drivers/staging/rdma/hfi1/ruc.c

   1 /*
   2  *
   3  * This file is provided under a dual BSD/GPLv2 license.  When using or
   4  * redistributing this file, you may do so under either license.
   5  *
   6  * GPL LICENSE SUMMARY
   7  *
   8  * Copyright(c) 2015 Intel Corporation.
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of version 2 of the GNU General Public License as
  12  * published by the Free Software Foundation.
  13  *
  14  * This program is distributed in the hope that it will be useful, but
  15  * WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * General Public License for more details.
  18  *
  19  * BSD LICENSE
  20  *
  21  * Copyright(c) 2015 Intel Corporation.
  22  *
  23  * Redistribution and use in source and binary forms, with or without
  24  * modification, are permitted provided that the following conditions
  25  * are met:
  26  *
  27  *  - Redistributions of source code must retain the above copyright
  28  *    notice, this list of conditions and the following disclaimer.
  29  *  - Redistributions in binary form must reproduce the above copyright
  30  *    notice, this list of conditions and the following disclaimer in
  31  *    the documentation and/or other materials provided with the
  32  *    distribution.
  33  *  - Neither the name of Intel Corporation nor the names of its
  34  *    contributors may be used to endorse or promote products derived
  35  *    from this software without specific prior written permission.
  36  *
  37  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  38  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  39  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  40  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  41  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  42  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  43  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  44  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  45  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  46  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  47  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  48  *
  49  */
  50
  51 #include <linux/spinlock.h>
  52
  53 #include "hfi.h"
  54 #include "mad.h"
  55 #include "qp.h"
  56 #include "sdma.h"
  57
  58 /*
  59  * Convert the AETH RNR timeout code into the number of microseconds.
  60  */
  61 const u32 ib_hfi1_rnr_table[32] = {
  62         655360, /* 00: 655.36 */
  63         10,     /* 01:    .01 */
  64         20,     /* 02     .02 */
  65         30,     /* 03:    .03 */
  66         40,     /* 04:    .04 */
  67         60,     /* 05:    .06 */
  68         80,     /* 06:    .08 */
  69         120,    /* 07:    .12 */
  70         160,    /* 08:    .16 */
  71         240,    /* 09:    .24 */
  72         320,    /* 0A:    .32 */
  73         480,    /* 0B:    .48 */
  74         640,    /* 0C:    .64 */
  75         960,    /* 0D:    .96 */
  76         1280,   /* 0E:   1.28 */
  77         1920,   /* 0F:   1.92 */
  78         2560,   /* 10:   2.56 */
  79         3840,   /* 11:   3.84 */
  80         5120,   /* 12:   5.12 */
  81         7680,   /* 13:   7.68 */
  82         10240,  /* 14:  10.24 */
  83         15360,  /* 15:  15.36 */
  84         20480,  /* 16:  20.48 */
  85         30720,  /* 17:  30.72 */
  86         40960,  /* 18:  40.96 */
  87         61440,  /* 19:  61.44 */
  88         81920,  /* 1A:  81.92 */
  89         122880, /* 1B: 122.88 */
  90         163840, /* 1C: 163.84 */
  91         245760, /* 1D: 245.76 */
  92         327680, /* 1E: 327.68 */
  93         491520  /* 1F: 491.52 */
  94 };
  95
  96 /*
  97  * Validate a RWQE and fill in the SGE state.
  98  * Return 1 if OK.
  99  */
 100 static int init_sge(struct hfi1_qp *qp, struct hfi1_rwqe *wqe)
 101 {
 102         int i, j, ret;
 103         struct ib_wc wc;
 104         struct hfi1_lkey_table *rkt;
 105         struct hfi1_pd *pd;
 106         struct hfi1_sge_state *ss;
 107
 108         rkt = &to_idev(qp->ibqp.device)->lk_table;
 109         pd = to_ipd(qp->ibqp.srq ? qp->ibqp.srq->pd : qp->ibqp.pd);
 110         ss = &qp->r_sge;
 111         ss->sg_list = qp->r_sg_list;
 112         qp->r_len = 0;
 113         for (i = j = 0; i < wqe->num_sge; i++) {
 114                 if (wqe->sg_list[i].length == 0)
 115                         continue;
 116                 /* Check LKEY */
 117                 if (!hfi1_lkey_ok(rkt, pd, j ? &ss->sg_list[j - 1] : &ss->sge,
 118                                   &wqe->sg_list[i], IB_ACCESS_LOCAL_WRITE))
 119                         goto bad_lkey;
 120                 qp->r_len += wqe->sg_list[i].length;
 121                 j++;
 122         }
 123         ss->num_sge = j;
 124         ss->total_len = qp->r_len;
 125         ret = 1;
 126         goto bail;
 127
 128 bad_lkey:
 129         while (j) {
 130                 struct hfi1_sge *sge = --j ? &ss->sg_list[j - 1] : &ss->sge;
 131
 132                 hfi1_put_mr(sge->mr);
 133         }
 134         ss->num_sge = 0;
 135         memset(&wc, 0, sizeof(wc));
 136         wc.wr_id = wqe->wr_id;
 137         wc.status = IB_WC_LOC_PROT_ERR;
 138         wc.opcode = IB_WC_RECV;
 139         wc.qp = &qp->ibqp;
 140         /* Signal solicited completion event. */
 141         hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc, 1);
 142         ret = 0;
 143 bail:
 144         return ret;
 145 }
 146
 147 /**
 148  * hfi1_get_rwqe - copy the next RWQE into the QP's RWQE
 149  * @qp: the QP
 150  * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
 151  *
 152  * Return -1 if there is a local error, 0 if no RWQE is available,
 153  * otherwise return 1.
 154  *
 155  * Can be called from interrupt level.
 156  */
 157 int hfi1_get_rwqe(struct hfi1_qp *qp, int wr_id_only)
 158 {
 159         unsigned long flags;
 160         struct hfi1_rq *rq;
 161         struct hfi1_rwq *wq;
 162         struct hfi1_srq *srq;
 163         struct hfi1_rwqe *wqe;
 164         void (*handler)(struct ib_event *, void *);
 165         u32 tail;
 166         int ret;
 167
 168         if (qp->ibqp.srq) {
 169                 srq = to_isrq(qp->ibqp.srq);
 170                 handler = srq->ibsrq.event_handler;
 171                 rq = &srq->rq;
 172         } else {
 173                 srq = NULL;
 174                 handler = NULL;
 175                 rq = &qp->r_rq;
 176         }
 177
 178         spin_lock_irqsave(&rq->lock, flags);
 179         if (!(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK)) {
 180                 ret = 0;
 181                 goto unlock;
 182         }
 183
 184         wq = rq->wq;
 185         tail = wq->tail;
 186         /* Validate tail before using it since it is user writable. */
 187         if (tail >= rq->size)
 188                 tail = 0;
 189         if (unlikely(tail == wq->head)) {
 190                 ret = 0;
 191                 goto unlock;
 192         }
 193         /* Make sure entry is read after head index is read. */
 194         smp_rmb();
 195         wqe = get_rwqe_ptr(rq, tail);
 196         /*
 197          * Even though we update the tail index in memory, the verbs
 198          * consumer is not supposed to post more entries until a
 199          * completion is generated.
 200          */
 201         if (++tail >= rq->size)
 202                 tail = 0;
 203         wq->tail = tail;
 204         if (!wr_id_only && !init_sge(qp, wqe)) {
 205                 ret = -1;
 206                 goto unlock;
 207         }
 208         qp->r_wr_id = wqe->wr_id;
 209
 210         ret = 1;
 211         set_bit(HFI1_R_WRID_VALID, &qp->r_aflags);
 212         if (handler) {
 213                 u32 n;
 214
 215                 /*
 216                  * Validate head pointer value and compute
 217                  * the number of remaining WQEs.
 218                  */
 219                 n = wq->head;
 220                 if (n >= rq->size)
 221                         n = 0;
 222                 if (n < tail)
 223                         n += rq->size - tail;
 224                 else
 225                         n -= tail;
 226                 if (n < srq->limit) {
 227                         struct ib_event ev;
 228
 229                         srq->limit = 0;
 230                         spin_unlock_irqrestore(&rq->lock, flags);
 231                         ev.device = qp->ibqp.device;
 232                         ev.element.srq = qp->ibqp.srq;
 233                         ev.event = IB_EVENT_SRQ_LIMIT_REACHED;
 234                         handler(&ev, srq->ibsrq.srq_context);
 235                         goto bail;
 236                 }
 237         }
 238 unlock:
 239         spin_unlock_irqrestore(&rq->lock, flags);
 240 bail:
 241         return ret;
 242 }
 243
 244 /*
 245  * Switch to alternate path.
 246  * The QP s_lock should be held and interrupts disabled.
 247  */
 248 void hfi1_migrate_qp(struct hfi1_qp *qp)
 249 {
 250         struct ib_event ev;
 251
 252         qp->s_mig_state = IB_MIG_MIGRATED;
 253         qp->remote_ah_attr = qp->alt_ah_attr;
 254         qp->port_num = qp->alt_ah_attr.port_num;
 255         qp->s_pkey_index = qp->s_alt_pkey_index;
 256         qp->s_flags |= HFI1_S_AHG_CLEAR;
 257
 258         ev.device = qp->ibqp.device;
 259         ev.element.qp = &qp->ibqp;
 260         ev.event = IB_EVENT_PATH_MIG;
 261         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 262 }
 263
 264 static __be64 get_sguid(struct hfi1_ibport *ibp, unsigned index)
 265 {
 266         if (!index) {
 267                 struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 268
 269                 return cpu_to_be64(ppd->guid);
 270         }
 271         return ibp->guids[index - 1];
 272 }
 273
 274 static int gid_ok(union ib_gid *gid, __be64 gid_prefix, __be64 id)
 275 {
 276         return (gid->global.interface_id == id &&
 277                 (gid->global.subnet_prefix == gid_prefix ||
 278                  gid->global.subnet_prefix == IB_DEFAULT_GID_PREFIX));
 279 }
 280
 281 /*
 282  *
 283  * This should be called with the QP r_lock held.
 284  *
 285  * The s_lock will be acquired around the hfi1_migrate_qp() call.
 286  */
 287 int hfi1_ruc_check_hdr(struct hfi1_ibport *ibp, struct hfi1_ib_header *hdr,
 288                        int has_grh, struct hfi1_qp *qp, u32 bth0)
 289 {
 290         __be64 guid;
 291         unsigned long flags;
 292         u8 sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
 293
 294         if (qp->s_mig_state == IB_MIG_ARMED && (bth0 & IB_BTH_MIG_REQ)) {
 295                 if (!has_grh) {
 296                         if (qp->alt_ah_attr.ah_flags & IB_AH_GRH)
 297                                 goto err;
 298                 } else {
 299                         if (!(qp->alt_ah_attr.ah_flags & IB_AH_GRH))
 300                                 goto err;
 301                         guid = get_sguid(ibp, qp->alt_ah_attr.grh.sgid_index);
 302                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
 303                                 goto err;
 304                         if (!gid_ok(&hdr->u.l.grh.sgid,
 305                             qp->alt_ah_attr.grh.dgid.global.subnet_prefix,
 306                             qp->alt_ah_attr.grh.dgid.global.interface_id))
 307                                 goto err;
 308                 }
 309                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
 310                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
 311                         hfi1_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
 312                                        (u16)bth0,
 313                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 314                                        0, qp->ibqp.qp_num,
 315                                        hdr->lrh[3], hdr->lrh[1]);
 316                         goto err;
 317                 }
 318                 /* Validate the SLID. See Ch. 9.6.1.5 and 17.2.8 */
 319                 if (be16_to_cpu(hdr->lrh[3]) != qp->alt_ah_attr.dlid ||
 320                     ppd_from_ibp(ibp)->port != qp->alt_ah_attr.port_num)
 321                         goto err;
 322                 spin_lock_irqsave(&qp->s_lock, flags);
 323                 hfi1_migrate_qp(qp);
 324                 spin_unlock_irqrestore(&qp->s_lock, flags);
 325         } else {
 326                 if (!has_grh) {
 327                         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
 328                                 goto err;
 329                 } else {
 330                         if (!(qp->remote_ah_attr.ah_flags & IB_AH_GRH))
 331                                 goto err;
 332                         guid = get_sguid(ibp,
 333                                          qp->remote_ah_attr.grh.sgid_index);
 334                         if (!gid_ok(&hdr->u.l.grh.dgid, ibp->gid_prefix, guid))
 335                                 goto err;
 336                         if (!gid_ok(&hdr->u.l.grh.sgid,
 337                             qp->remote_ah_attr.grh.dgid.global.subnet_prefix,
 338                             qp->remote_ah_attr.grh.dgid.global.interface_id))
 339                                 goto err;
 340                 }
 341                 if (unlikely(rcv_pkey_check(ppd_from_ibp(ibp), (u16)bth0,
 342                                             sc5, be16_to_cpu(hdr->lrh[3])))) {
 343                         hfi1_bad_pqkey(ibp, IB_NOTICE_TRAP_BAD_PKEY,
 344                                        (u16)bth0,
 345                                        (be16_to_cpu(hdr->lrh[0]) >> 4) & 0xF,
 346                                        0, qp->ibqp.qp_num,
 347                                        hdr->lrh[3], hdr->lrh[1]);
 348                         goto err;
 349                 }
 350                 /* Validate the SLID. See Ch. 9.6.1.5 */
 351                 if (be16_to_cpu(hdr->lrh[3]) != qp->remote_ah_attr.dlid ||
 352                     ppd_from_ibp(ibp)->port != qp->port_num)
 353                         goto err;
 354                 if (qp->s_mig_state == IB_MIG_REARM &&
 355                     !(bth0 & IB_BTH_MIG_REQ))
 356                         qp->s_mig_state = IB_MIG_ARMED;
 357         }
 358
 359         return 0;
 360
 361 err:
 362         return 1;
 363 }
 364
 365 /**
 366  * ruc_loopback - handle UC and RC loopback requests
 367  * @sqp: the sending QP
 368  *
 369  * This is called from hfi1_do_send() to
 370  * forward a WQE addressed to the same HFI.
 371  * Note that although we are single threaded due to the tasklet, we still
 372  * have to protect against post_send().  We don't have to worry about
 373  * receive interrupts since this is a connected protocol and all packets
 374  * will pass through here.
 375  */
 376 static void ruc_loopback(struct hfi1_qp *sqp)
 377 {
 378         struct hfi1_ibport *ibp = to_iport(sqp->ibqp.device, sqp->port_num);
 379         struct hfi1_qp *qp;
 380         struct hfi1_swqe *wqe;
 381         struct hfi1_sge *sge;
 382         unsigned long flags;
 383         struct ib_wc wc;
 384         u64 sdata;
 385         atomic64_t *maddr;
 386         enum ib_wc_status send_status;
 387         int release;
 388         int ret;
 389
 390         rcu_read_lock();
 391
 392         /*
 393          * Note that we check the responder QP state after
 394          * checking the requester's state.
 395          */
 396         qp = hfi1_lookup_qpn(ibp, sqp->remote_qpn);
 397
 398         spin_lock_irqsave(&sqp->s_lock, flags);
 399
 400         /* Return if we are already busy processing a work request. */
 401         if ((sqp->s_flags & (HFI1_S_BUSY | HFI1_S_ANY_WAIT)) ||
 402             !(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_OR_FLUSH_SEND))
 403                 goto unlock;
 404
 405         sqp->s_flags |= HFI1_S_BUSY;
 406
 407 again:
 408         if (sqp->s_last == sqp->s_head)
 409                 goto clr_busy;
 410         wqe = get_swqe_ptr(sqp, sqp->s_last);
 411
 412         /* Return if it is not OK to start a new work request. */
 413         if (!(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_NEXT_SEND_OK)) {
 414                 if (!(ib_hfi1_state_ops[sqp->state] & HFI1_FLUSH_SEND))
 415                         goto clr_busy;
 416                 /* We are in the error state, flush the work request. */
 417                 send_status = IB_WC_WR_FLUSH_ERR;
 418                 goto flush_send;
 419         }
 420
 421         /*
 422          * We can rely on the entry not changing without the s_lock
 423          * being held until we update s_last.
 424          * We increment s_cur to indicate s_last is in progress.
 425          */
 426         if (sqp->s_last == sqp->s_cur) {
 427                 if (++sqp->s_cur >= sqp->s_size)
 428                         sqp->s_cur = 0;
 429         }
 430         spin_unlock_irqrestore(&sqp->s_lock, flags);
 431
 432         if (!qp || !(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_RECV_OK) ||
 433             qp->ibqp.qp_type != sqp->ibqp.qp_type) {
 434                 ibp->n_pkt_drops++;
 435                 /*
 436                  * For RC, the requester would timeout and retry so
 437                  * shortcut the timeouts and just signal too many retries.
 438                  */
 439                 if (sqp->ibqp.qp_type == IB_QPT_RC)
 440                         send_status = IB_WC_RETRY_EXC_ERR;
 441                 else
 442                         send_status = IB_WC_SUCCESS;
 443                 goto serr;
 444         }
 445
 446         memset(&wc, 0, sizeof(wc));
 447         send_status = IB_WC_SUCCESS;
 448
 449         release = 1;
 450         sqp->s_sge.sge = wqe->sg_list[0];
 451         sqp->s_sge.sg_list = wqe->sg_list + 1;
 452         sqp->s_sge.num_sge = wqe->wr.num_sge;
 453         sqp->s_len = wqe->length;
 454         switch (wqe->wr.opcode) {
 455         case IB_WR_SEND_WITH_IMM:
 456                 wc.wc_flags = IB_WC_WITH_IMM;
 457                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 458                 /* FALLTHROUGH */
 459         case IB_WR_SEND:
 460                 ret = hfi1_get_rwqe(qp, 0);
 461                 if (ret < 0)
 462                         goto op_err;
 463                 if (!ret)
 464                         goto rnr_nak;
 465                 break;
 466
 467         case IB_WR_RDMA_WRITE_WITH_IMM:
 468                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 469                         goto inv_err;
 470                 wc.wc_flags = IB_WC_WITH_IMM;
 471                 wc.ex.imm_data = wqe->wr.ex.imm_data;
 472                 ret = hfi1_get_rwqe(qp, 1);
 473                 if (ret < 0)
 474                         goto op_err;
 475                 if (!ret)
 476                         goto rnr_nak;
 477                 /* FALLTHROUGH */
 478         case IB_WR_RDMA_WRITE:
 479                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE)))
 480                         goto inv_err;
 481                 if (wqe->length == 0)
 482                         break;
 483                 if (unlikely(!hfi1_rkey_ok(qp, &qp->r_sge.sge, wqe->length,
 484                                            wqe->rdma_wr.remote_addr,
 485                                            wqe->rdma_wr.rkey,
 486                                            IB_ACCESS_REMOTE_WRITE)))
 487                         goto acc_err;
 488                 qp->r_sge.sg_list = NULL;
 489                 qp->r_sge.num_sge = 1;
 490                 qp->r_sge.total_len = wqe->length;
 491                 break;
 492
 493         case IB_WR_RDMA_READ:
 494                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ)))
 495                         goto inv_err;
 496                 if (unlikely(!hfi1_rkey_ok(qp, &sqp->s_sge.sge, wqe->length,
 497                                            wqe->rdma_wr.remote_addr,
 498                                            wqe->rdma_wr.rkey,
 499                                            IB_ACCESS_REMOTE_READ)))
 500                         goto acc_err;
 501                 release = 0;
 502                 sqp->s_sge.sg_list = NULL;
 503                 sqp->s_sge.num_sge = 1;
 504                 qp->r_sge.sge = wqe->sg_list[0];
 505                 qp->r_sge.sg_list = wqe->sg_list + 1;
 506                 qp->r_sge.num_sge = wqe->wr.num_sge;
 507                 qp->r_sge.total_len = wqe->length;
 508                 break;
 509
 510         case IB_WR_ATOMIC_CMP_AND_SWP:
 511         case IB_WR_ATOMIC_FETCH_AND_ADD:
 512                 if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_ATOMIC)))
 513                         goto inv_err;
 514                 if (unlikely(!hfi1_rkey_ok(qp, &qp->r_sge.sge, sizeof(u64),
 515                                            wqe->atomic_wr.remote_addr,
 516                                            wqe->atomic_wr.rkey,
 517                                            IB_ACCESS_REMOTE_ATOMIC)))
 518                         goto acc_err;
 519                 /* Perform atomic OP and save result. */
 520                 maddr = (atomic64_t *) qp->r_sge.sge.vaddr;
 521                 sdata = wqe->atomic_wr.compare_add;
 522                 *(u64 *) sqp->s_sge.sge.vaddr =
 523                         (wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) ?
 524                         (u64) atomic64_add_return(sdata, maddr) - sdata :
 525                         (u64) cmpxchg((u64 *) qp->r_sge.sge.vaddr,
 526                                       sdata, wqe->atomic_wr.swap);
 527                 hfi1_put_mr(qp->r_sge.sge.mr);
 528                 qp->r_sge.num_sge = 0;
 529                 goto send_comp;
 530
 531         default:
 532                 send_status = IB_WC_LOC_QP_OP_ERR;
 533                 goto serr;
 534         }
 535
 536         sge = &sqp->s_sge.sge;
 537         while (sqp->s_len) {
 538                 u32 len = sqp->s_len;
 539
 540                 if (len > sge->length)
 541                         len = sge->length;
 542                 if (len > sge->sge_length)
 543                         len = sge->sge_length;
 544                 WARN_ON_ONCE(len == 0);
 545                 hfi1_copy_sge(&qp->r_sge, sge->vaddr, len, release);
 546                 sge->vaddr += len;
 547                 sge->length -= len;
 548                 sge->sge_length -= len;
 549                 if (sge->sge_length == 0) {
 550                         if (!release)
 551                                 hfi1_put_mr(sge->mr);
 552                         if (--sqp->s_sge.num_sge)
 553                                 *sge = *sqp->s_sge.sg_list++;
 554                 } else if (sge->length == 0 && sge->mr->lkey) {
 555                         if (++sge->n >= HFI1_SEGSZ) {
 556                                 if (++sge->m >= sge->mr->mapsz)
 557                                         break;
 558                                 sge->n = 0;
 559                         }
 560                         sge->vaddr =
 561                                 sge->mr->map[sge->m]->segs[sge->n].vaddr;
 562                         sge->length =
 563                                 sge->mr->map[sge->m]->segs[sge->n].length;
 564                 }
 565                 sqp->s_len -= len;
 566         }
 567         if (release)
 568                 hfi1_put_ss(&qp->r_sge);
 569
 570         if (!test_and_clear_bit(HFI1_R_WRID_VALID, &qp->r_aflags))
 571                 goto send_comp;
 572
 573         if (wqe->wr.opcode == IB_WR_RDMA_WRITE_WITH_IMM)
 574                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 575         else
 576                 wc.opcode = IB_WC_RECV;
 577         wc.wr_id = qp->r_wr_id;
 578         wc.status = IB_WC_SUCCESS;
 579         wc.byte_len = wqe->length;
 580         wc.qp = &qp->ibqp;
 581         wc.src_qp = qp->remote_qpn;
 582         wc.slid = qp->remote_ah_attr.dlid;
 583         wc.sl = qp->remote_ah_attr.sl;
 584         wc.port_num = 1;
 585         /* Signal completion event if the solicited bit is set. */
 586         hfi1_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 587                       wqe->wr.send_flags & IB_SEND_SOLICITED);
 588
 589 send_comp:
 590         spin_lock_irqsave(&sqp->s_lock, flags);
 591         ibp->n_loop_pkts++;
 592 flush_send:
 593         sqp->s_rnr_retry = sqp->s_rnr_retry_cnt;
 594         hfi1_send_complete(sqp, wqe, send_status);
 595         goto again;
 596
 597 rnr_nak:
 598         /* Handle RNR NAK */
 599         if (qp->ibqp.qp_type == IB_QPT_UC)
 600                 goto send_comp;
 601         ibp->n_rnr_naks++;
 602         /*
 603          * Note: we don't need the s_lock held since the BUSY flag
 604          * makes this single threaded.
 605          */
 606         if (sqp->s_rnr_retry == 0) {
 607                 send_status = IB_WC_RNR_RETRY_EXC_ERR;
 608                 goto serr;
 609         }
 610         if (sqp->s_rnr_retry_cnt < 7)
 611                 sqp->s_rnr_retry--;
 612         spin_lock_irqsave(&sqp->s_lock, flags);
 613         if (!(ib_hfi1_state_ops[sqp->state] & HFI1_PROCESS_RECV_OK))
 614                 goto clr_busy;
 615         sqp->s_flags |= HFI1_S_WAIT_RNR;
 616         sqp->s_timer.function = hfi1_rc_rnr_retry;
 617         sqp->s_timer.expires = jiffies +
 618                 usecs_to_jiffies(ib_hfi1_rnr_table[qp->r_min_rnr_timer]);
 619         add_timer(&sqp->s_timer);
 620         goto clr_busy;
 621
 622 op_err:
 623         send_status = IB_WC_REM_OP_ERR;
 624         wc.status = IB_WC_LOC_QP_OP_ERR;
 625         goto err;
 626
 627 inv_err:
 628         send_status = IB_WC_REM_INV_REQ_ERR;
 629         wc.status = IB_WC_LOC_QP_OP_ERR;
 630         goto err;
 631
 632 acc_err:
 633         send_status = IB_WC_REM_ACCESS_ERR;
 634         wc.status = IB_WC_LOC_PROT_ERR;
 635 err:
 636         /* responder goes to error state */
 637         hfi1_rc_error(qp, wc.status);
 638
 639 serr:
 640         spin_lock_irqsave(&sqp->s_lock, flags);
 641         hfi1_send_complete(sqp, wqe, send_status);
 642         if (sqp->ibqp.qp_type == IB_QPT_RC) {
 643                 int lastwqe = hfi1_error_qp(sqp, IB_WC_WR_FLUSH_ERR);
 644
 645                 sqp->s_flags &= ~HFI1_S_BUSY;
 646                 spin_unlock_irqrestore(&sqp->s_lock, flags);
 647                 if (lastwqe) {
 648                         struct ib_event ev;
 649
 650                         ev.device = sqp->ibqp.device;
 651                         ev.element.qp = &sqp->ibqp;
 652                         ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 653                         sqp->ibqp.event_handler(&ev, sqp->ibqp.qp_context);
 654                 }
 655                 goto done;
 656         }
 657 clr_busy:
 658         sqp->s_flags &= ~HFI1_S_BUSY;
 659 unlock:
 660         spin_unlock_irqrestore(&sqp->s_lock, flags);
 661 done:
 662         rcu_read_unlock();
 663 }
 664
 665 /**
 666  * hfi1_make_grh - construct a GRH header
 667  * @ibp: a pointer to the IB port
 668  * @hdr: a pointer to the GRH header being constructed
 669  * @grh: the global route address to send to
 670  * @hwords: the number of 32 bit words of header being sent
 671  * @nwords: the number of 32 bit words of data being sent
 672  *
 673  * Return the size of the header in 32 bit words.
 674  */
 675 u32 hfi1_make_grh(struct hfi1_ibport *ibp, struct ib_grh *hdr,
 676                   struct ib_global_route *grh, u32 hwords, u32 nwords)
 677 {
 678         hdr->version_tclass_flow =
 679                 cpu_to_be32((IB_GRH_VERSION << IB_GRH_VERSION_SHIFT) |
 680                             (grh->traffic_class << IB_GRH_TCLASS_SHIFT) |
 681                             (grh->flow_label << IB_GRH_FLOW_SHIFT));
 682         hdr->paylen = cpu_to_be16((hwords - 2 + nwords + SIZE_OF_CRC) << 2);
 683         /* next_hdr is defined by C8-7 in ch. 8.4.1 */
 684         hdr->next_hdr = IB_GRH_NEXT_HDR;
 685         hdr->hop_limit = grh->hop_limit;
 686         /* The SGID is 32-bit aligned. */
 687         hdr->sgid.global.subnet_prefix = ibp->gid_prefix;
 688         hdr->sgid.global.interface_id =
 689                 grh->sgid_index && grh->sgid_index < ARRAY_SIZE(ibp->guids) ?
 690                 ibp->guids[grh->sgid_index - 1] :
 691                         cpu_to_be64(ppd_from_ibp(ibp)->guid);
 692         hdr->dgid = grh->dgid;
 693
 694         /* GRH header size in 32-bit words. */
 695         return sizeof(struct ib_grh) / sizeof(u32);
 696 }
 697
 698 #define BTH2_OFFSET (offsetof(struct hfi1_pio_header, hdr.u.oth.bth[2]) / 4)
 699
 700 /**
 701  * build_ahg - create ahg in s_hdr
 702  * @qp: a pointer to QP
 703  * @npsn: the next PSN for the request/response
 704  *
 705  * This routine handles the AHG by allocating an ahg entry and causing the
 706  * copy of the first middle.
 707  *
 708  * Subsequent middles use the copied entry, editing the
 709  * PSN with 1 or 2 edits.
 710  */
 711 static inline void build_ahg(struct hfi1_qp *qp, u32 npsn)
 712 {
 713         if (unlikely(qp->s_flags & HFI1_S_AHG_CLEAR))
 714                 clear_ahg(qp);
 715         if (!(qp->s_flags & HFI1_S_AHG_VALID)) {
 716                 /* first middle that needs copy  */
 717                 if (qp->s_ahgidx < 0) {
 718                         if (!qp->s_sde)
 719                                 qp->s_sde = qp_to_sdma_engine(qp, qp->s_sc);
 720                         qp->s_ahgidx = sdma_ahg_alloc(qp->s_sde);
 721                 }
 722                 if (qp->s_ahgidx >= 0) {
 723                         qp->s_ahgpsn = npsn;
 724                         qp->s_hdr->tx_flags |= SDMA_TXREQ_F_AHG_COPY;
 725                         /* save to protect a change in another thread */
 726                         qp->s_hdr->sde = qp->s_sde;
 727                         qp->s_hdr->ahgidx = qp->s_ahgidx;
 728                         qp->s_flags |= HFI1_S_AHG_VALID;
 729                 }
 730         } else {
 731                 /* subsequent middle after valid */
 732                 if (qp->s_ahgidx >= 0) {
 733                         qp->s_hdr->tx_flags |= SDMA_TXREQ_F_USE_AHG;
 734                         qp->s_hdr->ahgidx = qp->s_ahgidx;
 735                         qp->s_hdr->ahgcount++;
 736                         qp->s_hdr->ahgdesc[0] =
 737                                 sdma_build_ahg_descriptor(
 738                                         (__force u16)cpu_to_be16((u16)npsn),
 739                                         BTH2_OFFSET,
 740                                         16,
 741                                         16);
 742                         if ((npsn & 0xffff0000) !=
 743                                         (qp->s_ahgpsn & 0xffff0000)) {
 744                                 qp->s_hdr->ahgcount++;
 745                                 qp->s_hdr->ahgdesc[1] =
 746                                         sdma_build_ahg_descriptor(
 747                                                 (__force u16)cpu_to_be16(
 748                                                         (u16)(npsn >> 16)),
 749                                                 BTH2_OFFSET,
 750                                                 0,
 751                                                 16);
 752                         }
 753                 }
 754         }
 755 }
 756
 757 void hfi1_make_ruc_header(struct hfi1_qp *qp, struct hfi1_other_headers *ohdr,
 758                           u32 bth0, u32 bth2, int middle)
 759 {
 760         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 761         u16 lrh0;
 762         u32 nwords;
 763         u32 extra_bytes;
 764         u8 sc5;
 765         u32 bth1;
 766
 767         /* Construct the header. */
 768         extra_bytes = -qp->s_cur_size & 3;
 769         nwords = (qp->s_cur_size + extra_bytes) >> 2;
 770         lrh0 = HFI1_LRH_BTH;
 771         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 772                 qp->s_hdrwords += hfi1_make_grh(ibp, &qp->s_hdr->ibh.u.l.grh,
 773                                                &qp->remote_ah_attr.grh,
 774                                                qp->s_hdrwords, nwords);
 775                 lrh0 = HFI1_LRH_GRH;
 776                 middle = 0;
 777         }
 778         sc5 = ibp->sl_to_sc[qp->remote_ah_attr.sl];
 779         lrh0 |= (sc5 & 0xf) << 12 | (qp->remote_ah_attr.sl & 0xf) << 4;
 780         qp->s_sc = sc5;
 781         /*
 782          * reset s_hdr/AHG fields
 783          *
 784          * This insures that the ahgentry/ahgcount
 785          * are at a non-AHG default to protect
 786          * build_verbs_tx_desc() from using
 787          * an include ahgidx.
 788          *
 789          * build_ahg() will modify as appropriate
 790          * to use the AHG feature.
 791          */
 792         qp->s_hdr->tx_flags = 0;
 793         qp->s_hdr->ahgcount = 0;
 794         qp->s_hdr->ahgidx = 0;
 795         qp->s_hdr->sde = NULL;
 796         if (qp->s_mig_state == IB_MIG_MIGRATED)
 797                 bth0 |= IB_BTH_MIG_REQ;
 798         else
 799                 middle = 0;
 800         if (middle)
 801                 build_ahg(qp, bth2);
 802         else
 803                 qp->s_flags &= ~HFI1_S_AHG_VALID;
 804         qp->s_hdr->ibh.lrh[0] = cpu_to_be16(lrh0);
 805         qp->s_hdr->ibh.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 806         qp->s_hdr->ibh.lrh[2] =
 807                 cpu_to_be16(qp->s_hdrwords + nwords + SIZE_OF_CRC);
 808         qp->s_hdr->ibh.lrh[3] = cpu_to_be16(ppd_from_ibp(ibp)->lid |
 809                                        qp->remote_ah_attr.src_path_bits);
 810         bth0 |= hfi1_get_pkey(ibp, qp->s_pkey_index);
 811         bth0 |= extra_bytes << 20;
 812         ohdr->bth[0] = cpu_to_be32(bth0);
 813         bth1 = qp->remote_qpn;
 814         if (qp->s_flags & HFI1_S_ECN) {
 815                 qp->s_flags &= ~HFI1_S_ECN;
 816                 /* we recently received a FECN, so return a BECN */
 817                 bth1 |= (HFI1_BECN_MASK << HFI1_BECN_SHIFT);
 818         }
 819         ohdr->bth[1] = cpu_to_be32(bth1);
 820         ohdr->bth[2] = cpu_to_be32(bth2);
 821 }
 822
 823 /* when sending, force a reschedule every one of these periods */
 824 #define SEND_RESCHED_TIMEOUT (5 * HZ)  /* 5s in jiffies */
 825
 826 /**
 827  * hfi1_do_send - perform a send on a QP
 828  * @work: contains a pointer to the QP
 829  *
 830  * Process entries in the send work queue until credit or queue is
 831  * exhausted.  Only allow one CPU to send a packet per QP (tasklet).
 832  * Otherwise, two threads could send packets out of order.
 833  */
 834 void hfi1_do_send(struct work_struct *work)
 835 {
 836         struct iowait *wait = container_of(work, struct iowait, iowork);
 837         struct hfi1_qp *qp = container_of(wait, struct hfi1_qp, s_iowait);
 838         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 839         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 840         int (*make_req)(struct hfi1_qp *qp);
 841         unsigned long flags;
 842         unsigned long timeout;
 843
 844         if ((qp->ibqp.qp_type == IB_QPT_RC ||
 845              qp->ibqp.qp_type == IB_QPT_UC) &&
 846             !loopback &&
 847             (qp->remote_ah_attr.dlid & ~((1 << ppd->lmc) - 1)) == ppd->lid) {
 848                 ruc_loopback(qp);
 849                 return;
 850         }
 851
 852         if (qp->ibqp.qp_type == IB_QPT_RC)
 853                 make_req = hfi1_make_rc_req;
 854         else if (qp->ibqp.qp_type == IB_QPT_UC)
 855                 make_req = hfi1_make_uc_req;
 856         else
 857                 make_req = hfi1_make_ud_req;
 858
 859         spin_lock_irqsave(&qp->s_lock, flags);
 860
 861         /* Return if we are already busy processing a work request. */
 862         if (!hfi1_send_ok(qp)) {
 863                 spin_unlock_irqrestore(&qp->s_lock, flags);
 864                 return;
 865         }
 866
 867         qp->s_flags |= HFI1_S_BUSY;
 868
 869         spin_unlock_irqrestore(&qp->s_lock, flags);
 870
 871         timeout = jiffies + SEND_RESCHED_TIMEOUT;
 872         do {
 873                 /* Check for a constructed packet to be sent. */
 874                 if (qp->s_hdrwords != 0) {
 875                         /*
 876                          * If the packet cannot be sent now, return and
 877                          * the send tasklet will be woken up later.
 878                          */
 879                         if (hfi1_verbs_send(qp, qp->s_hdr, qp->s_hdrwords,
 880                                             qp->s_cur_sge, qp->s_cur_size))
 881                                 break;
 882                         /* Record that s_hdr is empty. */
 883                         qp->s_hdrwords = 0;
 884                 }
 885
 886                 /* allow other tasks to run */
 887                 if (unlikely(time_after(jiffies, timeout))) {
 888                         cond_resched();
 889                         ppd->dd->verbs_dev.n_send_schedule++;
 890                         timeout = jiffies + SEND_RESCHED_TIMEOUT;
 891                 }
 892         } while (make_req(qp));
 893 }
 894
 895 /*
 896  * This should be called with s_lock held.
 897  */
 898 void hfi1_send_complete(struct hfi1_qp *qp, struct hfi1_swqe *wqe,
 899                         enum ib_wc_status status)
 900 {
 901         u32 old_last, last;
 902         unsigned i;
 903
 904         if (!(ib_hfi1_state_ops[qp->state] & HFI1_PROCESS_OR_FLUSH_SEND))
 905                 return;
 906
 907         for (i = 0; i < wqe->wr.num_sge; i++) {
 908                 struct hfi1_sge *sge = &wqe->sg_list[i];
 909
 910                 hfi1_put_mr(sge->mr);
 911         }
 912         if (qp->ibqp.qp_type == IB_QPT_UD ||
 913             qp->ibqp.qp_type == IB_QPT_SMI ||
 914             qp->ibqp.qp_type == IB_QPT_GSI)
 915                 atomic_dec(&to_iah(wqe->ud_wr.ah)->refcount);
 916
 917         /* See ch. 11.2.4.1 and 10.7.3.1 */
 918         if (!(qp->s_flags & HFI1_S_SIGNAL_REQ_WR) ||
 919             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 920             status != IB_WC_SUCCESS) {
 921                 struct ib_wc wc;
 922
 923                 memset(&wc, 0, sizeof(wc));
 924                 wc.wr_id = wqe->wr.wr_id;
 925                 wc.status = status;
 926                 wc.opcode = ib_hfi1_wc_opcode[wqe->wr.opcode];
 927                 wc.qp = &qp->ibqp;
 928                 if (status == IB_WC_SUCCESS)
 929                         wc.byte_len = wqe->length;
 930                 hfi1_cq_enter(to_icq(qp->ibqp.send_cq), &wc,
 931                               status != IB_WC_SUCCESS);
 932         }
 933
 934         last = qp->s_last;
 935         old_last = last;
 936         if (++last >= qp->s_size)
 937                 last = 0;
 938         qp->s_last = last;
 939         if (qp->s_acked == old_last)
 940                 qp->s_acked = last;
 941         if (qp->s_cur == old_last)
 942                 qp->s_cur = last;
 943         if (qp->s_tail == old_last)
 944                 qp->s_tail = last;
 945         if (qp->state == IB_QPS_SQD && last == qp->s_cur)
 946                 qp->s_draining = 0;
 947 }