drivers/infiniband/hw/qib/qib_uc.c

   1 /*
   2  * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
   3  * All rights reserved.
   4  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include "qib.h"
  36
  37 /* cut down ridiculously long IB macro names */
  38 #define OP(x) IB_OPCODE_UC_##x
  39
  40 /**
  41  * qib_make_uc_req - construct a request packet (SEND, RDMA write)
  42  * @qp: a pointer to the QP
  43  *
  44  * Return 1 if constructed; otherwise, return 0.
  45  */
  46 int qib_make_uc_req(struct rvt_qp *qp)
  47 {
  48         struct qib_qp_priv *priv = qp->priv;
  49         struct qib_other_headers *ohdr;
  50         struct rvt_swqe *wqe;
  51         unsigned long flags;
  52         u32 hwords;
  53         u32 bth0;
  54         u32 len;
  55         u32 pmtu = qp->pmtu;
  56         int ret = 0;
  57
  58         spin_lock_irqsave(&qp->s_lock, flags);
  59
  60         if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) {
  61                 if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND))
  62                         goto bail;
  63                 /* We are in the error state, flush the work request. */
  64                 if (qp->s_last == qp->s_head)
  65                         goto bail;
  66                 /* If DMAs are in progress, we can't flush immediately. */
  67                 if (atomic_read(&priv->s_dma_busy)) {
  68                         qp->s_flags |= RVT_S_WAIT_DMA;
  69                         goto bail;
  70                 }
  71                 wqe = rvt_get_swqe_ptr(qp, qp->s_last);
  72                 qib_send_complete(qp, wqe, IB_WC_WR_FLUSH_ERR);
  73                 goto done;
  74         }
  75
  76         ohdr = &priv->s_hdr->u.oth;
  77         if (qp->remote_ah_attr.ah_flags & IB_AH_GRH)
  78                 ohdr = &priv->s_hdr->u.l.oth;
  79
  80         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  81         hwords = 5;
  82         bth0 = 0;
  83
  84         /* Get the next send request. */
  85         wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
  86         qp->s_wqe = NULL;
  87         switch (qp->s_state) {
  88         default:
  89                 if (!(ib_rvt_state_ops[qp->state] &
  90                     RVT_PROCESS_NEXT_SEND_OK))
  91                         goto bail;
  92                 /* Check if send work queue is empty. */
  93                 if (qp->s_cur == qp->s_head)
  94                         goto bail;
  95                 /*
  96                  * Start a new request.
  97                  */
  98                 wqe->psn = qp->s_next_psn;
  99                 qp->s_psn = qp->s_next_psn;
 100                 qp->s_sge.sge = wqe->sg_list[0];
 101                 qp->s_sge.sg_list = wqe->sg_list + 1;
 102                 qp->s_sge.num_sge = wqe->wr.num_sge;
 103                 qp->s_sge.total_len = wqe->length;
 104                 len = wqe->length;
 105                 qp->s_len = len;
 106                 switch (wqe->wr.opcode) {
 107                 case IB_WR_SEND:
 108                 case IB_WR_SEND_WITH_IMM:
 109                         if (len > pmtu) {
 110                                 qp->s_state = OP(SEND_FIRST);
 111                                 len = pmtu;
 112                                 break;
 113                         }
 114                         if (wqe->wr.opcode == IB_WR_SEND)
 115                                 qp->s_state = OP(SEND_ONLY);
 116                         else {
 117                                 qp->s_state =
 118                                         OP(SEND_ONLY_WITH_IMMEDIATE);
 119                                 /* Immediate data comes after the BTH */
 120                                 ohdr->u.imm_data = wqe->wr.ex.imm_data;
 121                                 hwords += 1;
 122                         }
 123                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 124                                 bth0 |= IB_BTH_SOLICITED;
 125                         qp->s_wqe = wqe;
 126                         if (++qp->s_cur >= qp->s_size)
 127                                 qp->s_cur = 0;
 128                         break;
 129
 130                 case IB_WR_RDMA_WRITE:
 131                 case IB_WR_RDMA_WRITE_WITH_IMM:
 132                         ohdr->u.rc.reth.vaddr =
 133                                 cpu_to_be64(wqe->rdma_wr.remote_addr);
 134                         ohdr->u.rc.reth.rkey =
 135                                 cpu_to_be32(wqe->rdma_wr.rkey);
 136                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 137                         hwords += sizeof(struct ib_reth) / 4;
 138                         if (len > pmtu) {
 139                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 140                                 len = pmtu;
 141                                 break;
 142                         }
 143                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 144                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 145                         else {
 146                                 qp->s_state =
 147                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 148                                 /* Immediate data comes after the RETH */
 149                                 ohdr->u.rc.imm_data = wqe->wr.ex.imm_data;
 150                                 hwords += 1;
 151                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 152                                         bth0 |= IB_BTH_SOLICITED;
 153                         }
 154                         qp->s_wqe = wqe;
 155                         if (++qp->s_cur >= qp->s_size)
 156                                 qp->s_cur = 0;
 157                         break;
 158
 159                 default:
 160                         goto bail;
 161                 }
 162                 break;
 163
 164         case OP(SEND_FIRST):
 165                 qp->s_state = OP(SEND_MIDDLE);
 166                 /* FALLTHROUGH */
 167         case OP(SEND_MIDDLE):
 168                 len = qp->s_len;
 169                 if (len > pmtu) {
 170                         len = pmtu;
 171                         break;
 172                 }
 173                 if (wqe->wr.opcode == IB_WR_SEND)
 174                         qp->s_state = OP(SEND_LAST);
 175                 else {
 176                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 177                         /* Immediate data comes after the BTH */
 178                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 179                         hwords += 1;
 180                 }
 181                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 182                         bth0 |= IB_BTH_SOLICITED;
 183                 qp->s_wqe = wqe;
 184                 if (++qp->s_cur >= qp->s_size)
 185                         qp->s_cur = 0;
 186                 break;
 187
 188         case OP(RDMA_WRITE_FIRST):
 189                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 190                 /* FALLTHROUGH */
 191         case OP(RDMA_WRITE_MIDDLE):
 192                 len = qp->s_len;
 193                 if (len > pmtu) {
 194                         len = pmtu;
 195                         break;
 196                 }
 197                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 198                         qp->s_state = OP(RDMA_WRITE_LAST);
 199                 else {
 200                         qp->s_state =
 201                                 OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 202                         /* Immediate data comes after the BTH */
 203                         ohdr->u.imm_data = wqe->wr.ex.imm_data;
 204                         hwords += 1;
 205                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 206                                 bth0 |= IB_BTH_SOLICITED;
 207                 }
 208                 qp->s_wqe = wqe;
 209                 if (++qp->s_cur >= qp->s_size)
 210                         qp->s_cur = 0;
 211                 break;
 212         }
 213         qp->s_len -= len;
 214         qp->s_hdrwords = hwords;
 215         qp->s_cur_sge = &qp->s_sge;
 216         qp->s_cur_size = len;
 217         qib_make_ruc_header(qp, ohdr, bth0 | (qp->s_state << 24),
 218                             qp->s_next_psn++ & QIB_PSN_MASK);
 219 done:
 220         ret = 1;
 221         goto unlock;
 222
 223 bail:
 224         qp->s_flags &= ~RVT_S_BUSY;
 225 unlock:
 226         spin_unlock_irqrestore(&qp->s_lock, flags);
 227         return ret;
 228 }
 229
 230 /**
 231  * qib_uc_rcv - handle an incoming UC packet
 232  * @ibp: the port the packet came in on
 233  * @hdr: the header of the packet
 234  * @has_grh: true if the packet has a GRH
 235  * @data: the packet data
 236  * @tlen: the length of the packet
 237  * @qp: the QP for this packet.
 238  *
 239  * This is called from qib_qp_rcv() to process an incoming UC packet
 240  * for the given QP.
 241  * Called at interrupt level.
 242  */
 243 void qib_uc_rcv(struct qib_ibport *ibp, struct qib_ib_header *hdr,
 244                 int has_grh, void *data, u32 tlen, struct rvt_qp *qp)
 245 {
 246         struct qib_other_headers *ohdr;
 247         u32 opcode;
 248         u32 hdrsize;
 249         u32 psn;
 250         u32 pad;
 251         struct ib_wc wc;
 252         u32 pmtu = qp->pmtu;
 253         struct ib_reth *reth;
 254         int ret;
 255
 256         /* Check for GRH */
 257         if (!has_grh) {
 258                 ohdr = &hdr->u.oth;
 259                 hdrsize = 8 + 12;       /* LRH + BTH */
 260         } else {
 261                 ohdr = &hdr->u.l.oth;
 262                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
 263         }
 264
 265         opcode = be32_to_cpu(ohdr->bth[0]);
 266         if (qib_ruc_check_hdr(ibp, hdr, has_grh, qp, opcode))
 267                 return;
 268
 269         psn = be32_to_cpu(ohdr->bth[2]);
 270         opcode >>= 24;
 271
 272         /* Compare the PSN verses the expected PSN. */
 273         if (unlikely(qib_cmp24(psn, qp->r_psn) != 0)) {
 274                 /*
 275                  * Handle a sequence error.
 276                  * Silently drop any current message.
 277                  */
 278                 qp->r_psn = psn;
 279 inv:
 280                 if (qp->r_state == OP(SEND_FIRST) ||
 281                     qp->r_state == OP(SEND_MIDDLE)) {
 282                         set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
 283                         qp->r_sge.num_sge = 0;
 284                 } else
 285                         rvt_put_ss(&qp->r_sge);
 286                 qp->r_state = OP(SEND_LAST);
 287                 switch (opcode) {
 288                 case OP(SEND_FIRST):
 289                 case OP(SEND_ONLY):
 290                 case OP(SEND_ONLY_WITH_IMMEDIATE):
 291                         goto send_first;
 292
 293                 case OP(RDMA_WRITE_FIRST):
 294                 case OP(RDMA_WRITE_ONLY):
 295                 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
 296                         goto rdma_first;
 297
 298                 default:
 299                         goto drop;
 300                 }
 301         }
 302
 303         /* Check for opcode sequence errors. */
 304         switch (qp->r_state) {
 305         case OP(SEND_FIRST):
 306         case OP(SEND_MIDDLE):
 307                 if (opcode == OP(SEND_MIDDLE) ||
 308                     opcode == OP(SEND_LAST) ||
 309                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
 310                         break;
 311                 goto inv;
 312
 313         case OP(RDMA_WRITE_FIRST):
 314         case OP(RDMA_WRITE_MIDDLE):
 315                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
 316                     opcode == OP(RDMA_WRITE_LAST) ||
 317                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
 318                         break;
 319                 goto inv;
 320
 321         default:
 322                 if (opcode == OP(SEND_FIRST) ||
 323                     opcode == OP(SEND_ONLY) ||
 324                     opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
 325                     opcode == OP(RDMA_WRITE_FIRST) ||
 326                     opcode == OP(RDMA_WRITE_ONLY) ||
 327                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
 328                         break;
 329                 goto inv;
 330         }
 331
 332         if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) {
 333                 qp->r_flags |= RVT_R_COMM_EST;
 334                 if (qp->ibqp.event_handler) {
 335                         struct ib_event ev;
 336
 337                         ev.device = qp->ibqp.device;
 338                         ev.element.qp = &qp->ibqp;
 339                         ev.event = IB_EVENT_COMM_EST;
 340                         qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
 341                 }
 342         }
 343
 344         /* OK, process the packet. */
 345         switch (opcode) {
 346         case OP(SEND_FIRST):
 347         case OP(SEND_ONLY):
 348         case OP(SEND_ONLY_WITH_IMMEDIATE):
 349 send_first:
 350                 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
 351                         qp->r_sge = qp->s_rdma_read_sge;
 352                 else {
 353                         ret = qib_get_rwqe(qp, 0);
 354                         if (ret < 0)
 355                                 goto op_err;
 356                         if (!ret)
 357                                 goto drop;
 358                         /*
 359                          * qp->s_rdma_read_sge will be the owner
 360                          * of the mr references.
 361                          */
 362                         qp->s_rdma_read_sge = qp->r_sge;
 363                 }
 364                 qp->r_rcv_len = 0;
 365                 if (opcode == OP(SEND_ONLY))
 366                         goto no_immediate_data;
 367                 else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
 368                         goto send_last_imm;
 369                 /* FALLTHROUGH */
 370         case OP(SEND_MIDDLE):
 371                 /* Check for invalid length PMTU or posted rwqe len. */
 372                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
 373                         goto rewind;
 374                 qp->r_rcv_len += pmtu;
 375                 if (unlikely(qp->r_rcv_len > qp->r_len))
 376                         goto rewind;
 377                 qib_copy_sge(&qp->r_sge, data, pmtu, 0);
 378                 break;
 379
 380         case OP(SEND_LAST_WITH_IMMEDIATE):
 381 send_last_imm:
 382                 wc.ex.imm_data = ohdr->u.imm_data;
 383                 hdrsize += 4;
 384                 wc.wc_flags = IB_WC_WITH_IMM;
 385                 goto send_last;
 386         case OP(SEND_LAST):
 387 no_immediate_data:
 388                 wc.ex.imm_data = 0;
 389                 wc.wc_flags = 0;
 390 send_last:
 391                 /* Get the number of bytes the message was padded by. */
 392                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 393                 /* Check for invalid length. */
 394                 /* XXX LAST len should be >= 1 */
 395                 if (unlikely(tlen < (hdrsize + pad + 4)))
 396                         goto rewind;
 397                 /* Don't count the CRC. */
 398                 tlen -= (hdrsize + pad + 4);
 399                 wc.byte_len = tlen + qp->r_rcv_len;
 400                 if (unlikely(wc.byte_len > qp->r_len))
 401                         goto rewind;
 402                 wc.opcode = IB_WC_RECV;
 403                 qib_copy_sge(&qp->r_sge, data, tlen, 0);
 404                 rvt_put_ss(&qp->s_rdma_read_sge);
 405 last_imm:
 406                 wc.wr_id = qp->r_wr_id;
 407                 wc.status = IB_WC_SUCCESS;
 408                 wc.qp = &qp->ibqp;
 409                 wc.src_qp = qp->remote_qpn;
 410                 wc.slid = qp->remote_ah_attr.dlid;
 411                 wc.sl = qp->remote_ah_attr.sl;
 412                 /* zero fields that are N/A */
 413                 wc.vendor_err = 0;
 414                 wc.pkey_index = 0;
 415                 wc.dlid_path_bits = 0;
 416                 wc.port_num = 0;
 417                 /* Signal completion event if the solicited bit is set. */
 418                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc,
 419                              (ohdr->bth[0] &
 420                                 cpu_to_be32(IB_BTH_SOLICITED)) != 0);
 421                 break;
 422
 423         case OP(RDMA_WRITE_FIRST):
 424         case OP(RDMA_WRITE_ONLY):
 425         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE): /* consume RWQE */
 426 rdma_first:
 427                 if (unlikely(!(qp->qp_access_flags &
 428                                IB_ACCESS_REMOTE_WRITE))) {
 429                         goto drop;
 430                 }
 431                 reth = &ohdr->u.rc.reth;
 432                 hdrsize += sizeof(*reth);
 433                 qp->r_len = be32_to_cpu(reth->length);
 434                 qp->r_rcv_len = 0;
 435                 qp->r_sge.sg_list = NULL;
 436                 if (qp->r_len != 0) {
 437                         u32 rkey = be32_to_cpu(reth->rkey);
 438                         u64 vaddr = be64_to_cpu(reth->vaddr);
 439                         int ok;
 440
 441                         /* Check rkey */
 442                         ok = rvt_rkey_ok(qp, &qp->r_sge.sge, qp->r_len,
 443                                          vaddr, rkey, IB_ACCESS_REMOTE_WRITE);
 444                         if (unlikely(!ok))
 445                                 goto drop;
 446                         qp->r_sge.num_sge = 1;
 447                 } else {
 448                         qp->r_sge.num_sge = 0;
 449                         qp->r_sge.sge.mr = NULL;
 450                         qp->r_sge.sge.vaddr = NULL;
 451                         qp->r_sge.sge.length = 0;
 452                         qp->r_sge.sge.sge_length = 0;
 453                 }
 454                 if (opcode == OP(RDMA_WRITE_ONLY))
 455                         goto rdma_last;
 456                 else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE)) {
 457                         wc.ex.imm_data = ohdr->u.rc.imm_data;
 458                         goto rdma_last_imm;
 459                 }
 460                 /* FALLTHROUGH */
 461         case OP(RDMA_WRITE_MIDDLE):
 462                 /* Check for invalid length PMTU or posted rwqe len. */
 463                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
 464                         goto drop;
 465                 qp->r_rcv_len += pmtu;
 466                 if (unlikely(qp->r_rcv_len > qp->r_len))
 467                         goto drop;
 468                 qib_copy_sge(&qp->r_sge, data, pmtu, 1);
 469                 break;
 470
 471         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
 472                 wc.ex.imm_data = ohdr->u.imm_data;
 473 rdma_last_imm:
 474                 hdrsize += 4;
 475                 wc.wc_flags = IB_WC_WITH_IMM;
 476
 477                 /* Get the number of bytes the message was padded by. */
 478                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 479                 /* Check for invalid length. */
 480                 /* XXX LAST len should be >= 1 */
 481                 if (unlikely(tlen < (hdrsize + pad + 4)))
 482                         goto drop;
 483                 /* Don't count the CRC. */
 484                 tlen -= (hdrsize + pad + 4);
 485                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
 486                         goto drop;
 487                 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
 488                         rvt_put_ss(&qp->s_rdma_read_sge);
 489                 else {
 490                         ret = qib_get_rwqe(qp, 1);
 491                         if (ret < 0)
 492                                 goto op_err;
 493                         if (!ret)
 494                                 goto drop;
 495                 }
 496                 wc.byte_len = qp->r_len;
 497                 wc.opcode = IB_WC_RECV_RDMA_WITH_IMM;
 498                 qib_copy_sge(&qp->r_sge, data, tlen, 1);
 499                 rvt_put_ss(&qp->r_sge);
 500                 goto last_imm;
 501
 502         case OP(RDMA_WRITE_LAST):
 503 rdma_last:
 504                 /* Get the number of bytes the message was padded by. */
 505                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 506                 /* Check for invalid length. */
 507                 /* XXX LAST len should be >= 1 */
 508                 if (unlikely(tlen < (hdrsize + pad + 4)))
 509                         goto drop;
 510                 /* Don't count the CRC. */
 511                 tlen -= (hdrsize + pad + 4);
 512                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len))
 513                         goto drop;
 514                 qib_copy_sge(&qp->r_sge, data, tlen, 1);
 515                 rvt_put_ss(&qp->r_sge);
 516                 break;
 517
 518         default:
 519                 /* Drop packet for unknown opcodes. */
 520                 goto drop;
 521         }
 522         qp->r_psn++;
 523         qp->r_state = opcode;
 524         return;
 525
 526 rewind:
 527         set_bit(RVT_R_REWIND_SGE, &qp->r_aflags);
 528         qp->r_sge.num_sge = 0;
 529 drop:
 530         ibp->rvp.n_pkt_drops++;
 531         return;
 532
 533 op_err:
 534         qib_rc_error(qp, IB_WC_LOC_QP_OP_ERR);
 535         return;
 536
 537 }