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