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 = 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                 /* fall through */
 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 = dd_from_ibdev(qp->ibqp.device);
 371
 372         return iowait_schedule(&priv->s_iowait, ppd->hfi1_wq,
 373                                priv->s_sde ?
 374                                priv->s_sde->cpu :
 375                                cpumask_first(cpumask_of_node(dd->node)));
 376 }
 377
 378 static void qp_pio_drain(struct rvt_qp *qp)
 379 {
 380         struct hfi1_qp_priv *priv = qp->priv;
 381
 382         if (!priv->s_sendcontext)
 383                 return;
 384         while (iowait_pio_pending(&priv->s_iowait)) {
 385                 write_seqlock_irq(&priv->s_sendcontext->waitlock);
 386                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 1);
 387                 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
 388                 iowait_pio_drain(&priv->s_iowait);
 389                 write_seqlock_irq(&priv->s_sendcontext->waitlock);
 390                 hfi1_sc_wantpiobuf_intr(priv->s_sendcontext, 0);
 391                 write_sequnlock_irq(&priv->s_sendcontext->waitlock);
 392         }
 393 }
 394
 395 /**
 396  * hfi1_schedule_send - schedule progress
 397  * @qp: the QP
 398  *
 399  * This schedules qp progress and caller should hold
 400  * the s_lock.
 401  * @return true if the first leg is scheduled;
 402  * false if the first leg is not scheduled.
 403  */
 404 bool hfi1_schedule_send(struct rvt_qp *qp)
 405 {
 406         lockdep_assert_held(&qp->s_lock);
 407         if (hfi1_send_ok(qp)) {
 408                 _hfi1_schedule_send(qp);
 409                 return true;
 410         }
 411         if (qp->s_flags & HFI1_S_ANY_WAIT_IO)
 412                 iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait,
 413                                 IOWAIT_PENDING_IB);
 414         return false;
 415 }
 416
 417 static void hfi1_qp_schedule(struct rvt_qp *qp)
 418 {
 419         struct hfi1_qp_priv *priv = qp->priv;
 420         bool ret;
 421
 422         if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_IB)) {
 423                 ret = hfi1_schedule_send(qp);
 424                 if (ret)
 425                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
 426         }
 427         if (iowait_flag_set(&priv->s_iowait, IOWAIT_PENDING_TID)) {
 428                 ret = hfi1_schedule_tid_send(qp);
 429                 if (ret)
 430                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 431         }
 432 }
 433
 434 void hfi1_qp_wakeup(struct rvt_qp *qp, u32 flag)
 435 {
 436         unsigned long flags;
 437
 438         spin_lock_irqsave(&qp->s_lock, flags);
 439         if (qp->s_flags & flag) {
 440                 qp->s_flags &= ~flag;
 441                 trace_hfi1_qpwakeup(qp, flag);
 442                 hfi1_qp_schedule(qp);
 443         }
 444         spin_unlock_irqrestore(&qp->s_lock, flags);
 445         /* Notify hfi1_destroy_qp() if it is waiting. */
 446         rvt_put_qp(qp);
 447 }
 448
 449 void hfi1_qp_unbusy(struct rvt_qp *qp, struct iowait_work *wait)
 450 {
 451         struct hfi1_qp_priv *priv = qp->priv;
 452
 453         if (iowait_set_work_flag(wait) == IOWAIT_IB_SE) {
 454                 qp->s_flags &= ~RVT_S_BUSY;
 455                 /*
 456                  * If we are sending a first-leg packet from the second leg,
 457                  * we need to clear the busy flag from priv->s_flags to
 458                  * avoid a race condition when the qp wakes up before
 459                  * the call to hfi1_verbs_send() returns to the second
 460                  * leg. In that case, the second leg will terminate without
 461                  * being re-scheduled, resulting in failure to send TID RDMA
 462                  * WRITE DATA and TID RDMA ACK packets.
 463                  */
 464                 if (priv->s_flags & HFI1_S_TID_BUSY_SET) {
 465                         priv->s_flags &= ~(HFI1_S_TID_BUSY_SET |
 466                                            RVT_S_BUSY);
 467                         iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 468                 }
 469         } else {
 470                 priv->s_flags &= ~RVT_S_BUSY;
 471         }
 472 }
 473
 474 static int iowait_sleep(
 475         struct sdma_engine *sde,
 476         struct iowait_work *wait,
 477         struct sdma_txreq *stx,
 478         uint seq,
 479         bool pkts_sent)
 480 {
 481         struct verbs_txreq *tx = container_of(stx, struct verbs_txreq, txreq);
 482         struct rvt_qp *qp;
 483         struct hfi1_qp_priv *priv;
 484         unsigned long flags;
 485         int ret = 0;
 486
 487         qp = tx->qp;
 488         priv = qp->priv;
 489
 490         spin_lock_irqsave(&qp->s_lock, flags);
 491         if (ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK) {
 492                 /*
 493                  * If we couldn't queue the DMA request, save the info
 494                  * and try again later rather than destroying the
 495                  * buffer and undoing the side effects of the copy.
 496                  */
 497                 /* Make a common routine? */
 498                 list_add_tail(&stx->list, &wait->tx_head);
 499                 write_seqlock(&sde->waitlock);
 500                 if (sdma_progress(sde, seq, stx))
 501                         goto eagain;
 502                 if (list_empty(&priv->s_iowait.list)) {
 503                         struct hfi1_ibport *ibp =
 504                                 to_iport(qp->ibqp.device, qp->port_num);
 505
 506                         ibp->rvp.n_dmawait++;
 507                         qp->s_flags |= RVT_S_WAIT_DMA_DESC;
 508                         iowait_get_priority(&priv->s_iowait);
 509                         iowait_queue(pkts_sent, &priv->s_iowait,
 510                                      &sde->dmawait);
 511                         priv->s_iowait.lock = &sde->waitlock;
 512                         trace_hfi1_qpsleep(qp, RVT_S_WAIT_DMA_DESC);
 513                         rvt_get_qp(qp);
 514                 }
 515                 write_sequnlock(&sde->waitlock);
 516                 hfi1_qp_unbusy(qp, wait);
 517                 spin_unlock_irqrestore(&qp->s_lock, flags);
 518                 ret = -EBUSY;
 519         } else {
 520                 spin_unlock_irqrestore(&qp->s_lock, flags);
 521                 hfi1_put_txreq(tx);
 522         }
 523         return ret;
 524 eagain:
 525         write_sequnlock(&sde->waitlock);
 526         spin_unlock_irqrestore(&qp->s_lock, flags);
 527         list_del_init(&stx->list);
 528         return -EAGAIN;
 529 }
 530
 531 static void iowait_wakeup(struct iowait *wait, int reason)
 532 {
 533         struct rvt_qp *qp = iowait_to_qp(wait);
 534
 535         WARN_ON(reason != SDMA_AVAIL_REASON);
 536         hfi1_qp_wakeup(qp, RVT_S_WAIT_DMA_DESC);
 537 }
 538
 539 static void iowait_sdma_drained(struct iowait *wait)
 540 {
 541         struct rvt_qp *qp = iowait_to_qp(wait);
 542         unsigned long flags;
 543
 544         /*
 545          * This happens when the send engine notes
 546          * a QP in the error state and cannot
 547          * do the flush work until that QP's
 548          * sdma work has finished.
 549          */
 550         spin_lock_irqsave(&qp->s_lock, flags);
 551         if (qp->s_flags & RVT_S_WAIT_DMA) {
 552                 qp->s_flags &= ~RVT_S_WAIT_DMA;
 553                 hfi1_schedule_send(qp);
 554         }
 555         spin_unlock_irqrestore(&qp->s_lock, flags);
 556 }
 557
 558 static void hfi1_init_priority(struct iowait *w)
 559 {
 560         struct rvt_qp *qp = iowait_to_qp(w);
 561         struct hfi1_qp_priv *priv = qp->priv;
 562
 563         if (qp->s_flags & RVT_S_ACK_PENDING)
 564                 w->priority++;
 565         if (priv->s_flags & RVT_S_ACK_PENDING)
 566                 w->priority++;
 567 }
 568
 569 /**
 570  * qp_to_sdma_engine - map a qp to a send engine
 571  * @qp: the QP
 572  * @sc5: the 5 bit sc
 573  *
 574  * Return:
 575  * A send engine for the qp or NULL for SMI type qp.
 576  */
 577 struct sdma_engine *qp_to_sdma_engine(struct rvt_qp *qp, u8 sc5)
 578 {
 579         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 580         struct sdma_engine *sde;
 581
 582         if (!(dd->flags & HFI1_HAS_SEND_DMA))
 583                 return NULL;
 584         switch (qp->ibqp.qp_type) {
 585         case IB_QPT_SMI:
 586                 return NULL;
 587         default:
 588                 break;
 589         }
 590         sde = sdma_select_engine_sc(dd, qp->ibqp.qp_num >> dd->qos_shift, sc5);
 591         return sde;
 592 }
 593
 594 /*
 595  * qp_to_send_context - map a qp to a send context
 596  * @qp: the QP
 597  * @sc5: the 5 bit sc
 598  *
 599  * Return:
 600  * A send context for the qp
 601  */
 602 struct send_context *qp_to_send_context(struct rvt_qp *qp, u8 sc5)
 603 {
 604         struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device);
 605
 606         switch (qp->ibqp.qp_type) {
 607         case IB_QPT_SMI:
 608                 /* SMA packets to VL15 */
 609                 return dd->vld[15].sc;
 610         default:
 611                 break;
 612         }
 613
 614         return pio_select_send_context_sc(dd, qp->ibqp.qp_num >> dd->qos_shift,
 615                                           sc5);
 616 }
 617
 618 static const char * const qp_type_str[] = {
 619         "SMI", "GSI", "RC", "UC", "UD",
 620 };
 621
 622 static int qp_idle(struct rvt_qp *qp)
 623 {
 624         return
 625                 qp->s_last == qp->s_acked &&
 626                 qp->s_acked == qp->s_cur &&
 627                 qp->s_cur == qp->s_tail &&
 628                 qp->s_tail == qp->s_head;
 629 }
 630
 631 /**
 632  * qp_iter_print - print the qp information to seq_file
 633  * @s: the seq_file to emit the qp information on
 634  * @iter: the iterator for the qp hash list
 635  */
 636 void qp_iter_print(struct seq_file *s, struct rvt_qp_iter *iter)
 637 {
 638         struct rvt_swqe *wqe;
 639         struct rvt_qp *qp = iter->qp;
 640         struct hfi1_qp_priv *priv = qp->priv;
 641         struct sdma_engine *sde;
 642         struct send_context *send_context;
 643         struct rvt_ack_entry *e = NULL;
 644         struct rvt_srq *srq = qp->ibqp.srq ?
 645                 ibsrq_to_rvtsrq(qp->ibqp.srq) : NULL;
 646
 647         sde = qp_to_sdma_engine(qp, priv->s_sc);
 648         wqe = rvt_get_swqe_ptr(qp, qp->s_last);
 649         send_context = qp_to_send_context(qp, priv->s_sc);
 650         if (qp->s_ack_queue)
 651                 e = &qp->s_ack_queue[qp->s_tail_ack_queue];
 652         seq_printf(s,
 653                    "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",
 654                    iter->n,
 655                    qp_idle(qp) ? "I" : "B",
 656                    qp->ibqp.qp_num,
 657                    atomic_read(&qp->refcount),
 658                    qp_type_str[qp->ibqp.qp_type],
 659                    qp->state,
 660                    wqe ? wqe->wr.opcode : 0,
 661                    qp->s_flags,
 662                    iowait_sdma_pending(&priv->s_iowait),
 663                    iowait_pio_pending(&priv->s_iowait),
 664                    !list_empty(&priv->s_iowait.list),
 665                    qp->timeout,
 666                    wqe ? wqe->ssn : 0,
 667                    qp->s_lsn,
 668                    qp->s_last_psn,
 669                    qp->s_psn, qp->s_next_psn,
 670                    qp->s_sending_psn, qp->s_sending_hpsn,
 671                    qp->r_psn,
 672                    qp->s_last, qp->s_acked, qp->s_cur,
 673                    qp->s_tail, qp->s_head, qp->s_size,
 674                    qp->s_avail,
 675                    /* ack_queue ring pointers, size */
 676                    qp->s_tail_ack_queue, qp->r_head_ack_queue,
 677                    rvt_max_atomic(&to_idev(qp->ibqp.device)->rdi),
 678                    /* remote QP info  */
 679                    qp->remote_qpn,
 680                    rdma_ah_get_dlid(&qp->remote_ah_attr),
 681                    rdma_ah_get_sl(&qp->remote_ah_attr),
 682                    qp->pmtu,
 683                    qp->s_retry,
 684                    qp->s_retry_cnt,
 685                    qp->s_rnr_retry_cnt,
 686                    qp->s_rnr_retry,
 687                    sde,
 688                    sde ? sde->this_idx : 0,
 689                    send_context,
 690                    send_context ? send_context->sw_index : 0,
 691                    ib_cq_head(qp->ibqp.send_cq),
 692                    ib_cq_tail(qp->ibqp.send_cq),
 693                    qp->pid,
 694                    qp->s_state,
 695                    qp->s_ack_state,
 696                    /* ack queue information */
 697                    e ? e->opcode : 0,
 698                    e ? e->psn : 0,
 699                    e ? e->lpsn : 0,
 700                    qp->r_min_rnr_timer,
 701                    srq ? "SRQ" : "RQ",
 702                    srq ? srq->rq.size : qp->r_rq.size
 703                 );
 704 }
 705
 706 void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 707 {
 708         struct hfi1_qp_priv *priv;
 709
 710         priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, rdi->dparms.node);
 711         if (!priv)
 712                 return ERR_PTR(-ENOMEM);
 713
 714         priv->owner = qp;
 715
 716         priv->s_ahg = kzalloc_node(sizeof(*priv->s_ahg), GFP_KERNEL,
 717                                    rdi->dparms.node);
 718         if (!priv->s_ahg) {
 719                 kfree(priv);
 720                 return ERR_PTR(-ENOMEM);
 721         }
 722         iowait_init(
 723                 &priv->s_iowait,
 724                 1,
 725                 _hfi1_do_send,
 726                 _hfi1_do_tid_send,
 727                 iowait_sleep,
 728                 iowait_wakeup,
 729                 iowait_sdma_drained,
 730                 hfi1_init_priority);
 731         /* Init to a value to start the running average correctly */
 732         priv->s_running_pkt_size = piothreshold / 2;
 733         return priv;
 734 }
 735
 736 void qp_priv_free(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 737 {
 738         struct hfi1_qp_priv *priv = qp->priv;
 739
 740         hfi1_qp_priv_tid_free(rdi, qp);
 741         kfree(priv->s_ahg);
 742         kfree(priv);
 743 }
 744
 745 unsigned free_all_qps(struct rvt_dev_info *rdi)
 746 {
 747         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 748                                                     struct hfi1_ibdev,
 749                                                     rdi);
 750         struct hfi1_devdata *dd = container_of(verbs_dev,
 751                                                struct hfi1_devdata,
 752                                                verbs_dev);
 753         int n;
 754         unsigned qp_inuse = 0;
 755
 756         for (n = 0; n < dd->num_pports; n++) {
 757                 struct hfi1_ibport *ibp = &dd->pport[n].ibport_data;
 758
 759                 rcu_read_lock();
 760                 if (rcu_dereference(ibp->rvp.qp[0]))
 761                         qp_inuse++;
 762                 if (rcu_dereference(ibp->rvp.qp[1]))
 763                         qp_inuse++;
 764                 rcu_read_unlock();
 765         }
 766
 767         return qp_inuse;
 768 }
 769
 770 void flush_qp_waiters(struct rvt_qp *qp)
 771 {
 772         lockdep_assert_held(&qp->s_lock);
 773         flush_iowait(qp);
 774         hfi1_tid_rdma_flush_wait(qp);
 775 }
 776
 777 void stop_send_queue(struct rvt_qp *qp)
 778 {
 779         struct hfi1_qp_priv *priv = qp->priv;
 780
 781         iowait_cancel_work(&priv->s_iowait);
 782         if (cancel_work_sync(&priv->tid_rdma.trigger_work))
 783                 rvt_put_qp(qp);
 784 }
 785
 786 void quiesce_qp(struct rvt_qp *qp)
 787 {
 788         struct hfi1_qp_priv *priv = qp->priv;
 789
 790         hfi1_del_tid_reap_timer(qp);
 791         hfi1_del_tid_retry_timer(qp);
 792         iowait_sdma_drain(&priv->s_iowait);
 793         qp_pio_drain(qp);
 794         flush_tx_list(qp);
 795 }
 796
 797 void notify_qp_reset(struct rvt_qp *qp)
 798 {
 799         hfi1_qp_kern_exp_rcv_clear_all(qp);
 800         qp->r_adefered = 0;
 801         clear_ahg(qp);
 802
 803         /* Clear any OPFN state */
 804         if (qp->ibqp.qp_type == IB_QPT_RC)
 805                 opfn_conn_error(qp);
 806 }
 807
 808 /*
 809  * Switch to alternate path.
 810  * The QP s_lock should be held and interrupts disabled.
 811  */
 812 void hfi1_migrate_qp(struct rvt_qp *qp)
 813 {
 814         struct hfi1_qp_priv *priv = qp->priv;
 815         struct ib_event ev;
 816
 817         qp->s_mig_state = IB_MIG_MIGRATED;
 818         qp->remote_ah_attr = qp->alt_ah_attr;
 819         qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
 820         qp->s_pkey_index = qp->s_alt_pkey_index;
 821         qp->s_flags |= HFI1_S_AHG_CLEAR;
 822         priv->s_sc = ah_to_sc(qp->ibqp.device, &qp->remote_ah_attr);
 823         priv->s_sde = qp_to_sdma_engine(qp, priv->s_sc);
 824         qp_set_16b(qp);
 825
 826         ev.device = qp->ibqp.device;
 827         ev.element.qp = &qp->ibqp;
 828         ev.event = IB_EVENT_PATH_MIG;
 829         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 830 }
 831
 832 int mtu_to_path_mtu(u32 mtu)
 833 {
 834         return mtu_to_enum(mtu, OPA_MTU_8192);
 835 }
 836
 837 u32 mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu)
 838 {
 839         u32 mtu;
 840         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 841                                                     struct hfi1_ibdev,
 842                                                     rdi);
 843         struct hfi1_devdata *dd = container_of(verbs_dev,
 844                                                struct hfi1_devdata,
 845                                                verbs_dev);
 846         struct hfi1_ibport *ibp;
 847         u8 sc, vl;
 848
 849         ibp = &dd->pport[qp->port_num - 1].ibport_data;
 850         sc = ibp->sl_to_sc[rdma_ah_get_sl(&qp->remote_ah_attr)];
 851         vl = sc_to_vlt(dd, sc);
 852
 853         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, pmtu);
 854         if (vl < PER_VL_SEND_CONTEXTS)
 855                 mtu = min_t(u32, mtu, dd->vld[vl].mtu);
 856         return mtu;
 857 }
 858
 859 int get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
 860                        struct ib_qp_attr *attr)
 861 {
 862         int mtu, pidx = qp->port_num - 1;
 863         struct hfi1_ibdev *verbs_dev = container_of(rdi,
 864                                                     struct hfi1_ibdev,
 865                                                     rdi);
 866         struct hfi1_devdata *dd = container_of(verbs_dev,
 867                                                struct hfi1_devdata,
 868                                                verbs_dev);
 869         mtu = verbs_mtu_enum_to_int(qp->ibqp.device, attr->path_mtu);
 870         if (mtu == -1)
 871                 return -1; /* values less than 0 are error */
 872
 873         if (mtu > dd->pport[pidx].ibmtu)
 874                 return mtu_to_enum(dd->pport[pidx].ibmtu, IB_MTU_2048);
 875         else
 876                 return attr->path_mtu;
 877 }
 878
 879 void notify_error_qp(struct rvt_qp *qp)
 880 {
 881         struct hfi1_qp_priv *priv = qp->priv;
 882         seqlock_t *lock = priv->s_iowait.lock;
 883
 884         if (lock) {
 885                 write_seqlock(lock);
 886                 if (!list_empty(&priv->s_iowait.list) &&
 887                     !(qp->s_flags & RVT_S_BUSY) &&
 888                     !(priv->s_flags & RVT_S_BUSY)) {
 889                         qp->s_flags &= ~HFI1_S_ANY_WAIT_IO;
 890                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_IB);
 891                         iowait_clear_flag(&priv->s_iowait, IOWAIT_PENDING_TID);
 892                         list_del_init(&priv->s_iowait.list);
 893                         priv->s_iowait.lock = NULL;
 894                         rvt_put_qp(qp);
 895                 }
 896                 write_sequnlock(lock);
 897         }
 898
 899         if (!(qp->s_flags & RVT_S_BUSY) && !(priv->s_flags & RVT_S_BUSY)) {
 900                 qp->s_hdrwords = 0;
 901                 if (qp->s_rdma_mr) {
 902                         rvt_put_mr(qp->s_rdma_mr);
 903                         qp->s_rdma_mr = NULL;
 904                 }
 905                 flush_tx_list(qp);
 906         }
 907 }
 908
 909 /**
 910  * hfi1_qp_iter_cb - callback for iterator
 911  * @qp - the qp
 912  * @v - the sl in low bits of v
 913  *
 914  * This is called from the iterator callback to work
 915  * on an individual qp.
 916  */
 917 static void hfi1_qp_iter_cb(struct rvt_qp *qp, u64 v)
 918 {
 919         int lastwqe;
 920         struct ib_event ev;
 921         struct hfi1_ibport *ibp =
 922                 to_iport(qp->ibqp.device, qp->port_num);
 923         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 924         u8 sl = (u8)v;
 925
 926         if (qp->port_num != ppd->port ||
 927             (qp->ibqp.qp_type != IB_QPT_UC &&
 928              qp->ibqp.qp_type != IB_QPT_RC) ||
 929             rdma_ah_get_sl(&qp->remote_ah_attr) != sl ||
 930             !(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))
 931                 return;
 932
 933         spin_lock_irq(&qp->r_lock);
 934         spin_lock(&qp->s_hlock);
 935         spin_lock(&qp->s_lock);
 936         lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 937         spin_unlock(&qp->s_lock);
 938         spin_unlock(&qp->s_hlock);
 939         spin_unlock_irq(&qp->r_lock);
 940         if (lastwqe) {
 941                 ev.device = qp->ibqp.device;
 942                 ev.element.qp = &qp->ibqp;
 943                 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
 944                 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 945         }
 946 }
 947
 948 /**
 949  * hfi1_error_port_qps - put a port's RC/UC qps into error state
 950  * @ibp: the ibport.
 951  * @sl: the service level.
 952  *
 953  * This function places all RC/UC qps with a given service level into error
 954  * state. It is generally called to force upper lay apps to abandon stale qps
 955  * after an sl->sc mapping change.
 956  */
 957 void hfi1_error_port_qps(struct hfi1_ibport *ibp, u8 sl)
 958 {
 959         struct hfi1_pportdata *ppd = ppd_from_ibp(ibp);
 960         struct hfi1_ibdev *dev = &ppd->dd->verbs_dev;
 961
 962         rvt_qp_iter(&dev->rdi, sl, hfi1_qp_iter_cb);
 963 }