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         spin_lock_bh(&qp->state_lock);
 106         if (qp->valid) {
 107                 /* request a send queue retry */
 108                 qp->req.need_retry = 1;
 109                 qp->req.wait_for_rnr_timer = 0;
 110                 rxe_sched_task(&qp->req.task);
 111         }
 112         spin_unlock_bh(&qp->state_lock);
 113 }
 114
 115 static void req_check_sq_drain_done(struct rxe_qp *qp)
 116 {
 117         struct rxe_queue *q;
 118         unsigned int index;
 119         unsigned int cons;
 120         struct rxe_send_wqe *wqe;
 121
 122         spin_lock_bh(&qp->state_lock);
 123         if (qp_state(qp) == IB_QPS_SQD) {
 124                 q = qp->sq.queue;
 125                 index = qp->req.wqe_index;
 126                 cons = queue_get_consumer(q, QUEUE_TYPE_FROM_CLIENT);
 127                 wqe = queue_addr_from_index(q, cons);
 128
 129                 /* check to see if we are drained;
 130                  * state_lock used by requester and completer
 131                  */
 132                 do {
 133                         if (!qp->attr.sq_draining)
 134                                 /* comp just finished */
 135                                 break;
 136
 137                         if (wqe && ((index != cons) ||
 138                                 (wqe->state != wqe_state_posted)))
 139                                 /* comp not done yet */
 140                                 break;
 141
 142                         qp->attr.sq_draining = 0;
 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                         return;
 155                 } while (0);
 156         }
 157         spin_unlock_bh(&qp->state_lock);
 158 }
 159
 160 static struct rxe_send_wqe *__req_next_wqe(struct rxe_qp *qp)
 161 {
 162         struct rxe_queue *q = qp->sq.queue;
 163         unsigned int index = qp->req.wqe_index;
 164         unsigned int prod;
 165
 166         prod = queue_get_producer(q, QUEUE_TYPE_FROM_CLIENT);
 167         if (index == prod)
 168                 return NULL;
 169         else
 170                 return queue_addr_from_index(q, index);
 171 }
 172
 173 static struct rxe_send_wqe *req_next_wqe(struct rxe_qp *qp)
 174 {
 175         struct rxe_send_wqe *wqe;
 176
 177         req_check_sq_drain_done(qp);
 178
 179         wqe = __req_next_wqe(qp);
 180         if (wqe == NULL)
 181                 return NULL;
 182
 183         spin_lock_bh(&qp->state_lock);
 184         if (unlikely((qp_state(qp) == IB_QPS_SQD) &&
 185                      (wqe->state != wqe_state_processing))) {
 186                 spin_unlock_bh(&qp->state_lock);
 187                 return NULL;
 188         }
 189         spin_unlock_bh(&qp->state_lock);
 190
 191         wqe->mask = wr_opcode_mask(wqe->wr.opcode, qp);
 192         return wqe;
 193 }
 194
 195 /**
 196  * rxe_wqe_is_fenced - check if next wqe is fenced
 197  * @qp: the queue pair
 198  * @wqe: the next wqe
 199  *
 200  * Returns: 1 if wqe needs to wait
 201  *          0 if wqe is ready to go
 202  */
 203 static int rxe_wqe_is_fenced(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 204 {
 205         /* Local invalidate fence (LIF) see IBA 10.6.5.1
 206          * Requires ALL previous operations on the send queue
 207          * are complete. Make mandatory for the rxe driver.
 208          */
 209         if (wqe->wr.opcode == IB_WR_LOCAL_INV)
 210                 return qp->req.wqe_index != queue_get_consumer(qp->sq.queue,
 211                                                 QUEUE_TYPE_FROM_CLIENT);
 212
 213         /* Fence see IBA 10.8.3.3
 214          * Requires that all previous read and atomic operations
 215          * are complete.
 216          */
 217         return (wqe->wr.send_flags & IB_SEND_FENCE) &&
 218                 atomic_read(&qp->req.rd_atomic) != qp->attr.max_rd_atomic;
 219 }
 220
 221 static int next_opcode_rc(struct rxe_qp *qp, u32 opcode, int fits)
 222 {
 223         switch (opcode) {
 224         case IB_WR_RDMA_WRITE:
 225                 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
 226                     qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
 227                         return fits ?
 228                                 IB_OPCODE_RC_RDMA_WRITE_LAST :
 229                                 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
 230                 else
 231                         return fits ?
 232                                 IB_OPCODE_RC_RDMA_WRITE_ONLY :
 233                                 IB_OPCODE_RC_RDMA_WRITE_FIRST;
 234
 235         case IB_WR_RDMA_WRITE_WITH_IMM:
 236                 if (qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_FIRST ||
 237                     qp->req.opcode == IB_OPCODE_RC_RDMA_WRITE_MIDDLE)
 238                         return fits ?
 239                                 IB_OPCODE_RC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
 240                                 IB_OPCODE_RC_RDMA_WRITE_MIDDLE;
 241                 else
 242                         return fits ?
 243                                 IB_OPCODE_RC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
 244                                 IB_OPCODE_RC_RDMA_WRITE_FIRST;
 245
 246         case IB_WR_SEND:
 247                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 248                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 249                         return fits ?
 250                                 IB_OPCODE_RC_SEND_LAST :
 251                                 IB_OPCODE_RC_SEND_MIDDLE;
 252                 else
 253                         return fits ?
 254                                 IB_OPCODE_RC_SEND_ONLY :
 255                                 IB_OPCODE_RC_SEND_FIRST;
 256
 257         case IB_WR_SEND_WITH_IMM:
 258                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 259                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 260                         return fits ?
 261                                 IB_OPCODE_RC_SEND_LAST_WITH_IMMEDIATE :
 262                                 IB_OPCODE_RC_SEND_MIDDLE;
 263                 else
 264                         return fits ?
 265                                 IB_OPCODE_RC_SEND_ONLY_WITH_IMMEDIATE :
 266                                 IB_OPCODE_RC_SEND_FIRST;
 267
 268         case IB_WR_FLUSH:
 269                 return IB_OPCODE_RC_FLUSH;
 270
 271         case IB_WR_RDMA_READ:
 272                 return IB_OPCODE_RC_RDMA_READ_REQUEST;
 273
 274         case IB_WR_ATOMIC_CMP_AND_SWP:
 275                 return IB_OPCODE_RC_COMPARE_SWAP;
 276
 277         case IB_WR_ATOMIC_FETCH_AND_ADD:
 278                 return IB_OPCODE_RC_FETCH_ADD;
 279
 280         case IB_WR_SEND_WITH_INV:
 281                 if (qp->req.opcode == IB_OPCODE_RC_SEND_FIRST ||
 282                     qp->req.opcode == IB_OPCODE_RC_SEND_MIDDLE)
 283                         return fits ? IB_OPCODE_RC_SEND_LAST_WITH_INVALIDATE :
 284                                 IB_OPCODE_RC_SEND_MIDDLE;
 285                 else
 286                         return fits ? IB_OPCODE_RC_SEND_ONLY_WITH_INVALIDATE :
 287                                 IB_OPCODE_RC_SEND_FIRST;
 288
 289         case IB_WR_ATOMIC_WRITE:
 290                 return IB_OPCODE_RC_ATOMIC_WRITE;
 291
 292         case IB_WR_REG_MR:
 293         case IB_WR_LOCAL_INV:
 294                 return opcode;
 295         }
 296
 297         return -EINVAL;
 298 }
 299
 300 static int next_opcode_uc(struct rxe_qp *qp, u32 opcode, int fits)
 301 {
 302         switch (opcode) {
 303         case IB_WR_RDMA_WRITE:
 304                 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
 305                     qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
 306                         return fits ?
 307                                 IB_OPCODE_UC_RDMA_WRITE_LAST :
 308                                 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
 309                 else
 310                         return fits ?
 311                                 IB_OPCODE_UC_RDMA_WRITE_ONLY :
 312                                 IB_OPCODE_UC_RDMA_WRITE_FIRST;
 313
 314         case IB_WR_RDMA_WRITE_WITH_IMM:
 315                 if (qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_FIRST ||
 316                     qp->req.opcode == IB_OPCODE_UC_RDMA_WRITE_MIDDLE)
 317                         return fits ?
 318                                 IB_OPCODE_UC_RDMA_WRITE_LAST_WITH_IMMEDIATE :
 319                                 IB_OPCODE_UC_RDMA_WRITE_MIDDLE;
 320                 else
 321                         return fits ?
 322                                 IB_OPCODE_UC_RDMA_WRITE_ONLY_WITH_IMMEDIATE :
 323                                 IB_OPCODE_UC_RDMA_WRITE_FIRST;
 324
 325         case IB_WR_SEND:
 326                 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
 327                     qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
 328                         return fits ?
 329                                 IB_OPCODE_UC_SEND_LAST :
 330                                 IB_OPCODE_UC_SEND_MIDDLE;
 331                 else
 332                         return fits ?
 333                                 IB_OPCODE_UC_SEND_ONLY :
 334                                 IB_OPCODE_UC_SEND_FIRST;
 335
 336         case IB_WR_SEND_WITH_IMM:
 337                 if (qp->req.opcode == IB_OPCODE_UC_SEND_FIRST ||
 338                     qp->req.opcode == IB_OPCODE_UC_SEND_MIDDLE)
 339                         return fits ?
 340                                 IB_OPCODE_UC_SEND_LAST_WITH_IMMEDIATE :
 341                                 IB_OPCODE_UC_SEND_MIDDLE;
 342                 else
 343                         return fits ?
 344                                 IB_OPCODE_UC_SEND_ONLY_WITH_IMMEDIATE :
 345                                 IB_OPCODE_UC_SEND_FIRST;
 346         }
 347
 348         return -EINVAL;
 349 }
 350
 351 static int next_opcode(struct rxe_qp *qp, struct rxe_send_wqe *wqe,
 352                        u32 opcode)
 353 {
 354         int fits = (wqe->dma.resid <= qp->mtu);
 355
 356         switch (qp_type(qp)) {
 357         case IB_QPT_RC:
 358                 return next_opcode_rc(qp, opcode, fits);
 359
 360         case IB_QPT_UC:
 361                 return next_opcode_uc(qp, opcode, fits);
 362
 363         case IB_QPT_UD:
 364         case IB_QPT_GSI:
 365                 switch (opcode) {
 366                 case IB_WR_SEND:
 367                         return IB_OPCODE_UD_SEND_ONLY;
 368
 369                 case IB_WR_SEND_WITH_IMM:
 370                         return IB_OPCODE_UD_SEND_ONLY_WITH_IMMEDIATE;
 371                 }
 372                 break;
 373
 374         default:
 375                 break;
 376         }
 377
 378         return -EINVAL;
 379 }
 380
 381 static inline int check_init_depth(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 382 {
 383         int depth;
 384
 385         if (wqe->has_rd_atomic)
 386                 return 0;
 387
 388         qp->req.need_rd_atomic = 1;
 389         depth = atomic_dec_return(&qp->req.rd_atomic);
 390
 391         if (depth >= 0) {
 392                 qp->req.need_rd_atomic = 0;
 393                 wqe->has_rd_atomic = 1;
 394                 return 0;
 395         }
 396
 397         atomic_inc(&qp->req.rd_atomic);
 398         return -EAGAIN;
 399 }
 400
 401 static inline int get_mtu(struct rxe_qp *qp)
 402 {
 403         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 404
 405         if ((qp_type(qp) == IB_QPT_RC) || (qp_type(qp) == IB_QPT_UC))
 406                 return qp->mtu;
 407
 408         return rxe->port.mtu_cap;
 409 }
 410
 411 static struct sk_buff *init_req_packet(struct rxe_qp *qp,
 412                                        struct rxe_av *av,
 413                                        struct rxe_send_wqe *wqe,
 414                                        int opcode, u32 payload,
 415                                        struct rxe_pkt_info *pkt)
 416 {
 417         struct rxe_dev          *rxe = to_rdev(qp->ibqp.device);
 418         struct sk_buff          *skb;
 419         struct rxe_send_wr      *ibwr = &wqe->wr;
 420         int                     pad = (-payload) & 0x3;
 421         int                     paylen;
 422         int                     solicited;
 423         u32                     qp_num;
 424         int                     ack_req;
 425
 426         /* length from start of bth to end of icrc */
 427         paylen = rxe_opcode[opcode].length + payload + pad + RXE_ICRC_SIZE;
 428         pkt->paylen = paylen;
 429
 430         /* init skb */
 431         skb = rxe_init_packet(rxe, av, paylen, pkt);
 432         if (unlikely(!skb))
 433                 return NULL;
 434
 435         /* init bth */
 436         solicited = (ibwr->send_flags & IB_SEND_SOLICITED) &&
 437                         (pkt->mask & RXE_END_MASK) &&
 438                         ((pkt->mask & (RXE_SEND_MASK)) ||
 439                         (pkt->mask & (RXE_WRITE_MASK | RXE_IMMDT_MASK)) ==
 440                         (RXE_WRITE_MASK | RXE_IMMDT_MASK));
 441
 442         qp_num = (pkt->mask & RXE_DETH_MASK) ? ibwr->wr.ud.remote_qpn :
 443                                          qp->attr.dest_qp_num;
 444
 445         ack_req = ((pkt->mask & RXE_END_MASK) ||
 446                 (qp->req.noack_pkts++ > RXE_MAX_PKT_PER_ACK));
 447         if (ack_req)
 448                 qp->req.noack_pkts = 0;
 449
 450         bth_init(pkt, pkt->opcode, solicited, 0, pad, IB_DEFAULT_PKEY_FULL, qp_num,
 451                  ack_req, pkt->psn);
 452
 453         /* init optional headers */
 454         if (pkt->mask & RXE_RETH_MASK) {
 455                 if (pkt->mask & RXE_FETH_MASK)
 456                         reth_set_rkey(pkt, ibwr->wr.flush.rkey);
 457                 else
 458                         reth_set_rkey(pkt, ibwr->wr.rdma.rkey);
 459                 reth_set_va(pkt, wqe->iova);
 460                 reth_set_len(pkt, wqe->dma.resid);
 461         }
 462
 463         /* Fill Flush Extension Transport Header */
 464         if (pkt->mask & RXE_FETH_MASK)
 465                 feth_init(pkt, ibwr->wr.flush.type, ibwr->wr.flush.level);
 466
 467         if (pkt->mask & RXE_IMMDT_MASK)
 468                 immdt_set_imm(pkt, ibwr->ex.imm_data);
 469
 470         if (pkt->mask & RXE_IETH_MASK)
 471                 ieth_set_rkey(pkt, ibwr->ex.invalidate_rkey);
 472
 473         if (pkt->mask & RXE_ATMETH_MASK) {
 474                 atmeth_set_va(pkt, wqe->iova);
 475                 if (opcode == IB_OPCODE_RC_COMPARE_SWAP) {
 476                         atmeth_set_swap_add(pkt, ibwr->wr.atomic.swap);
 477                         atmeth_set_comp(pkt, ibwr->wr.atomic.compare_add);
 478                 } else {
 479                         atmeth_set_swap_add(pkt, ibwr->wr.atomic.compare_add);
 480                 }
 481                 atmeth_set_rkey(pkt, ibwr->wr.atomic.rkey);
 482         }
 483
 484         if (pkt->mask & RXE_DETH_MASK) {
 485                 if (qp->ibqp.qp_num == 1)
 486                         deth_set_qkey(pkt, GSI_QKEY);
 487                 else
 488                         deth_set_qkey(pkt, ibwr->wr.ud.remote_qkey);
 489                 deth_set_sqp(pkt, qp->ibqp.qp_num);
 490         }
 491
 492         return skb;
 493 }
 494
 495 static int finish_packet(struct rxe_qp *qp, struct rxe_av *av,
 496                          struct rxe_send_wqe *wqe, struct rxe_pkt_info *pkt,
 497                          struct sk_buff *skb, u32 payload)
 498 {
 499         int err;
 500
 501         err = rxe_prepare(av, pkt, skb);
 502         if (err)
 503                 return err;
 504
 505         if (pkt->mask & RXE_WRITE_OR_SEND_MASK) {
 506                 if (wqe->wr.send_flags & IB_SEND_INLINE) {
 507                         u8 *tmp = &wqe->dma.inline_data[wqe->dma.sge_offset];
 508
 509                         memcpy(payload_addr(pkt), tmp, payload);
 510
 511                         wqe->dma.resid -= payload;
 512                         wqe->dma.sge_offset += payload;
 513                 } else {
 514                         err = copy_data(qp->pd, 0, &wqe->dma,
 515                                         payload_addr(pkt), payload,
 516                                         RXE_FROM_MR_OBJ);
 517                         if (err)
 518                                 return err;
 519                 }
 520                 if (bth_pad(pkt)) {
 521                         u8 *pad = payload_addr(pkt) + payload;
 522
 523                         memset(pad, 0, bth_pad(pkt));
 524                 }
 525         } else if (pkt->mask & RXE_FLUSH_MASK) {
 526                 /* oA19-2: shall have no payload. */
 527                 wqe->dma.resid = 0;
 528         }
 529
 530         if (pkt->mask & RXE_ATOMIC_WRITE_MASK) {
 531                 memcpy(payload_addr(pkt), wqe->dma.atomic_wr, payload);
 532                 wqe->dma.resid -= payload;
 533         }
 534
 535         return 0;
 536 }
 537
 538 static void update_wqe_state(struct rxe_qp *qp,
 539                 struct rxe_send_wqe *wqe,
 540                 struct rxe_pkt_info *pkt)
 541 {
 542         if (pkt->mask & RXE_END_MASK) {
 543                 if (qp_type(qp) == IB_QPT_RC)
 544                         wqe->state = wqe_state_pending;
 545         } else {
 546                 wqe->state = wqe_state_processing;
 547         }
 548 }
 549
 550 static void update_wqe_psn(struct rxe_qp *qp,
 551                            struct rxe_send_wqe *wqe,
 552                            struct rxe_pkt_info *pkt,
 553                            u32 payload)
 554 {
 555         /* number of packets left to send including current one */
 556         int num_pkt = (wqe->dma.resid + payload + qp->mtu - 1) / qp->mtu;
 557
 558         /* handle zero length packet case */
 559         if (num_pkt == 0)
 560                 num_pkt = 1;
 561
 562         if (pkt->mask & RXE_START_MASK) {
 563                 wqe->first_psn = qp->req.psn;
 564                 wqe->last_psn = (qp->req.psn + num_pkt - 1) & BTH_PSN_MASK;
 565         }
 566
 567         if (pkt->mask & RXE_READ_MASK)
 568                 qp->req.psn = (wqe->first_psn + num_pkt) & BTH_PSN_MASK;
 569         else
 570                 qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
 571 }
 572
 573 static void save_state(struct rxe_send_wqe *wqe,
 574                        struct rxe_qp *qp,
 575                        struct rxe_send_wqe *rollback_wqe,
 576                        u32 *rollback_psn)
 577 {
 578         rollback_wqe->state     = wqe->state;
 579         rollback_wqe->first_psn = wqe->first_psn;
 580         rollback_wqe->last_psn  = wqe->last_psn;
 581         *rollback_psn           = qp->req.psn;
 582 }
 583
 584 static void rollback_state(struct rxe_send_wqe *wqe,
 585                            struct rxe_qp *qp,
 586                            struct rxe_send_wqe *rollback_wqe,
 587                            u32 rollback_psn)
 588 {
 589         wqe->state     = rollback_wqe->state;
 590         wqe->first_psn = rollback_wqe->first_psn;
 591         wqe->last_psn  = rollback_wqe->last_psn;
 592         qp->req.psn    = rollback_psn;
 593 }
 594
 595 static void update_state(struct rxe_qp *qp, struct rxe_pkt_info *pkt)
 596 {
 597         qp->req.opcode = pkt->opcode;
 598
 599         if (pkt->mask & RXE_END_MASK)
 600                 qp->req.wqe_index = queue_next_index(qp->sq.queue,
 601                                                      qp->req.wqe_index);
 602
 603         qp->need_req_skb = 0;
 604
 605         if (qp->qp_timeout_jiffies && !timer_pending(&qp->retrans_timer))
 606                 mod_timer(&qp->retrans_timer,
 607                           jiffies + qp->qp_timeout_jiffies);
 608 }
 609
 610 static int rxe_do_local_ops(struct rxe_qp *qp, struct rxe_send_wqe *wqe)
 611 {
 612         u8 opcode = wqe->wr.opcode;
 613         u32 rkey;
 614         int ret;
 615
 616         switch (opcode) {
 617         case IB_WR_LOCAL_INV:
 618                 rkey = wqe->wr.ex.invalidate_rkey;
 619                 if (rkey_is_mw(rkey))
 620                         ret = rxe_invalidate_mw(qp, rkey);
 621                 else
 622                         ret = rxe_invalidate_mr(qp, rkey);
 623
 624                 if (unlikely(ret)) {
 625                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 626                         return ret;
 627                 }
 628                 break;
 629         case IB_WR_REG_MR:
 630                 ret = rxe_reg_fast_mr(qp, wqe);
 631                 if (unlikely(ret)) {
 632                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 633                         return ret;
 634                 }
 635                 break;
 636         case IB_WR_BIND_MW:
 637                 ret = rxe_bind_mw(qp, wqe);
 638                 if (unlikely(ret)) {
 639                         wqe->status = IB_WC_MW_BIND_ERR;
 640                         return ret;
 641                 }
 642                 break;
 643         default:
 644                 rxe_dbg_qp(qp, "Unexpected send wqe opcode %d\n", opcode);
 645                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 646                 return -EINVAL;
 647         }
 648
 649         wqe->state = wqe_state_done;
 650         wqe->status = IB_WC_SUCCESS;
 651         qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
 652
 653         /* There is no ack coming for local work requests
 654          * which can lead to a deadlock. So go ahead and complete
 655          * it now.
 656          */
 657         rxe_sched_task(&qp->comp.task);
 658
 659         return 0;
 660 }
 661
 662 int rxe_requester(struct rxe_qp *qp)
 663 {
 664         struct rxe_dev *rxe = to_rdev(qp->ibqp.device);
 665         struct rxe_pkt_info pkt;
 666         struct sk_buff *skb;
 667         struct rxe_send_wqe *wqe;
 668         enum rxe_hdr_mask mask;
 669         u32 payload;
 670         int mtu;
 671         int opcode;
 672         int err;
 673         int ret;
 674         struct rxe_send_wqe rollback_wqe;
 675         u32 rollback_psn;
 676         struct rxe_queue *q = qp->sq.queue;
 677         struct rxe_ah *ah;
 678         struct rxe_av *av;
 679
 680         spin_lock_bh(&qp->state_lock);
 681         if (unlikely(!qp->valid)) {
 682                 spin_unlock_bh(&qp->state_lock);
 683                 goto exit;
 684         }
 685
 686         if (unlikely(qp_state(qp) == IB_QPS_ERR)) {
 687                 wqe = __req_next_wqe(qp);
 688                 spin_unlock_bh(&qp->state_lock);
 689                 if (wqe)
 690                         goto err;
 691                 else
 692                         goto exit;
 693         }
 694
 695         if (unlikely(qp_state(qp) == IB_QPS_RESET)) {
 696                 qp->req.wqe_index = queue_get_consumer(q,
 697                                                 QUEUE_TYPE_FROM_CLIENT);
 698                 qp->req.opcode = -1;
 699                 qp->req.need_rd_atomic = 0;
 700                 qp->req.wait_psn = 0;
 701                 qp->req.need_retry = 0;
 702                 qp->req.wait_for_rnr_timer = 0;
 703                 spin_unlock_bh(&qp->state_lock);
 704                 goto exit;
 705         }
 706         spin_unlock_bh(&qp->state_lock);
 707
 708         /* we come here if the retransmit timer has fired
 709          * or if the rnr timer has fired. If the retransmit
 710          * timer fires while we are processing an RNR NAK wait
 711          * until the rnr timer has fired before starting the
 712          * retry flow
 713          */
 714         if (unlikely(qp->req.need_retry && !qp->req.wait_for_rnr_timer)) {
 715                 req_retry(qp);
 716                 qp->req.need_retry = 0;
 717         }
 718
 719         wqe = req_next_wqe(qp);
 720         if (unlikely(!wqe))
 721                 goto exit;
 722
 723         if (rxe_wqe_is_fenced(qp, wqe)) {
 724                 qp->req.wait_fence = 1;
 725                 goto exit;
 726         }
 727
 728         if (wqe->mask & WR_LOCAL_OP_MASK) {
 729                 err = rxe_do_local_ops(qp, wqe);
 730                 if (unlikely(err))
 731                         goto err;
 732                 else
 733                         goto done;
 734         }
 735
 736         if (unlikely(qp_type(qp) == IB_QPT_RC &&
 737                 psn_compare(qp->req.psn, (qp->comp.psn +
 738                                 RXE_MAX_UNACKED_PSNS)) > 0)) {
 739                 qp->req.wait_psn = 1;
 740                 goto exit;
 741         }
 742
 743         /* Limit the number of inflight SKBs per QP */
 744         if (unlikely(atomic_read(&qp->skb_out) >
 745                      RXE_INFLIGHT_SKBS_PER_QP_HIGH)) {
 746                 qp->need_req_skb = 1;
 747                 goto exit;
 748         }
 749
 750         opcode = next_opcode(qp, wqe, wqe->wr.opcode);
 751         if (unlikely(opcode < 0)) {
 752                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 753                 goto err;
 754         }
 755
 756         mask = rxe_opcode[opcode].mask;
 757         if (unlikely(mask & (RXE_READ_OR_ATOMIC_MASK |
 758                         RXE_ATOMIC_WRITE_MASK))) {
 759                 if (check_init_depth(qp, wqe))
 760                         goto exit;
 761         }
 762
 763         mtu = get_mtu(qp);
 764         payload = (mask & (RXE_WRITE_OR_SEND_MASK | RXE_ATOMIC_WRITE_MASK)) ?
 765                         wqe->dma.resid : 0;
 766         if (payload > mtu) {
 767                 if (qp_type(qp) == IB_QPT_UD) {
 768                         /* C10-93.1.1: If the total sum of all the buffer lengths specified for a
 769                          * UD message exceeds the MTU of the port as returned by QueryHCA, the CI
 770                          * shall not emit any packets for this message. Further, the CI shall not
 771                          * generate an error due to this condition.
 772                          */
 773
 774                         /* fake a successful UD send */
 775                         wqe->first_psn = qp->req.psn;
 776                         wqe->last_psn = qp->req.psn;
 777                         qp->req.psn = (qp->req.psn + 1) & BTH_PSN_MASK;
 778                         qp->req.opcode = IB_OPCODE_UD_SEND_ONLY;
 779                         qp->req.wqe_index = queue_next_index(qp->sq.queue,
 780                                                        qp->req.wqe_index);
 781                         wqe->state = wqe_state_done;
 782                         wqe->status = IB_WC_SUCCESS;
 783                         rxe_sched_task(&qp->comp.task);
 784                         goto done;
 785                 }
 786                 payload = mtu;
 787         }
 788
 789         pkt.rxe = rxe;
 790         pkt.opcode = opcode;
 791         pkt.qp = qp;
 792         pkt.psn = qp->req.psn;
 793         pkt.mask = rxe_opcode[opcode].mask;
 794         pkt.wqe = wqe;
 795
 796         av = rxe_get_av(&pkt, &ah);
 797         if (unlikely(!av)) {
 798                 rxe_dbg_qp(qp, "Failed no address vector\n");
 799                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 800                 goto err;
 801         }
 802
 803         skb = init_req_packet(qp, av, wqe, opcode, payload, &pkt);
 804         if (unlikely(!skb)) {
 805                 rxe_dbg_qp(qp, "Failed allocating skb\n");
 806                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 807                 if (ah)
 808                         rxe_put(ah);
 809                 goto err;
 810         }
 811
 812         err = finish_packet(qp, av, wqe, &pkt, skb, payload);
 813         if (unlikely(err)) {
 814                 rxe_dbg_qp(qp, "Error during finish packet\n");
 815                 if (err == -EFAULT)
 816                         wqe->status = IB_WC_LOC_PROT_ERR;
 817                 else
 818                         wqe->status = IB_WC_LOC_QP_OP_ERR;
 819                 kfree_skb(skb);
 820                 if (ah)
 821                         rxe_put(ah);
 822                 goto err;
 823         }
 824
 825         if (ah)
 826                 rxe_put(ah);
 827
 828         /*
 829          * To prevent a race on wqe access between requester and completer,
 830          * wqe members state and psn need to be set before calling
 831          * rxe_xmit_packet().
 832          * Otherwise, completer might initiate an unjustified retry flow.
 833          */
 834         save_state(wqe, qp, &rollback_wqe, &rollback_psn);
 835         update_wqe_state(qp, wqe, &pkt);
 836         update_wqe_psn(qp, wqe, &pkt, payload);
 837
 838         err = rxe_xmit_packet(qp, &pkt, skb);
 839         if (err) {
 840                 qp->need_req_skb = 1;
 841
 842                 rollback_state(wqe, qp, &rollback_wqe, rollback_psn);
 843
 844                 if (err == -EAGAIN) {
 845                         rxe_sched_task(&qp->req.task);
 846                         goto exit;
 847                 }
 848
 849                 wqe->status = IB_WC_LOC_QP_OP_ERR;
 850                 goto err;
 851         }
 852
 853         update_state(qp, &pkt);
 854
 855         /* A non-zero return value will cause rxe_do_task to
 856          * exit its loop and end the tasklet. A zero return
 857          * will continue looping and return to rxe_requester
 858          */
 859 done:
 860         ret = 0;
 861         goto out;
 862 err:
 863         /* update wqe_index for each wqe completion */
 864         qp->req.wqe_index = queue_next_index(qp->sq.queue, qp->req.wqe_index);
 865         wqe->state = wqe_state_error;
 866         rxe_qp_error(qp);
 867 exit:
 868         ret = -EAGAIN;
 869 out:
 870         return ret;
 871 }