drivers/infiniband/hw/hfi1/ruc.c

   1 /*
   2  * Copyright(c) 2015 - 2018 Intel Corporation.
   3  *
   4  * This file is provided under a dual BSD/GPLv2 license.  When using or
   5  * redistributing this file, you may do so under either license.
   6  *
   7  * GPL LICENSE SUMMARY
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of version 2 of the GNU General Public License as
  11  * published by the Free Software Foundation.
  12  *
  13  * This program is distributed in the hope that it will be useful, but
  14  * WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  16  * General Public License for more details.
  17  *
  18  * BSD LICENSE
  19  *
  20  * Redistribution and use in source and binary forms, with or without
  21  * modification, are permitted provided that the following conditions
  22  * are met:
  23  *
  24  *  - Redistributions of source code must retain the above copyright
  25  *    notice, this list of conditions and the following disclaimer.
  26  *  - Redistributions in binary form must reproduce the above copyright
  27  *    notice, this list of conditions and the following disclaimer in
  28  *    the documentation and/or other materials provided with the
  29  *    distribution.
  30  *  - Neither the name of Intel Corporation nor the names of its
  31  *    contributors may be used to endorse or promote products derived
  32  *    from this software without specific prior written permission.
  33  *
  34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  45  *
  46  */
  47
  48 #include <linux/spinlock.h>
  49
  50 #include "hfi.h"
  51 #include "mad.h"
  52 #include "qp.h"
  53 #include "verbs_txreq.h"
  54 #include "trace.h"
  55
  56 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
  57 {
  58         return (gid->global.interface_id == id &&
  59                 (gid->global.subnet_prefix == gid_prefix ||
  60                  gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
  61 }
  62
  63 /*
  64  *
  65  * This should be called with the QP r_lock held.
  66  *
  67  * The s_lock will be acquired around the hfi1_migrate_qp() call.
  68  */
  69 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_packet *packet)
  70 {
  71         __be64 guid;
  72         unsigned long flags;
  73         struct rvt_qp *qp = packet->qp;
  74         u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
  75         u32 dlid = packet->dlid;
  76         u32 slid = packet->slid;
  77         u32 sl = packet->sl;
  78         bool migrated = packet->migrated;
  79         u16 pkey = packet->pkey;
  80
  81         if (qp->s_mig_state == IB_MIG_ARMED && migrated) {
  82                 if (!packet->grh) {
  83                         if ((rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
  84                              IB_AH_GRH) &&
  85                             (packet->etype != RHF_RCV_TYPE_BYPASS))
  86                                 return 1;
  87                 } else {
  88                         const struct ib_global_route *grh;
  89
  90                         if (!(rdma_ah_get_ah_flags(&qp->alt_ah_attr) &
  91                               IB_AH_GRH))
  92                                 return 1;
  93                         grh = rdma_ah_read_grh(&qp->alt_ah_attr);
  94                         guid = get_sguid(ibp, grh->sgid_index);
  95                         if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
  96                                     guid))
  97                                 return 1;
  98                         if (!gid_ok(
  99                                 &packet->grh->sgid,
 100                                 grh->dgid.global.subnet_prefix,
 101                                 grh->dgid.global.interface_id))
 102                                 return 1;
 103                 }
 104                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
 105                                             sc5, slid))) {
 106                         hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
 107                                       slid, dlid);
 108                         return 1;
 109                 }
 110                 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
 111                 if (slid != rdma_ah_get_dlid(&qp->alt_ah_attr) ||
 112                     ppd_from_ibp(ibp)->port !=
 113                         rdma_ah_get_port_num(&qp->alt_ah_attr))
 114                         return 1;
 115                 spin_lock_irqsave(&qp->s_lock, flags);
 116                 hfi1_migrate_qp(qp);
 117                 spin_unlock_irqrestore(&qp->s_lock, flags);
 118         } else {
 119                 if (!packet->grh) {
 120                         if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
 121                              IB_AH_GRH) &&
 122                             (packet->etype != RHF_RCV_TYPE_BYPASS))
 123                                 return 1;
 124                 } else {
 125                         const struct ib_global_route *grh;
 126
 127                         if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) &
 128                                                    IB_AH_GRH))
 129                                 return 1;
 130                         grh = rdma_ah_read_grh(&qp->remote_ah_attr);
 131                         guid = get_sguid(ibp, grh->sgid_index);
 132                         if (!gid_ok(&packet->grh->dgid, ibp->rvp.gid_prefix,
 133                                     guid))
 134                                 return 1;
 135                         if (!gid_ok(
 136                              &packet->grh->sgid,
 137                              grh->dgid.global.subnet_prefix,
 138                              grh->dgid.global.interface_id))
 139                                 return 1;
 140                 }
 141                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), pkey,
 142                                             sc5, slid))) {
 143                         hfi1_bad_pkey(ibp, pkey, sl, 0, qp->ibqp.qp_num,
 144                                       slid, dlid);
 145                         return 1;
 146                 }
 147                 /* Validate the SLID. See Ch. 9.6.1.5 */
 148                 if ((slid != rdma_ah_get_dlid(&qp->remote_ah_attr)) ||
 149                     ppd_from_ibp(ibp)->port != qp->port_num)
 150                         return 1;
 151                 if (qp->s_mig_state == IB_MIG_REARM && !migrated)
 152                         qp->s_mig_state = IB_MIG_ARMED;
 153         }
 154
 155         return 0;
 156 }
 157
 158 /**
 159  * ruc_loopback - handle UC and RC loopback requests
 160  * @sqp: the sending QP
 161  *
 162  * This is called from hfi1_do_send() to
 163  * forward a WQE addressed to the same HFI.
 164  * Note that although we are single threaded due to the send engine, we still
 165  * have to protect against post_send().  We don't have to worry about
 166  * receive interrupts since this is a connected protocol and all packets
 167  * will pass through here.
 168  */
 169 static void ruc_loopback(struct rvt_qp *sqp)
 170 {
 171         struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
 172         struct rvt_qp *qp;
 173         struct rvt_swqe *wqe;
 174         struct rvt_sge *sge;
 175         unsigned long flags;
 176         struct ib_wc wc;
 177         u64 sdata;
 178         atomic64_t *maddr;
 179         enum ib_wc_status send_status;
 180         bool release;
 181         int ret;
 182         bool copy_last = false;
 183         int local_ops = 0;
 184
 185         rcu_read_lock();
 186
 187         /*
 188          * Note that we check the responder QP state after
 189          * checking the requester's state.
 190          */
 191         qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
 192                             sqp->remote_qpn);
 193
 194         spin_lock_irqsave(&sqp->s_lock, flags);
 195
 196         /* Return if we are already busy processing a work request. */
 197         if ((sqp->s_flags & (RVT_S_BUSY | HFI1_S_ANY_WAIT)) ||
 198             !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 199                 goto unlock;
 200
 201         sqp->s_flags |= RVT_S_BUSY;
 202
 203 again:
 204         if (sqp->s_last == READ_ONCE(sqp->s_head))
 205                 goto clr_busy;
 206         wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
 207
 208         /* Return if it is not OK to start a new work request. */
 209         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
 210                 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
 211                         goto clr_busy;
 212                 /* We are in the error state, flush the work request. */
 213                 send_status = IB_WC_WR_FLUSH_ERR;
 214                 goto flush_send;
 215         }
 216
 217         /*
 218          * We can rely on the entry not changing without the s_lock
 219          * being held until we update s_last.
 220          * We increment s_cur to indicate s_last is in progress.
 221          */
 222         if (sqp->s_last == sqp->s_cur) {
 223                 if (++sqp->s_cur >= sqp->s_size)
 224                         sqp->s_cur = 0;
 225         }
 226         spin_unlock_irqrestore(&sqp->s_lock, flags);
 227
 228         if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
 229             qp->ibqp.qp_type != sqp->ibqp.qp_type) {
 230                 ibp->rvp.n_pkt_drops++;
 231                 /*
 232                  * For RC, the requester would timeout and retry so
 233                  * shortcut the timeouts and just signal too many retries.
 234                  */
 235                 if (sqp->ibqp.qp_type == IB_QPT_RC)
 236                         send_status = IB_WC_RETRY_EXC_ERR;
 237                 else
 238                         send_status = IB_WC_SUCCESS;
 239                 goto serr;
 240         }
 241
 242         memset(&wc, 0, sizeof(wc));
 243         send_status = IB_WC_SUCCESS;
 244
 245         release = true;
 246         sqp->s_sge.sge = wqe->sg_list[0];
 247         sqp->s_sge.sg_list = wqe->sg_list + 1;
 248         sqp->s_sge.num_sge = wqe->wr.num_sge;
 249         sqp->s_len = wqe->length;
 250         switch (wqe->wr.opcode) {
 251         case IB_WR_REG_MR:
 252                 goto send_comp;
 253
 254         case IB_WR_LOCAL_INV:
 255                 if (!(wqe->wr.send_flags & RVT_SEND_COMPLETION_ONLY)) {
 256                         if (rvt_invalidate_rkey(sqp,
 257                                                 wqe->wr.ex.invalidate_rkey))
 258                                 send_status = IB_WC_LOC_PROT_ERR;
 259                         local_ops = 1;
 260                 }
 261                 goto send_comp;
 262
 263         case IB_WR_SEND_WITH_INV:
 264                 if (!rvt_invalidate_rkey(qp, wqe->wr.ex.invalidate_rkey)) {
 265                         wc.wc_flags = IB_WC_WITH_INVALIDATE;
 266                         wc.ex.invalidate_rkey = wqe->wr.ex.invalidate_rkey;
 267                 }
 268                 goto send;
 269
 270         case IB_WR_SEND_WITH_IMM:
 271                 wc.wc_flags = IB_WC_WITH_IMM;
 272                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 273                 /* FALLTHROUGH */
 274         case IB_WR_SEND:
 275 send:
 276                 ret = rvt_get_rwqe(qp, false);
 277                 if (ret < 0)
 278                         goto op_err;
 279                 if (!ret)
 280                         goto rnr_nak;
 281                 if (wqe->length > qp->r_len)
 282                         goto inv_err;
 283                 break;
 284
 285         case IB_WR_RDMA_WRITE_WITH_IMM:
 286                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 287                         goto inv_err;
 288                 wc.wc_flags = IB_WC_WITH_IMM;
 289                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 290                 ret = rvt_get_rwqe(qp, true);
 291                 if (ret < 0)
 292                         goto op_err;
 293                 if (!ret)
 294                         goto rnr_nak;
 295                 /* skip copy_last set and qp_access_flags recheck */
 296                 goto do_write;
 297         case IB_WR_RDMA_WRITE:
 298                 copy_last = rvt_is_user_qp(qp);
 299                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 300                         goto inv_err;
 301 do_write:
 302                 if (wqe->length == 0)
 303                         break;
 304                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
 305                                           wqe->rdma_wr.remote_addr,
 306                                           wqe->rdma_wr.rkey,
 307                                           IB_ACCESS_REMOTE_WRITE)))
 308                         goto acc_err;
 309                 qp->r_sge.sg_list = NULL;
 310                 qp->r_sge.num_sge = 1;
 311                 qp->r_sge.total_len = wqe->length;
 312                 break;
 313
 314         case IB_WR_RDMA_READ:
 315                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 316                         goto inv_err;
 317                 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
 318                                           wqe->rdma_wr.remote_addr,
 319                                           wqe->rdma_wr.rkey,
 320                                           IB_ACCESS_REMOTE_READ)))
 321                         goto acc_err;
 322                 release = false;
 323                 sqp->s_sge.sg_list = NULL;
 324                 sqp->s_sge.num_sge = 1;
 325                 qp->r_sge.sge = wqe->sg_list[0];
 326                 qp->r_sge.sg_list = wqe->sg_list + 1;
 327                 qp->r_sge.num_sge = wqe->wr.num_sge;
 328                 qp->r_sge.total_len = wqe->length;
 329                 break;
 330
 331         case IB_WR_ATOMIC_CMP_AND_SWP:
 332         case IB_WR_ATOMIC_FETCH_AND_ADD:
 333                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 334                         goto inv_err;
 335                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
 336                                           wqe->atomic_wr.remote_addr,
 337                                           wqe->atomic_wr.rkey,
 338                                           IB_ACCESS_REMOTE_ATOMIC)))
 339                         goto acc_err;
 340                 /* Perform atomic OP and save result. */
 341                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
 342                 sdata = wqe->atomic_wr.compare_add;
 343                 *(u64 *)sqp->s_sge.sge.vaddr =
 344                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 345                         (u64)atomic64_add_return(sdata, maddr) - sdata :
 346                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
 347                                       sdata, wqe->atomic_wr.swap);
 348                 rvt_put_mr(qp->r_sge.sge.mr);
 349                 qp->r_sge.num_sge = 0;
 350                 goto send_comp;
 351
 352         default:
 353                 send_status = IB_WC_LOC_QP_OP_ERR;
 354                 goto serr;
 355         }
 356
 357         sge = &sqp->s_sge.sge;
 358         while (sqp->s_len) {
 359                 u32 len = sqp->s_len;
 360
 361                 if (len > sge->length)
 362                         len = sge->length;
 363                 if (len > sge->sge_length)
 364                         len = sge->sge_length;
 365                 WARN_ON_ONCE(len == 0);
 366                 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
 367                 sge->vaddr += len;
 368                 sge->length -= len;
 369                 sge->sge_length -= len;
 370                 if (sge->sge_length == 0) {
 371                         if (!release)
 372                                 rvt_put_mr(sge->mr);
 373                         if (--sqp->s_sge.num_sge)
 374                                 *sge = *sqp->s_sge.sg_list++;
 375                 } else if (sge->length == 0 && sge->mr->lkey) {
 376                         if (++sge->n >= RVT_SEGSZ) {
 377                                 if (++sge->m >= sge->mr->mapsz)
 378                                         break;
 379                                 sge->n = 0;
 380                         }
 381                         sge->vaddr =
 382                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 383                         sge->length =
 384                                 sge->mr->map[sge->m]->segs[sge->n].length;
 385                 }
 386                 sqp->s_len -= len;
 387         }
 388         if (release)
 389                 rvt_put_ss(&qp->r_sge);
 390
 391         if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
 392                 goto send_comp;
 393
 394         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 395                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 396         else
 397                 wc.opcode = IB_WC_RECV;
 398         wc.wr_id = qp->r_wr_id;
 399         wc.status = IB_WC_SUCCESS;
 400         wc.byte_len = wqe->length;
 401         wc.qp = &qp->ibqp;
 402         wc.src_qp = qp->remote_qpn;
 403         wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
 404         wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
 405         wc.port_num = 1;
 406         /* Signal completion event if the solicited bit is set. */
 407         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
 408                      wqe->wr.send_flags & IB_SEND_SOLICITED);
 409
 410 send_comp:
 411         spin_lock_irqsave(&sqp->s_lock, flags);
 412         ibp->rvp.n_loop_pkts++;
 413 flush_send:
 414         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 415         hfi1_send_complete(sqp, wqe, send_status);
 416         if (local_ops) {
 417                 atomic_dec(&sqp->local_ops_pending);
 418                 local_ops = 0;
 419         }
 420         goto again;
 421
 422 rnr_nak:
 423         /* Handle RNR NAK */
 424         if (qp->ibqp.qp_type == IB_QPT_UC)
 425                 goto send_comp;
 426         ibp->rvp.n_rnr_naks++;
 427         /*
 428          * Note: we don't need the s_lock held since the BUSY flag
 429          * makes this single threaded.
 430          */
 431         if (sqp->s_rnr_retry == 0) {
 432                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 433                 goto serr;
 434         }
 435         if (sqp->s_rnr_retry_cnt < 7)
 436                 sqp->s_rnr_retry--;
 437         spin_lock_irqsave(&sqp->s_lock, flags);
 438         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
 439                 goto clr_busy;
 440         rvt_add_rnr_timer(sqp, qp->r_min_rnr_timer <<
 441                                 IB_AETH_CREDIT_SHIFT);
 442         goto clr_busy;
 443
 444 op_err:
 445         send_status = IB_WC_REM_OP_ERR;
 446         wc.status = IB_WC_LOC_QP_OP_ERR;
 447         goto err;
 448
 449 inv_err:
 450         send_status =
 451                 sqp->ibqp.qp_type == IB_QPT_RC ?
 452                         IB_WC_REM_INV_REQ_ERR :
 453                         IB_WC_SUCCESS;
 454         wc.status = IB_WC_LOC_QP_OP_ERR;
 455         goto err;
 456
 457 acc_err:
 458         send_status = IB_WC_REM_ACCESS_ERR;
 459         wc.status = IB_WC_LOC_PROT_ERR;
 460 err:
 461         /* responder goes to error state */
 462         rvt_rc_error(qp, wc.status);
 463
 464 serr:
 465         spin_lock_irqsave(&sqp->s_lock, flags);
 466         hfi1_send_complete(sqp, wqe, send_status);
 467         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 468                 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 469
 470                 sqp->s_flags &= ~RVT_S_BUSY;
 471                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 472                 if (lastwqe) {
 473                         struct ib_event ev;
 474
 475                         ev.device = sqp->ibqp.device;
 476                         ev.element.qp = &sqp->ibqp;
 477                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 478                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 479                 }
 480                 goto done;
 481         }
 482 clr_busy:
 483         sqp->s_flags &= ~RVT_S_BUSY;
 484 unlock:
 485         spin_unlock_irqrestore(&sqp->s_lock, flags);
 486 done:
 487         rcu_read_unlock();
 488 }
 489
 490 /**
 491  * hfi1_make_grh - construct a GRH header
 492  * @ibp: a pointer to the IB port
 493  * @hdr: a pointer to the GRH header being constructed
 494  * @grh: the global route address to send to
 495  * @hwords: size of header after grh being sent in dwords
 496  * @nwords: the number of 32 bit words of data being sent
 497  *
 498  * Return the size of the header in 32 bit words.
 499  */
 500 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
 501                   const struct ib_global_route *grh, u32 hwords, u32 nwords)
 502 {
 503         hdr->version_tclass_flow =
 504                 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
 505                             (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
 506                             (grh->flow_label << IB_GRH_FLOW_SHIFT));
 507         hdr->paylen = cpu_to_be16((hwords + nwords) << 2);
 508         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 509         hdr->next_hdr = IB_GRH_NEXT_HDR;
 510         hdr->hop_limit = grh->hop_limit;
 511         /* The SGID is 32-bit aligned. */
 512         hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
 513         hdr->sgid.global.interface_id =
 514                 grh->sgid_index < HFI1_GUIDS_PER_PORT ?
 515                 get_sguid(ibp, grh->sgid_index) :
 516                 get_sguid(ibp, HFI1_PORT_GUID_INDEX);
 517         hdr->dgid = grh->dgid;
 518
 519         /* GRH header size in 32-bit words. */
 520         return sizeof(struct ib_grh) / sizeof(u32);
 521 }
 522
 523 #define BTH2_OFFSET (offsetof(struct hfi1_sdma_header, \
 524                               hdr.ibh.u.oth.bth[2]) / 4)
 525
 526 /**
 527  * build_ahg - create ahg in s_ahg
 528  * @qp: a pointer to QP
 529  * @npsn: the next PSN for the request/response
 530  *
 531  * This routine handles the AHG by allocating an ahg entry and causing the
 532  * copy of the first middle.
 533  *
 534  * Subsequent middles use the copied entry, editing the
 535  * PSN with 1 or 2 edits.
 536  */
 537 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
 538 {
 539         struct hfi1_qp_priv *priv = qp->priv;
 540
 541         if (unlikely(qp->s_flags & HFI1_S_AHG_CLEAR))
 542                 clear_ahg(qp);
 543         if (!(qp->s_flags & HFI1_S_AHG_VALID)) {
 544                 /* first middle that needs copy  */
 545                 if (qp->s_ahgidx < 0)
 546                         qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
 547                 if (qp->s_ahgidx >= 0) {
 548                         qp->s_ahgpsn = npsn;
 549                         priv->s_ahg->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
 550                         /* save to protect a change in another thread */
 551                         priv->s_ahg->ahgidx = qp->s_ahgidx;
 552                         qp->s_flags |= HFI1_S_AHG_VALID;
 553                 }
 554         } else {
 555                 /* subsequent middle after valid */
 556                 if (qp->s_ahgidx >= 0) {
 557                         priv->s_ahg->tx_flags |= SDMA_TXREQ_F_USE_AHG;
 558                         priv->s_ahg->ahgidx = qp->s_ahgidx;
 559                         priv->s_ahg->ahgcount++;
 560                         priv->s_ahg->ahgdesc[0] =
 561                                 sdma_build_ahg_descriptor(
 562                                         (__force u16)cpu_to_be16((u16)npsn),
 563                                         BTH2_OFFSET,
 564                                         16,
 565                                         16);
 566                         if ((npsn & 0xffff0000) !=
 567                                         (qp->s_ahgpsn & 0xffff0000)) {
 568                                 priv->s_ahg->ahgcount++;
 569                                 priv->s_ahg->ahgdesc[1] =
 570                                         sdma_build_ahg_descriptor(
 571                                                 (__force u16)cpu_to_be16(
 572                                                         (u16)(npsn >> 16)),
 573                                                 BTH2_OFFSET,
 574                                                 0,
 575                                                 16);
 576                         }
 577                 }
 578         }
 579 }
 580
 581 static inline void hfi1_make_ruc_bth(struct rvt_qp *qp,
 582                                      struct ib_other_headers *ohdr,
 583                                      u32 bth0, u32 bth1, u32 bth2)
 584 {
 585         bth1 |= qp->remote_qpn;
 586         ohdr->bth[0] = cpu_to_be32(bth0);
 587         ohdr->bth[1] = cpu_to_be32(bth1);
 588         ohdr->bth[2] = cpu_to_be32(bth2);
 589 }
 590
 591 /**
 592  * hfi1_make_ruc_header_16B - build a 16B header
 593  * @qp: the queue pair
 594  * @ohdr: a pointer to the destination header memory
 595  * @bth0: bth0 passed in from the RC/UC builder
 596  * @bth2: bth2 passed in from the RC/UC builder
 597  * @middle: non zero implies indicates ahg "could" be used
 598  * @ps: the current packet state
 599  *
 600  * This routine may disarm ahg under these situations:
 601  * - packet needs a GRH
 602  * - BECN needed
 603  * - migration state not IB_MIG_MIGRATED
 604  */
 605 static inline void hfi1_make_ruc_header_16B(struct rvt_qp *qp,
 606                                             struct ib_other_headers *ohdr,
 607                                             u32 bth0, u32 bth2, int middle,
 608                                             struct hfi1_pkt_state *ps)
 609 {
 610         struct hfi1_qp_priv *priv = qp->priv;
 611         struct hfi1_ibport *ibp = ps->ibp;
 612         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 613         u32 bth1 = 0;
 614         u32 slid;
 615         u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 616         u8 l4 = OPA_16B_L4_IB_LOCAL;
 617         u8 extra_bytes = hfi1_get_16b_padding(
 618                                 (ps->s_txreq->hdr_dwords << 2),
 619                                 ps->s_txreq->s_cur_size);
 620         u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
 621                                  extra_bytes + SIZE_OF_LT) >> 2);
 622         bool becn = false;
 623
 624         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
 625             hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
 626                 struct ib_grh *grh;
 627                 struct ib_global_route *grd =
 628                         rdma_ah_retrieve_grh(&qp->remote_ah_attr);
 629                 /*
 630                  * Ensure OPA GIDs are transformed to IB gids
 631                  * before creating the GRH.
 632                  */
 633                 if (grd->sgid_index == OPA_GID_INDEX)
 634                         grd->sgid_index = 0;
 635                 grh = &ps->s_txreq->phdr.hdr.opah.u.l.grh;
 636                 l4 = OPA_16B_L4_IB_GLOBAL;
 637                 ps->s_txreq->hdr_dwords +=
 638                         hfi1_make_grh(ibp, grh, grd,
 639                                       ps->s_txreq->hdr_dwords - LRH_16B_DWORDS,
 640                                       nwords);
 641                 middle = 0;
 642         }
 643
 644         if (qp->s_mig_state == IB_MIG_MIGRATED)
 645                 bth1 |= OPA_BTH_MIG_REQ;
 646         else
 647                 middle = 0;
 648
 649         if (qp->s_flags & RVT_S_ECN) {
 650                 qp->s_flags &= ~RVT_S_ECN;
 651                 /* we recently received a FECN, so return a BECN */
 652                 becn = true;
 653                 middle = 0;
 654         }
 655         if (middle)
 656                 build_ahg(qp, bth2);
 657         else
 658                 qp->s_flags &= ~HFI1_S_AHG_VALID;
 659
 660         bth0 |= pkey;
 661         bth0 |= extra_bytes << 20;
 662         hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
 663
 664         if (!ppd->lid)
 665                 slid = be32_to_cpu(OPA_LID_PERMISSIVE);
 666         else
 667                 slid = ppd->lid |
 668                         (rdma_ah_get_path_bits(&qp->remote_ah_attr) &
 669                         ((1 << ppd->lmc) - 1));
 670
 671         hfi1_make_16b_hdr(&ps->s_txreq->phdr.hdr.opah,
 672                           slid,
 673                           opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
 674                                       16B),
 675                           (ps->s_txreq->hdr_dwords + nwords) >> 1,
 676                           pkey, becn, 0, l4, priv->s_sc);
 677 }
 678
 679 /**
 680  * hfi1_make_ruc_header_9B - build a 9B header
 681  * @qp: the queue pair
 682  * @ohdr: a pointer to the destination header memory
 683  * @bth0: bth0 passed in from the RC/UC builder
 684  * @bth2: bth2 passed in from the RC/UC builder
 685  * @middle: non zero implies indicates ahg "could" be used
 686  * @ps: the current packet state
 687  *
 688  * This routine may disarm ahg under these situations:
 689  * - packet needs a GRH
 690  * - BECN needed
 691  * - migration state not IB_MIG_MIGRATED
 692  */
 693 static inline void hfi1_make_ruc_header_9B(struct rvt_qp *qp,
 694                                            struct ib_other_headers *ohdr,
 695                                            u32 bth0, u32 bth2, int middle,
 696                                            struct hfi1_pkt_state *ps)
 697 {
 698         struct hfi1_qp_priv *priv = qp->priv;
 699         struct hfi1_ibport *ibp = ps->ibp;
 700         u32 bth1 = 0;
 701         u16 pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 702         u16 lrh0 = HFI1_LRH_BTH;
 703         u8 extra_bytes = -ps->s_txreq->s_cur_size & 3;
 704         u32 nwords = SIZE_OF_CRC + ((ps->s_txreq->s_cur_size +
 705                                          extra_bytes) >> 2);
 706
 707         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
 708                 struct ib_grh *grh = &ps->s_txreq->phdr.hdr.ibh.u.l.grh;
 709
 710                 lrh0 = HFI1_LRH_GRH;
 711                 ps->s_txreq->hdr_dwords +=
 712                         hfi1_make_grh(ibp, grh,
 713                                       rdma_ah_read_grh(&qp->remote_ah_attr),
 714                                       ps->s_txreq->hdr_dwords - LRH_9B_DWORDS,
 715                                       nwords);
 716                 middle = 0;
 717         }
 718         lrh0 |= (priv->s_sc & 0xf) << 12 |
 719                 (rdma_ah_get_sl(&qp->remote_ah_attr) & 0xf) << 4;
 720
 721         if (qp->s_mig_state == IB_MIG_MIGRATED)
 722                 bth0 |= IB_BTH_MIG_REQ;
 723         else
 724                 middle = 0;
 725
 726         if (qp->s_flags & RVT_S_ECN) {
 727                 qp->s_flags &= ~RVT_S_ECN;
 728                 /* we recently received a FECN, so return a BECN */
 729                 bth1 |= (IB_BECN_MASK << IB_BECN_SHIFT);
 730                 middle = 0;
 731         }
 732         if (middle)
 733                 build_ahg(qp, bth2);
 734         else
 735                 qp->s_flags &= ~HFI1_S_AHG_VALID;
 736
 737         bth0 |= pkey;
 738         bth0 |= extra_bytes << 20;
 739         hfi1_make_ruc_bth(qp, ohdr, bth0, bth1, bth2);
 740         hfi1_make_ib_hdr(&ps->s_txreq->phdr.hdr.ibh,
 741                          lrh0,
 742                          ps->s_txreq->hdr_dwords + nwords,
 743                          opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
 744                          ppd_from_ibp(ibp)->lid |
 745                                 rdma_ah_get_path_bits(&qp->remote_ah_attr));
 746 }
 747
 748 typedef void (*hfi1_make_ruc_hdr)(struct rvt_qp *qp,
 749                                   struct ib_other_headers *ohdr,
 750                                   u32 bth0, u32 bth2, int middle,
 751                                   struct hfi1_pkt_state *ps);
 752
 753 /* We support only two types - 9B and 16B for now */
 754 static const hfi1_make_ruc_hdr hfi1_ruc_header_tbl[2] = {
 755         [HFI1_PKT_TYPE_9B] = &hfi1_make_ruc_header_9B,
 756         [HFI1_PKT_TYPE_16B] = &hfi1_make_ruc_header_16B
 757 };
 758
 759 void hfi1_make_ruc_header(struct rvt_qp *qp, struct ib_other_headers *ohdr,
 760                           u32 bth0, u32 bth2, int middle,
 761                           struct hfi1_pkt_state *ps)
 762 {
 763         struct hfi1_qp_priv *priv = qp->priv;
 764
 765         /*
 766          * reset s_ahg/AHG fields
 767          *
 768          * This insures that the ahgentry/ahgcount
 769          * are at a non-AHG default to protect
 770          * build_verbs_tx_desc() from using
 771          * an include ahgidx.
 772          *
 773          * build_ahg() will modify as appropriate
 774          * to use the AHG feature.
 775          */
 776         priv->s_ahg->tx_flags = 0;
 777         priv->s_ahg->ahgcount = 0;
 778         priv->s_ahg->ahgidx = 0;
 779
 780         /* Make the appropriate header */
 781         hfi1_ruc_header_tbl[priv->hdr_type](qp, ohdr, bth0, bth2, middle, ps);
 782 }
 783
 784 /* when sending, force a reschedule every one of these periods */
 785 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
 786
 787 /**
 788  * schedule_send_yield - test for a yield required for QP send engine
 789  * @timeout: Final time for timeout slice for jiffies
 790  * @qp: a pointer to QP
 791  * @ps: a pointer to a structure with commonly lookup values for
 792  *      the the send engine progress
 793  *
 794  * This routine checks if the time slice for the QP has expired
 795  * for RC QPs, if so an additional work entry is queued. At this
 796  * point, other QPs have an opportunity to be scheduled. It
 797  * returns true if a yield is required, otherwise, false
 798  * is returned.
 799  */
 800 static bool schedule_send_yield(struct rvt_qp *qp,
 801                                 struct hfi1_pkt_state *ps)
 802 {
 803         ps->pkts_sent = true;
 804
 805         if (unlikely(time_after(jiffies, ps->timeout))) {
 806                 if (!ps->in_thread ||
 807                     workqueue_congested(ps->cpu, ps->ppd->hfi1_wq)) {
 808                         spin_lock_irqsave(&qp->s_lock, ps->flags);
 809                         qp->s_flags &= ~RVT_S_BUSY;
 810                         hfi1_schedule_send(qp);
 811                         spin_unlock_irqrestore(&qp->s_lock, ps->flags);
 812                         this_cpu_inc(*ps->ppd->dd->send_schedule);
 813                         trace_hfi1_rc_expired_time_slice(qp, true);
 814                         return true;
 815                 }
 816
 817                 cond_resched();
 818                 this_cpu_inc(*ps->ppd->dd->send_schedule);
 819                 ps->timeout = jiffies + ps->timeout_int;
 820         }
 821
 822         trace_hfi1_rc_expired_time_slice(qp, false);
 823         return false;
 824 }
 825
 826 void hfi1_do_send_from_rvt(struct rvt_qp *qp)
 827 {
 828         hfi1_do_send(qp, false);
 829 }
 830
 831 void _hfi1_do_send(struct work_struct *work)
 832 {
 833         struct iowait *wait = container_of(work, struct iowait, iowork);
 834         struct rvt_qp *qp = iowait_to_qp(wait);
 835
 836         hfi1_do_send(qp, true);
 837 }
 838
 839 /**
 840  * hfi1_do_send - perform a send on a QP
 841  * @work: contains a pointer to the QP
 842  * @in_thread: true if in a workqueue thread
 843  *
 844  * Process entries in the send work queue until credit or queue is
 845  * exhausted.  Only allow one CPU to send a packet per QP.
 846  * Otherwise, two threads could send packets out of order.
 847  */
 848 void hfi1_do_send(struct rvt_qp *qp, bool in_thread)
 849 {
 850         struct hfi1_pkt_state ps;
 851         struct hfi1_qp_priv *priv = qp->priv;
 852         int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
 853
 854         ps.dev = to_idev(qp->ibqp.device);
 855         ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
 856         ps.ppd = ppd_from_ibp(ps.ibp);
 857         ps.in_thread = in_thread;
 858
 859         trace_hfi1_rc_do_send(qp, in_thread);
 860
 861         switch (qp->ibqp.qp_type) {
 862         case IB_QPT_RC:
 863                 if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
 864                                    ~((1 << ps.ppd->lmc) - 1)) ==
 865                                   ps.ppd->lid)) {
 866                         ruc_loopback(qp);
 867                         return;
 868                 }
 869                 make_req = hfi1_make_rc_req;
 870                 ps.timeout_int = qp->timeout_jiffies;
 871                 break;
 872         case IB_QPT_UC:
 873                 if (!loopback && ((rdma_ah_get_dlid(&qp->remote_ah_attr) &
 874                                    ~((1 << ps.ppd->lmc) - 1)) ==
 875                                   ps.ppd->lid)) {
 876                         ruc_loopback(qp);
 877                         return;
 878                 }
 879                 make_req = hfi1_make_uc_req;
 880                 ps.timeout_int = SEND_RESCHED_TIMEOUT;
 881                 break;
 882         default:
 883                 make_req = hfi1_make_ud_req;
 884                 ps.timeout_int = SEND_RESCHED_TIMEOUT;
 885         }
 886
 887         spin_lock_irqsave(&qp->s_lock, ps.flags);
 888
 889         /* Return if we are already busy processing a work request. */
 890         if (!hfi1_send_ok(qp)) {
 891                 spin_unlock_irqrestore(&qp->s_lock, ps.flags);
 892                 return;
 893         }
 894
 895         qp->s_flags |= RVT_S_BUSY;
 896
 897         ps.timeout_int = ps.timeout_int / 8;
 898         ps.timeout = jiffies + ps.timeout_int;
 899         ps.cpu = priv->s_sde ? priv->s_sde->cpu :
 900                         cpumask_first(cpumask_of_node(ps.ppd->dd->node));
 901         ps.pkts_sent = false;
 902
 903         /* insure a pre-built packet is handled  */
 904         ps.s_txreq = get_waiting_verbs_txreq(qp);
 905         do {
 906                 /* Check for a constructed packet to be sent. */
 907                 if (ps.s_txreq) {
 908                         spin_unlock_irqrestore(&qp->s_lock, ps.flags);
 909                         /*
 910                          * If the packet cannot be sent now, return and
 911                          * the send engine will be woken up later.
 912                          */
 913                         if (hfi1_verbs_send(qp, &ps))
 914                                 return;
 915                         /* allow other tasks to run */
 916                         if (schedule_send_yield(qp, &ps))
 917                                 return;
 918
 919                         spin_lock_irqsave(&qp->s_lock, ps.flags);
 920                 }
 921         } while (make_req(qp, &ps));
 922         iowait_starve_clear(ps.pkts_sent, &priv->s_iowait);
 923         spin_unlock_irqrestore(&qp->s_lock, ps.flags);
 924 }
 925
 926 /*
 927  * This should be called with s_lock held.
 928  */
 929 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
 930                         enum ib_wc_status status)
 931 {
 932         u32 old_last, last;
 933
 934         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 935                 return;
 936
 937         last = qp->s_last;
 938         old_last = last;
 939         trace_hfi1_qp_send_completion(qp, wqe, last);
 940         if (++last >= qp->s_size)
 941                 last = 0;
 942         trace_hfi1_qp_send_completion(qp, wqe, last);
 943         qp->s_last = last;
 944         /* See post_send() */
 945         barrier();
 946         rvt_put_swqe(wqe);
 947         if (qp->ibqp.qp_type == IB_QPT_UD ||
 948             qp->ibqp.qp_type == IB_QPT_SMI ||
 949             qp->ibqp.qp_type == IB_QPT_GSI)
 950                 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
 951
 952         rvt_qp_swqe_complete(qp,
 953                              wqe,
 954                              ib_hfi1_wc_opcode[wqe->wr.opcode],
 955                              status);
 956
 957         if (qp->s_acked == old_last)
 958                 qp->s_acked = last;
 959         if (qp->s_cur == old_last)
 960                 qp->s_cur = last;
 961         if (qp->s_tail == old_last)
 962                 qp->s_tail = last;
 963         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
 964                 qp->s_draining = 0;
 965 }