drivers/infiniband/hw/hfi1/qp.c

   1 /*
   2  * Copyright(c) 2015 - 2020 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_work *wait,
  70         struct sdma_txreq *stx,
  71         unsigned int seq,
  72         bool pkts_sent);
  73 static void iowait_wakeup(struct iowait *wait, int reason);
  74 static void iowait_sdma_drained(struct iowait *wait);
  75 static void qp_pio_drain(struct rvt_qp *qp);
  76
  77 const struct rvt_operation_params hfi1_post_parms[RVT_OPERATION_MAX] = {
  78 [IB_WR_RDMA_WRITE] = {
  79         .length = sizeof(struct ib_rdma_wr),
  80         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
  81 },
  82
  83 [IB_WR_RDMA_READ] = {
  84         .length = sizeof(struct ib_rdma_wr),
  85         .qpt_support = BIT(IB_QPT_RC),
  86         .flags = RVT_OPERATION_ATOMIC,
  87 },
  88
  89 [IB_WR_ATOMIC_CMP_AND_SWP] = {
  90         .length = sizeof(struct ib_atomic_wr),
  91         .qpt_support = BIT(IB_QPT_RC),
  92         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
  93 },
  94
  95 [IB_WR_ATOMIC_FETCH_AND_ADD] = {
  96         .length = sizeof(struct ib_atomic_wr),
  97         .qpt_support = BIT(IB_QPT_RC),
  98         .flags = RVT_OPERATION_ATOMIC | RVT_OPERATION_ATOMIC_SGE,
  99 },
 100
 101 [IB_WR_RDMA_WRITE_WITH_IMM] = {
 102         .length = sizeof(struct ib_rdma_wr),
 103         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 104 },
 105
 106 [IB_WR_SEND] = {
 107         .length = sizeof(struct ib_send_wr),
 108         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 109                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 110 },
 111
 112 [IB_WR_SEND_WITH_IMM] = {
 113         .length = sizeof(struct ib_send_wr),
 114         .qpt_support = BIT(IB_QPT_UD) | BIT(IB_QPT_SMI) | BIT(IB_QPT_GSI) |
 115                        BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 116 },
 117
 118 [IB_WR_REG_MR] = {
 119         .length = sizeof(struct ib_reg_wr),
 120         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 121         .flags = RVT_OPERATION_LOCAL,
 122 },
 123
 124 [IB_WR_LOCAL_INV] = {
 125         .length = sizeof(struct ib_send_wr),
 126         .qpt_support = BIT(IB_QPT_UC) | BIT(IB_QPT_RC),
 127         .flags = RVT_OPERATION_LOCAL,
 128 },
 129
 130 [IB_WR_SEND_WITH_INV] = {
 131         .length = sizeof(struct ib_send_wr),
 132         .qpt_support = BIT(IB_QPT_RC),
 133 },
 134
 135 [IB_WR_OPFN] = {
 136         .length = sizeof(struct ib_atomic_wr),
 137         .qpt_support = BIT(IB_QPT_RC),
 138         .flags = RVT_OPERATION_USE_RESERVE,
 139 },
 140
 141 [IB_WR_TID_RDMA_WRITE] = {
 142         .length = sizeof(struct ib_rdma_wr),
 143         .qpt_support = BIT(IB_QPT_RC),
 144         .flags = RVT_OPERATION_IGN_RNR_CNT,
 145 },
 146
 147 };
 148
 149 static void flush_list_head(struct list_head *l)
 150 {
 151         while (!list_empty(l)) {
 152                 struct sdma_txreq *tx;
 153
 154                 tx = list_first_entry(
 155                         l,
 156                         struct sdma_txreq,
 157                         list);
 158                 list_del_init(&tx->list);
 159                 hfi1_put_txreq(
 160                         container_of(tx, struct verbs_txreq, txreq));
 161         }
 162 }
 163
 164 static void flush_tx_list(struct rvt_qp *qp)
 165 {
 166         struct hfi1_qp_priv *priv = qp->priv;
 167
 168         flush_list_head(&iowait_get_ib_work(&priv->s_iowait)->tx_head);
 169         flush_list_head(&iowait_get_tid_work(&priv->s_iowait)->tx_head);
 170 }
 171
 172 static void flush_iowait(struct rvt_qp *qp)
 173 {
 174         struct hfi1_qp_priv *priv = qp->priv;
 175         unsigned long flags;
 176         seqlock_t *lock = priv->s_iowait.lock;
 177
 178         if (!lock)
 179                 return;
 180         write_seqlock_irqsave(lock, flags);
 181         if (!list_empty(&priv->s_iowait.list)) {
 182                 list_del_init(&priv->s_iowait.list);
 183                 priv->s_iowait.lock = NULL;
 184                 rvt_put_qp(qp);
 185         }
 186         write_sequnlock_irqrestore(lock, flags);
 187 }
 188
 189 /**
 190  * This function is what we would push to the core layer if we wanted to be a
 191  * "first class citizen".  Instead we hide this here and rely on Verbs ULPs
 192  * to blindly pass the MTU enum value from the PathRecord to us.
 193  */
 194 static inline int verbs_mtu_enum_to_int(struct ib_device *dev, enum ib_mtu mtu)
 195 {
 196         /* Constraining 10KB packets to 8KB packets */
 197         if (mtu == (enum ib_mtu)OPA_MTU_10240)
 198                 mtu = (enum ib_mtu)OPA_MTU_8192;
 199         return opa_mtu_enum_to_int((enum opa_mtu)mtu);
 200 }
 201
 202 int hfi1_check_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
 203                          int attr_mask, struct ib_udata *udata)
 204 {
 205         struct ib_qp *ibqp = &qp->ibqp;
 206         struct hfi1_ibdev *dev = to_idev(ibqp->device);
 207         struct hfi1_devdata *dd = dd_from_dev(dev);
 208         u8 sc;
 209
 210         if (attr_mask & IB_QP_AV) {
 211                 sc = ah_to_sc(ibqp->device, &attr->ah_attr);
 212                 if (sc == 0xf)
 213                         return -EINVAL;
 214
 215                 if (!qp_to_sdma_engine(qp, sc) &&
 216                     dd->flags & HFI1_HAS_SEND_DMA)
 217                         return -EINVAL;
 218
 219                 if (!qp_to_send_context(qp, sc))
 220                         return -EINVAL;
 221         }
 222
 223         if (attr_mask & IB_QP_ALT_PATH) {
 224                 sc = ah_to_sc(ibqp->device, &attr->alt_ah_attr);
 225                 if (sc == 0xf)
 226                         return -EINVAL;
 227
 228                 if (!qp_to_sdma_engine(qp, sc) &&
 229                     dd->flags & HFI1_HAS_SEND_DMA)
 230                         return -EINVAL;
 231
 232                 if (!qp_to_send_context(qp, sc))
 233                         return -EINVAL;
 234         }
 235
 236         return 0;
 237 }
 238
 239 /*
 240  * qp_set_16b - Set the hdr_type based on whether the slid or the
 241  * dlid in the connection is extended. Only applicable for RC and UC
 242  * QPs. UD QPs determine this on the fly from the ah in the wqe
 243  */
 244 static inline void qp_set_16b(struct rvt_qp *qp)
 245 {
 246         struct hfi1_pportdata *ppd;
 247         struct hfi1_ibport *ibp;
 248         struct hfi1_qp_priv *priv = qp->priv;
 249
 250         /* Update ah_attr to account for extended LIDs */
 251         hfi1_update_ah_attr(qp->ibqp.device, &qp->remote_ah_attr);
 252
 253         /* Create 32 bit LIDs */
 254         hfi1_make_opa_lid(&qp->remote_ah_attr);
 255
 256         if (!(rdma_ah_get_ah_flags(&qp->remote_ah_attr) & IB_AH_GRH))
 257                 return;
 258
 259         ibp = to_iport(qp->ibqp.device, qp->port_num);
 260         ppd = ppd_from_ibp(ibp);
 261         priv->hdr_type = hfi1_get_hdr_type(ppd->lid, &qp->remote_ah_attr);
 262 }
 263
 264 void hfi1_modify_qp(struct rvt_qp *qp, struct ib_qp_attr *attr,
 265                     int attr_mask, struct ib_udata *udata)
 266 {
 267         struct ib_qp *ibqp = &qp->ibqp;
 268         struct hfi1_qp_priv *priv = qp->priv;
 269
 270         if (attr_mask & IB_QP_AV) {
 271                 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
 272                 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 273                 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 274                 qp_set_16b(qp);
 275         }
 276
 277         if (attr_mask & IB_QP_PATH_MIG_STATE &&
 278             attr->path_mig_state == IB_MIG_MIGRATED &&
 279             qp->s_mig_state == IB_MIG_ARMED) {
 280                 qp->s_flags |= HFI1_S_AHG_CLEAR;
 281                 priv->s_sc = ah_to_sc(ibqp->device, &qp->remote_ah_attr);
 282                 priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 283                 priv->s_sendcontext = qp_to_send_context(qp, priv->s_sc);
 284                 qp_set_16b(qp);
 285         }
 286
 287         opfn_qp_init(qp, attr, attr_mask);
 288 }
 289
 290 /**
 291  * hfi1_setup_wqe - set up the wqe
 292  * @qp - The qp
 293  * @wqe - The built wqe
 294  * @call_send - Determine if the send should be posted or scheduled.
 295  *
 296  * Perform setup of the wqe.  This is called
 297  * prior to inserting the wqe into the ring but after
 298  * the wqe has been setup by RDMAVT. This function
 299  * allows the driver the opportunity to perform
 300  * validation and additional setup of the wqe.
 301  *
 302  * Returns 0 on success, -EINVAL on failure
 303  *
 304  */
 305 int hfi1_setup_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe, bool *call_send)
 306 {
 307         struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num);
 308         struct rvt_ah *ah;
 309         struct hfi1_pportdata *ppd;
 310         struct hfi1_devdata *dd;
 311
 312         switch (qp->ibqp.qp_type) {
 313         case IB_QPT_RC:
 314                 hfi1_setup_tid_rdma_wqe(qp, wqe);
 315                 fallthrough;
 316         case IB_QPT_UC:
 317                 if (wqe->length > 0x80000000U)
 318                         return -EINVAL;
 319                 if (wqe->length > qp->pmtu)
 320                         *call_send = false;
 321                 break;
 322         case IB_QPT_SMI:
 323                 /*
 324                  * SM packets should exclusively use VL15 and their SL is
 325                  * ignored (IBTA v1.3, Section 3.5.8.2). Therefore, when ah
 326                  * is created, SL is 0 in most cases and as a result some
 327                  * fields (vl and pmtu) in ah may not be set correctly,
 328                  * depending on the SL2SC and SC2VL tables at the time.
 329                  */
 330                 ppd = ppd_from_ibp(ibp);
 331                 dd = dd_from_ppd(ppd);
 332                 if (wqe->length > dd->vld[15].mtu)
 333                         return -EINVAL;
 334                 break;
 335         case IB_QPT_GSI:
 336         case IB_QPT_UD:
 337                 ah = rvt_get_swqe_ah(wqe);
 338                 if (wqe->length > (1 << ah->log_pmtu))
 339                         return -EINVAL;
 340                 if (ibp->sl_to_sc[rdma_ah_get_sl(&ah->attr)] == 0xf)
 341                         return -EINVAL;
 342         default:
 343                 break;
 344         }
 345
 346         /*
 347          * System latency between send and schedule is large enough that
 348          * forcing call_send to true for piothreshold packets is necessary.
 349          */
 350         if (wqe->length <= piothreshold)
 351                 *call_send = true;
 352         return 0;
 353 }
 354
 355 /**
 356  * _hfi1_schedule_send - schedule progress
 357  * @qp: the QP
 358  *
 359  * This schedules qp progress w/o regard to the s_flags.
 360  *
 361  * It is only used in the post send, which doesn't hold
 362  * the s_lock.
 363  */
 364 bool _hfi1_schedule_send(struct rvt_qp *qp)
 365 {
 366         struct hfi1_qp_priv *priv = qp->priv;
 367         struct hfi1_ibport *ibp =
 368                 to_iport(qp->ibqp.device, qp->port_num);
 369         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 370         struct hfi1_devdata *dd = ppd->dd;
 371
 372         if (dd->flags & HFI1_SHUTDOWN)
 373                 return true;
 374
 375         return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
 376                                priv->s_sde ?
 377                                priv->s_sde->cpu :
 378                                cpumask_first(cpumask_of_node(dd->node)));
 379 }
 380
 381 static void qp_pio_drain(struct rvt_qp *qp)
 382 {
 383         struct hfi1_qp_priv *priv = qp->priv;
 384
 385         if (!priv->s_sendcontext)
 386                 return;
 387         while (iowait_pio_pending(&priv->s_iowait)) {
 388                 write_seqlock_irq(&priv->s_sendcontext->waitlock);
 389                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
 390                 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
 391                 iowait_pio_drain(&priv->s_iowait);
 392                 write_seqlock_irq(&priv->s_sendcontext->waitlock);
 393                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
 394                 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
 395         }
 396 }
 397
 398 /**
 399  * hfi1_schedule_send - schedule progress
 400  * @qp: the QP
 401  *
 402  * This schedules qp progress and caller should hold
 403  * the s_lock.
 404  * @return true if the first leg is scheduled;
 405  * false if the first leg is not scheduled.
 406  */
 407 bool hfi1_schedule_send(struct rvt_qp *qp)
 408 {
 409         lockdep_assert_held(&qp->s_lock);
 410         if (hfi1_send_ok(qp)) {
 411                 _hfi1_schedule_send(qp);
 412                 return true;
 413         }
 414         if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
 415                 iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
 416                                 IOWAIT_PENDING_IB);
 417         return false;
 418 }
 419
 420 static void hfi1_qp_schedule(struct rvt_qp *qp)
 421 {
 422         struct hfi1_qp_priv *priv = qp->priv;
 423         bool ret;
 424
 425         if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
 426                 ret = hfi1_schedule_send(qp);
 427                 if (ret)
 428                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
 429         }
 430         if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
 431                 ret = hfi1_schedule_tid_send(qp);
 432                 if (ret)
 433                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 434         }
 435 }
 436
 437 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
 438 {
 439         unsigned long flags;
 440
 441         spin_lock_irqsave(&qp->s_lock, flags);
 442         if (qp->s_flags & flag) {
 443                 qp->s_flags &= ~flag;
 444                 trace_hfi1_qpwakeup(qp, flag);
 445                 hfi1_qp_schedule(qp);
 446         }
 447         spin_unlock_irqrestore(&qp->s_lock, flags);
 448         /* Notify hfi1_destroy_qp() if it is waiting. */
 449         rvt_put_qp(qp);
 450 }
 451
 452 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
 453 {
 454         struct hfi1_qp_priv *priv = qp->priv;
 455
 456         if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
 457                 qp->s_flags &= ~RVT_S_BUSY;
 458                 /*
 459                  * If we are sending a first-leg packet from the second leg,
 460                  * we need to clear the busy flag from priv->s_flags to
 461                  * avoid a race condition when the qp wakes up before
 462                  * the call to hfi1_verbs_send() returns to the second
 463                  * leg. In that case, the second leg will terminate without
 464                  * being re-scheduled, resulting in failure to send TID RDMA
 465                  * WRITE DATA and TID RDMA ACK packets.
 466                  */
 467                 if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
 468                         priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
 469                                            RVT_S_BUSY);
 470                         iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 471                 }
 472         } else {
 473                 priv->s_flags &= ~RVT_S_BUSY;
 474         }
 475 }
 476
 477 static int iowait_sleep(
 478         struct sdma_engine *sde,
 479         struct iowait_work *wait,
 480         struct sdma_txreq *stx,
 481         uint seq,
 482         bool pkts_sent)
 483 {
 484         struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
 485         struct rvt_qp *qp;
 486         struct hfi1_qp_priv *priv;
 487         unsigned long flags;
 488         int ret = 0;
 489
 490         qp = tx->qp;
 491         priv = qp->priv;
 492
 493         spin_lock_irqsave(&qp->s_lock, flags);
 494         if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 495                 /*
 496                  * If we couldn't queue the DMA request, save the info
 497                  * and try again later rather than destroying the
 498                  * buffer and undoing the side effects of the copy.
 499                  */
 500                 /* Make a common routine? */
 501                 list_add_tail(&stx->list, &wait->tx_head);
 502                 write_seqlock(&sde->waitlock);
 503                 if (sdma_progress(sde, seq, stx))
 504                         goto eagain;
 505                 if (list_empty(&priv->s_iowait.list)) {
 506                         struct hfi1_ibport *ibp =
 507                                 to_iport(qp->ibqp.device, qp->port_num);
 508
 509                         ibp->rvp.n_dmawait++;
 510                         qp->s_flags |= RVT_S_WAIT_DMA_DESC;
 511                         iowait_get_priority(&priv->s_iowait);
 512                         iowait_queue(pkts_sent, &priv->s_iowait,
 513                                      &sde->dmawait);
 514                         priv->s_iowait.lock = &sde->waitlock;
 515                         trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
 516                         rvt_get_qp(qp);
 517                 }
 518                 write_sequnlock(&sde->waitlock);
 519                 hfi1_qp_unbusy(qp, wait);
 520                 spin_unlock_irqrestore(&qp->s_lock, flags);
 521                 ret = -EBUSY;
 522         } else {
 523                 spin_unlock_irqrestore(&qp->s_lock, flags);
 524                 hfi1_put_txreq(tx);
 525         }
 526         return ret;
 527 eagain:
 528         write_sequnlock(&sde->waitlock);
 529         spin_unlock_irqrestore(&qp->s_lock, flags);
 530         list_del_init(&stx->list);
 531         return -EAGAIN;
 532 }
 533
 534 static void iowait_wakeup(struct iowait *wait, int reason)
 535 {
 536         struct rvt_qp *qp = iowait_to_qp(wait);
 537
 538         WARN_ON(reason != SDMA_AVAIL_REASON);
 539         hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
 540 }
 541
 542 static void iowait_sdma_drained(struct iowait *wait)
 543 {
 544         struct rvt_qp *qp = iowait_to_qp(wait);
 545         unsigned long flags;
 546
 547         /*
 548          * This happens when the send engine notes
 549          * a QP in the error state and cannot
 550          * do the flush work until that QP's
 551          * sdma work has finished.
 552          */
 553         spin_lock_irqsave(&qp->s_lock, flags);
 554         if (qp->s_flags & RVT_S_WAIT_DMA) {
 555                 qp->s_flags &= ~RVT_S_WAIT_DMA;
 556                 hfi1_schedule_send(qp);
 557         }
 558         spin_unlock_irqrestore(&qp->s_lock, flags);
 559 }
 560
 561 static void hfi1_init_priority(struct iowait *w)
 562 {
 563         struct rvt_qp *qp = iowait_to_qp(w);
 564         struct hfi1_qp_priv *priv = qp->priv;
 565
 566         if (qp->s_flags & RVT_S_ACK_PENDING)
 567                 w->priority++;
 568         if (priv->s_flags & RVT_S_ACK_PENDING)
 569                 w->priority++;
 570 }
 571
 572 /**
 573  * qp_to_sdma_engine - map a qp to a send engine
 574  * @qp: the QP
 575  * @sc5: the 5 bit sc
 576  *
 577  * Return:
 578  * A send engine for the qp or NULL for SMI type qp.
 579  */
 580 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
 581 {
 582         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 583         struct sdma_engine *sde;
 584
 585         if (!(dd->flags & HFI1_HAS_SEND_DMA))
 586                 return NULL;
 587         switch (qp->ibqp.qp_type) {
 588         case IB_QPT_SMI:
 589                 return NULL;
 590         default:
 591                 break;
 592         }
 593         sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
 594         return sde;
 595 }
 596
 597 /*
 598  * qp_to_send_context - map a qp to a send context
 599  * @qp: the QP
 600  * @sc5: the 5 bit sc
 601  *
 602  * Return:
 603  * A send context for the qp
 604  */
 605 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
 606 {
 607         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 608
 609         switch (qp->ibqp.qp_type) {
 610         case IB_QPT_SMI:
 611                 /* SMA packets to VL15 */
 612                 return dd->vld[15].sc;
 613         default:
 614                 break;
 615         }
 616
 617         return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
 618                                           sc5);
 619 }
 620
 621 static const char * const qp_type_str[] = {
 622         "SMI", "GSI", "RC", "UC", "UD",
 623 };
 624
 625 static int qp_idle(struct rvt_qp *qp)
 626 {
 627         return
 628                 qp->s_last == qp->s_acked &&
 629                 qp->s_acked == qp->s_cur &&
 630                 qp->s_cur == qp->s_tail &&
 631                 qp->s_tail == qp->s_head;
 632 }
 633
 634 /**
 635  * qp_iter_print - print the qp information to seq_file
 636  * @s: the seq_file to emit the qp information on
 637  * @iter: the iterator for the qp hash list
 638  */
 639 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
 640 {
 641         struct rvt_swqe *wqe;
 642         struct rvt_qp *qp = iter->qp;
 643         struct hfi1_qp_priv *priv = qp->priv;
 644         struct sdma_engine *sde;
 645         struct send_context *send_context;
 646         struct rvt_ack_entry *e = NULL;
 647         struct rvt_srq *srq = qp->ibqp.srq ?
 648                 ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
 649
 650         sde = qp_to_sdma_engine(qp, priv->s_sc);
 651         wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 652         send_context = qp_to_send_context(qp, priv->s_sc);
 653         if (qp->s_ack_queue)
 654                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 655         seq_printf(s,
 656                    "N %d %s QP %x R %u %s %u %u f=%x %u %u %u %u %u %u SPSN %x %x %x %x %x RPSN %x S(%u %u %u %u %u %u %u) R(%u %u %u) RQP %x LID %x SL %u MTU %u %u %u %u %u SDE %p,%u SC %p,%u SCQ %u %u PID %d OS %x %x E %x %x %x RNR %d %s %d\n",
 657                    iter->n,
 658                    qp_idle(qp) ? "I" : "B",
 659                    qp->ibqp.qp_num,
 660                    atomic_read(&qp->refcount),
 661                    qp_type_str[qp->ibqp.qp_type],
 662                    qp->state,
 663                    wqe ? wqe->wr.opcode : 0,
 664                    qp->s_flags,
 665                    iowait_sdma_pending(&priv->s_iowait),
 666                    iowait_pio_pending(&priv->s_iowait),
 667                    !list_empty(&priv->s_iowait.list),
 668                    qp->timeout,
 669                    wqe ? wqe->ssn : 0,
 670                    qp->s_lsn,
 671                    qp->s_last_psn,
 672                    qp->s_psn, qp->s_next_psn,
 673                    qp->s_sending_psn, qp->s_sending_hpsn,
 674                    qp->r_psn,
 675                    qp->s_last, qp->s_acked, qp->s_cur,
 676                    qp->s_tail, qp->s_head, qp->s_size,
 677                    qp->s_avail,
 678                    /* ack_queue ring pointers, size */
 679                    qp->s_tail_ack_queue, qp->r_head_ack_queue,
 680                    rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
 681                    /* remote QP info  */
 682                    qp->remote_qpn,
 683                    rdma_ah_get_dlid(&qp->remote_ah_attr),
 684                    rdma_ah_get_sl(&qp->remote_ah_attr),
 685                    qp->pmtu,
 686                    qp->s_retry,
 687                    qp->s_retry_cnt,
 688                    qp->s_rnr_retry_cnt,
 689                    qp->s_rnr_retry,
 690                    sde,
 691                    sde ? sde->this_idx : 0,
 692                    send_context,
 693                    send_context ? send_context->sw_index : 0,
 694                    ib_cq_head(qp->ibqp.send_cq),
 695                    ib_cq_tail(qp->ibqp.send_cq),
 696                    qp->pid,
 697                    qp->s_state,
 698                    qp->s_ack_state,
 699                    /* ack queue information */
 700                    e ? e->opcode : 0,
 701                    e ? e->psn : 0,
 702                    e ? e->lpsn : 0,
 703                    qp->r_min_rnr_timer,
 704                    srq ? "SRQ" : "RQ",
 705                    srq ? srq->rq.size : qp->r_rq.size
 706                 );
 707 }
 708
 709 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 710 {
 711         struct hfi1_qp_priv *priv;
 712
 713         priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
 714         if (!priv)
 715                 return ERR_PTR(-ENOMEM);
 716
 717         priv->owner = qp;
 718
 719         priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
 720                                    rdi->dparms.node);
 721         if (!priv->s_ahg) {
 722                 kfree(priv);
 723                 return ERR_PTR(-ENOMEM);
 724         }
 725         iowait_init(
 726                 &priv->s_iowait,
 727                 1,
 728                 _hfi1_do_send,
 729                 _hfi1_do_tid_send,
 730                 iowait_sleep,
 731                 iowait_wakeup,
 732                 iowait_sdma_drained,
 733                 hfi1_init_priority);
 734         /* Init to a value to start the running average correctly */
 735         priv->s_running_pkt_size = piothreshold / 2;
 736         return priv;
 737 }
 738
 739 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 740 {
 741         struct hfi1_qp_priv *priv = qp->priv;
 742
 743         hfi1_qp_priv_tid_free(rdi, qp);
 744         kfree(priv->s_ahg);
 745         kfree(priv);
 746 }
 747
 748 unsigned free_all_qps(struct rvt_dev_info *rdi)
 749 {
 750         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 751                                                     struct hfi1_ibdev,
 752                                                     rdi);
 753         struct hfi1_devdata *dd = container_of(verbs_dev,
 754                                                struct hfi1_devdata,
 755                                                verbs_dev);
 756         int n;
 757         unsigned qp_inuse = 0;
 758
 759         for (n = 0; n < dd->num_pports; n++) {
 760                 struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
 761
 762                 rcu_read_lock();
 763                 if (rcu_dereference(ibp->rvp.qp[0]))
 764                         qp_inuse++;
 765                 if (rcu_dereference(ibp->rvp.qp[1]))
 766                         qp_inuse++;
 767                 rcu_read_unlock();
 768         }
 769
 770         return qp_inuse;
 771 }
 772
 773 void flush_qp_waiters(struct rvt_qp *qp)
 774 {
 775         lockdep_assert_held(&qp->s_lock);
 776         flush_iowait(qp);
 777         hfi1_tid_rdma_flush_wait(qp);
 778 }
 779
 780 void stop_send_queue(struct rvt_qp *qp)
 781 {
 782         struct hfi1_qp_priv *priv = qp->priv;
 783
 784         iowait_cancel_work(&priv->s_iowait);
 785         if (cancel_work_sync(&priv->tid_rdma.trigger_work))
 786                 rvt_put_qp(qp);
 787 }
 788
 789 void quiesce_qp(struct rvt_qp *qp)
 790 {
 791         struct hfi1_qp_priv *priv = qp->priv;
 792
 793         hfi1_del_tid_reap_timer(qp);
 794         hfi1_del_tid_retry_timer(qp);
 795         iowait_sdma_drain(&priv->s_iowait);
 796         qp_pio_drain(qp);
 797         flush_tx_list(qp);
 798 }
 799
 800 void notify_qp_reset(struct rvt_qp *qp)
 801 {
 802         hfi1_qp_kern_exp_rcv_clear_all(qp);
 803         qp->r_adefered = 0;
 804         clear_ahg(qp);
 805
 806         /* Clear any OPFN state */
 807         if (qp->ibqp.qp_type == IB_QPT_RC)
 808                 opfn_conn_error(qp);
 809 }
 810
 811 /*
 812  * Switch to alternate path.
 813  * The QP s_lock should be held and interrupts disabled.
 814  */
 815 void hfi1_migrate_qp(struct rvt_qp *qp)
 816 {
 817         struct hfi1_qp_priv *priv = qp->priv;
 818         struct ib_event ev;
 819
 820         qp->s_mig_state = IB_MIG_MIGRATED;
 821         qp->remote_ah_attr = qp->alt_ah_attr;
 822         qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
 823         qp->s_pkey_index = qp->s_alt_pkey_index;
 824         qp->s_flags |= HFI1_S_AHG_CLEAR;
 825         priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
 826         priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 827         qp_set_16b(qp);
 828
 829         ev.device = qp->ibqp.device;
 830         ev.element.qp = &qp->ibqp;
 831         ev.event = IB_EVENT_PATH_MIG;
 832         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 833 }
 834
 835 int mtu_to_path_mtu(u32 mtu)
 836 {
 837         return mtu_to_enum(mtu, OPA_MTU_8192);
 838 }
 839
 840 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
 841 {
 842         u32 mtu;
 843         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 844                                                     struct hfi1_ibdev,
 845                                                     rdi);
 846         struct hfi1_devdata *dd = container_of(verbs_dev,
 847                                                struct hfi1_devdata,
 848                                                verbs_dev);
 849         struct hfi1_ibport *ibp;
 850         u8 sc, vl;
 851
 852         ibp = &dd->pport[qp->port_num - 1].ibport_data;
 853         sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
 854         vl = sc_to_vlt(dd, sc);
 855
 856         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
 857         if (vl < PER_VL_SEND_CONTEXTS)
 858                 mtu = min_t(u32, mtu, dd->vld[vl].mtu);
 859         return mtu;
 860 }
 861
 862 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 863                        struct ib_qp_attr *attr)
 864 {
 865         int mtu, pidx = qp->port_num - 1;
 866         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 867                                                     struct hfi1_ibdev,
 868                                                     rdi);
 869         struct hfi1_devdata *dd = container_of(verbs_dev,
 870                                                struct hfi1_devdata,
 871                                                verbs_dev);
 872         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
 873         if (mtu == -1)
 874                 return -1; /* values less than 0 are error */
 875
 876         if (mtu > dd->pport[pidx].ibmtu)
 877                 return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
 878         else
 879                 return attr->path_mtu;
 880 }
 881
 882 void notify_error_qp(struct rvt_qp *qp)
 883 {
 884         struct hfi1_qp_priv *priv = qp->priv;
 885         seqlock_t *lock = priv->s_iowait.lock;
 886
 887         if (lock) {
 888                 write_seqlock(lock);
 889                 if (!list_empty(&priv->s_iowait.list) &&
 890                     !(qp->s_flags & RVT_S_BUSY) &&
 891                     !(priv->s_flags & RVT_S_BUSY)) {
 892                         qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
 893                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
 894                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 895                         list_del_init(&priv->s_iowait.list);
 896                         priv->s_iowait.lock = NULL;
 897                         rvt_put_qp(qp);
 898                 }
 899                 write_sequnlock(lock);
 900         }
 901
 902         if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
 903                 qp->s_hdrwords = 0;
 904                 if (qp->s_rdma_mr) {
 905                         rvt_put_mr(qp->s_rdma_mr);
 906                         qp->s_rdma_mr = NULL;
 907                 }
 908                 flush_tx_list(qp);
 909         }
 910 }
 911
 912 /**
 913  * hfi1_qp_iter_cb - callback for iterator
 914  * @qp - the qp
 915  * @v - the sl in low bits of v
 916  *
 917  * This is called from the iterator callback to work
 918  * on an individual qp.
 919  */
 920 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
 921 {
 922         int lastwqe;
 923         struct ib_event ev;
 924         struct hfi1_ibport *ibp =
 925                 to_iport(qp->ibqp.device, qp->port_num);
 926         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 927         u8 sl = (u8)v;
 928
 929         if (qp->port_num != ppd->port ||
 930             (qp->ibqp.qp_type != IB_QPT_UC &&
 931              qp->ibqp.qp_type != IB_QPT_RC) ||
 932             rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
 933             !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
 934                 return;
 935
 936         spin_lock_irq(&qp->r_lock);
 937         spin_lock(&qp->s_hlock);
 938         spin_lock(&qp->s_lock);
 939         lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 940         spin_unlock(&qp->s_lock);
 941         spin_unlock(&qp->s_hlock);
 942         spin_unlock_irq(&qp->r_lock);
 943         if (lastwqe) {
 944                 ev.device = qp->ibqp.device;
 945                 ev.element.qp = &qp->ibqp;
 946                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 947                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 948         }
 949 }
 950
 951 /**
 952  * hfi1_error_port_qps - put a port's RC/UC qps into error state
 953  * @ibp: the ibport.
 954  * @sl: the service level.
 955  *
 956  * This function places all RC/UC qps with a given service level into error
 957  * state. It is generally called to force upper lay apps to abandon stale qps
 958  * after an sl->sc mapping change.
 959  */
 960 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
 961 {
 962         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 963         struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
 964
 965         rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
 966 }