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         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 104         hwords = 5;
 105
 106         switch (qp->s_ack_state) {
 107         case OP(RDMA_READ_RESPONSE_LAST):
 108         case OP(RDMA_READ_RESPONSE_ONLY):
 109                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 110                 if (e->rdma_sge.mr) {
 111                         rvt_put_mr(e->rdma_sge.mr);
 112                         e->rdma_sge.mr = NULL;
 113                 }
 114                 /* FALLTHROUGH */
 115         case OP(ATOMIC_ACKNOWLEDGE):
 116                 /*
 117                  * We can increment the tail pointer now that the last
 118                  * response has been sent instead of only being
 119                  * constructed.
 120                  */
 121                 if (++qp->s_tail_ack_queue > HFI1_MAX_RDMA_ATOMIC)
 122                         qp->s_tail_ack_queue = 0;
 123                 /* FALLTHROUGH */
 124         case OP(SEND_ONLY):
 125         case OP(ACKNOWLEDGE):
 126                 /* Check for no next entry in the queue. */
 127                 if (qp->r_head_ack_queue == qp->s_tail_ack_queue) {
 128                         if (qp->s_flags & RVT_S_ACK_PENDING)
 129                                 goto normal;
 130                         goto bail;
 131                 }
 132
 133                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 134                 if (e->opcode == OP(RDMA_READ_REQUEST)) {
 135                         /*
 136                          * If a RDMA read response is being resent and
 137                          * we haven't seen the duplicate request yet,
 138                          * then stop sending the remaining responses the
 139                          * responder has seen until the requester re-sends it.
 140                          */
 141                         len = e->rdma_sge.sge_length;
 142                         if (len && !e->rdma_sge.mr) {
 143                                 qp->s_tail_ack_queue = qp->r_head_ack_queue;
 144                                 goto bail;
 145                         }
 146                         /* Copy SGE state in case we need to resend */
 147                         ps->s_txreq->mr = e->rdma_sge.mr;
 148                         if (ps->s_txreq->mr)
 149                                 rvt_get_mr(ps->s_txreq->mr);
 150                         qp->s_ack_rdma_sge.sge = e->rdma_sge;
 151                         qp->s_ack_rdma_sge.num_sge = 1;
 152                         ps->s_txreq->ss = &qp->s_ack_rdma_sge;
 153                         if (len > pmtu) {
 154                                 len = pmtu;
 155                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
 156                         } else {
 157                                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 158                                 e->sent = 1;
 159                         }
 160                         ohdr->u.aeth = rvt_compute_aeth(qp);
 161                         hwords++;
 162                         qp->s_ack_rdma_psn = e->psn;
 163                         bth2 = mask_psn(qp->s_ack_rdma_psn++);
 164                 } else {
 165                         /* COMPARE_SWAP or FETCH_ADD */
 166                         ps->s_txreq->ss = NULL;
 167                         len = 0;
 168                         qp->s_ack_state = OP(ATOMIC_ACKNOWLEDGE);
 169                         ohdr->u.at.aeth = rvt_compute_aeth(qp);
 170                         ib_u64_put(e->atomic_data, &ohdr->u.at.atomic_ack_eth);
 171                         hwords += sizeof(ohdr->u.at) / sizeof(u32);
 172                         bth2 = mask_psn(e->psn);
 173                         e->sent = 1;
 174                 }
 175                 bth0 = qp->s_ack_state << 24;
 176                 break;
 177
 178         case OP(RDMA_READ_RESPONSE_FIRST):
 179                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 180                 /* FALLTHROUGH */
 181         case OP(RDMA_READ_RESPONSE_MIDDLE):
 182                 ps->s_txreq->ss = &qp->s_ack_rdma_sge;
 183                 ps->s_txreq->mr = qp->s_ack_rdma_sge.sge.mr;
 184                 if (ps->s_txreq->mr)
 185                         rvt_get_mr(ps->s_txreq->mr);
 186                 len = qp->s_ack_rdma_sge.sge.sge_length;
 187                 if (len > pmtu) {
 188                         len = pmtu;
 189                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 190                 } else {
 191                         ohdr->u.aeth = rvt_compute_aeth(qp);
 192                         hwords++;
 193                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 194                         e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 195                         e->sent = 1;
 196                 }
 197                 bth0 = qp->s_ack_state << 24;
 198                 bth2 = mask_psn(qp->s_ack_rdma_psn++);
 199                 break;
 200
 201         default:
 202 normal:
 203                 /*
 204                  * Send a regular ACK.
 205                  * Set the s_ack_state so we wait until after sending
 206                  * the ACK before setting s_ack_state to ACKNOWLEDGE
 207                  * (see above).
 208                  */
 209                 qp->s_ack_state = OP(SEND_ONLY);
 210                 qp->s_flags &= ~RVT_S_ACK_PENDING;
 211                 ps->s_txreq->ss = NULL;
 212                 if (qp->s_nak_state)
 213                         ohdr->u.aeth =
 214                                 cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
 215                                             (qp->s_nak_state <<
 216                                              IB_AETH_CREDIT_SHIFT));
 217                 else
 218                         ohdr->u.aeth = rvt_compute_aeth(qp);
 219                 hwords++;
 220                 len = 0;
 221                 bth0 = OP(ACKNOWLEDGE) << 24;
 222                 bth2 = mask_psn(qp->s_ack_psn);
 223         }
 224         qp->s_rdma_ack_cnt++;
 225         qp->s_hdrwords = hwords;
 226         ps->s_txreq->sde = priv->s_sde;
 227         ps->s_txreq->s_cur_size = len;
 228         hfi1_make_ruc_header(qp, ohdr, bth0, bth2, middle, ps);
 229         /* pbc */
 230         ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
 231         return 1;
 232
 233 bail:
 234         qp->s_ack_state = OP(ACKNOWLEDGE);
 235         /*
 236          * Ensure s_rdma_ack_cnt changes are committed prior to resetting
 237          * RVT_S_RESP_PENDING
 238          */
 239         smp_wmb();
 240         qp->s_flags &= ~(RVT_S_RESP_PENDING
 241                                 | RVT_S_ACK_PENDING
 242                                 | RVT_S_AHG_VALID);
 243         return 0;
 244 }
 245
 246 /**
 247  * hfi1_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 248  * @qp: a pointer to the QP
 249  *
 250  * Assumes s_lock is held.
 251  *
 252  * Return 1 if constructed; otherwise, return 0.
 253  */
 254 int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 255 {
 256         struct hfi1_qp_priv *priv = qp->priv;
 257         struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 258         struct ib_other_headers *ohdr;
 259         struct rvt_sge_state *ss;
 260         struct rvt_swqe *wqe;
 261         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 262         u32 hwords = 5;
 263         u32 len;
 264         u32 bth0 = 0;
 265         u32 bth2;
 266         u32 pmtu = qp->pmtu;
 267         char newreq;
 268         int middle = 0;
 269         int delta;
 270
 271         lockdep_assert_held(&qp->s_lock);
 272         ps->s_txreq = get_txreq(ps->dev, qp);
 273         if (IS_ERR(ps->s_txreq))
 274                 goto bail_no_tx;
 275
 276         ohdr = &ps->s_txreq->phdr.hdr.u.oth;
 277         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 278                 ohdr = &ps->s_txreq->phdr.hdr.u.l.oth;
 279
 280         /* Sending responses has higher priority over sending requests. */
 281         if ((qp->s_flags & RVT_S_RESP_PENDING) &&
 282             make_rc_ack(dev, qp, ohdr, ps))
 283                 return 1;
 284
 285         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
 286                 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
 287                         goto bail;
 288                 /* We are in the error state, flush the work request. */
 289                 smp_read_barrier_depends(); /* see post_one_send() */
 290                 if (qp->s_last == READ_ONCE(qp->s_head))
 291                         goto bail;
 292                 /* If DMAs are in progress, we can't flush immediately. */
 293                 if (iowait_sdma_pending(&priv->s_iowait)) {
 294                         qp->s_flags |= RVT_S_WAIT_DMA;
 295                         goto bail;
 296                 }
 297                 clear_ahg(qp);
 298                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 299                 hfi1_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
 300                         IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
 301                 /* will get called again */
 302                 goto done_free_tx;
 303         }
 304
 305         if (qp->s_flags & (RVT_S_WAIT_RNR | RVT_S_WAIT_ACK))
 306                 goto bail;
 307
 308         if (cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) {
 309                 if (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0) {
 310                         qp->s_flags |= RVT_S_WAIT_PSN;
 311                         goto bail;
 312                 }
 313                 qp->s_sending_psn = qp->s_psn;
 314                 qp->s_sending_hpsn = qp->s_psn - 1;
 315         }
 316
 317         /* Send a request. */
 318         wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
 319         switch (qp->s_state) {
 320         default:
 321                 if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
 322                         goto bail;
 323                 /*
 324                  * Resend an old request or start a new one.
 325                  *
 326                  * We keep track of the current SWQE so that
 327                  * we don't reset the "furthest progress" state
 328                  * if we need to back up.
 329                  */
 330                 newreq = 0;
 331                 if (qp->s_cur == qp->s_tail) {
 332                         /* Check if send work queue is empty. */
 333                         smp_read_barrier_depends(); /* see post_one_send() */
 334                         if (qp->s_tail == READ_ONCE(qp->s_head)) {
 335                                 clear_ahg(qp);
 336                                 goto bail;
 337                         }
 338                         /*
 339                          * If a fence is requested, wait for previous
 340                          * RDMA read and atomic operations to finish.
 341                          */
 342                         if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
 343                             qp->s_num_rd_atomic) {
 344                                 qp->s_flags |= RVT_S_WAIT_FENCE;
 345                                 goto bail;
 346                         }
 347                         /*
 348                          * Local operations are processed immediately
 349                          * after all prior requests have completed
 350                          */
 351                         if (wqe->wr.opcode == IB_WR_REG_MR ||
 352                             wqe->wr.opcode == IB_WR_LOCAL_INV) {
 353                                 int local_ops = 0;
 354                                 int err = 0;
 355
 356                                 if (qp->s_last != qp->s_cur)
 357                                         goto bail;
 358                                 if (++qp->s_cur == qp->s_size)
 359                                         qp->s_cur = 0;
 360                                 if (++qp->s_tail == qp->s_size)
 361                                         qp->s_tail = 0;
 362                                 if (!(wqe->wr.send_flags &
 363                                       RVT_SEND_COMPLETION_ONLY)) {
 364                                         err = rvt_invalidate_rkey(
 365                                                 qp,
 366                                                 wqe->wr.ex.invalidate_rkey);
 367                                         local_ops = 1;
 368                                 }
 369                                 hfi1_send_complete(qp, wqe,
 370                                                    err ? IB_WC_LOC_PROT_ERR
 371                                                        : IB_WC_SUCCESS);
 372                                 if (local_ops)
 373                                         atomic_dec(&qp->local_ops_pending);
 374                                 qp->s_hdrwords = 0;
 375                                 goto done_free_tx;
 376                         }
 377
 378                         newreq = 1;
 379                         qp->s_psn = wqe->psn;
 380                 }
 381                 /*
 382                  * Note that we have to be careful not to modify the
 383                  * original work request since we may need to resend
 384                  * it.
 385                  */
 386                 len = wqe->length;
 387                 ss = &qp->s_sge;
 388                 bth2 = mask_psn(qp->s_psn);
 389                 switch (wqe->wr.opcode) {
 390                 case IB_WR_SEND:
 391                 case IB_WR_SEND_WITH_IMM:
 392                 case IB_WR_SEND_WITH_INV:
 393                         /* If no credit, return. */
 394                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 395                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 396                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 397                                 goto bail;
 398                         }
 399                         if (len > pmtu) {
 400                                 qp->s_state = OP(SEND_FIRST);
 401                                 len = pmtu;
 402                                 break;
 403                         }
 404                         if (wqe->wr.opcode == IB_WR_SEND) {
 405                                 qp->s_state = OP(SEND_ONLY);
 406                         } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 407                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 408                                 /* Immediate data comes after the BTH */
 409                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 410                                 hwords += 1;
 411                         } else {
 412                                 qp->s_state = OP(SEND_ONLY_WITH_INVALIDATE);
 413                                 /* Invalidate rkey comes after the BTH */
 414                                 ohdr->u.ieth = cpu_to_be32(
 415                                                 wqe->wr.ex.invalidate_rkey);
 416                                 hwords += 1;
 417                         }
 418                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 419                                 bth0 |= IB_BTH_SOLICITED;
 420                         bth2 |= IB_BTH_REQ_ACK;
 421                         if (++qp->s_cur == qp->s_size)
 422                                 qp->s_cur = 0;
 423                         break;
 424
 425                 case IB_WR_RDMA_WRITE:
 426                         if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 427                                 qp->s_lsn++;
 428                         /* FALLTHROUGH */
 429                 case IB_WR_RDMA_WRITE_WITH_IMM:
 430                         /* If no credit, return. */
 431                         if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
 432                             rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
 433                                 qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
 434                                 goto bail;
 435                         }
 436                         put_ib_reth_vaddr(
 437                                 wqe->rdma_wr.remote_addr,
 438                                 &ohdr->u.rc.reth);
 439                         ohdr->u.rc.reth.rkey =
 440                                 cpu_to_be32(wqe->rdma_wr.rkey);
 441                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 442                         hwords += sizeof(struct ib_reth) / sizeof(u32);
 443                         if (len > pmtu) {
 444                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 445                                 len = pmtu;
 446                                 break;
 447                         }
 448                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 449                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 450                         } else {
 451                                 qp->s_state =
 452                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 453                                 /* Immediate data comes after RETH */
 454                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 455                                 hwords += 1;
 456                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 457                                         bth0 |= IB_BTH_SOLICITED;
 458                         }
 459                         bth2 |= IB_BTH_REQ_ACK;
 460                         if (++qp->s_cur == qp->s_size)
 461                                 qp->s_cur = 0;
 462                         break;
 463
 464                 case IB_WR_RDMA_READ:
 465                         /*
 466                          * Don't allow more operations to be started
 467                          * than the QP limits allow.
 468                          */
 469                         if (newreq) {
 470                                 if (qp->s_num_rd_atomic >=
 471                                     qp->s_max_rd_atomic) {
 472                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 473                                         goto bail;
 474                                 }
 475                                 qp->s_num_rd_atomic++;
 476                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 477                                         qp->s_lsn++;
 478                         }
 479                         put_ib_reth_vaddr(
 480                                 wqe->rdma_wr.remote_addr,
 481                                 &ohdr->u.rc.reth);
 482                         ohdr->u.rc.reth.rkey =
 483                                 cpu_to_be32(wqe->rdma_wr.rkey);
 484                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 485                         qp->s_state = OP(RDMA_READ_REQUEST);
 486                         hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 487                         ss = NULL;
 488                         len = 0;
 489                         bth2 |= IB_BTH_REQ_ACK;
 490                         if (++qp->s_cur == qp->s_size)
 491                                 qp->s_cur = 0;
 492                         break;
 493
 494                 case IB_WR_ATOMIC_CMP_AND_SWP:
 495                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 496                         /*
 497                          * Don't allow more operations to be started
 498                          * than the QP limits allow.
 499                          */
 500                         if (newreq) {
 501                                 if (qp->s_num_rd_atomic >=
 502                                     qp->s_max_rd_atomic) {
 503                                         qp->s_flags |= RVT_S_WAIT_RDMAR;
 504                                         goto bail;
 505                                 }
 506                                 qp->s_num_rd_atomic++;
 507                                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 508                                         qp->s_lsn++;
 509                         }
 510                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP) {
 511                                 qp->s_state = OP(COMPARE_SWAP);
 512                                 put_ib_ateth_swap(wqe->atomic_wr.swap,
 513                                                   &ohdr->u.atomic_eth);
 514                                 put_ib_ateth_compare(wqe->atomic_wr.compare_add,
 515                                                      &ohdr->u.atomic_eth);
 516                         } else {
 517                                 qp->s_state = OP(FETCH_ADD);
 518                                 put_ib_ateth_swap(wqe->atomic_wr.compare_add,
 519                                                   &ohdr->u.atomic_eth);
 520                                 put_ib_ateth_compare(0, &ohdr->u.atomic_eth);
 521                         }
 522                         put_ib_ateth_vaddr(wqe->atomic_wr.remote_addr,
 523                                            &ohdr->u.atomic_eth);
 524                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 525                                 wqe->atomic_wr.rkey);
 526                         hwords += sizeof(struct ib_atomic_eth) / sizeof(u32);
 527                         ss = NULL;
 528                         len = 0;
 529                         bth2 |= IB_BTH_REQ_ACK;
 530                         if (++qp->s_cur == qp->s_size)
 531                                 qp->s_cur = 0;
 532                         break;
 533
 534                 default:
 535                         goto bail;
 536                 }
 537                 qp->s_sge.sge = wqe->sg_list[0];
 538                 qp->s_sge.sg_list = wqe->sg_list + 1;
 539                 qp->s_sge.num_sge = wqe->wr.num_sge;
 540                 qp->s_sge.total_len = wqe->length;
 541                 qp->s_len = wqe->length;
 542                 if (newreq) {
 543                         qp->s_tail++;
 544                         if (qp->s_tail >= qp->s_size)
 545                                 qp->s_tail = 0;
 546                 }
 547                 if (wqe->wr.opcode == IB_WR_RDMA_READ)
 548                         qp->s_psn = wqe->lpsn + 1;
 549                 else
 550                         qp->s_psn++;
 551                 break;
 552
 553         case OP(RDMA_READ_RESPONSE_FIRST):
 554                 /*
 555                  * qp->s_state is normally set to the opcode of the
 556                  * last packet constructed for new requests and therefore
 557                  * is never set to RDMA read response.
 558                  * RDMA_READ_RESPONSE_FIRST is used by the ACK processing
 559                  * thread to indicate a SEND needs to be restarted from an
 560                  * earlier PSN without interfering with the sending thread.
 561                  * See restart_rc().
 562                  */
 563                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 564                 /* FALLTHROUGH */
 565         case OP(SEND_FIRST):
 566                 qp->s_state = OP(SEND_MIDDLE);
 567                 /* FALLTHROUGH */
 568         case OP(SEND_MIDDLE):
 569                 bth2 = mask_psn(qp->s_psn++);
 570                 ss = &qp->s_sge;
 571                 len = qp->s_len;
 572                 if (len > pmtu) {
 573                         len = pmtu;
 574                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 575                         break;
 576                 }
 577                 if (wqe->wr.opcode == IB_WR_SEND) {
 578                         qp->s_state = OP(SEND_LAST);
 579                 } else if (wqe->wr.opcode == IB_WR_SEND_WITH_IMM) {
 580                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 581                         /* Immediate data comes after the BTH */
 582                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 583                         hwords += 1;
 584                 } else {
 585                         qp->s_state = OP(SEND_LAST_WITH_INVALIDATE);
 586                         /* invalidate data comes after the BTH */
 587                         ohdr->u.ieth = cpu_to_be32(wqe->wr.ex.invalidate_rkey);
 588                         hwords += 1;
 589                 }
 590                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 591                         bth0 |= IB_BTH_SOLICITED;
 592                 bth2 |= IB_BTH_REQ_ACK;
 593                 qp->s_cur++;
 594                 if (qp->s_cur >= qp->s_size)
 595                         qp->s_cur = 0;
 596                 break;
 597
 598         case OP(RDMA_READ_RESPONSE_LAST):
 599                 /*
 600                  * qp->s_state is normally set to the opcode of the
 601                  * last packet constructed for new requests and therefore
 602                  * is never set to RDMA read response.
 603                  * RDMA_READ_RESPONSE_LAST is used by the ACK processing
 604                  * thread to indicate a RDMA write needs to be restarted from
 605                  * an earlier PSN without interfering with the sending thread.
 606                  * See restart_rc().
 607                  */
 608                 qp->s_len = restart_sge(&qp->s_sge, wqe, qp->s_psn, pmtu);
 609                 /* FALLTHROUGH */
 610         case OP(RDMA_WRITE_FIRST):
 611                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 612                 /* FALLTHROUGH */
 613         case OP(RDMA_WRITE_MIDDLE):
 614                 bth2 = mask_psn(qp->s_psn++);
 615                 ss = &qp->s_sge;
 616                 len = qp->s_len;
 617                 if (len > pmtu) {
 618                         len = pmtu;
 619                         middle = HFI1_CAP_IS_KSET(SDMA_AHG);
 620                         break;
 621                 }
 622                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE) {
 623                         qp->s_state = OP(RDMA_WRITE_LAST);
 624                 } else {
 625                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 626                         /* Immediate data comes after the BTH */
 627                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 628                         hwords += 1;
 629                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 630                                 bth0 |= IB_BTH_SOLICITED;
 631                 }
 632                 bth2 |= IB_BTH_REQ_ACK;
 633                 qp->s_cur++;
 634                 if (qp->s_cur >= qp->s_size)
 635                         qp->s_cur = 0;
 636                 break;
 637
 638         case OP(RDMA_READ_RESPONSE_MIDDLE):
 639                 /*
 640                  * qp->s_state is normally set to the opcode of the
 641                  * last packet constructed for new requests and therefore
 642                  * is never set to RDMA read response.
 643                  * RDMA_READ_RESPONSE_MIDDLE is used by the ACK processing
 644                  * thread to indicate a RDMA read needs to be restarted from
 645                  * an earlier PSN without interfering with the sending thread.
 646                  * See restart_rc().
 647                  */
 648                 len = (delta_psn(qp->s_psn, wqe->psn)) * pmtu;
 649                 put_ib_reth_vaddr(
 650                         wqe->rdma_wr.remote_addr + len,
 651                         &ohdr->u.rc.reth);
 652                 ohdr->u.rc.reth.rkey =
 653                         cpu_to_be32(wqe->rdma_wr.rkey);
 654                 ohdr->u.rc.reth.length = cpu_to_be32(wqe->length - len);
 655                 qp->s_state = OP(RDMA_READ_REQUEST);
 656                 hwords += sizeof(ohdr->u.rc.reth) / sizeof(u32);
 657                 bth2 = mask_psn(qp->s_psn) | IB_BTH_REQ_ACK;
 658                 qp->s_psn = wqe->lpsn + 1;
 659                 ss = NULL;
 660                 len = 0;
 661                 qp->s_cur++;
 662                 if (qp->s_cur == qp->s_size)
 663                         qp->s_cur = 0;
 664                 break;
 665         }
 666         qp->s_sending_hpsn = bth2;
 667         delta = delta_psn(bth2, wqe->psn);
 668         if (delta && delta % HFI1_PSN_CREDIT == 0)
 669                 bth2 |= IB_BTH_REQ_ACK;
 670         if (qp->s_flags & RVT_S_SEND_ONE) {
 671                 qp->s_flags &= ~RVT_S_SEND_ONE;
 672                 qp->s_flags |= RVT_S_WAIT_ACK;
 673                 bth2 |= IB_BTH_REQ_ACK;
 674         }
 675         qp->s_len -= len;
 676         qp->s_hdrwords = hwords;
 677         ps->s_txreq->sde = priv->s_sde;
 678         ps->s_txreq->ss = ss;
 679         ps->s_txreq->s_cur_size = len;
 680         hfi1_make_ruc_header(
 681                 qp,
 682                 ohdr,
 683                 bth0 | (qp->s_state << 24),
 684                 bth2,
 685                 middle,
 686                 ps);
 687         /* pbc */
 688         ps->s_txreq->hdr_dwords = qp->s_hdrwords + 2;
 689         return 1;
 690
 691 done_free_tx:
 692         hfi1_put_txreq(ps->s_txreq);
 693         ps->s_txreq = NULL;
 694         return 1;
 695
 696 bail:
 697         hfi1_put_txreq(ps->s_txreq);
 698
 699 bail_no_tx:
 700         ps->s_txreq = NULL;
 701         qp->s_flags &= ~RVT_S_BUSY;
 702         qp->s_hdrwords = 0;
 703         return 0;
 704 }
 705
 706 /**
 707  * hfi1_send_rc_ack - Construct an ACK packet and send it
 708  * @qp: a pointer to the QP
 709  *
 710  * This is called from hfi1_rc_rcv() and handle_receive_interrupt().
 711  * Note that RDMA reads and atomics are handled in the
 712  * send side QP state and send engine.
 713  */
 714 void hfi1_send_rc_ack(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp,
 715                       int is_fecn)
 716 {
 717         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
 718         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 719         u64 pbc, pbc_flags = 0;
 720         u16 lrh0;
 721         u16 sc5;
 722         u32 bth0;
 723         u32 hwords;
 724         u32 vl, plen;
 725         struct send_context *sc;
 726         struct pio_buf *pbuf;
 727         struct ib_header hdr;
 728         struct ib_other_headers *ohdr;
 729         unsigned long flags;
 730         struct hfi1_qp_priv *priv = qp->priv;
 731
 732         /* clear the defer count */
 733         priv->r_adefered = 0;
 734
 735         /* Don't send ACK or NAK if a RDMA read or atomic is pending. */
 736         if (qp->s_flags & RVT_S_RESP_PENDING)
 737                 goto queue_ack;
 738
 739         /* Ensure s_rdma_ack_cnt changes are committed */
 740         smp_read_barrier_depends();
 741         if (qp->s_rdma_ack_cnt)
 742                 goto queue_ack;
 743
 744         /* Construct the header */
 745         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4 */
 746         hwords = 6;
 747         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 748                 hwords += hfi1_make_grh(ibp, &hdr.u.l.grh,
 749                                        &qp->remote_ah_attr.grh, hwords, 0);
 750                 ohdr = &hdr.u.l.oth;
 751                 lrh0 = HFI1_LRH_GRH;
 752         } else {
 753                 ohdr = &hdr.u.oth;
 754                 lrh0 = HFI1_LRH_BTH;
 755         }
 756         /* read pkey_index w/o lock (its atomic) */
 757         bth0 = hfi1_get_pkey(ibp, qp->s_pkey_index) | (OP(ACKNOWLEDGE) << 24);
 758         if (qp->s_mig_state == IB_MIG_MIGRATED)
 759                 bth0 |= IB_BTH_MIG_REQ;
 760         if (qp->r_nak_state)
 761                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IB_MSN_MASK) |
 762                                             (qp->r_nak_state <<
 763                                              IB_AETH_CREDIT_SHIFT));
 764         else
 765                 ohdr->u.aeth = rvt_compute_aeth(qp);
 766         sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
 767         /* set PBC_DC_INFO bit (aka SC[4]) in pbc_flags */
 768         pbc_flags |= ((!!(sc5 & 0x10)) << PBC_DC_INFO_SHIFT);
 769         lrh0 |= (sc5 & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
 770         hdr.lrh[0] = cpu_to_be16(lrh0);
 771         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 772         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 773         hdr.lrh[3] = cpu_to_be16(ppd->lid | qp->remote_ah_attr.src_path_bits);
 774         ohdr->bth[0] = cpu_to_be32(bth0);
 775         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 776         ohdr->bth[1] |= cpu_to_be32((!!is_fecn) << HFI1_BECN_SHIFT);
 777         ohdr->bth[2] = cpu_to_be32(mask_psn(qp->r_ack_psn));
 778
 779         /* Don't try to send ACKs if the link isn't ACTIVE */
 780         if (driver_lstate(ppd) != IB_PORT_ACTIVE)
 781                 return;
 782
 783         sc = rcd->sc;
 784         plen = 2 /* PBC */ + hwords;
 785         vl = sc_to_vlt(ppd->dd, sc5);
 786         pbc = create_pbc(ppd, pbc_flags, qp->srate_mbps, vl, plen);
 787
 788         pbuf = sc_buffer_alloc(sc, plen, NULL, NULL);
 789         if (!pbuf) {
 790                 /*
 791                  * We have no room to send at the moment.  Pass
 792                  * responsibility for sending the ACK to the send engine
 793                  * so that when enough buffer space becomes available,
 794                  * the ACK is sent ahead of other outgoing packets.
 795                  */
 796                 goto queue_ack;
 797         }
 798
 799         trace_ack_output_ibhdr(dd_from_ibdev(qp->ibqp.device), &hdr);
 800
 801         /* write the pbc and data */
 802         ppd->dd->pio_inline_send(ppd->dd, pbuf, pbc, &hdr, hwords);
 803
 804         return;
 805
 806 queue_ack:
 807         spin_lock_irqsave(&qp->s_lock, flags);
 808         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
 809                 goto unlock;
 810         this_cpu_inc(*ibp->rvp.rc_qacks);
 811         qp->s_flags |= RVT_S_ACK_PENDING | RVT_S_RESP_PENDING;
 812         qp->s_nak_state = qp->r_nak_state;
 813         qp->s_ack_psn = qp->r_ack_psn;
 814         if (is_fecn)
 815                 qp->s_flags |= RVT_S_ECN;
 816
 817         /* Schedule the send engine. */
 818         hfi1_schedule_send(qp);
 819 unlock:
 820         spin_unlock_irqrestore(&qp->s_lock, flags);
 821 }
 822
 823 /**
 824  * reset_psn - reset the QP state to send starting from PSN
 825  * @qp: the QP
 826  * @psn: the packet sequence number to restart at
 827  *
 828  * This is called from hfi1_rc_rcv() to process an incoming RC ACK
 829  * for the given QP.
 830  * Called at interrupt level with the QP s_lock held.
 831  */
 832 static void reset_psn(struct rvt_qp *qp, u32 psn)
 833 {
 834         u32 n = qp->s_acked;
 835         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
 836         u32 opcode;
 837
 838         lockdep_assert_held(&qp->s_lock);
 839         qp->s_cur = n;
 840
 841         /*
 842          * If we are starting the request from the beginning,
 843          * let the normal send code handle initialization.
 844          */
 845         if (cmp_psn(psn, wqe->psn) <= 0) {
 846                 qp->s_state = OP(SEND_LAST);
 847                 goto done;
 848         }
 849
 850         /* Find the work request opcode corresponding to the given PSN. */
 851         opcode = wqe->wr.opcode;
 852         for (;;) {
 853                 int diff;
 854
 855                 if (++n == qp->s_size)
 856                         n = 0;
 857                 if (n == qp->s_tail)
 858                         break;
 859                 wqe = rvt_get_swqe_ptr(qp, n);
 860                 diff = cmp_psn(psn, wqe->psn);
 861                 if (diff < 0)
 862                         break;
 863                 qp->s_cur = n;
 864                 /*
 865                  * If we are starting the request from the beginning,
 866                  * let the normal send code handle initialization.
 867                  */
 868                 if (diff == 0) {
 869                         qp->s_state = OP(SEND_LAST);
 870                         goto done;
 871                 }
 872                 opcode = wqe->wr.opcode;
 873         }
 874
 875         /*
 876          * Set the state to restart in the middle of a request.
 877          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 878          * See hfi1_make_rc_req().
 879          */
 880         switch (opcode) {
 881         case IB_WR_SEND:
 882         case IB_WR_SEND_WITH_IMM:
 883                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 884                 break;
 885
 886         case IB_WR_RDMA_WRITE:
 887         case IB_WR_RDMA_WRITE_WITH_IMM:
 888                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 889                 break;
 890
 891         case IB_WR_RDMA_READ:
 892                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 893                 break;
 894
 895         default:
 896                 /*
 897                  * This case shouldn't happen since its only
 898                  * one PSN per req.
 899                  */
 900                 qp->s_state = OP(SEND_LAST);
 901         }
 902 done:
 903         qp->s_psn = psn;
 904         /*
 905          * Set RVT_S_WAIT_PSN as rc_complete() may start the timer
 906          * asynchronously before the send engine can get scheduled.
 907          * Doing it in hfi1_make_rc_req() is too late.
 908          */
 909         if ((cmp_psn(qp->s_psn, qp->s_sending_hpsn) <= 0) &&
 910             (cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0))
 911                 qp->s_flags |= RVT_S_WAIT_PSN;
 912         qp->s_flags &= ~RVT_S_AHG_VALID;
 913 }
 914
 915 /*
 916  * Back up requester to resend the last un-ACKed request.
 917  * The QP r_lock and s_lock should be held and interrupts disabled.
 918  */
 919 void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
 920 {
 921         struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
 922         struct hfi1_ibport *ibp;
 923
 924         lockdep_assert_held(&qp->r_lock);
 925         lockdep_assert_held(&qp->s_lock);
 926         if (qp->s_retry == 0) {
 927                 if (qp->s_mig_state == IB_MIG_ARMED) {
 928                         hfi1_migrate_qp(qp);
 929                         qp->s_retry = qp->s_retry_cnt;
 930                 } else if (qp->s_last == qp->s_acked) {
 931                         hfi1_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR);
 932                         rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 933                         return;
 934                 } else { /* need to handle delayed completion */
 935                         return;
 936                 }
 937         } else {
 938                 qp->s_retry--;
 939         }
 940
 941         ibp = to_iport(qp->ibqp.device, qp->port_num);
 942         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 943                 ibp->rvp.n_rc_resends++;
 944         else
 945                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
 946
 947         qp->s_flags &= ~(RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR |
 948                          RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_PSN |
 949                          RVT_S_WAIT_ACK);
 950         if (wait)
 951                 qp->s_flags |= RVT_S_SEND_ONE;
 952         reset_psn(qp, psn);
 953 }
 954
 955 /*
 956  * Set qp->s_sending_psn to the next PSN after the given one.
 957  * This would be psn+1 except when RDMA reads are present.
 958  */
 959 static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
 960 {
 961         struct rvt_swqe *wqe;
 962         u32 n = qp->s_last;
 963
 964         lockdep_assert_held(&qp->s_lock);
 965         /* Find the work request corresponding to the given PSN. */
 966         for (;;) {
 967                 wqe = rvt_get_swqe_ptr(qp, n);
 968                 if (cmp_psn(psn, wqe->lpsn) <= 0) {
 969                         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 970                                 qp->s_sending_psn = wqe->lpsn + 1;
 971                         else
 972                                 qp->s_sending_psn = psn + 1;
 973                         break;
 974                 }
 975                 if (++n == qp->s_size)
 976                         n = 0;
 977                 if (n == qp->s_tail)
 978                         break;
 979         }
 980 }
 981
 982 /*
 983  * This should be called with the QP s_lock held and interrupts disabled.
 984  */
 985 void hfi1_rc_send_complete(struct rvt_qp *qp, struct ib_header *hdr)
 986 {
 987         struct ib_other_headers *ohdr;
 988         struct rvt_swqe *wqe;
 989         u32 opcode;
 990         u32 psn;
 991
 992         lockdep_assert_held(&qp->s_lock);
 993         if (!(ib_rvt_state_ops[qp->state] & RVT_SEND_OR_FLUSH_OR_RECV_OK))
 994                 return;
 995
 996         /* Find out where the BTH is */
 997         if ((be16_to_cpu(hdr->lrh[0]) & 3) == HFI1_LRH_BTH)
 998                 ohdr = &hdr->u.oth;
 999         else
1000                 ohdr = &hdr->u.l.oth;
1001
1002         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1003         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1004             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1005                 WARN_ON(!qp->s_rdma_ack_cnt);
1006                 qp->s_rdma_ack_cnt--;
1007                 return;
1008         }
1009
1010         psn = be32_to_cpu(ohdr->bth[2]);
1011         reset_sending_psn(qp, psn);
1012
1013         /*
1014          * Start timer after a packet requesting an ACK has been sent and
1015          * there are still requests that haven't been acked.
1016          */
1017         if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
1018             !(qp->s_flags &
1019                 (RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
1020                 (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
1021                 rvt_add_retry_timer(qp);
1022
1023         while (qp->s_last != qp->s_acked) {
1024                 u32 s_last;
1025
1026                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
1027                 if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
1028                     cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
1029                         break;
1030                 s_last = qp->s_last;
1031                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1032                 if (++s_last >= qp->s_size)
1033                         s_last = 0;
1034                 qp->s_last = s_last;
1035                 /* see post_send() */
1036                 barrier();
1037                 rvt_put_swqe(wqe);
1038                 rvt_qp_swqe_complete(qp,
1039                                      wqe,
1040                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1041                                      IB_WC_SUCCESS);
1042         }
1043         /*
1044          * If we were waiting for sends to complete before re-sending,
1045          * and they are now complete, restart sending.
1046          */
1047         trace_hfi1_sendcomplete(qp, psn);
1048         if (qp->s_flags & RVT_S_WAIT_PSN &&
1049             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1050                 qp->s_flags &= ~RVT_S_WAIT_PSN;
1051                 qp->s_sending_psn = qp->s_psn;
1052                 qp->s_sending_hpsn = qp->s_psn - 1;
1053                 hfi1_schedule_send(qp);
1054         }
1055 }
1056
1057 static inline void update_last_psn(struct rvt_qp *qp, u32 psn)
1058 {
1059         qp->s_last_psn = psn;
1060 }
1061
1062 /*
1063  * Generate a SWQE completion.
1064  * This is similar to hfi1_send_complete but has to check to be sure
1065  * that the SGEs are not being referenced if the SWQE is being resent.
1066  */
1067 static struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
1068                                          struct rvt_swqe *wqe,
1069                                          struct hfi1_ibport *ibp)
1070 {
1071         lockdep_assert_held(&qp->s_lock);
1072         /*
1073          * Don't decrement refcount and don't generate a
1074          * completion if the SWQE is being resent until the send
1075          * is finished.
1076          */
1077         if (cmp_psn(wqe->lpsn, qp->s_sending_psn) < 0 ||
1078             cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
1079                 u32 s_last;
1080
1081                 rvt_put_swqe(wqe);
1082                 s_last = qp->s_last;
1083                 trace_hfi1_qp_send_completion(qp, wqe, s_last);
1084                 if (++s_last >= qp->s_size)
1085                         s_last = 0;
1086                 qp->s_last = s_last;
1087                 /* see post_send() */
1088                 barrier();
1089                 rvt_qp_swqe_complete(qp,
1090                                      wqe,
1091                                      ib_hfi1_wc_opcode[wqe->wr.opcode],
1092                                      IB_WC_SUCCESS);
1093         } else {
1094                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
1095
1096                 this_cpu_inc(*ibp->rvp.rc_delayed_comp);
1097                 /*
1098                  * If send progress not running attempt to progress
1099                  * SDMA queue.
1100                  */
1101                 if (ppd->dd->flags & HFI1_HAS_SEND_DMA) {
1102                         struct sdma_engine *engine;
1103                         u8 sc5;
1104
1105                         /* For now use sc to find engine */
1106                         sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
1107                         engine = qp_to_sdma_engine(qp, sc5);
1108                         sdma_engine_progress_schedule(engine);
1109                 }
1110         }
1111
1112         qp->s_retry = qp->s_retry_cnt;
1113         update_last_psn(qp, wqe->lpsn);
1114
1115         /*
1116          * If we are completing a request which is in the process of
1117          * being resent, we can stop re-sending it since we know the
1118          * responder has already seen it.
1119          */
1120         if (qp->s_acked == qp->s_cur) {
1121                 if (++qp->s_cur >= qp->s_size)
1122                         qp->s_cur = 0;
1123                 qp->s_acked = qp->s_cur;
1124                 wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
1125                 if (qp->s_acked != qp->s_tail) {
1126                         qp->s_state = OP(SEND_LAST);
1127                         qp->s_psn = wqe->psn;
1128                 }
1129         } else {
1130                 if (++qp->s_acked >= qp->s_size)
1131                         qp->s_acked = 0;
1132                 if (qp->state == IB_QPS_SQD && qp->s_acked == qp->s_cur)
1133                         qp->s_draining = 0;
1134                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1135         }
1136         return wqe;
1137 }
1138
1139 /**
1140  * do_rc_ack - process an incoming RC ACK
1141  * @qp: the QP the ACK came in on
1142  * @psn: the packet sequence number of the ACK
1143  * @opcode: the opcode of the request that resulted in the ACK
1144  *
1145  * This is called from rc_rcv_resp() to process an incoming RC ACK
1146  * for the given QP.
1147  * May be called at interrupt level, with the QP s_lock held.
1148  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
1149  */
1150 static int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
1151                      u64 val, struct hfi1_ctxtdata *rcd)
1152 {
1153         struct hfi1_ibport *ibp;
1154         enum ib_wc_status status;
1155         struct rvt_swqe *wqe;
1156         int ret = 0;
1157         u32 ack_psn;
1158         int diff;
1159
1160         lockdep_assert_held(&qp->s_lock);
1161         /*
1162          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
1163          * requests and implicitly NAK RDMA read and atomic requests issued
1164          * before the NAK'ed request.  The MSN won't include the NAK'ed
1165          * request but will include an ACK'ed request(s).
1166          */
1167         ack_psn = psn;
1168         if (aeth >> IB_AETH_NAK_SHIFT)
1169                 ack_psn--;
1170         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1171         ibp = rcd_to_iport(rcd);
1172
1173         /*
1174          * The MSN might be for a later WQE than the PSN indicates so
1175          * only complete WQEs that the PSN finishes.
1176          */
1177         while ((diff = delta_psn(ack_psn, wqe->lpsn)) >= 0) {
1178                 /*
1179                  * RDMA_READ_RESPONSE_ONLY is a special case since
1180                  * we want to generate completion events for everything
1181                  * before the RDMA read, copy the data, then generate
1182                  * the completion for the read.
1183                  */
1184                 if (wqe->wr.opcode == IB_WR_RDMA_READ &&
1185                     opcode == OP(RDMA_READ_RESPONSE_ONLY) &&
1186                     diff == 0) {
1187                         ret = 1;
1188                         goto bail_stop;
1189                 }
1190                 /*
1191                  * If this request is a RDMA read or atomic, and the ACK is
1192                  * for a later operation, this ACK NAKs the RDMA read or
1193                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
1194                  * can ACK a RDMA read and likewise for atomic ops.  Note
1195                  * that the NAK case can only happen if relaxed ordering is
1196                  * used and requests are sent after an RDMA read or atomic
1197                  * is sent but before the response is received.
1198                  */
1199                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
1200                      (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
1201                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1202                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
1203                      (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
1204                         /* Retry this request. */
1205                         if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) {
1206                                 qp->r_flags |= RVT_R_RDMAR_SEQ;
1207                                 hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1208                                 if (list_empty(&qp->rspwait)) {
1209                                         qp->r_flags |= RVT_R_RSP_SEND;
1210                                         rvt_get_qp(qp);
1211                                         list_add_tail(&qp->rspwait,
1212                                                       &rcd->qp_wait_list);
1213                                 }
1214                         }
1215                         /*
1216                          * No need to process the ACK/NAK since we are
1217                          * restarting an earlier request.
1218                          */
1219                         goto bail_stop;
1220                 }
1221                 if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1222                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
1223                         u64 *vaddr = wqe->sg_list[0].vaddr;
1224                         *vaddr = val;
1225                 }
1226                 if (qp->s_num_rd_atomic &&
1227                     (wqe->wr.opcode == IB_WR_RDMA_READ ||
1228                      wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1229                      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)) {
1230                         qp->s_num_rd_atomic--;
1231                         /* Restart sending task if fence is complete */
1232                         if ((qp->s_flags & RVT_S_WAIT_FENCE) &&
1233                             !qp->s_num_rd_atomic) {
1234                                 qp->s_flags &= ~(RVT_S_WAIT_FENCE |
1235                                                  RVT_S_WAIT_ACK);
1236                                 hfi1_schedule_send(qp);
1237                         } else if (qp->s_flags & RVT_S_WAIT_RDMAR) {
1238                                 qp->s_flags &= ~(RVT_S_WAIT_RDMAR |
1239                                                  RVT_S_WAIT_ACK);
1240                                 hfi1_schedule_send(qp);
1241                         }
1242                 }
1243                 wqe = do_rc_completion(qp, wqe, ibp);
1244                 if (qp->s_acked == qp->s_tail)
1245                         break;
1246         }
1247
1248         switch (aeth >> IB_AETH_NAK_SHIFT) {
1249         case 0:         /* ACK */
1250                 this_cpu_inc(*ibp->rvp.rc_acks);
1251                 if (qp->s_acked != qp->s_tail) {
1252                         /*
1253                          * We are expecting more ACKs so
1254                          * mod the retry timer.
1255                          */
1256                         rvt_mod_retry_timer(qp);
1257                         /*
1258                          * We can stop re-sending the earlier packets and
1259                          * continue with the next packet the receiver wants.
1260                          */
1261                         if (cmp_psn(qp->s_psn, psn) <= 0)
1262                                 reset_psn(qp, psn + 1);
1263                 } else {
1264                         /* No more acks - kill all timers */
1265                         rvt_stop_rc_timers(qp);
1266                         if (cmp_psn(qp->s_psn, psn) <= 0) {
1267                                 qp->s_state = OP(SEND_LAST);
1268                                 qp->s_psn = psn + 1;
1269                         }
1270                 }
1271                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1272                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1273                         hfi1_schedule_send(qp);
1274                 }
1275                 rvt_get_credit(qp, aeth);
1276                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1277                 qp->s_retry = qp->s_retry_cnt;
1278                 update_last_psn(qp, psn);
1279                 return 1;
1280
1281         case 1:         /* RNR NAK */
1282                 ibp->rvp.n_rnr_naks++;
1283                 if (qp->s_acked == qp->s_tail)
1284                         goto bail_stop;
1285                 if (qp->s_flags & RVT_S_WAIT_RNR)
1286                         goto bail_stop;
1287                 if (qp->s_rnr_retry == 0) {
1288                         status = IB_WC_RNR_RETRY_EXC_ERR;
1289                         goto class_b;
1290                 }
1291                 if (qp->s_rnr_retry_cnt < 7)
1292                         qp->s_rnr_retry--;
1293
1294                 /* The last valid PSN is the previous PSN. */
1295                 update_last_psn(qp, psn - 1);
1296
1297                 ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);
1298
1299                 reset_psn(qp, psn);
1300
1301                 qp->s_flags &= ~(RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_ACK);
1302                 rvt_stop_rc_timers(qp);
1303                 rvt_add_rnr_timer(qp, aeth);
1304                 return 0;
1305
1306         case 3:         /* NAK */
1307                 if (qp->s_acked == qp->s_tail)
1308                         goto bail_stop;
1309                 /* The last valid PSN is the previous PSN. */
1310                 update_last_psn(qp, psn - 1);
1311                 switch ((aeth >> IB_AETH_CREDIT_SHIFT) &
1312                         IB_AETH_CREDIT_MASK) {
1313                 case 0: /* PSN sequence error */
1314                         ibp->rvp.n_seq_naks++;
1315                         /*
1316                          * Back up to the responder's expected PSN.
1317                          * Note that we might get a NAK in the middle of an
1318                          * RDMA READ response which terminates the RDMA
1319                          * READ.
1320                          */
1321                         hfi1_restart_rc(qp, psn, 0);
1322                         hfi1_schedule_send(qp);
1323                         break;
1324
1325                 case 1: /* Invalid Request */
1326                         status = IB_WC_REM_INV_REQ_ERR;
1327                         ibp->rvp.n_other_naks++;
1328                         goto class_b;
1329
1330                 case 2: /* Remote Access Error */
1331                         status = IB_WC_REM_ACCESS_ERR;
1332                         ibp->rvp.n_other_naks++;
1333                         goto class_b;
1334
1335                 case 3: /* Remote Operation Error */
1336                         status = IB_WC_REM_OP_ERR;
1337                         ibp->rvp.n_other_naks++;
1338 class_b:
1339                         if (qp->s_last == qp->s_acked) {
1340                                 hfi1_send_complete(qp, wqe, status);
1341                                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1342                         }
1343                         break;
1344
1345                 default:
1346                         /* Ignore other reserved NAK error codes */
1347                         goto reserved;
1348                 }
1349                 qp->s_retry = qp->s_retry_cnt;
1350                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
1351                 goto bail_stop;
1352
1353         default:                /* 2: reserved */
1354 reserved:
1355                 /* Ignore reserved NAK codes. */
1356                 goto bail_stop;
1357         }
1358         /* cannot be reached  */
1359 bail_stop:
1360         rvt_stop_rc_timers(qp);
1361         return ret;
1362 }
1363
1364 /*
1365  * We have seen an out of sequence RDMA read middle or last packet.
1366  * This ACKs SENDs and RDMA writes up to the first RDMA read or atomic SWQE.
1367  */
1368 static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
1369                          struct hfi1_ctxtdata *rcd)
1370 {
1371         struct rvt_swqe *wqe;
1372
1373         lockdep_assert_held(&qp->s_lock);
1374         /* Remove QP from retry timer */
1375         rvt_stop_rc_timers(qp);
1376
1377         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1378
1379         while (cmp_psn(psn, wqe->lpsn) > 0) {
1380                 if (wqe->wr.opcode == IB_WR_RDMA_READ ||
1381                     wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
1382                     wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
1383                         break;
1384                 wqe = do_rc_completion(qp, wqe, ibp);
1385         }
1386
1387         ibp->rvp.n_rdma_seq++;
1388         qp->r_flags |= RVT_R_RDMAR_SEQ;
1389         hfi1_restart_rc(qp, qp->s_last_psn + 1, 0);
1390         if (list_empty(&qp->rspwait)) {
1391                 qp->r_flags |= RVT_R_RSP_SEND;
1392                 rvt_get_qp(qp);
1393                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1394         }
1395 }
1396
1397 /**
1398  * rc_rcv_resp - process an incoming RC response packet
1399  * @ibp: the port this packet came in on
1400  * @ohdr: the other headers for this packet
1401  * @data: the packet data
1402  * @tlen: the packet length
1403  * @qp: the QP for this packet
1404  * @opcode: the opcode for this packet
1405  * @psn: the packet sequence number for this packet
1406  * @hdrsize: the header length
1407  * @pmtu: the path MTU
1408  *
1409  * This is called from hfi1_rc_rcv() to process an incoming RC response
1410  * packet for the given QP.
1411  * Called at interrupt level.
1412  */
1413 static void rc_rcv_resp(struct hfi1_ibport *ibp,
1414                         struct ib_other_headers *ohdr,
1415                         void *data, u32 tlen, struct rvt_qp *qp,
1416                         u32 opcode, u32 psn, u32 hdrsize, u32 pmtu,
1417                         struct hfi1_ctxtdata *rcd)
1418 {
1419         struct rvt_swqe *wqe;
1420         enum ib_wc_status status;
1421         unsigned long flags;
1422         int diff;
1423         u32 pad;
1424         u32 aeth;
1425         u64 val;
1426
1427         spin_lock_irqsave(&qp->s_lock, flags);
1428
1429         trace_hfi1_ack(qp, psn);
1430
1431         /* Ignore invalid responses. */
1432         smp_read_barrier_depends(); /* see post_one_send */
1433         if (cmp_psn(psn, READ_ONCE(qp->s_next_psn)) >= 0)
1434                 goto ack_done;
1435
1436         /* Ignore duplicate responses. */
1437         diff = cmp_psn(psn, qp->s_last_psn);
1438         if (unlikely(diff <= 0)) {
1439                 /* Update credits for "ghost" ACKs */
1440                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1441                         aeth = be32_to_cpu(ohdr->u.aeth);
1442                         if ((aeth >> IB_AETH_NAK_SHIFT) == 0)
1443                                 rvt_get_credit(qp, aeth);
1444                 }
1445                 goto ack_done;
1446         }
1447
1448         /*
1449          * Skip everything other than the PSN we expect, if we are waiting
1450          * for a reply to a restarted RDMA read or atomic op.
1451          */
1452         if (qp->r_flags & RVT_R_RDMAR_SEQ) {
1453                 if (cmp_psn(psn, qp->s_last_psn + 1) != 0)
1454                         goto ack_done;
1455                 qp->r_flags &= ~RVT_R_RDMAR_SEQ;
1456         }
1457
1458         if (unlikely(qp->s_acked == qp->s_tail))
1459                 goto ack_done;
1460         wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1461         status = IB_WC_SUCCESS;
1462
1463         switch (opcode) {
1464         case OP(ACKNOWLEDGE):
1465         case OP(ATOMIC_ACKNOWLEDGE):
1466         case OP(RDMA_READ_RESPONSE_FIRST):
1467                 aeth = be32_to_cpu(ohdr->u.aeth);
1468                 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1469                         val = ib_u64_get(&ohdr->u.at.atomic_ack_eth);
1470                 else
1471                         val = 0;
1472                 if (!do_rc_ack(qp, aeth, psn, opcode, val, rcd) ||
1473                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1474                         goto ack_done;
1475                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1476                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1477                         goto ack_op_err;
1478                 /*
1479                  * If this is a response to a resent RDMA read, we
1480                  * have to be careful to copy the data to the right
1481                  * location.
1482                  */
1483                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1484                                                   wqe, psn, pmtu);
1485                 goto read_middle;
1486
1487         case OP(RDMA_READ_RESPONSE_MIDDLE):
1488                 /* no AETH, no ACK */
1489                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1490                         goto ack_seq_err;
1491                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1492                         goto ack_op_err;
1493 read_middle:
1494                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1495                         goto ack_len_err;
1496                 if (unlikely(pmtu >= qp->s_rdma_read_len))
1497                         goto ack_len_err;
1498
1499                 /*
1500                  * We got a response so update the timeout.
1501                  * 4.096 usec. * (1 << qp->timeout)
1502                  */
1503                 rvt_mod_retry_timer(qp);
1504                 if (qp->s_flags & RVT_S_WAIT_ACK) {
1505                         qp->s_flags &= ~RVT_S_WAIT_ACK;
1506                         hfi1_schedule_send(qp);
1507                 }
1508
1509                 if (opcode == OP(RDMA_READ_RESPONSE_MIDDLE))
1510                         qp->s_retry = qp->s_retry_cnt;
1511
1512                 /*
1513                  * Update the RDMA receive state but do the copy w/o
1514                  * holding the locks and blocking interrupts.
1515                  */
1516                 qp->s_rdma_read_len -= pmtu;
1517                 update_last_psn(qp, psn);
1518                 spin_unlock_irqrestore(&qp->s_lock, flags);
1519                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, pmtu, false, false);
1520                 goto bail;
1521
1522         case OP(RDMA_READ_RESPONSE_ONLY):
1523                 aeth = be32_to_cpu(ohdr->u.aeth);
1524                 if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd))
1525                         goto ack_done;
1526                 /* Get the number of bytes the message was padded by. */
1527                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1528                 /*
1529                  * Check that the data size is >= 0 && <= pmtu.
1530                  * Remember to account for ICRC (4).
1531                  */
1532                 if (unlikely(tlen < (hdrsize + pad + 4)))
1533                         goto ack_len_err;
1534                 /*
1535                  * If this is a response to a resent RDMA read, we
1536                  * have to be careful to copy the data to the right
1537                  * location.
1538                  */
1539                 wqe = rvt_get_swqe_ptr(qp, qp->s_acked);
1540                 qp->s_rdma_read_len = restart_sge(&qp->s_rdma_read_sge,
1541                                                   wqe, psn, pmtu);
1542                 goto read_last;
1543
1544         case OP(RDMA_READ_RESPONSE_LAST):
1545                 /* ACKs READ req. */
1546                 if (unlikely(cmp_psn(psn, qp->s_last_psn + 1)))
1547                         goto ack_seq_err;
1548                 if (unlikely(wqe->wr.opcode != IB_WR_RDMA_READ))
1549                         goto ack_op_err;
1550                 /* Get the number of bytes the message was padded by. */
1551                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1552                 /*
1553                  * Check that the data size is >= 1 && <= pmtu.
1554                  * Remember to account for ICRC (4).
1555                  */
1556                 if (unlikely(tlen <= (hdrsize + pad + 4)))
1557                         goto ack_len_err;
1558 read_last:
1559                 tlen -= hdrsize + pad + 4;
1560                 if (unlikely(tlen != qp->s_rdma_read_len))
1561                         goto ack_len_err;
1562                 aeth = be32_to_cpu(ohdr->u.aeth);
1563                 hfi1_copy_sge(&qp->s_rdma_read_sge, data, tlen, false, false);
1564                 WARN_ON(qp->s_rdma_read_sge.num_sge);
1565                 (void)do_rc_ack(qp, aeth, psn,
1566                                  OP(RDMA_READ_RESPONSE_LAST), 0, rcd);
1567                 goto ack_done;
1568         }
1569
1570 ack_op_err:
1571         status = IB_WC_LOC_QP_OP_ERR;
1572         goto ack_err;
1573
1574 ack_seq_err:
1575         rdma_seq_err(qp, ibp, psn, rcd);
1576         goto ack_done;
1577
1578 ack_len_err:
1579         status = IB_WC_LOC_LEN_ERR;
1580 ack_err:
1581         if (qp->s_last == qp->s_acked) {
1582                 hfi1_send_complete(qp, wqe, status);
1583                 rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1584         }
1585 ack_done:
1586         spin_unlock_irqrestore(&qp->s_lock, flags);
1587 bail:
1588         return;
1589 }
1590
1591 static inline void rc_defered_ack(struct hfi1_ctxtdata *rcd,
1592                                   struct rvt_qp *qp)
1593 {
1594         if (list_empty(&qp->rspwait)) {
1595                 qp->r_flags |= RVT_R_RSP_NAK;
1596                 rvt_get_qp(qp);
1597                 list_add_tail(&qp->rspwait, &rcd->qp_wait_list);
1598         }
1599 }
1600
1601 static inline void rc_cancel_ack(struct rvt_qp *qp)
1602 {
1603         struct hfi1_qp_priv *priv = qp->priv;
1604
1605         priv->r_adefered = 0;
1606         if (list_empty(&qp->rspwait))
1607                 return;
1608         list_del_init(&qp->rspwait);
1609         qp->r_flags &= ~RVT_R_RSP_NAK;
1610         rvt_put_qp(qp);
1611 }
1612
1613 /**
1614  * rc_rcv_error - process an incoming duplicate or error RC packet
1615  * @ohdr: the other headers for this packet
1616  * @data: the packet data
1617  * @qp: the QP for this packet
1618  * @opcode: the opcode for this packet
1619  * @psn: the packet sequence number for this packet
1620  * @diff: the difference between the PSN and the expected PSN
1621  *
1622  * This is called from hfi1_rc_rcv() to process an unexpected
1623  * incoming RC packet for the given QP.
1624  * Called at interrupt level.
1625  * Return 1 if no more processing is needed; otherwise return 0 to
1626  * schedule a response to be sent.
1627  */
1628 static noinline int rc_rcv_error(struct ib_other_headers *ohdr, void *data,
1629                                  struct rvt_qp *qp, u32 opcode, u32 psn,
1630                                  int diff, struct hfi1_ctxtdata *rcd)
1631 {
1632         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1633         struct rvt_ack_entry *e;
1634         unsigned long flags;
1635         u8 i, prev;
1636         int old_req;
1637
1638         trace_hfi1_rcv_error(qp, psn);
1639         if (diff > 0) {
1640                 /*
1641                  * Packet sequence error.
1642                  * A NAK will ACK earlier sends and RDMA writes.
1643                  * Don't queue the NAK if we already sent one.
1644                  */
1645                 if (!qp->r_nak_state) {
1646                         ibp->rvp.n_rc_seqnak++;
1647                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1648                         /* Use the expected PSN. */
1649                         qp->r_ack_psn = qp->r_psn;
1650                         /*
1651                          * Wait to send the sequence NAK until all packets
1652                          * in the receive queue have been processed.
1653                          * Otherwise, we end up propagating congestion.
1654                          */
1655                         rc_defered_ack(rcd, qp);
1656                 }
1657                 goto done;
1658         }
1659
1660         /*
1661          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1662          * write or atomic op.  Don't NAK errors, just silently drop
1663          * the duplicate request.  Note that r_sge, r_len, and
1664          * r_rcv_len may be in use so don't modify them.
1665          *
1666          * We are supposed to ACK the earliest duplicate PSN but we
1667          * can coalesce an outstanding duplicate ACK.  We have to
1668          * send the earliest so that RDMA reads can be restarted at
1669          * the requester's expected PSN.
1670          *
1671          * First, find where this duplicate PSN falls within the
1672          * ACKs previously sent.
1673          * old_req is true if there is an older response that is scheduled
1674          * to be sent before sending this one.
1675          */
1676         e = NULL;
1677         old_req = 1;
1678         ibp->rvp.n_rc_dupreq++;
1679
1680         spin_lock_irqsave(&qp->s_lock, flags);
1681
1682         for (i = qp->r_head_ack_queue; ; i = prev) {
1683                 if (i == qp->s_tail_ack_queue)
1684                         old_req = 0;
1685                 if (i)
1686                         prev = i - 1;
1687                 else
1688                         prev = HFI1_MAX_RDMA_ATOMIC;
1689                 if (prev == qp->r_head_ack_queue) {
1690                         e = NULL;
1691                         break;
1692                 }
1693                 e = &qp->s_ack_queue[prev];
1694                 if (!e->opcode) {
1695                         e = NULL;
1696                         break;
1697                 }
1698                 if (cmp_psn(psn, e->psn) >= 0) {
1699                         if (prev == qp->s_tail_ack_queue &&
1700                             cmp_psn(psn, e->lpsn) <= 0)
1701                                 old_req = 0;
1702                         break;
1703                 }
1704         }
1705         switch (opcode) {
1706         case OP(RDMA_READ_REQUEST): {
1707                 struct ib_reth *reth;
1708                 u32 offset;
1709                 u32 len;
1710
1711                 /*
1712                  * If we didn't find the RDMA read request in the ack queue,
1713                  * we can ignore this request.
1714                  */
1715                 if (!e || e->opcode != OP(RDMA_READ_REQUEST))
1716                         goto unlock_done;
1717                 /* RETH comes after BTH */
1718                 reth = &ohdr->u.rc.reth;
1719                 /*
1720                  * Address range must be a subset of the original
1721                  * request and start on pmtu boundaries.
1722                  * We reuse the old ack_queue slot since the requester
1723                  * should not back up and request an earlier PSN for the
1724                  * same request.
1725                  */
1726                 offset = delta_psn(psn, e->psn) * qp->pmtu;
1727                 len = be32_to_cpu(reth->length);
1728                 if (unlikely(offset + len != e->rdma_sge.sge_length))
1729                         goto unlock_done;
1730                 if (e->rdma_sge.mr) {
1731                         rvt_put_mr(e->rdma_sge.mr);
1732                         e->rdma_sge.mr = NULL;
1733                 }
1734                 if (len != 0) {
1735                         u32 rkey = be32_to_cpu(reth->rkey);
1736                         u64 vaddr = get_ib_reth_vaddr(reth);
1737                         int ok;
1738
1739                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey,
1740                                          IB_ACCESS_REMOTE_READ);
1741                         if (unlikely(!ok))
1742                                 goto unlock_done;
1743                 } else {
1744                         e->rdma_sge.vaddr = NULL;
1745                         e->rdma_sge.length = 0;
1746                         e->rdma_sge.sge_length = 0;
1747                 }
1748                 e->psn = psn;
1749                 if (old_req)
1750                         goto unlock_done;
1751                 qp->s_tail_ack_queue = prev;
1752                 break;
1753         }
1754
1755         case OP(COMPARE_SWAP):
1756         case OP(FETCH_ADD): {
1757                 /*
1758                  * If we didn't find the atomic request in the ack queue
1759                  * or the send engine is already backed up to send an
1760                  * earlier entry, we can ignore this request.
1761                  */
1762                 if (!e || e->opcode != (u8)opcode || old_req)
1763                         goto unlock_done;
1764                 qp->s_tail_ack_queue = prev;
1765                 break;
1766         }
1767
1768         default:
1769                 /*
1770                  * Ignore this operation if it doesn't request an ACK
1771                  * or an earlier RDMA read or atomic is going to be resent.
1772                  */
1773                 if (!(psn & IB_BTH_REQ_ACK) || old_req)
1774                         goto unlock_done;
1775                 /*
1776                  * Resend the most recent ACK if this request is
1777                  * after all the previous RDMA reads and atomics.
1778                  */
1779                 if (i == qp->r_head_ack_queue) {
1780                         spin_unlock_irqrestore(&qp->s_lock, flags);
1781                         qp->r_nak_state = 0;
1782                         qp->r_ack_psn = qp->r_psn - 1;
1783                         goto send_ack;
1784                 }
1785
1786                 /*
1787                  * Resend the RDMA read or atomic op which
1788                  * ACKs this duplicate request.
1789                  */
1790                 qp->s_tail_ack_queue = i;
1791                 break;
1792         }
1793         qp->s_ack_state = OP(ACKNOWLEDGE);
1794         qp->s_flags |= RVT_S_RESP_PENDING;
1795         qp->r_nak_state = 0;
1796         hfi1_schedule_send(qp);
1797
1798 unlock_done:
1799         spin_unlock_irqrestore(&qp->s_lock, flags);
1800 done:
1801         return 1;
1802
1803 send_ack:
1804         return 0;
1805 }
1806
1807 static inline void update_ack_queue(struct rvt_qp *qp, unsigned n)
1808 {
1809         unsigned next;
1810
1811         next = n + 1;
1812         if (next > HFI1_MAX_RDMA_ATOMIC)
1813                 next = 0;
1814         qp->s_tail_ack_queue = next;
1815         qp->s_ack_state = OP(ACKNOWLEDGE);
1816 }
1817
1818 static void log_cca_event(struct hfi1_pportdata *ppd, u8 sl, u32 rlid,
1819                           u32 lqpn, u32 rqpn, u8 svc_type)
1820 {
1821         struct opa_hfi1_cong_log_event_internal *cc_event;
1822         unsigned long flags;
1823
1824         if (sl >= OPA_MAX_SLS)
1825                 return;
1826
1827         spin_lock_irqsave(&ppd->cc_log_lock, flags);
1828
1829         ppd->threshold_cong_event_map[sl / 8] |= 1 << (sl % 8);
1830         ppd->threshold_event_counter++;
1831
1832         cc_event = &ppd->cc_events[ppd->cc_log_idx++];
1833         if (ppd->cc_log_idx == OPA_CONG_LOG_ELEMS)
1834                 ppd->cc_log_idx = 0;
1835         cc_event->lqpn = lqpn & RVT_QPN_MASK;
1836         cc_event->rqpn = rqpn & RVT_QPN_MASK;
1837         cc_event->sl = sl;
1838         cc_event->svc_type = svc_type;
1839         cc_event->rlid = rlid;
1840         /* keep timestamp in units of 1.024 usec */
1841         cc_event->timestamp = ktime_to_ns(ktime_get()) / 1024;
1842
1843         spin_unlock_irqrestore(&ppd->cc_log_lock, flags);
1844 }
1845
1846 void process_becn(struct hfi1_pportdata *ppd, u8 sl, u16 rlid, u32 lqpn,
1847                   u32 rqpn, u8 svc_type)
1848 {
1849         struct cca_timer *cca_timer;
1850         u16 ccti, ccti_incr, ccti_timer, ccti_limit;
1851         u8 trigger_threshold;
1852         struct cc_state *cc_state;
1853         unsigned long flags;
1854
1855         if (sl >= OPA_MAX_SLS)
1856                 return;
1857
1858         cc_state = get_cc_state(ppd);
1859
1860         if (!cc_state)
1861                 return;
1862
1863         /*
1864          * 1) increase CCTI (for this SL)
1865          * 2) select IPG (i.e., call set_link_ipg())
1866          * 3) start timer
1867          */
1868         ccti_limit = cc_state->cct.ccti_limit;
1869         ccti_incr = cc_state->cong_setting.entries[sl].ccti_increase;
1870         ccti_timer = cc_state->cong_setting.entries[sl].ccti_timer;
1871         trigger_threshold =
1872                 cc_state->cong_setting.entries[sl].trigger_threshold;
1873
1874         spin_lock_irqsave(&ppd->cca_timer_lock, flags);
1875
1876         cca_timer = &ppd->cca_timer[sl];
1877         if (cca_timer->ccti < ccti_limit) {
1878                 if (cca_timer->ccti + ccti_incr <= ccti_limit)
1879                         cca_timer->ccti += ccti_incr;
1880                 else
1881                         cca_timer->ccti = ccti_limit;
1882                 set_link_ipg(ppd);
1883         }
1884
1885         ccti = cca_timer->ccti;
1886
1887         if (!hrtimer_active(&cca_timer->hrtimer)) {
1888                 /* ccti_timer is in units of 1.024 usec */
1889                 unsigned long nsec = 1024 * ccti_timer;
1890
1891                 hrtimer_start(&cca_timer->hrtimer, ns_to_ktime(nsec),
1892                               HRTIMER_MODE_REL);
1893         }
1894
1895         spin_unlock_irqrestore(&ppd->cca_timer_lock, flags);
1896
1897         if ((trigger_threshold != 0) && (ccti >= trigger_threshold))
1898                 log_cca_event(ppd, sl, rlid, lqpn, rqpn, svc_type);
1899 }
1900
1901 /**
1902  * hfi1_rc_rcv - process an incoming RC packet
1903  * @rcd: the context pointer
1904  * @hdr: the header of this packet
1905  * @rcv_flags: flags relevant to rcv processing
1906  * @data: the packet data
1907  * @tlen: the packet length
1908  * @qp: the QP for this packet
1909  *
1910  * This is called from qp_rcv() to process an incoming RC packet
1911  * for the given QP.
1912  * May be called at interrupt level.
1913  */
1914 void hfi1_rc_rcv(struct hfi1_packet *packet)
1915 {
1916         struct hfi1_ctxtdata *rcd = packet->rcd;
1917         struct ib_header *hdr = packet->hdr;
1918         u32 rcv_flags = packet->rcv_flags;
1919         void *data = packet->ebuf;
1920         u32 tlen = packet->tlen;
1921         struct rvt_qp *qp = packet->qp;
1922         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
1923         struct ib_other_headers *ohdr = packet->ohdr;
1924         u32 bth0, opcode;
1925         u32 hdrsize = packet->hlen;
1926         u32 psn;
1927         u32 pad;
1928         struct ib_wc wc;
1929         u32 pmtu = qp->pmtu;
1930         int diff;
1931         struct ib_reth *reth;
1932         unsigned long flags;
1933         int ret, is_fecn = 0;
1934         bool copy_last = false;
1935         u32 rkey;
1936
1937         lockdep_assert_held(&qp->r_lock);
1938         bth0 = be32_to_cpu(ohdr->bth[0]);
1939         if (hfi1_ruc_check_hdr(ibp, hdr, rcv_flags & HFI1_HAS_GRH, qp, bth0))
1940                 return;
1941
1942         is_fecn = process_ecn(qp, packet, false);
1943
1944         psn = be32_to_cpu(ohdr->bth[2]);
1945         opcode = (bth0 >> 24) & 0xff;
1946
1947         /*
1948          * Process responses (ACKs) before anything else.  Note that the
1949          * packet sequence number will be for something in the send work
1950          * queue rather than the expected receive packet sequence number.
1951          * In other words, this QP is the requester.
1952          */
1953         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1954             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1955                 rc_rcv_resp(ibp, ohdr, data, tlen, qp, opcode, psn,
1956                             hdrsize, pmtu, rcd);
1957                 if (is_fecn)
1958                         goto send_ack;
1959                 return;
1960         }
1961
1962         /* Compute 24 bits worth of difference. */
1963         diff = delta_psn(psn, qp->r_psn);
1964         if (unlikely(diff)) {
1965                 if (rc_rcv_error(ohdr, data, qp, opcode, psn, diff, rcd))
1966                         return;
1967                 goto send_ack;
1968         }
1969
1970         /* Check for opcode sequence errors. */
1971         switch (qp->r_state) {
1972         case OP(SEND_FIRST):
1973         case OP(SEND_MIDDLE):
1974                 if (opcode == OP(SEND_MIDDLE) ||
1975                     opcode == OP(SEND_LAST) ||
1976                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1977                     opcode == OP(SEND_LAST_WITH_INVALIDATE))
1978                         break;
1979                 goto nack_inv;
1980
1981         case OP(RDMA_WRITE_FIRST):
1982         case OP(RDMA_WRITE_MIDDLE):
1983                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1984                     opcode == OP(RDMA_WRITE_LAST) ||
1985                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1986                         break;
1987                 goto nack_inv;
1988
1989         default:
1990                 if (opcode == OP(SEND_MIDDLE) ||
1991                     opcode == OP(SEND_LAST) ||
1992                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1993                     opcode == OP(SEND_LAST_WITH_INVALIDATE) ||
1994                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1995                     opcode == OP(RDMA_WRITE_LAST) ||
1996                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1997                         goto nack_inv;
1998                 /*
1999                  * Note that it is up to the requester to not send a new
2000                  * RDMA read or atomic operation before receiving an ACK
2001                  * for the previous operation.
2002                  */
2003                 break;
2004         }
2005
2006         if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST))
2007                 rvt_comm_est(qp);
2008
2009         /* OK, process the packet. */
2010         switch (opcode) {
2011         case OP(SEND_FIRST):
2012                 ret = hfi1_rvt_get_rwqe(qp, 0);
2013                 if (ret < 0)
2014                         goto nack_op_err;
2015                 if (!ret)
2016                         goto rnr_nak;
2017                 qp->r_rcv_len = 0;
2018                 /* FALLTHROUGH */
2019         case OP(SEND_MIDDLE):
2020         case OP(RDMA_WRITE_MIDDLE):
2021 send_middle:
2022                 /* Check for invalid length PMTU or posted rwqe len. */
2023                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
2024                         goto nack_inv;
2025                 qp->r_rcv_len += pmtu;
2026                 if (unlikely(qp->r_rcv_len > qp->r_len))
2027                         goto nack_inv;
2028                 hfi1_copy_sge(&qp->r_sge, data, pmtu, true, false);
2029                 break;
2030
2031         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
2032                 /* consume RWQE */
2033                 ret = hfi1_rvt_get_rwqe(qp, 1);
2034                 if (ret < 0)
2035                         goto nack_op_err;
2036                 if (!ret)
2037                         goto rnr_nak;
2038                 goto send_last_imm;
2039
2040         case OP(SEND_ONLY):
2041         case OP(SEND_ONLY_WITH_IMMEDIATE):
2042         case OP(SEND_ONLY_WITH_INVALIDATE):
2043                 ret = hfi1_rvt_get_rwqe(qp, 0);
2044                 if (ret < 0)
2045                         goto nack_op_err;
2046                 if (!ret)
2047                         goto rnr_nak;
2048                 qp->r_rcv_len = 0;
2049                 if (opcode == OP(SEND_ONLY))
2050                         goto no_immediate_data;
2051                 if (opcode == OP(SEND_ONLY_WITH_INVALIDATE))
2052                         goto send_last_inv;
2053                 /* FALLTHROUGH for SEND_ONLY_WITH_IMMEDIATE */
2054         case OP(SEND_LAST_WITH_IMMEDIATE):
2055 send_last_imm:
2056                 wc.ex.imm_data = ohdr->u.imm_data;
2057                 wc.wc_flags = IB_WC_WITH_IMM;
2058                 goto send_last;
2059         case OP(SEND_LAST_WITH_INVALIDATE):
2060 send_last_inv:
2061                 rkey = be32_to_cpu(ohdr->u.ieth);
2062                 if (rvt_invalidate_rkey(qp, rkey))
2063                         goto no_immediate_data;
2064                 wc.ex.invalidate_rkey = rkey;
2065                 wc.wc_flags = IB_WC_WITH_INVALIDATE;
2066                 goto send_last;
2067         case OP(RDMA_WRITE_LAST):
2068                 copy_last = rvt_is_user_qp(qp);
2069                 /* fall through */
2070         case OP(SEND_LAST):
2071 no_immediate_data:
2072                 wc.wc_flags = 0;
2073                 wc.ex.imm_data = 0;
2074 send_last:
2075                 /* Get the number of bytes the message was padded by. */
2076                 pad = (bth0 >> 20) & 3;
2077                 /* Check for invalid length. */
2078                 /* LAST len should be >= 1 */
2079                 if (unlikely(tlen < (hdrsize + pad + 4)))
2080                         goto nack_inv;
2081                 /* Don't count the CRC. */
2082                 tlen -= (hdrsize + pad + 4);
2083                 wc.byte_len = tlen + qp->r_rcv_len;
2084                 if (unlikely(wc.byte_len > qp->r_len))
2085                         goto nack_inv;
2086                 hfi1_copy_sge(&qp->r_sge, data, tlen, true, copy_last);
2087                 rvt_put_ss(&qp->r_sge);
2088                 qp->r_msn++;
2089                 if (!__test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
2090                         break;
2091                 wc.wr_id = qp->r_wr_id;
2092                 wc.status = IB_WC_SUCCESS;
2093                 if (opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE) ||
2094                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
2095                         wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
2096                 else
2097                         wc.opcode = IB_WC_RECV;
2098                 wc.qp = &qp->ibqp;
2099                 wc.src_qp = qp->remote_qpn;
2100                 wc.slid = qp->remote_ah_attr.dlid;
2101                 /*
2102                  * It seems that IB mandates the presence of an SL in a
2103                  * work completion only for the UD transport (see section
2104                  * 11.4.2 of IBTA Vol. 1).
2105                  *
2106                  * However, the way the SL is chosen below is consistent
2107                  * with the way that IB/qib works and is trying avoid
2108                  * introducing incompatibilities.
2109                  *
2110                  * See also OPA Vol. 1, section 9.7.6, and table 9-17.
2111                  */
2112                 wc.sl = qp->remote_ah_attr.sl;
2113                 /* zero fields that are N/A */
2114                 wc.vendor_err = 0;
2115                 wc.pkey_index = 0;
2116                 wc.dlid_path_bits = 0;
2117                 wc.port_num = 0;
2118                 /* Signal completion event if the solicited bit is set. */
2119                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
2120                              (bth0 & IB_BTH_SOLICITED) != 0);
2121                 break;
2122
2123         case OP(RDMA_WRITE_ONLY):
2124                 copy_last = rvt_is_user_qp(qp);
2125                 /* fall through */
2126         case OP(RDMA_WRITE_FIRST):
2127         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
2128                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
2129                         goto nack_inv;
2130                 /* consume RWQE */
2131                 reth = &ohdr->u.rc.reth;
2132                 qp->r_len = be32_to_cpu(reth->length);
2133                 qp->r_rcv_len = 0;
2134                 qp->r_sge.sg_list = NULL;
2135                 if (qp->r_len != 0) {
2136                         u32 rkey = be32_to_cpu(reth->rkey);
2137                         u64 vaddr = get_ib_reth_vaddr(reth);
2138                         int ok;
2139
2140                         /* Check rkey & NAK */
2141                         ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len, vaddr,
2142                                          rkey, IB_ACCESS_REMOTE_WRITE);
2143                         if (unlikely(!ok))
2144                                 goto nack_acc;
2145                         qp->r_sge.num_sge = 1;
2146                 } else {
2147                         qp->r_sge.num_sge = 0;
2148                         qp->r_sge.sge.mr = NULL;
2149                         qp->r_sge.sge.vaddr = NULL;
2150                         qp->r_sge.sge.length = 0;
2151                         qp->r_sge.sge.sge_length = 0;
2152                 }
2153                 if (opcode == OP(RDMA_WRITE_FIRST))
2154                         goto send_middle;
2155                 else if (opcode == OP(RDMA_WRITE_ONLY))
2156                         goto no_immediate_data;
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                 wc.ex.imm_data = ohdr->u.rc.imm_data;
2163                 wc.wc_flags = IB_WC_WITH_IMM;
2164                 goto send_last;
2165
2166         case OP(RDMA_READ_REQUEST): {
2167                 struct rvt_ack_entry *e;
2168                 u32 len;
2169                 u8 next;
2170
2171                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
2172                         goto nack_inv;
2173                 next = qp->r_head_ack_queue + 1;
2174                 /* s_ack_queue is size HFI1_MAX_RDMA_ATOMIC+1 so use > not >= */
2175                 if (next > HFI1_MAX_RDMA_ATOMIC)
2176                         next = 0;
2177                 spin_lock_irqsave(&qp->s_lock, flags);
2178                 if (unlikely(next == qp->s_tail_ack_queue)) {
2179                         if (!qp->s_ack_queue[next].sent)
2180                                 goto nack_inv_unlck;
2181                         update_ack_queue(qp, next);
2182                 }
2183                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2184                 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2185                         rvt_put_mr(e->rdma_sge.mr);
2186                         e->rdma_sge.mr = NULL;
2187                 }
2188                 reth = &ohdr->u.rc.reth;
2189                 len = be32_to_cpu(reth->length);
2190                 if (len) {
2191                         u32 rkey = be32_to_cpu(reth->rkey);
2192                         u64 vaddr = get_ib_reth_vaddr(reth);
2193                         int ok;
2194
2195                         /* Check rkey & NAK */
2196                         ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr,
2197                                          rkey, IB_ACCESS_REMOTE_READ);
2198                         if (unlikely(!ok))
2199                                 goto nack_acc_unlck;
2200                         /*
2201                          * Update the next expected PSN.  We add 1 later
2202                          * below, so only add the remainder here.
2203                          */
2204                         qp->r_psn += rvt_div_mtu(qp, len - 1);
2205                 } else {
2206                         e->rdma_sge.mr = NULL;
2207                         e->rdma_sge.vaddr = NULL;
2208                         e->rdma_sge.length = 0;
2209                         e->rdma_sge.sge_length = 0;
2210                 }
2211                 e->opcode = opcode;
2212                 e->sent = 0;
2213                 e->psn = psn;
2214                 e->lpsn = qp->r_psn;
2215                 /*
2216                  * We need to increment the MSN here instead of when we
2217                  * finish sending the result since a duplicate request would
2218                  * increment it more than once.
2219                  */
2220                 qp->r_msn++;
2221                 qp->r_psn++;
2222                 qp->r_state = opcode;
2223                 qp->r_nak_state = 0;
2224                 qp->r_head_ack_queue = next;
2225
2226                 /* Schedule the send engine. */
2227                 qp->s_flags |= RVT_S_RESP_PENDING;
2228                 hfi1_schedule_send(qp);
2229
2230                 spin_unlock_irqrestore(&qp->s_lock, flags);
2231                 if (is_fecn)
2232                         goto send_ack;
2233                 return;
2234         }
2235
2236         case OP(COMPARE_SWAP):
2237         case OP(FETCH_ADD): {
2238                 struct ib_atomic_eth *ateth;
2239                 struct rvt_ack_entry *e;
2240                 u64 vaddr;
2241                 atomic64_t *maddr;
2242                 u64 sdata;
2243                 u32 rkey;
2244                 u8 next;
2245
2246                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
2247                         goto nack_inv;
2248                 next = qp->r_head_ack_queue + 1;
2249                 if (next > HFI1_MAX_RDMA_ATOMIC)
2250                         next = 0;
2251                 spin_lock_irqsave(&qp->s_lock, flags);
2252                 if (unlikely(next == qp->s_tail_ack_queue)) {
2253                         if (!qp->s_ack_queue[next].sent)
2254                                 goto nack_inv_unlck;
2255                         update_ack_queue(qp, next);
2256                 }
2257                 e = &qp->s_ack_queue[qp->r_head_ack_queue];
2258                 if (e->opcode == OP(RDMA_READ_REQUEST) && e->rdma_sge.mr) {
2259                         rvt_put_mr(e->rdma_sge.mr);
2260                         e->rdma_sge.mr = NULL;
2261                 }
2262                 ateth = &ohdr->u.atomic_eth;
2263                 vaddr = get_ib_ateth_vaddr(ateth);
2264                 if (unlikely(vaddr & (sizeof(u64) - 1)))
2265                         goto nack_inv_unlck;
2266                 rkey = be32_to_cpu(ateth->rkey);
2267                 /* Check rkey & NAK */
2268                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
2269                                           vaddr, rkey,
2270                                           IB_ACCESS_REMOTE_ATOMIC)))
2271                         goto nack_acc_unlck;
2272                 /* Perform atomic OP and save result. */
2273                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
2274                 sdata = get_ib_ateth_swap(ateth);
2275                 e->atomic_data = (opcode == OP(FETCH_ADD)) ?
2276                         (u64)atomic64_add_return(sdata, maddr) - sdata :
2277                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
2278                                       get_ib_ateth_compare(ateth),
2279                                       sdata);
2280                 rvt_put_mr(qp->r_sge.sge.mr);
2281                 qp->r_sge.num_sge = 0;
2282                 e->opcode = opcode;
2283                 e->sent = 0;
2284                 e->psn = psn;
2285                 e->lpsn = psn;
2286                 qp->r_msn++;
2287                 qp->r_psn++;
2288                 qp->r_state = opcode;
2289                 qp->r_nak_state = 0;
2290                 qp->r_head_ack_queue = next;
2291
2292                 /* Schedule the send engine. */
2293                 qp->s_flags |= RVT_S_RESP_PENDING;
2294                 hfi1_schedule_send(qp);
2295
2296                 spin_unlock_irqrestore(&qp->s_lock, flags);
2297                 if (is_fecn)
2298                         goto send_ack;
2299                 return;
2300         }
2301
2302         default:
2303                 /* NAK unknown opcodes. */
2304                 goto nack_inv;
2305         }
2306         qp->r_psn++;
2307         qp->r_state = opcode;
2308         qp->r_ack_psn = psn;
2309         qp->r_nak_state = 0;
2310         /* Send an ACK if requested or required. */
2311         if (psn & IB_BTH_REQ_ACK) {
2312                 struct hfi1_qp_priv *priv = qp->priv;
2313
2314                 if (packet->numpkt == 0) {
2315                         rc_cancel_ack(qp);
2316                         goto send_ack;
2317                 }
2318                 if (priv->r_adefered >= HFI1_PSN_CREDIT) {
2319                         rc_cancel_ack(qp);
2320                         goto send_ack;
2321                 }
2322                 if (unlikely(is_fecn)) {
2323                         rc_cancel_ack(qp);
2324                         goto send_ack;
2325                 }
2326                 priv->r_adefered++;
2327                 rc_defered_ack(rcd, qp);
2328         }
2329         return;
2330
2331 rnr_nak:
2332         qp->r_nak_state = qp->r_min_rnr_timer | IB_RNR_NAK;
2333         qp->r_ack_psn = qp->r_psn;
2334         /* Queue RNR NAK for later */
2335         rc_defered_ack(rcd, qp);
2336         return;
2337
2338 nack_op_err:
2339         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2340         qp->r_nak_state = IB_NAK_REMOTE_OPERATIONAL_ERROR;
2341         qp->r_ack_psn = qp->r_psn;
2342         /* Queue NAK for later */
2343         rc_defered_ack(rcd, qp);
2344         return;
2345
2346 nack_inv_unlck:
2347         spin_unlock_irqrestore(&qp->s_lock, flags);
2348 nack_inv:
2349         rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
2350         qp->r_nak_state = IB_NAK_INVALID_REQUEST;
2351         qp->r_ack_psn = qp->r_psn;
2352         /* Queue NAK for later */
2353         rc_defered_ack(rcd, qp);
2354         return;
2355
2356 nack_acc_unlck:
2357         spin_unlock_irqrestore(&qp->s_lock, flags);
2358 nack_acc:
2359         rvt_rc_error(qp, IB_WC_LOC_PROT_ERR);
2360         qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
2361         qp->r_ack_psn = qp->r_psn;
2362 send_ack:
2363         hfi1_send_rc_ack(rcd, qp, is_fecn);
2364 }
2365
2366 void hfi1_rc_hdrerr(
2367         struct hfi1_ctxtdata *rcd,
2368         struct ib_header *hdr,
2369         u32 rcv_flags,
2370         struct rvt_qp *qp)
2371 {
2372         int has_grh = rcv_flags & HFI1_HAS_GRH;
2373         struct ib_other_headers *ohdr;
2374         struct hfi1_ibport *ibp = rcd_to_iport(rcd);
2375         int diff;
2376         u32 opcode;
2377         u32 psn, bth0;
2378
2379         /* Check for GRH */
2380         ohdr = &hdr->u.oth;
2381         if (has_grh)
2382                 ohdr = &hdr->u.l.oth;
2383
2384         bth0 = be32_to_cpu(ohdr->bth[0]);
2385         if (hfi1_ruc_check_hdr(ibp, hdr, has_grh, qp, bth0))
2386                 return;
2387
2388         psn = be32_to_cpu(ohdr->bth[2]);
2389         opcode = (bth0 >> 24) & 0xff;
2390
2391         /* Only deal with RDMA Writes for now */
2392         if (opcode < IB_OPCODE_RC_RDMA_READ_RESPONSE_FIRST) {
2393                 diff = delta_psn(psn, qp->r_psn);
2394                 if (!qp->r_nak_state && diff >= 0) {
2395                         ibp->rvp.n_rc_seqnak++;
2396                         qp->r_nak_state = IB_NAK_PSN_ERROR;
2397                         /* Use the expected PSN. */
2398                         qp->r_ack_psn = qp->r_psn;
2399                         /*
2400                          * Wait to send the sequence
2401                          * NAK until all packets
2402                          * in the receive queue have
2403                          * been processed.
2404                          * Otherwise, we end up
2405                          * propagating congestion.
2406                          */
2407                         rc_defered_ack(rcd, qp);
2408                 } /* Out of sequence NAK */
2409         } /* QP Request NAKs */
2410 }