drivers/infiniband/sw/rxe/rxe_verbs.c

   1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
   2 /*
   3  * Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved.
   4  * Copyright (c) 2015 System Fabric Works, Inc. All rights reserved.
   5  */
   6
   7 #include <linux/dma-mapping.h>
   8 #include <net/addrconf.h>
   9 #include <rdma/uverbs_ioctl.h>
  10 #include "rxe.h"
  11 #include "rxe_loc.h"
  12 #include "rxe_queue.h"
  13 #include "rxe_hw_counters.h"
  14
  15 static int rxe_query_device(struct ib_device *dev,
  16                             struct ib_device_attr *attr,
  17                             struct ib_udata *uhw)
  18 {
  19         struct rxe_dev *rxe = to_rdev(dev);
  20
  21         if (uhw->inlen || uhw->outlen)
  22                 return -EINVAL;
  23
  24         *attr = rxe->attr;
  25         return 0;
  26 }
  27
  28 static int rxe_query_port(struct ib_device *dev,
  29                           u8 port_num, struct ib_port_attr *attr)
  30 {
  31         struct rxe_dev *rxe = to_rdev(dev);
  32         struct rxe_port *port;
  33         int rc;
  34
  35         port = &rxe->port;
  36
  37         /* *attr being zeroed by the caller, avoid zeroing it here */
  38         *attr = port->attr;
  39
  40         mutex_lock(&rxe->usdev_lock);
  41         rc = ib_get_eth_speed(dev, port_num, &attr->active_speed,
  42                               &attr->active_width);
  43
  44         if (attr->state == IB_PORT_ACTIVE)
  45                 attr->phys_state = IB_PORT_PHYS_STATE_LINK_UP;
  46         else if (dev_get_flags(rxe->ndev) & IFF_UP)
  47                 attr->phys_state = IB_PORT_PHYS_STATE_POLLING;
  48         else
  49                 attr->phys_state = IB_PORT_PHYS_STATE_DISABLED;
  50
  51         mutex_unlock(&rxe->usdev_lock);
  52
  53         return rc;
  54 }
  55
  56 static int rxe_query_pkey(struct ib_device *device,
  57                           u8 port_num, u16 index, u16 *pkey)
  58 {
  59         if (index > 0)
  60                 return -EINVAL;
  61
  62         *pkey = IB_DEFAULT_PKEY_FULL;
  63         return 0;
  64 }
  65
  66 static int rxe_modify_device(struct ib_device *dev,
  67                              int mask, struct ib_device_modify *attr)
  68 {
  69         struct rxe_dev *rxe = to_rdev(dev);
  70
  71         if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
  72                      IB_DEVICE_MODIFY_NODE_DESC))
  73                 return -EOPNOTSUPP;
  74
  75         if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID)
  76                 rxe->attr.sys_image_guid = cpu_to_be64(attr->sys_image_guid);
  77
  78         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
  79                 memcpy(rxe->ib_dev.node_desc,
  80                        attr->node_desc, sizeof(rxe->ib_dev.node_desc));
  81         }
  82
  83         return 0;
  84 }
  85
  86 static int rxe_modify_port(struct ib_device *dev,
  87                            u8 port_num, int mask, struct ib_port_modify *attr)
  88 {
  89         struct rxe_dev *rxe = to_rdev(dev);
  90         struct rxe_port *port;
  91
  92         port = &rxe->port;
  93
  94         port->attr.port_cap_flags |= attr->set_port_cap_mask;
  95         port->attr.port_cap_flags &= ~attr->clr_port_cap_mask;
  96
  97         if (mask & IB_PORT_RESET_QKEY_CNTR)
  98                 port->attr.qkey_viol_cntr = 0;
  99
 100         return 0;
 101 }
 102
 103 static enum rdma_link_layer rxe_get_link_layer(struct ib_device *dev,
 104                                                u8 port_num)
 105 {
 106         return IB_LINK_LAYER_ETHERNET;
 107 }
 108
 109 static int rxe_alloc_ucontext(struct ib_ucontext *uctx, struct ib_udata *udata)
 110 {
 111         struct rxe_dev *rxe = to_rdev(uctx->device);
 112         struct rxe_ucontext *uc = to_ruc(uctx);
 113
 114         return rxe_add_to_pool(&rxe->uc_pool, &uc->pelem);
 115 }
 116
 117 static void rxe_dealloc_ucontext(struct ib_ucontext *ibuc)
 118 {
 119         struct rxe_ucontext *uc = to_ruc(ibuc);
 120
 121         rxe_drop_ref(uc);
 122 }
 123
 124 static int rxe_port_immutable(struct ib_device *dev, u8 port_num,
 125                               struct ib_port_immutable *immutable)
 126 {
 127         int err;
 128         struct ib_port_attr attr;
 129
 130         immutable->core_cap_flags = RDMA_CORE_PORT_IBA_ROCE_UDP_ENCAP;
 131
 132         err = ib_query_port(dev, port_num, &attr);
 133         if (err)
 134                 return err;
 135
 136         immutable->pkey_tbl_len = attr.pkey_tbl_len;
 137         immutable->gid_tbl_len = attr.gid_tbl_len;
 138         immutable->max_mad_size = IB_MGMT_MAD_SIZE;
 139
 140         return 0;
 141 }
 142
 143 static int rxe_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 144 {
 145         struct rxe_dev *rxe = to_rdev(ibpd->device);
 146         struct rxe_pd *pd = to_rpd(ibpd);
 147
 148         return rxe_add_to_pool(&rxe->pd_pool, &pd->pelem);
 149 }
 150
 151 static int rxe_dealloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
 152 {
 153         struct rxe_pd *pd = to_rpd(ibpd);
 154
 155         rxe_drop_ref(pd);
 156         return 0;
 157 }
 158
 159 static int rxe_create_ah(struct ib_ah *ibah,
 160                          struct rdma_ah_init_attr *init_attr,
 161                          struct ib_udata *udata)
 162
 163 {
 164         int err;
 165         struct rxe_dev *rxe = to_rdev(ibah->device);
 166         struct rxe_ah *ah = to_rah(ibah);
 167
 168         err = rxe_av_chk_attr(rxe, init_attr->ah_attr);
 169         if (err)
 170                 return err;
 171
 172         err = rxe_add_to_pool(&rxe->ah_pool, &ah->pelem);
 173         if (err)
 174                 return err;
 175
 176         rxe_init_av(init_attr->ah_attr, &ah->av);
 177         return 0;
 178 }
 179
 180 static int rxe_modify_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
 181 {
 182         int err;
 183         struct rxe_dev *rxe = to_rdev(ibah->device);
 184         struct rxe_ah *ah = to_rah(ibah);
 185
 186         err = rxe_av_chk_attr(rxe, attr);
 187         if (err)
 188                 return err;
 189
 190         rxe_init_av(attr, &ah->av);
 191         return 0;
 192 }
 193
 194 static int rxe_query_ah(struct ib_ah *ibah, struct rdma_ah_attr *attr)
 195 {
 196         struct rxe_ah *ah = to_rah(ibah);
 197
 198         memset(attr, 0, sizeof(*attr));
 199         attr->type = ibah->type;
 200         rxe_av_to_attr(&ah->av, attr);
 201         return 0;
 202 }
 203
 204 static int rxe_destroy_ah(struct ib_ah *ibah, u32 flags)
 205 {
 206         struct rxe_ah *ah = to_rah(ibah);
 207
 208         rxe_drop_ref(ah);
 209         return 0;
 210 }
 211
 212 static int post_one_recv(struct rxe_rq *rq, const struct ib_recv_wr *ibwr)
 213 {
 214         int err;
 215         int i;
 216         u32 length;
 217         struct rxe_recv_wqe *recv_wqe;
 218         int num_sge = ibwr->num_sge;
 219
 220         if (unlikely(queue_full(rq->queue))) {
 221                 err = -ENOMEM;
 222                 goto err1;
 223         }
 224
 225         if (unlikely(num_sge > rq->max_sge)) {
 226                 err = -EINVAL;
 227                 goto err1;
 228         }
 229
 230         length = 0;
 231         for (i = 0; i < num_sge; i++)
 232                 length += ibwr->sg_list[i].length;
 233
 234         recv_wqe = producer_addr(rq->queue);
 235         recv_wqe->wr_id = ibwr->wr_id;
 236         recv_wqe->num_sge = num_sge;
 237
 238         memcpy(recv_wqe->dma.sge, ibwr->sg_list,
 239                num_sge * sizeof(struct ib_sge));
 240
 241         recv_wqe->dma.length            = length;
 242         recv_wqe->dma.resid             = length;
 243         recv_wqe->dma.num_sge           = num_sge;
 244         recv_wqe->dma.cur_sge           = 0;
 245         recv_wqe->dma.sge_offset        = 0;
 246
 247         advance_producer(rq->queue);
 248         return 0;
 249
 250 err1:
 251         return err;
 252 }
 253
 254 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
 255                           struct ib_udata *udata)
 256 {
 257         int err;
 258         struct rxe_dev *rxe = to_rdev(ibsrq->device);
 259         struct rxe_pd *pd = to_rpd(ibsrq->pd);
 260         struct rxe_srq *srq = to_rsrq(ibsrq);
 261         struct rxe_create_srq_resp __user *uresp = NULL;
 262
 263         if (init->srq_type != IB_SRQT_BASIC)
 264                 return -EOPNOTSUPP;
 265
 266         if (udata) {
 267                 if (udata->outlen < sizeof(*uresp))
 268                         return -EINVAL;
 269                 uresp = udata->outbuf;
 270         }
 271
 272         err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK);
 273         if (err)
 274                 goto err1;
 275
 276         err = rxe_add_to_pool(&rxe->srq_pool, &srq->pelem);
 277         if (err)
 278                 goto err1;
 279
 280         rxe_add_ref(pd);
 281         srq->pd = pd;
 282
 283         err = rxe_srq_from_init(rxe, srq, init, udata, uresp);
 284         if (err)
 285                 goto err2;
 286
 287         return 0;
 288
 289 err2:
 290         rxe_drop_ref(pd);
 291         rxe_drop_ref(srq);
 292 err1:
 293         return err;
 294 }
 295
 296 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 297                           enum ib_srq_attr_mask mask,
 298                           struct ib_udata *udata)
 299 {
 300         int err;
 301         struct rxe_srq *srq = to_rsrq(ibsrq);
 302         struct rxe_dev *rxe = to_rdev(ibsrq->device);
 303         struct rxe_modify_srq_cmd ucmd = {};
 304
 305         if (udata) {
 306                 if (udata->inlen < sizeof(ucmd))
 307                         return -EINVAL;
 308
 309                 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
 310                 if (err)
 311                         return err;
 312         }
 313
 314         err = rxe_srq_chk_attr(rxe, srq, attr, mask);
 315         if (err)
 316                 goto err1;
 317
 318         err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd, udata);
 319         if (err)
 320                 goto err1;
 321
 322         return 0;
 323
 324 err1:
 325         return err;
 326 }
 327
 328 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 329 {
 330         struct rxe_srq *srq = to_rsrq(ibsrq);
 331
 332         if (srq->error)
 333                 return -EINVAL;
 334
 335         attr->max_wr = srq->rq.queue->buf->index_mask;
 336         attr->max_sge = srq->rq.max_sge;
 337         attr->srq_limit = srq->limit;
 338         return 0;
 339 }
 340
 341 static int rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
 342 {
 343         struct rxe_srq *srq = to_rsrq(ibsrq);
 344
 345         if (srq->rq.queue)
 346                 rxe_queue_cleanup(srq->rq.queue);
 347
 348         rxe_drop_ref(srq->pd);
 349         rxe_drop_ref(srq);
 350         return 0;
 351 }
 352
 353 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
 354                              const struct ib_recv_wr **bad_wr)
 355 {
 356         int err = 0;
 357         unsigned long flags;
 358         struct rxe_srq *srq = to_rsrq(ibsrq);
 359
 360         spin_lock_irqsave(&srq->rq.producer_lock, flags);
 361
 362         while (wr) {
 363                 err = post_one_recv(&srq->rq, wr);
 364                 if (unlikely(err))
 365                         break;
 366                 wr = wr->next;
 367         }
 368
 369         spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
 370
 371         if (err)
 372                 *bad_wr = wr;
 373
 374         return err;
 375 }
 376
 377 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd,
 378                                    struct ib_qp_init_attr *init,
 379                                    struct ib_udata *udata)
 380 {
 381         int err;
 382         struct rxe_dev *rxe = to_rdev(ibpd->device);
 383         struct rxe_pd *pd = to_rpd(ibpd);
 384         struct rxe_qp *qp;
 385         struct rxe_create_qp_resp __user *uresp = NULL;
 386
 387         if (udata) {
 388                 if (udata->outlen < sizeof(*uresp))
 389                         return ERR_PTR(-EINVAL);
 390                 uresp = udata->outbuf;
 391         }
 392
 393         if (init->create_flags)
 394                 return ERR_PTR(-EOPNOTSUPP);
 395
 396         err = rxe_qp_chk_init(rxe, init);
 397         if (err)
 398                 goto err1;
 399
 400         qp = rxe_alloc(&rxe->qp_pool);
 401         if (!qp) {
 402                 err = -ENOMEM;
 403                 goto err1;
 404         }
 405
 406         if (udata) {
 407                 if (udata->inlen) {
 408                         err = -EINVAL;
 409                         goto err2;
 410                 }
 411                 qp->is_user = 1;
 412         }
 413
 414         rxe_add_index(qp);
 415
 416         err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd, udata);
 417         if (err)
 418                 goto err3;
 419
 420         return &qp->ibqp;
 421
 422 err3:
 423         rxe_drop_index(qp);
 424 err2:
 425         rxe_drop_ref(qp);
 426 err1:
 427         return ERR_PTR(err);
 428 }
 429
 430 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 431                          int mask, struct ib_udata *udata)
 432 {
 433         int err;
 434         struct rxe_dev *rxe = to_rdev(ibqp->device);
 435         struct rxe_qp *qp = to_rqp(ibqp);
 436
 437         if (mask & ~IB_QP_ATTR_STANDARD_BITS)
 438                 return -EOPNOTSUPP;
 439
 440         err = rxe_qp_chk_attr(rxe, qp, attr, mask);
 441         if (err)
 442                 goto err1;
 443
 444         err = rxe_qp_from_attr(qp, attr, mask, udata);
 445         if (err)
 446                 goto err1;
 447
 448         return 0;
 449
 450 err1:
 451         return err;
 452 }
 453
 454 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 455                         int mask, struct ib_qp_init_attr *init)
 456 {
 457         struct rxe_qp *qp = to_rqp(ibqp);
 458
 459         rxe_qp_to_init(qp, init);
 460         rxe_qp_to_attr(qp, attr, mask);
 461
 462         return 0;
 463 }
 464
 465 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 466 {
 467         struct rxe_qp *qp = to_rqp(ibqp);
 468
 469         rxe_qp_destroy(qp);
 470         rxe_drop_index(qp);
 471         rxe_drop_ref(qp);
 472         return 0;
 473 }
 474
 475 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 476                             unsigned int mask, unsigned int length)
 477 {
 478         int num_sge = ibwr->num_sge;
 479         struct rxe_sq *sq = &qp->sq;
 480
 481         if (unlikely(num_sge > sq->max_sge))
 482                 goto err1;
 483
 484         if (unlikely(mask & WR_ATOMIC_MASK)) {
 485                 if (length < 8)
 486                         goto err1;
 487
 488                 if (atomic_wr(ibwr)->remote_addr & 0x7)
 489                         goto err1;
 490         }
 491
 492         if (unlikely((ibwr->send_flags & IB_SEND_INLINE) &&
 493                      (length > sq->max_inline)))
 494                 goto err1;
 495
 496         return 0;
 497
 498 err1:
 499         return -EINVAL;
 500 }
 501
 502 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
 503                          const struct ib_send_wr *ibwr)
 504 {
 505         wr->wr_id = ibwr->wr_id;
 506         wr->num_sge = ibwr->num_sge;
 507         wr->opcode = ibwr->opcode;
 508         wr->send_flags = ibwr->send_flags;
 509
 510         if (qp_type(qp) == IB_QPT_UD ||
 511             qp_type(qp) == IB_QPT_SMI ||
 512             qp_type(qp) == IB_QPT_GSI) {
 513                 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
 514                 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
 515                 if (qp_type(qp) == IB_QPT_GSI)
 516                         wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
 517                 if (wr->opcode == IB_WR_SEND_WITH_IMM)
 518                         wr->ex.imm_data = ibwr->ex.imm_data;
 519         } else {
 520                 switch (wr->opcode) {
 521                 case IB_WR_RDMA_WRITE_WITH_IMM:
 522                         wr->ex.imm_data = ibwr->ex.imm_data;
 523                         fallthrough;
 524                 case IB_WR_RDMA_READ:
 525                 case IB_WR_RDMA_WRITE:
 526                         wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
 527                         wr->wr.rdma.rkey        = rdma_wr(ibwr)->rkey;
 528                         break;
 529                 case IB_WR_SEND_WITH_IMM:
 530                         wr->ex.imm_data = ibwr->ex.imm_data;
 531                         break;
 532                 case IB_WR_SEND_WITH_INV:
 533                         wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
 534                         break;
 535                 case IB_WR_ATOMIC_CMP_AND_SWP:
 536                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 537                         wr->wr.atomic.remote_addr =
 538                                 atomic_wr(ibwr)->remote_addr;
 539                         wr->wr.atomic.compare_add =
 540                                 atomic_wr(ibwr)->compare_add;
 541                         wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
 542                         wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
 543                         break;
 544                 case IB_WR_LOCAL_INV:
 545                         wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
 546                 break;
 547                 case IB_WR_REG_MR:
 548                         wr->wr.reg.mr = reg_wr(ibwr)->mr;
 549                         wr->wr.reg.key = reg_wr(ibwr)->key;
 550                         wr->wr.reg.access = reg_wr(ibwr)->access;
 551                 break;
 552                 default:
 553                         break;
 554                 }
 555         }
 556 }
 557
 558 static int init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 559                          unsigned int mask, unsigned int length,
 560                          struct rxe_send_wqe *wqe)
 561 {
 562         int num_sge = ibwr->num_sge;
 563         struct ib_sge *sge;
 564         int i;
 565         u8 *p;
 566
 567         init_send_wr(qp, &wqe->wr, ibwr);
 568
 569         if (qp_type(qp) == IB_QPT_UD ||
 570             qp_type(qp) == IB_QPT_SMI ||
 571             qp_type(qp) == IB_QPT_GSI)
 572                 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av));
 573
 574         if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) {
 575                 p = wqe->dma.inline_data;
 576
 577                 sge = ibwr->sg_list;
 578                 for (i = 0; i < num_sge; i++, sge++) {
 579                         memcpy(p, (void *)(uintptr_t)sge->addr,
 580                                         sge->length);
 581
 582                         p += sge->length;
 583                 }
 584         } else if (mask & WR_REG_MASK) {
 585                 wqe->mask = mask;
 586                 wqe->state = wqe_state_posted;
 587                 return 0;
 588         } else
 589                 memcpy(wqe->dma.sge, ibwr->sg_list,
 590                        num_sge * sizeof(struct ib_sge));
 591
 592         wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
 593                 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
 594         wqe->mask               = mask;
 595         wqe->dma.length         = length;
 596         wqe->dma.resid          = length;
 597         wqe->dma.num_sge        = num_sge;
 598         wqe->dma.cur_sge        = 0;
 599         wqe->dma.sge_offset     = 0;
 600         wqe->state              = wqe_state_posted;
 601         wqe->ssn                = atomic_add_return(1, &qp->ssn);
 602
 603         return 0;
 604 }
 605
 606 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 607                          unsigned int mask, u32 length)
 608 {
 609         int err;
 610         struct rxe_sq *sq = &qp->sq;
 611         struct rxe_send_wqe *send_wqe;
 612         unsigned long flags;
 613
 614         err = validate_send_wr(qp, ibwr, mask, length);
 615         if (err)
 616                 return err;
 617
 618         spin_lock_irqsave(&qp->sq.sq_lock, flags);
 619
 620         if (unlikely(queue_full(sq->queue))) {
 621                 err = -ENOMEM;
 622                 goto err1;
 623         }
 624
 625         send_wqe = producer_addr(sq->queue);
 626
 627         err = init_send_wqe(qp, ibwr, mask, length, send_wqe);
 628         if (unlikely(err))
 629                 goto err1;
 630
 631         advance_producer(sq->queue);
 632         spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
 633
 634         return 0;
 635
 636 err1:
 637         spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
 638         return err;
 639 }
 640
 641 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr,
 642                                 const struct ib_send_wr **bad_wr)
 643 {
 644         int err = 0;
 645         unsigned int mask;
 646         unsigned int length = 0;
 647         int i;
 648         struct ib_send_wr *next;
 649
 650         while (wr) {
 651                 mask = wr_opcode_mask(wr->opcode, qp);
 652                 if (unlikely(!mask)) {
 653                         err = -EINVAL;
 654                         *bad_wr = wr;
 655                         break;
 656                 }
 657
 658                 if (unlikely((wr->send_flags & IB_SEND_INLINE) &&
 659                              !(mask & WR_INLINE_MASK))) {
 660                         err = -EINVAL;
 661                         *bad_wr = wr;
 662                         break;
 663                 }
 664
 665                 next = wr->next;
 666
 667                 length = 0;
 668                 for (i = 0; i < wr->num_sge; i++)
 669                         length += wr->sg_list[i].length;
 670
 671                 err = post_one_send(qp, wr, mask, length);
 672
 673                 if (err) {
 674                         *bad_wr = wr;
 675                         break;
 676                 }
 677                 wr = next;
 678         }
 679
 680         rxe_run_task(&qp->req.task, 1);
 681         if (unlikely(qp->req.state == QP_STATE_ERROR))
 682                 rxe_run_task(&qp->comp.task, 1);
 683
 684         return err;
 685 }
 686
 687 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 688                          const struct ib_send_wr **bad_wr)
 689 {
 690         struct rxe_qp *qp = to_rqp(ibqp);
 691
 692         if (unlikely(!qp->valid)) {
 693                 *bad_wr = wr;
 694                 return -EINVAL;
 695         }
 696
 697         if (unlikely(qp->req.state < QP_STATE_READY)) {
 698                 *bad_wr = wr;
 699                 return -EINVAL;
 700         }
 701
 702         if (qp->is_user) {
 703                 /* Utilize process context to do protocol processing */
 704                 rxe_run_task(&qp->req.task, 0);
 705                 return 0;
 706         } else
 707                 return rxe_post_send_kernel(qp, wr, bad_wr);
 708 }
 709
 710 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
 711                          const struct ib_recv_wr **bad_wr)
 712 {
 713         int err = 0;
 714         struct rxe_qp *qp = to_rqp(ibqp);
 715         struct rxe_rq *rq = &qp->rq;
 716         unsigned long flags;
 717
 718         if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) {
 719                 *bad_wr = wr;
 720                 err = -EINVAL;
 721                 goto err1;
 722         }
 723
 724         if (unlikely(qp->srq)) {
 725                 *bad_wr = wr;
 726                 err = -EINVAL;
 727                 goto err1;
 728         }
 729
 730         spin_lock_irqsave(&rq->producer_lock, flags);
 731
 732         while (wr) {
 733                 err = post_one_recv(rq, wr);
 734                 if (unlikely(err)) {
 735                         *bad_wr = wr;
 736                         break;
 737                 }
 738                 wr = wr->next;
 739         }
 740
 741         spin_unlock_irqrestore(&rq->producer_lock, flags);
 742
 743         if (qp->resp.state == QP_STATE_ERROR)
 744                 rxe_run_task(&qp->resp.task, 1);
 745
 746 err1:
 747         return err;
 748 }
 749
 750 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 751                          struct ib_udata *udata)
 752 {
 753         int err;
 754         struct ib_device *dev = ibcq->device;
 755         struct rxe_dev *rxe = to_rdev(dev);
 756         struct rxe_cq *cq = to_rcq(ibcq);
 757         struct rxe_create_cq_resp __user *uresp = NULL;
 758
 759         if (udata) {
 760                 if (udata->outlen < sizeof(*uresp))
 761                         return -EINVAL;
 762                 uresp = udata->outbuf;
 763         }
 764
 765         if (attr->flags)
 766                 return -EOPNOTSUPP;
 767
 768         err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector);
 769         if (err)
 770                 return err;
 771
 772         err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata,
 773                                uresp);
 774         if (err)
 775                 return err;
 776
 777         return rxe_add_to_pool(&rxe->cq_pool, &cq->pelem);
 778 }
 779
 780 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 781 {
 782         struct rxe_cq *cq = to_rcq(ibcq);
 783
 784         rxe_cq_disable(cq);
 785
 786         rxe_drop_ref(cq);
 787         return 0;
 788 }
 789
 790 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
 791 {
 792         int err;
 793         struct rxe_cq *cq = to_rcq(ibcq);
 794         struct rxe_dev *rxe = to_rdev(ibcq->device);
 795         struct rxe_resize_cq_resp __user *uresp = NULL;
 796
 797         if (udata) {
 798                 if (udata->outlen < sizeof(*uresp))
 799                         return -EINVAL;
 800                 uresp = udata->outbuf;
 801         }
 802
 803         err = rxe_cq_chk_attr(rxe, cq, cqe, 0);
 804         if (err)
 805                 goto err1;
 806
 807         err = rxe_cq_resize_queue(cq, cqe, uresp, udata);
 808         if (err)
 809                 goto err1;
 810
 811         return 0;
 812
 813 err1:
 814         return err;
 815 }
 816
 817 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 818 {
 819         int i;
 820         struct rxe_cq *cq = to_rcq(ibcq);
 821         struct rxe_cqe *cqe;
 822         unsigned long flags;
 823
 824         spin_lock_irqsave(&cq->cq_lock, flags);
 825         for (i = 0; i < num_entries; i++) {
 826                 cqe = queue_head(cq->queue);
 827                 if (!cqe)
 828                         break;
 829
 830                 memcpy(wc++, &cqe->ibwc, sizeof(*wc));
 831                 advance_consumer(cq->queue);
 832         }
 833         spin_unlock_irqrestore(&cq->cq_lock, flags);
 834
 835         return i;
 836 }
 837
 838 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
 839 {
 840         struct rxe_cq *cq = to_rcq(ibcq);
 841         int count = queue_count(cq->queue);
 842
 843         return (count > wc_cnt) ? wc_cnt : count;
 844 }
 845
 846 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
 847 {
 848         struct rxe_cq *cq = to_rcq(ibcq);
 849         unsigned long irq_flags;
 850         int ret = 0;
 851
 852         spin_lock_irqsave(&cq->cq_lock, irq_flags);
 853         if (cq->notify != IB_CQ_NEXT_COMP)
 854                 cq->notify = flags & IB_CQ_SOLICITED_MASK;
 855
 856         if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue))
 857                 ret = 1;
 858
 859         spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
 860
 861         return ret;
 862 }
 863
 864 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
 865 {
 866         struct rxe_dev *rxe = to_rdev(ibpd->device);
 867         struct rxe_pd *pd = to_rpd(ibpd);
 868         struct rxe_mem *mr;
 869
 870         mr = rxe_alloc(&rxe->mr_pool);
 871         if (!mr)
 872                 return ERR_PTR(-ENOMEM);
 873
 874         rxe_add_index(mr);
 875         rxe_add_ref(pd);
 876         rxe_mem_init_dma(pd, access, mr);
 877
 878         return &mr->ibmr;
 879 }
 880
 881 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
 882                                      u64 start,
 883                                      u64 length,
 884                                      u64 iova,
 885                                      int access, struct ib_udata *udata)
 886 {
 887         int err;
 888         struct rxe_dev *rxe = to_rdev(ibpd->device);
 889         struct rxe_pd *pd = to_rpd(ibpd);
 890         struct rxe_mem *mr;
 891
 892         mr = rxe_alloc(&rxe->mr_pool);
 893         if (!mr) {
 894                 err = -ENOMEM;
 895                 goto err2;
 896         }
 897
 898         rxe_add_index(mr);
 899
 900         rxe_add_ref(pd);
 901
 902         err = rxe_mem_init_user(pd, start, length, iova,
 903                                 access, udata, mr);
 904         if (err)
 905                 goto err3;
 906
 907         return &mr->ibmr;
 908
 909 err3:
 910         rxe_drop_ref(pd);
 911         rxe_drop_index(mr);
 912         rxe_drop_ref(mr);
 913 err2:
 914         return ERR_PTR(err);
 915 }
 916
 917 static int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 918 {
 919         struct rxe_mem *mr = to_rmr(ibmr);
 920
 921         mr->state = RXE_MEM_STATE_ZOMBIE;
 922         rxe_drop_ref(mr_pd(mr));
 923         rxe_drop_index(mr);
 924         rxe_drop_ref(mr);
 925         return 0;
 926 }
 927
 928 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
 929                                   u32 max_num_sg)
 930 {
 931         struct rxe_dev *rxe = to_rdev(ibpd->device);
 932         struct rxe_pd *pd = to_rpd(ibpd);
 933         struct rxe_mem *mr;
 934         int err;
 935
 936         if (mr_type != IB_MR_TYPE_MEM_REG)
 937                 return ERR_PTR(-EINVAL);
 938
 939         mr = rxe_alloc(&rxe->mr_pool);
 940         if (!mr) {
 941                 err = -ENOMEM;
 942                 goto err1;
 943         }
 944
 945         rxe_add_index(mr);
 946
 947         rxe_add_ref(pd);
 948
 949         err = rxe_mem_init_fast(pd, max_num_sg, mr);
 950         if (err)
 951                 goto err2;
 952
 953         return &mr->ibmr;
 954
 955 err2:
 956         rxe_drop_ref(pd);
 957         rxe_drop_index(mr);
 958         rxe_drop_ref(mr);
 959 err1:
 960         return ERR_PTR(err);
 961 }
 962
 963 static int rxe_set_page(struct ib_mr *ibmr, u64 addr)
 964 {
 965         struct rxe_mem *mr = to_rmr(ibmr);
 966         struct rxe_map *map;
 967         struct rxe_phys_buf *buf;
 968
 969         if (unlikely(mr->nbuf == mr->num_buf))
 970                 return -ENOMEM;
 971
 972         map = mr->map[mr->nbuf / RXE_BUF_PER_MAP];
 973         buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP];
 974
 975         buf->addr = addr;
 976         buf->size = ibmr->page_size;
 977         mr->nbuf++;
 978
 979         return 0;
 980 }
 981
 982 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
 983                          int sg_nents, unsigned int *sg_offset)
 984 {
 985         struct rxe_mem *mr = to_rmr(ibmr);
 986         int n;
 987
 988         mr->nbuf = 0;
 989
 990         n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page);
 991
 992         mr->va = ibmr->iova;
 993         mr->iova = ibmr->iova;
 994         mr->length = ibmr->length;
 995         mr->page_shift = ilog2(ibmr->page_size);
 996         mr->page_mask = ibmr->page_size - 1;
 997         mr->offset = mr->iova & mr->page_mask;
 998
 999         return n;
1000 }
1001
1002 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1003 {
1004         int err;
1005         struct rxe_dev *rxe = to_rdev(ibqp->device);
1006         struct rxe_qp *qp = to_rqp(ibqp);
1007         struct rxe_mc_grp *grp;
1008
1009         /* takes a ref on grp if successful */
1010         err = rxe_mcast_get_grp(rxe, mgid, &grp);
1011         if (err)
1012                 return err;
1013
1014         err = rxe_mcast_add_grp_elem(rxe, qp, grp);
1015
1016         rxe_drop_ref(grp);
1017         return err;
1018 }
1019
1020 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1021 {
1022         struct rxe_dev *rxe = to_rdev(ibqp->device);
1023         struct rxe_qp *qp = to_rqp(ibqp);
1024
1025         return rxe_mcast_drop_grp_elem(rxe, qp, mgid);
1026 }
1027
1028 static ssize_t parent_show(struct device *device,
1029                            struct device_attribute *attr, char *buf)
1030 {
1031         struct rxe_dev *rxe =
1032                 rdma_device_to_drv_device(device, struct rxe_dev, ib_dev);
1033
1034         return sysfs_emit(buf, "%s\n", rxe_parent_name(rxe, 1));
1035 }
1036
1037 static DEVICE_ATTR_RO(parent);
1038
1039 static struct attribute *rxe_dev_attributes[] = {
1040         &dev_attr_parent.attr,
1041         NULL
1042 };
1043
1044 static const struct attribute_group rxe_attr_group = {
1045         .attrs = rxe_dev_attributes,
1046 };
1047
1048 static int rxe_enable_driver(struct ib_device *ib_dev)
1049 {
1050         struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
1051
1052         rxe_set_port_state(rxe);
1053         dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(rxe->ndev));
1054         return 0;
1055 }
1056
1057 static const struct ib_device_ops rxe_dev_ops = {
1058         .owner = THIS_MODULE,
1059         .driver_id = RDMA_DRIVER_RXE,
1060         .uverbs_abi_ver = RXE_UVERBS_ABI_VERSION,
1061
1062         .alloc_hw_stats = rxe_ib_alloc_hw_stats,
1063         .alloc_mr = rxe_alloc_mr,
1064         .alloc_pd = rxe_alloc_pd,
1065         .alloc_ucontext = rxe_alloc_ucontext,
1066         .attach_mcast = rxe_attach_mcast,
1067         .create_ah = rxe_create_ah,
1068         .create_cq = rxe_create_cq,
1069         .create_qp = rxe_create_qp,
1070         .create_srq = rxe_create_srq,
1071         .create_user_ah = rxe_create_ah,
1072         .dealloc_driver = rxe_dealloc,
1073         .dealloc_pd = rxe_dealloc_pd,
1074         .dealloc_ucontext = rxe_dealloc_ucontext,
1075         .dereg_mr = rxe_dereg_mr,
1076         .destroy_ah = rxe_destroy_ah,
1077         .destroy_cq = rxe_destroy_cq,
1078         .destroy_qp = rxe_destroy_qp,
1079         .destroy_srq = rxe_destroy_srq,
1080         .detach_mcast = rxe_detach_mcast,
1081         .enable_driver = rxe_enable_driver,
1082         .get_dma_mr = rxe_get_dma_mr,
1083         .get_hw_stats = rxe_ib_get_hw_stats,
1084         .get_link_layer = rxe_get_link_layer,
1085         .get_port_immutable = rxe_port_immutable,
1086         .map_mr_sg = rxe_map_mr_sg,
1087         .mmap = rxe_mmap,
1088         .modify_ah = rxe_modify_ah,
1089         .modify_device = rxe_modify_device,
1090         .modify_port = rxe_modify_port,
1091         .modify_qp = rxe_modify_qp,
1092         .modify_srq = rxe_modify_srq,
1093         .peek_cq = rxe_peek_cq,
1094         .poll_cq = rxe_poll_cq,
1095         .post_recv = rxe_post_recv,
1096         .post_send = rxe_post_send,
1097         .post_srq_recv = rxe_post_srq_recv,
1098         .query_ah = rxe_query_ah,
1099         .query_device = rxe_query_device,
1100         .query_pkey = rxe_query_pkey,
1101         .query_port = rxe_query_port,
1102         .query_qp = rxe_query_qp,
1103         .query_srq = rxe_query_srq,
1104         .reg_user_mr = rxe_reg_user_mr,
1105         .req_notify_cq = rxe_req_notify_cq,
1106         .resize_cq = rxe_resize_cq,
1107
1108         INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah),
1109         INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq),
1110         INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd),
1111         INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq),
1112         INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc),
1113 };
1114
1115 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name)
1116 {
1117         int err;
1118         struct ib_device *dev = &rxe->ib_dev;
1119         struct crypto_shash *tfm;
1120
1121         strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1122
1123         dev->node_type = RDMA_NODE_IB_CA;
1124         dev->phys_port_cnt = 1;
1125         dev->num_comp_vectors = num_possible_cpus();
1126         dev->local_dma_lkey = 0;
1127         addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1128                             rxe->ndev->dev_addr);
1129
1130         dev->uverbs_cmd_mask |= BIT_ULL(IB_USER_VERBS_CMD_POST_SEND) |
1131                                 BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ);
1132
1133         ib_set_device_ops(dev, &rxe_dev_ops);
1134         err = ib_device_set_netdev(&rxe->ib_dev, rxe->ndev, 1);
1135         if (err)
1136                 return err;
1137
1138         tfm = crypto_alloc_shash("crc32", 0, 0);
1139         if (IS_ERR(tfm)) {
1140                 pr_err("failed to allocate crc algorithm err:%ld\n",
1141                        PTR_ERR(tfm));
1142                 return PTR_ERR(tfm);
1143         }
1144         rxe->tfm = tfm;
1145
1146         rdma_set_device_sysfs_group(dev, &rxe_attr_group);
1147         err = ib_register_device(dev, ibdev_name, NULL);
1148         if (err)
1149                 pr_warn("%s failed with error %d\n", __func__, err);
1150
1151         /*
1152          * Note that rxe may be invalid at this point if another thread
1153          * unregistered it.
1154          */
1155         return err;
1156 }