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         /* make sure all changes to the work queue are written before we
 248          * update the producer pointer
 249          */
 250         smp_wmb();
 251
 252         advance_producer(rq->queue);
 253         return 0;
 254
 255 err1:
 256         return err;
 257 }
 258
 259 static int rxe_create_srq(struct ib_srq *ibsrq, struct ib_srq_init_attr *init,
 260                           struct ib_udata *udata)
 261 {
 262         int err;
 263         struct rxe_dev *rxe = to_rdev(ibsrq->device);
 264         struct rxe_pd *pd = to_rpd(ibsrq->pd);
 265         struct rxe_srq *srq = to_rsrq(ibsrq);
 266         struct rxe_create_srq_resp __user *uresp = NULL;
 267
 268         if (udata) {
 269                 if (udata->outlen < sizeof(*uresp))
 270                         return -EINVAL;
 271                 uresp = udata->outbuf;
 272         }
 273
 274         err = rxe_srq_chk_attr(rxe, NULL, &init->attr, IB_SRQ_INIT_MASK);
 275         if (err)
 276                 goto err1;
 277
 278         err = rxe_add_to_pool(&rxe->srq_pool, &srq->pelem);
 279         if (err)
 280                 goto err1;
 281
 282         rxe_add_ref(pd);
 283         srq->pd = pd;
 284
 285         err = rxe_srq_from_init(rxe, srq, init, udata, uresp);
 286         if (err)
 287                 goto err2;
 288
 289         return 0;
 290
 291 err2:
 292         rxe_drop_ref(pd);
 293         rxe_drop_ref(srq);
 294 err1:
 295         return err;
 296 }
 297
 298 static int rxe_modify_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr,
 299                           enum ib_srq_attr_mask mask,
 300                           struct ib_udata *udata)
 301 {
 302         int err;
 303         struct rxe_srq *srq = to_rsrq(ibsrq);
 304         struct rxe_dev *rxe = to_rdev(ibsrq->device);
 305         struct rxe_modify_srq_cmd ucmd = {};
 306
 307         if (udata) {
 308                 if (udata->inlen < sizeof(ucmd))
 309                         return -EINVAL;
 310
 311                 err = ib_copy_from_udata(&ucmd, udata, sizeof(ucmd));
 312                 if (err)
 313                         return err;
 314         }
 315
 316         err = rxe_srq_chk_attr(rxe, srq, attr, mask);
 317         if (err)
 318                 goto err1;
 319
 320         err = rxe_srq_from_attr(rxe, srq, attr, mask, &ucmd, udata);
 321         if (err)
 322                 goto err1;
 323
 324         return 0;
 325
 326 err1:
 327         return err;
 328 }
 329
 330 static int rxe_query_srq(struct ib_srq *ibsrq, struct ib_srq_attr *attr)
 331 {
 332         struct rxe_srq *srq = to_rsrq(ibsrq);
 333
 334         if (srq->error)
 335                 return -EINVAL;
 336
 337         attr->max_wr = srq->rq.queue->buf->index_mask;
 338         attr->max_sge = srq->rq.max_sge;
 339         attr->srq_limit = srq->limit;
 340         return 0;
 341 }
 342
 343 static int rxe_destroy_srq(struct ib_srq *ibsrq, struct ib_udata *udata)
 344 {
 345         struct rxe_srq *srq = to_rsrq(ibsrq);
 346
 347         if (srq->rq.queue)
 348                 rxe_queue_cleanup(srq->rq.queue);
 349
 350         rxe_drop_ref(srq->pd);
 351         rxe_drop_ref(srq);
 352         return 0;
 353 }
 354
 355 static int rxe_post_srq_recv(struct ib_srq *ibsrq, const struct ib_recv_wr *wr,
 356                              const struct ib_recv_wr **bad_wr)
 357 {
 358         int err = 0;
 359         unsigned long flags;
 360         struct rxe_srq *srq = to_rsrq(ibsrq);
 361
 362         spin_lock_irqsave(&srq->rq.producer_lock, flags);
 363
 364         while (wr) {
 365                 err = post_one_recv(&srq->rq, wr);
 366                 if (unlikely(err))
 367                         break;
 368                 wr = wr->next;
 369         }
 370
 371         spin_unlock_irqrestore(&srq->rq.producer_lock, flags);
 372
 373         if (err)
 374                 *bad_wr = wr;
 375
 376         return err;
 377 }
 378
 379 static struct ib_qp *rxe_create_qp(struct ib_pd *ibpd,
 380                                    struct ib_qp_init_attr *init,
 381                                    struct ib_udata *udata)
 382 {
 383         int err;
 384         struct rxe_dev *rxe = to_rdev(ibpd->device);
 385         struct rxe_pd *pd = to_rpd(ibpd);
 386         struct rxe_qp *qp;
 387         struct rxe_create_qp_resp __user *uresp = NULL;
 388
 389         if (udata) {
 390                 if (udata->outlen < sizeof(*uresp))
 391                         return ERR_PTR(-EINVAL);
 392                 uresp = udata->outbuf;
 393         }
 394
 395         err = rxe_qp_chk_init(rxe, init);
 396         if (err)
 397                 goto err1;
 398
 399         qp = rxe_alloc(&rxe->qp_pool);
 400         if (!qp) {
 401                 err = -ENOMEM;
 402                 goto err1;
 403         }
 404
 405         if (udata) {
 406                 if (udata->inlen) {
 407                         err = -EINVAL;
 408                         goto err2;
 409                 }
 410                 qp->is_user = 1;
 411         }
 412
 413         rxe_add_index(qp);
 414
 415         err = rxe_qp_from_init(rxe, qp, pd, init, uresp, ibpd, udata);
 416         if (err)
 417                 goto err3;
 418
 419         return &qp->ibqp;
 420
 421 err3:
 422         rxe_drop_index(qp);
 423 err2:
 424         rxe_drop_ref(qp);
 425 err1:
 426         return ERR_PTR(err);
 427 }
 428
 429 static int rxe_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 430                          int mask, struct ib_udata *udata)
 431 {
 432         int err;
 433         struct rxe_dev *rxe = to_rdev(ibqp->device);
 434         struct rxe_qp *qp = to_rqp(ibqp);
 435
 436         err = rxe_qp_chk_attr(rxe, qp, attr, mask);
 437         if (err)
 438                 goto err1;
 439
 440         err = rxe_qp_from_attr(qp, attr, mask, udata);
 441         if (err)
 442                 goto err1;
 443
 444         return 0;
 445
 446 err1:
 447         return err;
 448 }
 449
 450 static int rxe_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 451                         int mask, struct ib_qp_init_attr *init)
 452 {
 453         struct rxe_qp *qp = to_rqp(ibqp);
 454
 455         rxe_qp_to_init(qp, init);
 456         rxe_qp_to_attr(qp, attr, mask);
 457
 458         return 0;
 459 }
 460
 461 static int rxe_destroy_qp(struct ib_qp *ibqp, struct ib_udata *udata)
 462 {
 463         struct rxe_qp *qp = to_rqp(ibqp);
 464
 465         rxe_qp_destroy(qp);
 466         rxe_drop_index(qp);
 467         rxe_drop_ref(qp);
 468         return 0;
 469 }
 470
 471 static int validate_send_wr(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 472                             unsigned int mask, unsigned int length)
 473 {
 474         int num_sge = ibwr->num_sge;
 475         struct rxe_sq *sq = &qp->sq;
 476
 477         if (unlikely(num_sge > sq->max_sge))
 478                 goto err1;
 479
 480         if (unlikely(mask & WR_ATOMIC_MASK)) {
 481                 if (length < 8)
 482                         goto err1;
 483
 484                 if (atomic_wr(ibwr)->remote_addr & 0x7)
 485                         goto err1;
 486         }
 487
 488         if (unlikely((ibwr->send_flags & IB_SEND_INLINE) &&
 489                      (length > sq->max_inline)))
 490                 goto err1;
 491
 492         return 0;
 493
 494 err1:
 495         return -EINVAL;
 496 }
 497
 498 static void init_send_wr(struct rxe_qp *qp, struct rxe_send_wr *wr,
 499                          const struct ib_send_wr *ibwr)
 500 {
 501         wr->wr_id = ibwr->wr_id;
 502         wr->num_sge = ibwr->num_sge;
 503         wr->opcode = ibwr->opcode;
 504         wr->send_flags = ibwr->send_flags;
 505
 506         if (qp_type(qp) == IB_QPT_UD ||
 507             qp_type(qp) == IB_QPT_SMI ||
 508             qp_type(qp) == IB_QPT_GSI) {
 509                 wr->wr.ud.remote_qpn = ud_wr(ibwr)->remote_qpn;
 510                 wr->wr.ud.remote_qkey = ud_wr(ibwr)->remote_qkey;
 511                 if (qp_type(qp) == IB_QPT_GSI)
 512                         wr->wr.ud.pkey_index = ud_wr(ibwr)->pkey_index;
 513                 if (wr->opcode == IB_WR_SEND_WITH_IMM)
 514                         wr->ex.imm_data = ibwr->ex.imm_data;
 515         } else {
 516                 switch (wr->opcode) {
 517                 case IB_WR_RDMA_WRITE_WITH_IMM:
 518                         wr->ex.imm_data = ibwr->ex.imm_data;
 519                         fallthrough;
 520                 case IB_WR_RDMA_READ:
 521                 case IB_WR_RDMA_WRITE:
 522                         wr->wr.rdma.remote_addr = rdma_wr(ibwr)->remote_addr;
 523                         wr->wr.rdma.rkey        = rdma_wr(ibwr)->rkey;
 524                         break;
 525                 case IB_WR_SEND_WITH_IMM:
 526                         wr->ex.imm_data = ibwr->ex.imm_data;
 527                         break;
 528                 case IB_WR_SEND_WITH_INV:
 529                         wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
 530                         break;
 531                 case IB_WR_ATOMIC_CMP_AND_SWP:
 532                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 533                         wr->wr.atomic.remote_addr =
 534                                 atomic_wr(ibwr)->remote_addr;
 535                         wr->wr.atomic.compare_add =
 536                                 atomic_wr(ibwr)->compare_add;
 537                         wr->wr.atomic.swap = atomic_wr(ibwr)->swap;
 538                         wr->wr.atomic.rkey = atomic_wr(ibwr)->rkey;
 539                         break;
 540                 case IB_WR_LOCAL_INV:
 541                         wr->ex.invalidate_rkey = ibwr->ex.invalidate_rkey;
 542                 break;
 543                 case IB_WR_REG_MR:
 544                         wr->wr.reg.mr = reg_wr(ibwr)->mr;
 545                         wr->wr.reg.key = reg_wr(ibwr)->key;
 546                         wr->wr.reg.access = reg_wr(ibwr)->access;
 547                 break;
 548                 default:
 549                         break;
 550                 }
 551         }
 552 }
 553
 554 static int init_send_wqe(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 555                          unsigned int mask, unsigned int length,
 556                          struct rxe_send_wqe *wqe)
 557 {
 558         int num_sge = ibwr->num_sge;
 559         struct ib_sge *sge;
 560         int i;
 561         u8 *p;
 562
 563         init_send_wr(qp, &wqe->wr, ibwr);
 564
 565         if (qp_type(qp) == IB_QPT_UD ||
 566             qp_type(qp) == IB_QPT_SMI ||
 567             qp_type(qp) == IB_QPT_GSI)
 568                 memcpy(&wqe->av, &to_rah(ud_wr(ibwr)->ah)->av, sizeof(wqe->av));
 569
 570         if (unlikely(ibwr->send_flags & IB_SEND_INLINE)) {
 571                 p = wqe->dma.inline_data;
 572
 573                 sge = ibwr->sg_list;
 574                 for (i = 0; i < num_sge; i++, sge++) {
 575                         memcpy(p, (void *)(uintptr_t)sge->addr,
 576                                         sge->length);
 577
 578                         p += sge->length;
 579                 }
 580         } else if (mask & WR_REG_MASK) {
 581                 wqe->mask = mask;
 582                 wqe->state = wqe_state_posted;
 583                 return 0;
 584         } else
 585                 memcpy(wqe->dma.sge, ibwr->sg_list,
 586                        num_sge * sizeof(struct ib_sge));
 587
 588         wqe->iova = mask & WR_ATOMIC_MASK ? atomic_wr(ibwr)->remote_addr :
 589                 mask & WR_READ_OR_WRITE_MASK ? rdma_wr(ibwr)->remote_addr : 0;
 590         wqe->mask               = mask;
 591         wqe->dma.length         = length;
 592         wqe->dma.resid          = length;
 593         wqe->dma.num_sge        = num_sge;
 594         wqe->dma.cur_sge        = 0;
 595         wqe->dma.sge_offset     = 0;
 596         wqe->state              = wqe_state_posted;
 597         wqe->ssn                = atomic_add_return(1, &qp->ssn);
 598
 599         return 0;
 600 }
 601
 602 static int post_one_send(struct rxe_qp *qp, const struct ib_send_wr *ibwr,
 603                          unsigned int mask, u32 length)
 604 {
 605         int err;
 606         struct rxe_sq *sq = &qp->sq;
 607         struct rxe_send_wqe *send_wqe;
 608         unsigned long flags;
 609
 610         err = validate_send_wr(qp, ibwr, mask, length);
 611         if (err)
 612                 return err;
 613
 614         spin_lock_irqsave(&qp->sq.sq_lock, flags);
 615
 616         if (unlikely(queue_full(sq->queue))) {
 617                 err = -ENOMEM;
 618                 goto err1;
 619         }
 620
 621         send_wqe = producer_addr(sq->queue);
 622
 623         err = init_send_wqe(qp, ibwr, mask, length, send_wqe);
 624         if (unlikely(err))
 625                 goto err1;
 626
 627         /*
 628          * make sure all changes to the work queue are
 629          * written before we update the producer pointer
 630          */
 631         smp_wmb();
 632
 633         advance_producer(sq->queue);
 634         spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
 635
 636         return 0;
 637
 638 err1:
 639         spin_unlock_irqrestore(&qp->sq.sq_lock, flags);
 640         return err;
 641 }
 642
 643 static int rxe_post_send_kernel(struct rxe_qp *qp, const struct ib_send_wr *wr,
 644                                 const struct ib_send_wr **bad_wr)
 645 {
 646         int err = 0;
 647         unsigned int mask;
 648         unsigned int length = 0;
 649         int i;
 650         struct ib_send_wr *next;
 651
 652         while (wr) {
 653                 mask = wr_opcode_mask(wr->opcode, qp);
 654                 if (unlikely(!mask)) {
 655                         err = -EINVAL;
 656                         *bad_wr = wr;
 657                         break;
 658                 }
 659
 660                 if (unlikely((wr->send_flags & IB_SEND_INLINE) &&
 661                              !(mask & WR_INLINE_MASK))) {
 662                         err = -EINVAL;
 663                         *bad_wr = wr;
 664                         break;
 665                 }
 666
 667                 next = wr->next;
 668
 669                 length = 0;
 670                 for (i = 0; i < wr->num_sge; i++)
 671                         length += wr->sg_list[i].length;
 672
 673                 err = post_one_send(qp, wr, mask, length);
 674
 675                 if (err) {
 676                         *bad_wr = wr;
 677                         break;
 678                 }
 679                 wr = next;
 680         }
 681
 682         rxe_run_task(&qp->req.task, 1);
 683         if (unlikely(qp->req.state == QP_STATE_ERROR))
 684                 rxe_run_task(&qp->comp.task, 1);
 685
 686         return err;
 687 }
 688
 689 static int rxe_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 690                          const struct ib_send_wr **bad_wr)
 691 {
 692         struct rxe_qp *qp = to_rqp(ibqp);
 693
 694         if (unlikely(!qp->valid)) {
 695                 *bad_wr = wr;
 696                 return -EINVAL;
 697         }
 698
 699         if (unlikely(qp->req.state < QP_STATE_READY)) {
 700                 *bad_wr = wr;
 701                 return -EINVAL;
 702         }
 703
 704         if (qp->is_user) {
 705                 /* Utilize process context to do protocol processing */
 706                 rxe_run_task(&qp->req.task, 0);
 707                 return 0;
 708         } else
 709                 return rxe_post_send_kernel(qp, wr, bad_wr);
 710 }
 711
 712 static int rxe_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
 713                          const struct ib_recv_wr **bad_wr)
 714 {
 715         int err = 0;
 716         struct rxe_qp *qp = to_rqp(ibqp);
 717         struct rxe_rq *rq = &qp->rq;
 718         unsigned long flags;
 719
 720         if (unlikely((qp_state(qp) < IB_QPS_INIT) || !qp->valid)) {
 721                 *bad_wr = wr;
 722                 err = -EINVAL;
 723                 goto err1;
 724         }
 725
 726         if (unlikely(qp->srq)) {
 727                 *bad_wr = wr;
 728                 err = -EINVAL;
 729                 goto err1;
 730         }
 731
 732         spin_lock_irqsave(&rq->producer_lock, flags);
 733
 734         while (wr) {
 735                 err = post_one_recv(rq, wr);
 736                 if (unlikely(err)) {
 737                         *bad_wr = wr;
 738                         break;
 739                 }
 740                 wr = wr->next;
 741         }
 742
 743         spin_unlock_irqrestore(&rq->producer_lock, flags);
 744
 745         if (qp->resp.state == QP_STATE_ERROR)
 746                 rxe_run_task(&qp->resp.task, 1);
 747
 748 err1:
 749         return err;
 750 }
 751
 752 static int rxe_create_cq(struct ib_cq *ibcq, const struct ib_cq_init_attr *attr,
 753                          struct ib_udata *udata)
 754 {
 755         int err;
 756         struct ib_device *dev = ibcq->device;
 757         struct rxe_dev *rxe = to_rdev(dev);
 758         struct rxe_cq *cq = to_rcq(ibcq);
 759         struct rxe_create_cq_resp __user *uresp = NULL;
 760
 761         if (udata) {
 762                 if (udata->outlen < sizeof(*uresp))
 763                         return -EINVAL;
 764                 uresp = udata->outbuf;
 765         }
 766
 767         if (attr->flags)
 768                 return -EINVAL;
 769
 770         err = rxe_cq_chk_attr(rxe, NULL, attr->cqe, attr->comp_vector);
 771         if (err)
 772                 return err;
 773
 774         err = rxe_cq_from_init(rxe, cq, attr->cqe, attr->comp_vector, udata,
 775                                uresp);
 776         if (err)
 777                 return err;
 778
 779         return rxe_add_to_pool(&rxe->cq_pool, &cq->pelem);
 780 }
 781
 782 static int rxe_destroy_cq(struct ib_cq *ibcq, struct ib_udata *udata)
 783 {
 784         struct rxe_cq *cq = to_rcq(ibcq);
 785
 786         rxe_cq_disable(cq);
 787
 788         rxe_drop_ref(cq);
 789         return 0;
 790 }
 791
 792 static int rxe_resize_cq(struct ib_cq *ibcq, int cqe, struct ib_udata *udata)
 793 {
 794         int err;
 795         struct rxe_cq *cq = to_rcq(ibcq);
 796         struct rxe_dev *rxe = to_rdev(ibcq->device);
 797         struct rxe_resize_cq_resp __user *uresp = NULL;
 798
 799         if (udata) {
 800                 if (udata->outlen < sizeof(*uresp))
 801                         return -EINVAL;
 802                 uresp = udata->outbuf;
 803         }
 804
 805         err = rxe_cq_chk_attr(rxe, cq, cqe, 0);
 806         if (err)
 807                 goto err1;
 808
 809         err = rxe_cq_resize_queue(cq, cqe, uresp, udata);
 810         if (err)
 811                 goto err1;
 812
 813         return 0;
 814
 815 err1:
 816         return err;
 817 }
 818
 819 static int rxe_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 820 {
 821         int i;
 822         struct rxe_cq *cq = to_rcq(ibcq);
 823         struct rxe_cqe *cqe;
 824         unsigned long flags;
 825
 826         spin_lock_irqsave(&cq->cq_lock, flags);
 827         for (i = 0; i < num_entries; i++) {
 828                 cqe = queue_head(cq->queue);
 829                 if (!cqe)
 830                         break;
 831
 832                 memcpy(wc++, &cqe->ibwc, sizeof(*wc));
 833                 advance_consumer(cq->queue);
 834         }
 835         spin_unlock_irqrestore(&cq->cq_lock, flags);
 836
 837         return i;
 838 }
 839
 840 static int rxe_peek_cq(struct ib_cq *ibcq, int wc_cnt)
 841 {
 842         struct rxe_cq *cq = to_rcq(ibcq);
 843         int count = queue_count(cq->queue);
 844
 845         return (count > wc_cnt) ? wc_cnt : count;
 846 }
 847
 848 static int rxe_req_notify_cq(struct ib_cq *ibcq, enum ib_cq_notify_flags flags)
 849 {
 850         struct rxe_cq *cq = to_rcq(ibcq);
 851         unsigned long irq_flags;
 852         int ret = 0;
 853
 854         spin_lock_irqsave(&cq->cq_lock, irq_flags);
 855         if (cq->notify != IB_CQ_NEXT_COMP)
 856                 cq->notify = flags & IB_CQ_SOLICITED_MASK;
 857
 858         if ((flags & IB_CQ_REPORT_MISSED_EVENTS) && !queue_empty(cq->queue))
 859                 ret = 1;
 860
 861         spin_unlock_irqrestore(&cq->cq_lock, irq_flags);
 862
 863         return ret;
 864 }
 865
 866 static struct ib_mr *rxe_get_dma_mr(struct ib_pd *ibpd, int access)
 867 {
 868         struct rxe_dev *rxe = to_rdev(ibpd->device);
 869         struct rxe_pd *pd = to_rpd(ibpd);
 870         struct rxe_mem *mr;
 871
 872         mr = rxe_alloc(&rxe->mr_pool);
 873         if (!mr)
 874                 return ERR_PTR(-ENOMEM);
 875
 876         rxe_add_index(mr);
 877         rxe_add_ref(pd);
 878         rxe_mem_init_dma(pd, access, mr);
 879
 880         return &mr->ibmr;
 881 }
 882
 883 static struct ib_mr *rxe_reg_user_mr(struct ib_pd *ibpd,
 884                                      u64 start,
 885                                      u64 length,
 886                                      u64 iova,
 887                                      int access, struct ib_udata *udata)
 888 {
 889         int err;
 890         struct rxe_dev *rxe = to_rdev(ibpd->device);
 891         struct rxe_pd *pd = to_rpd(ibpd);
 892         struct rxe_mem *mr;
 893
 894         mr = rxe_alloc(&rxe->mr_pool);
 895         if (!mr) {
 896                 err = -ENOMEM;
 897                 goto err2;
 898         }
 899
 900         rxe_add_index(mr);
 901
 902         rxe_add_ref(pd);
 903
 904         err = rxe_mem_init_user(pd, start, length, iova,
 905                                 access, udata, mr);
 906         if (err)
 907                 goto err3;
 908
 909         return &mr->ibmr;
 910
 911 err3:
 912         rxe_drop_ref(pd);
 913         rxe_drop_index(mr);
 914         rxe_drop_ref(mr);
 915 err2:
 916         return ERR_PTR(err);
 917 }
 918
 919 static int rxe_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 920 {
 921         struct rxe_mem *mr = to_rmr(ibmr);
 922
 923         mr->state = RXE_MEM_STATE_ZOMBIE;
 924         rxe_drop_ref(mr_pd(mr));
 925         rxe_drop_index(mr);
 926         rxe_drop_ref(mr);
 927         return 0;
 928 }
 929
 930 static struct ib_mr *rxe_alloc_mr(struct ib_pd *ibpd, enum ib_mr_type mr_type,
 931                                   u32 max_num_sg)
 932 {
 933         struct rxe_dev *rxe = to_rdev(ibpd->device);
 934         struct rxe_pd *pd = to_rpd(ibpd);
 935         struct rxe_mem *mr;
 936         int err;
 937
 938         if (mr_type != IB_MR_TYPE_MEM_REG)
 939                 return ERR_PTR(-EINVAL);
 940
 941         mr = rxe_alloc(&rxe->mr_pool);
 942         if (!mr) {
 943                 err = -ENOMEM;
 944                 goto err1;
 945         }
 946
 947         rxe_add_index(mr);
 948
 949         rxe_add_ref(pd);
 950
 951         err = rxe_mem_init_fast(pd, max_num_sg, mr);
 952         if (err)
 953                 goto err2;
 954
 955         return &mr->ibmr;
 956
 957 err2:
 958         rxe_drop_ref(pd);
 959         rxe_drop_index(mr);
 960         rxe_drop_ref(mr);
 961 err1:
 962         return ERR_PTR(err);
 963 }
 964
 965 static int rxe_set_page(struct ib_mr *ibmr, u64 addr)
 966 {
 967         struct rxe_mem *mr = to_rmr(ibmr);
 968         struct rxe_map *map;
 969         struct rxe_phys_buf *buf;
 970
 971         if (unlikely(mr->nbuf == mr->num_buf))
 972                 return -ENOMEM;
 973
 974         map = mr->map[mr->nbuf / RXE_BUF_PER_MAP];
 975         buf = &map->buf[mr->nbuf % RXE_BUF_PER_MAP];
 976
 977         buf->addr = addr;
 978         buf->size = ibmr->page_size;
 979         mr->nbuf++;
 980
 981         return 0;
 982 }
 983
 984 static int rxe_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg,
 985                          int sg_nents, unsigned int *sg_offset)
 986 {
 987         struct rxe_mem *mr = to_rmr(ibmr);
 988         int n;
 989
 990         mr->nbuf = 0;
 991
 992         n = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset, rxe_set_page);
 993
 994         mr->va = ibmr->iova;
 995         mr->iova = ibmr->iova;
 996         mr->length = ibmr->length;
 997         mr->page_shift = ilog2(ibmr->page_size);
 998         mr->page_mask = ibmr->page_size - 1;
 999         mr->offset = mr->iova & mr->page_mask;
1000
1001         return n;
1002 }
1003
1004 static int rxe_attach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1005 {
1006         int err;
1007         struct rxe_dev *rxe = to_rdev(ibqp->device);
1008         struct rxe_qp *qp = to_rqp(ibqp);
1009         struct rxe_mc_grp *grp;
1010
1011         /* takes a ref on grp if successful */
1012         err = rxe_mcast_get_grp(rxe, mgid, &grp);
1013         if (err)
1014                 return err;
1015
1016         err = rxe_mcast_add_grp_elem(rxe, qp, grp);
1017
1018         rxe_drop_ref(grp);
1019         return err;
1020 }
1021
1022 static int rxe_detach_mcast(struct ib_qp *ibqp, union ib_gid *mgid, u16 mlid)
1023 {
1024         struct rxe_dev *rxe = to_rdev(ibqp->device);
1025         struct rxe_qp *qp = to_rqp(ibqp);
1026
1027         return rxe_mcast_drop_grp_elem(rxe, qp, mgid);
1028 }
1029
1030 static ssize_t parent_show(struct device *device,
1031                            struct device_attribute *attr, char *buf)
1032 {
1033         struct rxe_dev *rxe =
1034                 rdma_device_to_drv_device(device, struct rxe_dev, ib_dev);
1035
1036         return scnprintf(buf, PAGE_SIZE, "%s\n", rxe_parent_name(rxe, 1));
1037 }
1038
1039 static DEVICE_ATTR_RO(parent);
1040
1041 static struct attribute *rxe_dev_attributes[] = {
1042         &dev_attr_parent.attr,
1043         NULL
1044 };
1045
1046 static const struct attribute_group rxe_attr_group = {
1047         .attrs = rxe_dev_attributes,
1048 };
1049
1050 static int rxe_enable_driver(struct ib_device *ib_dev)
1051 {
1052         struct rxe_dev *rxe = container_of(ib_dev, struct rxe_dev, ib_dev);
1053
1054         rxe_set_port_state(rxe);
1055         dev_info(&rxe->ib_dev.dev, "added %s\n", netdev_name(rxe->ndev));
1056         return 0;
1057 }
1058
1059 static const struct ib_device_ops rxe_dev_ops = {
1060         .owner = THIS_MODULE,
1061         .driver_id = RDMA_DRIVER_RXE,
1062         .uverbs_abi_ver = RXE_UVERBS_ABI_VERSION,
1063
1064         .alloc_hw_stats = rxe_ib_alloc_hw_stats,
1065         .alloc_mr = rxe_alloc_mr,
1066         .alloc_pd = rxe_alloc_pd,
1067         .alloc_ucontext = rxe_alloc_ucontext,
1068         .attach_mcast = rxe_attach_mcast,
1069         .create_ah = rxe_create_ah,
1070         .create_cq = rxe_create_cq,
1071         .create_qp = rxe_create_qp,
1072         .create_srq = rxe_create_srq,
1073         .dealloc_driver = rxe_dealloc,
1074         .dealloc_pd = rxe_dealloc_pd,
1075         .dealloc_ucontext = rxe_dealloc_ucontext,
1076         .dereg_mr = rxe_dereg_mr,
1077         .destroy_ah = rxe_destroy_ah,
1078         .destroy_cq = rxe_destroy_cq,
1079         .destroy_qp = rxe_destroy_qp,
1080         .destroy_srq = rxe_destroy_srq,
1081         .detach_mcast = rxe_detach_mcast,
1082         .enable_driver = rxe_enable_driver,
1083         .get_dma_mr = rxe_get_dma_mr,
1084         .get_hw_stats = rxe_ib_get_hw_stats,
1085         .get_link_layer = rxe_get_link_layer,
1086         .get_port_immutable = rxe_port_immutable,
1087         .map_mr_sg = rxe_map_mr_sg,
1088         .mmap = rxe_mmap,
1089         .modify_ah = rxe_modify_ah,
1090         .modify_device = rxe_modify_device,
1091         .modify_port = rxe_modify_port,
1092         .modify_qp = rxe_modify_qp,
1093         .modify_srq = rxe_modify_srq,
1094         .peek_cq = rxe_peek_cq,
1095         .poll_cq = rxe_poll_cq,
1096         .post_recv = rxe_post_recv,
1097         .post_send = rxe_post_send,
1098         .post_srq_recv = rxe_post_srq_recv,
1099         .query_ah = rxe_query_ah,
1100         .query_device = rxe_query_device,
1101         .query_pkey = rxe_query_pkey,
1102         .query_port = rxe_query_port,
1103         .query_qp = rxe_query_qp,
1104         .query_srq = rxe_query_srq,
1105         .reg_user_mr = rxe_reg_user_mr,
1106         .req_notify_cq = rxe_req_notify_cq,
1107         .resize_cq = rxe_resize_cq,
1108
1109         INIT_RDMA_OBJ_SIZE(ib_ah, rxe_ah, ibah),
1110         INIT_RDMA_OBJ_SIZE(ib_cq, rxe_cq, ibcq),
1111         INIT_RDMA_OBJ_SIZE(ib_pd, rxe_pd, ibpd),
1112         INIT_RDMA_OBJ_SIZE(ib_srq, rxe_srq, ibsrq),
1113         INIT_RDMA_OBJ_SIZE(ib_ucontext, rxe_ucontext, ibuc),
1114 };
1115
1116 int rxe_register_device(struct rxe_dev *rxe, const char *ibdev_name)
1117 {
1118         int err;
1119         struct ib_device *dev = &rxe->ib_dev;
1120         struct crypto_shash *tfm;
1121         u64 dma_mask;
1122
1123         strlcpy(dev->node_desc, "rxe", sizeof(dev->node_desc));
1124
1125         dev->node_type = RDMA_NODE_IB_CA;
1126         dev->phys_port_cnt = 1;
1127         dev->num_comp_vectors = num_possible_cpus();
1128         dev->dev.parent = rxe_dma_device(rxe);
1129         dev->local_dma_lkey = 0;
1130         addrconf_addr_eui48((unsigned char *)&dev->node_guid,
1131                             rxe->ndev->dev_addr);
1132         dev->dev.dma_parms = &rxe->dma_parms;
1133         dma_set_max_seg_size(&dev->dev, UINT_MAX);
1134         dma_mask = IS_ENABLED(CONFIG_64BIT) ? DMA_BIT_MASK(64) : DMA_BIT_MASK(32);
1135         err = dma_coerce_mask_and_coherent(&dev->dev, dma_mask);
1136         if (err)
1137                 return err;
1138
1139         dev->uverbs_cmd_mask = BIT_ULL(IB_USER_VERBS_CMD_GET_CONTEXT)
1140             | BIT_ULL(IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)
1141             | BIT_ULL(IB_USER_VERBS_CMD_QUERY_DEVICE)
1142             | BIT_ULL(IB_USER_VERBS_CMD_QUERY_PORT)
1143             | BIT_ULL(IB_USER_VERBS_CMD_ALLOC_PD)
1144             | BIT_ULL(IB_USER_VERBS_CMD_DEALLOC_PD)
1145             | BIT_ULL(IB_USER_VERBS_CMD_CREATE_SRQ)
1146             | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_SRQ)
1147             | BIT_ULL(IB_USER_VERBS_CMD_QUERY_SRQ)
1148             | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_SRQ)
1149             | BIT_ULL(IB_USER_VERBS_CMD_POST_SRQ_RECV)
1150             | BIT_ULL(IB_USER_VERBS_CMD_CREATE_QP)
1151             | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_QP)
1152             | BIT_ULL(IB_USER_VERBS_CMD_QUERY_QP)
1153             | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_QP)
1154             | BIT_ULL(IB_USER_VERBS_CMD_POST_SEND)
1155             | BIT_ULL(IB_USER_VERBS_CMD_POST_RECV)
1156             | BIT_ULL(IB_USER_VERBS_CMD_CREATE_CQ)
1157             | BIT_ULL(IB_USER_VERBS_CMD_RESIZE_CQ)
1158             | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_CQ)
1159             | BIT_ULL(IB_USER_VERBS_CMD_POLL_CQ)
1160             | BIT_ULL(IB_USER_VERBS_CMD_PEEK_CQ)
1161             | BIT_ULL(IB_USER_VERBS_CMD_REQ_NOTIFY_CQ)
1162             | BIT_ULL(IB_USER_VERBS_CMD_REG_MR)
1163             | BIT_ULL(IB_USER_VERBS_CMD_DEREG_MR)
1164             | BIT_ULL(IB_USER_VERBS_CMD_CREATE_AH)
1165             | BIT_ULL(IB_USER_VERBS_CMD_MODIFY_AH)
1166             | BIT_ULL(IB_USER_VERBS_CMD_QUERY_AH)
1167             | BIT_ULL(IB_USER_VERBS_CMD_DESTROY_AH)
1168             | BIT_ULL(IB_USER_VERBS_CMD_ATTACH_MCAST)
1169             | BIT_ULL(IB_USER_VERBS_CMD_DETACH_MCAST)
1170             ;
1171
1172         ib_set_device_ops(dev, &rxe_dev_ops);
1173         err = ib_device_set_netdev(&rxe->ib_dev, rxe->ndev, 1);
1174         if (err)
1175                 return err;
1176
1177         tfm = crypto_alloc_shash("crc32", 0, 0);
1178         if (IS_ERR(tfm)) {
1179                 pr_err("failed to allocate crc algorithm err:%ld\n",
1180                        PTR_ERR(tfm));
1181                 return PTR_ERR(tfm);
1182         }
1183         rxe->tfm = tfm;
1184
1185         rdma_set_device_sysfs_group(dev, &rxe_attr_group);
1186         err = ib_register_device(dev, ibdev_name, NULL);
1187         if (err)
1188                 pr_warn("%s failed with error %d\n", __func__, err);
1189
1190         /*
1191          * Note that rxe may be invalid at this point if another thread
1192          * unregistered it.
1193          */
1194         return err;
1195 }