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 (IS_ERR(ps->s_txreq))
 277                 goto bail_no_tx;
 278
 279         if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
 280                 /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 281                 hwords = 5;
 282                 if (rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)
 283                         ohdr = &ps->s_txreq->phdr.hdr.ibh.u.l.oth;
 284                 else
 285                         ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth;
 286         } else {
 287                 /* header size in 32-bit words 16B LRH+BTH = (16+12)/4. */
 288                 hwords = 7;
 289                 if ((rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
 290                     (hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))))
 291                         ohdr = &ps->s_txreq->phdr.hdr.opah.u.l.oth;
 292                 else
 293                         ohdr = &ps->s_txreq->phdr.hdr.opah.u.oth;
 294         }
 295
 296         /* Sending responses has higher priority over sending requests. */
 297         if ((qp->s_flags & RVT_S_RESP_PENDING) &&
 298             make_rc_ack(dev, qp, ohdr, ps))
 299                 return 1;
 300
 301         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
 302                 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
 303                         goto bail;
 304                 /* We are in the error state, flush the work request. */
 305                 smp_read_barrier_depends(); /* see post_one_send() */
 306                 if (qp->s_last == READ_ONCE(qp->s_head))
 307                         goto bail;
 308                 /* If DMAs are in progress, we can't flush immediately. */
 309                 if (iowait_sdma_pending(&priv->s_iowait)) {
 310                         qp->s_flags |= RVT_S_WAIT_DMA;
 311                         goto bail;
 312                 }
 313                 clear_ahg(qp);
 314                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 315                 hfi1_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
 316                         IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
 317                 /* will get called again */
 318                 goto done_free_tx;
 319         }
 320
 321         if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
 322                 goto bail;
 323
 324         if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
 325                 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
 326                         qp->s_flags |= RVT_S_WAIT_PSN;
 327                         goto bail;
 328                 }
 329                 qp->s_sending_psn = qp->s_psn;
 330                 qp->s_sending_hpsn = qp->s_psn - 1;
 331         }
 332
 333         /* Send a request. */
 334         wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
 335         switch (qp->s_state) {
 336         default:
 337                 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
 338                         goto bail;
 339                 /*
 340                  * Resend an old request or start a new one.
 341                  *
 342                  * We keep track of the current SWQE so that
 343                  * we don't reset the "furthest progress" state
 344                  * if we need to back up.
 345                  */
 346                 newreq = 0;
 347                 if (qp->s_cur == qp->s_tail) {
 348                         /* Check if send work queue is empty. */
 349                         smp_read_barrier_depends(); /* see post_one_send() */
 350                         if (qp->s_tail == READ_ONCE(qp->s_head)) {
 351                                 clear_ahg(qp);
 352                                 goto bail;
 353                         }
 354                         /*
 355                          * If a fence is requested, wait for previous
 356                          * RDMA read and atomic operations to finish.
 357                          */
 358                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 359                             qp->s_num_rd_atomic) {
 360                                 qp->s_flags |= RVT_S_WAIT_FENCE;
 361                                 goto bail;
 362                         }
 363                         /*
 364                          * Local operations are processed immediately
 365                          * after all prior requests have completed
 366                          */
 367                         if (wqe->wr.opcode == IB_WR_REG_MR ||
 368                             wqe->wr.opcode == IB_WR_LOCAL_INV) {
 369                                 int local_ops = 0;
 370                                 int err = 0;
 371
 372                                 if (qp->s_last != qp->s_cur)
 373                                         goto bail;
 374                                 if (++qp->s_cur == qp->s_size)
 375                                         qp->s_cur = 0;
 376                                 if (++qp->s_tail == qp->s_size)
 377                                         qp->s_tail = 0;
 378                                 if (!(wqe->wr.send_flags &
 379                                       RVT_SEND_COMPLETION_ONLY)) {
 380                                         err = rvt_invalidate_rkey(
 381                                                 qp,
 382                                                 wqe->wr.ex.invalidate_rkey);
 383                                         local_ops = 1;
 384                                 }
 385                                 hfi1_send_complete(qp, wqe,
 386                                                    err ? IB_WC_LOC_PROT_ERR
 387                                                        : IB_WC_SUCCESS);
 388                                 if (local_ops)
 389                                         atomic_dec(&qp->local_ops_pending);
 390                                 qp->s_hdrwords = 0;
 391                                 goto done_free_tx;
 392                         }
 393
 394                         newreq = 1;
 395                         qp->s_psn = wqe->psn;
 396                 }
 397                 /*
 398                  * Note that we have to be careful not to modify the
 399                  * original work request since we may need to resend
 400                  * it.
 401                  */
 402                 len = wqe->length;
 403                 ss = &qp->s_sge;
 404                 bth2 = mask_psn(qp->s_psn);
 405                 switch (wqe->wr.opcode) {
 406                 case IB_WR_SEND:
 407                 case IB_WR_SEND_WITH_IMM:
 408                 case IB_WR_SEND_WITH_INV:
 409                         /* If no credit, return. */
 410                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 411                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 412                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 413                                 goto bail;
 414                         }
 415                         if (len > pmtu) {
 416                                 qp->s_state = OP(SEND_FIRST);
 417                                 len = pmtu;
 418                                 break;
 419                         }
 420                         if (wqe->wr.opcode == IB_WR_SEND) {
 421                                 qp->s_state = OP(SEND_ONLY);
 422                         } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 423                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 424                                 /* Immediate data comes after the BTH */
 425                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 426                                 hwords += 1;
 427                         } else {
 428                                 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
 429                                 /* Invalidate rkey comes after the BTH */
 430                                 ohdr->u.ieth = cpu_to_be32(
 431                                                 wqe->wr.ex.invalidate_rkey);
 432                                 hwords += 1;
 433                         }
 434                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 435                                 bth0 |= IB_BTH_SOLICITED;
 436                         bth2 |= IB_BTH_REQ_ACK;
 437                         if (++qp->s_cur == qp->s_size)
 438                                 qp->s_cur = 0;
 439                         break;
 440
 441                 case IB_WR_RDMA_WRITE:
 442                         if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 443                                 qp->s_lsn++;
 444                         goto no_flow_control;
 445                 case IB_WR_RDMA_WRITE_WITH_IMM:
 446                         /* If no credit, return. */
 447                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 448                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 449                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 450                                 goto bail;
 451                         }
 452 no_flow_control:
 453                         put_ib_reth_vaddr(
 454                                 wqe->rdma_wr.remote_addr,
 455                                 &ohdr->u.rc.reth);
 456                         ohdr->u.rc.reth.rkey =
 457                                 cpu_to_be32(wqe->rdma_wr.rkey);
 458                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 459                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 460                         if (len > pmtu) {
 461                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 462                                 len = pmtu;
 463                                 break;
 464                         }
 465                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 466                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 467                         } else {
 468                                 qp->s_state =
 469                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 470                                 /* Immediate data comes after RETH */
 471                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 472                                 hwords += 1;
 473                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 474                                         bth0 |= IB_BTH_SOLICITED;
 475                         }
 476                         bth2 |= IB_BTH_REQ_ACK;
 477                         if (++qp->s_cur == qp->s_size)
 478                                 qp->s_cur = 0;
 479                         break;
 480
 481                 case IB_WR_RDMA_READ:
 482                         /*
 483                          * Don't allow more operations to be started
 484                          * than the QP limits allow.
 485                          */
 486                         if (newreq) {
 487                                 if (qp->s_num_rd_atomic >=
 488                                     qp->s_max_rd_atomic) {
 489                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 490                                         goto bail;
 491                                 }
 492                                 qp->s_num_rd_atomic++;
 493                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 494                                         qp->s_lsn++;
 495                         }
 496                         put_ib_reth_vaddr(
 497                                 wqe->rdma_wr.remote_addr,
 498                                 &ohdr->u.rc.reth);
 499                         ohdr->u.rc.reth.rkey =
 500                                 cpu_to_be32(wqe->rdma_wr.rkey);
 501                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 502                         qp->s_state = OP(RDMA_READ_REQUEST);
 503                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 504                         ss = NULL;
 505                         len = 0;
 506                         bth2 |= IB_BTH_REQ_ACK;
 507                         if (++qp->s_cur == qp->s_size)
 508                                 qp->s_cur = 0;
 509                         break;
 510
 511                 case IB_WR_ATOMIC_CMP_AND_SWP:
 512                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 513                         /*
 514                          * Don't allow more operations to be started
 515                          * than the QP limits allow.
 516                          */
 517                         if (newreq) {
 518                                 if (qp->s_num_rd_atomic >=
 519                                     qp->s_max_rd_atomic) {
 520                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 521                                         goto bail;
 522                                 }
 523                                 qp->s_num_rd_atomic++;
 524                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 525                                         qp->s_lsn++;
 526                         }
 527                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 528                                 qp->s_state = OP(COMPARE_SWAP);
 529                                 put_ib_ateth_swap(wqe->atomic_wr.swap,
 530                                                   &ohdr->u.atomic_eth);
 531                                 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
 532                                                      &ohdr->u.atomic_eth);
 533                         } else {
 534                                 qp->s_state = OP(FETCH_ADD);
 535                                 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
 536                                                   &ohdr->u.atomic_eth);
 537                                 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
 538                         }
 539                         put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
 540                                            &ohdr->u.atomic_eth);
 541                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 542                                 wqe->atomic_wr.rkey);
 543                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 544                         ss = NULL;
 545                         len = 0;
 546                         bth2 |= IB_BTH_REQ_ACK;
 547                         if (++qp->s_cur == qp->s_size)
 548                                 qp->s_cur = 0;
 549                         break;
 550
 551                 default:
 552                         goto bail;
 553                 }
 554                 qp->s_sge.sge = wqe->sg_list[0];
 555                 qp->s_sge.sg_list = wqe->sg_list + 1;
 556                 qp->s_sge.num_sge = wqe->wr.num_sge;
 557                 qp->s_sge.total_len = wqe->length;
 558                 qp->s_len = wqe->length;
 559                 if (newreq) {
 560                         qp->s_tail++;
 561                         if (qp->s_tail >= qp->s_size)
 562                                 qp->s_tail = 0;
 563                 }
 564                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 565                         qp->s_psn = wqe->lpsn + 1;
 566                 else
 567                         qp->s_psn++;
 568                 break;
 569
 570         case OP(RDMA_READ_RESPONSE_FIRST):
 571                 /*
 572                  * qp->s_state is normally set to the opcode of the
 573                  * last packet constructed for new requests and therefore
 574                  * is never set to RDMA read response.
 575                  * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
 576                  * thread to indicate a SEND needs to be restarted from an
 577                  * earlier PSN without interfering with the sending thread.
 578                  * See restart_rc().
 579                  */
 580                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 581                 /* FALLTHROUGH */
 582         case OP(SEND_FIRST):
 583                 qp->s_state = OP(SEND_MIDDLE);
 584                 /* FALLTHROUGH */
 585         case OP(SEND_MIDDLE):
 586                 bth2 = mask_psn(qp->s_psn++);
 587                 ss = &qp->s_sge;
 588                 len = qp->s_len;
 589                 if (len > pmtu) {
 590                         len = pmtu;
 591                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 592                         break;
 593                 }
 594                 if (wqe->wr.opcode == IB_WR_SEND) {
 595                         qp->s_state = OP(SEND_LAST);
 596                 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 597                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 598                         /* Immediate data comes after the BTH */
 599                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 600                         hwords += 1;
 601                 } else {
 602                         qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
 603                         /* invalidate data comes after the BTH */
 604                         ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
 605                         hwords += 1;
 606                 }
 607                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 608                         bth0 |= IB_BTH_SOLICITED;
 609                 bth2 |= IB_BTH_REQ_ACK;
 610                 qp->s_cur++;
 611                 if (qp->s_cur >= qp->s_size)
 612                         qp->s_cur = 0;
 613                 break;
 614
 615         case OP(RDMA_READ_RESPONSE_LAST):
 616                 /*
 617                  * qp->s_state is normally set to the opcode of the
 618                  * last packet constructed for new requests and therefore
 619                  * is never set to RDMA read response.
 620                  * RDMA_READ_RESPONSE_LAST is used by the ACK processing
 621                  * thread to indicate a RDMA write needs to be restarted from
 622                  * an earlier PSN without interfering with the sending thread.
 623                  * See restart_rc().
 624                  */
 625                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 626                 /* FALLTHROUGH */
 627         case OP(RDMA_WRITE_FIRST):
 628                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 629                 /* FALLTHROUGH */
 630         case OP(RDMA_WRITE_MIDDLE):
 631                 bth2 = mask_psn(qp->s_psn++);
 632                 ss = &qp->s_sge;
 633                 len = qp->s_len;
 634                 if (len > pmtu) {
 635                         len = pmtu;
 636                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 637                         break;
 638                 }
 639                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 640                         qp->s_state = OP(RDMA_WRITE_LAST);
 641                 } else {
 642                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 643                         /* Immediate data comes after the BTH */
 644                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 645                         hwords += 1;
 646                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 647                                 bth0 |= IB_BTH_SOLICITED;
 648                 }
 649                 bth2 |= IB_BTH_REQ_ACK;
 650                 qp->s_cur++;
 651                 if (qp->s_cur >= qp->s_size)
 652                         qp->s_cur = 0;
 653                 break;
 654
 655         case OP(RDMA_READ_RESPONSE_MIDDLE):
 656                 /*
 657                  * qp->s_state is normally set to the opcode of the
 658                  * last packet constructed for new requests and therefore
 659                  * is never set to RDMA read response.
 660                  * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
 661                  * thread to indicate a RDMA read needs to be restarted from
 662                  * an earlier PSN without interfering with the sending thread.
 663                  * See restart_rc().
 664                  */
 665                 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
 666                 put_ib_reth_vaddr(
 667                         wqe->rdma_wr.remote_addr + len,
 668                         &ohdr->u.rc.reth);
 669                 ohdr->u.rc.reth.rkey =
 670                         cpu_to_be32(wqe->rdma_wr.rkey);
 671                 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
 672                 qp->s_state = OP(RDMA_READ_REQUEST);
 673                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 674                 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
 675                 qp->s_psn = wqe->lpsn + 1;
 676                 ss = NULL;
 677                 len = 0;
 678                 qp->s_cur++;
 679                 if (qp->s_cur == qp->s_size)
 680                         qp->s_cur = 0;
 681                 break;
 682         }
 683         qp->s_sending_hpsn = bth2;
 684         delta = delta_psn(bth2, wqe->psn);
 685         if (delta && delta % HFI1_PSN_CREDIT == 0)
 686                 bth2 |= IB_BTH_REQ_ACK;
 687         if (qp->s_flags & RVT_S_SEND_ONE) {
 688                 qp->s_flags &= ~RVT_S_SEND_ONE;
 689                 qp->s_flags |= RVT_S_WAIT_ACK;
 690                 bth2 |= IB_BTH_REQ_ACK;
 691         }
 692         qp->s_len -= len;
 693         qp->s_hdrwords = hwords;
 694         ps->s_txreq->sde = priv->s_sde;
 695         ps->s_txreq->ss = ss;
 696         ps->s_txreq->s_cur_size = len;
 697         hfi1_make_ruc_header(
 698                 qp,
 699                 ohdr,
 700                 bth0 | (qp->s_state << 24),
 701                 bth2,
 702                 middle,
 703                 ps);
 704         /* pbc */
 705         ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
 706         return 1;
 707
 708 done_free_tx:
 709         hfi1_put_txreq(ps->s_txreq);
 710         ps->s_txreq = NULL;
 711         return 1;
 712
 713 bail:
 714         hfi1_put_txreq(ps->s_txreq);
 715
 716 bail_no_tx:
 717         ps->s_txreq = NULL;
 718         qp->s_flags &= ~RVT_S_BUSY;
 719         qp->s_hdrwords = 0;
 720         return 0;
 721 }
 722
 723 static inline void hfi1_make_bth_aeth(struct rvt_qp *qp,
 724                                       struct ib_other_headers *ohdr,
 725                                       u32 bth0, u32 bth1)
 726 {
 727         if (qp->r_nak_state)
 728                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
 729                                             (qp->r_nak_state <<
 730                                              IB_AETH_CREDIT_SHIFT));
 731         else
 732                 ohdr->u.aeth = rvt_compute_aeth(qp);
 733
 734         ohdr->bth[0] = cpu_to_be32(bth0);
 735         ohdr->bth[1] = cpu_to_be32(bth1 | qp->remote_qpn);
 736         ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
 737 }
 738
 739 static inline void hfi1_queue_rc_ack(struct rvt_qp *qp, bool is_fecn)
 740 {
 741         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 742         unsigned long flags;
 743
 744         spin_lock_irqsave(&qp->s_lock, flags);
 745         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
 746                 goto unlock;
 747         this_cpu_inc(*ibp->rvp.rc_qacks);
 748         qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
 749         qp->s_nak_state = qp->r_nak_state;
 750         qp->s_ack_psn = qp->r_ack_psn;
 751         if (is_fecn)
 752                 qp->s_flags |= RVT_S_ECN;
 753
 754         /* Schedule the send tasklet. */
 755         hfi1_schedule_send(qp);
 756 unlock:
 757         spin_unlock_irqrestore(&qp->s_lock, flags);
 758 }
 759
 760 static inline void hfi1_make_rc_ack_9B(struct rvt_qp *qp,
 761                                        struct hfi1_opa_header *opa_hdr,
 762                                        u8 sc5, bool is_fecn,
 763                                        u64 *pbc_flags, u32 *hwords,
 764                                        u32 *nwords)
 765 {
 766         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 767         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 768         struct ib_header *hdr = &opa_hdr->ibh;
 769         struct ib_other_headers *ohdr;
 770         u16 lrh0 = HFI1_LRH_BTH;
 771         u16 pkey;
 772         u32 bth0, bth1;
 773
 774         opa_hdr->hdr_type = HFI1_PKT_TYPE_9B;
 775         ohdr = &hdr->u.oth;
 776         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
 777         *hwords = 6;
 778
 779         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH)) {
 780                 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
 781                                          rdma_ah_read_grh(&qp->remote_ah_attr),
 782                                          *hwords - 2, SIZE_OF_CRC);
 783                 ohdr = &hdr->u.l.oth;
 784                 lrh0 = HFI1_LRH_GRH;
 785         }
 786         /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
 787         *pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
 788
 789         /* read pkey_index w/o lock (its atomic) */
 790         pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 791
 792         lrh0 |= (sc5 & IB_SC_MASK) << IB_SC_SHIFT |
 793                 (rdma_ah_get_sl(&qp->remote_ah_attr) & IB_SL_MASK) <<
 794                         IB_SL_SHIFT;
 795
 796         hfi1_make_ib_hdr(hdr, lrh0, *hwords + SIZE_OF_CRC,
 797                          opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr), 9B),
 798                          ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr));
 799
 800         bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
 801         if (qp->s_mig_state == IB_MIG_MIGRATED)
 802                 bth0 |= IB_BTH_MIG_REQ;
 803         bth1 = (!!is_fecn) << IB_BECN_SHIFT;
 804         hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
 805 }
 806
 807 static inline void hfi1_make_rc_ack_16B(struct rvt_qp *qp,
 808                                         struct hfi1_opa_header *opa_hdr,
 809                                         u8 sc5, bool is_fecn,
 810                                         u64 *pbc_flags, u32 *hwords,
 811                                         u32 *nwords)
 812 {
 813         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 814         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 815         struct hfi1_16b_header *hdr = &opa_hdr->opah;
 816         struct ib_other_headers *ohdr;
 817         u32 bth0, bth1;
 818         u16 len, pkey;
 819         u8 becn = !!is_fecn;
 820         u8 l4 = OPA_16B_L4_IB_LOCAL;
 821         u8 extra_bytes;
 822
 823         opa_hdr->hdr_type = HFI1_PKT_TYPE_16B;
 824         ohdr = &hdr->u.oth;
 825         /* header size in 32-bit words 16B LRH+BTH+AETH = (16+12+4)/4 */
 826         *hwords = 8;
 827         extra_bytes = hfi1_get_16b_padding(*hwords << 2, 0);
 828         *nwords = SIZE_OF_CRC + ((extra_bytes + SIZE_OF_LT) >> 2);
 829
 830         if (unlikely(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH) &&
 831             hfi1_check_mcast(rdma_ah_get_dlid(&qp->remote_ah_attr))) {
 832                 *hwords += hfi1_make_grh(ibp, &hdr->u.l.grh,
 833                                          rdma_ah_read_grh(&qp->remote_ah_attr),
 834                                          *hwords - 4, *nwords);
 835                 ohdr = &hdr->u.l.oth;
 836                 l4 = OPA_16B_L4_IB_GLOBAL;
 837         }
 838         *pbc_flags |= PBC_PACKET_BYPASS | PBC_INSERT_BYPASS_ICRC;
 839
 840         /* read pkey_index w/o lock (its atomic) */
 841         pkey = hfi1_get_pkey(ibp, qp->s_pkey_index);
 842
 843         /* Convert dwords to flits */
 844         len = (*hwords + *nwords) >> 1;
 845
 846         hfi1_make_16b_hdr(hdr,
 847                           ppd->lid | rdma_ah_get_path_bits(&qp->remote_ah_attr),
 848                           opa_get_lid(rdma_ah_get_dlid(&qp->remote_ah_attr),
 849                                       16B),
 850                           len, pkey, becn, 0, l4, sc5);
 851
 852         bth0 = pkey | (OP(ACKNOWLEDGE) << 24);
 853         bth0 |= extra_bytes << 20;
 854         if (qp->s_mig_state == IB_MIG_MIGRATED)
 855                 bth1 = OPA_BTH_MIG_REQ;
 856         hfi1_make_bth_aeth(qp, ohdr, bth0, bth1);
 857 }
 858
 859 typedef void (*hfi1_make_rc_ack)(struct rvt_qp *qp,
 860                                  struct hfi1_opa_header *opa_hdr,
 861                                  u8 sc5, bool is_fecn,
 862                                  u64 *pbc_flags, u32 *hwords,
 863                                  u32 *nwords);
 864
 865 /* We support only two types - 9B and 16B for now */
 866 static const hfi1_make_rc_ack hfi1_make_rc_ack_tbl[2] = {
 867         [HFI1_PKT_TYPE_9B] = &hfi1_make_rc_ack_9B,
 868         [HFI1_PKT_TYPE_16B] = &hfi1_make_rc_ack_16B
 869 };
 870
 871 /**
 872  * hfi1_send_rc_ack - Construct an ACK packet and send it
 873  * @qp: a pointer to the QP
 874  *
 875  * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
 876  * Note that RDMA reads and atomics are handled in the
 877  * send side QP state and send engine.
 878  */
 879 void hfi1_send_rc_ack(struct hfi1_ctxtdata *rcd,
 880                       struct rvt_qp *qp, bool is_fecn)
 881 {
 882         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 883         struct hfi1_qp_priv *priv = qp->priv;
 884         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 885         u8 sc5 = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
 886         u64 pbc, pbc_flags = 0;
 887         u32 hwords = 0;
 888         u32 nwords = 0;
 889         u32 plen;
 890         struct pio_buf *pbuf;
 891         struct hfi1_opa_header opa_hdr;
 892
 893         /* clear the defer count */
 894         qp->r_adefered = 0;
 895
 896         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 897         if (qp->s_flags & RVT_S_RESP_PENDING) {
 898                 hfi1_queue_rc_ack(qp, is_fecn);
 899                 return;
 900         }
 901
 902         /* Ensure s_rdma_ack_cnt changes are committed */
 903         smp_read_barrier_depends();
 904         if (qp->s_rdma_ack_cnt) {
 905                 hfi1_queue_rc_ack(qp, is_fecn);
 906                 return;
 907         }
 908
 909         /* Don't try to send ACKs if the link isn't ACTIVE */
 910         if (driver_lstate(ppd) != IB_PORT_ACTIVE)
 911                 return;
 912
 913         /* Make the appropriate header */
 914         hfi1_make_rc_ack_tbl[priv->hdr_type](qp, &opa_hdr, sc5, is_fecn,
 915                                              &pbc_flags, &hwords, &nwords);
 916
 917         plen = 2 /* PBC */ + hwords + nwords;
 918         pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps,
 919                          sc_to_vlt(ppd->dd, sc5), plen);
 920         pbuf = sc_buffer_alloc(rcd->sc, plen, NULL, NULL);
 921         if (!pbuf) {
 922                 /*
 923                  * We have no room to send at the moment.  Pass
 924                  * responsibility for sending the ACK to the send engine
 925                  * so that when enough buffer space becomes available,
 926                  * the ACK is sent ahead of other outgoing packets.
 927                  */
 928                 hfi1_queue_rc_ack(qp, is_fecn);
 929                 return;
 930         }
 931         trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device),
 932                                &opa_hdr, ib_is_sc5(sc5));
 933
 934         /* write the pbc and data */
 935         ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc,
 936                                  (priv->hdr_type == HFI1_PKT_TYPE_9B ?
 937                                  (void *)&opa_hdr.ibh :
 938                                  (void *)&opa_hdr.opah), hwords);
 939         return;
 940 }
 941
 942 /**
 943  * reset_psn - reset the QP state to send starting from PSN
 944  * @qp: the QP
 945  * @psn: the packet sequence number to restart at
 946  *
 947  * This is called from hfi1_rc_rcv() to process an incoming RC ACK
 948  * for the given QP.
 949  * Called at interrupt level with the QP s_lock held.
 950  */
 951 static void reset_psn(struct rvt_qp *qp, u32 psn)
 952 {
 953         u32 n = qp->s_acked;
 954         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
 955         u32 opcode;
 956
 957         lockdep_assert_held(&qp->s_lock);
 958         qp->s_cur = n;
 959
 960         /*
 961          * If we are starting the request from the beginning,
 962          * let the normal send code handle initialization.
 963          */
 964         if (cmp_psn(psn, wqe->psn) <= 0) {
 965                 qp->s_state = OP(SEND_LAST);
 966                 goto done;
 967         }
 968
 969         /* Find the work request opcode corresponding to the given PSN. */
 970         opcode = wqe->wr.opcode;
 971         for (;;) {
 972                 int diff;
 973
 974                 if (++n == qp->s_size)
 975                         n = 0;
 976                 if (n == qp->s_tail)
 977                         break;
 978                 wqe = rvt_get_swqe_ptr(qp, n);
 979                 diff = cmp_psn(psn, wqe->psn);
 980                 if (diff < 0)
 981                         break;
 982                 qp->s_cur = n;
 983                 /*
 984                  * If we are starting the request from the beginning,
 985                  * let the normal send code handle initialization.
 986                  */
 987                 if (diff == 0) {
 988                         qp->s_state = OP(SEND_LAST);
 989                         goto done;
 990                 }
 991                 opcode = wqe->wr.opcode;
 992         }
 993
 994         /*
 995          * Set the state to restart in the middle of a request.
 996          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 997          * See hfi1_make_rc_req().
 998          */
 999         switch (opcode) {
1000         case IB_WR_SEND:
1001         case IB_WR_SEND_WITH_IMM:
1002                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
1003                 break;
1004
1005         case IB_WR_RDMA_WRITE:
1006         case IB_WR_RDMA_WRITE_WITH_IMM:
1007                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
1008                 break;
1009
1010         case IB_WR_RDMA_READ:
1011                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
1012                 break;
1013
1014         default:
1015                 /*
1016                  * This case shouldn't happen since its only
1017                  * one PSN per req.
1018                  */
1019                 qp->s_state = OP(SEND_LAST);
1020         }
1021 done:
1022         qp->s_psn = psn;
1023         /*
1024          * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
1025          * asynchronously before the send engine can get scheduled.
1026          * Doing it in hfi1_make_rc_req() is too late.
1027          */
1028         if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
1029             (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
1030                 qp->s_flags |= RVT_S_WAIT_PSN;
1031         qp->s_flags &= ~RVT_S_AHG_VALID;
1032 }
1033
1034 /*
1035  * Back up requester to resend the last un-ACKed request.
1036  * The QP r_lock and s_lock should be held and interrupts disabled.
1037  */
1038 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
1039 {
1040         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1041         struct hfi1_ibport *ibp;
1042
1043         lockdep_assert_held(&qp->r_lock);
1044         lockdep_assert_held(&qp->s_lock);
1045         if (qp->s_retry == 0) {
1046                 if (qp->s_mig_state == IB_MIG_ARMED) {
1047                         hfi1_migrate_qp(qp);
1048                         qp->s_retry = qp->s_retry_cnt;
1049                 } else if (qp->s_last == qp->s_acked) {
1050                         hfi1_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
1051                         rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1052                         return;
1053                 } else { /* need to handle delayed completion */
1054                         return;
1055                 }
1056         } else {
1057                 qp->s_retry--;
1058         }
1059
1060         ibp = to_iport(qp->ibqp.device, qp->port_num);
1061         if (wqe->wr.opcode == IB_WR_RDMA_READ)
1062                 ibp->rvp.n_rc_resends++;
1063         else
1064                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1065
1066         qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
1067                          RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
1068                          RVT_S_WAIT_ACK);
1069         if (wait)
1070                 qp->s_flags |= RVT_S_SEND_ONE;
1071         reset_psn(qp, psn);
1072 }
1073
1074 /*
1075  * Set qp->s_sending_psn to the next PSN after the given one.
1076  * This would be psn+1 except when RDMA reads are present.
1077  */
1078 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
1079 {
1080         struct rvt_swqe *wqe;
1081         u32 n = qp->s_last;
1082
1083         lockdep_assert_held(&qp->s_lock);
1084         /* Find the work request corresponding to the given PSN. */
1085         for (;;) {
1086                 wqe = rvt_get_swqe_ptr(qp, n);
1087                 if (cmp_psn(psn, wqe->lpsn) <= 0) {
1088                         if (wqe->wr.opcode == IB_WR_RDMA_READ)
1089                                 qp->s_sending_psn = wqe->lpsn + 1;
1090                         else
1091                                 qp->s_sending_psn = psn + 1;
1092                         break;
1093                 }
1094                 if (++n == qp->s_size)
1095                         n = 0;
1096                 if (n == qp->s_tail)
1097                         break;
1098         }
1099 }
1100
1101 /*
1102  * This should be called with the QP s_lock held and interrupts disabled.
1103  */
1104 void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
1105 {
1106         struct ib_other_headers *ohdr;
1107         struct hfi1_qp_priv *priv = qp->priv;
1108         struct rvt_swqe *wqe;
1109         struct ib_header *hdr = NULL;
1110         struct hfi1_16b_header *hdr_16b = NULL;
1111         u32 opcode;
1112         u32 psn;
1113
1114         lockdep_assert_held(&qp->s_lock);
1115         if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
1116                 return;
1117
1118         /* Find out where the BTH is */
1119         if (priv->hdr_type == HFI1_PKT_TYPE_9B) {
1120                 hdr = &opah->ibh;
1121                 if (ib_get_lnh(hdr) == HFI1_LRH_BTH)
1122                         ohdr = &hdr->u.oth;
1123                 else
1124                         ohdr = &hdr->u.l.oth;
1125         } else {
1126                 u8 l4;
1127
1128                 hdr_16b = &opah->opah;
1129                 l4  = hfi1_16B_get_l4(hdr_16b);
1130                 if (l4 == OPA_16B_L4_IB_LOCAL)
1131                         ohdr = &hdr_16b->u.oth;
1132                 else
1133                         ohdr = &hdr_16b->u.l.oth;
1134         }
1135
1136         opcode = ib_bth_get_opcode(ohdr);
1137         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1138             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1139                 WARN_ON(!qp->s_rdma_ack_cnt);
1140                 qp->s_rdma_ack_cnt--;
1141                 return;
1142         }
1143
1144         psn = ib_bth_get_psn(ohdr);
1145         reset_sending_psn(qp, psn);
1146
1147         /*
1148          * Start timer after a packet requesting an ACK has been sent and
1149          * there are still requests that haven't been acked.
1150          */
1151         if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
1152             !(qp->s_flags &
1153                 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1154                 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1155                 rvt_add_retry_timer(qp);
1156
1157         while (qp->s_last != qp->s_acked) {
1158                 u32 s_last;
1159
1160                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1161                 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1162                     cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1163                         break;
1164                 s_last = qp->s_last;
1165                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1166                 if (++s_last >= qp->s_size)
1167                         s_last = 0;
1168                 qp->s_last = s_last;
1169                 /* see post_send() */
1170                 barrier();
1171                 rvt_put_swqe(wqe);
1172                 rvt_qp_swqe_complete(qp,
1173                                      wqe,
1174                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1175                                      IB_WC_SUCCESS);
1176         }
1177         /*
1178          * If we were waiting for sends to complete before re-sending,
1179          * and they are now complete, restart sending.
1180          */
1181         trace_hfi1_sendcomplete(qp, psn);
1182         if (qp->s_flags & RVT_S_WAIT_PSN &&
1183             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1184                 qp->s_flags &= ~RVT_S_WAIT_PSN;
1185                 qp->s_sending_psn = qp->s_psn;
1186                 qp->s_sending_hpsn = qp->s_psn - 1;
1187                 hfi1_schedule_send(qp);
1188         }
1189 }
1190
1191 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1192 {
1193         qp->s_last_psn = psn;
1194 }
1195
1196 /*
1197  * Generate a SWQE completion.
1198  * This is similar to hfi1_send_complete but has to check to be sure
1199  * that the SGEs are not being referenced if the SWQE is being resent.
1200  */
1201 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1202                                          struct rvt_swqe *wqe,
1203                                          struct hfi1_ibport *ibp)
1204 {
1205         lockdep_assert_held(&qp->s_lock);
1206         /*
1207          * Don't decrement refcount and don't generate a
1208          * completion if the SWQE is being resent until the send
1209          * is finished.
1210          */
1211         if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1212             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1213                 u32 s_last;
1214
1215                 rvt_put_swqe(wqe);
1216                 s_last = qp->s_last;
1217                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1218                 if (++s_last >= qp->s_size)
1219                         s_last = 0;
1220                 qp->s_last = s_last;
1221                 /* see post_send() */
1222                 barrier();
1223                 rvt_qp_swqe_complete(qp,
1224                                      wqe,
1225                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1226                                      IB_WC_SUCCESS);
1227         } else {
1228                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1229
1230                 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1231                 /*
1232                  * If send progress not running attempt to progress
1233                  * SDMA queue.
1234                  */
1235                 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1236                         struct sdma_engine *engine;
1237                         u8 sl = rdma_ah_get_sl(&qp->remote_ah_attr);
1238                         u8 sc5;
1239
1240                         /* For now use sc to find engine */
1241                         sc5 = ibp->sl_to_sc[sl];
1242                         engine = qp_to_sdma_engine(qp, sc5);
1243                         sdma_engine_progress_schedule(engine);
1244                 }
1245         }
1246
1247         qp->s_retry = qp->s_retry_cnt;
1248         update_last_psn(qp, wqe->lpsn);
1249
1250         /*
1251          * If we are completing a request which is in the process of
1252          * being resent, we can stop re-sending it since we know the
1253          * responder has already seen it.
1254          */
1255         if (qp->s_acked == qp->s_cur) {
1256                 if (++qp->s_cur >= qp->s_size)
1257                         qp->s_cur = 0;
1258                 qp->s_acked = qp->s_cur;
1259                 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1260                 if (qp->s_acked != qp->s_tail) {
1261                         qp->s_state = OP(SEND_LAST);
1262                         qp->s_psn = wqe->psn;
1263                 }
1264         } else {
1265                 if (++qp->s_acked >= qp->s_size)
1266                         qp->s_acked = 0;
1267                 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1268                         qp->s_draining = 0;
1269                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1270         }
1271         return wqe;
1272 }
1273
1274 /**
1275  * do_rc_ack - process an incoming RC ACK
1276  * @qp: the QP the ACK came in on
1277  * @psn: the packet sequence number of the ACK
1278  * @opcode: the opcode of the request that resulted in the ACK
1279  *
1280  * This is called from rc_rcv_resp() to process an incoming RC ACK
1281  * for the given QP.
1282  * May be called at interrupt level, with the QP s_lock held.
1283  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1284  */
1285 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1286                      u64 val, struct hfi1_ctxtdata *rcd)
1287 {
1288         struct hfi1_ibport *ibp;
1289         enum ib_wc_status status;
1290         struct rvt_swqe *wqe;
1291         int ret = 0;
1292         u32 ack_psn;
1293         int diff;
1294
1295         lockdep_assert_held(&qp->s_lock);
1296         /*
1297          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1298          * requests and implicitly NAK RDMA read and atomic requests issued
1299          * before the NAK'ed request.  The MSN won't include the NAK'ed
1300          * request but will include an ACK'ed request(s).
1301          */
1302         ack_psn = psn;
1303         if (aeth >> IB_AETH_NAK_SHIFT)
1304                 ack_psn--;
1305         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1306         ibp = rcd_to_iport(rcd);
1307
1308         /*
1309          * The MSN might be for a later WQE than the PSN indicates so
1310          * only complete WQEs that the PSN finishes.
1311          */
1312         while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
1313                 /*
1314                  * RDMA_READ_RESPONSE_ONLY is a special case since
1315                  * we want to generate completion events for everything
1316                  * before the RDMA read, copy the data, then generate
1317                  * the completion for the read.
1318                  */
1319                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1320                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1321                     diff == 0) {
1322                         ret = 1;
1323                         goto bail_stop;
1324                 }
1325                 /*
1326                  * If this request is a RDMA read or atomic, and the ACK is
1327                  * for a later operation, this ACK NAKs the RDMA read or
1328                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1329                  * can ACK a RDMA read and likewise for atomic ops.  Note
1330                  * that the NAK case can only happen if relaxed ordering is
1331                  * used and requests are sent after an RDMA read or atomic
1332                  * is sent but before the response is received.
1333                  */
1334                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1335                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1336                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1337                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1338                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1339                         /* Retry this request. */
1340                         if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1341                                 qp->r_flags |= RVT_R_RDMAR_SEQ;
1342                                 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1343                                 if (list_empty(&qp->rspwait)) {
1344                                         qp->r_flags |= RVT_R_RSP_SEND;
1345                                         rvt_get_qp(qp);
1346                                         list_add_tail(&qp->rspwait,
1347                                                       &rcd->qp_wait_list);
1348                                 }
1349                         }
1350                         /*
1351                          * No need to process the ACK/NAK since we are
1352                          * restarting an earlier request.
1353                          */
1354                         goto bail_stop;
1355                 }
1356                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1357                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1358                         u64 *vaddr = wqe->sg_list[0].vaddr;
1359                         *vaddr = val;
1360                 }
1361                 if (qp->s_num_rd_atomic &&
1362                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
1363                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1364                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1365                         qp->s_num_rd_atomic--;
1366                         /* Restart sending task if fence is complete */
1367                         if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1368                             !qp->s_num_rd_atomic) {
1369                                 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1370                                                  RVT_S_WAIT_ACK);
1371                                 hfi1_schedule_send(qp);
1372                         } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1373                                 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1374                                                  RVT_S_WAIT_ACK);
1375                                 hfi1_schedule_send(qp);
1376                         }
1377                 }
1378                 wqe = do_rc_completion(qp, wqe, ibp);
1379                 if (qp->s_acked == qp->s_tail)
1380                         break;
1381         }
1382
1383         switch (aeth >> IB_AETH_NAK_SHIFT) {
1384         case 0:         /* ACK */
1385                 this_cpu_inc(*ibp->rvp.rc_acks);
1386                 if (qp->s_acked != qp->s_tail) {
1387                         /*
1388                          * We are expecting more ACKs so
1389                          * mod the retry timer.
1390                          */
1391                         rvt_mod_retry_timer(qp);
1392                         /*
1393                          * We can stop re-sending the earlier packets and
1394                          * continue with the next packet the receiver wants.
1395                          */
1396                         if (cmp_psn(qp->s_psn, psn) <= 0)
1397                                 reset_psn(qp, psn + 1);
1398                 } else {
1399                         /* No more acks - kill all timers */
1400                         rvt_stop_rc_timers(qp);
1401                         if (cmp_psn(qp->s_psn, psn) <= 0) {
1402                                 qp->s_state = OP(SEND_LAST);
1403                                 qp->s_psn = psn + 1;
1404                         }
1405                 }
1406                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1407                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1408                         hfi1_schedule_send(qp);
1409                 }
1410                 rvt_get_credit(qp, aeth);
1411                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1412                 qp->s_retry = qp->s_retry_cnt;
1413                 update_last_psn(qp, psn);
1414                 return 1;
1415
1416         case 1:         /* RNR NAK */
1417                 ibp->rvp.n_rnr_naks++;
1418                 if (qp->s_acked == qp->s_tail)
1419                         goto bail_stop;
1420                 if (qp->s_flags & RVT_S_WAIT_RNR)
1421                         goto bail_stop;
1422                 if (qp->s_rnr_retry == 0) {
1423                         status = IB_WC_RNR_RETRY_EXC_ERR;
1424                         goto class_b;
1425                 }
1426                 if (qp->s_rnr_retry_cnt < 7)
1427                         qp->s_rnr_retry--;
1428
1429                 /* The last valid PSN is the previous PSN. */
1430                 update_last_psn(qp, psn - 1);
1431
1432                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1433
1434                 reset_psn(qp, psn);
1435
1436                 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1437                 rvt_stop_rc_timers(qp);
1438                 rvt_add_rnr_timer(qp, aeth);
1439                 return 0;
1440
1441         case 3:         /* NAK */
1442                 if (qp->s_acked == qp->s_tail)
1443                         goto bail_stop;
1444                 /* The last valid PSN is the previous PSN. */
1445                 update_last_psn(qp, psn - 1);
1446                 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1447                         IB_AETH_CREDIT_MASK) {
1448                 case 0: /* PSN sequence error */
1449                         ibp->rvp.n_seq_naks++;
1450                         /*
1451                          * Back up to the responder's expected PSN.
1452                          * Note that we might get a NAK in the middle of an
1453                          * RDMA READ response which terminates the RDMA
1454                          * READ.
1455                          */
1456                         hfi1_restart_rc(qp, psn, 0);
1457                         hfi1_schedule_send(qp);
1458                         break;
1459
1460                 case 1: /* Invalid Request */
1461                         status = IB_WC_REM_INV_REQ_ERR;
1462                         ibp->rvp.n_other_naks++;
1463                         goto class_b;
1464
1465                 case 2: /* Remote Access Error */
1466                         status = IB_WC_REM_ACCESS_ERR;
1467                         ibp->rvp.n_other_naks++;
1468                         goto class_b;
1469
1470                 case 3: /* Remote Operation Error */
1471                         status = IB_WC_REM_OP_ERR;
1472                         ibp->rvp.n_other_naks++;
1473 class_b:
1474                         if (qp->s_last == qp->s_acked) {
1475                                 hfi1_send_complete(qp, wqe, status);
1476                                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1477                         }
1478                         break;
1479
1480                 default:
1481                         /* Ignore other reserved NAK error codes */
1482                         goto reserved;
1483                 }
1484                 qp->s_retry = qp->s_retry_cnt;
1485                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1486                 goto bail_stop;
1487
1488         default:                /* 2: reserved */
1489 reserved:
1490                 /* Ignore reserved NAK codes. */
1491                 goto bail_stop;
1492         }
1493         /* cannot be reached  */
1494 bail_stop:
1495         rvt_stop_rc_timers(qp);
1496         return ret;
1497 }
1498
1499 /*
1500  * We have seen an out of sequence RDMA read middle or last packet.
1501  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1502  */
1503 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
1504                          struct hfi1_ctxtdata *rcd)
1505 {
1506         struct rvt_swqe *wqe;
1507
1508         lockdep_assert_held(&qp->s_lock);
1509         /* Remove QP from retry timer */
1510         rvt_stop_rc_timers(qp);
1511
1512         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1513
1514         while (cmp_psn(psn, wqe->lpsn) > 0) {
1515                 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1516                     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1517                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1518                         break;
1519                 wqe = do_rc_completion(qp, wqe, ibp);
1520         }
1521
1522         ibp->rvp.n_rdma_seq++;
1523         qp->r_flags |= RVT_R_RDMAR_SEQ;
1524         hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1525         if (list_empty(&qp->rspwait)) {
1526                 qp->r_flags |= RVT_R_RSP_SEND;
1527                 rvt_get_qp(qp);
1528                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1529         }
1530 }
1531
1532 /**
1533  * rc_rcv_resp - process an incoming RC response packet
1534  * @packet: data packet information
1535  *
1536  * This is called from hfi1_rc_rcv() to process an incoming RC response
1537  * packet for the given QP.
1538  * Called at interrupt level.
1539  */
1540 static void rc_rcv_resp(struct hfi1_packet *packet)
1541 {
1542         struct hfi1_ctxtdata *rcd = packet->rcd;
1543         void *data = packet->payload;
1544         u32 tlen = packet->tlen;
1545         struct rvt_qp *qp = packet->qp;
1546         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
1547         struct ib_other_headers *ohdr = packet->ohdr;
1548         struct rvt_swqe *wqe;
1549         enum ib_wc_status status;
1550         unsigned long flags;
1551         int diff;
1552         u64 val;
1553         u32 aeth;
1554         u32 psn = ib_bth_get_psn(packet->ohdr);
1555         u32 pmtu = qp->pmtu;
1556         u16 hdrsize = packet->hlen;
1557         u8 opcode = packet->opcode;
1558         u8 pad = packet->pad;
1559         u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
1560
1561         spin_lock_irqsave(&qp->s_lock, flags);
1562         trace_hfi1_ack(qp, psn);
1563
1564         /* Ignore invalid responses. */
1565         smp_read_barrier_depends(); /* see post_one_send */
1566         if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1567                 goto ack_done;
1568
1569         /* Ignore duplicate responses. */
1570         diff = cmp_psn(psn, qp->s_last_psn);
1571         if (unlikely(diff <= 0)) {
1572                 /* Update credits for "ghost" ACKs */
1573                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1574                         aeth = be32_to_cpu(ohdr->u.aeth);
1575                         if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1576                                 rvt_get_credit(qp, aeth);
1577                 }
1578                 goto ack_done;
1579         }
1580
1581         /*
1582          * Skip everything other than the PSN we expect, if we are waiting
1583          * for a reply to a restarted RDMA read or atomic op.
1584          */
1585         if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1586                 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
1587                         goto ack_done;
1588                 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1589         }
1590
1591         if (unlikely(qp->s_acked == qp->s_tail))
1592                 goto ack_done;
1593         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1594         status = IB_WC_SUCCESS;
1595
1596         switch (opcode) {
1597         case OP(ACKNOWLEDGE):
1598         case OP(ATOMIC_ACKNOWLEDGE):
1599         case OP(RDMA_READ_RESPONSE_FIRST):
1600                 aeth = be32_to_cpu(ohdr->u.aeth);
1601                 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1602                         val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1603                 else
1604                         val = 0;
1605                 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1606                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1607                         goto ack_done;
1608                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1609                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1610                         goto ack_op_err;
1611                 /*
1612                  * If this is a response to a resent RDMA read, we
1613                  * have to be careful to copy the data to the right
1614                  * location.
1615                  */
1616                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1617                                                   wqe, psn, pmtu);
1618                 goto read_middle;
1619
1620         case OP(RDMA_READ_RESPONSE_MIDDLE):
1621                 /* no AETH, no ACK */
1622                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1623                         goto ack_seq_err;
1624                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1625                         goto ack_op_err;
1626 read_middle:
1627                 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
1628                         goto ack_len_err;
1629                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1630                         goto ack_len_err;
1631
1632                 /*
1633                  * We got a response so update the timeout.
1634                  * 4.096 usec. * (1 << qp->timeout)
1635                  */
1636                 rvt_mod_retry_timer(qp);
1637                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1638                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1639                         hfi1_schedule_send(qp);
1640                 }
1641
1642                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1643                         qp->s_retry = qp->s_retry_cnt;
1644
1645                 /*
1646                  * Update the RDMA receive state but do the copy w/o
1647                  * holding the locks and blocking interrupts.
1648                  */
1649                 qp->s_rdma_read_len -= pmtu;
1650                 update_last_psn(qp, psn);
1651                 spin_unlock_irqrestore(&qp->s_lock, flags);
1652                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, false, false);
1653                 goto bail;
1654
1655         case OP(RDMA_READ_RESPONSE_ONLY):
1656                 aeth = be32_to_cpu(ohdr->u.aeth);
1657                 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1658                         goto ack_done;
1659                 /*
1660                  * Check that the data size is >= 0 && <= pmtu.
1661                  * Remember to account for ICRC (4).
1662                  */
1663                 if (unlikely(tlen < (hdrsize + extra_bytes)))
1664                         goto ack_len_err;
1665                 /*
1666                  * If this is a response to a resent RDMA read, we
1667                  * have to be careful to copy the data to the right
1668                  * location.
1669                  */
1670                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1671                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1672                                                   wqe, psn, pmtu);
1673                 goto read_last;
1674
1675         case OP(RDMA_READ_RESPONSE_LAST):
1676                 /* ACKs READ req. */
1677                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1678                         goto ack_seq_err;
1679                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1680                         goto ack_op_err;
1681                 /*
1682                  * Check that the data size is >= 1 && <= pmtu.
1683                  * Remember to account for ICRC (4).
1684                  */
1685                 if (unlikely(tlen <= (hdrsize + extra_bytes)))
1686                         goto ack_len_err;
1687 read_last:
1688                 tlen -= hdrsize + extra_bytes;
1689                 if (unlikely(tlen != qp->s_rdma_read_len))
1690                         goto ack_len_err;
1691                 aeth = be32_to_cpu(ohdr->u.aeth);
1692                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, false, false);
1693                 WARN_ON(qp->s_rdma_read_sge.num_sge);
1694                 (void)do_rc_ack(qp, aeth, psn,
1695                                  OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1696                 goto ack_done;
1697         }
1698
1699 ack_op_err:
1700         status = IB_WC_LOC_QP_OP_ERR;
1701         goto ack_err;
1702
1703 ack_seq_err:
1704         rdma_seq_err(qp, ibp, psn, rcd);
1705         goto ack_done;
1706
1707 ack_len_err:
1708         status = IB_WC_LOC_LEN_ERR;
1709 ack_err:
1710         if (qp->s_last == qp->s_acked) {
1711                 hfi1_send_complete(qp, wqe, status);
1712                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1713         }
1714 ack_done:
1715         spin_unlock_irqrestore(&qp->s_lock, flags);
1716 bail:
1717         return;
1718 }
1719
1720 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
1721                                   struct rvt_qp *qp)
1722 {
1723         if (list_empty(&qp->rspwait)) {
1724                 qp->r_flags |= RVT_R_RSP_NAK;
1725                 rvt_get_qp(qp);
1726                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1727         }
1728 }
1729
1730 static inline void rc_cancel_ack(struct rvt_qp *qp)
1731 {
1732         qp->r_adefered = 0;
1733         if (list_empty(&qp->rspwait))
1734                 return;
1735         list_del_init(&qp->rspwait);
1736         qp->r_flags &= ~RVT_R_RSP_NAK;
1737         rvt_put_qp(qp);
1738 }
1739
1740 /**
1741  * rc_rcv_error - process an incoming duplicate or error RC packet
1742  * @ohdr: the other headers for this packet
1743  * @data: the packet data
1744  * @qp: the QP for this packet
1745  * @opcode: the opcode for this packet
1746  * @psn: the packet sequence number for this packet
1747  * @diff: the difference between the PSN and the expected PSN
1748  *
1749  * This is called from hfi1_rc_rcv() to process an unexpected
1750  * incoming RC packet for the given QP.
1751  * Called at interrupt level.
1752  * Return 1 if no more processing is needed; otherwise return 0 to
1753  * schedule a response to be sent.
1754  */
1755 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
1756                                  struct rvt_qp *qp, u32 opcode, u32 psn,
1757                                  int diff, struct hfi1_ctxtdata *rcd)
1758 {
1759         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1760         struct rvt_ack_entry *e;
1761         unsigned long flags;
1762         u8 i, prev;
1763         int old_req;
1764
1765         trace_hfi1_rcv_error(qp, psn);
1766         if (diff > 0) {
1767                 /*
1768                  * Packet sequence error.
1769                  * A NAK will ACK earlier sends and RDMA writes.
1770                  * Don't queue the NAK if we already sent one.
1771                  */
1772                 if (!qp->r_nak_state) {
1773                         ibp->rvp.n_rc_seqnak++;
1774                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1775                         /* Use the expected PSN. */
1776                         qp->r_ack_psn = qp->r_psn;
1777                         /*
1778                          * Wait to send the sequence NAK until all packets
1779                          * in the receive queue have been processed.
1780                          * Otherwise, we end up propagating congestion.
1781                          */
1782                         rc_defered_ack(rcd, qp);
1783                 }
1784                 goto done;
1785         }
1786
1787         /*
1788          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1789          * write or atomic op.  Don't NAK errors, just silently drop
1790          * the duplicate request.  Note that r_sge, r_len, and
1791          * r_rcv_len may be in use so don't modify them.
1792          *
1793          * We are supposed to ACK the earliest duplicate PSN but we
1794          * can coalesce an outstanding duplicate ACK.  We have to
1795          * send the earliest so that RDMA reads can be restarted at
1796          * the requester's expected PSN.
1797          *
1798          * First, find where this duplicate PSN falls within the
1799          * ACKs previously sent.
1800          * old_req is true if there is an older response that is scheduled
1801          * to be sent before sending this one.
1802          */
1803         e = NULL;
1804         old_req = 1;
1805         ibp->rvp.n_rc_dupreq++;
1806
1807         spin_lock_irqsave(&qp->s_lock, flags);
1808
1809         for (i = qp->r_head_ack_queue; ; i = prev) {
1810                 if (i == qp->s_tail_ack_queue)
1811                         old_req = 0;
1812                 if (i)
1813                         prev = i - 1;
1814                 else
1815                         prev = HFI1_MAX_RDMA_ATOMIC;
1816                 if (prev == qp->r_head_ack_queue) {
1817                         e = NULL;
1818                         break;
1819                 }
1820                 e = &qp->s_ack_queue[prev];
1821                 if (!e->opcode) {
1822                         e = NULL;
1823                         break;
1824                 }
1825                 if (cmp_psn(psn, e->psn) >= 0) {
1826                         if (prev == qp->s_tail_ack_queue &&
1827                             cmp_psn(psn, e->lpsn) <= 0)
1828                                 old_req = 0;
1829                         break;
1830                 }
1831         }
1832         switch (opcode) {
1833         case OP(RDMA_READ_REQUEST): {
1834                 struct ib_reth *reth;
1835                 u32 offset;
1836                 u32 len;
1837
1838                 /*
1839                  * If we didn't find the RDMA read request in the ack queue,
1840                  * we can ignore this request.
1841                  */
1842                 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1843                         goto unlock_done;
1844                 /* RETH comes after BTH */
1845                 reth = &ohdr->u.rc.reth;
1846                 /*
1847                  * Address range must be a subset of the original
1848                  * request and start on pmtu boundaries.
1849                  * We reuse the old ack_queue slot since the requester
1850                  * should not back up and request an earlier PSN for the
1851                  * same request.
1852                  */
1853                 offset = delta_psn(psn, e->psn) * qp->pmtu;
1854                 len = be32_to_cpu(reth->length);
1855                 if (unlikely(offset + len != e->rdma_sge.sge_length))
1856                         goto unlock_done;
1857                 if (e->rdma_sge.mr) {
1858                         rvt_put_mr(e->rdma_sge.mr);
1859                         e->rdma_sge.mr = NULL;
1860                 }
1861                 if (len != 0) {
1862                         u32 rkey = be32_to_cpu(reth->rkey);
1863                         u64 vaddr = get_ib_reth_vaddr(reth);
1864                         int ok;
1865
1866                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1867                                          IB_ACCESS_REMOTE_READ);
1868                         if (unlikely(!ok))
1869                                 goto unlock_done;
1870                 } else {
1871                         e->rdma_sge.vaddr = NULL;
1872                         e->rdma_sge.length = 0;
1873                         e->rdma_sge.sge_length = 0;
1874                 }
1875                 e->psn = psn;
1876                 if (old_req)
1877                         goto unlock_done;
1878                 qp->s_tail_ack_queue = prev;
1879                 break;
1880         }
1881
1882         case OP(COMPARE_SWAP):
1883         case OP(FETCH_ADD): {
1884                 /*
1885                  * If we didn't find the atomic request in the ack queue
1886                  * or the send engine is already backed up to send an
1887                  * earlier entry, we can ignore this request.
1888                  */
1889                 if (!e || e->opcode != (u8)opcode || old_req)
1890                         goto unlock_done;
1891                 qp->s_tail_ack_queue = prev;
1892                 break;
1893         }
1894
1895         default:
1896                 /*
1897                  * Ignore this operation if it doesn't request an ACK
1898                  * or an earlier RDMA read or atomic is going to be resent.
1899                  */
1900                 if (!(psn & IB_BTH_REQ_ACK) || old_req)
1901                         goto unlock_done;
1902                 /*
1903                  * Resend the most recent ACK if this request is
1904                  * after all the previous RDMA reads and atomics.
1905                  */
1906                 if (i == qp->r_head_ack_queue) {
1907                         spin_unlock_irqrestore(&qp->s_lock, flags);
1908                         qp->r_nak_state = 0;
1909                         qp->r_ack_psn = qp->r_psn - 1;
1910                         goto send_ack;
1911                 }
1912
1913                 /*
1914                  * Resend the RDMA read or atomic op which
1915                  * ACKs this duplicate request.
1916                  */
1917                 qp->s_tail_ack_queue = i;
1918                 break;
1919         }
1920         qp->s_ack_state = OP(ACKNOWLEDGE);
1921         qp->s_flags |= RVT_S_RESP_PENDING;
1922         qp->r_nak_state = 0;
1923         hfi1_schedule_send(qp);
1924
1925 unlock_done:
1926         spin_unlock_irqrestore(&qp->s_lock, flags);
1927 done:
1928         return 1;
1929
1930 send_ack:
1931         return 0;
1932 }
1933
1934 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n)
1935 {
1936         unsigned next;
1937
1938         next = n + 1;
1939         if (next > HFI1_MAX_RDMA_ATOMIC)
1940                 next = 0;
1941         qp->s_tail_ack_queue = next;
1942         qp->s_ack_state = OP(ACKNOWLEDGE);
1943 }
1944
1945 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
1946                           u32 lqpn, u32 rqpn, u8 svc_type)
1947 {
1948         struct opa_hfi1_cong_log_event_internal *cc_event;
1949         unsigned long flags;
1950
1951         if (sl >= OPA_MAX_SLS)
1952                 return;
1953
1954         spin_lock_irqsave(&ppd->cc_log_lock, flags);
1955
1956         ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
1957         ppd->threshold_event_counter++;
1958
1959         cc_event = &ppd->cc_events[ppd->cc_log_idx++];
1960         if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
1961                 ppd->cc_log_idx = 0;
1962         cc_event->lqpn = lqpn & RVT_QPN_MASK;
1963         cc_event->rqpn = rqpn & RVT_QPN_MASK;
1964         cc_event->sl = sl;
1965         cc_event->svc_type = svc_type;
1966         cc_event->rlid = rlid;
1967         /* keep timestamp in units of 1.024 usec */
1968         cc_event->timestamp = ktime_get_ns() / 1024;
1969
1970         spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
1971 }
1972
1973 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u32 rlid, u32 lqpn,
1974                   u32 rqpn, u8 svc_type)
1975 {
1976         struct cca_timer *cca_timer;
1977         u16 ccti, ccti_incr, ccti_timer, ccti_limit;
1978         u8 trigger_threshold;
1979         struct cc_state *cc_state;
1980         unsigned long flags;
1981
1982         if (sl >= OPA_MAX_SLS)
1983                 return;
1984
1985         cc_state = get_cc_state(ppd);
1986
1987         if (!cc_state)
1988                 return;
1989
1990         /*
1991          * 1) increase CCTI (for this SL)
1992          * 2) select IPG (i.e., call set_link_ipg())
1993          * 3) start timer
1994          */
1995         ccti_limit = cc_state->cct.ccti_limit;
1996         ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
1997         ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
1998         trigger_threshold =
1999                 cc_state->cong_setting.entries[sl].trigger_threshold;
2000
2001         spin_lock_irqsave(&ppd->cca_timer_lock, flags);
2002
2003         cca_timer = &ppd->cca_timer[sl];
2004         if (cca_timer->ccti < ccti_limit) {
2005                 if (cca_timer->ccti + ccti_incr <= ccti_limit)
2006                         cca_timer->ccti += ccti_incr;
2007                 else
2008                         cca_timer->ccti = ccti_limit;
2009                 set_link_ipg(ppd);
2010         }
2011
2012         ccti = cca_timer->ccti;
2013
2014         if (!hrtimer_active(&cca_timer->hrtimer)) {
2015                 /* ccti_timer is in units of 1.024 usec */
2016                 unsigned long nsec = 1024 * ccti_timer;
2017
2018                 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
2019                               HRTIMER_MODE_REL);
2020         }
2021
2022         spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
2023
2024         if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
2025                 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
2026 }
2027
2028 /**
2029  * hfi1_rc_rcv - process an incoming RC packet
2030  * @packet: data packet information
2031  *
2032  * This is called from qp_rcv() to process an incoming RC packet
2033  * for the given QP.
2034  * May be called at interrupt level.
2035  */
2036 void hfi1_rc_rcv(struct hfi1_packet *packet)
2037 {
2038         struct hfi1_ctxtdata *rcd = packet->rcd;
2039         void *data = packet->payload;
2040         u32 tlen = packet->tlen;
2041         struct rvt_qp *qp = packet->qp;
2042         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2043         struct ib_other_headers *ohdr = packet->ohdr;
2044         u32 bth0 = be32_to_cpu(ohdr->bth[0]);
2045         u32 opcode = packet->opcode;
2046         u32 hdrsize = packet->hlen;
2047         u32 psn = ib_bth_get_psn(packet->ohdr);
2048         u32 pad = packet->pad;
2049         struct ib_wc wc;
2050         u32 pmtu = qp->pmtu;
2051         int diff;
2052         struct ib_reth *reth;
2053         unsigned long flags;
2054         int ret;
2055         bool is_fecn = false;
2056         bool copy_last = false;
2057         u32 rkey;
2058         u8 extra_bytes = pad + packet->extra_byte + (SIZE_OF_CRC << 2);
2059
2060         lockdep_assert_held(&qp->r_lock);
2061
2062         if (hfi1_ruc_check_hdr(ibp, packet))
2063                 return;
2064
2065         is_fecn = process_ecn(qp, packet, false);
2066
2067         /*
2068          * Process responses (ACKs) before anything else.  Note that the
2069          * packet sequence number will be for something in the send work
2070          * queue rather than the expected receive packet sequence number.
2071          * In other words, this QP is the requester.
2072          */
2073         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
2074             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
2075                 rc_rcv_resp(packet);
2076                 if (is_fecn)
2077                         goto send_ack;
2078                 return;
2079         }
2080
2081         /* Compute 24 bits worth of difference. */
2082         diff = delta_psn(psn, qp->r_psn);
2083         if (unlikely(diff)) {
2084                 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
2085                         return;
2086                 goto send_ack;
2087         }
2088
2089         /* Check for opcode sequence errors. */
2090         switch (qp->r_state) {
2091         case OP(SEND_FIRST):
2092         case OP(SEND_MIDDLE):
2093                 if (opcode == OP(SEND_MIDDLE) ||
2094                     opcode == OP(SEND_LAST) ||
2095                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2096                     opcode == OP(SEND_LAST_WITH_INVALIDATE))
2097                         break;
2098                 goto nack_inv;
2099
2100         case OP(RDMA_WRITE_FIRST):
2101         case OP(RDMA_WRITE_MIDDLE):
2102                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
2103                     opcode == OP(RDMA_WRITE_LAST) ||
2104                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2105                         break;
2106                 goto nack_inv;
2107
2108         default:
2109                 if (opcode == OP(SEND_MIDDLE) ||
2110                     opcode == OP(SEND_LAST) ||
2111                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
2112                     opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
2113                     opcode == OP(RDMA_WRITE_MIDDLE) ||
2114                     opcode == OP(RDMA_WRITE_LAST) ||
2115                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
2116                         goto nack_inv;
2117                 /*
2118                  * Note that it is up to the requester to not send a new
2119                  * RDMA read or atomic operation before receiving an ACK
2120                  * for the previous operation.
2121                  */
2122                 break;
2123         }
2124
2125         if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2126                 rvt_comm_est(qp);
2127
2128         /* OK, process the packet. */
2129         switch (opcode) {
2130         case OP(SEND_FIRST):
2131                 ret = hfi1_rvt_get_rwqe(qp, 0);
2132                 if (ret < 0)
2133                         goto nack_op_err;
2134                 if (!ret)
2135                         goto rnr_nak;
2136                 qp->r_rcv_len = 0;
2137                 /* FALLTHROUGH */
2138         case OP(SEND_MIDDLE):
2139         case OP(RDMA_WRITE_MIDDLE):
2140 send_middle:
2141                 /* Check for invalid length PMTU or posted rwqe len. */
2142                 /*
2143                  * There will be no padding for 9B packet but 16B packets
2144                  * will come in with some padding since we always add
2145                  * CRC and LT bytes which will need to be flit aligned
2146                  */
2147                 if (unlikely(tlen != (hdrsize + pmtu + extra_bytes)))
2148                         goto nack_inv;
2149                 qp->r_rcv_len += pmtu;
2150                 if (unlikely(qp->r_rcv_len > qp->r_len))
2151                         goto nack_inv;
2152                 hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
2153                 break;
2154
2155         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2156                 /* consume RWQE */
2157                 ret = hfi1_rvt_get_rwqe(qp, 1);
2158                 if (ret < 0)
2159                         goto nack_op_err;
2160                 if (!ret)
2161                         goto rnr_nak;
2162                 goto send_last_imm;
2163
2164         case OP(SEND_ONLY):
2165         case OP(SEND_ONLY_WITH_IMMEDIATE):
2166         case OP(SEND_ONLY_WITH_INVALIDATE):
2167                 ret = hfi1_rvt_get_rwqe(qp, 0);
2168                 if (ret < 0)
2169                         goto nack_op_err;
2170                 if (!ret)
2171                         goto rnr_nak;
2172                 qp->r_rcv_len = 0;
2173                 if (opcode == OP(SEND_ONLY))
2174                         goto no_immediate_data;
2175                 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2176                         goto send_last_inv;
2177                 /* FALLTHROUGH -- for SEND_ONLY_WITH_IMMEDIATE */
2178         case OP(SEND_LAST_WITH_IMMEDIATE):
2179 send_last_imm:
2180                 wc.ex.imm_data = ohdr->u.imm_data;
2181                 wc.wc_flags = IB_WC_WITH_IMM;
2182                 goto send_last;
2183         case OP(SEND_LAST_WITH_INVALIDATE):
2184 send_last_inv:
2185                 rkey = be32_to_cpu(ohdr->u.ieth);
2186                 if (rvt_invalidate_rkey(qp, rkey))
2187                         goto no_immediate_data;
2188                 wc.ex.invalidate_rkey = rkey;
2189                 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2190                 goto send_last;
2191         case OP(RDMA_WRITE_LAST):
2192                 copy_last = rvt_is_user_qp(qp);
2193                 /* fall through */
2194         case OP(SEND_LAST):
2195 no_immediate_data:
2196                 wc.wc_flags = 0;
2197                 wc.ex.imm_data = 0;
2198 send_last:
2199                 /* Check for invalid length. */
2200                 /* LAST len should be >= 1 */
2201                 if (unlikely(tlen < (hdrsize + extra_bytes)))
2202                         goto nack_inv;
2203                 /* Don't count the CRC(and padding and LT byte for 16B). */
2204                 tlen -= (hdrsize + extra_bytes);
2205                 wc.byte_len = tlen + qp->r_rcv_len;
2206                 if (unlikely(wc.byte_len > qp->r_len))
2207                         goto nack_inv;
2208                 hfi1_copy_sge(&qp->r_sge, data, tlen, true, copy_last);
2209                 rvt_put_ss(&qp->r_sge);
2210                 qp->r_msn++;
2211                 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2212                         break;
2213                 wc.wr_id = qp->r_wr_id;
2214                 wc.status = IB_WC_SUCCESS;
2215                 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2216                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2217                         wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2218                 else
2219                         wc.opcode = IB_WC_RECV;
2220                 wc.qp = &qp->ibqp;
2221                 wc.src_qp = qp->remote_qpn;
2222                 wc.slid = rdma_ah_get_dlid(&qp->remote_ah_attr) & U16_MAX;
2223                 /*
2224                  * It seems that IB mandates the presence of an SL in a
2225                  * work completion only for the UD transport (see section
2226                  * 11.4.2 of IBTA Vol. 1).
2227                  *
2228                  * However, the way the SL is chosen below is consistent
2229                  * with the way that IB/qib works and is trying avoid
2230                  * introducing incompatibilities.
2231                  *
2232                  * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2233                  */
2234                 wc.sl = rdma_ah_get_sl(&qp->remote_ah_attr);
2235                 /* zero fields that are N/A */
2236                 wc.vendor_err = 0;
2237                 wc.pkey_index = 0;
2238                 wc.dlid_path_bits = 0;
2239                 wc.port_num = 0;
2240                 /* Signal completion event if the solicited bit is set. */
2241                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
2242                              (bth0 & IB_BTH_SOLICITED) != 0);
2243                 break;
2244
2245         case OP(RDMA_WRITE_ONLY):
2246                 copy_last = rvt_is_user_qp(qp);
2247                 /* fall through */
2248         case OP(RDMA_WRITE_FIRST):
2249         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
2250                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
2251                         goto nack_inv;
2252                 /* consume RWQE */
2253                 reth = &ohdr->u.rc.reth;
2254                 qp->r_len = be32_to_cpu(reth->length);
2255                 qp->r_rcv_len = 0;
2256                 qp->r_sge.sg_list = NULL;
2257                 if (qp->r_len != 0) {
2258                         u32 rkey = be32_to_cpu(reth->rkey);
2259                         u64 vaddr = get_ib_reth_vaddr(reth);
2260                         int ok;
2261
2262                         /* Check rkey & NAK */
2263                         ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
2264                                          rkey, IB_ACCESS_REMOTE_WRITE);
2265                         if (unlikely(!ok))
2266                                 goto nack_acc;
2267                         qp->r_sge.num_sge = 1;
2268                 } else {
2269                         qp->r_sge.num_sge = 0;
2270                         qp->r_sge.sge.mr = NULL;
2271                         qp->r_sge.sge.vaddr = NULL;
2272                         qp->r_sge.sge.length = 0;
2273                         qp->r_sge.sge.sge_length = 0;
2274                 }
2275                 if (opcode == OP(RDMA_WRITE_FIRST))
2276                         goto send_middle;
2277                 else if (opcode == OP(RDMA_WRITE_ONLY))
2278                         goto no_immediate_data;
2279                 ret = hfi1_rvt_get_rwqe(qp, 1);
2280                 if (ret < 0)
2281                         goto nack_op_err;
2282                 if (!ret) {
2283                         /* peer will send again */
2284                         rvt_put_ss(&qp->r_sge);
2285                         goto rnr_nak;
2286                 }
2287                 wc.ex.imm_data = ohdr->u.rc.imm_data;
2288                 wc.wc_flags = IB_WC_WITH_IMM;
2289                 goto send_last;
2290
2291         case OP(RDMA_READ_REQUEST): {
2292                 struct rvt_ack_entry *e;
2293                 u32 len;
2294                 u8 next;
2295
2296                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2297                         goto nack_inv;
2298                 next = qp->r_head_ack_queue + 1;
2299                 /* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */
2300                 if (next > HFI1_MAX_RDMA_ATOMIC)
2301                         next = 0;
2302                 spin_lock_irqsave(&qp->s_lock, flags);
2303                 if (unlikely(next == qp->s_tail_ack_queue)) {
2304                         if (!qp->s_ack_queue[next].sent)
2305                                 goto nack_inv_unlck;
2306                         update_ack_queue(qp, next);
2307                 }
2308                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2309                 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2310                         rvt_put_mr(e->rdma_sge.mr);
2311                         e->rdma_sge.mr = NULL;
2312                 }
2313                 reth = &ohdr->u.rc.reth;
2314                 len = be32_to_cpu(reth->length);
2315                 if (len) {
2316                         u32 rkey = be32_to_cpu(reth->rkey);
2317                         u64 vaddr = get_ib_reth_vaddr(reth);
2318                         int ok;
2319
2320                         /* Check rkey & NAK */
2321                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
2322                                          rkey, IB_ACCESS_REMOTE_READ);
2323                         if (unlikely(!ok))
2324                                 goto nack_acc_unlck;
2325                         /*
2326                          * Update the next expected PSN.  We add 1 later
2327                          * below, so only add the remainder here.
2328                          */
2329                         qp->r_psn += rvt_div_mtu(qp, len - 1);
2330                 } else {
2331                         e->rdma_sge.mr = NULL;
2332                         e->rdma_sge.vaddr = NULL;
2333                         e->rdma_sge.length = 0;
2334                         e->rdma_sge.sge_length = 0;
2335                 }
2336                 e->opcode = opcode;
2337                 e->sent = 0;
2338                 e->psn = psn;
2339                 e->lpsn = qp->r_psn;
2340                 /*
2341                  * We need to increment the MSN here instead of when we
2342                  * finish sending the result since a duplicate request would
2343                  * increment it more than once.
2344                  */
2345                 qp->r_msn++;
2346                 qp->r_psn++;
2347                 qp->r_state = opcode;
2348                 qp->r_nak_state = 0;
2349                 qp->r_head_ack_queue = next;
2350
2351                 /* Schedule the send engine. */
2352                 qp->s_flags |= RVT_S_RESP_PENDING;
2353                 hfi1_schedule_send(qp);
2354
2355                 spin_unlock_irqrestore(&qp->s_lock, flags);
2356                 if (is_fecn)
2357                         goto send_ack;
2358                 return;
2359         }
2360
2361         case OP(COMPARE_SWAP):
2362         case OP(FETCH_ADD): {
2363                 struct ib_atomic_eth *ateth;
2364                 struct rvt_ack_entry *e;
2365                 u64 vaddr;
2366                 atomic64_t *maddr;
2367                 u64 sdata;
2368                 u32 rkey;
2369                 u8 next;
2370
2371                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2372                         goto nack_inv;
2373                 next = qp->r_head_ack_queue + 1;
2374                 if (next > HFI1_MAX_RDMA_ATOMIC)
2375                         next = 0;
2376                 spin_lock_irqsave(&qp->s_lock, flags);
2377                 if (unlikely(next == qp->s_tail_ack_queue)) {
2378                         if (!qp->s_ack_queue[next].sent)
2379                                 goto nack_inv_unlck;
2380                         update_ack_queue(qp, next);
2381                 }
2382                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2383                 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2384                         rvt_put_mr(e->rdma_sge.mr);
2385                         e->rdma_sge.mr = NULL;
2386                 }
2387                 ateth = &ohdr->u.atomic_eth;
2388                 vaddr = get_ib_ateth_vaddr(ateth);
2389                 if (unlikely(vaddr & (sizeof(u64) - 1)))
2390                         goto nack_inv_unlck;
2391                 rkey = be32_to_cpu(ateth->rkey);
2392                 /* Check rkey & NAK */
2393                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2394                                           vaddr, rkey,
2395                                           IB_ACCESS_REMOTE_ATOMIC)))
2396                         goto nack_acc_unlck;
2397                 /* Perform atomic OP and save result. */
2398                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
2399                 sdata = get_ib_ateth_swap(ateth);
2400                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2401                         (u64)atomic64_add_return(sdata, maddr) - sdata :
2402                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
2403                                       get_ib_ateth_compare(ateth),
2404                                       sdata);
2405                 rvt_put_mr(qp->r_sge.sge.mr);
2406                 qp->r_sge.num_sge = 0;
2407                 e->opcode = opcode;
2408                 e->sent = 0;
2409                 e->psn = psn;
2410                 e->lpsn = psn;
2411                 qp->r_msn++;
2412                 qp->r_psn++;
2413                 qp->r_state = opcode;
2414                 qp->r_nak_state = 0;
2415                 qp->r_head_ack_queue = next;
2416
2417                 /* Schedule the send engine. */
2418                 qp->s_flags |= RVT_S_RESP_PENDING;
2419                 hfi1_schedule_send(qp);
2420
2421                 spin_unlock_irqrestore(&qp->s_lock, flags);
2422                 if (is_fecn)
2423                         goto send_ack;
2424                 return;
2425         }
2426
2427         default:
2428                 /* NAK unknown opcodes. */
2429                 goto nack_inv;
2430         }
2431         qp->r_psn++;
2432         qp->r_state = opcode;
2433         qp->r_ack_psn = psn;
2434         qp->r_nak_state = 0;
2435         /* Send an ACK if requested or required. */
2436         if (psn & IB_BTH_REQ_ACK) {
2437                 if (packet->numpkt == 0) {
2438                         rc_cancel_ack(qp);
2439                         goto send_ack;
2440                 }
2441                 if (qp->r_adefered >= HFI1_PSN_CREDIT) {
2442                         rc_cancel_ack(qp);
2443                         goto send_ack;
2444                 }
2445                 if (unlikely(is_fecn)) {
2446                         rc_cancel_ack(qp);
2447                         goto send_ack;
2448                 }
2449                 qp->r_adefered++;
2450                 rc_defered_ack(rcd, qp);
2451         }
2452         return;
2453
2454 rnr_nak:
2455         qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
2456         qp->r_ack_psn = qp->r_psn;
2457         /* Queue RNR NAK for later */
2458         rc_defered_ack(rcd, qp);
2459         return;
2460
2461 nack_op_err:
2462         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2463         qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2464         qp->r_ack_psn = qp->r_psn;
2465         /* Queue NAK for later */
2466         rc_defered_ack(rcd, qp);
2467         return;
2468
2469 nack_inv_unlck:
2470         spin_unlock_irqrestore(&qp->s_lock, flags);
2471 nack_inv:
2472         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2473         qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2474         qp->r_ack_psn = qp->r_psn;
2475         /* Queue NAK for later */
2476         rc_defered_ack(rcd, qp);
2477         return;
2478
2479 nack_acc_unlck:
2480         spin_unlock_irqrestore(&qp->s_lock, flags);
2481 nack_acc:
2482         rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2483         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2484         qp->r_ack_psn = qp->r_psn;
2485 send_ack:
2486         hfi1_send_rc_ack(rcd, qp, is_fecn);
2487 }
2488
2489 void hfi1_rc_hdrerr(
2490         struct hfi1_ctxtdata *rcd,
2491         struct hfi1_packet *packet,
2492         struct rvt_qp *qp)
2493 {
2494         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2495         int diff;
2496         u32 opcode;
2497         u32 psn;
2498
2499         if (hfi1_ruc_check_hdr(ibp, packet))
2500                 return;
2501
2502         psn = ib_bth_get_psn(packet->ohdr);
2503         opcode = ib_bth_get_opcode(packet->ohdr);
2504
2505         /* Only deal with RDMA Writes for now */
2506         if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
2507                 diff = delta_psn(psn, qp->r_psn);
2508                 if (!qp->r_nak_state && diff >= 0) {
2509                         ibp->rvp.n_rc_seqnak++;
2510                         qp->r_nak_state = IB_NAK_PSN_ERROR;
2511                         /* Use the expected PSN. */
2512                         qp->r_ack_psn = qp->r_psn;
2513                         /*
2514                          * Wait to send the sequence
2515                          * NAK until all packets
2516                          * in the receive queue have
2517                          * been processed.
2518                          * Otherwise, we end up
2519                          * propagating congestion.
2520                          */
2521                         rc_defered_ack(rcd, qp);
2522                 } /* Out of sequence NAK */
2523         } /* QP Request NAKs */
2524 }