                 }
1845                 qp->alt.candidate_vid = 0xFFFF;
1846                 qp->alt.update_vid = 0;
1847         }
1848
1849         return err;
1850 }
1851
1852 int mlx4_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1853                       int attr_mask, struct ib_udata *udata)
1854 {
1855         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
1856         struct mlx4_ib_qp *qp = to_mqp(ibqp);
1857         enum ib_qp_state cur_state, new_state;
1858         int err = -EINVAL;
1859         int ll;
1860         mutex_lock(&qp->mutex);
1861
1862         cur_state = attr_mask & IB_QP_CUR_STATE ? attr->cur_qp_state : qp->state;
1863         new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1864
1865         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1866                 ll = IB_LINK_LAYER_UNSPECIFIED;
1867         } else {
1868                 int port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1869                 ll = rdma_port_get_link_layer(&dev->ib_dev, port);
1870         }
1871
1872         if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1873                                 attr_mask, ll)) {
1874                 pr_debug("qpn 0x%x: invalid attribute mask specified "
1875                          "for transition %d to %d. qp_type %d,"
1876                          " attr_mask 0x%x\n",
1877                          ibqp->qp_num, cur_state, new_state,
1878                          ibqp->qp_type, attr_mask);
1879                 goto out;
1880         }
1881
1882         if ((attr_mask & IB_QP_PORT) &&
1883             (attr->port_num == 0 || attr->port_num > dev->num_ports)) {
1884                 pr_debug("qpn 0x%x: invalid port number (%d) specified "
1885                          "for transition %d to %d. qp_type %d\n",
1886                          ibqp->qp_num, attr->port_num, cur_state,
1887                          new_state, ibqp->qp_type);
1888                 goto out;
1889         }
1890
1891         if ((attr_mask & IB_QP_PORT) && (ibqp->qp_type == IB_QPT_RAW_PACKET) &&
1892             (rdma_port_get_link_layer(&dev->ib_dev, attr->port_num) !=
1893              IB_LINK_LAYER_ETHERNET))
1894                 goto out;
1895
1896         if (attr_mask & IB_QP_PKEY_INDEX) {
1897                 int p = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1898                 if (attr->pkey_index >= dev->dev->caps.pkey_table_len[p]) {
1899                         pr_debug("qpn 0x%x: invalid pkey index (%d) specified "
1900                                  "for transition %d to %d. qp_type %d\n",
1901                                  ibqp->qp_num, attr->pkey_index, cur_state,
1902                                  new_state, ibqp->qp_type);
1903                         goto out;
1904                 }
1905         }
1906
1907         if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC &&
1908             attr->max_rd_atomic > dev->dev->caps.max_qp_init_rdma) {
1909                 pr_debug("qpn 0x%x: max_rd_atomic (%d) too large. "
1910                          "Transition %d to %d. qp_type %d\n",
1911                          ibqp->qp_num, attr->max_rd_atomic, cur_state,
1912                          new_state, ibqp->qp_type);
1913                 goto out;
1914         }
1915
1916         if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC &&
1917             attr->max_dest_rd_atomic > dev->dev->caps.max_qp_dest_rdma) {
1918                 pr_debug("qpn 0x%x: max_dest_rd_atomic (%d) too large. "
1919                          "Transition %d to %d. qp_type %d\n",
1920                          ibqp->qp_num, attr->max_dest_rd_atomic, cur_state,
1921                          new_state, ibqp->qp_type);
1922                 goto out;
1923         }
1924
1925         if (cur_state == new_state && cur_state == IB_QPS_RESET) {
1926                 err = 0;
1927                 goto out;
1928         }
1929
1930         err = __mlx4_ib_modify_qp(ibqp, attr, attr_mask, cur_state, new_state);
1931
1932 out:
1933         mutex_unlock(&qp->mutex);
1934         return err;
1935 }
1936
1937 static int build_sriov_qp0_header(struct mlx4_ib_sqp *sqp,
1938                                   struct ib_send_wr *wr,
1939                                   void *wqe, unsigned *mlx_seg_len)
1940 {
1941         struct mlx4_ib_dev *mdev = to_mdev(sqp->qp.ibqp.device);
1942         struct ib_device *ib_dev = &mdev->ib_dev;
1943         struct mlx4_wqe_mlx_seg *mlx = wqe;
1944         struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
1945         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
1946         u16 pkey;
1947         u32 qkey;
1948         int send_size;
1949         int header_size;
1950         int spc;
1951         int i;
1952
1953         if (wr->opcode != IB_WR_SEND)
1954                 return -EINVAL;
1955
1956         send_size = 0;
1957
1958         for (i = 0; i < wr->num_sge; ++i)
1959                 send_size += wr->sg_list[i].length;
1960
1961         /* for proxy-qp0 sends, need to add in size of tunnel header */
1962         /* for tunnel-qp0 sends, tunnel header is already in s/g list */
1963         if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER)
1964                 send_size += sizeof (struct mlx4_ib_tunnel_header);
1965
1966         ib_ud_header_init(send_size, 1, 0, 0, 0, 0, &sqp->ud_header);
1967
1968         if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_PROXY_SMI_OWNER) {
1969                 sqp->ud_header.lrh.service_level =
1970                         be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
1971                 sqp->ud_header.lrh.destination_lid =
1972                         cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1973                 sqp->ud_header.lrh.source_lid =
1974                         cpu_to_be16(ah->av.ib.g_slid & 0x7f);
1975         }
1976
1977         mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
1978
1979         /* force loopback */
1980         mlx->flags |= cpu_to_be32(MLX4_WQE_MLX_VL15 | 0x1 | MLX4_WQE_MLX_SLR);
1981         mlx->rlid = sqp->ud_header.lrh.destination_lid;
1982
1983         sqp->ud_header.lrh.virtual_lane    = 0;
1984         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
1985         ib_get_cached_pkey(ib_dev, sqp->qp.port, 0, &pkey);
1986         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
1987         if (sqp->qp.mlx4_ib_qp_type == MLX4_IB_QPT_TUN_SMI_OWNER)
1988                 sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
1989         else
1990                 sqp->ud_header.bth.destination_qpn =
1991                         cpu_to_be32(mdev->dev->caps.qp0_tunnel[sqp->qp.port - 1]);
1992
1993         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
1994         if (mlx4_get_parav_qkey(mdev->dev, sqp->qp.mqp.qpn, &qkey))
1995                 return -EINVAL;
1996         sqp->ud_header.deth.qkey = cpu_to_be32(qkey);
1997         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.mqp.qpn);
1998
1999         sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2000         sqp->ud_header.immediate_present = 0;
2001
2002         header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2003
2004         /*
2005          * Inline data segments may not cross a 64 byte boundary.  If
2006          * our UD header is bigger than the space available up to the
2007          * next 64 byte boundary in the WQE, use two inline data
2008          * segments to hold the UD header.
2009          */
2010         spc = MLX4_INLINE_ALIGN -
2011               ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2012         if (header_size <= spc) {
2013                 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2014                 memcpy(inl + 1, sqp->header_buf, header_size);
2015                 i = 1;
2016         } else {
2017                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2018                 memcpy(inl + 1, sqp->header_buf, spc);
2019
2020                 inl = (void *) (inl + 1) + spc;
2021                 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2022                 /*
2023                  * Need a barrier here to make sure all the data is
2024                  * visible before the byte_count field is set.
2025                  * Otherwise the HCA prefetcher could grab the 64-byte
2026                  * chunk with this inline segment and get a valid (!=
2027                  * 0xffffffff) byte count but stale data, and end up
2028                  * generating a packet with bad headers.
2029                  *
2030                  * The first inline segment's byte_count field doesn't
2031                  * need a barrier, because it comes after a
2032                  * control/MLX segment and therefore is at an offset
2033                  * of 16 mod 64.
2034                  */
2035                 wmb();
2036                 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2037                 i = 2;
2038         }
2039
2040         *mlx_seg_len =
2041         ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2042         return 0;
2043 }
2044
2045 static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
2046                             void *wqe, unsigned *mlx_seg_len)
2047 {
2048         struct ib_device *ib_dev = sqp->qp.ibqp.device;
2049         struct mlx4_wqe_mlx_seg *mlx = wqe;
2050         struct mlx4_wqe_ctrl_seg *ctrl = wqe;
2051         struct mlx4_wqe_inline_seg *inl = wqe + sizeof *mlx;
2052         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2053         union ib_gid sgid;
2054         u16 pkey;
2055         int send_size;
2056         int header_size;
2057         int spc;
2058         int i;
2059         int err = 0;
2060         u16 vlan = 0xffff;
2061         bool is_eth;
2062         bool is_vlan = false;
2063         bool is_grh;
2064
2065         send_size = 0;
2066         for (i = 0; i < wr->num_sge; ++i)
2067                 send_size += wr->sg_list[i].length;
2068
2069         is_eth = rdma_port_get_link_layer(sqp->qp.ibqp.device, sqp->qp.port) == IB_LINK_LAYER_ETHERNET;
2070         is_grh = mlx4_ib_ah_grh_present(ah);
2071         if (is_eth) {
2072                 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2073                         /* When multi-function is enabled, the ib_core gid
2074                          * indexes don't necessarily match the hw ones, so
2075                          * we must use our own cache */
2076                         err = mlx4_get_roce_gid_from_slave(to_mdev(ib_dev)->dev,
2077                                                            be32_to_cpu(ah->av.ib.port_pd) >> 24,
2078                                                            ah->av.ib.gid_index, &sgid.raw[0]);
2079                         if (err)
2080                                 return err;
2081                 } else  {
2082                         err = ib_get_cached_gid(ib_dev,
2083                                                 be32_to_cpu(ah->av.ib.port_pd) >> 24,
2084                                                 ah->av.ib.gid_index, &sgid);
2085                         if (err)
2086                                 return err;
2087                 }
2088
2089                 if (ah->av.eth.vlan != cpu_to_be16(0xffff)) {
2090                         vlan = be16_to_cpu(ah->av.eth.vlan) & 0x0fff;
2091                         is_vlan = 1;
2092                 }
2093         }
2094         ib_ud_header_init(send_size, !is_eth, is_eth, is_vlan, is_grh, 0, &sqp->ud_header);
2095
2096         if (!is_eth) {
2097                 sqp->ud_header.lrh.service_level =
2098                         be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 28;
2099                 sqp->ud_header.lrh.destination_lid = ah->av.ib.dlid;
2100                 sqp->ud_header.lrh.source_lid = cpu_to_be16(ah->av.ib.g_slid & 0x7f);
2101         }
2102
2103         if (is_grh) {
2104                 sqp->ud_header.grh.traffic_class =
2105                         (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 20) & 0xff;
2106                 sqp->ud_header.grh.flow_label    =
2107                         ah->av.ib.sl_tclass_flowlabel & cpu_to_be32(0xfffff);
2108                 sqp->ud_header.grh.hop_limit     = ah->av.ib.hop_limit;
2109                 if (is_eth)
2110                         memcpy(sqp->ud_header.grh.source_gid.raw, sgid.raw, 16);
2111                 else {
2112                 if (mlx4_is_mfunc(to_mdev(ib_dev)->dev)) {
2113                         /* When multi-function is enabled, the ib_core gid
2114                          * indexes don't necessarily match the hw ones, so
2115                          * we must use our own cache */
2116                         sqp->ud_header.grh.source_gid.global.subnet_prefix =
2117                                 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2118                                                        subnet_prefix;
2119                         sqp->ud_header.grh.source_gid.global.interface_id =
2120                                 to_mdev(ib_dev)->sriov.demux[sqp->qp.port - 1].
2121                                                guid_cache[ah->av.ib.gid_index];
2122                 } else
2123                         ib_get_cached_gid(ib_dev,
2124                                           be32_to_cpu(ah->av.ib.port_pd) >> 24,
2125                                           ah->av.ib.gid_index,
2126                                           &sqp->ud_header.grh.source_gid);
2127                 }
2128                 memcpy(sqp->ud_header.grh.destination_gid.raw,
2129                        ah->av.ib.dgid, 16);
2130         }
2131
2132         mlx->flags &= cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE);
2133
2134         if (!is_eth) {
2135                 mlx->flags |= cpu_to_be32((!sqp->qp.ibqp.qp_num ? MLX4_WQE_MLX_VL15 : 0) |
2136                                           (sqp->ud_header.lrh.destination_lid ==
2137                                            IB_LID_PERMISSIVE ? MLX4_WQE_MLX_SLR : 0) |
2138                                           (sqp->ud_header.lrh.service_level << 8));
2139                 if (ah->av.ib.port_pd & cpu_to_be32(0x80000000))
2140                         mlx->flags |= cpu_to_be32(0x1); /* force loopback */
2141                 mlx->rlid = sqp->ud_header.lrh.destination_lid;
2142         }
2143
2144         switch (wr->opcode) {
2145         case IB_WR_SEND:
2146                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY;
2147                 sqp->ud_header.immediate_present = 0;
2148                 break;
2149         case IB_WR_SEND_WITH_IMM:
2150                 sqp->ud_header.bth.opcode        = IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
2151                 sqp->ud_header.immediate_present = 1;
2152                 sqp->ud_header.immediate_data    = wr->ex.imm_data;
2153                 break;
2154         default:
2155                 return -EINVAL;
2156         }
2157
2158         if (is_eth) {
2159                 u8 *smac;
2160                 struct in6_addr in6;
2161
2162                 u16 pcp = (be32_to_cpu(ah->av.ib.sl_tclass_flowlabel) >> 29) << 13;
2163
2164                 mlx->sched_prio = cpu_to_be16(pcp);
2165
2166                 memcpy(sqp->ud_header.eth.dmac_h, ah->av.eth.mac, 6);
2167                 /* FIXME: cache smac value? */
2168                 memcpy(&ctrl->srcrb_flags16[0], ah->av.eth.mac, 2);
2169                 memcpy(&ctrl->imm, ah->av.eth.mac + 2, 4);
2170                 memcpy(&in6, sgid.raw, sizeof(in6));
2171
2172                 if (!mlx4_is_mfunc(to_mdev(ib_dev)->dev))
2173                         smac = to_mdev(sqp->qp.ibqp.device)->
2174                                 iboe.netdevs[sqp->qp.port - 1]->dev_addr;
2175                 else    /* use the src mac of the tunnel */
2176                         smac = ah->av.eth.s_mac;
2177                 memcpy(sqp->ud_header.eth.smac_h, smac, 6);
2178                 if (!memcmp(sqp->ud_header.eth.smac_h, sqp->ud_header.eth.dmac_h, 6))
2179                         mlx->flags |= cpu_to_be32(MLX4_WQE_CTRL_FORCE_LOOPBACK);
2180                 if (!is_vlan) {
2181                         sqp->ud_header.eth.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2182                 } else {
2183                         sqp->ud_header.vlan.type = cpu_to_be16(MLX4_IB_IBOE_ETHERTYPE);
2184                         sqp->ud_header.vlan.tag = cpu_to_be16(vlan | pcp);
2185                 }
2186         } else {
2187                 sqp->ud_header.lrh.virtual_lane    = !sqp->qp.ibqp.qp_num ? 15 : 0;
2188                 if (sqp->ud_header.lrh.destination_lid == IB_LID_PERMISSIVE)
2189                         sqp->ud_header.lrh.source_lid = IB_LID_PERMISSIVE;
2190         }
2191         sqp->ud_header.bth.solicited_event = !!(wr->send_flags & IB_SEND_SOLICITED);
2192         if (!sqp->qp.ibqp.qp_num)
2193                 ib_get_cached_pkey(ib_dev, sqp->qp.port, sqp->pkey_index, &pkey);
2194         else
2195                 ib_get_cached_pkey(ib_dev, sqp->qp.port, wr->wr.ud.pkey_index, &pkey);
2196         sqp->ud_header.bth.pkey = cpu_to_be16(pkey);
2197         sqp->ud_header.bth.destination_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2198         sqp->ud_header.bth.psn = cpu_to_be32((sqp->send_psn++) & ((1 << 24) - 1));
2199         sqp->ud_header.deth.qkey = cpu_to_be32(wr->wr.ud.remote_qkey & 0x80000000 ?
2200                                                sqp->qkey : wr->wr.ud.remote_qkey);
2201         sqp->ud_header.deth.source_qpn = cpu_to_be32(sqp->qp.ibqp.qp_num);
2202
2203         header_size = ib_ud_header_pack(&sqp->ud_header, sqp->header_buf);
2204
2205         if (0) {
2206                 pr_err("built UD header of size %d:\n", header_size);
2207                 for (i = 0; i < header_size / 4; ++i) {
2208                         if (i % 8 == 0)
2209                                 pr_err("  [%02x] ", i * 4);
2210                         pr_cont(" %08x",
2211                                 be32_to_cpu(((__be32 *) sqp->header_buf)[i]));
2212                         if ((i + 1) % 8 == 0)
2213                                 pr_cont("\n");
2214                 }
2215                 pr_err("\n");
2216         }
2217
2218         /*
2219          * Inline data segments may not cross a 64 byte boundary.  If
2220          * our UD header is bigger than the space available up to the
2221          * next 64 byte boundary in the WQE, use two inline data
2222          * segments to hold the UD header.
2223          */
2224         spc = MLX4_INLINE_ALIGN -
2225                 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2226         if (header_size <= spc) {
2227                 inl->byte_count = cpu_to_be32(1 << 31 | header_size);
2228                 memcpy(inl + 1, sqp->header_buf, header_size);
2229                 i = 1;
2230         } else {
2231                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2232                 memcpy(inl + 1, sqp->header_buf, spc);
2233
2234                 inl = (void *) (inl + 1) + spc;
2235                 memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
2236                 /*
2237                  * Need a barrier here to make sure all the data is
2238                  * visible before the byte_count field is set.
2239                  * Otherwise the HCA prefetcher could grab the 64-byte
2240                  * chunk with this inline segment and get a valid (!=
2241                  * 0xffffffff) byte count but stale data, and end up
2242                  * generating a packet with bad headers.
2243                  *
2244                  * The first inline segment's byte_count field doesn't
2245                  * need a barrier, because it comes after a
2246                  * control/MLX segment and therefore is at an offset
2247                  * of 16 mod 64.
2248                  */
2249                 wmb();
2250                 inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
2251                 i = 2;
2252         }
2253
2254         *mlx_seg_len =
2255                 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
2256         return 0;
2257 }
2258
2259 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
2260 {
2261         unsigned cur;
2262         struct mlx4_ib_cq *cq;
2263
2264         cur = wq->head - wq->tail;
2265         if (likely(cur + nreq < wq->max_post))
2266                 return 0;
2267
2268         cq = to_mcq(ib_cq);
2269         spin_lock(&cq->lock);
2270         cur = wq->head - wq->tail;
2271         spin_unlock(&cq->lock);
2272
2273         return cur + nreq >= wq->max_post;
2274 }
2275
2276 static __be32 convert_access(int acc)
2277 {
2278         return (acc & IB_ACCESS_REMOTE_ATOMIC ?
2279                 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC)       : 0) |
2280                (acc & IB_ACCESS_REMOTE_WRITE  ?
2281                 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE) : 0) |
2282                (acc & IB_ACCESS_REMOTE_READ   ?
2283                 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ)  : 0) |
2284                (acc & IB_ACCESS_LOCAL_WRITE   ? cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_WRITE)  : 0) |
2285                 cpu_to_be32(MLX4_WQE_FMR_PERM_LOCAL_READ);
2286 }
2287
2288 static void set_fmr_seg(struct mlx4_wqe_fmr_seg *fseg, struct ib_send_wr *wr)
2289 {
2290         struct mlx4_ib_fast_reg_page_list *mfrpl = to_mfrpl(wr->wr.fast_reg.page_list);
2291         int i;
2292
2293         for (i = 0; i < wr->wr.fast_reg.page_list_len; ++i)
2294                 mfrpl->mapped_page_list[i] =
2295                         cpu_to_be64(wr->wr.fast_reg.page_list->page_list[i] |
2296                                     MLX4_MTT_FLAG_PRESENT);
2297
2298         fseg->flags             = convert_access(wr->wr.fast_reg.access_flags);
2299         fseg->mem_key           = cpu_to_be32(wr->wr.fast_reg.rkey);
2300         fseg->buf_list          = cpu_to_be64(mfrpl->map);
2301         fseg->start_addr        = cpu_to_be64(wr->wr.fast_reg.iova_start);
2302         fseg->reg_len           = cpu_to_be64(wr->wr.fast_reg.length);
2303         fseg->offset            = 0; /* XXX -- is this just for ZBVA? */
2304         fseg->page_size         = cpu_to_be32(wr->wr.fast_reg.page_shift);
2305         fseg->reserved[0]       = 0;
2306         fseg->reserved[1]       = 0;
2307 }
2308
2309 static void set_bind_seg(struct mlx4_wqe_bind_seg *bseg, struct ib_send_wr *wr)
2310 {
2311         bseg->flags1 =
2312                 convert_access(wr->wr.bind_mw.bind_info.mw_access_flags) &
2313                 cpu_to_be32(MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_READ  |
2314                             MLX4_WQE_FMR_AND_BIND_PERM_REMOTE_WRITE |
2315                             MLX4_WQE_FMR_AND_BIND_PERM_ATOMIC);
2316         bseg->flags2 = 0;
2317         if (wr->wr.bind_mw.mw->type == IB_MW_TYPE_2)
2318                 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_TYPE_2);
2319         if (wr->wr.bind_mw.bind_info.mw_access_flags & IB_ZERO_BASED)
2320                 bseg->flags2 |= cpu_to_be32(MLX4_WQE_BIND_ZERO_BASED);
2321         bseg->new_rkey = cpu_to_be32(wr->wr.bind_mw.rkey);
2322         bseg->lkey = cpu_to_be32(wr->wr.bind_mw.bind_info.mr->lkey);
2323         bseg->addr = cpu_to_be64(wr->wr.bind_mw.bind_info.addr);
2324         bseg->length = cpu_to_be64(wr->wr.bind_mw.bind_info.length);
2325 }
2326
2327 static void set_local_inv_seg(struct mlx4_wqe_local_inval_seg *iseg, u32 rkey)
2328 {
2329         memset(iseg, 0, sizeof(*iseg));
2330         iseg->mem_key = cpu_to_be32(rkey);
2331 }
2332
2333 static __always_inline void set_raddr_seg(struct mlx4_wqe_raddr_seg *rseg,
2334                                           u64 remote_addr, u32 rkey)
2335 {
2336         rseg->raddr    = cpu_to_be64(remote_addr);
2337         rseg->rkey     = cpu_to_be32(rkey);
2338         rseg->reserved = 0;
2339 }
2340
2341 static void set_atomic_seg(struct mlx4_wqe_atomic_seg *aseg, struct ib_send_wr *wr)
2342 {
2343         if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
2344                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.swap);
2345                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add);
2346         } else if (wr->opcode == IB_WR_MASKED_ATOMIC_FETCH_AND_ADD) {
2347                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2348                 aseg->compare  = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2349         } else {
2350                 aseg->swap_add = cpu_to_be64(wr->wr.atomic.compare_add);
2351                 aseg->compare  = 0;
2352         }
2353
2354 }
2355
2356 static void set_masked_atomic_seg(struct mlx4_wqe_masked_atomic_seg *aseg,
2357                                   struct ib_send_wr *wr)
2358 {
2359         aseg->swap_add          = cpu_to_be64(wr->wr.atomic.swap);
2360         aseg->swap_add_mask     = cpu_to_be64(wr->wr.atomic.swap_mask);
2361         aseg->compare           = cpu_to_be64(wr->wr.atomic.compare_add);
2362         aseg->compare_mask      = cpu_to_be64(wr->wr.atomic.compare_add_mask);
2363 }
2364
2365 static void set_datagram_seg(struct mlx4_wqe_datagram_seg *dseg,
2366                              struct ib_send_wr *wr)
2367 {
2368         memcpy(dseg->av, &to_mah(wr->wr.ud.ah)->av, sizeof (struct mlx4_av));
2369         dseg->dqpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2370         dseg->qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2371         dseg->vlan = to_mah(wr->wr.ud.ah)->av.eth.vlan;
2372         memcpy(dseg->mac, to_mah(wr->wr.ud.ah)->av.eth.mac, 6);
2373 }
2374
2375 static void set_tunnel_datagram_seg(struct mlx4_ib_dev *dev,
2376                                     struct mlx4_wqe_datagram_seg *dseg,
2377                                     struct ib_send_wr *wr, enum ib_qp_type qpt)
2378 {
2379         union mlx4_ext_av *av = &to_mah(wr->wr.ud.ah)->av;
2380         struct mlx4_av sqp_av = {0};
2381         int port = *((u8 *) &av->ib.port_pd) & 0x3;
2382
2383         /* force loopback */
2384         sqp_av.port_pd = av->ib.port_pd | cpu_to_be32(0x80000000);
2385         sqp_av.g_slid = av->ib.g_slid & 0x7f; /* no GRH */
2386         sqp_av.sl_tclass_flowlabel = av->ib.sl_tclass_flowlabel &
2387                         cpu_to_be32(0xf0000000);
2388
2389         memcpy(dseg->av, &sqp_av, sizeof (struct mlx4_av));
2390         /* This function used only for sending on QP1 proxies */
2391         dseg->dqpn = cpu_to_be32(dev->dev->caps.qp1_tunnel[port - 1]);
2392         /* Use QKEY from the QP context, which is set by master */
2393         dseg->qkey = cpu_to_be32(IB_QP_SET_QKEY);
2394 }
2395
2396 static void build_tunnel_header(struct ib_send_wr *wr, void *wqe, unsigned *mlx_seg_len)
2397 {
2398         struct mlx4_wqe_inline_seg *inl = wqe;
2399         struct mlx4_ib_tunnel_header hdr;
2400         struct mlx4_ib_ah *ah = to_mah(wr->wr.ud.ah);
2401         int spc;
2402         int i;
2403
2404         memcpy(&hdr.av, &ah->av, sizeof hdr.av);
2405         hdr.remote_qpn = cpu_to_be32(wr->wr.ud.remote_qpn);
2406         hdr.pkey_index = cpu_to_be16(wr->wr.ud.pkey_index);
2407         hdr.qkey = cpu_to_be32(wr->wr.ud.remote_qkey);
2408         memcpy(hdr.mac, ah->av.eth.mac, 6);
2409         hdr.vlan = ah->av.eth.vlan;
2410
2411         spc = MLX4_INLINE_ALIGN -
2412                 ((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
2413         if (sizeof (hdr) <= spc) {
2414                 memcpy(inl + 1, &hdr, sizeof (hdr));
2415                 wmb();
2416                 inl->byte_count = cpu_to_be32(1 << 31 | sizeof (hdr));
2417                 i = 1;
2418         } else {
2419                 memcpy(inl + 1, &hdr, spc);
2420                 wmb();
2421                 inl->byte_count = cpu_to_be32(1 << 31 | spc);
2422
2423                 inl = (void *) (inl + 1) + spc;
2424                 memcpy(inl + 1, (void *) &hdr + spc, sizeof (hdr) - spc);
2425                 wmb();
2426                 inl->byte_count = cpu_to_be32(1 << 31 | (sizeof (hdr) - spc));
2427                 i = 2;
2428         }
2429
2430         *mlx_seg_len =
2431                 ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + sizeof (hdr), 16);
2432 }
2433
2434 static void set_mlx_icrc_seg(void *dseg)
2435 {
2436         u32 *t = dseg;
2437         struct mlx4_wqe_inline_seg *iseg = dseg;
2438
2439         t[1] = 0;
2440
2441         /*
2442          * Need a barrier here before writing the byte_count field to
2443          * make sure that all the data is visible before the
2444          * byte_count field is set.  Otherwise, if the segment begins
2445          * a new cacheline, the HCA prefetcher could grab the 64-byte
2446          * chunk and get a valid (!= * 0xffffffff) byte count but
2447          * stale data, and end up sending the wrong data.
2448          */
2449         wmb();
2450
2451         iseg->byte_count = cpu_to_be32((1 << 31) | 4);
2452 }
2453
2454 static void set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2455 {
2456         dseg->lkey       = cpu_to_be32(sg->lkey);
2457         dseg->addr       = cpu_to_be64(sg->addr);
2458
2459         /*
2460          * Need a barrier here before writing the byte_count field to
2461          * make sure that all the data is visible before the
2462          * byte_count field is set.  Otherwise, if the segment begins
2463          * a new cacheline, the HCA prefetcher could grab the 64-byte
2464          * chunk and get a valid (!= * 0xffffffff) byte count but
2465          * stale data, and end up sending the wrong data.
2466          */
2467         wmb();
2468
2469         dseg->byte_count = cpu_to_be32(sg->length);
2470 }
2471
2472 static void __set_data_seg(struct mlx4_wqe_data_seg *dseg, struct ib_sge *sg)
2473 {
2474         dseg->byte_count = cpu_to_be32(sg->length);
2475         dseg->lkey       = cpu_to_be32(sg->lkey);
2476         dseg->addr       = cpu_to_be64(sg->addr);
2477 }
2478
2479 static int build_lso_seg(struct mlx4_wqe_lso_seg *wqe, struct ib_send_wr *wr,
2480                          struct mlx4_ib_qp *qp, unsigned *lso_seg_len,
2481                          __be32 *lso_hdr_sz, __be32 *blh)
2482 {
2483         unsigned halign = ALIGN(sizeof *wqe + wr->wr.ud.hlen, 16);
2484
2485         if (unlikely(halign > MLX4_IB_CACHE_LINE_SIZE))
2486                 *blh = cpu_to_be32(1 << 6);
2487
2488         if (unlikely(!(qp->flags & MLX4_IB_QP_LSO) &&
2489                      wr->num_sge > qp->sq.max_gs - (halign >> 4)))
2490                 return -EINVAL;
2491
2492         memcpy(wqe->header, wr->wr.ud.header, wr->wr.ud.hlen);
2493
2494         *lso_hdr_sz  = cpu_to_be32((wr->wr.ud.mss - wr->wr.ud.hlen) << 16 |
2495                                    wr->wr.ud.hlen);
2496         *lso_seg_len = halign;
2497         return 0;
2498 }
2499
2500 static __be32 send_ieth(struct ib_send_wr *wr)
2501 {
2502         switch (wr->opcode) {
2503         case IB_WR_SEND_WITH_IMM:
2504         case IB_WR_RDMA_WRITE_WITH_IMM:
2505                 return wr->ex.imm_data;
2506
2507         case IB_WR_SEND_WITH_INV:
2508                 return cpu_to_be32(wr->ex.invalidate_rkey);
2509
2510         default:
2511                 return 0;
2512         }
2513 }
2514
2515 static void add_zero_len_inline(void *wqe)
2516 {
2517         struct mlx4_wqe_inline_seg *inl = wqe;
2518         memset(wqe, 0, 16);
2519         inl->byte_count = cpu_to_be32(1 << 31);
2520 }
2521
2522 int mlx4_ib_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
2523                       struct ib_send_wr **bad_wr)
2524 {
2525         struct mlx4_ib_qp *qp = to_mqp(ibqp);
2526         void *wqe;
2527         struct mlx4_wqe_ctrl_seg *ctrl;
2528         struct mlx4_wqe_data_seg *dseg;
2529         unsigned long flags;
2530         int nreq;
2531         int err = 0;
2532         unsigned ind;
2533         int uninitialized_var(stamp);
2534         int uninitialized_var(size);
2535         unsigned uninitialized_var(seglen);
2536         __be32 dummy;
2537         __be32 *lso_wqe;
2538         __be32 uninitialized_var(lso_hdr_sz);
2539         __be32 blh;
2540         int i;
2541
2542         spin_lock_irqsave(&qp->sq.lock, flags);
2543
2544         ind = qp->sq_next_wqe;
2545
2546         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2547                 lso_wqe = &dummy;
2548                 blh = 0;
2549
2550                 if (mlx4_wq_overflow(&qp->sq, nreq, qp->ibqp.send_cq)) {
2551                         err = -ENOMEM;
2552                         *bad_wr = wr;
2553                         goto out;
2554                 }
2555
2556                 if (unlikely(wr->num_sge > qp->sq.max_gs)) {
2557                         err = -EINVAL;
2558                         *bad_wr = wr;
2559                         goto out;
2560                 }
2561
2562                 ctrl = wqe = get_send_wqe(qp, ind & (qp->sq.wqe_cnt - 1));
2563                 qp->sq.wrid[(qp->sq.head + nreq) & (qp->sq.wqe_cnt - 1)] = wr->wr_id;
2564
2565                 ctrl->srcrb_flags =
2566                         (wr->send_flags & IB_SEND_SIGNALED ?
2567                          cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) : 0) |
2568                         (wr->send_flags & IB_SEND_SOLICITED ?
2569                          cpu_to_be32(MLX4_WQE_CTRL_SOLICITED) : 0) |
2570                         ((wr->send_flags & IB_SEND_IP_CSUM) ?
2571                          cpu_to_be32(MLX4_WQE_CTRL_IP_CSUM |
2572                                      MLX4_WQE_CTRL_TCP_UDP_CSUM) : 0) |
2573                         qp->sq_signal_bits;
2574
2575                 ctrl->imm = send_ieth(wr);
2576
2577                 wqe += sizeof *ctrl;
2578                 size = sizeof *ctrl / 16;
2579
2580                 switch (qp->mlx4_ib_qp_type) {
2581                 case MLX4_IB_QPT_RC:
2582                 case MLX4_IB_QPT_UC:
2583                         switch (wr->opcode) {
2584                         case IB_WR_ATOMIC_CMP_AND_SWP:
2585                         case IB_WR_ATOMIC_FETCH_AND_ADD:
2586                         case IB_WR_MASKED_ATOMIC_FETCH_AND_ADD:
2587                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2588                                               wr->wr.atomic.rkey);
2589                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2590
2591                                 set_atomic_seg(wqe, wr);
2592                                 wqe  += sizeof (struct mlx4_wqe_atomic_seg);
2593
2594                                 size += (sizeof (struct mlx4_wqe_raddr_seg) +
2595                                          sizeof (struct mlx4_wqe_atomic_seg)) / 16;
2596
2597                                 break;
2598
2599                         case IB_WR_MASKED_ATOMIC_CMP_AND_SWP:
2600                                 set_raddr_seg(wqe, wr->wr.atomic.remote_addr,
2601                                               wr->wr.atomic.rkey);
2602                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2603
2604                                 set_masked_atomic_seg(wqe, wr);
2605                                 wqe  += sizeof (struct mlx4_wqe_masked_atomic_seg);
2606
2607                                 size += (sizeof (struct mlx4_wqe_raddr_seg) +
2608                                          sizeof (struct mlx4_wqe_masked_atomic_seg)) / 16;
2609
2610                                 break;
2611
2612                         case IB_WR_RDMA_READ:
2613                         case IB_WR_RDMA_WRITE:
2614                         case IB_WR_RDMA_WRITE_WITH_IMM:
2615                                 set_raddr_seg(wqe, wr->wr.rdma.remote_addr,
2616                                               wr->wr.rdma.rkey);
2617                                 wqe  += sizeof (struct mlx4_wqe_raddr_seg);
2618                                 size += sizeof (struct mlx4_wqe_raddr_seg) / 16;
2619                                 break;
2620
2621                         case IB_WR_LOCAL_INV:
2622                                 ctrl->srcrb_flags |=
2623                                         cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2624                                 set_local_inv_seg(wqe, wr->ex.invalidate_rkey);
2625                                 wqe  += sizeof (struct mlx4_wqe_local_inval_seg);
2626                                 size += sizeof (struct mlx4_wqe_local_inval_seg) / 16;
2627                                 break;
2628
2629                         case IB_WR_FAST_REG_MR:
2630                                 ctrl->srcrb_flags |=
2631                                         cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2632                                 set_fmr_seg(wqe, wr);
2633                                 wqe  += sizeof (struct mlx4_wqe_fmr_seg);
2634                                 size += sizeof (struct mlx4_wqe_fmr_seg) / 16;
2635                                 break;
2636
2637                         case IB_WR_BIND_MW:
2638                                 ctrl->srcrb_flags |=
2639                                         cpu_to_be32(MLX4_WQE_CTRL_STRONG_ORDER);
2640                                 set_bind_seg(wqe, wr);
2641                                 wqe  += sizeof(struct mlx4_wqe_bind_seg);
2642                                 size += sizeof(struct mlx4_wqe_bind_seg) / 16;
2643                                 break;
2644                         default:
2645                                 /* No extra segments required for sends */
2646                                 break;
2647                         }
2648                         break;
2649
2650                 case MLX4_IB_QPT_TUN_SMI_OWNER:
2651                         err =  build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2652                         if (unlikely(err)) {
2653                                 *bad_wr = wr;
2654                                 goto out;
2655                         }
2656                         wqe  += seglen;
2657                         size += seglen / 16;
2658                         break;
2659                 case MLX4_IB_QPT_TUN_SMI:
2660                 case MLX4_IB_QPT_TUN_GSI:
2661                         /* this is a UD qp used in MAD responses to slaves. */
2662                         set_datagram_seg(wqe, wr);
2663                         /* set the forced-loopback bit in the data seg av */
2664                         *(__be32 *) wqe |= cpu_to_be32(0x80000000);
2665                         wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2666                         size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2667                         break;
2668                 case MLX4_IB_QPT_UD:
2669                         set_datagram_seg(wqe, wr);
2670                         wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2671                         size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2672
2673                         if (wr->opcode == IB_WR_LSO) {
2674                                 err = build_lso_seg(wqe, wr, qp, &seglen, &lso_hdr_sz, &blh);
2675                                 if (unlikely(err)) {
2676                                         *bad_wr = wr;
2677                                         goto out;
2678                                 }
2679                                 lso_wqe = (__be32 *) wqe;
2680                                 wqe  += seglen;
2681                                 size += seglen / 16;
2682                         }
2683                         break;
2684
2685                 case MLX4_IB_QPT_PROXY_SMI_OWNER:
2686                         if (unlikely(!mlx4_is_master(to_mdev(ibqp->device)->dev))) {
2687                                 err = -ENOSYS;
2688                                 *bad_wr = wr;
2689                                 goto out;
2690                         }
2691                         err = build_sriov_qp0_header(to_msqp(qp), wr, ctrl, &seglen);
2692                         if (unlikely(err)) {
2693                                 *bad_wr = wr;
2694                                 goto out;
2695                         }
2696                         wqe  += seglen;
2697                         size += seglen / 16;
2698                         /* to start tunnel header on a cache-line boundary */
2699                         add_zero_len_inline(wqe);
2700                         wqe += 16;
2701                         size++;
2702                         build_tunnel_header(wr, wqe, &seglen);
2703                         wqe  += seglen;
2704                         size += seglen / 16;
2705                         break;
2706                 case MLX4_IB_QPT_PROXY_SMI:
2707                         /* don't allow QP0 sends on guests */
2708                         err = -ENOSYS;
2709                         *bad_wr = wr;
2710                         goto out;
2711                 case MLX4_IB_QPT_PROXY_GSI:
2712                         /* If we are tunneling special qps, this is a UD qp.
2713                          * In this case we first add a UD segment targeting
2714                          * the tunnel qp, and then add a header with address
2715                          * information */
2716                         set_tunnel_datagram_seg(to_mdev(ibqp->device), wqe, wr, ibqp->qp_type);
2717                         wqe  += sizeof (struct mlx4_wqe_datagram_seg);
2718                         size += sizeof (struct mlx4_wqe_datagram_seg) / 16;
2719                         build_tunnel_header(wr, wqe, &seglen);
2720                         wqe  += seglen;
2721                         size += seglen / 16;
2722                         break;
2723
2724                 case MLX4_IB_QPT_SMI:
2725                 case MLX4_IB_QPT_GSI:
2726                         err = build_mlx_header(to_msqp(qp), wr, ctrl, &seglen);
2727                         if (unlikely(err)) {
2728                                 *bad_wr = wr;
2729                                 goto out;
2730                         }
2731                         wqe  += seglen;
2732                         size += seglen / 16;
2733                         break;
2734
2735                 default:
2736                         break;
2737                 }
2738
2739                 /*
2740                  * Write data segments in reverse order, so as to
2741                  * overwrite cacheline stamp last within each
2742                  * cacheline.  This avoids issues with WQE
2743                  * prefetching.
2744                  */
2745
2746                 dseg = wqe;
2747                 dseg += wr->num_sge - 1;
2748                 size += wr->num_sge * (sizeof (struct mlx4_wqe_data_seg) / 16);
2749
2750                 /* Add one more inline data segment for ICRC for MLX sends */
2751                 if (unlikely(qp->mlx4_ib_qp_type == MLX4_IB_QPT_SMI ||
2752                              qp->mlx4_ib_qp_type == MLX4_IB_QPT_GSI ||
2753                              qp->mlx4_ib_qp_type &
2754                              (MLX4_IB_QPT_PROXY_SMI_OWNER | MLX4_IB_QPT_TUN_SMI_OWNER))) {
2755                         set_mlx_icrc_seg(dseg + 1);
2756                         size += sizeof (struct mlx4_wqe_data_seg) / 16;
2757                 }
2758
2759                 for (i = wr->num_sge - 1; i >= 0; --i, --dseg)
2760                         set_data_seg(dseg, wr->sg_list + i);
2761
2762                 /*
2763                  * Possibly overwrite stamping in cacheline with LSO
2764                  * segment only after making sure all data segments
2765                  * are written.
2766                  */
2767                 wmb();
2768                 *lso_wqe = lso_hdr_sz;
2769
2770                 ctrl->fence_size = (wr->send_flags & IB_SEND_FENCE ?
2771                                     MLX4_WQE_CTRL_FENCE : 0) | size;
2772
2773                 /*
2774                  * Make sure descriptor is fully written before
2775                  * setting ownership bit (because HW can start
2776                  * executing as soon as we do).
2777                  */
2778                 wmb();
2779
2780                 if (wr->opcode < 0 || wr->opcode >= ARRAY_SIZE(mlx4_ib_opcode)) {
2781                         *bad_wr = wr;
2782                         err = -EINVAL;
2783                         goto out;
2784                 }
2785
2786                 ctrl->owner_opcode = mlx4_ib_opcode[wr->opcode] |
2787                         (ind & qp->sq.wqe_cnt ? cpu_to_be32(1 << 31) : 0) | blh;
2788
2789                 stamp = ind + qp->sq_spare_wqes;
2790                 ind += DIV_ROUND_UP(size * 16, 1U << qp->sq.wqe_shift);
2791
2792                 /*
2793                  * We can improve latency by not stamping the last
2794                  * send queue WQE until after ringing the doorbell, so
2795                  * only stamp here if there are still more WQEs to post.
2796                  *
2797                  * Same optimization applies to padding with NOP wqe
2798                  * in case of WQE shrinking (used to prevent wrap-around
2799                  * in the middle of WR).
2800                  */
2801                 if (wr->next) {
2802                         stamp_send_wqe(qp, stamp, size * 16);
2803                         ind = pad_wraparound(qp, ind);
2804                 }
2805         }
2806
2807 out:
2808         if (likely(nreq)) {
2809                 qp->sq.head += nreq;
2810
2811                 /*
2812                  * Make sure that descriptors are written before
2813                  * doorbell record.
2814                  */
2815                 wmb();
2816
2817                 writel(qp->doorbell_qpn,
2818                        to_mdev(ibqp->device)->uar_map + MLX4_SEND_DOORBELL);
2819
2820                 /*
2821                  * Make sure doorbells don't leak out of SQ spinlock
2822                  * and reach the HCA out of order.
2823                  */
2824                 mmiowb();
2825
2826                 stamp_send_wqe(qp, stamp, size * 16);
2827
2828                 ind = pad_wraparound(qp, ind);
2829                 qp->sq_next_wqe = ind;
2830         }
2831
2832         spin_unlock_irqrestore(&qp->sq.lock, flags);
2833
2834         return err;
2835 }
2836
2837 int mlx4_ib_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
2838                       struct ib_recv_wr **bad_wr)
2839 {
2840         struct mlx4_ib_qp *qp = to_mqp(ibqp);
2841         struct mlx4_wqe_data_seg *scat;
2842         unsigned long flags;
2843         int err = 0;
2844         int nreq;
2845         int ind;
2846         int max_gs;
2847         int i;
2848
2849         max_gs = qp->rq.max_gs;
2850         spin_lock_irqsave(&qp->rq.lock, flags);
2851
2852         ind = qp->rq.head & (qp->rq.wqe_cnt - 1);
2853
2854         for (nreq = 0; wr; ++nreq, wr = wr->next) {
2855                 if (mlx4_wq_overflow(&qp->rq, nreq, qp->ibqp.recv_cq)) {
2856                         err = -ENOMEM;
2857                         *bad_wr = wr;
2858                         goto out;
2859                 }
2860
2861                 if (unlikely(wr->num_sge > qp->rq.max_gs)) {
2862                         err = -EINVAL;
2863                         *bad_wr = wr;
2864                         goto out;
2865                 }
2866
2867                 scat = get_recv_wqe(qp, ind);
2868
2869                 if (qp->mlx4_ib_qp_type & (MLX4_IB_QPT_PROXY_SMI_OWNER |
2870                     MLX4_IB_QPT_PROXY_SMI | MLX4_IB_QPT_PROXY_GSI)) {
2871                         ib_dma_sync_single_for_device(ibqp->device,
2872                                                       qp->sqp_proxy_rcv[ind].map,
2873                                                       sizeof (struct mlx4_ib_proxy_sqp_hdr),
2874                                                       DMA_FROM_DEVICE);
2875                         scat->byte_count =
2876                                 cpu_to_be32(sizeof (struct mlx4_ib_proxy_sqp_hdr));
2877                         /* use dma lkey from upper layer entry */
2878                         scat->lkey = cpu_to_be32(wr->sg_list->lkey);
2879                         scat->addr = cpu_to_be64(qp->sqp_proxy_rcv[ind].map);
2880                         scat++;
2881                         max_gs--;
2882                 }
2883
2884                 for (i = 0; i < wr->num_sge; ++i)
2885                         __set_data_seg(scat + i, wr->sg_list + i);
2886
2887                 if (i < max_gs) {
2888                         scat[i].byte_count = 0;
2889                         scat[i].lkey       = cpu_to_be32(MLX4_INVALID_LKEY);
2890                         scat[i].addr       = 0;
2891                 }
2892
2893                 qp->rq.wrid[ind] = wr->wr_id;
2894
2895                 ind = (ind + 1) & (qp->rq.wqe_cnt - 1);
2896         }
2897
2898 out:
2899         if (likely(nreq)) {
2900                 qp->rq.head += nreq;
2901
2902                 /*
2903                  * Make sure that descriptors are written before
2904                  * doorbell record.
2905                  */
2906                 wmb();
2907
2908                 *qp->db.db = cpu_to_be32(qp->rq.head & 0xffff);
2909         }
2910
2911         spin_unlock_irqrestore(&qp->rq.lock, flags);
2912
2913         return err;
2914 }
2915
2916 static inline enum ib_qp_state to_ib_qp_state(enum mlx4_qp_state mlx4_state)
2917 {
2918         switch (mlx4_state) {
2919         case MLX4_QP_STATE_RST:      return IB_QPS_RESET;
2920         case MLX4_QP_STATE_INIT:     return IB_QPS_INIT;
2921         case MLX4_QP_STATE_RTR:      return IB_QPS_RTR;
2922         case MLX4_QP_STATE_RTS:      return IB_QPS_RTS;
2923         case MLX4_QP_STATE_SQ_DRAINING:
2924         case MLX4_QP_STATE_SQD:      return IB_QPS_SQD;
2925         case MLX4_QP_STATE_SQER:     return IB_QPS_SQE;
2926         case MLX4_QP_STATE_ERR:      return IB_QPS_ERR;
2927         default:                     return -1;
2928         }
2929 }
2930
2931 static inline enum ib_mig_state to_ib_mig_state(int mlx4_mig_state)
2932 {
2933         switch (mlx4_mig_state) {
2934         case MLX4_QP_PM_ARMED:          return IB_MIG_ARMED;
2935         case MLX4_QP_PM_REARM:          return IB_MIG_REARM;
2936         case MLX4_QP_PM_MIGRATED:       return IB_MIG_MIGRATED;
2937         default: return -1;
2938         }
2939 }
2940
2941 static int to_ib_qp_access_flags(int mlx4_flags)
2942 {
2943         int ib_flags = 0;
2944
2945         if (mlx4_flags & MLX4_QP_BIT_RRE)
2946                 ib_flags |= IB_ACCESS_REMOTE_READ;
2947         if (mlx4_flags & MLX4_QP_BIT_RWE)
2948                 ib_flags |= IB_ACCESS_REMOTE_WRITE;
2949         if (mlx4_flags & MLX4_QP_BIT_RAE)
2950                 ib_flags |= IB_ACCESS_REMOTE_ATOMIC;
2951
2952         return ib_flags;
2953 }
2954
2955 static void to_ib_ah_attr(struct mlx4_ib_dev *ibdev, struct ib_ah_attr *ib_ah_attr,
2956                                 struct mlx4_qp_path *path)
2957 {
2958         struct mlx4_dev *dev = ibdev->dev;
2959         int is_eth;
2960
2961         memset(ib_ah_attr, 0, sizeof *ib_ah_attr);
2962         ib_ah_attr->port_num      = path->sched_queue & 0x40 ? 2 : 1;
2963
2964         if (ib_ah_attr->port_num == 0 || ib_ah_attr->port_num > dev->caps.num_ports)
2965                 return;
2966
2967         is_eth = rdma_port_get_link_layer(&ibdev->ib_dev, ib_ah_attr->port_num) ==
2968                 IB_LINK_LAYER_ETHERNET;
2969         if (is_eth)
2970                 ib_ah_attr->sl = ((path->sched_queue >> 3) & 0x7) |
2971                 ((path->sched_queue & 4) << 1);
2972         else
2973                 ib_ah_attr->sl = (path->sched_queue >> 2) & 0xf;
2974
2975         ib_ah_attr->dlid          = be16_to_cpu(path->rlid);
2976         ib_ah_attr->src_path_bits = path->grh_mylmc & 0x7f;
2977         ib_ah_attr->static_rate   = path->static_rate ? path->static_rate - 5 : 0;
2978         ib_ah_attr->ah_flags      = (path->grh_mylmc & (1 << 7)) ? IB_AH_GRH : 0;
2979         if (ib_ah_attr->ah_flags) {
2980                 ib_ah_attr->grh.sgid_index = path->mgid_index;
2981                 ib_ah_attr->grh.hop_limit  = path->hop_limit;
2982                 ib_ah_attr->grh.traffic_class =
2983                         (be32_to_cpu(path->tclass_flowlabel) >> 20) & 0xff;
2984                 ib_ah_attr->grh.flow_label =
2985                         be32_to_cpu(path->tclass_flowlabel) & 0xfffff;
2986                 memcpy(ib_ah_attr->grh.dgid.raw,
2987                         path->rgid, sizeof ib_ah_attr->grh.dgid.raw);
2988         }
2989 }
2990
2991 int mlx4_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *qp_attr, int qp_attr_mask,
2992                      struct ib_qp_init_attr *qp_init_attr)
2993 {
2994         struct mlx4_ib_dev *dev = to_mdev(ibqp->device);
2995         struct mlx4_ib_qp *qp = to_mqp(ibqp);
2996         struct mlx4_qp_context context;
2997         int mlx4_state;
2998         int err = 0;
2999
3000         mutex_lock(&qp->mutex);
3001
3002         if (qp->state == IB_QPS_RESET) {
3003                 qp_attr->qp_state = IB_QPS_RESET;
3004                 goto done;
3005         }
3006
3007         err = mlx4_qp_query(dev->dev, &qp->mqp, &context);
3008         if (err) {
3009                 err = -EINVAL;
3010                 goto out;
3011         }
3012
3013         mlx4_state = be32_to_cpu(context.flags) >> 28;
3014
3015         qp->state                    = to_ib_qp_state(mlx4_state);
3016         qp_attr->qp_state            = qp->state;
3017         qp_attr->path_mtu            = context.mtu_msgmax >> 5;
3018         qp_attr->path_mig_state      =
3019                 to_ib_mig_state((be32_to_cpu(context.flags) >> 11) & 0x3);
3020         qp_attr->qkey                = be32_to_cpu(context.qkey);
3021         qp_attr->rq_psn              = be32_to_cpu(context.rnr_nextrecvpsn) & 0xffffff;
3022         qp_attr->sq_psn              = be32_to_cpu(context.next_send_psn) & 0xffffff;
3023         qp_attr->dest_qp_num         = be32_to_cpu(context.remote_qpn) & 0xffffff;
3024         qp_attr->qp_access_flags     =
3025                 to_ib_qp_access_flags(be32_to_cpu(context.params2));
3026
3027         if (qp->ibqp.qp_type == IB_QPT_RC || qp->ibqp.qp_type == IB_QPT_UC) {
3028                 to_ib_ah_attr(dev, &qp_attr->ah_attr, &context.pri_path);
3029                 to_ib_ah_attr(dev, &qp_attr->alt_ah_attr, &context.alt_path);
3030                 qp_attr->alt_pkey_index = context.alt_path.pkey_index & 0x7f;
3031                 qp_attr->alt_port_num   = qp_attr->alt_ah_attr.port_num;
3032         }
3033
3034         qp_attr->pkey_index = context.pri_path.pkey_index & 0x7f;
3035         if (qp_attr->qp_state == IB_QPS_INIT)
3036                 qp_attr->port_num = qp->port;
3037         else
3038                 qp_attr->port_num = context.pri_path.sched_queue & 0x40 ? 2 : 1;
3039
3040         /* qp_attr->en_sqd_async_notify is only applicable in modify qp */
3041         qp_attr->sq_draining = mlx4_state == MLX4_QP_STATE_SQ_DRAINING;
3042
3043         qp_attr->max_rd_atomic = 1 << ((be32_to_cpu(context.params1) >> 21) & 0x7);
3044
3045         qp_attr->max_dest_rd_atomic =
3046                 1 << ((be32_to_cpu(context.params2) >> 21) & 0x7);
3047         qp_attr->min_rnr_timer      =
3048                 (be32_to_cpu(context.rnr_nextrecvpsn) >> 24) & 0x1f;
3049         qp_attr->timeout            = context.pri_path.ackto >> 3;
3050         qp_attr->retry_cnt          = (be32_to_cpu(context.params1) >> 16) & 0x7;
3051         qp_attr->rnr_retry          = (be32_to_cpu(context.params1) >> 13) & 0x7;
3052         qp_attr->alt_timeout        = context.alt_path.ackto >> 3;
3053
3054 done:
3055         qp_attr->cur_qp_state        = qp_attr->qp_state;
3056         qp_attr->cap.max_recv_wr     = qp->rq.wqe_cnt;
3057         qp_attr->cap.max_recv_sge    = qp->rq.max_gs;
3058
3059         if (!ibqp->uobject) {
3060                 qp_attr->cap.max_send_wr  = qp->sq.wqe_cnt;
3061                 qp_attr->cap.max_send_sge = qp->sq.max_gs;
3062         } else {
3063                 qp_attr->cap.max_send_wr  = 0;
3064                 qp_attr->cap.max_send_sge = 0;
3065         }
3066
3067         /*
3068          * We don't support inline sends for kernel QPs (yet), and we
3069          * don't know what userspace's value should be.
3070          */
3071         qp_attr->cap.max_inline_data = 0;
3072
3073         qp_init_attr->cap            = qp_attr->cap;
3074
3075         qp_init_attr->create_flags = 0;
3076         if (qp->flags & MLX4_IB_QP_BLOCK_MULTICAST_LOOPBACK)
3077                 qp_init_attr->create_flags |= IB_QP_CREATE_BLOCK_MULTICAST_LOOPBACK;
3078
3079         if (qp->flags & MLX4_IB_QP_LSO)
3080                 qp_init_attr->create_flags |= IB_QP_CREATE_IPOIB_UD_LSO;
3081
3082         if (qp->flags & MLX4_IB_QP_NETIF)
3083                 qp_init_attr->create_flags |= IB_QP_CREATE_NETIF_QP;
3084
3085         qp_init_attr->sq_sig_type =
3086                 qp->sq_signal_bits == cpu_to_be32(MLX4_WQE_CTRL_CQ_UPDATE) ?
3087                 IB_SIGNAL_ALL_WR : IB_SIGNAL_REQ_WR;
3088
3089 out:
3090         mutex_unlock(&qp->mutex);
3091         return err;
3092 }
3093