drivers/staging/rdma/hfi1/ruc.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/spinlock.h>
  49
  50 #include "hfi.h"
  51 #include "mad.h"
  52 #include "qp.h"
  53 #include "verbs_txreq.h"
  54 #include "trace.h"
  55
  56 /*
  57  * Convert the AETH RNR timeout code into the number of microseconds.
  58  */
  59 const u32 ib_hfi1_rnr_table[32] = {
  60         655360, /* 00: 655.36 */
  61         10,     /* 01:    .01 */
  62         20,     /* 02     .02 */
  63         30,     /* 03:    .03 */
  64         40,     /* 04:    .04 */
  65         60,     /* 05:    .06 */
  66         80,     /* 06:    .08 */
  67         120,    /* 07:    .12 */
  68         160,    /* 08:    .16 */
  69         240,    /* 09:    .24 */
  70         320,    /* 0A:    .32 */
  71         480,    /* 0B:    .48 */
  72         640,    /* 0C:    .64 */
  73         960,    /* 0D:    .96 */
  74         1280,   /* 0E:   1.28 */
  75         1920,   /* 0F:   1.92 */
  76         2560,   /* 10:   2.56 */
  77         3840,   /* 11:   3.84 */
  78         5120,   /* 12:   5.12 */
  79         7680,   /* 13:   7.68 */
  80         10240,  /* 14:  10.24 */
  81         15360,  /* 15:  15.36 */
  82         20480,  /* 16:  20.48 */
  83         30720,  /* 17:  30.72 */
  84         40960,  /* 18:  40.96 */
  85         61440,  /* 19:  61.44 */
  86         81920,  /* 1A:  81.92 */
  87         122880, /* 1B: 122.88 */
  88         163840, /* 1C: 163.84 */
  89         245760, /* 1D: 245.76 */
  90         327680, /* 1E: 327.68 */
  91         491520  /* 1F: 491.52 */
  92 };
  93
  94 /*
  95  * Validate a RWQE and fill in the SGE state.
  96  * Return 1 if OK.
  97  */
  98 static int init_sge(struct rvt_qp *qp, struct rvt_rwqe *wqe)
  99 {
 100         int i, j, ret;
 101         struct ib_wc wc;
 102         struct rvt_lkey_table *rkt;
 103         struct rvt_pd *pd;
 104         struct rvt_sge_state *ss;
 105
 106         rkt = &to_idev(qp->ibqp.device)->rdi.lkey_table;
 107         pd = ibpd_to_rvtpd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
 108         ss = &qp->r_sge;
 109         ss->sg_list = qp->r_sg_list;
 110         qp->r_len = 0;
 111         for (i = j = 0; i < wqe->num_sge; i++) {
 112                 if (wqe->sg_list[i].length == 0)
 113                         continue;
 114                 /* Check LKEY */
 115                 if (!rvt_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
 116                                  &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
 117                         goto bad_lkey;
 118                 qp->r_len += wqe->sg_list[i].length;
 119                 j++;
 120         }
 121         ss->num_sge = j;
 122         ss->total_len = qp->r_len;
 123         ret = 1;
 124         goto bail;
 125
 126 bad_lkey:
 127         while (j) {
 128                 struct rvt_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
 129
 130                 rvt_put_mr(sge->mr);
 131         }
 132         ss->num_sge = 0;
 133         memset(&wc, 0, sizeof(wc));
 134         wc.wr_id = wqe->wr_id;
 135         wc.status = IB_WC_LOC_PROT_ERR;
 136         wc.opcode = IB_WC_RECV;
 137         wc.qp = &qp->ibqp;
 138         /* Signal solicited completion event. */
 139         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
 140         ret = 0;
 141 bail:
 142         return ret;
 143 }
 144
 145 /**
 146  * hfi1_rvt_get_rwqe - copy the next RWQE into the QP's RWQE
 147  * @qp: the QP
 148  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
 149  *
 150  * Return -1 if there is a local error, 0 if no RWQE is available,
 151  * otherwise return 1.
 152  *
 153  * Can be called from interrupt level.
 154  */
 155 int hfi1_rvt_get_rwqe(struct rvt_qp *qp, int wr_id_only)
 156 {
 157         unsigned long flags;
 158         struct rvt_rq *rq;
 159         struct rvt_rwq *wq;
 160         struct rvt_srq *srq;
 161         struct rvt_rwqe *wqe;
 162         void (*handler)(struct ib_event *, void *);
 163         u32 tail;
 164         int ret;
 165
 166         if (qp->ibqp.srq) {
 167                 srq = ibsrq_to_rvtsrq(qp->ibqp.srq);
 168                 handler = srq->ibsrq.event_handler;
 169                 rq = &srq->rq;
 170         } else {
 171                 srq = NULL;
 172                 handler = NULL;
 173                 rq = &qp->r_rq;
 174         }
 175
 176         spin_lock_irqsave(&rq->lock, flags);
 177         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
 178                 ret = 0;
 179                 goto unlock;
 180         }
 181
 182         wq = rq->wq;
 183         tail = wq->tail;
 184         /* Validate tail before using it since it is user writable. */
 185         if (tail >= rq->size)
 186                 tail = 0;
 187         if (unlikely(tail == wq->head)) {
 188                 ret = 0;
 189                 goto unlock;
 190         }
 191         /* Make sure entry is read after head index is read. */
 192         smp_rmb();
 193         wqe = rvt_get_rwqe_ptr(rq, tail);
 194         /*
 195          * Even though we update the tail index in memory, the verbs
 196          * consumer is not supposed to post more entries until a
 197          * completion is generated.
 198          */
 199         if (++tail >= rq->size)
 200                 tail = 0;
 201         wq->tail = tail;
 202         if (!wr_id_only && !init_sge(qp, wqe)) {
 203                 ret = -1;
 204                 goto unlock;
 205         }
 206         qp->r_wr_id = wqe->wr_id;
 207
 208         ret = 1;
 209         set_bit(RVT_R_WRID_VALID, &qp->r_aflags);
 210         if (handler) {
 211                 u32 n;
 212
 213                 /*
 214                  * Validate head pointer value and compute
 215                  * the number of remaining WQEs.
 216                  */
 217                 n = wq->head;
 218                 if (n >= rq->size)
 219                         n = 0;
 220                 if (n < tail)
 221                         n += rq->size - tail;
 222                 else
 223                         n -= tail;
 224                 if (n < srq->limit) {
 225                         struct ib_event ev;
 226
 227                         srq->limit = 0;
 228                         spin_unlock_irqrestore(&rq->lock, flags);
 229                         ev.device = qp->ibqp.device;
 230                         ev.element.srq = qp->ibqp.srq;
 231                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 232                         handler(&ev, srq->ibsrq.srq_context);
 233                         goto bail;
 234                 }
 235         }
 236 unlock:
 237         spin_unlock_irqrestore(&rq->lock, flags);
 238 bail:
 239         return ret;
 240 }
 241
 242 static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index)
 243 {
 244         if (!index) {
 245                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 246
 247                 return cpu_to_be64(ppd->guid);
 248         }
 249         return ibp->guids[index - 1];
 250 }
 251
 252 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
 253 {
 254         return (gid->global.interface_id == id &&
 255                 (gid->global.subnet_prefix == gid_prefix ||
 256                  gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
 257 }
 258
 259 /*
 260  *
 261  * This should be called with the QP r_lock held.
 262  *
 263  * The s_lock will be acquired around the hfi1_migrate_qp() call.
 264  */
 265 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_ib_header *hdr,
 266                        int has_grh, struct rvt_qp *qp, u32 bth0)
 267 {
 268         __be64 guid;
 269         unsigned long flags;
 270         u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
 271
 272         if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
 273                 if (!has_grh) {
 274                         if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
 275                                 goto err;
 276                 } else {
 277                         if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
 278                                 goto err;
 279                         guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
 280                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
 281                                     guid))
 282                                 goto err;
 283                         if (!gid_ok(
 284                                 &hdr->u.l.grh.sgid,
 285                                 qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
 286                                 qp->alt_ah_attr.grh.dgid.global.interface_id))
 287                                 goto err;
 288                 }
 289                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
 290                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
 291                         hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
 292                                        (u16)bth0,
 293                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 294                                        0, qp->ibqp.qp_num,
 295                                        be16_to_cpu(hdr->lrh[3]),
 296                                        be16_to_cpu(hdr->lrh[1]));
 297                         goto err;
 298                 }
 299                 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
 300                 if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
 301                     ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
 302                         goto err;
 303                 spin_lock_irqsave(&qp->s_lock, flags);
 304                 hfi1_migrate_qp(qp);
 305                 spin_unlock_irqrestore(&qp->s_lock, flags);
 306         } else {
 307                 if (!has_grh) {
 308                         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 309                                 goto err;
 310                 } else {
 311                         if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
 312                                 goto err;
 313                         guid = get_sguid(ibp,
 314                                          qp->remote_ah_attr.grh.sgid_index);
 315                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->rvp.gid_prefix,
 316                                     guid))
 317                                 goto err;
 318                         if (!gid_ok(
 319                              &hdr->u.l.grh.sgid,
 320                              qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
 321                              qp->remote_ah_attr.grh.dgid.global.interface_id))
 322                                 goto err;
 323                 }
 324                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
 325                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
 326                         hfi1_bad_pqkey(ibp, OPA_TRAP_BAD_P_KEY,
 327                                        (u16)bth0,
 328                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 329                                        0, qp->ibqp.qp_num,
 330                                        be16_to_cpu(hdr->lrh[3]),
 331                                        be16_to_cpu(hdr->lrh[1]));
 332                         goto err;
 333                 }
 334                 /* Validate the SLID. See Ch. 9.6.1.5 */
 335                 if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
 336                     ppd_from_ibp(ibp)->port != qp->port_num)
 337                         goto err;
 338                 if (qp->s_mig_state == IB_MIG_REARM &&
 339                     !(bth0 & IB_BTH_MIG_REQ))
 340                         qp->s_mig_state = IB_MIG_ARMED;
 341         }
 342
 343         return 0;
 344
 345 err:
 346         return 1;
 347 }
 348
 349 /**
 350  * ruc_loopback - handle UC and RC loopback requests
 351  * @sqp: the sending QP
 352  *
 353  * This is called from hfi1_do_send() to
 354  * forward a WQE addressed to the same HFI.
 355  * Note that although we are single threaded due to the tasklet, we still
 356  * have to protect against post_send().  We don't have to worry about
 357  * receive interrupts since this is a connected protocol and all packets
 358  * will pass through here.
 359  */
 360 static void ruc_loopback(struct rvt_qp *sqp)
 361 {
 362         struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
 363         struct rvt_qp *qp;
 364         struct rvt_swqe *wqe;
 365         struct rvt_sge *sge;
 366         unsigned long flags;
 367         struct ib_wc wc;
 368         u64 sdata;
 369         atomic64_t *maddr;
 370         enum ib_wc_status send_status;
 371         int release;
 372         int ret;
 373         int copy_last = 0;
 374         u32 to;
 375
 376         rcu_read_lock();
 377
 378         /*
 379          * Note that we check the responder QP state after
 380          * checking the requester's state.
 381          */
 382         qp = rvt_lookup_qpn(ib_to_rvt(sqp->ibqp.device), &ibp->rvp,
 383                             sqp->remote_qpn);
 384
 385         spin_lock_irqsave(&sqp->s_lock, flags);
 386
 387         /* Return if we are already busy processing a work request. */
 388         if ((sqp->s_flags & (RVT_S_BUSY | RVT_S_ANY_WAIT)) ||
 389             !(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 390                 goto unlock;
 391
 392         sqp->s_flags |= RVT_S_BUSY;
 393
 394 again:
 395         smp_read_barrier_depends(); /* see post_one_send() */
 396         if (sqp->s_last == ACCESS_ONCE(sqp->s_head))
 397                 goto clr_busy;
 398         wqe = rvt_get_swqe_ptr(sqp, sqp->s_last);
 399
 400         /* Return if it is not OK to start a new work request. */
 401         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_NEXT_SEND_OK)) {
 402                 if (!(ib_rvt_state_ops[sqp->state] & RVT_FLUSH_SEND))
 403                         goto clr_busy;
 404                 /* We are in the error state, flush the work request. */
 405                 send_status = IB_WC_WR_FLUSH_ERR;
 406                 goto flush_send;
 407         }
 408
 409         /*
 410          * We can rely on the entry not changing without the s_lock
 411          * being held until we update s_last.
 412          * We increment s_cur to indicate s_last is in progress.
 413          */
 414         if (sqp->s_last == sqp->s_cur) {
 415                 if (++sqp->s_cur >= sqp->s_size)
 416                         sqp->s_cur = 0;
 417         }
 418         spin_unlock_irqrestore(&sqp->s_lock, flags);
 419
 420         if (!qp || !(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) ||
 421             qp->ibqp.qp_type != sqp->ibqp.qp_type) {
 422                 ibp->rvp.n_pkt_drops++;
 423                 /*
 424                  * For RC, the requester would timeout and retry so
 425                  * shortcut the timeouts and just signal too many retries.
 426                  */
 427                 if (sqp->ibqp.qp_type == IB_QPT_RC)
 428                         send_status = IB_WC_RETRY_EXC_ERR;
 429                 else
 430                         send_status = IB_WC_SUCCESS;
 431                 goto serr;
 432         }
 433
 434         memset(&wc, 0, sizeof(wc));
 435         send_status = IB_WC_SUCCESS;
 436
 437         release = 1;
 438         sqp->s_sge.sge = wqe->sg_list[0];
 439         sqp->s_sge.sg_list = wqe->sg_list + 1;
 440         sqp->s_sge.num_sge = wqe->wr.num_sge;
 441         sqp->s_len = wqe->length;
 442         switch (wqe->wr.opcode) {
 443         case IB_WR_SEND_WITH_IMM:
 444                 wc.wc_flags = IB_WC_WITH_IMM;
 445                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 446                 /* FALLTHROUGH */
 447         case IB_WR_SEND:
 448                 ret = hfi1_rvt_get_rwqe(qp, 0);
 449                 if (ret < 0)
 450                         goto op_err;
 451                 if (!ret)
 452                         goto rnr_nak;
 453                 break;
 454
 455         case IB_WR_RDMA_WRITE_WITH_IMM:
 456                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 457                         goto inv_err;
 458                 wc.wc_flags = IB_WC_WITH_IMM;
 459                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 460                 ret = hfi1_rvt_get_rwqe(qp, 1);
 461                 if (ret < 0)
 462                         goto op_err;
 463                 if (!ret)
 464                         goto rnr_nak;
 465                 /* skip copy_last set and qp_access_flags recheck */
 466                 goto do_write;
 467         case IB_WR_RDMA_WRITE:
 468                 copy_last = ibpd_to_rvtpd(qp->ibqp.pd)->user;
 469                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 470                         goto inv_err;
 471 do_write:
 472                 if (wqe->length == 0)
 473                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
 474                                           wqe->rdma_wr.remote_addr,
 475                                           wqe->rdma_wr.rkey,
 476                                           IB_ACCESS_REMOTE_WRITE)))
 477                         goto acc_err;
 478                 qp->r_sge.sg_list = NULL;
 479                 qp->r_sge.num_sge = 1;
 480                 qp->r_sge.total_len = wqe->length;
 481                 break;
 482
 483         case IB_WR_RDMA_READ:
 484                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 485                         goto inv_err;
 486                 if (unlikely(!rvt_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
 487                                           wqe->rdma_wr.remote_addr,
 488                                           wqe->rdma_wr.rkey,
 489                                           IB_ACCESS_REMOTE_READ)))
 490                         goto acc_err;
 491                 release = 0;
 492                 sqp->s_sge.sg_list = NULL;
 493                 sqp->s_sge.num_sge = 1;
 494                 qp->r_sge.sge = wqe->sg_list[0];
 495                 qp->r_sge.sg_list = wqe->sg_list + 1;
 496                 qp->r_sge.num_sge = wqe->wr.num_sge;
 497                 qp->r_sge.total_len = wqe->length;
 498                 break;
 499
 500         case IB_WR_ATOMIC_CMP_AND_SWP:
 501         case IB_WR_ATOMIC_FETCH_AND_ADD:
 502                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 503                         goto inv_err;
 504                 if (unlikely(!rvt_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
 505                                           wqe->atomic_wr.remote_addr,
 506                                           wqe->atomic_wr.rkey,
 507                                           IB_ACCESS_REMOTE_ATOMIC)))
 508                         goto acc_err;
 509                 /* Perform atomic OP and save result. */
 510                 maddr = (atomic64_t *)qp->r_sge.sge.vaddr;
 511                 sdata = wqe->atomic_wr.compare_add;
 512                 *(u64 *)sqp->s_sge.sge.vaddr =
 513                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 514                         (u64)atomic64_add_return(sdata, maddr) - sdata :
 515                         (u64)cmpxchg((u64 *)qp->r_sge.sge.vaddr,
 516                                       sdata, wqe->atomic_wr.swap);
 517                 rvt_put_mr(qp->r_sge.sge.mr);
 518                 qp->r_sge.num_sge = 0;
 519                 goto send_comp;
 520
 521         default:
 522                 send_status = IB_WC_LOC_QP_OP_ERR;
 523                 goto serr;
 524         }
 525
 526         sge = &sqp->s_sge.sge;
 527         while (sqp->s_len) {
 528                 u32 len = sqp->s_len;
 529
 530                 if (len > sge->length)
 531                         len = sge->length;
 532                 if (len > sge->sge_length)
 533                         len = sge->sge_length;
 534                 WARN_ON_ONCE(len == 0);
 535                 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release, copy_last);
 536                 sge->vaddr += len;
 537                 sge->length -= len;
 538                 sge->sge_length -= len;
 539                 if (sge->sge_length == 0) {
 540                         if (!release)
 541                                 rvt_put_mr(sge->mr);
 542                         if (--sqp->s_sge.num_sge)
 543                                 *sge = *sqp->s_sge.sg_list++;
 544                 } else if (sge->length == 0 && sge->mr->lkey) {
 545                         if (++sge->n >= RVT_SEGSZ) {
 546                                 if (++sge->m >= sge->mr->mapsz)
 547                                         break;
 548                                 sge->n = 0;
 549                         }
 550                         sge->vaddr =
 551                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 552                         sge->length =
 553                                 sge->mr->map[sge->m]->segs[sge->n].length;
 554                 }
 555                 sqp->s_len -= len;
 556         }
 557         if (release)
 558                 rvt_put_ss(&qp->r_sge);
 559
 560         if (!test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags))
 561                 goto send_comp;
 562
 563         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 564                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 565         else
 566                 wc.opcode = IB_WC_RECV;
 567         wc.wr_id = qp->r_wr_id;
 568         wc.status = IB_WC_SUCCESS;
 569         wc.byte_len = wqe->length;
 570         wc.qp = &qp->ibqp;
 571         wc.src_qp = qp->remote_qpn;
 572         wc.slid = qp->remote_ah_attr.dlid;
 573         wc.sl = qp->remote_ah_attr.sl;
 574         wc.port_num = 1;
 575         /* Signal completion event if the solicited bit is set. */
 576         rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
 577                      wqe->wr.send_flags & IB_SEND_SOLICITED);
 578
 579 send_comp:
 580         spin_lock_irqsave(&sqp->s_lock, flags);
 581         ibp->rvp.n_loop_pkts++;
 582 flush_send:
 583         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 584         hfi1_send_complete(sqp, wqe, send_status);
 585         goto again;
 586
 587 rnr_nak:
 588         /* Handle RNR NAK */
 589         if (qp->ibqp.qp_type == IB_QPT_UC)
 590                 goto send_comp;
 591         ibp->rvp.n_rnr_naks++;
 592         /*
 593          * Note: we don't need the s_lock held since the BUSY flag
 594          * makes this single threaded.
 595          */
 596         if (sqp->s_rnr_retry == 0) {
 597                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 598                 goto serr;
 599         }
 600         if (sqp->s_rnr_retry_cnt < 7)
 601                 sqp->s_rnr_retry--;
 602         spin_lock_irqsave(&sqp->s_lock, flags);
 603         if (!(ib_rvt_state_ops[sqp->state] & RVT_PROCESS_RECV_OK))
 604                 goto clr_busy;
 605         to = ib_hfi1_rnr_table[qp->r_min_rnr_timer];
 606         hfi1_add_rnr_timer(sqp, to);
 607         goto clr_busy;
 608
 609 op_err:
 610         send_status = IB_WC_REM_OP_ERR;
 611         wc.status = IB_WC_LOC_QP_OP_ERR;
 612         goto err;
 613
 614 inv_err:
 615         send_status = IB_WC_REM_INV_REQ_ERR;
 616         wc.status = IB_WC_LOC_QP_OP_ERR;
 617         goto err;
 618
 619 acc_err:
 620         send_status = IB_WC_REM_ACCESS_ERR;
 621         wc.status = IB_WC_LOC_PROT_ERR;
 622 err:
 623         /* responder goes to error state */
 624         hfi1_rc_error(qp, wc.status);
 625
 626 serr:
 627         spin_lock_irqsave(&sqp->s_lock, flags);
 628         hfi1_send_complete(sqp, wqe, send_status);
 629         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 630                 int lastwqe = rvt_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 631
 632                 sqp->s_flags &= ~RVT_S_BUSY;
 633                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 634                 if (lastwqe) {
 635                         struct ib_event ev;
 636
 637                         ev.device = sqp->ibqp.device;
 638                         ev.element.qp = &sqp->ibqp;
 639                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 640                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 641                 }
 642                 goto done;
 643         }
 644 clr_busy:
 645         sqp->s_flags &= ~RVT_S_BUSY;
 646 unlock:
 647         spin_unlock_irqrestore(&sqp->s_lock, flags);
 648 done:
 649         rcu_read_unlock();
 650 }
 651
 652 /**
 653  * hfi1_make_grh - construct a GRH header
 654  * @ibp: a pointer to the IB port
 655  * @hdr: a pointer to the GRH header being constructed
 656  * @grh: the global route address to send to
 657  * @hwords: the number of 32 bit words of header being sent
 658  * @nwords: the number of 32 bit words of data being sent
 659  *
 660  * Return the size of the header in 32 bit words.
 661  */
 662 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
 663                   struct ib_global_route *grh, u32 hwords, u32 nwords)
 664 {
 665         hdr->version_tclass_flow =
 666                 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
 667                             (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
 668                             (grh->flow_label << IB_GRH_FLOW_SHIFT));
 669         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 670         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 671         hdr->next_hdr = IB_GRH_NEXT_HDR;
 672         hdr->hop_limit = grh->hop_limit;
 673         /* The SGID is 32-bit aligned. */
 674         hdr->sgid.global.subnet_prefix = ibp->rvp.gid_prefix;
 675         hdr->sgid.global.interface_id =
 676                 grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ?
 677                 ibp->guids[grh->sgid_index - 1] :
 678                         cpu_to_be64(ppd_from_ibp(ibp)->guid);
 679         hdr->dgid = grh->dgid;
 680
 681         /* GRH header size in 32-bit words. */
 682         return sizeof(struct ib_grh) / sizeof(u32);
 683 }
 684
 685 #define BTH2_OFFSET (offsetof(struct hfi1_pio_header, hdr.u.oth.bth[2]) / 4)
 686
 687 /**
 688  * build_ahg - create ahg in s_hdr
 689  * @qp: a pointer to QP
 690  * @npsn: the next PSN for the request/response
 691  *
 692  * This routine handles the AHG by allocating an ahg entry and causing the
 693  * copy of the first middle.
 694  *
 695  * Subsequent middles use the copied entry, editing the
 696  * PSN with 1 or 2 edits.
 697  */
 698 static inline void build_ahg(struct rvt_qp *qp, u32 npsn)
 699 {
 700         struct hfi1_qp_priv *priv = qp->priv;
 701
 702         if (unlikely(qp->s_flags & RVT_S_AHG_CLEAR))
 703                 clear_ahg(qp);
 704         if (!(qp->s_flags & RVT_S_AHG_VALID)) {
 705                 /* first middle that needs copy  */
 706                 if (qp->s_ahgidx < 0)
 707                         qp->s_ahgidx = sdma_ahg_alloc(priv->s_sde);
 708                 if (qp->s_ahgidx >= 0) {
 709                         qp->s_ahgpsn = npsn;
 710                         priv->s_hdr->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
 711                         /* save to protect a change in another thread */
 712                         priv->s_hdr->sde = priv->s_sde;
 713                         priv->s_hdr->ahgidx = qp->s_ahgidx;
 714                         qp->s_flags |= RVT_S_AHG_VALID;
 715                 }
 716         } else {
 717                 /* subsequent middle after valid */
 718                 if (qp->s_ahgidx >= 0) {
 719                         priv->s_hdr->tx_flags |= SDMA_TXREQ_F_USE_AHG;
 720                         priv->s_hdr->ahgidx = qp->s_ahgidx;
 721                         priv->s_hdr->ahgcount++;
 722                         priv->s_hdr->ahgdesc[0] =
 723                                 sdma_build_ahg_descriptor(
 724                                         (__force u16)cpu_to_be16((u16)npsn),
 725                                         BTH2_OFFSET,
 726                                         16,
 727                                         16);
 728                         if ((npsn & 0xffff0000) !=
 729                                         (qp->s_ahgpsn & 0xffff0000)) {
 730                                 priv->s_hdr->ahgcount++;
 731                                 priv->s_hdr->ahgdesc[1] =
 732                                         sdma_build_ahg_descriptor(
 733                                                 (__force u16)cpu_to_be16(
 734                                                         (u16)(npsn >> 16)),
 735                                                 BTH2_OFFSET,
 736                                                 0,
 737                                                 16);
 738                         }
 739                 }
 740         }
 741 }
 742
 743 void hfi1_make_ruc_header(struct rvt_qp *qp, struct hfi1_other_headers *ohdr,
 744                           u32 bth0, u32 bth2, int middle,
 745                           struct hfi1_pkt_state *ps)
 746 {
 747         struct hfi1_qp_priv *priv = qp->priv;
 748         struct hfi1_ibport *ibp = ps->ibp;
 749         u16 lrh0;
 750         u32 nwords;
 751         u32 extra_bytes;
 752         u32 bth1;
 753
 754         /* Construct the header. */
 755         extra_bytes = -qp->s_cur_size & 3;
 756         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 757         lrh0 = HFI1_LRH_BTH;
 758         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 759                 qp->s_hdrwords += hfi1_make_grh(ibp,
 760                                                 &ps->s_txreq->phdr.hdr.u.l.grh,
 761                                                 &qp->remote_ah_attr.grh,
 762                                                 qp->s_hdrwords, nwords);
 763                 lrh0 = HFI1_LRH_GRH;
 764                 middle = 0;
 765         }
 766         lrh0 |= (priv->s_sc & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
 767         /*
 768          * reset s_hdr/AHG fields
 769          *
 770          * This insures that the ahgentry/ahgcount
 771          * are at a non-AHG default to protect
 772          * build_verbs_tx_desc() from using
 773          * an include ahgidx.
 774          *
 775          * build_ahg() will modify as appropriate
 776          * to use the AHG feature.
 777          */
 778         priv->s_hdr->tx_flags = 0;
 779         priv->s_hdr->ahgcount = 0;
 780         priv->s_hdr->ahgidx = 0;
 781         priv->s_hdr->sde = NULL;
 782         if (qp->s_mig_state == IB_MIG_MIGRATED)
 783                 bth0 |= IB_BTH_MIG_REQ;
 784         else
 785                 middle = 0;
 786         if (middle)
 787                 build_ahg(qp, bth2);
 788         else
 789                 qp->s_flags &= ~RVT_S_AHG_VALID;
 790         ps->s_txreq->phdr.hdr.lrh[0] = cpu_to_be16(lrh0);
 791         ps->s_txreq->phdr.hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 792         ps->s_txreq->phdr.hdr.lrh[2] =
 793                 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 794         ps->s_txreq->phdr.hdr.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
 795                                        qp->remote_ah_attr.src_path_bits);
 796         bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
 797         bth0 |= extra_bytes << 20;
 798         ohdr->bth[0] = cpu_to_be32(bth0);
 799         bth1 = qp->remote_qpn;
 800         if (qp->s_flags & RVT_S_ECN) {
 801                 qp->s_flags &= ~RVT_S_ECN;
 802                 /* we recently received a FECN, so return a BECN */
 803                 bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT);
 804         }
 805         ohdr->bth[1] = cpu_to_be32(bth1);
 806         ohdr->bth[2] = cpu_to_be32(bth2);
 807 }
 808
 809 /* when sending, force a reschedule every one of these periods */
 810 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
 811
 812 void _hfi1_do_send(struct work_struct *work)
 813 {
 814         struct iowait *wait = container_of(work, struct iowait, iowork);
 815         struct rvt_qp *qp = iowait_to_qp(wait);
 816
 817         hfi1_do_send(qp);
 818 }
 819
 820 /**
 821  * hfi1_do_send - perform a send on a QP
 822  * @work: contains a pointer to the QP
 823  *
 824  * Process entries in the send work queue until credit or queue is
 825  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 826  * Otherwise, two threads could send packets out of order.
 827  */
 828 void hfi1_do_send(struct rvt_qp *qp)
 829 {
 830         struct hfi1_pkt_state ps;
 831         struct hfi1_qp_priv *priv = qp->priv;
 832         int (*make_req)(struct rvt_qp *qp, struct hfi1_pkt_state *ps);
 833         unsigned long flags;
 834         unsigned long timeout;
 835         unsigned long timeout_int;
 836         int cpu;
 837
 838         ps.dev = to_idev(qp->ibqp.device);
 839         ps.ibp = to_iport(qp->ibqp.device, qp->port_num);
 840         ps.ppd = ppd_from_ibp(ps.ibp);
 841
 842         switch (qp->ibqp.qp_type) {
 843         case IB_QPT_RC:
 844                 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
 845                                                                 ) - 1)) ==
 846                                  ps.ppd->lid)) {
 847                         ruc_loopback(qp);
 848                         return;
 849                 }
 850                 make_req = hfi1_make_rc_req;
 851                 timeout_int = (qp->timeout_jiffies);
 852                 break;
 853         case IB_QPT_UC:
 854                 if (!loopback && ((qp->remote_ah_attr.dlid & ~((1 << ps.ppd->lmc
 855                                                                 ) - 1)) ==
 856                                  ps.ppd->lid)) {
 857                         ruc_loopback(qp);
 858                         return;
 859                 }
 860                 make_req = hfi1_make_uc_req;
 861                 timeout_int = SEND_RESCHED_TIMEOUT;
 862                 break;
 863         default:
 864                 make_req = hfi1_make_ud_req;
 865                 timeout_int = SEND_RESCHED_TIMEOUT;
 866         }
 867
 868         spin_lock_irqsave(&qp->s_lock, flags);
 869
 870         /* Return if we are already busy processing a work request. */
 871         if (!hfi1_send_ok(qp)) {
 872                 spin_unlock_irqrestore(&qp->s_lock, flags);
 873                 return;
 874         }
 875
 876         qp->s_flags |= RVT_S_BUSY;
 877
 878         timeout = jiffies + (timeout_int) / 8;
 879         cpu = priv->s_sde ? priv->s_sde->cpu :
 880                         cpumask_first(cpumask_of_node(ps.ppd->dd->node));
 881         /* insure a pre-built packet is handled  */
 882         ps.s_txreq = get_waiting_verbs_txreq(qp);
 883         do {
 884                 /* Check for a constructed packet to be sent. */
 885                 if (qp->s_hdrwords != 0) {
 886                         spin_unlock_irqrestore(&qp->s_lock, flags);
 887                         /*
 888                          * If the packet cannot be sent now, return and
 889                          * the send tasklet will be woken up later.
 890                          */
 891                         if (hfi1_verbs_send(qp, &ps))
 892                                 return;
 893                         /* Record that s_hdr is empty. */
 894                         qp->s_hdrwords = 0;
 895                         /* allow other tasks to run */
 896                         if (unlikely(time_after(jiffies, timeout))) {
 897                                 if (workqueue_congested(cpu,
 898                                                         ps.ppd->hfi1_wq)) {
 899                                         spin_lock_irqsave(&qp->s_lock, flags);
 900                                         qp->s_flags &= ~RVT_S_BUSY;
 901                                         hfi1_schedule_send(qp);
 902                                         spin_unlock_irqrestore(&qp->s_lock,
 903                                                                flags);
 904                                         this_cpu_inc(
 905                                                 *ps.ppd->dd->send_schedule);
 906                                         return;
 907                                 }
 908                                 cond_resched();
 909                                 this_cpu_inc(*ps.ppd->dd->send_schedule);
 910                                 timeout = jiffies + (timeout_int) / 8;
 911                         }
 912                         spin_lock_irqsave(&qp->s_lock, flags);
 913                 }
 914         } while (make_req(qp, &ps));
 915
 916         spin_unlock_irqrestore(&qp->s_lock, flags);
 917 }
 918
 919 /*
 920  * This should be called with s_lock held.
 921  */
 922 void hfi1_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
 923                         enum ib_wc_status status)
 924 {
 925         u32 old_last, last;
 926         unsigned i;
 927
 928         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_OR_FLUSH_SEND))
 929                 return;
 930
 931         last = qp->s_last;
 932         old_last = last;
 933         if (++last >= qp->s_size)
 934                 last = 0;
 935         qp->s_last = last;
 936         /* See post_send() */
 937         barrier();
 938         for (i = 0; i < wqe->wr.num_sge; i++) {
 939                 struct rvt_sge *sge = &wqe->sg_list[i];
 940
 941                 rvt_put_mr(sge->mr);
 942         }
 943         if (qp->ibqp.qp_type == IB_QPT_UD ||
 944             qp->ibqp.qp_type == IB_QPT_SMI ||
 945             qp->ibqp.qp_type == IB_QPT_GSI)
 946                 atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
 947
 948         /* See ch. 11.2.4.1 and 10.7.3.1 */
 949         if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
 950             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 951             status != IB_WC_SUCCESS) {
 952                 struct ib_wc wc;
 953
 954                 memset(&wc, 0, sizeof(wc));
 955                 wc.wr_id = wqe->wr.wr_id;
 956                 wc.status = status;
 957                 wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
 958                 wc.qp = &qp->ibqp;
 959                 if (status == IB_WC_SUCCESS)
 960                         wc.byte_len = wqe->length;
 961                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
 962                              status != IB_WC_SUCCESS);
 963         }
 964
 965         if (qp->s_acked == old_last)
 966                 qp->s_acked = last;
 967         if (qp->s_cur == old_last)
 968                 qp->s_cur = last;
 969         if (qp->s_tail == old_last)
 970                 qp->s_tail = last;
 971         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
 972                 qp->s_draining = 0;
 973 }