drivers/infiniband/sw/rxe/rxe_req.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/skbuff.h>
   8 #include <crypto/hash.h>
   9
  10 #include "rxe.h"
  11 #include "rxe_loc.h"
  12 #include "rxe_queue.h"
  13
  14 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
  15                        u32 opcode);
  16
  17 static inline void retry_first_write_send(struct rxe_qp *qp,
  18                                           struct rxe_send_wqe *wqe, int npsn)
  19 {
  20         int i;
  21
  22         for (i = 0; i < npsn; i++) {
  23                 int to_send = (wqe->dma.resid > qp->mtu) ?
  24                                 qp->mtu : wqe->dma.resid;
  25
  26                 qp->req.opcode = next_opcode(qp, wqe,
  27                                              wqe->wr.opcode);
  28
  29                 if (wqe->wr.send_flags & IB_SEND_INLINE) {
  30                         wqe->dma.resid -= to_send;
  31                         wqe->dma.sge_offset += to_send;
  32                 } else {
  33                         advance_dma_data(&wqe->dma, to_send);
  34                 }
  35         }
  36 }
  37
  38 static void req_retry(struct rxe_qp *qp)
  39 {
  40         struct rxe_send_wqe *wqe;
  41         unsigned int wqe_index;
  42         unsigned int mask;
  43         int npsn;
  44         int first = 1;
  45         struct rxe_queue *q = qp->sq.queue;
  46         unsigned int cons;
  47         unsigned int prod;
  48
  49         cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
  50         prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
  51
  52         qp->req.wqe_index       = cons;
  53         qp->req.psn             = qp->comp.psn;
  54         qp->req.opcode          = -1;
  55
  56         for (wqe_index = cons; wqe_index != prod;
  57                         wqe_index = queue_next_index(q, wqe_index)) {
  58                 wqe = queue_addr_from_index(qp->sq.queue, wqe_index);
  59                 mask = wr_opcode_mask(wqe->wr.opcode, qp);
  60
  61                 if (wqe->state == wqe_state_posted)
  62                         break;
  63
  64                 if (wqe->state == wqe_state_done)
  65                         continue;
  66
  67                 wqe->iova = (mask & WR_ATOMIC_MASK) ?
  68                              wqe->wr.wr.atomic.remote_addr :
  69                              (mask & WR_READ_OR_WRITE_MASK) ?
  70                              wqe->wr.wr.rdma.remote_addr :
  71                              0;
  72
  73                 if (!first || (mask & WR_READ_MASK) == 0) {
  74                         wqe->dma.resid = wqe->dma.length;
  75                         wqe->dma.cur_sge = 0;
  76                         wqe->dma.sge_offset = 0;
  77                 }
  78
  79                 if (first) {
  80                         first = 0;
  81
  82                         if (mask & WR_WRITE_OR_SEND_MASK) {
  83                                 npsn = (qp->comp.psn - wqe->first_psn) &
  84                                         BTH_PSN_MASK;
  85                                 retry_first_write_send(qp, wqe, npsn);
  86                         }
  87
  88                         if (mask & WR_READ_MASK) {
  89                                 npsn = (wqe->dma.length - wqe->dma.resid) /
  90                                         qp->mtu;
  91                                 wqe->iova += npsn * qp->mtu;
  92                         }
  93                 }
  94
  95                 wqe->state = wqe_state_posted;
  96         }
  97 }
  98
  99 void rnr_nak_timer(struct timer_list *t)
 100 {
 101         struct rxe_qp *qp = from_timer(qp, t, rnr_nak_timer);
 102
 103         rxe_dbg_qp(qp, "nak timer fired\n");
 104
 105         /* request a send queue retry */
 106         qp->req.need_retry = 1;
 107         qp->req.wait_for_rnr_timer = 0;
 108         rxe_sched_task(&qp->req.task);
 109 }
 110
 111 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
 112 {
 113         struct rxe_send_wqe *wqe;
 114         struct rxe_queue *q = qp->sq.queue;
 115         unsigned int index = qp->req.wqe_index;
 116         unsigned int cons;
 117         unsigned int prod;
 118
 119         wqe = queue_head(q, QUEUE_TYPE_FROM_CLIENT);
 120         cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
 121         prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
 122
 123         if (unlikely(qp->req.state == QP_STATE_DRAIN)) {
 124                 /* check to see if we are drained;
 125                  * state_lock used by requester and completer
 126                  */
 127                 spin_lock_bh(&qp->state_lock);
 128                 do {
 129                         if (qp->req.state != QP_STATE_DRAIN) {
 130                                 /* comp just finished */
 131                                 spin_unlock_bh(&qp->state_lock);
 132                                 break;
 133                         }
 134
 135                         if (wqe && ((index != cons) ||
 136                                 (wqe->state != wqe_state_posted))) {
 137                                 /* comp not done yet */
 138                                 spin_unlock_bh(&qp->state_lock);
 139                                 break;
 140                         }
 141
 142                         qp->req.state = QP_STATE_DRAINED;
 143                         spin_unlock_bh(&qp->state_lock);
 144
 145                         if (qp->ibqp.event_handler) {
 146                                 struct ib_event ev;
 147
 148                                 ev.device = qp->ibqp.device;
 149                                 ev.element.qp = &qp->ibqp;
 150                                 ev.event = IB_EVENT_SQ_DRAINED;
 151                                 qp->ibqp.event_handler(&ev,
 152                                         qp->ibqp.qp_context);
 153                         }
 154                 } while (0);
 155         }
 156
 157         if (index == prod)
 158                 return NULL;
 159
 160         wqe = queue_addr_from_index(q, index);
 161
 162         if (unlikely((qp->req.state == QP_STATE_DRAIN ||
 163                       qp->req.state == QP_STATE_DRAINED) &&
 164                      (wqe->state != wqe_state_processing)))
 165                 return NULL;
 166
 167         wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
 168         return wqe;
 169 }
 170
 171 /**
 172  * rxe_wqe_is_fenced - check if next wqe is fenced
 173  * @qp: the queue pair
 174  * @wqe: the next wqe
 175  *
 176  * Returns: 1 if wqe needs to wait
 177  *          0 if wqe is ready to go
 178  */
 179 static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 180 {
 181         /* Local invalidate fence (LIF) see IBA 10.6.5.1
 182          * Requires ALL previous operations on the send queue
 183          * are complete. Make mandatory for the rxe driver.
 184          */
 185         if (wqe->wr.opcode == IB_WR_LOCAL_INV)
 186                 return qp->req.wqe_index != queue_get_consumer(qp->sq.queue,
 187                                                 QUEUE_TYPE_FROM_CLIENT);
 188
 189         /* Fence see IBA 10.8.3.3
 190          * Requires that all previous read and atomic operations
 191          * are complete.
 192          */
 193         return (wqe->wr.send_flags & IB_SEND_FENCE) &&
 194                 atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic;
 195 }
 196
 197 static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
 198 {
 199         switch (opcode) {
 200         case IB_WR_RDMA_WRITE:
 201                 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
 202                     qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
 203                         return fits ?
 204                                 IB_OPCODE_RC_RDMA_WRITE_LAST :
 205                                 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
 206                 else
 207                         return fits ?
 208                                 IB_OPCODE_RC_RDMA_WRITE_ONLY :
 209                                 IB_OPCODE_RC_RDMA_WRITE_FIRST;
 210
 211         case IB_WR_RDMA_WRITE_WITH_IMM:
 212                 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
 213                     qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
 214                         return fits ?
 215                                 IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
 216                                 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
 217                 else
 218                         return fits ?
 219                                 IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
 220                                 IB_OPCODE_RC_RDMA_WRITE_FIRST;
 221
 222         case IB_WR_SEND:
 223                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 224                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 225                         return fits ?
 226                                 IB_OPCODE_RC_SEND_LAST :
 227                                 IB_OPCODE_RC_SEND_MIDDLE;
 228                 else
 229                         return fits ?
 230                                 IB_OPCODE_RC_SEND_ONLY :
 231                                 IB_OPCODE_RC_SEND_FIRST;
 232
 233         case IB_WR_SEND_WITH_IMM:
 234                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 235                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 236                         return fits ?
 237                                 IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
 238                                 IB_OPCODE_RC_SEND_MIDDLE;
 239                 else
 240                         return fits ?
 241                                 IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
 242                                 IB_OPCODE_RC_SEND_FIRST;
 243
 244         case IB_WR_RDMA_READ:
 245                 return IB_OPCODE_RC_RDMA_READ_REQUEST;
 246
 247         case IB_WR_ATOMIC_CMP_AND_SWP:
 248                 return IB_OPCODE_RC_COMPARE_SWAP;
 249
 250         case IB_WR_ATOMIC_FETCH_AND_ADD:
 251                 return IB_OPCODE_RC_FETCH_ADD;
 252
 253         case IB_WR_SEND_WITH_INV:
 254                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 255                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 256                         return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
 257                                 IB_OPCODE_RC_SEND_MIDDLE;
 258                 else
 259                         return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
 260                                 IB_OPCODE_RC_SEND_FIRST;
 261
 262         case IB_WR_ATOMIC_WRITE:
 263                 return IB_OPCODE_RC_ATOMIC_WRITE;
 264
 265         case IB_WR_REG_MR:
 266         case IB_WR_LOCAL_INV:
 267                 return opcode;
 268         }
 269
 270         return -EINVAL;
 271 }
 272
 273 static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits)
 274 {
 275         switch (opcode) {
 276         case IB_WR_RDMA_WRITE:
 277                 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
 278                     qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
 279                         return fits ?
 280                                 IB_OPCODE_UC_RDMA_WRITE_LAST :
 281                                 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
 282                 else
 283                         return fits ?
 284                                 IB_OPCODE_UC_RDMA_WRITE_ONLY :
 285                                 IB_OPCODE_UC_RDMA_WRITE_FIRST;
 286
 287         case IB_WR_RDMA_WRITE_WITH_IMM:
 288                 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
 289                     qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
 290                         return fits ?
 291                                 IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
 292                                 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
 293                 else
 294                         return fits ?
 295                                 IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
 296                                 IB_OPCODE_UC_RDMA_WRITE_FIRST;
 297
 298         case IB_WR_SEND:
 299                 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
 300                     qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
 301                         return fits ?
 302                                 IB_OPCODE_UC_SEND_LAST :
 303                                 IB_OPCODE_UC_SEND_MIDDLE;
 304                 else
 305                         return fits ?
 306                                 IB_OPCODE_UC_SEND_ONLY :
 307                                 IB_OPCODE_UC_SEND_FIRST;
 308
 309         case IB_WR_SEND_WITH_IMM:
 310                 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
 311                     qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
 312                         return fits ?
 313                                 IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
 314                                 IB_OPCODE_UC_SEND_MIDDLE;
 315                 else
 316                         return fits ?
 317                                 IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
 318                                 IB_OPCODE_UC_SEND_FIRST;
 319         }
 320
 321         return -EINVAL;
 322 }
 323
 324 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
 325                        u32 opcode)
 326 {
 327         int fits = (wqe->dma.resid <= qp->mtu);
 328
 329         switch (qp_type(qp)) {
 330         case IB_QPT_RC:
 331                 return next_opcode_rc(qp, opcode, fits);
 332
 333         case IB_QPT_UC:
 334                 return next_opcode_uc(qp, opcode, fits);
 335
 336         case IB_QPT_UD:
 337         case IB_QPT_GSI:
 338                 switch (opcode) {
 339                 case IB_WR_SEND:
 340                         return IB_OPCODE_UD_SEND_ONLY;
 341
 342                 case IB_WR_SEND_WITH_IMM:
 343                         return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
 344                 }
 345                 break;
 346
 347         default:
 348                 break;
 349         }
 350
 351         return -EINVAL;
 352 }
 353
 354 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 355 {
 356         int depth;
 357
 358         if (wqe->has_rd_atomic)
 359                 return 0;
 360
 361         qp->req.need_rd_atomic = 1;
 362         depth = atomic_dec_return(&qp->req.rd_atomic);
 363
 364         if (depth >= 0) {
 365                 qp->req.need_rd_atomic = 0;
 366                 wqe->has_rd_atomic = 1;
 367                 return 0;
 368         }
 369
 370         atomic_inc(&qp->req.rd_atomic);
 371         return -EAGAIN;
 372 }
 373
 374 static inline int get_mtu(struct rxe_qp *qp)
 375 {
 376         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 377
 378         if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
 379                 return qp->mtu;
 380
 381         return rxe->port.mtu_cap;
 382 }
 383
 384 static struct sk_buff *init_req_packet(struct rxe_qp *qp,
 385                                        struct rxe_av *av,
 386                                        struct rxe_send_wqe *wqe,
 387                                        int opcode, u32 payload,
 388                                        struct rxe_pkt_info *pkt)
 389 {
 390         struct rxe_dev          *rxe = to_rdev(qp->ibqp.device);
 391         struct sk_buff          *skb;
 392         struct rxe_send_wr      *ibwr = &wqe->wr;
 393         int                     pad = (-payload) & 0x3;
 394         int                     paylen;
 395         int                     solicited;
 396         u32                     qp_num;
 397         int                     ack_req;
 398
 399         /* length from start of bth to end of icrc */
 400         paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
 401         pkt->paylen = paylen;
 402
 403         /* init skb */
 404         skb = rxe_init_packet(rxe, av, paylen, pkt);
 405         if (unlikely(!skb))
 406                 return NULL;
 407
 408         /* init bth */
 409         solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
 410                         (pkt->mask & RXE_END_MASK) &&
 411                         ((pkt->mask & (RXE_SEND_MASK)) ||
 412                         (pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
 413                         (RXE_WRITE_MASK | RXE_IMMDT_MASK));
 414
 415         qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
 416                                          qp->attr.dest_qp_num;
 417
 418         ack_req = ((pkt->mask & RXE_END_MASK) ||
 419                 (qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
 420         if (ack_req)
 421                 qp->req.noack_pkts = 0;
 422
 423         bth_init(pkt, pkt->opcode, solicited, 0, pad, IB_DEFAULT_PKEY_FULL, qp_num,
 424                  ack_req, pkt->psn);
 425
 426         /* init optional headers */
 427         if (pkt->mask & RXE_RETH_MASK) {
 428                 reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
 429                 reth_set_va(pkt, wqe->iova);
 430                 reth_set_len(pkt, wqe->dma.resid);
 431         }
 432
 433         if (pkt->mask & RXE_IMMDT_MASK)
 434                 immdt_set_imm(pkt, ibwr->ex.imm_data);
 435
 436         if (pkt->mask & RXE_IETH_MASK)
 437                 ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
 438
 439         if (pkt->mask & RXE_ATMETH_MASK) {
 440                 atmeth_set_va(pkt, wqe->iova);
 441                 if (opcode == IB_OPCODE_RC_COMPARE_SWAP) {
 442                         atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
 443                         atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
 444                 } else {
 445                         atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
 446                 }
 447                 atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
 448         }
 449
 450         if (pkt->mask & RXE_DETH_MASK) {
 451                 if (qp->ibqp.qp_num == 1)
 452                         deth_set_qkey(pkt, GSI_QKEY);
 453                 else
 454                         deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
 455                 deth_set_sqp(pkt, qp->ibqp.qp_num);
 456         }
 457
 458         return skb;
 459 }
 460
 461 static int finish_packet(struct rxe_qp *qp, struct rxe_av *av,
 462                          struct rxe_send_wqe *wqe, struct rxe_pkt_info *pkt,
 463                          struct sk_buff *skb, u32 payload)
 464 {
 465         int err;
 466
 467         err = rxe_prepare(av, pkt, skb);
 468         if (err)
 469                 return err;
 470
 471         if (pkt->mask & RXE_WRITE_OR_SEND_MASK) {
 472                 if (wqe->wr.send_flags & IB_SEND_INLINE) {
 473                         u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
 474
 475                         memcpy(payload_addr(pkt), tmp, payload);
 476
 477                         wqe->dma.resid -= payload;
 478                         wqe->dma.sge_offset += payload;
 479                 } else {
 480                         err = copy_data(qp->pd, 0, &wqe->dma,
 481                                         payload_addr(pkt), payload,
 482                                         RXE_FROM_MR_OBJ);
 483                         if (err)
 484                                 return err;
 485                 }
 486                 if (bth_pad(pkt)) {
 487                         u8 *pad = payload_addr(pkt) + payload;
 488
 489                         memset(pad, 0, bth_pad(pkt));
 490                 }
 491         }
 492
 493         if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
 494                 memcpy(payload_addr(pkt), wqe->dma.atomic_wr, payload);
 495                 wqe->dma.resid -= payload;
 496         }
 497
 498         return 0;
 499 }
 500
 501 static void update_wqe_state(struct rxe_qp *qp,
 502                 struct rxe_send_wqe *wqe,
 503                 struct rxe_pkt_info *pkt)
 504 {
 505         if (pkt->mask & RXE_END_MASK) {
 506                 if (qp_type(qp) == IB_QPT_RC)
 507                         wqe->state = wqe_state_pending;
 508         } else {
 509                 wqe->state = wqe_state_processing;
 510         }
 511 }
 512
 513 static void update_wqe_psn(struct rxe_qp *qp,
 514                            struct rxe_send_wqe *wqe,
 515                            struct rxe_pkt_info *pkt,
 516                            u32 payload)
 517 {
 518         /* number of packets left to send including current one */
 519         int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
 520
 521         /* handle zero length packet case */
 522         if (num_pkt == 0)
 523                 num_pkt = 1;
 524
 525         if (pkt->mask & RXE_START_MASK) {
 526                 wqe->first_psn = qp->req.psn;
 527                 wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
 528         }
 529
 530         if (pkt->mask & RXE_READ_MASK)
 531                 qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
 532         else
 533                 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
 534 }
 535
 536 static void save_state(struct rxe_send_wqe *wqe,
 537                        struct rxe_qp *qp,
 538                        struct rxe_send_wqe *rollback_wqe,
 539                        u32 *rollback_psn)
 540 {
 541         rollback_wqe->state     = wqe->state;
 542         rollback_wqe->first_psn = wqe->first_psn;
 543         rollback_wqe->last_psn  = wqe->last_psn;
 544         *rollback_psn           = qp->req.psn;
 545 }
 546
 547 static void rollback_state(struct rxe_send_wqe *wqe,
 548                            struct rxe_qp *qp,
 549                            struct rxe_send_wqe *rollback_wqe,
 550                            u32 rollback_psn)
 551 {
 552         wqe->state     = rollback_wqe->state;
 553         wqe->first_psn = rollback_wqe->first_psn;
 554         wqe->last_psn  = rollback_wqe->last_psn;
 555         qp->req.psn    = rollback_psn;
 556 }
 557
 558 static void update_state(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 559 {
 560         qp->req.opcode = pkt->opcode;
 561
 562         if (pkt->mask & RXE_END_MASK)
 563                 qp->req.wqe_index = queue_next_index(qp->sq.queue,
 564                                                      qp->req.wqe_index);
 565
 566         qp->need_req_skb = 0;
 567
 568         if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
 569                 mod_timer(&qp->retrans_timer,
 570                           jiffies + qp->qp_timeout_jiffies);
 571 }
 572
 573 static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 574 {
 575         u8 opcode = wqe->wr.opcode;
 576         u32 rkey;
 577         int ret;
 578
 579         switch (opcode) {
 580         case IB_WR_LOCAL_INV:
 581                 rkey = wqe->wr.ex.invalidate_rkey;
 582                 if (rkey_is_mw(rkey))
 583                         ret = rxe_invalidate_mw(qp, rkey);
 584                 else
 585                         ret = rxe_invalidate_mr(qp, rkey);
 586
 587                 if (unlikely(ret)) {
 588                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 589                         return ret;
 590                 }
 591                 break;
 592         case IB_WR_REG_MR:
 593                 ret = rxe_reg_fast_mr(qp, wqe);
 594                 if (unlikely(ret)) {
 595                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 596                         return ret;
 597                 }
 598                 break;
 599         case IB_WR_BIND_MW:
 600                 ret = rxe_bind_mw(qp, wqe);
 601                 if (unlikely(ret)) {
 602                         wqe->status = IB_WC_MW_BIND_ERR;
 603                         return ret;
 604                 }
 605                 break;
 606         default:
 607                 rxe_dbg_qp(qp, "Unexpected send wqe opcode %d\n", opcode);
 608                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 609                 return -EINVAL;
 610         }
 611
 612         wqe->state = wqe_state_done;
 613         wqe->status = IB_WC_SUCCESS;
 614         qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
 615
 616         /* There is no ack coming for local work requests
 617          * which can lead to a deadlock. So go ahead and complete
 618          * it now.
 619          */
 620         rxe_sched_task(&qp->comp.task);
 621
 622         return 0;
 623 }
 624
 625 int rxe_requester(void *arg)
 626 {
 627         struct rxe_qp *qp = (struct rxe_qp *)arg;
 628         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 629         struct rxe_pkt_info pkt;
 630         struct sk_buff *skb;
 631         struct rxe_send_wqe *wqe;
 632         enum rxe_hdr_mask mask;
 633         u32 payload;
 634         int mtu;
 635         int opcode;
 636         int err;
 637         int ret;
 638         struct rxe_send_wqe rollback_wqe;
 639         u32 rollback_psn;
 640         struct rxe_queue *q = qp->sq.queue;
 641         struct rxe_ah *ah;
 642         struct rxe_av *av;
 643
 644         if (!rxe_get(qp))
 645                 return -EAGAIN;
 646
 647         if (unlikely(!qp->valid))
 648                 goto exit;
 649
 650         if (unlikely(qp->req.state == QP_STATE_ERROR)) {
 651                 wqe = req_next_wqe(qp);
 652                 if (wqe)
 653                         /*
 654                          * Generate an error completion for error qp state
 655                          */
 656                         goto err;
 657                 else
 658                         goto exit;
 659         }
 660
 661         if (unlikely(qp->req.state == QP_STATE_RESET)) {
 662                 qp->req.wqe_index = queue_get_consumer(q,
 663                                                 QUEUE_TYPE_FROM_CLIENT);
 664                 qp->req.opcode = -1;
 665                 qp->req.need_rd_atomic = 0;
 666                 qp->req.wait_psn = 0;
 667                 qp->req.need_retry = 0;
 668                 qp->req.wait_for_rnr_timer = 0;
 669                 goto exit;
 670         }
 671
 672         /* we come here if the retransmit timer has fired
 673          * or if the rnr timer has fired. If the retransmit
 674          * timer fires while we are processing an RNR NAK wait
 675          * until the rnr timer has fired before starting the
 676          * retry flow
 677          */
 678         if (unlikely(qp->req.need_retry && !qp->req.wait_for_rnr_timer)) {
 679                 req_retry(qp);
 680                 qp->req.need_retry = 0;
 681         }
 682
 683         wqe = req_next_wqe(qp);
 684         if (unlikely(!wqe))
 685                 goto exit;
 686
 687         if (rxe_wqe_is_fenced(qp, wqe)) {
 688                 qp->req.wait_fence = 1;
 689                 goto exit;
 690         }
 691
 692         if (wqe->mask & WR_LOCAL_OP_MASK) {
 693                 err = rxe_do_local_ops(qp, wqe);
 694                 if (unlikely(err))
 695                         goto err;
 696                 else
 697                         goto done;
 698         }
 699
 700         if (unlikely(qp_type(qp) == IB_QPT_RC &&
 701                 psn_compare(qp->req.psn, (qp->comp.psn +
 702                                 RXE_MAX_UNACKED_PSNS)) > 0)) {
 703                 qp->req.wait_psn = 1;
 704                 goto exit;
 705         }
 706
 707         /* Limit the number of inflight SKBs per QP */
 708         if (unlikely(atomic_read(&qp->skb_out) >
 709                      RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
 710                 qp->need_req_skb = 1;
 711                 goto exit;
 712         }
 713
 714         opcode = next_opcode(qp, wqe, wqe->wr.opcode);
 715         if (unlikely(opcode < 0)) {
 716                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 717                 goto err;
 718         }
 719
 720         mask = rxe_opcode[opcode].mask;
 721         if (unlikely(mask & (RXE_READ_OR_ATOMIC_MASK |
 722                         RXE_ATOMIC_WRITE_MASK))) {
 723                 if (check_init_depth(qp, wqe))
 724                         goto exit;
 725         }
 726
 727         mtu = get_mtu(qp);
 728         payload = (mask & (RXE_WRITE_OR_SEND_MASK | RXE_ATOMIC_WRITE_MASK)) ?
 729                         wqe->dma.resid : 0;
 730         if (payload > mtu) {
 731                 if (qp_type(qp) == IB_QPT_UD) {
 732                         /* C10-93.1.1: If the total sum of all the buffer lengths specified for a
 733                          * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
 734                          * shall not emit any packets for this message. Further, the CI shall not
 735                          * generate an error due to this condition.
 736                          */
 737
 738                         /* fake a successful UD send */
 739                         wqe->first_psn = qp->req.psn;
 740                         wqe->last_psn = qp->req.psn;
 741                         qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
 742                         qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
 743                         qp->req.wqe_index = queue_next_index(qp->sq.queue,
 744                                                        qp->req.wqe_index);
 745                         wqe->state = wqe_state_done;
 746                         wqe->status = IB_WC_SUCCESS;
 747                         rxe_run_task(&qp->comp.task);
 748                         goto done;
 749                 }
 750                 payload = mtu;
 751         }
 752
 753         pkt.rxe = rxe;
 754         pkt.opcode = opcode;
 755         pkt.qp = qp;
 756         pkt.psn = qp->req.psn;
 757         pkt.mask = rxe_opcode[opcode].mask;
 758         pkt.wqe = wqe;
 759
 760         av = rxe_get_av(&pkt, &ah);
 761         if (unlikely(!av)) {
 762                 rxe_dbg_qp(qp, "Failed no address vector\n");
 763                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 764                 goto err;
 765         }
 766
 767         skb = init_req_packet(qp, av, wqe, opcode, payload, &pkt);
 768         if (unlikely(!skb)) {
 769                 rxe_dbg_qp(qp, "Failed allocating skb\n");
 770                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 771                 if (ah)
 772                         rxe_put(ah);
 773                 goto err;
 774         }
 775
 776         err = finish_packet(qp, av, wqe, &pkt, skb, payload);
 777         if (unlikely(err)) {
 778                 rxe_dbg_qp(qp, "Error during finish packet\n");
 779                 if (err == -EFAULT)
 780                         wqe->status = IB_WC_LOC_PROT_ERR;
 781                 else
 782                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 783                 kfree_skb(skb);
 784                 if (ah)
 785                         rxe_put(ah);
 786                 goto err;
 787         }
 788
 789         if (ah)
 790                 rxe_put(ah);
 791
 792         /*
 793          * To prevent a race on wqe access between requester and completer,
 794          * wqe members state and psn need to be set before calling
 795          * rxe_xmit_packet().
 796          * Otherwise, completer might initiate an unjustified retry flow.
 797          */
 798         save_state(wqe, qp, &rollback_wqe, &rollback_psn);
 799         update_wqe_state(qp, wqe, &pkt);
 800         update_wqe_psn(qp, wqe, &pkt, payload);
 801
 802         err = rxe_xmit_packet(qp, &pkt, skb);
 803         if (err) {
 804                 qp->need_req_skb = 1;
 805
 806                 rollback_state(wqe, qp, &rollback_wqe, rollback_psn);
 807
 808                 if (err == -EAGAIN) {
 809                         rxe_sched_task(&qp->req.task);
 810                         goto exit;
 811                 }
 812
 813                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 814                 goto err;
 815         }
 816
 817         update_state(qp, &pkt);
 818
 819         /* A non-zero return value will cause rxe_do_task to
 820          * exit its loop and end the tasklet. A zero return
 821          * will continue looping and return to rxe_requester
 822          */
 823 done:
 824         ret = 0;
 825         goto out;
 826 err:
 827         /* update wqe_index for each wqe completion */
 828         qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
 829         wqe->state = wqe_state_error;
 830         qp->req.state = QP_STATE_ERROR;
 831         rxe_run_task(&qp->comp.task);
 832 exit:
 833         ret = -EAGAIN;
 834 out:
 835         rxe_put(qp);
 836
 837         return ret;
 838 }