drivers/infiniband/hw/hfi1/rc.c

   1 /*
   2  * Copyright(c) 2015, 2016 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/io.h>
  49 #include <rdma/rdma_vt.h>
  50 #include <rdma/rdmavt_qp.h>
  51
  52 #include "hfi.h"
  53 #include "qp.h"
  54 #include "verbs_txreq.h"
  55 #include "trace.h"
  56
  57 /* cut down ridiculously long IB macro names */
  58 #define OP(x) RC_OP(x)
  59
  60 static u32 restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe,
  61                        u32 psn, u32 pmtu)
  62 {
  63         u32 len;
  64
  65         len = delta_psn(psn, wqe->psn) * pmtu;
  66         ss->sge = wqe->sg_list[0];
  67         ss->sg_list = wqe->sg_list + 1;
  68         ss->num_sge = wqe->wr.num_sge;
  69         ss->total_len = wqe->length;
  70         rvt_skip_sge(ss, len, false);
  71         return wqe->length - len;
  72 }
  73
  74 /**
  75  * make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
  76  * @dev: the device for this QP
  77  * @qp: a pointer to the QP
  78  * @ohdr: a pointer to the IB header being constructed
  79  * @ps: the xmit packet state
  80  *
  81  * Return 1 if constructed; otherwise, return 0.
  82  * Note that we are in the responder's side of the QP context.
  83  * Note the QP s_lock must be held.
  84  */
  85 static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
  86                        struct ib_other_headers *ohdr,
  87                        struct hfi1_pkt_state *ps)
  88 {
  89         struct rvt_ack_entry *e;
  90         u32 hwords;
  91         u32 len;
  92         u32 bth0;
  93         u32 bth2;
  94         int middle = 0;
  95         u32 pmtu = qp->pmtu;
  96         struct hfi1_qp_priv *priv = qp->priv;
  97
  98         lockdep_assert_held(&qp->s_lock);
  99         /* Don't send an ACK if we aren't supposed to. */
 100         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
 101                 goto bail;
 102
 103         if (priv->hdr_type == HFI1_PKT_TYPE_9B)
 104                 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 105                 hwords = 5;
 106         else
 107                 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
 108                 hwords = 7;
 109
 110         switch (qp->s_ack_state) {
 111         case OP(RDMA_READ_RESPONSE_LAST):
 112         case OP(RDMA_READ_RESPONSE_ONLY):
 113                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 114                 if (e->rdma_sge.mr) {
 115                         rvt_put_mr(e->rdma_sge.mr);
 116                         e->rdma_sge.mr = NULL;
 117                 }
 118                 /* FALLTHROUGH */
 119         case OP(ATOMIC_ACKNOWLEDGE):
 120                 /*
 121                  * We can increment the tail pointer now that the last
 122                  * response has been sent instead of only being
 123                  * constructed.
 124                  */
 125                 if (++qp->s_tail_ack_queue > HFI1_MAX_RDMA_ATOMIC)
 126                         qp->s_tail_ack_queue = 0;
 127                 /* FALLTHROUGH */
 128         case OP(SEND_ONLY):
 129         case OP(ACKNOWLEDGE):
 130                 /* Check for no next entry in the queue. */
 131                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 132                         if (qp->s_flags & RVT_S_ACK_PENDING)
 133                                 goto normal;
 134                         goto bail;
 135                 }
 136
 137                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 138                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 139                         /*
 140                          * If a RDMA read response is being resent and
 141                          * we haven't seen the duplicate request yet,
 142                          * then stop sending the remaining responses the
 143                          * responder has seen until the requester re-sends it.
 144                          */
 145                         len = e->rdma_sge.sge_length;
 146                         if (len && !e->rdma_sge.mr) {
 147                                 qp->s_tail_ack_queue = qp->r_head_ack_queue;
 148                                 goto bail;
 149                         }
 150                         /* Copy SGE state in case we need to resend */
 151                         ps->s_txreq->mr = e->rdma_sge.mr;
 152                         if (ps->s_txreq->mr)
 153                                 rvt_get_mr(ps->s_txreq->mr);
 154                         qp->s_ack_rdma_sge.sge = e->rdma_sge;
 155                         qp->s_ack_rdma_sge.num_sge = 1;
 156                         ps->s_txreq->ss = &qp->s_ack_rdma_sge;
 157                         if (len > pmtu) {
 158                                 len = pmtu;
 159                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 160                         } else {
 161                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 162                                 e->sent = 1;
 163                         }
 164                         ohdr->u.aeth = rvt_compute_aeth(qp);
 165                         hwords++;
 166                         qp->s_ack_rdma_psn = e->psn;
 167                         bth2 = mask_psn(qp->s_ack_rdma_psn++);
 168                 } else {
 169                         /* COMPARE_SWAP or FETCH_ADD */
 170                         ps->s_txreq->ss = NULL;
 171                         len = 0;
 172                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 173                         ohdr->u.at.aeth = rvt_compute_aeth(qp);
 174                         ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
 175                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 176                         bth2 = mask_psn(e->psn);
 177                         e->sent = 1;
 178                 }
 179                 bth0 = qp->s_ack_state << 24;
 180                 break;
 181
 182         case OP(RDMA_READ_RESPONSE_FIRST):
 183                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 184                 /* FALLTHROUGH */
 185         case OP(RDMA_READ_RESPONSE_MIDDLE):
 186                 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
 187                 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
 188                 if (ps->s_txreq->mr)
 189                         rvt_get_mr(ps->s_txreq->mr);
 190                 len = qp->s_ack_rdma_sge.sge.sge_length;
 191                 if (len > pmtu) {
 192                         len = pmtu;
 193                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 194                 } else {
 195                         ohdr->u.aeth = rvt_compute_aeth(qp);
 196                         hwords++;
 197                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 198                         e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 199                         e->sent = 1;
 200                 }
 201                 bth0 = qp->s_ack_state << 24;
 202                 bth2 = mask_psn(qp->s_ack_rdma_psn++);
 203                 break;
 204
 205         default:
 206 normal:
 207                 /*
 208                  * Send a regular ACK.
 209                  * Set the s_ack_state so we wait until after sending
 210                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 211                  * (see above).
 212                  */
 213                 qp->s_ack_state = OP(SEND_ONLY);
 214                 qp->s_flags &= ~RVT_S_ACK_PENDING;
 215                 ps->s_txreq->ss = NULL;
 216                 if (qp->s_nak_state)
 217                         ohdr->u.aeth =
 218                                 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
 219                                             (qp->s_nak_state <<
 220                                              IB_AETH_CREDIT_SHIFT));
 221                 else
 222                         ohdr->u.aeth = rvt_compute_aeth(qp);
 223                 hwords++;
 224                 len = 0;
 225                 bth0 = OP(ACKNOWLEDGE) << 24;
 226                 bth2 = mask_psn(qp->s_ack_psn);
 227         }
 228         qp->s_rdma_ack_cnt++;
 229         qp->s_hdrwords = hwords;
 230         ps->s_txreq->sde = priv->s_sde;
 231         ps->s_txreq->s_cur_size = len;
 232         hfi1_make_ruc_header(qp, ohdr, bth0, bth2, middle, ps);
 233         /* pbc */
 234         ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
 235         return 1;
 236
 237 bail:
 238         qp->s_ack_state = OP(ACKNOWLEDGE);
 239         /*
 240          * Ensure s_rdma_ack_cnt changes are committed prior to resetting
 241          * RVT_S_RESP_PENDING
 242          */
 243         smp_wmb();
 244         qp->s_flags &= ~(RVT_S_RESP_PENDING
 245                                 | RVT_S_ACK_PENDING
 246                                 | RVT_S_AHG_VALID);
 247         return 0;
 248 }
 249
 250 /**
 251  * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 252  * @qp: a pointer to the QP
 253  *
 254  * Assumes s_lock is held.
 255  *
 256  * Return 1 if constructed; otherwise, return 0.
 257  */
 258 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 259 {
 260         struct hfi1_qp_priv *priv = qp->priv;
 261         struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 262         struct ib_other_headers *ohdr;
 263         struct rvt_sge_state *ss;
 264         struct rvt_swqe *wqe;
 265         u32 hwords;
 266         u32 len;
 267         u32 bth0 = 0;
 268         u32 bth2;
 269         u32 pmtu = qp->pmtu;
 270         char newreq;
 271         int middle = 0;
 272         int delta;
 273
 274         lockdep_assert_held(&qp->s_lock);
 275         ps->s_txreq = get_txreq(ps->dev, qp);
 276         if (!ps->s_txreq)
 277                 goto bail_no_tx;
 278
 279         ps->s_txreq->phdr.hdr.hdr_type = priv->hdr_type;
 280         if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
 281                 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 282                 hwords = 5;
 283                 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
 284                         ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
 285                 else
 286                         ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
 287         } else {
 288                 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
 289                 hwords = 7;
 290                 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
 291                     (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
 292                         ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
 293                 else
 294                         ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
 295         }
 296
 297         /* Sending responses has higher priority over sending requests. */
 298         if ((qp->s_flags & RVT_S_RESP_PENDING) &&
 299             make_rc_ack(dev, qp, ohdr, ps))
 300                 return 1;
 301
 302         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
 303                 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
 304                         goto bail;
 305                 /* We are in the error state, flush the work request. */
 306                 smp_read_barrier_depends(); /* see post_one_send() */
 307                 if (qp->s_last == READ_ONCE(qp->s_head))
 308                         goto bail;
 309                 /* If DMAs are in progress, we can't flush immediately. */
 310                 if (iowait_sdma_pending(&priv->s_iowait)) {
 311                         qp->s_flags |= RVT_S_WAIT_DMA;
 312                         goto bail;
 313                 }
 314                 clear_ahg(qp);
 315                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 316                 hfi1_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
 317                         IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
 318                 /* will get called again */
 319                 goto done_free_tx;
 320         }
 321
 322         if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
 323                 goto bail;
 324
 325         if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
 326                 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
 327                         qp->s_flags |= RVT_S_WAIT_PSN;
 328                         goto bail;
 329                 }
 330                 qp->s_sending_psn = qp->s_psn;
 331                 qp->s_sending_hpsn = qp->s_psn - 1;
 332         }
 333
 334         /* Send a request. */
 335         wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
 336         switch (qp->s_state) {
 337         default:
 338                 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
 339                         goto bail;
 340                 /*
 341                  * Resend an old request or start a new one.
 342                  *
 343                  * We keep track of the current SWQE so that
 344                  * we don't reset the "furthest progress" state
 345                  * if we need to back up.
 346                  */
 347                 newreq = 0;
 348                 if (qp->s_cur == qp->s_tail) {
 349                         /* Check if send work queue is empty. */
 350                         smp_read_barrier_depends(); /* see post_one_send() */
 351                         if (qp->s_tail == READ_ONCE(qp->s_head)) {
 352                                 clear_ahg(qp);
 353                                 goto bail;
 354                         }
 355                         /*
 356                          * If a fence is requested, wait for previous
 357                          * RDMA read and atomic operations to finish.
 358                          */
 359                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 360                             qp->s_num_rd_atomic) {
 361                                 qp->s_flags |= RVT_S_WAIT_FENCE;
 362                                 goto bail;
 363                         }
 364                         /*
 365                          * Local operations are processed immediately
 366                          * after all prior requests have completed
 367                          */
 368                         if (wqe->wr.opcode == IB_WR_REG_MR ||
 369                             wqe->wr.opcode == IB_WR_LOCAL_INV) {
 370                                 int local_ops = 0;
 371                                 int err = 0;
 372
 373                                 if (qp->s_last != qp->s_cur)
 374                                         goto bail;
 375                                 if (++qp->s_cur == qp->s_size)
 376                                         qp->s_cur = 0;
 377                                 if (++qp->s_tail == qp->s_size)
 378                                         qp->s_tail = 0;
 379                                 if (!(wqe->wr.send_flags &
 380                                       RVT_SEND_COMPLETION_ONLY)) {
 381                                         err = rvt_invalidate_rkey(
 382                                                 qp,
 383                                                 wqe->wr.ex.invalidate_rkey);
 384                                         local_ops = 1;
 385                                 }
 386                                 hfi1_send_complete(qp, wqe,
 387                                                    err ? IB_WC_LOC_PROT_ERR
 388                                                        : IB_WC_SUCCESS);
 389                                 if (local_ops)
 390                                         atomic_dec(&qp->local_ops_pending);
 391                                 qp->s_hdrwords = 0;
 392                                 goto done_free_tx;
 393                         }
 394
 395                         newreq = 1;
 396                         qp->s_psn = wqe->psn;
 397                 }
 398                 /*
 399                  * Note that we have to be careful not to modify the
 400                  * original work request since we may need to resend
 401                  * it.
 402                  */
 403                 len = wqe->length;
 404                 ss = &qp->s_sge;
 405                 bth2 = mask_psn(qp->s_psn);
 406                 switch (wqe->wr.opcode) {
 407                 case IB_WR_SEND:
 408                 case IB_WR_SEND_WITH_IMM:
 409                 case IB_WR_SEND_WITH_INV:
 410                         /* If no credit, return. */
 411                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 412                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 413                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 414                                 goto bail;
 415                         }
 416                         if (len > pmtu) {
 417                                 qp->s_state = OP(SEND_FIRST);
 418                                 len = pmtu;
 419                                 break;
 420                         }
 421                         if (wqe->wr.opcode == IB_WR_SEND) {
 422                                 qp->s_state = OP(SEND_ONLY);
 423                         } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 424                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 425                                 /* Immediate data comes after the BTH */
 426                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 427                                 hwords += 1;
 428                         } else {
 429                                 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
 430                                 /* Invalidate rkey comes after the BTH */
 431                                 ohdr->u.ieth = cpu_to_be32(
 432                                                 wqe->wr.ex.invalidate_rkey);
 433                                 hwords += 1;
 434                         }
 435                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 436                                 bth0 |= IB_BTH_SOLICITED;
 437                         bth2 |= IB_BTH_REQ_ACK;
 438                         if (++qp->s_cur == qp->s_size)
 439                                 qp->s_cur = 0;
 440                         break;
 441
 442                 case IB_WR_RDMA_WRITE:
 443                         if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 444                                 qp->s_lsn++;
 445                         goto no_flow_control;
 446                 case IB_WR_RDMA_WRITE_WITH_IMM:
 447                         /* If no credit, return. */
 448                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 449                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 450                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 451                                 goto bail;
 452                         }
 453 no_flow_control:
 454                         put_ib_reth_vaddr(
 455                                 wqe->rdma_wr.remote_addr,
 456                                 &ohdr->u.rc.reth);
 457                         ohdr->u.rc.reth.rkey =
 458                                 cpu_to_be32(wqe->rdma_wr.rkey);
 459                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 460                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 461                         if (len > pmtu) {
 462                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 463                                 len = pmtu;
 464                                 break;
 465                         }
 466                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 467                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 468                         } else {
 469                                 qp->s_state =
 470                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 471                                 /* Immediate data comes after RETH */
 472                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 473                                 hwords += 1;
 474                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 475                                         bth0 |= IB_BTH_SOLICITED;
 476                         }
 477                         bth2 |= IB_BTH_REQ_ACK;
 478                         if (++qp->s_cur == qp->s_size)
 479                                 qp->s_cur = 0;
 480                         break;
 481
 482                 case IB_WR_RDMA_READ:
 483                         /*
 484                          * Don't allow more operations to be started
 485                          * than the QP limits allow.
 486                          */
 487                         if (newreq) {
 488                                 if (qp->s_num_rd_atomic >=
 489                                     qp->s_max_rd_atomic) {
 490                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 491                                         goto bail;
 492                                 }
 493                                 qp->s_num_rd_atomic++;
 494                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 495                                         qp->s_lsn++;
 496                         }
 497                         put_ib_reth_vaddr(
 498                                 wqe->rdma_wr.remote_addr,
 499                                 &ohdr->u.rc.reth);
 500                         ohdr->u.rc.reth.rkey =
 501                                 cpu_to_be32(wqe->rdma_wr.rkey);
 502                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 503                         qp->s_state = OP(RDMA_READ_REQUEST);
 504                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 505                         ss = NULL;
 506                         len = 0;
 507                         bth2 |= IB_BTH_REQ_ACK;
 508                         if (++qp->s_cur == qp->s_size)
 509                                 qp->s_cur = 0;
 510                         break;
 511
 512                 case IB_WR_ATOMIC_CMP_AND_SWP:
 513                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 514                         /*
 515                          * Don't allow more operations to be started
 516                          * than the QP limits allow.
 517                          */
 518                         if (newreq) {
 519                                 if (qp->s_num_rd_atomic >=
 520                                     qp->s_max_rd_atomic) {
 521                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 522                                         goto bail;
 523                                 }
 524                                 qp->s_num_rd_atomic++;
 525                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 526                                         qp->s_lsn++;
 527                         }
 528                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 529                                 qp->s_state = OP(COMPARE_SWAP);
 530                                 put_ib_ateth_swap(wqe->atomic_wr.swap,
 531                                                   &ohdr->u.atomic_eth);
 532                                 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
 533                                                      &ohdr->u.atomic_eth);
 534                         } else {
 535                                 qp->s_state = OP(FETCH_ADD);
 536                                 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
 537                                                   &ohdr->u.atomic_eth);
 538                                 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
 539                         }
 540                         put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
 541                                            &ohdr->u.atomic_eth);
 542                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 543                                 wqe->atomic_wr.rkey);
 544                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 545                         ss = NULL;
 546                         len = 0;
 547                         bth2 |= IB_BTH_REQ_ACK;
 548                         if (++qp->s_cur == qp->s_size)
 549                                 qp->s_cur = 0;
 550                         break;
 551
 552                 default:
 553                         goto bail;
 554                 }
 555                 qp->s_sge.sge = wqe->sg_list[0];
 556                 qp->s_sge.sg_list = wqe->sg_list + 1;
 557                 qp->s_sge.num_sge = wqe->wr.num_sge;
 558                 qp->s_sge.total_len = wqe->length;
 559                 qp->s_len = wqe->length;
 560                 if (newreq) {
 561                         qp->s_tail++;
 562                         if (qp->s_tail >= qp->s_size)
 563                                 qp->s_tail = 0;
 564                 }
 565                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 566                         qp->s_psn = wqe->lpsn + 1;
 567                 else
 568                         qp->s_psn++;
 569                 break;
 570
 571         case OP(RDMA_READ_RESPONSE_FIRST):
 572                 /*
 573                  * qp->s_state is normally set to the opcode of the
 574                  * last packet constructed for new requests and therefore
 575                  * is never set to RDMA read response.
 576                  * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
 577                  * thread to indicate a SEND needs to be restarted from an
 578                  * earlier PSN without interfering with the sending thread.
 579                  * See restart_rc().
 580                  */
 581                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 582                 /* FALLTHROUGH */
 583         case OP(SEND_FIRST):
 584                 qp->s_state = OP(SEND_MIDDLE);
 585                 /* FALLTHROUGH */
 586         case OP(SEND_MIDDLE):
 587                 bth2 = mask_psn(qp->s_psn++);
 588                 ss = &qp->s_sge;
 589                 len = qp->s_len;
 590                 if (len > pmtu) {
 591                         len = pmtu;
 592                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 593                         break;
 594                 }
 595                 if (wqe->wr.opcode == IB_WR_SEND) {
 596                         qp->s_state = OP(SEND_LAST);
 597                 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 598                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 599                         /* Immediate data comes after the BTH */
 600                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 601                         hwords += 1;
 602                 } else {
 603                         qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
 604                         /* invalidate data comes after the BTH */
 605                         ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
 606                         hwords += 1;
 607                 }
 608                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 609                         bth0 |= IB_BTH_SOLICITED;
 610                 bth2 |= IB_BTH_REQ_ACK;
 611                 qp->s_cur++;
 612                 if (qp->s_cur >= qp->s_size)
 613                         qp->s_cur = 0;
 614                 break;
 615
 616         case OP(RDMA_READ_RESPONSE_LAST):
 617                 /*
 618                  * qp->s_state is normally set to the opcode of the
 619                  * last packet constructed for new requests and therefore
 620                  * is never set to RDMA read response.
 621                  * RDMA_READ_RESPONSE_LAST is used by the ACK processing
 622                  * thread to indicate a RDMA write needs to be restarted from
 623                  * an earlier PSN without interfering with the sending thread.
 624                  * See restart_rc().
 625                  */
 626                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 627                 /* FALLTHROUGH */
 628         case OP(RDMA_WRITE_FIRST):
 629                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 630                 /* FALLTHROUGH */
 631         case OP(RDMA_WRITE_MIDDLE):
 632                 bth2 = mask_psn(qp->s_psn++);
 633                 ss = &qp->s_sge;
 634                 len = qp->s_len;
 635                 if (len > pmtu) {
 636                         len = pmtu;
 637                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 638                         break;
 639                 }
 640                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 641                         qp->s_state = OP(RDMA_WRITE_LAST);
 642                 } else {
 643                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 644                         /* Immediate data comes after the BTH */
 645                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 646                         hwords += 1;
 647                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 648                                 bth0 |= IB_BTH_SOLICITED;
 649                 }
 650                 bth2 |= IB_BTH_REQ_ACK;
 651                 qp->s_cur++;
 652                 if (qp->s_cur >= qp->s_size)
 653                         qp->s_cur = 0;
 654                 break;
 655
 656         case OP(RDMA_READ_RESPONSE_MIDDLE):
 657                 /*
 658                  * qp->s_state is normally set to the opcode of the
 659                  * last packet constructed for new requests and therefore
 660                  * is never set to RDMA read response.
 661                  * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
 662                  * thread to indicate a RDMA read needs to be restarted from
 663                  * an earlier PSN without interfering with the sending thread.
 664                  * See restart_rc().
 665                  */
 666                 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
 667                 put_ib_reth_vaddr(
 668                         wqe->rdma_wr.remote_addr + len,
 669                         &ohdr->u.rc.reth);
 670                 ohdr->u.rc.reth.rkey =
 671                         cpu_to_be32(wqe->rdma_wr.rkey);
 672                 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
 673                 qp->s_state = OP(RDMA_READ_REQUEST);
 674                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 675                 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
 676                 qp->s_psn = wqe->lpsn + 1;
 677                 ss = NULL;
 678                 len = 0;
 679                 qp->s_cur++;
 680                 if (qp->s_cur == qp->s_size)
 681                         qp->s_cur = 0;
 682                 break;
 683         }
 684         qp->s_sending_hpsn = bth2;
 685         delta = delta_psn(bth2, wqe->psn);
 686         if (delta && delta % HFI1_PSN_CREDIT == 0)
 687                 bth2 |= IB_BTH_REQ_ACK;
 688         if (qp->s_flags & RVT_S_SEND_ONE) {
 689                 qp->s_flags &= ~RVT_S_SEND_ONE;
 690                 qp->s_flags |= RVT_S_WAIT_ACK;
 691                 bth2 |= IB_BTH_REQ_ACK;
 692         }
 693         qp->s_len -= len;
 694         qp->s_hdrwords = hwords;
 695         ps->s_txreq->sde = priv->s_sde;
 696         ps->s_txreq->ss = ss;
 697         ps->s_txreq->s_cur_size = len;
 698         hfi1_make_ruc_header(
 699                 qp,
 700                 ohdr,
 701                 bth0 | (qp->s_state << 24),
 702                 bth2,
 703                 middle,
 704                 ps);
 705         /* pbc */
 706         ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
 707         return 1;
 708
 709 done_free_tx:
 710         hfi1_put_txreq(ps->s_txreq);
 711         ps->s_txreq = NULL;
 712         return 1;
 713
 714 bail:
 715         hfi1_put_txreq(ps->s_txreq);
 716
 717 bail_no_tx:
 718         ps->s_txreq = NULL;
 719         qp->s_flags &= ~RVT_S_BUSY;
 720         qp->s_hdrwords = 0;
 721         return 0;
 722 }
 723
 724 static inline void hfi1_make_bth_aeth(struct rvt_qp *qp,
 725                                       struct ib_other_headers *ohdr,
 726                                       u32 bth0, u32 bth1)
 727 {
 728         if (qp->r_nak_state)
 729                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
 730                                             (qp->r_nak_state <<
 731                                              IB_AETH_CREDIT_SHIFT));
 732         else
 733                 ohdr->u.aeth = rvt_compute_aeth(qp);
 734
 735         ohdr->bth[0] = cpu_to_be32(bth0);
 736         ohdr->bth[1] = cpu_to_be32(bth1 | qp->remote_qpn);
 737         ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
 738 }
 739
 740 static inline void hfi1_queue_rc_ack(struct rvt_qp *qp, bool is_fecn)
 741 {
 742         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 743         unsigned long flags;
 744
 745         spin_lock_irqsave(&qp->s_lock, flags);
 746         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
 747                 goto unlock;
 748         this_cpu_inc(*ibp->rvp.rc_qacks);
 749         qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
 750         qp->s_nak_state = qp->r_nak_state;
 751         qp->s_ack_psn = qp->r_ack_psn;
 752         if (is_fecn)
 753                 qp->s_flags |= RVT_S_ECN;
 754
 755         /* Schedule the send tasklet. */
 756         hfi1_schedule_send(qp);
 757 unlock:
 758         spin_unlock_irqrestore(&qp->s_lock, flags);
 759 }
 760
 761 static inline void hfi1_make_rc_ack_9B(struct rvt_qp *qp,
 762                                        struct hfi1_opa_header *opa_hdr,
 763                                        u8 sc5, bool is_fecn,
 764                                        u64 *pbc_flags, u32 *hwords,
 765                                        u32 *nwords)
 766 {
 767         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 768         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 769         struct ib_header *hdr = &opa_hdr->ibh;
 770         struct ib_other_headers *ohdr;
 771         u16 lrh0 = HFI1_LRH_BTH;
 772         u16 pkey;
 773         u32 bth0, bth1;
 774
 775         opa_hdr->hdr_type = HFI1_PKT_TYPE_9B;
 776         ohdr = &hdr->u.oth;
 777         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
 778         *hwords = 6;
 779
 780         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
 781                 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
 782                                          rdma_ah_read_grh(&qp->remote_ah_attr),
 783                                          *hwords - 2, SIZE_OF_CRC);
 784                 ohdr = &hdr->u.l.oth;
 785                 lrh0 = HFI1_LRH_GRH;
 786         }
 787         /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
 788         *pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
 789
 790         /* read pkey_index w/o lock (its atomic) */
 791         pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 792
 793         lrh0 |= (sc5 & IB_SC_MASK) << IB_SC_SHIFT |
 794                 (rdma_ah_get_sl(&qp->remote_ah_attr) & IB_SL_MASK) <<
 795                         IB_SL_SHIFT;
 796
 797         hfi1_make_ib_hdr(hdr, lrh0, *hwords + SIZE_OF_CRC,
 798                          opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
 799                          ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr));
 800
 801         bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
 802         if (qp->s_mig_state == IB_MIG_MIGRATED)
 803                 bth0 |= IB_BTH_MIG_REQ;
 804         bth1 = (!!is_fecn) << IB_BECN_SHIFT;
 805         hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
 806 }
 807
 808 static inline void hfi1_make_rc_ack_16B(struct rvt_qp *qp,
 809                                         struct hfi1_opa_header *opa_hdr,
 810                                         u8 sc5, bool is_fecn,
 811                                         u64 *pbc_flags, u32 *hwords,
 812                                         u32 *nwords)
 813 {
 814         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 815         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 816         struct hfi1_16b_header *hdr = &opa_hdr->opah;
 817         struct ib_other_headers *ohdr;
 818         u32 bth0, bth1 = 0;
 819         u16 len, pkey;
 820         u8 becn = !!is_fecn;
 821         u8 l4 = OPA_16B_L4_IB_LOCAL;
 822         u8 extra_bytes;
 823
 824         opa_hdr->hdr_type = HFI1_PKT_TYPE_16B;
 825         ohdr = &hdr->u.oth;
 826         /* header size in 32-bit words 16B LRH+BTH+AETH = (16+12+4)/4 */
 827         *hwords = 8;
 828         extra_bytes = hfi1_get_16b_padding(*hwords << 2, 0);
 829         *nwords = SIZE_OF_CRC + ((extra_bytes + SIZE_OF_LT) >> 2);
 830
 831         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
 832             hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
 833                 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
 834                                          rdma_ah_read_grh(&qp->remote_ah_attr),
 835                                          *hwords - 4, *nwords);
 836                 ohdr = &hdr->u.l.oth;
 837                 l4 = OPA_16B_L4_IB_GLOBAL;
 838         }
 839         *pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
 840
 841         /* read pkey_index w/o lock (its atomic) */
 842         pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 843
 844         /* Convert dwords to flits */
 845         len = (*hwords + *nwords) >> 1;
 846
 847         hfi1_make_16b_hdr(hdr,
 848                           ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr),
 849                           opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
 850                                       16B),
 851                           len, pkey, becn, 0, l4, sc5);
 852
 853         bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
 854         bth0 |= extra_bytes << 20;
 855         if (qp->s_mig_state == IB_MIG_MIGRATED)
 856                 bth1 = OPA_BTH_MIG_REQ;
 857         hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
 858 }
 859
 860 typedef void (*hfi1_make_rc_ack)(struct rvt_qp *qp,
 861                                  struct hfi1_opa_header *opa_hdr,
 862                                  u8 sc5, bool is_fecn,
 863                                  u64 *pbc_flags, u32 *hwords,
 864                                  u32 *nwords);
 865
 866 /* We support only two types - 9B and 16B for now */
 867 static const hfi1_make_rc_ack hfi1_make_rc_ack_tbl[2] = {
 868         [HFI1_PKT_TYPE_9B] = &hfi1_make_rc_ack_9B,
 869         [HFI1_PKT_TYPE_16B] = &hfi1_make_rc_ack_16B
 870 };
 871
 872 /**
 873  * hfi1_send_rc_ack - Construct an ACK packet and send it
 874  * @qp: a pointer to the QP
 875  *
 876  * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
 877  * Note that RDMA reads and atomics are handled in the
 878  * send side QP state and send engine.
 879  */
 880 void hfi1_send_rc_ack(struct hfi1_ctxtdata *rcd,
 881                       struct rvt_qp *qp, bool is_fecn)
 882 {
 883         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 884         struct hfi1_qp_priv *priv = qp->priv;
 885         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 886         u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
 887         u64 pbc, pbc_flags = 0;
 888         u32 hwords = 0;
 889         u32 nwords = 0;
 890         u32 plen;
 891         struct pio_buf *pbuf;
 892         struct hfi1_opa_header opa_hdr;
 893
 894         /* clear the defer count */
 895         qp->r_adefered = 0;
 896
 897         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 898         if (qp->s_flags & RVT_S_RESP_PENDING) {
 899                 hfi1_queue_rc_ack(qp, is_fecn);
 900                 return;
 901         }
 902
 903         /* Ensure s_rdma_ack_cnt changes are committed */
 904         smp_read_barrier_depends();
 905         if (qp->s_rdma_ack_cnt) {
 906                 hfi1_queue_rc_ack(qp, is_fecn);
 907                 return;
 908         }
 909
 910         /* Don't try to send ACKs if the link isn't ACTIVE */
 911         if (driver_lstate(ppd) != IB_PORT_ACTIVE)
 912                 return;
 913
 914         /* Make the appropriate header */
 915         hfi1_make_rc_ack_tbl[priv->hdr_type](qp, &opa_hdr, sc5, is_fecn,
 916                                              &pbc_flags, &hwords, &nwords);
 917
 918         plen = 2 /* PBC */ + hwords + nwords;
 919         pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps,
 920                          sc_to_vlt(ppd->dd, sc5), plen);
 921         pbuf = sc_buffer_alloc(rcd->sc, plen, NULL, NULL);
 922         if (!pbuf) {
 923                 /*
 924                  * We have no room to send at the moment.  Pass
 925                  * responsibility for sending the ACK to the send engine
 926                  * so that when enough buffer space becomes available,
 927                  * the ACK is sent ahead of other outgoing packets.
 928                  */
 929                 hfi1_queue_rc_ack(qp, is_fecn);
 930                 return;
 931         }
 932         trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
 933                                &opa_hdr, ib_is_sc5(sc5));
 934
 935         /* write the pbc and data */
 936         ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
 937                                  (priv->hdr_type == HFI1_PKT_TYPE_9B ?
 938                                  (void *)&opa_hdr.ibh :
 939                                  (void *)&opa_hdr.opah), hwords);
 940         return;
 941 }
 942
 943 /**
 944  * reset_psn - reset the QP state to send starting from PSN
 945  * @qp: the QP
 946  * @psn: the packet sequence number to restart at
 947  *
 948  * This is called from hfi1_rc_rcv() to process an incoming RC ACK
 949  * for the given QP.
 950  * Called at interrupt level with the QP s_lock held.
 951  */
 952 static void reset_psn(struct rvt_qp *qp, u32 psn)
 953 {
 954         u32 n = qp->s_acked;
 955         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
 956         u32 opcode;
 957
 958         lockdep_assert_held(&qp->s_lock);
 959         qp->s_cur = n;
 960
 961         /*
 962          * If we are starting the request from the beginning,
 963          * let the normal send code handle initialization.
 964          */
 965         if (cmp_psn(psn, wqe->psn) <= 0) {
 966                 qp->s_state = OP(SEND_LAST);
 967                 goto done;
 968         }
 969
 970         /* Find the work request opcode corresponding to the given PSN. */
 971         opcode = wqe->wr.opcode;
 972         for (;;) {
 973                 int diff;
 974
 975                 if (++n == qp->s_size)
 976                         n = 0;
 977                 if (n == qp->s_tail)
 978                         break;
 979                 wqe = rvt_get_swqe_ptr(qp, n);
 980                 diff = cmp_psn(psn, wqe->psn);
 981                 if (diff < 0)
 982                         break;
 983                 qp->s_cur = n;
 984                 /*
 985                  * If we are starting the request from the beginning,
 986                  * let the normal send code handle initialization.
 987                  */
 988                 if (diff == 0) {
 989                         qp->s_state = OP(SEND_LAST);
 990                         goto done;
 991                 }
 992                 opcode = wqe->wr.opcode;
 993         }
 994
 995         /*
 996          * Set the state to restart in the middle of a request.
 997          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 998          * See hfi1_make_rc_req().
 999          */
1000         switch (opcode) {
1001         case IB_WR_SEND:
1002         case IB_WR_SEND_WITH_IMM:
1003                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
1004                 break;
1005
1006         case IB_WR_RDMA_WRITE:
1007         case IB_WR_RDMA_WRITE_WITH_IMM:
1008                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
1009                 break;
1010
1011         case IB_WR_RDMA_READ:
1012                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
1013                 break;
1014
1015         default:
1016                 /*
1017                  * This case shouldn't happen since its only
1018                  * one PSN per req.
1019                  */
1020                 qp->s_state = OP(SEND_LAST);
1021         }
1022 done:
1023         qp->s_psn = psn;
1024         /*
1025          * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
1026          * asynchronously before the send engine can get scheduled.
1027          * Doing it in hfi1_make_rc_req() is too late.
1028          */
1029         if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
1030             (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
1031                 qp->s_flags |= RVT_S_WAIT_PSN;
1032         qp->s_flags &= ~RVT_S_AHG_VALID;
1033 }
1034
1035 /*
1036  * Back up requester to resend the last un-ACKed request.
1037  * The QP r_lock and s_lock should be held and interrupts disabled.
1038  */
1039 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
1040 {
1041         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1042         struct hfi1_ibport *ibp;
1043
1044         lockdep_assert_held(&qp->r_lock);
1045         lockdep_assert_held(&qp->s_lock);
1046         if (qp->s_retry == 0) {
1047                 if (qp->s_mig_state == IB_MIG_ARMED) {
1048                         hfi1_migrate_qp(qp);
1049                         qp->s_retry = qp->s_retry_cnt;
1050                 } else if (qp->s_last == qp->s_acked) {
1051                         hfi1_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
1052                         rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1053                         return;
1054                 } else { /* need to handle delayed completion */
1055                         return;
1056                 }
1057         } else {
1058                 qp->s_retry--;
1059         }
1060
1061         ibp = to_iport(qp->ibqp.device, qp->port_num);
1062         if (wqe->wr.opcode == IB_WR_RDMA_READ)
1063                 ibp->rvp.n_rc_resends++;
1064         else
1065                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1066
1067         qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
1068                          RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
1069                          RVT_S_WAIT_ACK);
1070         if (wait)
1071                 qp->s_flags |= RVT_S_SEND_ONE;
1072         reset_psn(qp, psn);
1073 }
1074
1075 /*
1076  * Set qp->s_sending_psn to the next PSN after the given one.
1077  * This would be psn+1 except when RDMA reads are present.
1078  */
1079 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
1080 {
1081         struct rvt_swqe *wqe;
1082         u32 n = qp->s_last;
1083
1084         lockdep_assert_held(&qp->s_lock);
1085         /* Find the work request corresponding to the given PSN. */
1086         for (;;) {
1087                 wqe = rvt_get_swqe_ptr(qp, n);
1088                 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1089                         if (wqe->wr.opcode == IB_WR_RDMA_READ)
1090                                 qp->s_sending_psn = wqe->lpsn + 1;
1091                         else
1092                                 qp->s_sending_psn = psn + 1;
1093                         break;
1094                 }
1095                 if (++n == qp->s_size)
1096                         n = 0;
1097                 if (n == qp->s_tail)
1098                         break;
1099         }
1100 }
1101
1102 /*
1103  * This should be called with the QP s_lock held and interrupts disabled.
1104  */
1105 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1106 {
1107         struct ib_other_headers *ohdr;
1108         struct hfi1_qp_priv *priv = qp->priv;
1109         struct rvt_swqe *wqe;
1110         struct ib_header *hdr = NULL;
1111         struct hfi1_16b_header *hdr_16b = NULL;
1112         u32 opcode;
1113         u32 psn;
1114
1115         lockdep_assert_held(&qp->s_lock);
1116         if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
1117                 return;
1118
1119         /* Find out where the BTH is */
1120         if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
1121                 hdr = &opah->ibh;
1122                 if (ib_get_lnh(hdr) == HFI1_LRH_BTH)
1123                         ohdr = &hdr->u.oth;
1124                 else
1125                         ohdr = &hdr->u.l.oth;
1126         } else {
1127                 u8 l4;
1128
1129                 hdr_16b = &opah->opah;
1130                 l4  = hfi1_16B_get_l4(hdr_16b);
1131                 if (l4 == OPA_16B_L4_IB_LOCAL)
1132                         ohdr = &hdr_16b->u.oth;
1133                 else
1134                         ohdr = &hdr_16b->u.l.oth;
1135         }
1136
1137         opcode = ib_bth_get_opcode(ohdr);
1138         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1139             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1140                 WARN_ON(!qp->s_rdma_ack_cnt);
1141                 qp->s_rdma_ack_cnt--;
1142                 return;
1143         }
1144
1145         psn = ib_bth_get_psn(ohdr);
1146         reset_sending_psn(qp, psn);
1147
1148         /*
1149          * Start timer after a packet requesting an ACK has been sent and
1150          * there are still requests that haven't been acked.
1151          */
1152         if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
1153             !(qp->s_flags &
1154                 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1155                 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1156                 rvt_add_retry_timer(qp);
1157
1158         while (qp->s_last != qp->s_acked) {
1159                 u32 s_last;
1160
1161                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1162                 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1163                     cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1164                         break;
1165                 rvt_qp_wqe_unreserve(qp, wqe);
1166                 s_last = qp->s_last;
1167                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1168                 if (++s_last >= qp->s_size)
1169                         s_last = 0;
1170                 qp->s_last = s_last;
1171                 /* see post_send() */
1172                 barrier();
1173                 rvt_put_swqe(wqe);
1174                 rvt_qp_swqe_complete(qp,
1175                                      wqe,
1176                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1177                                      IB_WC_SUCCESS);
1178         }
1179         /*
1180          * If we were waiting for sends to complete before re-sending,
1181          * and they are now complete, restart sending.
1182          */
1183         trace_hfi1_sendcomplete(qp, psn);
1184         if (qp->s_flags & RVT_S_WAIT_PSN &&
1185             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1186                 qp->s_flags &= ~RVT_S_WAIT_PSN;
1187                 qp->s_sending_psn = qp->s_psn;
1188                 qp->s_sending_hpsn = qp->s_psn - 1;
1189                 hfi1_schedule_send(qp);
1190         }
1191 }
1192
1193 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1194 {
1195         qp->s_last_psn = psn;
1196 }
1197
1198 /*
1199  * Generate a SWQE completion.
1200  * This is similar to hfi1_send_complete but has to check to be sure
1201  * that the SGEs are not being referenced if the SWQE is being resent.
1202  */
1203 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1204                                          struct rvt_swqe *wqe,
1205                                          struct hfi1_ibport *ibp)
1206 {
1207         lockdep_assert_held(&qp->s_lock);
1208         /*
1209          * Don't decrement refcount and don't generate a
1210          * completion if the SWQE is being resent until the send
1211          * is finished.
1212          */
1213         if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1214             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1215                 u32 s_last;
1216
1217                 rvt_put_swqe(wqe);
1218                 rvt_qp_wqe_unreserve(qp, wqe);
1219                 s_last = qp->s_last;
1220                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1221                 if (++s_last >= qp->s_size)
1222                         s_last = 0;
1223                 qp->s_last = s_last;
1224                 /* see post_send() */
1225                 barrier();
1226                 rvt_qp_swqe_complete(qp,
1227                                      wqe,
1228                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1229                                      IB_WC_SUCCESS);
1230         } else {
1231                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1232
1233                 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1234                 /*
1235                  * If send progress not running attempt to progress
1236                  * SDMA queue.
1237                  */
1238                 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1239                         struct sdma_engine *engine;
1240                         u8 sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1241                         u8 sc5;
1242
1243                         /* For now use sc to find engine */
1244                         sc5 = ibp->sl_to_sc[sl];
1245                         engine = qp_to_sdma_engine(qp, sc5);
1246                         sdma_engine_progress_schedule(engine);
1247                 }
1248         }
1249
1250         qp->s_retry = qp->s_retry_cnt;
1251         update_last_psn(qp, wqe->lpsn);
1252
1253         /*
1254          * If we are completing a request which is in the process of
1255          * being resent, we can stop re-sending it since we know the
1256          * responder has already seen it.
1257          */
1258         if (qp->s_acked == qp->s_cur) {
1259                 if (++qp->s_cur >= qp->s_size)
1260                         qp->s_cur = 0;
1261                 qp->s_acked = qp->s_cur;
1262                 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1263                 if (qp->s_acked != qp->s_tail) {
1264                         qp->s_state = OP(SEND_LAST);
1265                         qp->s_psn = wqe->psn;
1266                 }
1267         } else {
1268                 if (++qp->s_acked >= qp->s_size)
1269                         qp->s_acked = 0;
1270                 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1271                         qp->s_draining = 0;
1272                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1273         }
1274         return wqe;
1275 }
1276
1277 /**
1278  * do_rc_ack - process an incoming RC ACK
1279  * @qp: the QP the ACK came in on
1280  * @psn: the packet sequence number of the ACK
1281  * @opcode: the opcode of the request that resulted in the ACK
1282  *
1283  * This is called from rc_rcv_resp() to process an incoming RC ACK
1284  * for the given QP.
1285  * May be called at interrupt level, with the QP s_lock held.
1286  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1287  */
1288 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1289                      u64 val, struct hfi1_ctxtdata *rcd)
1290 {
1291         struct hfi1_ibport *ibp;
1292         enum ib_wc_status status;
1293         struct rvt_swqe *wqe;
1294         int ret = 0;
1295         u32 ack_psn;
1296         int diff;
1297
1298         lockdep_assert_held(&qp->s_lock);
1299         /*
1300          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1301          * requests and implicitly NAK RDMA read and atomic requests issued
1302          * before the NAK'ed request.  The MSN won't include the NAK'ed
1303          * request but will include an ACK'ed request(s).
1304          */
1305         ack_psn = psn;
1306         if (aeth >> IB_AETH_NAK_SHIFT)
1307                 ack_psn--;
1308         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1309         ibp = rcd_to_iport(rcd);
1310
1311         /*
1312          * The MSN might be for a later WQE than the PSN indicates so
1313          * only complete WQEs that the PSN finishes.
1314          */
1315         while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
1316                 /*
1317                  * RDMA_READ_RESPONSE_ONLY is a special case since
1318                  * we want to generate completion events for everything
1319                  * before the RDMA read, copy the data, then generate
1320                  * the completion for the read.
1321                  */
1322                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1323                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1324                     diff == 0) {
1325                         ret = 1;
1326                         goto bail_stop;
1327                 }
1328                 /*
1329                  * If this request is a RDMA read or atomic, and the ACK is
1330                  * for a later operation, this ACK NAKs the RDMA read or
1331                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1332                  * can ACK a RDMA read and likewise for atomic ops.  Note
1333                  * that the NAK case can only happen if relaxed ordering is
1334                  * used and requests are sent after an RDMA read or atomic
1335                  * is sent but before the response is received.
1336                  */
1337                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1338                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1339                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1340                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1341                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1342                         /* Retry this request. */
1343                         if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1344                                 qp->r_flags |= RVT_R_RDMAR_SEQ;
1345                                 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1346                                 if (list_empty(&qp->rspwait)) {
1347                                         qp->r_flags |= RVT_R_RSP_SEND;
1348                                         rvt_get_qp(qp);
1349                                         list_add_tail(&qp->rspwait,
1350                                                       &rcd->qp_wait_list);
1351                                 }
1352                         }
1353                         /*
1354                          * No need to process the ACK/NAK since we are
1355                          * restarting an earlier request.
1356                          */
1357                         goto bail_stop;
1358                 }
1359                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1360                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1361                         u64 *vaddr = wqe->sg_list[0].vaddr;
1362                         *vaddr = val;
1363                 }
1364                 if (qp->s_num_rd_atomic &&
1365                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
1366                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1367                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1368                         qp->s_num_rd_atomic--;
1369                         /* Restart sending task if fence is complete */
1370                         if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1371                             !qp->s_num_rd_atomic) {
1372                                 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1373                                                  RVT_S_WAIT_ACK);
1374                                 hfi1_schedule_send(qp);
1375                         } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1376                                 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1377                                                  RVT_S_WAIT_ACK);
1378                                 hfi1_schedule_send(qp);
1379                         }
1380                 }
1381                 wqe = do_rc_completion(qp, wqe, ibp);
1382                 if (qp->s_acked == qp->s_tail)
1383                         break;
1384         }
1385
1386         switch (aeth >> IB_AETH_NAK_SHIFT) {
1387         case 0:         /* ACK */
1388                 this_cpu_inc(*ibp->rvp.rc_acks);
1389                 if (qp->s_acked != qp->s_tail) {
1390                         /*
1391                          * We are expecting more ACKs so
1392                          * mod the retry timer.
1393                          */
1394                         rvt_mod_retry_timer(qp);
1395                         /*
1396                          * We can stop re-sending the earlier packets and
1397                          * continue with the next packet the receiver wants.
1398                          */
1399                         if (cmp_psn(qp->s_psn, psn) <= 0)
1400                                 reset_psn(qp, psn + 1);
1401                 } else {
1402                         /* No more acks - kill all timers */
1403                         rvt_stop_rc_timers(qp);
1404                         if (cmp_psn(qp->s_psn, psn) <= 0) {
1405                                 qp->s_state = OP(SEND_LAST);
1406                                 qp->s_psn = psn + 1;
1407                         }
1408                 }
1409                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1410                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1411                         hfi1_schedule_send(qp);
1412                 }
1413                 rvt_get_credit(qp, aeth);
1414                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1415                 qp->s_retry = qp->s_retry_cnt;
1416                 update_last_psn(qp, psn);
1417                 return 1;
1418
1419         case 1:         /* RNR NAK */
1420                 ibp->rvp.n_rnr_naks++;
1421                 if (qp->s_acked == qp->s_tail)
1422                         goto bail_stop;
1423                 if (qp->s_flags & RVT_S_WAIT_RNR)
1424                         goto bail_stop;
1425                 if (qp->s_rnr_retry == 0) {
1426                         status = IB_WC_RNR_RETRY_EXC_ERR;
1427                         goto class_b;
1428                 }
1429                 if (qp->s_rnr_retry_cnt < 7)
1430                         qp->s_rnr_retry--;
1431
1432                 /* The last valid PSN is the previous PSN. */
1433                 update_last_psn(qp, psn - 1);
1434
1435                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1436
1437                 reset_psn(qp, psn);
1438
1439                 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1440                 rvt_stop_rc_timers(qp);
1441                 rvt_add_rnr_timer(qp, aeth);
1442                 return 0;
1443
1444         case 3:         /* NAK */
1445                 if (qp->s_acked == qp->s_tail)
1446                         goto bail_stop;
1447                 /* The last valid PSN is the previous PSN. */
1448                 update_last_psn(qp, psn - 1);
1449                 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1450                         IB_AETH_CREDIT_MASK) {
1451                 case 0: /* PSN sequence error */
1452                         ibp->rvp.n_seq_naks++;
1453                         /*
1454                          * Back up to the responder's expected PSN.
1455                          * Note that we might get a NAK in the middle of an
1456                          * RDMA READ response which terminates the RDMA
1457                          * READ.
1458                          */
1459                         hfi1_restart_rc(qp, psn, 0);
1460                         hfi1_schedule_send(qp);
1461                         break;
1462
1463                 case 1: /* Invalid Request */
1464                         status = IB_WC_REM_INV_REQ_ERR;
1465                         ibp->rvp.n_other_naks++;
1466                         goto class_b;
1467
1468                 case 2: /* Remote Access Error */
1469                         status = IB_WC_REM_ACCESS_ERR;
1470                         ibp->rvp.n_other_naks++;
1471                         goto class_b;
1472
1473                 case 3: /* Remote Operation Error */
1474                         status = IB_WC_REM_OP_ERR;
1475                         ibp->rvp.n_other_naks++;
1476 class_b:
1477                         if (qp->s_last == qp->s_acked) {
1478                                 hfi1_send_complete(qp, wqe, status);
1479                                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1480                         }
1481                         break;
1482
1483                 default:
1484                         /* Ignore other reserved NAK error codes */
1485                         goto reserved;
1486                 }
1487                 qp->s_retry = qp->s_retry_cnt;
1488                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1489                 goto bail_stop;
1490
1491         default:                /* 2: reserved */
1492 reserved:
1493                 /* Ignore reserved NAK codes. */
1494                 goto bail_stop;
1495         }
1496         /* cannot be reached  */
1497 bail_stop:
1498         rvt_stop_rc_timers(qp);
1499         return ret;
1500 }
1501
1502 /*
1503  * We have seen an out of sequence RDMA read middle or last packet.
1504  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1505  */
1506 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
1507                          struct hfi1_ctxtdata *rcd)
1508 {
1509         struct rvt_swqe *wqe;
1510
1511         lockdep_assert_held(&qp->s_lock);
1512         /* Remove QP from retry timer */
1513         rvt_stop_rc_timers(qp);
1514
1515         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1516
1517         while (cmp_psn(psn, wqe->lpsn) > 0) {
1518                 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1519                     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1520                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1521                         break;
1522                 wqe = do_rc_completion(qp, wqe, ibp);
1523         }
1524
1525         ibp->rvp.n_rdma_seq++;
1526         qp->r_flags |= RVT_R_RDMAR_SEQ;
1527         hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1528         if (list_empty(&qp->rspwait)) {
1529                 qp->r_flags |= RVT_R_RSP_SEND;
1530                 rvt_get_qp(qp);
1531                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1532         }
1533 }
1534
1535 /**
1536  * rc_rcv_resp - process an incoming RC response packet
1537  * @packet: data packet information
1538  *
1539  * This is called from hfi1_rc_rcv() to process an incoming RC response
1540  * packet for the given QP.
1541  * Called at interrupt level.
1542  */
1543 static void rc_rcv_resp(struct hfi1_packet *packet)
1544 {
1545         struct hfi1_ctxtdata *rcd = packet->rcd;
1546         void *data = packet->payload;
1547         u32 tlen = packet->tlen;
1548         struct rvt_qp *qp = packet->qp;
1549         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1550         struct ib_other_headers *ohdr = packet->ohdr;
1551         struct rvt_swqe *wqe;
1552         enum ib_wc_status status;
1553         unsigned long flags;
1554         int diff;
1555         u64 val;
1556         u32 aeth;
1557         u32 psn = ib_bth_get_psn(packet->ohdr);
1558         u32 pmtu = qp->pmtu;
1559         u16 hdrsize = packet->hlen;
1560         u8 opcode = packet->opcode;
1561         u8 pad = packet->pad;
1562         u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
1563
1564         spin_lock_irqsave(&qp->s_lock, flags);
1565         trace_hfi1_ack(qp, psn);
1566
1567         /* Ignore invalid responses. */
1568         smp_read_barrier_depends(); /* see post_one_send */
1569         if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1570                 goto ack_done;
1571
1572         /* Ignore duplicate responses. */
1573         diff = cmp_psn(psn, qp->s_last_psn);
1574         if (unlikely(diff <= 0)) {
1575                 /* Update credits for "ghost" ACKs */
1576                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1577                         aeth = be32_to_cpu(ohdr->u.aeth);
1578                         if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1579                                 rvt_get_credit(qp, aeth);
1580                 }
1581                 goto ack_done;
1582         }
1583
1584         /*
1585          * Skip everything other than the PSN we expect, if we are waiting
1586          * for a reply to a restarted RDMA read or atomic op.
1587          */
1588         if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1589                 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
1590                         goto ack_done;
1591                 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1592         }
1593
1594         if (unlikely(qp->s_acked == qp->s_tail))
1595                 goto ack_done;
1596         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1597         status = IB_WC_SUCCESS;
1598
1599         switch (opcode) {
1600         case OP(ACKNOWLEDGE):
1601         case OP(ATOMIC_ACKNOWLEDGE):
1602         case OP(RDMA_READ_RESPONSE_FIRST):
1603                 aeth = be32_to_cpu(ohdr->u.aeth);
1604                 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1605                         val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1606                 else
1607                         val = 0;
1608                 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1609                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1610                         goto ack_done;
1611                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1612                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1613                         goto ack_op_err;
1614                 /*
1615                  * If this is a response to a resent RDMA read, we
1616                  * have to be careful to copy the data to the right
1617                  * location.
1618                  */
1619                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1620                                                   wqe, psn, pmtu);
1621                 goto read_middle;
1622
1623         case OP(RDMA_READ_RESPONSE_MIDDLE):
1624                 /* no AETH, no ACK */
1625                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1626                         goto ack_seq_err;
1627                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1628                         goto ack_op_err;
1629 read_middle:
1630                 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
1631                         goto ack_len_err;
1632                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1633                         goto ack_len_err;
1634
1635                 /*
1636                  * We got a response so update the timeout.
1637                  * 4.096 usec. * (1 << qp->timeout)
1638                  */
1639                 rvt_mod_retry_timer(qp);
1640                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1641                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1642                         hfi1_schedule_send(qp);
1643                 }
1644
1645                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1646                         qp->s_retry = qp->s_retry_cnt;
1647
1648                 /*
1649                  * Update the RDMA receive state but do the copy w/o
1650                  * holding the locks and blocking interrupts.
1651                  */
1652                 qp->s_rdma_read_len -= pmtu;
1653                 update_last_psn(qp, psn);
1654                 spin_unlock_irqrestore(&qp->s_lock, flags);
1655                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, false, false);
1656                 goto bail;
1657
1658         case OP(RDMA_READ_RESPONSE_ONLY):
1659                 aeth = be32_to_cpu(ohdr->u.aeth);
1660                 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1661                         goto ack_done;
1662                 /*
1663                  * Check that the data size is >= 0 && <= pmtu.
1664                  * Remember to account for ICRC (4).
1665                  */
1666                 if (unlikely(tlen < (hdrsize + extra_bytes)))
1667                         goto ack_len_err;
1668                 /*
1669                  * If this is a response to a resent RDMA read, we
1670                  * have to be careful to copy the data to the right
1671                  * location.
1672                  */
1673                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1674                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1675                                                   wqe, psn, pmtu);
1676                 goto read_last;
1677
1678         case OP(RDMA_READ_RESPONSE_LAST):
1679                 /* ACKs READ req. */
1680                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1681                         goto ack_seq_err;
1682                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1683                         goto ack_op_err;
1684                 /*
1685                  * Check that the data size is >= 1 && <= pmtu.
1686                  * Remember to account for ICRC (4).
1687                  */
1688                 if (unlikely(tlen <= (hdrsize + extra_bytes)))
1689                         goto ack_len_err;
1690 read_last:
1691                 tlen -= hdrsize + extra_bytes;
1692                 if (unlikely(tlen != qp->s_rdma_read_len))
1693                         goto ack_len_err;
1694                 aeth = be32_to_cpu(ohdr->u.aeth);
1695                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, false, false);
1696                 WARN_ON(qp->s_rdma_read_sge.num_sge);
1697                 (void)do_rc_ack(qp, aeth, psn,
1698                                  OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1699                 goto ack_done;
1700         }
1701
1702 ack_op_err:
1703         status = IB_WC_LOC_QP_OP_ERR;
1704         goto ack_err;
1705
1706 ack_seq_err:
1707         rdma_seq_err(qp, ibp, psn, rcd);
1708         goto ack_done;
1709
1710 ack_len_err:
1711         status = IB_WC_LOC_LEN_ERR;
1712 ack_err:
1713         if (qp->s_last == qp->s_acked) {
1714                 hfi1_send_complete(qp, wqe, status);
1715                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1716         }
1717 ack_done:
1718         spin_unlock_irqrestore(&qp->s_lock, flags);
1719 bail:
1720         return;
1721 }
1722
1723 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
1724                                   struct rvt_qp *qp)
1725 {
1726         if (list_empty(&qp->rspwait)) {
1727                 qp->r_flags |= RVT_R_RSP_NAK;
1728                 rvt_get_qp(qp);
1729                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1730         }
1731 }
1732
1733 static inline void rc_cancel_ack(struct rvt_qp *qp)
1734 {
1735         qp->r_adefered = 0;
1736         if (list_empty(&qp->rspwait))
1737                 return;
1738         list_del_init(&qp->rspwait);
1739         qp->r_flags &= ~RVT_R_RSP_NAK;
1740         rvt_put_qp(qp);
1741 }
1742
1743 /**
1744  * rc_rcv_error - process an incoming duplicate or error RC packet
1745  * @ohdr: the other headers for this packet
1746  * @data: the packet data
1747  * @qp: the QP for this packet
1748  * @opcode: the opcode for this packet
1749  * @psn: the packet sequence number for this packet
1750  * @diff: the difference between the PSN and the expected PSN
1751  *
1752  * This is called from hfi1_rc_rcv() to process an unexpected
1753  * incoming RC packet for the given QP.
1754  * Called at interrupt level.
1755  * Return 1 if no more processing is needed; otherwise return 0 to
1756  * schedule a response to be sent.
1757  */
1758 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
1759                                  struct rvt_qp *qp, u32 opcode, u32 psn,
1760                                  int diff, struct hfi1_ctxtdata *rcd)
1761 {
1762         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1763         struct rvt_ack_entry *e;
1764         unsigned long flags;
1765         u8 i, prev;
1766         int old_req;
1767
1768         trace_hfi1_rcv_error(qp, psn);
1769         if (diff > 0) {
1770                 /*
1771                  * Packet sequence error.
1772                  * A NAK will ACK earlier sends and RDMA writes.
1773                  * Don't queue the NAK if we already sent one.
1774                  */
1775                 if (!qp->r_nak_state) {
1776                         ibp->rvp.n_rc_seqnak++;
1777                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1778                         /* Use the expected PSN. */
1779                         qp->r_ack_psn = qp->r_psn;
1780                         /*
1781                          * Wait to send the sequence NAK until all packets
1782                          * in the receive queue have been processed.
1783                          * Otherwise, we end up propagating congestion.
1784                          */
1785                         rc_defered_ack(rcd, qp);
1786                 }
1787                 goto done;
1788         }
1789
1790         /*
1791          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1792          * write or atomic op.  Don't NAK errors, just silently drop
1793          * the duplicate request.  Note that r_sge, r_len, and
1794          * r_rcv_len may be in use so don't modify them.
1795          *
1796          * We are supposed to ACK the earliest duplicate PSN but we
1797          * can coalesce an outstanding duplicate ACK.  We have to
1798          * send the earliest so that RDMA reads can be restarted at
1799          * the requester's expected PSN.
1800          *
1801          * First, find where this duplicate PSN falls within the
1802          * ACKs previously sent.
1803          * old_req is true if there is an older response that is scheduled
1804          * to be sent before sending this one.
1805          */
1806         e = NULL;
1807         old_req = 1;
1808         ibp->rvp.n_rc_dupreq++;
1809
1810         spin_lock_irqsave(&qp->s_lock, flags);
1811
1812         for (i = qp->r_head_ack_queue; ; i = prev) {
1813                 if (i == qp->s_tail_ack_queue)
1814                         old_req = 0;
1815                 if (i)
1816                         prev = i - 1;
1817                 else
1818                         prev = HFI1_MAX_RDMA_ATOMIC;
1819                 if (prev == qp->r_head_ack_queue) {
1820                         e = NULL;
1821                         break;
1822                 }
1823                 e = &qp->s_ack_queue[prev];
1824                 if (!e->opcode) {
1825                         e = NULL;
1826                         break;
1827                 }
1828                 if (cmp_psn(psn, e->psn) >= 0) {
1829                         if (prev == qp->s_tail_ack_queue &&
1830                             cmp_psn(psn, e->lpsn) <= 0)
1831                                 old_req = 0;
1832                         break;
1833                 }
1834         }
1835         switch (opcode) {
1836         case OP(RDMA_READ_REQUEST): {
1837                 struct ib_reth *reth;
1838                 u32 offset;
1839                 u32 len;
1840
1841                 /*
1842                  * If we didn't find the RDMA read request in the ack queue,
1843                  * we can ignore this request.
1844                  */
1845                 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1846                         goto unlock_done;
1847                 /* RETH comes after BTH */
1848                 reth = &ohdr->u.rc.reth;
1849                 /*
1850                  * Address range must be a subset of the original
1851                  * request and start on pmtu boundaries.
1852                  * We reuse the old ack_queue slot since the requester
1853                  * should not back up and request an earlier PSN for the
1854                  * same request.
1855                  */
1856                 offset = delta_psn(psn, e->psn) * qp->pmtu;
1857                 len = be32_to_cpu(reth->length);
1858                 if (unlikely(offset + len != e->rdma_sge.sge_length))
1859                         goto unlock_done;
1860                 if (e->rdma_sge.mr) {
1861                         rvt_put_mr(e->rdma_sge.mr);
1862                         e->rdma_sge.mr = NULL;
1863                 }
1864                 if (len != 0) {
1865                         u32 rkey = be32_to_cpu(reth->rkey);
1866                         u64 vaddr = get_ib_reth_vaddr(reth);
1867                         int ok;
1868
1869                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1870                                          IB_ACCESS_REMOTE_READ);
1871                         if (unlikely(!ok))
1872                                 goto unlock_done;
1873                 } else {
1874                         e->rdma_sge.vaddr = NULL;
1875                         e->rdma_sge.length = 0;
1876                         e->rdma_sge.sge_length = 0;
1877                 }
1878                 e->psn = psn;
1879                 if (old_req)
1880                         goto unlock_done;
1881                 qp->s_tail_ack_queue = prev;
1882                 break;
1883         }
1884
1885         case OP(COMPARE_SWAP):
1886         case OP(FETCH_ADD): {
1887                 /*
1888                  * If we didn't find the atomic request in the ack queue
1889                  * or the send engine is already backed up to send an
1890                  * earlier entry, we can ignore this request.
1891                  */
1892                 if (!e || e->opcode != (u8)opcode || old_req)
1893                         goto unlock_done;
1894                 qp->s_tail_ack_queue = prev;
1895                 break;
1896         }
1897
1898         default:
1899                 /*
1900                  * Ignore this operation if it doesn't request an ACK
1901                  * or an earlier RDMA read or atomic is going to be resent.
1902                  */
1903                 if (!(psn & IB_BTH_REQ_ACK) || old_req)
1904                         goto unlock_done;
1905                 /*
1906                  * Resend the most recent ACK if this request is
1907                  * after all the previous RDMA reads and atomics.
1908                  */
1909                 if (i == qp->r_head_ack_queue) {
1910                         spin_unlock_irqrestore(&qp->s_lock, flags);
1911                         qp->r_nak_state = 0;
1912                         qp->r_ack_psn = qp->r_psn - 1;
1913                         goto send_ack;
1914                 }
1915
1916                 /*
1917                  * Resend the RDMA read or atomic op which
1918                  * ACKs this duplicate request.
1919                  */
1920                 qp->s_tail_ack_queue = i;
1921                 break;
1922         }
1923         qp->s_ack_state = OP(ACKNOWLEDGE);
1924         qp->s_flags |= RVT_S_RESP_PENDING;
1925         qp->r_nak_state = 0;
1926         hfi1_schedule_send(qp);
1927
1928 unlock_done:
1929         spin_unlock_irqrestore(&qp->s_lock, flags);
1930 done:
1931         return 1;
1932
1933 send_ack:
1934         return 0;
1935 }
1936
1937 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n)
1938 {
1939         unsigned next;
1940
1941         next = n + 1;
1942         if (next > HFI1_MAX_RDMA_ATOMIC)
1943                 next = 0;
1944         qp->s_tail_ack_queue = next;
1945         qp->s_ack_state = OP(ACKNOWLEDGE);
1946 }
1947
1948 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
1949                           u32 lqpn, u32 rqpn, u8 svc_type)
1950 {
1951         struct opa_hfi1_cong_log_event_internal *cc_event;
1952         unsigned long flags;
1953
1954         if (sl >= OPA_MAX_SLS)
1955                 return;
1956
1957         spin_lock_irqsave(&ppd->cc_log_lock, flags);
1958
1959         ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
1960         ppd->threshold_event_counter++;
1961
1962         cc_event = &ppd->cc_events[ppd->cc_log_idx++];
1963         if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
1964                 ppd->cc_log_idx = 0;
1965         cc_event->lqpn = lqpn & RVT_QPN_MASK;
1966         cc_event->rqpn = rqpn & RVT_QPN_MASK;
1967         cc_event->sl = sl;
1968         cc_event->svc_type = svc_type;
1969         cc_event->rlid = rlid;
1970         /* keep timestamp in units of 1.024 usec */
1971         cc_event->timestamp = ktime_to_ns(ktime_get()) / 1024;
1972
1973         spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
1974 }
1975
1976 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, u32 lqpn,
1977                   u32 rqpn, u8 svc_type)
1978 {
1979         struct cca_timer *cca_timer;
1980         u16 ccti, ccti_incr, ccti_timer, ccti_limit;
1981         u8 trigger_threshold;
1982         struct cc_state *cc_state;
1983         unsigned long flags;
1984
1985         if (sl >= OPA_MAX_SLS)
1986                 return;
1987
1988         cc_state = get_cc_state(ppd);
1989
1990         if (!cc_state)
1991                 return;
1992
1993         /*
1994          * 1) increase CCTI (for this SL)
1995          * 2) select IPG (i.e., call set_link_ipg())
1996          * 3) start timer
1997          */
1998         ccti_limit = cc_state->cct.ccti_limit;
1999         ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
2000         ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
2001         trigger_threshold =
2002                 cc_state->cong_setting.entries[sl].trigger_threshold;
2003
2004         spin_lock_irqsave(&ppd->cca_timer_lock, flags);
2005
2006         cca_timer = &ppd->cca_timer[sl];
2007         if (cca_timer->ccti < ccti_limit) {
2008                 if (cca_timer->ccti + ccti_incr <= ccti_limit)
2009                         cca_timer->ccti += ccti_incr;
2010                 else
2011                         cca_timer->ccti = ccti_limit;
2012                 set_link_ipg(ppd);
2013         }
2014
2015         ccti = cca_timer->ccti;
2016
2017         if (!hrtimer_active(&cca_timer->hrtimer)) {
2018                 /* ccti_timer is in units of 1.024 usec */
2019                 unsigned long nsec = 1024 * ccti_timer;
2020
2021                 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
2022                               HRTIMER_MODE_REL);
2023         }
2024
2025         spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
2026
2027         if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
2028                 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
2029 }
2030
2031 /**
2032  * hfi1_rc_rcv - process an incoming RC packet
2033  * @packet: data packet information
2034  *
2035  * This is called from qp_rcv() to process an incoming RC packet
2036  * for the given QP.
2037  * May be called at interrupt level.
2038  */
2039 void hfi1_rc_rcv(struct hfi1_packet *packet)
2040 {
2041         struct hfi1_ctxtdata *rcd = packet->rcd;
2042         void *data = packet->payload;
2043         u32 tlen = packet->tlen;
2044         struct rvt_qp *qp = packet->qp;
2045         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2046         struct ib_other_headers *ohdr = packet->ohdr;
2047         u32 bth0 = be32_to_cpu(ohdr->bth[0]);
2048         u32 opcode = packet->opcode;
2049         u32 hdrsize = packet->hlen;
2050         u32 psn = ib_bth_get_psn(packet->ohdr);
2051         u32 pad = packet->pad;
2052         struct ib_wc wc;
2053         u32 pmtu = qp->pmtu;
2054         int diff;
2055         struct ib_reth *reth;
2056         unsigned long flags;
2057         int ret;
2058         bool is_fecn = false;
2059         bool copy_last = false;
2060         u32 rkey;
2061         u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2062
2063         lockdep_assert_held(&qp->r_lock);
2064
2065         if (hfi1_ruc_check_hdr(ibp, packet))
2066                 return;
2067
2068         is_fecn = process_ecn(qp, packet, false);
2069
2070         /*
2071          * Process responses (ACKs) before anything else.  Note that the
2072          * packet sequence number will be for something in the send work
2073          * queue rather than the expected receive packet sequence number.
2074          * In other words, this QP is the requester.
2075          */
2076         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
2077             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
2078                 rc_rcv_resp(packet);
2079                 if (is_fecn)
2080                         goto send_ack;
2081                 return;
2082         }
2083
2084         /* Compute 24 bits worth of difference. */
2085         diff = delta_psn(psn, qp->r_psn);
2086         if (unlikely(diff)) {
2087                 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
2088                         return;
2089                 goto send_ack;
2090         }
2091
2092         /* Check for opcode sequence errors. */
2093         switch (qp->r_state) {
2094         case OP(SEND_FIRST):
2095         case OP(SEND_MIDDLE):
2096                 if (opcode == OP(SEND_MIDDLE) ||
2097                     opcode == OP(SEND_LAST) ||
2098                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2099                     opcode == OP(SEND_LAST_WITH_INVALIDATE))
2100                         break;
2101                 goto nack_inv;
2102
2103         case OP(RDMA_WRITE_FIRST):
2104         case OP(RDMA_WRITE_MIDDLE):
2105                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
2106                     opcode == OP(RDMA_WRITE_LAST) ||
2107                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2108                         break;
2109                 goto nack_inv;
2110
2111         default:
2112                 if (opcode == OP(SEND_MIDDLE) ||
2113                     opcode == OP(SEND_LAST) ||
2114                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2115                     opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
2116                     opcode == OP(RDMA_WRITE_MIDDLE) ||
2117                     opcode == OP(RDMA_WRITE_LAST) ||
2118                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2119                         goto nack_inv;
2120                 /*
2121                  * Note that it is up to the requester to not send a new
2122                  * RDMA read or atomic operation before receiving an ACK
2123                  * for the previous operation.
2124                  */
2125                 break;
2126         }
2127
2128         if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2129                 rvt_comm_est(qp);
2130
2131         /* OK, process the packet. */
2132         switch (opcode) {
2133         case OP(SEND_FIRST):
2134                 ret = hfi1_rvt_get_rwqe(qp, 0);
2135                 if (ret < 0)
2136                         goto nack_op_err;
2137                 if (!ret)
2138                         goto rnr_nak;
2139                 qp->r_rcv_len = 0;
2140                 /* FALLTHROUGH */
2141         case OP(SEND_MIDDLE):
2142         case OP(RDMA_WRITE_MIDDLE):
2143 send_middle:
2144                 /* Check for invalid length PMTU or posted rwqe len. */
2145                 /*
2146                  * There will be no padding for 9B packet but 16B packets
2147                  * will come in with some padding since we always add
2148                  * CRC and LT bytes which will need to be flit aligned
2149                  */
2150                 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2151                         goto nack_inv;
2152                 qp->r_rcv_len += pmtu;
2153                 if (unlikely(qp->r_rcv_len > qp->r_len))
2154                         goto nack_inv;
2155                 hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
2156                 break;
2157
2158         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2159                 /* consume RWQE */
2160                 ret = hfi1_rvt_get_rwqe(qp, 1);
2161                 if (ret < 0)
2162                         goto nack_op_err;
2163                 if (!ret)
2164                         goto rnr_nak;
2165                 goto send_last_imm;
2166
2167         case OP(SEND_ONLY):
2168         case OP(SEND_ONLY_WITH_IMMEDIATE):
2169         case OP(SEND_ONLY_WITH_INVALIDATE):
2170                 ret = hfi1_rvt_get_rwqe(qp, 0);
2171                 if (ret < 0)
2172                         goto nack_op_err;
2173                 if (!ret)
2174                         goto rnr_nak;
2175                 qp->r_rcv_len = 0;
2176                 if (opcode == OP(SEND_ONLY))
2177                         goto no_immediate_data;
2178                 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2179                         goto send_last_inv;
2180                 /* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
2181         case OP(SEND_LAST_WITH_IMMEDIATE):
2182 send_last_imm:
2183                 wc.ex.imm_data = ohdr->u.imm_data;
2184                 wc.wc_flags = IB_WC_WITH_IMM;
2185                 goto send_last;
2186         case OP(SEND_LAST_WITH_INVALIDATE):
2187 send_last_inv:
2188                 rkey = be32_to_cpu(ohdr->u.ieth);
2189                 if (rvt_invalidate_rkey(qp, rkey))
2190                         goto no_immediate_data;
2191                 wc.ex.invalidate_rkey = rkey;
2192                 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2193                 goto send_last;
2194         case OP(RDMA_WRITE_LAST):
2195                 copy_last = rvt_is_user_qp(qp);
2196                 /* fall through */
2197         case OP(SEND_LAST):
2198 no_immediate_data:
2199                 wc.wc_flags = 0;
2200                 wc.ex.imm_data = 0;
2201 send_last:
2202                 /* Check for invalid length. */
2203                 /* LAST len should be >= 1 */
2204                 if (unlikely(tlen < (hdrsize + extra_bytes)))
2205                         goto nack_inv;
2206                 /* Don't count the CRC(and padding and LT byte for 16B). */
2207                 tlen -= (hdrsize + extra_bytes);
2208                 wc.byte_len = tlen + qp->r_rcv_len;
2209                 if (unlikely(wc.byte_len > qp->r_len))
2210                         goto nack_inv;
2211                 hfi1_copy_sge(&qp->r_sge, data, tlen, true, copy_last);
2212                 rvt_put_ss(&qp->r_sge);
2213                 qp->r_msn++;
2214                 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2215                         break;
2216                 wc.wr_id = qp->r_wr_id;
2217                 wc.status = IB_WC_SUCCESS;
2218                 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2219                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2220                         wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2221                 else
2222                         wc.opcode = IB_WC_RECV;
2223                 wc.qp = &qp->ibqp;
2224                 wc.src_qp = qp->remote_qpn;
2225                 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr);
2226                 /*
2227                  * It seems that IB mandates the presence of an SL in a
2228                  * work completion only for the UD transport (see section
2229                  * 11.4.2 of IBTA Vol. 1).
2230                  *
2231                  * However, the way the SL is chosen below is consistent
2232                  * with the way that IB/qib works and is trying avoid
2233                  * introducing incompatibilities.
2234                  *
2235                  * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2236                  */
2237                 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
2238                 /* zero fields that are N/A */
2239                 wc.vendor_err = 0;
2240                 wc.pkey_index = 0;
2241                 wc.dlid_path_bits = 0;
2242                 wc.port_num = 0;
2243                 /* Signal completion event if the solicited bit is set. */
2244                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
2245                              (bth0 & IB_BTH_SOLICITED) != 0);
2246                 break;
2247
2248         case OP(RDMA_WRITE_ONLY):
2249                 copy_last = rvt_is_user_qp(qp);
2250                 /* fall through */
2251         case OP(RDMA_WRITE_FIRST):
2252         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
2253                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
2254                         goto nack_inv;
2255                 /* consume RWQE */
2256                 reth = &ohdr->u.rc.reth;
2257                 qp->r_len = be32_to_cpu(reth->length);
2258                 qp->r_rcv_len = 0;
2259                 qp->r_sge.sg_list = NULL;
2260                 if (qp->r_len != 0) {
2261                         u32 rkey = be32_to_cpu(reth->rkey);
2262                         u64 vaddr = get_ib_reth_vaddr(reth);
2263                         int ok;
2264
2265                         /* Check rkey & NAK */
2266                         ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
2267                                          rkey, IB_ACCESS_REMOTE_WRITE);
2268                         if (unlikely(!ok))
2269                                 goto nack_acc;
2270                         qp->r_sge.num_sge = 1;
2271                 } else {
2272                         qp->r_sge.num_sge = 0;
2273                         qp->r_sge.sge.mr = NULL;
2274                         qp->r_sge.sge.vaddr = NULL;
2275                         qp->r_sge.sge.length = 0;
2276                         qp->r_sge.sge.sge_length = 0;
2277                 }
2278                 if (opcode == OP(RDMA_WRITE_FIRST))
2279                         goto send_middle;
2280                 else if (opcode == OP(RDMA_WRITE_ONLY))
2281                         goto no_immediate_data;
2282                 ret = hfi1_rvt_get_rwqe(qp, 1);
2283                 if (ret < 0)
2284                         goto nack_op_err;
2285                 if (!ret) {
2286                         /* peer will send again */
2287                         rvt_put_ss(&qp->r_sge);
2288                         goto rnr_nak;
2289                 }
2290                 wc.ex.imm_data = ohdr->u.rc.imm_data;
2291                 wc.wc_flags = IB_WC_WITH_IMM;
2292                 goto send_last;
2293
2294         case OP(RDMA_READ_REQUEST): {
2295                 struct rvt_ack_entry *e;
2296                 u32 len;
2297                 u8 next;
2298
2299                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2300                         goto nack_inv;
2301                 next = qp->r_head_ack_queue + 1;
2302                 /* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */
2303                 if (next > HFI1_MAX_RDMA_ATOMIC)
2304                         next = 0;
2305                 spin_lock_irqsave(&qp->s_lock, flags);
2306                 if (unlikely(next == qp->s_tail_ack_queue)) {
2307                         if (!qp->s_ack_queue[next].sent)
2308                                 goto nack_inv_unlck;
2309                         update_ack_queue(qp, next);
2310                 }
2311                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2312                 if (e->rdma_sge.mr) {
2313                         rvt_put_mr(e->rdma_sge.mr);
2314                         e->rdma_sge.mr = NULL;
2315                 }
2316                 reth = &ohdr->u.rc.reth;
2317                 len = be32_to_cpu(reth->length);
2318                 if (len) {
2319                         u32 rkey = be32_to_cpu(reth->rkey);
2320                         u64 vaddr = get_ib_reth_vaddr(reth);
2321                         int ok;
2322
2323                         /* Check rkey & NAK */
2324                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
2325                                          rkey, IB_ACCESS_REMOTE_READ);
2326                         if (unlikely(!ok))
2327                                 goto nack_acc_unlck;
2328                         /*
2329                          * Update the next expected PSN.  We add 1 later
2330                          * below, so only add the remainder here.
2331                          */
2332                         qp->r_psn += rvt_div_mtu(qp, len - 1);
2333                 } else {
2334                         e->rdma_sge.mr = NULL;
2335                         e->rdma_sge.vaddr = NULL;
2336                         e->rdma_sge.length = 0;
2337                         e->rdma_sge.sge_length = 0;
2338                 }
2339                 e->opcode = opcode;
2340                 e->sent = 0;
2341                 e->psn = psn;
2342                 e->lpsn = qp->r_psn;
2343                 /*
2344                  * We need to increment the MSN here instead of when we
2345                  * finish sending the result since a duplicate request would
2346                  * increment it more than once.
2347                  */
2348                 qp->r_msn++;
2349                 qp->r_psn++;
2350                 qp->r_state = opcode;
2351                 qp->r_nak_state = 0;
2352                 qp->r_head_ack_queue = next;
2353
2354                 /* Schedule the send engine. */
2355                 qp->s_flags |= RVT_S_RESP_PENDING;
2356                 hfi1_schedule_send(qp);
2357
2358                 spin_unlock_irqrestore(&qp->s_lock, flags);
2359                 if (is_fecn)
2360                         goto send_ack;
2361                 return;
2362         }
2363
2364         case OP(COMPARE_SWAP):
2365         case OP(FETCH_ADD): {
2366                 struct ib_atomic_eth *ateth;
2367                 struct rvt_ack_entry *e;
2368                 u64 vaddr;
2369                 atomic64_t *maddr;
2370                 u64 sdata;
2371                 u32 rkey;
2372                 u8 next;
2373
2374                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2375                         goto nack_inv;
2376                 next = qp->r_head_ack_queue + 1;
2377                 if (next > HFI1_MAX_RDMA_ATOMIC)
2378                         next = 0;
2379                 spin_lock_irqsave(&qp->s_lock, flags);
2380                 if (unlikely(next == qp->s_tail_ack_queue)) {
2381                         if (!qp->s_ack_queue[next].sent)
2382                                 goto nack_inv_unlck;
2383                         update_ack_queue(qp, next);
2384                 }
2385                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2386                 if (e->rdma_sge.mr) {
2387                         rvt_put_mr(e->rdma_sge.mr);
2388                         e->rdma_sge.mr = NULL;
2389                 }
2390                 ateth = &ohdr->u.atomic_eth;
2391                 vaddr = get_ib_ateth_vaddr(ateth);
2392                 if (unlikely(vaddr & (sizeof(u64) - 1)))
2393                         goto nack_inv_unlck;
2394                 rkey = be32_to_cpu(ateth->rkey);
2395                 /* Check rkey & NAK */
2396                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2397                                           vaddr, rkey,
2398                                           IB_ACCESS_REMOTE_ATOMIC)))
2399                         goto nack_acc_unlck;
2400                 /* Perform atomic OP and save result. */
2401                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
2402                 sdata = get_ib_ateth_swap(ateth);
2403                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2404                         (u64)atomic64_add_return(sdata, maddr) - sdata :
2405                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
2406                                       get_ib_ateth_compare(ateth),
2407                                       sdata);
2408                 rvt_put_mr(qp->r_sge.sge.mr);
2409                 qp->r_sge.num_sge = 0;
2410                 e->opcode = opcode;
2411                 e->sent = 0;
2412                 e->psn = psn;
2413                 e->lpsn = psn;
2414                 qp->r_msn++;
2415                 qp->r_psn++;
2416                 qp->r_state = opcode;
2417                 qp->r_nak_state = 0;
2418                 qp->r_head_ack_queue = next;
2419
2420                 /* Schedule the send engine. */
2421                 qp->s_flags |= RVT_S_RESP_PENDING;
2422                 hfi1_schedule_send(qp);
2423
2424                 spin_unlock_irqrestore(&qp->s_lock, flags);
2425                 if (is_fecn)
2426                         goto send_ack;
2427                 return;
2428         }
2429
2430         default:
2431                 /* NAK unknown opcodes. */
2432                 goto nack_inv;
2433         }
2434         qp->r_psn++;
2435         qp->r_state = opcode;
2436         qp->r_ack_psn = psn;
2437         qp->r_nak_state = 0;
2438         /* Send an ACK if requested or required. */
2439         if (psn & IB_BTH_REQ_ACK) {
2440                 if (packet->numpkt == 0) {
2441                         rc_cancel_ack(qp);
2442                         goto send_ack;
2443                 }
2444                 if (qp->r_adefered >= HFI1_PSN_CREDIT) {
2445                         rc_cancel_ack(qp);
2446                         goto send_ack;
2447                 }
2448                 if (unlikely(is_fecn)) {
2449                         rc_cancel_ack(qp);
2450                         goto send_ack;
2451                 }
2452                 qp->r_adefered++;
2453                 rc_defered_ack(rcd, qp);
2454         }
2455         return;
2456
2457 rnr_nak:
2458         qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
2459         qp->r_ack_psn = qp->r_psn;
2460         /* Queue RNR NAK for later */
2461         rc_defered_ack(rcd, qp);
2462         return;
2463
2464 nack_op_err:
2465         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2466         qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2467         qp->r_ack_psn = qp->r_psn;
2468         /* Queue NAK for later */
2469         rc_defered_ack(rcd, qp);
2470         return;
2471
2472 nack_inv_unlck:
2473         spin_unlock_irqrestore(&qp->s_lock, flags);
2474 nack_inv:
2475         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2476         qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2477         qp->r_ack_psn = qp->r_psn;
2478         /* Queue NAK for later */
2479         rc_defered_ack(rcd, qp);
2480         return;
2481
2482 nack_acc_unlck:
2483         spin_unlock_irqrestore(&qp->s_lock, flags);
2484 nack_acc:
2485         rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2486         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2487         qp->r_ack_psn = qp->r_psn;
2488 send_ack:
2489         hfi1_send_rc_ack(rcd, qp, is_fecn);
2490 }
2491
2492 void hfi1_rc_hdrerr(
2493         struct hfi1_ctxtdata *rcd,
2494         struct hfi1_packet *packet,
2495         struct rvt_qp *qp)
2496 {
2497         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2498         int diff;
2499         u32 opcode;
2500         u32 psn;
2501
2502         if (hfi1_ruc_check_hdr(ibp, packet))
2503                 return;
2504
2505         psn = ib_bth_get_psn(packet->ohdr);
2506         opcode = ib_bth_get_opcode(packet->ohdr);
2507
2508         /* Only deal with RDMA Writes for now */
2509         if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
2510                 diff = delta_psn(psn, qp->r_psn);
2511                 if (!qp->r_nak_state && diff >= 0) {
2512                         ibp->rvp.n_rc_seqnak++;
2513                         qp->r_nak_state = IB_NAK_PSN_ERROR;
2514                         /* Use the expected PSN. */
2515                         qp->r_ack_psn = qp->r_psn;
2516                         /*
2517                          * Wait to send the sequence
2518                          * NAK until all packets
2519                          * in the receive queue have
2520                          * been processed.
2521                          * Otherwise, we end up
2522                          * propagating congestion.
2523                          */
2524                         rc_defered_ack(rcd, qp);
2525                 } /* Out of sequence NAK */
2526         } /* QP Request NAKs */
2527 }