drivers/infiniband/hw/hfi1/qp.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/err.h>
  49 #include <linux/vmalloc.h>
  50 #include <linux/hash.h>
  51 #include <linux/module.h>
  52 #include <linux/seq_file.h>
  53 #include <rdma/rdma_vt.h>
  54 #include <rdma/rdmavt_qp.h>
  55 #include <rdma/ib_verbs.h>
  56
  57 #include "hfi.h"
  58 #include "qp.h"
  59 #include "trace.h"
  60 #include "verbs_txreq.h"
  61
  62 unsigned int hfi1_qp_table_size = 256;
  63 module_param_named(qp_table_size, hfi1_qp_table_size, uint, S_IRUGO);
  64 MODULE_PARM_DESC(qp_table_size, "QP table size");
  65
  66 static void flush_tx_list(struct rvt_qp *qp);
  67 static int iowait_sleep(
  68         struct sdma_engine *sde,
  69         struct iowait *wait,
  70         struct sdma_txreq *stx,
  71         unsigned seq);
  72 static void iowait_wakeup(struct iowait *wait, int reason);
  73 static void iowait_sdma_drained(struct iowait *wait);
  74 static void qp_pio_drain(struct rvt_qp *qp);
  75
  76 static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
  77                               struct rvt_qpn_map *map, unsigned off)
  78 {
  79         return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
  80 }
  81
  82 /*
  83  * Convert the AETH credit code into the number of credits.
  84  */
  85 static const u16 credit_table[31] = {
  86         0,                      /* 0 */
  87         1,                      /* 1 */
  88         2,                      /* 2 */
  89         3,                      /* 3 */
  90         4,                      /* 4 */
  91         6,                      /* 5 */
  92         8,                      /* 6 */
  93         12,                     /* 7 */
  94         16,                     /* 8 */
  95         24,                     /* 9 */
  96         32,                     /* A */
  97         48,                     /* B */
  98         64,                     /* C */
  99         96,                     /* D */
 100         128,                    /* E */
 101         192,                    /* F */
 102         256,                    /* 10 */
 103         384,                    /* 11 */
 104         512,                    /* 12 */
 105         768,                    /* 13 */
 106         1024,                   /* 14 */
 107         1536,                   /* 15 */
 108         2048,                   /* 16 */
 109         3072,                   /* 17 */
 110         4096,                   /* 18 */
 111         6144,                   /* 19 */
 112         8192,                   /* 1A */
 113         12288,                  /* 1B */
 114         16384,                  /* 1C */
 115         24576,                  /* 1D */
 116         32768                   /* 1E */
 117 };
 118
 119 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
 120 [IB_WR_RDMA_WRITE] = {
 121         .length = sizeof(struct ib_rdma_wr),
 122         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 123 },
 124
 125 [IB_WR_RDMA_READ] = {
 126         .length = sizeof(struct ib_rdma_wr),
 127         .qpt_support = BIT(IB_QPT_RC),
 128         .flags = RVT_OPERATION_ATOMIC,
 129 },
 130
 131 [IB_WR_ATOMIC_CMP_AND_SWP] = {
 132         .length = sizeof(struct ib_atomic_wr),
 133         .qpt_support = BIT(IB_QPT_RC),
 134         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 135 },
 136
 137 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
 138         .length = sizeof(struct ib_atomic_wr),
 139         .qpt_support = BIT(IB_QPT_RC),
 140         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
 141 },
 142
 143 [IB_WR_RDMA_WRITE_WITH_IMM] = {
 144         .length = sizeof(struct ib_rdma_wr),
 145         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 146 },
 147
 148 [IB_WR_SEND] = {
 149         .length = sizeof(struct ib_send_wr),
 150         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 151                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 152 },
 153
 154 [IB_WR_SEND_WITH_IMM] = {
 155         .length = sizeof(struct ib_send_wr),
 156         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 157                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 158 },
 159
 160 [IB_WR_REG_MR] = {
 161         .length = sizeof(struct ib_reg_wr),
 162         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 163         .flags = RVT_OPERATION_LOCAL,
 164 },
 165
 166 [IB_WR_LOCAL_INV] = {
 167         .length = sizeof(struct ib_send_wr),
 168         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 169         .flags = RVT_OPERATION_LOCAL,
 170 },
 171
 172 [IB_WR_SEND_WITH_INV] = {
 173         .length = sizeof(struct ib_send_wr),
 174         .qpt_support = BIT(IB_QPT_RC),
 175 },
 176
 177 };
 178
 179 static void flush_tx_list(struct rvt_qp *qp)
 180 {
 181         struct hfi1_qp_priv *priv = qp->priv;
 182
 183         while (!list_empty(&priv->s_iowait.tx_head)) {
 184                 struct sdma_txreq *tx;
 185
 186                 tx = list_first_entry(
 187                         &priv->s_iowait.tx_head,
 188                         struct sdma_txreq,
 189                         list);
 190                 list_del_init(&tx->list);
 191                 hfi1_put_txreq(
 192                         container_of(tx, struct verbs_txreq, txreq));
 193         }
 194 }
 195
 196 static void flush_iowait(struct rvt_qp *qp)
 197 {
 198         struct hfi1_qp_priv *priv = qp->priv;
 199         struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 200         unsigned long flags;
 201
 202         write_seqlock_irqsave(&dev->iowait_lock, flags);
 203         if (!list_empty(&priv->s_iowait.list)) {
 204                 list_del_init(&priv->s_iowait.list);
 205                 if (atomic_dec_and_test(&qp->refcount))
 206                         wake_up(&qp->wait);
 207         }
 208         write_sequnlock_irqrestore(&dev->iowait_lock, flags);
 209 }
 210
 211 static inline int opa_mtu_enum_to_int(int mtu)
 212 {
 213         switch (mtu) {
 214         case OPA_MTU_8192:  return 8192;
 215         case OPA_MTU_10240: return 10240;
 216         default:            return -1;
 217         }
 218 }
 219
 220 /**
 221  * This function is what we would push to the core layer if we wanted to be a
 222  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
 223  * to blindly pass the MTU enum value from the PathRecord to us.
 224  */
 225 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
 226 {
 227         int val;
 228
 229         /* Constraining 10KB packets to 8KB packets */
 230         if (mtu == (enum ib_mtu)OPA_MTU_10240)
 231                 mtu = OPA_MTU_8192;
 232         val = opa_mtu_enum_to_int((int)mtu);
 233         if (val > 0)
 234                 return val;
 235         return ib_mtu_enum_to_int(mtu);
 236 }
 237
 238 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
 239                          int attr_mask, struct ib_udata *udata)
 240 {
 241         struct ib_qp *ibqp = &qp->ibqp;
 242         struct hfi1_ibdev *dev = to_idev(ibqp->device);
 243         struct hfi1_devdata *dd = dd_from_dev(dev);
 244         u8 sc;
 245
 246         if (attr_mask & IB_QP_AV) {
 247                 sc = ah_to_sc(ibqp->device, &attr->ah_attr);
 248                 if (sc == 0xf)
 249                         return -EINVAL;
 250
 251                 if (!qp_to_sdma_engine(qp, sc) &&
 252                     dd->flags & HFI1_HAS_SEND_DMA)
 253                         return -EINVAL;
 254
 255                 if (!qp_to_send_context(qp, sc))
 256                         return -EINVAL;
 257         }
 258
 259         if (attr_mask & IB_QP_ALT_PATH) {
 260                 sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
 261                 if (sc == 0xf)
 262                         return -EINVAL;
 263
 264                 if (!qp_to_sdma_engine(qp, sc) &&
 265                     dd->flags & HFI1_HAS_SEND_DMA)
 266                         return -EINVAL;
 267
 268                 if (!qp_to_send_context(qp, sc))
 269                         return -EINVAL;
 270         }
 271
 272         return 0;
 273 }
 274
 275 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
 276                     int attr_mask, struct ib_udata *udata)
 277 {
 278         struct ib_qp *ibqp = &qp->ibqp;
 279         struct hfi1_qp_priv *priv = qp->priv;
 280
 281         if (attr_mask & IB_QP_AV) {
 282                 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
 283                 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 284                 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 285         }
 286
 287         if (attr_mask & IB_QP_PATH_MIG_STATE &&
 288             attr->path_mig_state == IB_MIG_MIGRATED &&
 289             qp->s_mig_state == IB_MIG_ARMED) {
 290                 qp->s_flags |= RVT_S_AHG_CLEAR;
 291                 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
 292                 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 293                 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 294         }
 295 }
 296
 297 /**
 298  * hfi1_check_send_wqe - validate wqe
 299  * @qp - The qp
 300  * @wqe - The built wqe
 301  *
 302  * validate wqe.  This is called
 303  * prior to inserting the wqe into
 304  * the ring but after the wqe has been
 305  * setup.
 306  *
 307  * Returns 0 on success, -EINVAL on failure
 308  *
 309  */
 310 int hfi1_check_send_wqe(struct rvt_qp *qp,
 311                         struct rvt_swqe *wqe)
 312 {
 313         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 314         struct rvt_ah *ah;
 315
 316         switch (qp->ibqp.qp_type) {
 317         case IB_QPT_RC:
 318         case IB_QPT_UC:
 319                 if (wqe->length > 0x80000000U)
 320                         return -EINVAL;
 321                 break;
 322         case IB_QPT_SMI:
 323                 ah = ibah_to_rvtah(wqe->ud_wr.ah);
 324                 if (wqe->length > (1 << ah->log_pmtu))
 325                         return -EINVAL;
 326                 break;
 327         case IB_QPT_GSI:
 328         case IB_QPT_UD:
 329                 ah = ibah_to_rvtah(wqe->ud_wr.ah);
 330                 if (wqe->length > (1 << ah->log_pmtu))
 331                         return -EINVAL;
 332                 if (ibp->sl_to_sc[ah->attr.sl] == 0xf)
 333                         return -EINVAL;
 334         default:
 335                 break;
 336         }
 337         return wqe->length <= piothreshold;
 338 }
 339
 340 /**
 341  * hfi1_compute_aeth - compute the AETH (syndrome + MSN)
 342  * @qp: the queue pair to compute the AETH for
 343  *
 344  * Returns the AETH.
 345  */
 346 __be32 hfi1_compute_aeth(struct rvt_qp *qp)
 347 {
 348         u32 aeth = qp->r_msn & HFI1_MSN_MASK;
 349
 350         if (qp->ibqp.srq) {
 351                 /*
 352                  * Shared receive queues don't generate credits.
 353                  * Set the credit field to the invalid value.
 354                  */
 355                 aeth |= HFI1_AETH_CREDIT_INVAL << HFI1_AETH_CREDIT_SHIFT;
 356         } else {
 357                 u32 min, max, x;
 358                 u32 credits;
 359                 struct rvt_rwq *wq = qp->r_rq.wq;
 360                 u32 head;
 361                 u32 tail;
 362
 363                 /* sanity check pointers before trusting them */
 364                 head = wq->head;
 365                 if (head >= qp->r_rq.size)
 366                         head = 0;
 367                 tail = wq->tail;
 368                 if (tail >= qp->r_rq.size)
 369                         tail = 0;
 370                 /*
 371                  * Compute the number of credits available (RWQEs).
 372                  * There is a small chance that the pair of reads are
 373                  * not atomic, which is OK, since the fuzziness is
 374                  * resolved as further ACKs go out.
 375                  */
 376                 credits = head - tail;
 377                 if ((int)credits < 0)
 378                         credits += qp->r_rq.size;
 379                 /*
 380                  * Binary search the credit table to find the code to
 381                  * use.
 382                  */
 383                 min = 0;
 384                 max = 31;
 385                 for (;;) {
 386                         x = (min + max) / 2;
 387                         if (credit_table[x] == credits)
 388                                 break;
 389                         if (credit_table[x] > credits) {
 390                                 max = x;
 391                         } else {
 392                                 if (min == x)
 393                                         break;
 394                                 min = x;
 395                         }
 396                 }
 397                 aeth |= x << HFI1_AETH_CREDIT_SHIFT;
 398         }
 399         return cpu_to_be32(aeth);
 400 }
 401
 402 /**
 403  * _hfi1_schedule_send - schedule progress
 404  * @qp: the QP
 405  *
 406  * This schedules qp progress w/o regard to the s_flags.
 407  *
 408  * It is only used in the post send, which doesn't hold
 409  * the s_lock.
 410  */
 411 void _hfi1_schedule_send(struct rvt_qp *qp)
 412 {
 413         struct hfi1_qp_priv *priv = qp->priv;
 414         struct hfi1_ibport *ibp =
 415                 to_iport(qp->ibqp.device, qp->port_num);
 416         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 417         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 418
 419         iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
 420                         priv->s_sde ?
 421                         priv->s_sde->cpu :
 422                         cpumask_first(cpumask_of_node(dd->node)));
 423 }
 424
 425 static void qp_pio_drain(struct rvt_qp *qp)
 426 {
 427         struct hfi1_ibdev *dev;
 428         struct hfi1_qp_priv *priv = qp->priv;
 429
 430         if (!priv->s_sendcontext)
 431                 return;
 432         dev = to_idev(qp->ibqp.device);
 433         while (iowait_pio_pending(&priv->s_iowait)) {
 434                 write_seqlock_irq(&dev->iowait_lock);
 435                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
 436                 write_sequnlock_irq(&dev->iowait_lock);
 437                 iowait_pio_drain(&priv->s_iowait);
 438                 write_seqlock_irq(&dev->iowait_lock);
 439                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
 440                 write_sequnlock_irq(&dev->iowait_lock);
 441         }
 442 }
 443
 444 /**
 445  * hfi1_schedule_send - schedule progress
 446  * @qp: the QP
 447  *
 448  * This schedules qp progress and caller should hold
 449  * the s_lock.
 450  */
 451 void hfi1_schedule_send(struct rvt_qp *qp)
 452 {
 453         if (hfi1_send_ok(qp))
 454                 _hfi1_schedule_send(qp);
 455 }
 456
 457 /**
 458  * hfi1_get_credit - flush the send work queue of a QP
 459  * @qp: the qp who's send work queue to flush
 460  * @aeth: the Acknowledge Extended Transport Header
 461  *
 462  * The QP s_lock should be held.
 463  */
 464 void hfi1_get_credit(struct rvt_qp *qp, u32 aeth)
 465 {
 466         u32 credit = (aeth >> HFI1_AETH_CREDIT_SHIFT) & HFI1_AETH_CREDIT_MASK;
 467
 468         /*
 469          * If the credit is invalid, we can send
 470          * as many packets as we like.  Otherwise, we have to
 471          * honor the credit field.
 472          */
 473         if (credit == HFI1_AETH_CREDIT_INVAL) {
 474                 if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
 475                         qp->s_flags |= RVT_S_UNLIMITED_CREDIT;
 476                         if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
 477                                 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
 478                                 hfi1_schedule_send(qp);
 479                         }
 480                 }
 481         } else if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT)) {
 482                 /* Compute new LSN (i.e., MSN + credit) */
 483                 credit = (aeth + credit_table[credit]) & HFI1_MSN_MASK;
 484                 if (cmp_msn(credit, qp->s_lsn) > 0) {
 485                         qp->s_lsn = credit;
 486                         if (qp->s_flags & RVT_S_WAIT_SSN_CREDIT) {
 487                                 qp->s_flags &= ~RVT_S_WAIT_SSN_CREDIT;
 488                                 hfi1_schedule_send(qp);
 489                         }
 490                 }
 491         }
 492 }
 493
 494 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
 495 {
 496         unsigned long flags;
 497
 498         spin_lock_irqsave(&qp->s_lock, flags);
 499         if (qp->s_flags & flag) {
 500                 qp->s_flags &= ~flag;
 501                 trace_hfi1_qpwakeup(qp, flag);
 502                 hfi1_schedule_send(qp);
 503         }
 504         spin_unlock_irqrestore(&qp->s_lock, flags);
 505         /* Notify hfi1_destroy_qp() if it is waiting. */
 506         if (atomic_dec_and_test(&qp->refcount))
 507                 wake_up(&qp->wait);
 508 }
 509
 510 static int iowait_sleep(
 511         struct sdma_engine *sde,
 512         struct iowait *wait,
 513         struct sdma_txreq *stx,
 514         unsigned seq)
 515 {
 516         struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
 517         struct rvt_qp *qp;
 518         struct hfi1_qp_priv *priv;
 519         unsigned long flags;
 520         int ret = 0;
 521         struct hfi1_ibdev *dev;
 522
 523         qp = tx->qp;
 524         priv = qp->priv;
 525
 526         spin_lock_irqsave(&qp->s_lock, flags);
 527         if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 528                 /*
 529                  * If we couldn't queue the DMA request, save the info
 530                  * and try again later rather than destroying the
 531                  * buffer and undoing the side effects of the copy.
 532                  */
 533                 /* Make a common routine? */
 534                 dev = &sde->dd->verbs_dev;
 535                 list_add_tail(&stx->list, &wait->tx_head);
 536                 write_seqlock(&dev->iowait_lock);
 537                 if (sdma_progress(sde, seq, stx))
 538                         goto eagain;
 539                 if (list_empty(&priv->s_iowait.list)) {
 540                         struct hfi1_ibport *ibp =
 541                                 to_iport(qp->ibqp.device, qp->port_num);
 542
 543                         ibp->rvp.n_dmawait++;
 544                         qp->s_flags |= RVT_S_WAIT_DMA_DESC;
 545                         list_add_tail(&priv->s_iowait.list, &sde->dmawait);
 546                         trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
 547                         atomic_inc(&qp->refcount);
 548                 }
 549                 write_sequnlock(&dev->iowait_lock);
 550                 qp->s_flags &= ~RVT_S_BUSY;
 551                 spin_unlock_irqrestore(&qp->s_lock, flags);
 552                 ret = -EBUSY;
 553         } else {
 554                 spin_unlock_irqrestore(&qp->s_lock, flags);
 555                 hfi1_put_txreq(tx);
 556         }
 557         return ret;
 558 eagain:
 559         write_sequnlock(&dev->iowait_lock);
 560         spin_unlock_irqrestore(&qp->s_lock, flags);
 561         list_del_init(&stx->list);
 562         return -EAGAIN;
 563 }
 564
 565 static void iowait_wakeup(struct iowait *wait, int reason)
 566 {
 567         struct rvt_qp *qp = iowait_to_qp(wait);
 568
 569         WARN_ON(reason != SDMA_AVAIL_REASON);
 570         hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
 571 }
 572
 573 static void iowait_sdma_drained(struct iowait *wait)
 574 {
 575         struct rvt_qp *qp = iowait_to_qp(wait);
 576         unsigned long flags;
 577
 578         /*
 579          * This happens when the send engine notes
 580          * a QP in the error state and cannot
 581          * do the flush work until that QP's
 582          * sdma work has finished.
 583          */
 584         spin_lock_irqsave(&qp->s_lock, flags);
 585         if (qp->s_flags & RVT_S_WAIT_DMA) {
 586                 qp->s_flags &= ~RVT_S_WAIT_DMA;
 587                 hfi1_schedule_send(qp);
 588         }
 589         spin_unlock_irqrestore(&qp->s_lock, flags);
 590 }
 591
 592 /**
 593  *
 594  * qp_to_sdma_engine - map a qp to a send engine
 595  * @qp: the QP
 596  * @sc5: the 5 bit sc
 597  *
 598  * Return:
 599  * A send engine for the qp or NULL for SMI type qp.
 600  */
 601 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
 602 {
 603         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 604         struct sdma_engine *sde;
 605
 606         if (!(dd->flags & HFI1_HAS_SEND_DMA))
 607                 return NULL;
 608         switch (qp->ibqp.qp_type) {
 609         case IB_QPT_SMI:
 610                 return NULL;
 611         default:
 612                 break;
 613         }
 614         sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
 615         return sde;
 616 }
 617
 618 /*
 619  * qp_to_send_context - map a qp to a send context
 620  * @qp: the QP
 621  * @sc5: the 5 bit sc
 622  *
 623  * Return:
 624  * A send context for the qp
 625  */
 626 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
 627 {
 628         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 629
 630         switch (qp->ibqp.qp_type) {
 631         case IB_QPT_SMI:
 632                 /* SMA packets to VL15 */
 633                 return dd->vld[15].sc;
 634         default:
 635                 break;
 636         }
 637
 638         return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
 639                                           sc5);
 640 }
 641
 642 struct qp_iter {
 643         struct hfi1_ibdev *dev;
 644         struct rvt_qp *qp;
 645         int specials;
 646         int n;
 647 };
 648
 649 struct qp_iter *qp_iter_init(struct hfi1_ibdev *dev)
 650 {
 651         struct qp_iter *iter;
 652
 653         iter = kzalloc(sizeof(*iter), GFP_KERNEL);
 654         if (!iter)
 655                 return NULL;
 656
 657         iter->dev = dev;
 658         iter->specials = dev->rdi.ibdev.phys_port_cnt * 2;
 659         if (qp_iter_next(iter)) {
 660                 kfree(iter);
 661                 return NULL;
 662         }
 663
 664         return iter;
 665 }
 666
 667 int qp_iter_next(struct qp_iter *iter)
 668 {
 669         struct hfi1_ibdev *dev = iter->dev;
 670         int n = iter->n;
 671         int ret = 1;
 672         struct rvt_qp *pqp = iter->qp;
 673         struct rvt_qp *qp;
 674
 675         /*
 676          * The approach is to consider the special qps
 677          * as an additional table entries before the
 678          * real hash table.  Since the qp code sets
 679          * the qp->next hash link to NULL, this works just fine.
 680          *
 681          * iter->specials is 2 * # ports
 682          *
 683          * n = 0..iter->specials is the special qp indices
 684          *
 685          * n = iter->specials..dev->rdi.qp_dev->qp_table_size+iter->specials are
 686          * the potential hash bucket entries
 687          *
 688          */
 689         for (; n <  dev->rdi.qp_dev->qp_table_size + iter->specials; n++) {
 690                 if (pqp) {
 691                         qp = rcu_dereference(pqp->next);
 692                 } else {
 693                         if (n < iter->specials) {
 694                                 struct hfi1_pportdata *ppd;
 695                                 struct hfi1_ibport *ibp;
 696                                 int pidx;
 697
 698                                 pidx = n % dev->rdi.ibdev.phys_port_cnt;
 699                                 ppd = &dd_from_dev(dev)->pport[pidx];
 700                                 ibp = &ppd->ibport_data;
 701
 702                                 if (!(n & 1))
 703                                         qp = rcu_dereference(ibp->rvp.qp[0]);
 704                                 else
 705                                         qp = rcu_dereference(ibp->rvp.qp[1]);
 706                         } else {
 707                                 qp = rcu_dereference(
 708                                         dev->rdi.qp_dev->qp_table[
 709                                                 (n - iter->specials)]);
 710                         }
 711                 }
 712                 pqp = qp;
 713                 if (qp) {
 714                         iter->qp = qp;
 715                         iter->n = n;
 716                         return 0;
 717                 }
 718         }
 719         return ret;
 720 }
 721
 722 static const char * const qp_type_str[] = {
 723         "SMI", "GSI", "RC", "UC", "UD",
 724 };
 725
 726 static int qp_idle(struct rvt_qp *qp)
 727 {
 728         return
 729                 qp->s_last == qp->s_acked &&
 730                 qp->s_acked == qp->s_cur &&
 731                 qp->s_cur == qp->s_tail &&
 732                 qp->s_tail == qp->s_head;
 733 }
 734
 735 void qp_iter_print(struct seq_file *s, struct qp_iter *iter)
 736 {
 737         struct rvt_swqe *wqe;
 738         struct rvt_qp *qp = iter->qp;
 739         struct hfi1_qp_priv *priv = qp->priv;
 740         struct sdma_engine *sde;
 741         struct send_context *send_context;
 742
 743         sde = qp_to_sdma_engine(qp, priv->s_sc);
 744         wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 745         send_context = qp_to_send_context(qp, priv->s_sc);
 746         seq_printf(s,
 747                    "N %d %s QP %x R %u %s %u %u %u f=%x %u %u %u %u %u %u PSN %x %x %x %x %x (%u %u %u %u %u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d\n",
 748                    iter->n,
 749                    qp_idle(qp) ? "I" : "B",
 750                    qp->ibqp.qp_num,
 751                    atomic_read(&qp->refcount),
 752                    qp_type_str[qp->ibqp.qp_type],
 753                    qp->state,
 754                    wqe ? wqe->wr.opcode : 0,
 755                    qp->s_hdrwords,
 756                    qp->s_flags,
 757                    iowait_sdma_pending(&priv->s_iowait),
 758                    iowait_pio_pending(&priv->s_iowait),
 759                    !list_empty(&priv->s_iowait.list),
 760                    qp->timeout,
 761                    wqe ? wqe->ssn : 0,
 762                    qp->s_lsn,
 763                    qp->s_last_psn,
 764                    qp->s_psn, qp->s_next_psn,
 765                    qp->s_sending_psn, qp->s_sending_hpsn,
 766                    qp->s_last, qp->s_acked, qp->s_cur,
 767                    qp->s_tail, qp->s_head, qp->s_size,
 768                    qp->s_avail,
 769                    qp->remote_qpn,
 770                    qp->remote_ah_attr.dlid,
 771                    qp->remote_ah_attr.sl,
 772                    qp->pmtu,
 773                    qp->s_retry,
 774                    qp->s_retry_cnt,
 775                    qp->s_rnr_retry_cnt,
 776                    sde,
 777                    sde ? sde->this_idx : 0,
 778                    send_context,
 779                    send_context ? send_context->sw_index : 0,
 780                    ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->head,
 781                    ibcq_to_rvtcq(qp->ibqp.send_cq)->queue->tail,
 782                    qp->pid);
 783 }
 784
 785 void qp_comm_est(struct rvt_qp *qp)
 786 {
 787         qp->r_flags |= RVT_R_COMM_EST;
 788         if (qp->ibqp.event_handler) {
 789                 struct ib_event ev;
 790
 791                 ev.device = qp->ibqp.device;
 792                 ev.element.qp = &qp->ibqp;
 793                 ev.event = IB_EVENT_COMM_EST;
 794                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 795         }
 796 }
 797
 798 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 799                     gfp_t gfp)
 800 {
 801         struct hfi1_qp_priv *priv;
 802
 803         priv = kzalloc_node(sizeof(*priv), gfp, rdi->dparms.node);
 804         if (!priv)
 805                 return ERR_PTR(-ENOMEM);
 806
 807         priv->owner = qp;
 808
 809         priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), gfp,
 810                                    rdi->dparms.node);
 811         if (!priv->s_ahg) {
 812                 kfree(priv);
 813                 return ERR_PTR(-ENOMEM);
 814         }
 815         setup_timer(&priv->s_rnr_timer, hfi1_rc_rnr_retry, (unsigned long)qp);
 816         qp->s_timer.function = hfi1_rc_timeout;
 817         return priv;
 818 }
 819
 820 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 821 {
 822         struct hfi1_qp_priv *priv = qp->priv;
 823
 824         kfree(priv->s_ahg);
 825         kfree(priv);
 826 }
 827
 828 unsigned free_all_qps(struct rvt_dev_info *rdi)
 829 {
 830         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 831                                                     struct hfi1_ibdev,
 832                                                     rdi);
 833         struct hfi1_devdata *dd = container_of(verbs_dev,
 834                                                struct hfi1_devdata,
 835                                                verbs_dev);
 836         int n;
 837         unsigned qp_inuse = 0;
 838
 839         for (n = 0; n < dd->num_pports; n++) {
 840                 struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
 841
 842                 rcu_read_lock();
 843                 if (rcu_dereference(ibp->rvp.qp[0]))
 844                         qp_inuse++;
 845                 if (rcu_dereference(ibp->rvp.qp[1]))
 846                         qp_inuse++;
 847                 rcu_read_unlock();
 848         }
 849
 850         return qp_inuse;
 851 }
 852
 853 void flush_qp_waiters(struct rvt_qp *qp)
 854 {
 855         flush_iowait(qp);
 856         hfi1_stop_rc_timers(qp);
 857 }
 858
 859 void stop_send_queue(struct rvt_qp *qp)
 860 {
 861         struct hfi1_qp_priv *priv = qp->priv;
 862
 863         cancel_work_sync(&priv->s_iowait.iowork);
 864         hfi1_del_timers_sync(qp);
 865 }
 866
 867 void quiesce_qp(struct rvt_qp *qp)
 868 {
 869         struct hfi1_qp_priv *priv = qp->priv;
 870
 871         iowait_sdma_drain(&priv->s_iowait);
 872         qp_pio_drain(qp);
 873         flush_tx_list(qp);
 874 }
 875
 876 void notify_qp_reset(struct rvt_qp *qp)
 877 {
 878         struct hfi1_qp_priv *priv = qp->priv;
 879
 880         iowait_init(
 881                 &priv->s_iowait,
 882                 1,
 883                 _hfi1_do_send,
 884                 iowait_sleep,
 885                 iowait_wakeup,
 886                 iowait_sdma_drained);
 887         priv->r_adefered = 0;
 888         clear_ahg(qp);
 889 }
 890
 891 /*
 892  * Switch to alternate path.
 893  * The QP s_lock should be held and interrupts disabled.
 894  */
 895 void hfi1_migrate_qp(struct rvt_qp *qp)
 896 {
 897         struct hfi1_qp_priv *priv = qp->priv;
 898         struct ib_event ev;
 899
 900         qp->s_mig_state = IB_MIG_MIGRATED;
 901         qp->remote_ah_attr = qp->alt_ah_attr;
 902         qp->port_num = qp->alt_ah_attr.port_num;
 903         qp->s_pkey_index = qp->s_alt_pkey_index;
 904         qp->s_flags |= RVT_S_AHG_CLEAR;
 905         priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
 906         priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 907
 908         ev.device = qp->ibqp.device;
 909         ev.element.qp = &qp->ibqp;
 910         ev.event = IB_EVENT_PATH_MIG;
 911         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 912 }
 913
 914 int mtu_to_path_mtu(u32 mtu)
 915 {
 916         return mtu_to_enum(mtu, OPA_MTU_8192);
 917 }
 918
 919 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
 920 {
 921         u32 mtu;
 922         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 923                                                     struct hfi1_ibdev,
 924                                                     rdi);
 925         struct hfi1_devdata *dd = container_of(verbs_dev,
 926                                                struct hfi1_devdata,
 927                                                verbs_dev);
 928         struct hfi1_ibport *ibp;
 929         u8 sc, vl;
 930
 931         ibp = &dd->pport[qp->port_num - 1].ibport_data;
 932         sc = ibp->sl_to_sc[qp->remote_ah_attr.sl];
 933         vl = sc_to_vlt(dd, sc);
 934
 935         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
 936         if (vl < PER_VL_SEND_CONTEXTS)
 937                 mtu = min_t(u32, mtu, dd->vld[vl].mtu);
 938         return mtu;
 939 }
 940
 941 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 942                        struct ib_qp_attr *attr)
 943 {
 944         int mtu, pidx = qp->port_num - 1;
 945         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 946                                                     struct hfi1_ibdev,
 947                                                     rdi);
 948         struct hfi1_devdata *dd = container_of(verbs_dev,
 949                                                struct hfi1_devdata,
 950                                                verbs_dev);
 951         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
 952         if (mtu == -1)
 953                 return -1; /* values less than 0 are error */
 954
 955         if (mtu > dd->pport[pidx].ibmtu)
 956                 return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
 957         else
 958                 return attr->path_mtu;
 959 }
 960
 961 void notify_error_qp(struct rvt_qp *qp)
 962 {
 963         struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 964         struct hfi1_qp_priv *priv = qp->priv;
 965
 966         write_seqlock(&dev->iowait_lock);
 967         if (!list_empty(&priv->s_iowait.list) && !(qp->s_flags & RVT_S_BUSY)) {
 968                 qp->s_flags &= ~RVT_S_ANY_WAIT_IO;
 969                 list_del_init(&priv->s_iowait.list);
 970                 if (atomic_dec_and_test(&qp->refcount))
 971                         wake_up(&qp->wait);
 972         }
 973         write_sequnlock(&dev->iowait_lock);
 974
 975         if (!(qp->s_flags & RVT_S_BUSY)) {
 976                 qp->s_hdrwords = 0;
 977                 if (qp->s_rdma_mr) {
 978                         rvt_put_mr(qp->s_rdma_mr);
 979                         qp->s_rdma_mr = NULL;
 980                 }
 981                 flush_tx_list(qp);
 982         }
 983 }
 984
 985 /**
 986  * hfi1_error_port_qps - put a port's RC/UC qps into error state
 987  * @ibp: the ibport.
 988  * @sl: the service level.
 989  *
 990  * This function places all RC/UC qps with a given service level into error
 991  * state. It is generally called to force upper lay apps to abandon stale qps
 992  * after an sl->sc mapping change.
 993  */
 994 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
 995 {
 996         struct rvt_qp *qp = NULL;
 997         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 998         struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
 999         int n;
1000         int lastwqe;
1001         struct ib_event ev;
1002
1003         rcu_read_lock();
1004
1005         /* Deal only with RC/UC qps that use the given SL. */
1006         for (n = 0; n < dev->rdi.qp_dev->qp_table_size; n++) {
1007                 for (qp = rcu_dereference(dev->rdi.qp_dev->qp_table[n]); qp;
1008                         qp = rcu_dereference(qp->next)) {
1009                         if (qp->port_num == ppd->port &&
1010                             (qp->ibqp.qp_type == IB_QPT_UC ||
1011                              qp->ibqp.qp_type == IB_QPT_RC) &&
1012                             qp->remote_ah_attr.sl == sl &&
1013                             (ib_rvt_state_ops[qp->state] &
1014                              RVT_POST_SEND_OK)) {
1015                                 spin_lock_irq(&qp->r_lock);
1016                                 spin_lock(&qp->s_hlock);
1017                                 spin_lock(&qp->s_lock);
1018                                 lastwqe = rvt_error_qp(qp,
1019                                                        IB_WC_WR_FLUSH_ERR);
1020                                 spin_unlock(&qp->s_lock);
1021                                 spin_unlock(&qp->s_hlock);
1022                                 spin_unlock_irq(&qp->r_lock);
1023                                 if (lastwqe) {
1024                                         ev.device = qp->ibqp.device;
1025                                         ev.element.qp = &qp->ibqp;
1026                                         ev.event =
1027                                                 IB_EVENT_QP_LAST_WQE_REACHED;
1028                                         qp->ibqp.event_handler(&ev,
1029                                                 qp->ibqp.qp_context);
1030                                 }
1031                         }
1032                 }
1033         }
1034
1035         rcu_read_unlock();
1036 }