git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_uc.c

   1 /*
   2  * Copyright (c) 2006 QLogic, Inc. All rights reserved.
   3  * Copyright (c) 2005, 2006 PathScale, Inc. All rights reserved.
   4  *
   5  * This software is available to you under a choice of one of two
   6  * licenses.  You may choose to be licensed under the terms of the GNU
   7  * General Public License (GPL) Version 2, available from the file
   8  * COPYING in the main directory of this source tree, or the
   9  * OpenIB.org BSD license below:
  10  *
  11  *     Redistribution and use in source and binary forms, with or
  12  *     without modification, are permitted provided that the following
  13  *     conditions are met:
  14  *
  15  *      - Redistributions of source code must retain the above
  16  *        copyright notice, this list of conditions and the following
  17  *        disclaimer.
  18  *
  19  *      - Redistributions in binary form must reproduce the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer in the documentation and/or other materials
  22  *        provided with the distribution.
  23  *
  24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31  * SOFTWARE.
  32  */
  33
  34 #include "ipath_verbs.h"
  35 #include "ipath_kernel.h"
  36
  37 /* cut down ridiculously long IB macro names */
  38 #define OP(x) IB_OPCODE_UC_##x
  39
  40 static void complete_last_send(struct ipath_qp *qp, struct ipath_swqe *wqe,
  41                                struct ib_wc *wc)
  42 {
  43         if (++qp->s_last == qp->s_size)
  44                 qp->s_last = 0;
  45         if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
  46             (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
  47                 wc->wr_id = wqe->wr.wr_id;
  48                 wc->status = IB_WC_SUCCESS;
  49                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
  50                 wc->vendor_err = 0;
  51                 wc->byte_len = wqe->length;
  52                 wc->qp_num = qp->ibqp.qp_num;
  53                 wc->src_qp = qp->remote_qpn;
  54                 wc->pkey_index = 0;
  55                 wc->slid = qp->remote_ah_attr.dlid;
  56                 wc->sl = qp->remote_ah_attr.sl;
  57                 wc->dlid_path_bits = 0;
  58                 wc->port_num = 0;
  59                 ipath_cq_enter(to_icq(qp->ibqp.send_cq), wc, 0);
  60         }
  61         wqe = get_swqe_ptr(qp, qp->s_last);
  62 }
  63
  64 /**
  65  * ipath_make_uc_req - construct a request packet (SEND, RDMA write)
  66  * @qp: a pointer to the QP
  67  * @ohdr: a pointer to the IB header being constructed
  68  * @pmtu: the path MTU
  69  * @bth0p: pointer to the BTH opcode word
  70  * @bth2p: pointer to the BTH PSN word
  71  *
  72  * Return 1 if constructed; otherwise, return 0.
  73  * Note the QP s_lock must be held and interrupts disabled.
  74  */
  75 int ipath_make_uc_req(struct ipath_qp *qp,
  76                       struct ipath_other_headers *ohdr,
  77                       u32 pmtu, u32 *bth0p, u32 *bth2p)
  78 {
  79         struct ipath_swqe *wqe;
  80         u32 hwords;
  81         u32 bth0;
  82         u32 len;
  83         struct ib_wc wc;
  84
  85         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK))
  86                 goto done;
  87
  88         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  89         hwords = 5;
  90         bth0 = 0;
  91
  92         /* Get the next send request. */
  93         wqe = get_swqe_ptr(qp, qp->s_last);
  94         switch (qp->s_state) {
  95         default:
  96                 /*
  97                  * Signal the completion of the last send
  98                  * (if there is one).
  99                  */
 100                 if (qp->s_last != qp->s_tail)
 101                         complete_last_send(qp, wqe, &wc);
 102
 103                 /* Check if send work queue is empty. */
 104                 if (qp->s_tail == qp->s_head)
 105                         goto done;
 106                 /*
 107                  * Start a new request.
 108                  */
 109                 qp->s_psn = wqe->psn = qp->s_next_psn;
 110                 qp->s_sge.sge = wqe->sg_list[0];
 111                 qp->s_sge.sg_list = wqe->sg_list + 1;
 112                 qp->s_sge.num_sge = wqe->wr.num_sge;
 113                 qp->s_len = len = wqe->length;
 114                 switch (wqe->wr.opcode) {
 115                 case IB_WR_SEND:
 116                 case IB_WR_SEND_WITH_IMM:
 117                         if (len > pmtu) {
 118                                 qp->s_state = OP(SEND_FIRST);
 119                                 len = pmtu;
 120                                 break;
 121                         }
 122                         if (wqe->wr.opcode == IB_WR_SEND)
 123                                 qp->s_state = OP(SEND_ONLY);
 124                         else {
 125                                 qp->s_state =
 126                                         OP(SEND_ONLY_WITH_IMMEDIATE);
 127                                 /* Immediate data comes after the BTH */
 128                                 ohdr->u.imm_data = wqe->wr.imm_data;
 129                                 hwords += 1;
 130                         }
 131                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 132                                 bth0 |= 1 << 23;
 133                         break;
 134
 135                 case IB_WR_RDMA_WRITE:
 136                 case IB_WR_RDMA_WRITE_WITH_IMM:
 137                         ohdr->u.rc.reth.vaddr =
 138                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 139                         ohdr->u.rc.reth.rkey =
 140                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 141                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 142                         hwords += sizeof(struct ib_reth) / 4;
 143                         if (len > pmtu) {
 144                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 145                                 len = pmtu;
 146                                 break;
 147                         }
 148                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 149                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 150                         else {
 151                                 qp->s_state =
 152                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 153                                 /* Immediate data comes after the RETH */
 154                                 ohdr->u.rc.imm_data = wqe->wr.imm_data;
 155                                 hwords += 1;
 156                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 157                                         bth0 |= 1 << 23;
 158                         }
 159                         break;
 160
 161                 default:
 162                         goto done;
 163                 }
 164                 if (++qp->s_tail >= qp->s_size)
 165                         qp->s_tail = 0;
 166                 break;
 167
 168         case OP(SEND_FIRST):
 169                 qp->s_state = OP(SEND_MIDDLE);
 170                 /* FALLTHROUGH */
 171         case OP(SEND_MIDDLE):
 172                 len = qp->s_len;
 173                 if (len > pmtu) {
 174                         len = pmtu;
 175                         break;
 176                 }
 177                 if (wqe->wr.opcode == IB_WR_SEND)
 178                         qp->s_state = OP(SEND_LAST);
 179                 else {
 180                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 181                         /* Immediate data comes after the BTH */
 182                         ohdr->u.imm_data = wqe->wr.imm_data;
 183                         hwords += 1;
 184                 }
 185                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 186                         bth0 |= 1 << 23;
 187                 break;
 188
 189         case OP(RDMA_WRITE_FIRST):
 190                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 191                 /* FALLTHROUGH */
 192         case OP(RDMA_WRITE_MIDDLE):
 193                 len = qp->s_len;
 194                 if (len > pmtu) {
 195                         len = pmtu;
 196                         break;
 197                 }
 198                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 199                         qp->s_state = OP(RDMA_WRITE_LAST);
 200                 else {
 201                         qp->s_state =
 202                                 OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 203                         /* Immediate data comes after the BTH */
 204                         ohdr->u.imm_data = wqe->wr.imm_data;
 205                         hwords += 1;
 206                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 207                                 bth0 |= 1 << 23;
 208                 }
 209                 break;
 210         }
 211         qp->s_len -= len;
 212         qp->s_hdrwords = hwords;
 213         qp->s_cur_sge = &qp->s_sge;
 214         qp->s_cur_size = len;
 215         *bth0p = bth0 | (qp->s_state << 24);
 216         *bth2p = qp->s_next_psn++ & IPATH_PSN_MASK;
 217         return 1;
 218
 219 done:
 220         return 0;
 221 }
 222
 223 /**
 224  * ipath_uc_rcv - handle an incoming UC packet
 225  * @dev: the device the packet came in on
 226  * @hdr: the header of the packet
 227  * @has_grh: true if the packet has a GRH
 228  * @data: the packet data
 229  * @tlen: the length of the packet
 230  * @qp: the QP for this packet.
 231  *
 232  * This is called from ipath_qp_rcv() to process an incoming UC packet
 233  * for the given QP.
 234  * Called at interrupt level.
 235  */
 236 void ipath_uc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
 237                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
 238 {
 239         struct ipath_other_headers *ohdr;
 240         int opcode;
 241         u32 hdrsize;
 242         u32 psn;
 243         u32 pad;
 244         struct ib_wc wc;
 245         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
 246         struct ib_reth *reth;
 247         int header_in_data;
 248
 249         /* Check for GRH */
 250         if (!has_grh) {
 251                 ohdr = &hdr->u.oth;
 252                 hdrsize = 8 + 12;       /* LRH + BTH */
 253                 psn = be32_to_cpu(ohdr->bth[2]);
 254                 header_in_data = 0;
 255         } else {
 256                 ohdr = &hdr->u.l.oth;
 257                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
 258                 /*
 259                  * The header with GRH is 60 bytes and the
 260                  * core driver sets the eager header buffer
 261                  * size to 56 bytes so the last 4 bytes of
 262                  * the BTH header (PSN) is in the data buffer.
 263                  */
 264                 header_in_data = dev->dd->ipath_rcvhdrentsize == 16;
 265                 if (header_in_data) {
 266                         psn = be32_to_cpu(((__be32 *) data)[0]);
 267                         data += sizeof(__be32);
 268                 } else
 269                         psn = be32_to_cpu(ohdr->bth[2]);
 270         }
 271         /*
 272          * The opcode is in the low byte when its in network order
 273          * (top byte when in host order).
 274          */
 275         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
 276
 277         wc.imm_data = 0;
 278         wc.wc_flags = 0;
 279
 280         /* Compare the PSN verses the expected PSN. */
 281         if (unlikely(ipath_cmp24(psn, qp->r_psn) != 0)) {
 282                 /*
 283                  * Handle a sequence error.
 284                  * Silently drop any current message.
 285                  */
 286                 qp->r_psn = psn;
 287         inv:
 288                 qp->r_state = OP(SEND_LAST);
 289                 switch (opcode) {
 290                 case OP(SEND_FIRST):
 291                 case OP(SEND_ONLY):
 292                 case OP(SEND_ONLY_WITH_IMMEDIATE):
 293                         goto send_first;
 294
 295                 case OP(RDMA_WRITE_FIRST):
 296                 case OP(RDMA_WRITE_ONLY):
 297                 case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
 298                         goto rdma_first;
 299
 300                 default:
 301                         dev->n_pkt_drops++;
 302                         goto done;
 303                 }
 304         }
 305
 306         /* Check for opcode sequence errors. */
 307         switch (qp->r_state) {
 308         case OP(SEND_FIRST):
 309         case OP(SEND_MIDDLE):
 310                 if (opcode == OP(SEND_MIDDLE) ||
 311                     opcode == OP(SEND_LAST) ||
 312                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
 313                         break;
 314                 goto inv;
 315
 316         case OP(RDMA_WRITE_FIRST):
 317         case OP(RDMA_WRITE_MIDDLE):
 318                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
 319                     opcode == OP(RDMA_WRITE_LAST) ||
 320                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
 321                         break;
 322                 goto inv;
 323
 324         default:
 325                 if (opcode == OP(SEND_FIRST) ||
 326                     opcode == OP(SEND_ONLY) ||
 327                     opcode == OP(SEND_ONLY_WITH_IMMEDIATE) ||
 328                     opcode == OP(RDMA_WRITE_FIRST) ||
 329                     opcode == OP(RDMA_WRITE_ONLY) ||
 330                     opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
 331                         break;
 332                 goto inv;
 333         }
 334
 335         /* OK, process the packet. */
 336         switch (opcode) {
 337         case OP(SEND_FIRST):
 338         case OP(SEND_ONLY):
 339         case OP(SEND_ONLY_WITH_IMMEDIATE):
 340         send_first:
 341                 if (qp->r_reuse_sge) {
 342                         qp->r_reuse_sge = 0;
 343                         qp->r_sge = qp->s_rdma_sge;
 344                 } else if (!ipath_get_rwqe(qp, 0)) {
 345                         dev->n_pkt_drops++;
 346                         goto done;
 347                 }
 348                 /* Save the WQE so we can reuse it in case of an error. */
 349                 qp->s_rdma_sge = qp->r_sge;
 350                 qp->r_rcv_len = 0;
 351                 if (opcode == OP(SEND_ONLY))
 352                         goto send_last;
 353                 else if (opcode == OP(SEND_ONLY_WITH_IMMEDIATE))
 354                         goto send_last_imm;
 355                 /* FALLTHROUGH */
 356         case OP(SEND_MIDDLE):
 357                 /* Check for invalid length PMTU or posted rwqe len. */
 358                 if (unlikely(tlen != (hdrsize + pmtu + 4))) {
 359                         qp->r_reuse_sge = 1;
 360                         dev->n_pkt_drops++;
 361                         goto done;
 362                 }
 363                 qp->r_rcv_len += pmtu;
 364                 if (unlikely(qp->r_rcv_len > qp->r_len)) {
 365                         qp->r_reuse_sge = 1;
 366                         dev->n_pkt_drops++;
 367                         goto done;
 368                 }
 369                 ipath_copy_sge(&qp->r_sge, data, pmtu);
 370                 break;
 371
 372         case OP(SEND_LAST_WITH_IMMEDIATE):
 373         send_last_imm:
 374                 if (header_in_data) {
 375                         wc.imm_data = *(__be32 *) data;
 376                         data += sizeof(__be32);
 377                 } else {
 378                         /* Immediate data comes after BTH */
 379                         wc.imm_data = ohdr->u.imm_data;
 380                 }
 381                 hdrsize += 4;
 382                 wc.wc_flags = IB_WC_WITH_IMM;
 383                 /* FALLTHROUGH */
 384         case OP(SEND_LAST):
 385         send_last:
 386                 /* Get the number of bytes the message was padded by. */
 387                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 388                 /* Check for invalid length. */
 389                 /* XXX LAST len should be >= 1 */
 390                 if (unlikely(tlen < (hdrsize + pad + 4))) {
 391                         qp->r_reuse_sge = 1;
 392                         dev->n_pkt_drops++;
 393                         goto done;
 394                 }
 395                 /* Don't count the CRC. */
 396                 tlen -= (hdrsize + pad + 4);
 397                 wc.byte_len = tlen + qp->r_rcv_len;
 398                 if (unlikely(wc.byte_len > qp->r_len)) {
 399                         qp->r_reuse_sge = 1;
 400                         dev->n_pkt_drops++;
 401                         goto done;
 402                 }
 403                 /* XXX Need to free SGEs */
 404         last_imm:
 405                 ipath_copy_sge(&qp->r_sge, data, tlen);
 406                 wc.wr_id = qp->r_wr_id;
 407                 wc.status = IB_WC_SUCCESS;
 408                 wc.opcode = IB_WC_RECV;
 409                 wc.vendor_err = 0;
 410                 wc.qp_num = qp->ibqp.qp_num;
 411                 wc.src_qp = qp->remote_qpn;
 412                 wc.pkey_index = 0;
 413                 wc.slid = qp->remote_ah_attr.dlid;
 414                 wc.sl = qp->remote_ah_attr.sl;
 415                 wc.dlid_path_bits = 0;
 416                 wc.port_num = 0;
 417                 /* Signal completion event if the solicited bit is set. */
 418                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
 419                                (ohdr->bth[0] &
 420                                 __constant_cpu_to_be32(1 << 23)) != 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                 /* RETH comes after BTH */
 428                 if (!header_in_data)
 429                         reth = &ohdr->u.rc.reth;
 430                 else {
 431                         reth = (struct ib_reth *)data;
 432                         data += sizeof(*reth);
 433                 }
 434                 hdrsize += sizeof(*reth);
 435                 qp->r_len = be32_to_cpu(reth->length);
 436                 qp->r_rcv_len = 0;
 437                 if (qp->r_len != 0) {
 438                         u32 rkey = be32_to_cpu(reth->rkey);
 439                         u64 vaddr = be64_to_cpu(reth->vaddr);
 440                         int ok;
 441
 442                         /* Check rkey */
 443                         ok = ipath_rkey_ok(dev, &qp->r_sge, qp->r_len,
 444                                            vaddr, rkey,
 445                                            IB_ACCESS_REMOTE_WRITE);
 446                         if (unlikely(!ok)) {
 447                                 dev->n_pkt_drops++;
 448                                 goto done;
 449                         }
 450                 } else {
 451                         qp->r_sge.sg_list = NULL;
 452                         qp->r_sge.sge.mr = NULL;
 453                         qp->r_sge.sge.vaddr = NULL;
 454                         qp->r_sge.sge.length = 0;
 455                         qp->r_sge.sge.sge_length = 0;
 456                 }
 457                 if (unlikely(!(qp->qp_access_flags &
 458                                IB_ACCESS_REMOTE_WRITE))) {
 459                         dev->n_pkt_drops++;
 460                         goto done;
 461                 }
 462                 if (opcode == OP(RDMA_WRITE_ONLY))
 463                         goto rdma_last;
 464                 else if (opcode == OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE))
 465                         goto rdma_last_imm;
 466                 /* FALLTHROUGH */
 467         case OP(RDMA_WRITE_MIDDLE):
 468                 /* Check for invalid length PMTU or posted rwqe len. */
 469                 if (unlikely(tlen != (hdrsize + pmtu + 4))) {
 470                         dev->n_pkt_drops++;
 471                         goto done;
 472                 }
 473                 qp->r_rcv_len += pmtu;
 474                 if (unlikely(qp->r_rcv_len > qp->r_len)) {
 475                         dev->n_pkt_drops++;
 476                         goto done;
 477                 }
 478                 ipath_copy_sge(&qp->r_sge, data, pmtu);
 479                 break;
 480
 481         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
 482         rdma_last_imm:
 483                 /* Get the number of bytes the message was padded by. */
 484                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 485                 /* Check for invalid length. */
 486                 /* XXX LAST len should be >= 1 */
 487                 if (unlikely(tlen < (hdrsize + pad + 4))) {
 488                         dev->n_pkt_drops++;
 489                         goto done;
 490                 }
 491                 /* Don't count the CRC. */
 492                 tlen -= (hdrsize + pad + 4);
 493                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) {
 494                         dev->n_pkt_drops++;
 495                         goto done;
 496                 }
 497                 if (qp->r_reuse_sge)
 498                         qp->r_reuse_sge = 0;
 499                 else if (!ipath_get_rwqe(qp, 1)) {
 500                         dev->n_pkt_drops++;
 501                         goto done;
 502                 }
 503                 if (header_in_data) {
 504                         wc.imm_data = *(__be32 *) data;
 505                         data += sizeof(__be32);
 506                 } else {
 507                         /* Immediate data comes after BTH */
 508                         wc.imm_data = ohdr->u.imm_data;
 509                 }
 510                 hdrsize += 4;
 511                 wc.wc_flags = IB_WC_WITH_IMM;
 512                 wc.byte_len = 0;
 513                 goto last_imm;
 514
 515         case OP(RDMA_WRITE_LAST):
 516         rdma_last:
 517                 /* Get the number of bytes the message was padded by. */
 518                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
 519                 /* Check for invalid length. */
 520                 /* XXX LAST len should be >= 1 */
 521                 if (unlikely(tlen < (hdrsize + pad + 4))) {
 522                         dev->n_pkt_drops++;
 523                         goto done;
 524                 }
 525                 /* Don't count the CRC. */
 526                 tlen -= (hdrsize + pad + 4);
 527                 if (unlikely(tlen + qp->r_rcv_len != qp->r_len)) {
 528                         dev->n_pkt_drops++;
 529                         goto done;
 530                 }
 531                 ipath_copy_sge(&qp->r_sge, data, tlen);
 532                 break;
 533
 534         default:
 535                 /* Drop packet for unknown opcodes. */
 536                 dev->n_pkt_drops++;
 537                 goto done;
 538         }
 539         qp->r_psn++;
 540         qp->r_state = opcode;
 541 done:
 542         return;
 543 }