git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/ipath/ipath_rc.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_common.h"
  36
  37 /* cut down ridiculously long IB macro names */
  38 #define OP(x) IB_OPCODE_RC_##x
  39
  40 /**
  41  * ipath_init_restart- initialize the qp->s_sge after a restart
  42  * @qp: the QP who's SGE we're restarting
  43  * @wqe: the work queue to initialize the QP's SGE from
  44  *
  45  * The QP s_lock should be held and interrupts disabled.
  46  */
  47 static void ipath_init_restart(struct ipath_qp *qp, struct ipath_swqe *wqe)
  48 {
  49         struct ipath_ibdev *dev;
  50         u32 len;
  51
  52         len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) *
  53                 ib_mtu_enum_to_int(qp->path_mtu);
  54         qp->s_sge.sge = wqe->sg_list[0];
  55         qp->s_sge.sg_list = wqe->sg_list + 1;
  56         qp->s_sge.num_sge = wqe->wr.num_sge;
  57         ipath_skip_sge(&qp->s_sge, len);
  58         qp->s_len = wqe->length - len;
  59         dev = to_idev(qp->ibqp.device);
  60         spin_lock(&dev->pending_lock);
  61         if (list_empty(&qp->timerwait))
  62                 list_add_tail(&qp->timerwait,
  63                               &dev->pending[dev->pending_index]);
  64         spin_unlock(&dev->pending_lock);
  65 }
  66
  67 /**
  68  * ipath_make_rc_ack - construct a response packet (ACK, NAK, or RDMA read)
  69  * @qp: a pointer to the QP
  70  * @ohdr: a pointer to the IB header being constructed
  71  * @pmtu: the path MTU
  72  *
  73  * Return bth0 if constructed; otherwise, return 0.
  74  * Note the QP s_lock must be held.
  75  */
  76 u32 ipath_make_rc_ack(struct ipath_qp *qp,
  77                       struct ipath_other_headers *ohdr,
  78                       u32 pmtu)
  79 {
  80         u32 hwords;
  81         u32 len;
  82         u32 bth0;
  83
  84         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
  85         hwords = 5;
  86
  87         /*
  88          * Send a response.  Note that we are in the responder's
  89          * side of the QP context.
  90          */
  91         switch (qp->s_ack_state) {
  92         case OP(RDMA_READ_REQUEST):
  93                 qp->s_cur_sge = &qp->s_rdma_sge;
  94                 len = qp->s_rdma_len;
  95                 if (len > pmtu) {
  96                         len = pmtu;
  97                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_FIRST);
  98                 } else
  99                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_ONLY);
 100                 qp->s_rdma_len -= len;
 101                 bth0 = qp->s_ack_state << 24;
 102                 ohdr->u.aeth = ipath_compute_aeth(qp);
 103                 hwords++;
 104                 break;
 105
 106         case OP(RDMA_READ_RESPONSE_FIRST):
 107                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 108                 /* FALLTHROUGH */
 109         case OP(RDMA_READ_RESPONSE_MIDDLE):
 110                 qp->s_cur_sge = &qp->s_rdma_sge;
 111                 len = qp->s_rdma_len;
 112                 if (len > pmtu)
 113                         len = pmtu;
 114                 else {
 115                         ohdr->u.aeth = ipath_compute_aeth(qp);
 116                         hwords++;
 117                         qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 118                 }
 119                 qp->s_rdma_len -= len;
 120                 bth0 = qp->s_ack_state << 24;
 121                 break;
 122
 123         case OP(RDMA_READ_RESPONSE_LAST):
 124         case OP(RDMA_READ_RESPONSE_ONLY):
 125                 /*
 126                  * We have to prevent new requests from changing
 127                  * the r_sge state while a ipath_verbs_send()
 128                  * is in progress.
 129                  */
 130                 qp->s_ack_state = OP(ACKNOWLEDGE);
 131                 bth0 = 0;
 132                 goto bail;
 133
 134         case OP(COMPARE_SWAP):
 135         case OP(FETCH_ADD):
 136                 qp->s_cur_sge = NULL;
 137                 len = 0;
 138                 /*
 139                  * Set the s_ack_state so the receive interrupt handler
 140                  * won't try to send an ACK (out of order) until this one
 141                  * is actually sent.
 142                  */
 143                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 144                 bth0 = OP(ATOMIC_ACKNOWLEDGE) << 24;
 145                 ohdr->u.at.aeth = ipath_compute_aeth(qp);
 146                 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->r_atomic_data);
 147                 hwords += sizeof(ohdr->u.at) / 4;
 148                 break;
 149
 150         default:
 151                 /* Send a regular ACK. */
 152                 qp->s_cur_sge = NULL;
 153                 len = 0;
 154                 /*
 155                  * Set the s_ack_state so the receive interrupt handler
 156                  * won't try to send an ACK (out of order) until this one
 157                  * is actually sent.
 158                  */
 159                 qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
 160                 bth0 = OP(ACKNOWLEDGE) << 24;
 161                 if (qp->s_nak_state)
 162                         ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 163                                                     (qp->s_nak_state <<
 164                                                      IPATH_AETH_CREDIT_SHIFT));
 165                 else
 166                         ohdr->u.aeth = ipath_compute_aeth(qp);
 167                 hwords++;
 168         }
 169         qp->s_hdrwords = hwords;
 170         qp->s_cur_size = len;
 171
 172 bail:
 173         return bth0;
 174 }
 175
 176 /**
 177  * ipath_make_rc_req - construct a request packet (SEND, RDMA r/w, ATOMIC)
 178  * @qp: a pointer to the QP
 179  * @ohdr: a pointer to the IB header being constructed
 180  * @pmtu: the path MTU
 181  * @bth0p: pointer to the BTH opcode word
 182  * @bth2p: pointer to the BTH PSN word
 183  *
 184  * Return 1 if constructed; otherwise, return 0.
 185  * Note the QP s_lock must be held and interrupts disabled.
 186  */
 187 int ipath_make_rc_req(struct ipath_qp *qp,
 188                       struct ipath_other_headers *ohdr,
 189                       u32 pmtu, u32 *bth0p, u32 *bth2p)
 190 {
 191         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 192         struct ipath_sge_state *ss;
 193         struct ipath_swqe *wqe;
 194         u32 hwords;
 195         u32 len;
 196         u32 bth0;
 197         u32 bth2;
 198         char newreq;
 199
 200         if (!(ib_ipath_state_ops[qp->state] & IPATH_PROCESS_SEND_OK) ||
 201             qp->s_rnr_timeout)
 202                 goto done;
 203
 204         /* header size in 32-bit words LRH+BTH = (8+12)/4. */
 205         hwords = 5;
 206         bth0 = 0;
 207
 208         /* Send a request. */
 209         wqe = get_swqe_ptr(qp, qp->s_cur);
 210         switch (qp->s_state) {
 211         default:
 212                 /*
 213                  * Resend an old request or start a new one.
 214                  *
 215                  * We keep track of the current SWQE so that
 216                  * we don't reset the "furthest progress" state
 217                  * if we need to back up.
 218                  */
 219                 newreq = 0;
 220                 if (qp->s_cur == qp->s_tail) {
 221                         /* Check if send work queue is empty. */
 222                         if (qp->s_tail == qp->s_head)
 223                                 goto done;
 224                         qp->s_psn = wqe->psn = qp->s_next_psn;
 225                         newreq = 1;
 226                 }
 227                 /*
 228                  * Note that we have to be careful not to modify the
 229                  * original work request since we may need to resend
 230                  * it.
 231                  */
 232                 qp->s_sge.sge = wqe->sg_list[0];
 233                 qp->s_sge.sg_list = wqe->sg_list + 1;
 234                 qp->s_sge.num_sge = wqe->wr.num_sge;
 235                 qp->s_len = len = wqe->length;
 236                 ss = &qp->s_sge;
 237                 bth2 = 0;
 238                 switch (wqe->wr.opcode) {
 239                 case IB_WR_SEND:
 240                 case IB_WR_SEND_WITH_IMM:
 241                         /* If no credit, return. */
 242                         if (qp->s_lsn != (u32) -1 &&
 243                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 244                                 goto done;
 245                         wqe->lpsn = wqe->psn;
 246                         if (len > pmtu) {
 247                                 wqe->lpsn += (len - 1) / pmtu;
 248                                 qp->s_state = OP(SEND_FIRST);
 249                                 len = pmtu;
 250                                 break;
 251                         }
 252                         if (wqe->wr.opcode == IB_WR_SEND)
 253                                 qp->s_state = OP(SEND_ONLY);
 254                         else {
 255                                 qp->s_state = OP(SEND_ONLY_WITH_IMMEDIATE);
 256                                 /* Immediate data comes after the BTH */
 257                                 ohdr->u.imm_data = wqe->wr.imm_data;
 258                                 hwords += 1;
 259                         }
 260                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 261                                 bth0 |= 1 << 23;
 262                         bth2 = 1 << 31; /* Request ACK. */
 263                         if (++qp->s_cur == qp->s_size)
 264                                 qp->s_cur = 0;
 265                         break;
 266
 267                 case IB_WR_RDMA_WRITE:
 268                         if (newreq && qp->s_lsn != (u32) -1)
 269                                 qp->s_lsn++;
 270                         /* FALLTHROUGH */
 271                 case IB_WR_RDMA_WRITE_WITH_IMM:
 272                         /* If no credit, return. */
 273                         if (qp->s_lsn != (u32) -1 &&
 274                             ipath_cmp24(wqe->ssn, qp->s_lsn + 1) > 0)
 275                                 goto done;
 276                         ohdr->u.rc.reth.vaddr =
 277                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 278                         ohdr->u.rc.reth.rkey =
 279                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 280                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 281                         hwords += sizeof(struct ib_reth) / 4;
 282                         wqe->lpsn = wqe->psn;
 283                         if (len > pmtu) {
 284                                 wqe->lpsn += (len - 1) / pmtu;
 285                                 qp->s_state = OP(RDMA_WRITE_FIRST);
 286                                 len = pmtu;
 287                                 break;
 288                         }
 289                         if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 290                                 qp->s_state = OP(RDMA_WRITE_ONLY);
 291                         else {
 292                                 qp->s_state =
 293                                         OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE);
 294                                 /* Immediate data comes after RETH */
 295                                 ohdr->u.rc.imm_data = wqe->wr.imm_data;
 296                                 hwords += 1;
 297                                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 298                                         bth0 |= 1 << 23;
 299                         }
 300                         bth2 = 1 << 31; /* Request ACK. */
 301                         if (++qp->s_cur == qp->s_size)
 302                                 qp->s_cur = 0;
 303                         break;
 304
 305                 case IB_WR_RDMA_READ:
 306                         ohdr->u.rc.reth.vaddr =
 307                                 cpu_to_be64(wqe->wr.wr.rdma.remote_addr);
 308                         ohdr->u.rc.reth.rkey =
 309                                 cpu_to_be32(wqe->wr.wr.rdma.rkey);
 310                         ohdr->u.rc.reth.length = cpu_to_be32(len);
 311                         qp->s_state = OP(RDMA_READ_REQUEST);
 312                         hwords += sizeof(ohdr->u.rc.reth) / 4;
 313                         if (newreq) {
 314                                 if (qp->s_lsn != (u32) -1)
 315                                         qp->s_lsn++;
 316                                 /*
 317                                  * Adjust s_next_psn to count the
 318                                  * expected number of responses.
 319                                  */
 320                                 if (len > pmtu)
 321                                         qp->s_next_psn += (len - 1) / pmtu;
 322                                 wqe->lpsn = qp->s_next_psn++;
 323                         }
 324                         ss = NULL;
 325                         len = 0;
 326                         if (++qp->s_cur == qp->s_size)
 327                                 qp->s_cur = 0;
 328                         break;
 329
 330                 case IB_WR_ATOMIC_CMP_AND_SWP:
 331                 case IB_WR_ATOMIC_FETCH_AND_ADD:
 332                         if (wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP)
 333                                 qp->s_state = OP(COMPARE_SWAP);
 334                         else
 335                                 qp->s_state = OP(FETCH_ADD);
 336                         ohdr->u.atomic_eth.vaddr = cpu_to_be64(
 337                                 wqe->wr.wr.atomic.remote_addr);
 338                         ohdr->u.atomic_eth.rkey = cpu_to_be32(
 339                                 wqe->wr.wr.atomic.rkey);
 340                         ohdr->u.atomic_eth.swap_data = cpu_to_be64(
 341                                 wqe->wr.wr.atomic.swap);
 342                         ohdr->u.atomic_eth.compare_data = cpu_to_be64(
 343                                 wqe->wr.wr.atomic.compare_add);
 344                         hwords += sizeof(struct ib_atomic_eth) / 4;
 345                         if (newreq) {
 346                                 if (qp->s_lsn != (u32) -1)
 347                                         qp->s_lsn++;
 348                                 wqe->lpsn = wqe->psn;
 349                         }
 350                         if (++qp->s_cur == qp->s_size)
 351                                 qp->s_cur = 0;
 352                         ss = NULL;
 353                         len = 0;
 354                         break;
 355
 356                 default:
 357                         goto done;
 358                 }
 359                 if (newreq) {
 360                         qp->s_tail++;
 361                         if (qp->s_tail >= qp->s_size)
 362                                 qp->s_tail = 0;
 363                 }
 364                 bth2 |= qp->s_psn++ & IPATH_PSN_MASK;
 365                 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
 366                         qp->s_next_psn = qp->s_psn;
 367                 /*
 368                  * Put the QP on the pending list so lost ACKs will cause
 369                  * a retry.  More than one request can be pending so the
 370                  * QP may already be on the dev->pending list.
 371                  */
 372                 spin_lock(&dev->pending_lock);
 373                 if (list_empty(&qp->timerwait))
 374                         list_add_tail(&qp->timerwait,
 375                                       &dev->pending[dev->pending_index]);
 376                 spin_unlock(&dev->pending_lock);
 377                 break;
 378
 379         case OP(RDMA_READ_RESPONSE_FIRST):
 380                 /*
 381                  * This case can only happen if a send is restarted.
 382                  * See ipath_restart_rc().
 383                  */
 384                 ipath_init_restart(qp, wqe);
 385                 /* FALLTHROUGH */
 386         case OP(SEND_FIRST):
 387                 qp->s_state = OP(SEND_MIDDLE);
 388                 /* FALLTHROUGH */
 389         case OP(SEND_MIDDLE):
 390                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 391                 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
 392                         qp->s_next_psn = qp->s_psn;
 393                 ss = &qp->s_sge;
 394                 len = qp->s_len;
 395                 if (len > pmtu) {
 396                         /*
 397                          * Request an ACK every 1/2 MB to avoid retransmit
 398                          * timeouts.
 399                          */
 400                         if (((wqe->length - len) % (512 * 1024)) == 0)
 401                                 bth2 |= 1 << 31;
 402                         len = pmtu;
 403                         break;
 404                 }
 405                 if (wqe->wr.opcode == IB_WR_SEND)
 406                         qp->s_state = OP(SEND_LAST);
 407                 else {
 408                         qp->s_state = OP(SEND_LAST_WITH_IMMEDIATE);
 409                         /* Immediate data comes after the BTH */
 410                         ohdr->u.imm_data = wqe->wr.imm_data;
 411                         hwords += 1;
 412                 }
 413                 if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 414                         bth0 |= 1 << 23;
 415                 bth2 |= 1 << 31;        /* Request ACK. */
 416                 qp->s_cur++;
 417                 if (qp->s_cur >= qp->s_size)
 418                         qp->s_cur = 0;
 419                 break;
 420
 421         case OP(RDMA_READ_RESPONSE_LAST):
 422                 /*
 423                  * This case can only happen if a RDMA write is restarted.
 424                  * See ipath_restart_rc().
 425                  */
 426                 ipath_init_restart(qp, wqe);
 427                 /* FALLTHROUGH */
 428         case OP(RDMA_WRITE_FIRST):
 429                 qp->s_state = OP(RDMA_WRITE_MIDDLE);
 430                 /* FALLTHROUGH */
 431         case OP(RDMA_WRITE_MIDDLE):
 432                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 433                 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
 434                         qp->s_next_psn = qp->s_psn;
 435                 ss = &qp->s_sge;
 436                 len = qp->s_len;
 437                 if (len > pmtu) {
 438                         /*
 439                          * Request an ACK every 1/2 MB to avoid retransmit
 440                          * timeouts.
 441                          */
 442                         if (((wqe->length - len) % (512 * 1024)) == 0)
 443                                 bth2 |= 1 << 31;
 444                         len = pmtu;
 445                         break;
 446                 }
 447                 if (wqe->wr.opcode == IB_WR_RDMA_WRITE)
 448                         qp->s_state = OP(RDMA_WRITE_LAST);
 449                 else {
 450                         qp->s_state = OP(RDMA_WRITE_LAST_WITH_IMMEDIATE);
 451                         /* Immediate data comes after the BTH */
 452                         ohdr->u.imm_data = wqe->wr.imm_data;
 453                         hwords += 1;
 454                         if (wqe->wr.send_flags & IB_SEND_SOLICITED)
 455                                 bth0 |= 1 << 23;
 456                 }
 457                 bth2 |= 1 << 31;        /* Request ACK. */
 458                 qp->s_cur++;
 459                 if (qp->s_cur >= qp->s_size)
 460                         qp->s_cur = 0;
 461                 break;
 462
 463         case OP(RDMA_READ_RESPONSE_MIDDLE):
 464                 /*
 465                  * This case can only happen if a RDMA read is restarted.
 466                  * See ipath_restart_rc().
 467                  */
 468                 ipath_init_restart(qp, wqe);
 469                 len = ((qp->s_psn - wqe->psn) & IPATH_PSN_MASK) * pmtu;
 470                 ohdr->u.rc.reth.vaddr =
 471                         cpu_to_be64(wqe->wr.wr.rdma.remote_addr + len);
 472                 ohdr->u.rc.reth.rkey =
 473                         cpu_to_be32(wqe->wr.wr.rdma.rkey);
 474                 ohdr->u.rc.reth.length = cpu_to_be32(qp->s_len);
 475                 qp->s_state = OP(RDMA_READ_REQUEST);
 476                 hwords += sizeof(ohdr->u.rc.reth) / 4;
 477                 bth2 = qp->s_psn++ & IPATH_PSN_MASK;
 478                 if ((int)(qp->s_psn - qp->s_next_psn) > 0)
 479                         qp->s_next_psn = qp->s_psn;
 480                 ss = NULL;
 481                 len = 0;
 482                 qp->s_cur++;
 483                 if (qp->s_cur == qp->s_size)
 484                         qp->s_cur = 0;
 485                 break;
 486
 487         case OP(RDMA_READ_REQUEST):
 488         case OP(COMPARE_SWAP):
 489         case OP(FETCH_ADD):
 490                 /*
 491                  * We shouldn't start anything new until this request is
 492                  * finished.  The ACK will handle rescheduling us.  XXX The
 493                  * number of outstanding ones is negotiated at connection
 494                  * setup time (see pg. 258,289)?  XXX Also, if we support
 495                  * multiple outstanding requests, we need to check the WQE
 496                  * IB_SEND_FENCE flag and not send a new request if a RDMA
 497                  * read or atomic is pending.
 498                  */
 499                 goto done;
 500         }
 501         qp->s_len -= len;
 502         qp->s_hdrwords = hwords;
 503         qp->s_cur_sge = ss;
 504         qp->s_cur_size = len;
 505         *bth0p = bth0 | (qp->s_state << 24);
 506         *bth2p = bth2;
 507         return 1;
 508
 509 done:
 510         return 0;
 511 }
 512
 513 /**
 514  * send_rc_ack - Construct an ACK packet and send it
 515  * @qp: a pointer to the QP
 516  *
 517  * This is called from ipath_rc_rcv() and only uses the receive
 518  * side QP state.
 519  * Note that RDMA reads are handled in the send side QP state and tasklet.
 520  */
 521 static void send_rc_ack(struct ipath_qp *qp)
 522 {
 523         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 524         u16 lrh0;
 525         u32 bth0;
 526         u32 hwords;
 527         struct ipath_ib_header hdr;
 528         struct ipath_other_headers *ohdr;
 529
 530         /* Construct the header. */
 531         ohdr = &hdr.u.oth;
 532         lrh0 = IPATH_LRH_BTH;
 533         /* header size in 32-bit words LRH+BTH+AETH = (8+12+4)/4. */
 534         hwords = 6;
 535         if (unlikely(qp->remote_ah_attr.ah_flags & IB_AH_GRH)) {
 536                 hwords += ipath_make_grh(dev, &hdr.u.l.grh,
 537                                          &qp->remote_ah_attr.grh,
 538                                          hwords, 0);
 539                 ohdr = &hdr.u.l.oth;
 540                 lrh0 = IPATH_LRH_GRH;
 541         }
 542         /* read pkey_index w/o lock (its atomic) */
 543         bth0 = ipath_layer_get_pkey(dev->dd, qp->s_pkey_index);
 544         if (qp->r_nak_state)
 545                 ohdr->u.aeth = cpu_to_be32((qp->r_msn & IPATH_MSN_MASK) |
 546                                             (qp->r_nak_state <<
 547                                              IPATH_AETH_CREDIT_SHIFT));
 548         else
 549                 ohdr->u.aeth = ipath_compute_aeth(qp);
 550         if (qp->r_ack_state >= OP(COMPARE_SWAP)) {
 551                 bth0 |= OP(ATOMIC_ACKNOWLEDGE) << 24;
 552                 ohdr->u.at.atomic_ack_eth = cpu_to_be64(qp->r_atomic_data);
 553                 hwords += sizeof(ohdr->u.at.atomic_ack_eth) / 4;
 554         } else
 555                 bth0 |= OP(ACKNOWLEDGE) << 24;
 556         lrh0 |= qp->remote_ah_attr.sl << 4;
 557         hdr.lrh[0] = cpu_to_be16(lrh0);
 558         hdr.lrh[1] = cpu_to_be16(qp->remote_ah_attr.dlid);
 559         hdr.lrh[2] = cpu_to_be16(hwords + SIZE_OF_CRC);
 560         hdr.lrh[3] = cpu_to_be16(ipath_layer_get_lid(dev->dd));
 561         ohdr->bth[0] = cpu_to_be32(bth0);
 562         ohdr->bth[1] = cpu_to_be32(qp->remote_qpn);
 563         ohdr->bth[2] = cpu_to_be32(qp->r_ack_psn & IPATH_PSN_MASK);
 564
 565         /*
 566          * If we can send the ACK, clear the ACK state.
 567          */
 568         if (ipath_verbs_send(dev->dd, hwords, (u32 *) &hdr, 0, NULL) == 0) {
 569                 qp->r_ack_state = OP(ACKNOWLEDGE);
 570                 dev->n_unicast_xmit++;
 571         } else {
 572                 /*
 573                  * We are out of PIO buffers at the moment.
 574                  * Pass responsibility for sending the ACK to the
 575                  * send tasklet so that when a PIO buffer becomes
 576                  * available, the ACK is sent ahead of other outgoing
 577                  * packets.
 578                  */
 579                 dev->n_rc_qacks++;
 580                 spin_lock_irq(&qp->s_lock);
 581                 /* Don't coalesce if a RDMA read or atomic is pending. */
 582                 if (qp->s_ack_state == OP(ACKNOWLEDGE) ||
 583                     qp->s_ack_state < OP(RDMA_READ_REQUEST)) {
 584                         qp->s_ack_state = qp->r_ack_state;
 585                         qp->s_nak_state = qp->r_nak_state;
 586                         qp->s_ack_psn = qp->r_ack_psn;
 587                         qp->r_ack_state = OP(ACKNOWLEDGE);
 588                 }
 589                 spin_unlock_irq(&qp->s_lock);
 590
 591                 /* Call ipath_do_rc_send() in another thread. */
 592                 tasklet_hi_schedule(&qp->s_task);
 593         }
 594 }
 595
 596 /**
 597  * reset_psn - reset the QP state to send starting from PSN
 598  * @qp: the QP
 599  * @psn: the packet sequence number to restart at
 600  *
 601  * This is called from ipath_rc_rcv() to process an incoming RC ACK
 602  * for the given QP.
 603  * Called at interrupt level with the QP s_lock held.
 604  */
 605 static void reset_psn(struct ipath_qp *qp, u32 psn)
 606 {
 607         u32 n = qp->s_last;
 608         struct ipath_swqe *wqe = get_swqe_ptr(qp, n);
 609         u32 opcode;
 610
 611         qp->s_cur = n;
 612
 613         /*
 614          * If we are starting the request from the beginning,
 615          * let the normal send code handle initialization.
 616          */
 617         if (ipath_cmp24(psn, wqe->psn) <= 0) {
 618                 qp->s_state = OP(SEND_LAST);
 619                 goto done;
 620         }
 621
 622         /* Find the work request opcode corresponding to the given PSN. */
 623         opcode = wqe->wr.opcode;
 624         for (;;) {
 625                 int diff;
 626
 627                 if (++n == qp->s_size)
 628                         n = 0;
 629                 if (n == qp->s_tail)
 630                         break;
 631                 wqe = get_swqe_ptr(qp, n);
 632                 diff = ipath_cmp24(psn, wqe->psn);
 633                 if (diff < 0)
 634                         break;
 635                 qp->s_cur = n;
 636                 /*
 637                  * If we are starting the request from the beginning,
 638                  * let the normal send code handle initialization.
 639                  */
 640                 if (diff == 0) {
 641                         qp->s_state = OP(SEND_LAST);
 642                         goto done;
 643                 }
 644                 opcode = wqe->wr.opcode;
 645         }
 646
 647         /*
 648          * Set the state to restart in the middle of a request.
 649          * Don't change the s_sge, s_cur_sge, or s_cur_size.
 650          * See ipath_do_rc_send().
 651          */
 652         switch (opcode) {
 653         case IB_WR_SEND:
 654         case IB_WR_SEND_WITH_IMM:
 655                 qp->s_state = OP(RDMA_READ_RESPONSE_FIRST);
 656                 break;
 657
 658         case IB_WR_RDMA_WRITE:
 659         case IB_WR_RDMA_WRITE_WITH_IMM:
 660                 qp->s_state = OP(RDMA_READ_RESPONSE_LAST);
 661                 break;
 662
 663         case IB_WR_RDMA_READ:
 664                 qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 665                 break;
 666
 667         default:
 668                 /*
 669                  * This case shouldn't happen since its only
 670                  * one PSN per req.
 671                  */
 672                 qp->s_state = OP(SEND_LAST);
 673         }
 674 done:
 675         qp->s_psn = psn;
 676 }
 677
 678 /**
 679  * ipath_restart_rc - back up requester to resend the last un-ACKed request
 680  * @qp: the QP to restart
 681  * @psn: packet sequence number for the request
 682  * @wc: the work completion request
 683  *
 684  * The QP s_lock should be held and interrupts disabled.
 685  */
 686 void ipath_restart_rc(struct ipath_qp *qp, u32 psn, struct ib_wc *wc)
 687 {
 688         struct ipath_swqe *wqe = get_swqe_ptr(qp, qp->s_last);
 689         struct ipath_ibdev *dev;
 690
 691         /*
 692          * If there are no requests pending, we are done.
 693          */
 694         if (ipath_cmp24(psn, qp->s_next_psn) >= 0 ||
 695             qp->s_last == qp->s_tail)
 696                 goto done;
 697
 698         if (qp->s_retry == 0) {
 699                 wc->wr_id = wqe->wr.wr_id;
 700                 wc->status = IB_WC_RETRY_EXC_ERR;
 701                 wc->opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 702                 wc->vendor_err = 0;
 703                 wc->byte_len = 0;
 704                 wc->qp_num = qp->ibqp.qp_num;
 705                 wc->src_qp = qp->remote_qpn;
 706                 wc->pkey_index = 0;
 707                 wc->slid = qp->remote_ah_attr.dlid;
 708                 wc->sl = qp->remote_ah_attr.sl;
 709                 wc->dlid_path_bits = 0;
 710                 wc->port_num = 0;
 711                 ipath_sqerror_qp(qp, wc);
 712                 goto bail;
 713         }
 714         qp->s_retry--;
 715
 716         /*
 717          * Remove the QP from the timeout queue.
 718          * Note: it may already have been removed by ipath_ib_timer().
 719          */
 720         dev = to_idev(qp->ibqp.device);
 721         spin_lock(&dev->pending_lock);
 722         if (!list_empty(&qp->timerwait))
 723                 list_del_init(&qp->timerwait);
 724         spin_unlock(&dev->pending_lock);
 725
 726         if (wqe->wr.opcode == IB_WR_RDMA_READ)
 727                 dev->n_rc_resends++;
 728         else
 729                 dev->n_rc_resends += (int)qp->s_psn - (int)psn;
 730
 731         reset_psn(qp, psn);
 732
 733 done:
 734         tasklet_hi_schedule(&qp->s_task);
 735
 736 bail:
 737         return;
 738 }
 739
 740 /**
 741  * do_rc_ack - process an incoming RC ACK
 742  * @qp: the QP the ACK came in on
 743  * @psn: the packet sequence number of the ACK
 744  * @opcode: the opcode of the request that resulted in the ACK
 745  *
 746  * This is called from ipath_rc_rcv_resp() to process an incoming RC ACK
 747  * for the given QP.
 748  * Called at interrupt level with the QP s_lock held and interrupts disabled.
 749  * Returns 1 if OK, 0 if current operation should be aborted (NAK).
 750  */
 751 static int do_rc_ack(struct ipath_qp *qp, u32 aeth, u32 psn, int opcode)
 752 {
 753         struct ipath_ibdev *dev = to_idev(qp->ibqp.device);
 754         struct ib_wc wc;
 755         struct ipath_swqe *wqe;
 756         int ret = 0;
 757
 758         /*
 759          * Remove the QP from the timeout queue (or RNR timeout queue).
 760          * If ipath_ib_timer() has already removed it,
 761          * it's OK since we hold the QP s_lock and ipath_restart_rc()
 762          * just won't find anything to restart if we ACK everything.
 763          */
 764         spin_lock(&dev->pending_lock);
 765         if (!list_empty(&qp->timerwait))
 766                 list_del_init(&qp->timerwait);
 767         spin_unlock(&dev->pending_lock);
 768
 769         /*
 770          * Note that NAKs implicitly ACK outstanding SEND and RDMA write
 771          * requests and implicitly NAK RDMA read and atomic requests issued
 772          * before the NAK'ed request.  The MSN won't include the NAK'ed
 773          * request but will include an ACK'ed request(s).
 774          */
 775         wqe = get_swqe_ptr(qp, qp->s_last);
 776
 777         /* Nothing is pending to ACK/NAK. */
 778         if (qp->s_last == qp->s_tail)
 779                 goto bail;
 780
 781         /*
 782          * The MSN might be for a later WQE than the PSN indicates so
 783          * only complete WQEs that the PSN finishes.
 784          */
 785         while (ipath_cmp24(psn, wqe->lpsn) >= 0) {
 786                 /* If we are ACKing a WQE, the MSN should be >= the SSN. */
 787                 if (ipath_cmp24(aeth, wqe->ssn) < 0)
 788                         break;
 789                 /*
 790                  * If this request is a RDMA read or atomic, and the ACK is
 791                  * for a later operation, this ACK NAKs the RDMA read or
 792                  * atomic.  In other words, only a RDMA_READ_LAST or ONLY
 793                  * can ACK a RDMA read and likewise for atomic ops.  Note
 794                  * that the NAK case can only happen if relaxed ordering is
 795                  * used and requests are sent after an RDMA read or atomic
 796                  * is sent but before the response is received.
 797                  */
 798                 if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 799                      opcode != OP(RDMA_READ_RESPONSE_LAST)) ||
 800                     ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 801                       wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 802                      (opcode != OP(ATOMIC_ACKNOWLEDGE) ||
 803                       ipath_cmp24(wqe->psn, psn) != 0))) {
 804                         /*
 805                          * The last valid PSN seen is the previous
 806                          * request's.
 807                          */
 808                         qp->s_last_psn = wqe->psn - 1;
 809                         /* Retry this request. */
 810                         ipath_restart_rc(qp, wqe->psn, &wc);
 811                         /*
 812                          * No need to process the ACK/NAK since we are
 813                          * restarting an earlier request.
 814                          */
 815                         goto bail;
 816                 }
 817                 /* Post a send completion queue entry if requested. */
 818                 if (!test_bit(IPATH_S_SIGNAL_REQ_WR, &qp->s_flags) ||
 819                     (wqe->wr.send_flags & IB_SEND_SIGNALED)) {
 820                         wc.wr_id = wqe->wr.wr_id;
 821                         wc.status = IB_WC_SUCCESS;
 822                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 823                         wc.vendor_err = 0;
 824                         wc.byte_len = wqe->length;
 825                         wc.qp_num = qp->ibqp.qp_num;
 826                         wc.src_qp = qp->remote_qpn;
 827                         wc.pkey_index = 0;
 828                         wc.slid = qp->remote_ah_attr.dlid;
 829                         wc.sl = qp->remote_ah_attr.sl;
 830                         wc.dlid_path_bits = 0;
 831                         wc.port_num = 0;
 832                         ipath_cq_enter(to_icq(qp->ibqp.send_cq), &wc, 0);
 833                 }
 834                 qp->s_retry = qp->s_retry_cnt;
 835                 /*
 836                  * If we are completing a request which is in the process of
 837                  * being resent, we can stop resending it since we know the
 838                  * responder has already seen it.
 839                  */
 840                 if (qp->s_last == qp->s_cur) {
 841                         if (++qp->s_cur >= qp->s_size)
 842                                 qp->s_cur = 0;
 843                         wqe = get_swqe_ptr(qp, qp->s_cur);
 844                         qp->s_state = OP(SEND_LAST);
 845                         qp->s_psn = wqe->psn;
 846                 }
 847                 if (++qp->s_last >= qp->s_size)
 848                         qp->s_last = 0;
 849                 wqe = get_swqe_ptr(qp, qp->s_last);
 850                 if (qp->s_last == qp->s_tail)
 851                         break;
 852         }
 853
 854         switch (aeth >> 29) {
 855         case 0:         /* ACK */
 856                 dev->n_rc_acks++;
 857                 /* If this is a partial ACK, reset the retransmit timer. */
 858                 if (qp->s_last != qp->s_tail) {
 859                         spin_lock(&dev->pending_lock);
 860                         list_add_tail(&qp->timerwait,
 861                                       &dev->pending[dev->pending_index]);
 862                         spin_unlock(&dev->pending_lock);
 863                 }
 864                 ipath_get_credit(qp, aeth);
 865                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
 866                 qp->s_retry = qp->s_retry_cnt;
 867                 qp->s_last_psn = psn;
 868                 ret = 1;
 869                 goto bail;
 870
 871         case 1:         /* RNR NAK */
 872                 dev->n_rnr_naks++;
 873                 if (qp->s_rnr_retry == 0) {
 874                         if (qp->s_last == qp->s_tail)
 875                                 goto bail;
 876
 877                         wc.status = IB_WC_RNR_RETRY_EXC_ERR;
 878                         goto class_b;
 879                 }
 880                 if (qp->s_rnr_retry_cnt < 7)
 881                         qp->s_rnr_retry--;
 882                 if (qp->s_last == qp->s_tail)
 883                         goto bail;
 884
 885                 /* The last valid PSN is the previous PSN. */
 886                 qp->s_last_psn = psn - 1;
 887
 888                 dev->n_rc_resends += (int)qp->s_psn - (int)psn;
 889
 890                 reset_psn(qp, psn);
 891
 892                 qp->s_rnr_timeout =
 893                         ib_ipath_rnr_table[(aeth >> IPATH_AETH_CREDIT_SHIFT) &
 894                                            IPATH_AETH_CREDIT_MASK];
 895                 ipath_insert_rnr_queue(qp);
 896                 goto bail;
 897
 898         case 3:         /* NAK */
 899                 /* The last valid PSN seen is the previous request's. */
 900                 if (qp->s_last != qp->s_tail)
 901                         qp->s_last_psn = wqe->psn - 1;
 902                 switch ((aeth >> IPATH_AETH_CREDIT_SHIFT) &
 903                         IPATH_AETH_CREDIT_MASK) {
 904                 case 0: /* PSN sequence error */
 905                         dev->n_seq_naks++;
 906                         /*
 907                          * Back up to the responder's expected PSN.  XXX
 908                          * Note that we might get a NAK in the middle of an
 909                          * RDMA READ response which terminates the RDMA
 910                          * READ.
 911                          */
 912                         if (qp->s_last == qp->s_tail)
 913                                 break;
 914
 915                         if (ipath_cmp24(psn, wqe->psn) < 0)
 916                                 break;
 917
 918                         /* Retry the request. */
 919                         ipath_restart_rc(qp, psn, &wc);
 920                         break;
 921
 922                 case 1: /* Invalid Request */
 923                         wc.status = IB_WC_REM_INV_REQ_ERR;
 924                         dev->n_other_naks++;
 925                         goto class_b;
 926
 927                 case 2: /* Remote Access Error */
 928                         wc.status = IB_WC_REM_ACCESS_ERR;
 929                         dev->n_other_naks++;
 930                         goto class_b;
 931
 932                 case 3: /* Remote Operation Error */
 933                         wc.status = IB_WC_REM_OP_ERR;
 934                         dev->n_other_naks++;
 935                 class_b:
 936                         wc.wr_id = wqe->wr.wr_id;
 937                         wc.opcode = ib_ipath_wc_opcode[wqe->wr.opcode];
 938                         wc.vendor_err = 0;
 939                         wc.byte_len = 0;
 940                         wc.qp_num = qp->ibqp.qp_num;
 941                         wc.src_qp = qp->remote_qpn;
 942                         wc.pkey_index = 0;
 943                         wc.slid = qp->remote_ah_attr.dlid;
 944                         wc.sl = qp->remote_ah_attr.sl;
 945                         wc.dlid_path_bits = 0;
 946                         wc.port_num = 0;
 947                         ipath_sqerror_qp(qp, &wc);
 948                         break;
 949
 950                 default:
 951                         /* Ignore other reserved NAK error codes */
 952                         goto reserved;
 953                 }
 954                 qp->s_rnr_retry = qp->s_rnr_retry_cnt;
 955                 goto bail;
 956
 957         default:                /* 2: reserved */
 958         reserved:
 959                 /* Ignore reserved NAK codes. */
 960                 goto bail;
 961         }
 962
 963 bail:
 964         return ret;
 965 }
 966
 967 /**
 968  * ipath_rc_rcv_resp - process an incoming RC response packet
 969  * @dev: the device this packet came in on
 970  * @ohdr: the other headers for this packet
 971  * @data: the packet data
 972  * @tlen: the packet length
 973  * @qp: the QP for this packet
 974  * @opcode: the opcode for this packet
 975  * @psn: the packet sequence number for this packet
 976  * @hdrsize: the header length
 977  * @pmtu: the path MTU
 978  * @header_in_data: true if part of the header data is in the data buffer
 979  *
 980  * This is called from ipath_rc_rcv() to process an incoming RC response
 981  * packet for the given QP.
 982  * Called at interrupt level.
 983  */
 984 static inline void ipath_rc_rcv_resp(struct ipath_ibdev *dev,
 985                                      struct ipath_other_headers *ohdr,
 986                                      void *data, u32 tlen,
 987                                      struct ipath_qp *qp,
 988                                      u32 opcode,
 989                                      u32 psn, u32 hdrsize, u32 pmtu,
 990                                      int header_in_data)
 991 {
 992         unsigned long flags;
 993         struct ib_wc wc;
 994         int diff;
 995         u32 pad;
 996         u32 aeth;
 997
 998         spin_lock_irqsave(&qp->s_lock, flags);
 999
1000         /* Ignore invalid responses. */
1001         if (ipath_cmp24(psn, qp->s_next_psn) >= 0)
1002                 goto ack_done;
1003
1004         /* Ignore duplicate responses. */
1005         diff = ipath_cmp24(psn, qp->s_last_psn);
1006         if (unlikely(diff <= 0)) {
1007                 /* Update credits for "ghost" ACKs */
1008                 if (diff == 0 && opcode == OP(ACKNOWLEDGE)) {
1009                         if (!header_in_data)
1010                                 aeth = be32_to_cpu(ohdr->u.aeth);
1011                         else {
1012                                 aeth = be32_to_cpu(((__be32 *) data)[0]);
1013                                 data += sizeof(__be32);
1014                         }
1015                         if ((aeth >> 29) == 0)
1016                                 ipath_get_credit(qp, aeth);
1017                 }
1018                 goto ack_done;
1019         }
1020
1021         switch (opcode) {
1022         case OP(ACKNOWLEDGE):
1023         case OP(ATOMIC_ACKNOWLEDGE):
1024         case OP(RDMA_READ_RESPONSE_FIRST):
1025                 if (!header_in_data)
1026                         aeth = be32_to_cpu(ohdr->u.aeth);
1027                 else {
1028                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1029                         data += sizeof(__be32);
1030                 }
1031                 if (opcode == OP(ATOMIC_ACKNOWLEDGE))
1032                         *(u64 *) qp->s_sge.sge.vaddr = *(u64 *) data;
1033                 if (!do_rc_ack(qp, aeth, psn, opcode) ||
1034                     opcode != OP(RDMA_READ_RESPONSE_FIRST))
1035                         goto ack_done;
1036                 hdrsize += 4;
1037                 /*
1038                  * do_rc_ack() has already checked the PSN so skip
1039                  * the sequence check.
1040                  */
1041                 goto rdma_read;
1042
1043         case OP(RDMA_READ_RESPONSE_MIDDLE):
1044                 /* no AETH, no ACK */
1045                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1046                         dev->n_rdma_seq++;
1047                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1048                         goto ack_done;
1049                 }
1050         rdma_read:
1051                 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1052                         goto ack_done;
1053                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1054                         goto ack_done;
1055                 if (unlikely(pmtu >= qp->s_len))
1056                         goto ack_done;
1057                 /* We got a response so update the timeout. */
1058                 if (unlikely(qp->s_last == qp->s_tail ||
1059                              get_swqe_ptr(qp, qp->s_last)->wr.opcode !=
1060                              IB_WR_RDMA_READ))
1061                         goto ack_done;
1062                 spin_lock(&dev->pending_lock);
1063                 if (qp->s_rnr_timeout == 0 && !list_empty(&qp->timerwait))
1064                         list_move_tail(&qp->timerwait,
1065                                        &dev->pending[dev->pending_index]);
1066                 spin_unlock(&dev->pending_lock);
1067                 /*
1068                  * Update the RDMA receive state but do the copy w/o
1069                  * holding the locks and blocking interrupts.
1070                  * XXX Yet another place that affects relaxed RDMA order
1071                  * since we don't want s_sge modified.
1072                  */
1073                 qp->s_len -= pmtu;
1074                 qp->s_last_psn = psn;
1075                 spin_unlock_irqrestore(&qp->s_lock, flags);
1076                 ipath_copy_sge(&qp->s_sge, data, pmtu);
1077                 goto bail;
1078
1079         case OP(RDMA_READ_RESPONSE_LAST):
1080                 /* ACKs READ req. */
1081                 if (unlikely(ipath_cmp24(psn, qp->s_last_psn + 1))) {
1082                         dev->n_rdma_seq++;
1083                         ipath_restart_rc(qp, qp->s_last_psn + 1, &wc);
1084                         goto ack_done;
1085                 }
1086                 /* FALLTHROUGH */
1087         case OP(RDMA_READ_RESPONSE_ONLY):
1088                 if (unlikely(qp->s_state != OP(RDMA_READ_REQUEST)))
1089                         goto ack_done;
1090                 /*
1091                  * Get the number of bytes the message was padded by.
1092                  */
1093                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1094                 /*
1095                  * Check that the data size is >= 1 && <= pmtu.
1096                  * Remember to account for the AETH header (4) and
1097                  * ICRC (4).
1098                  */
1099                 if (unlikely(tlen <= (hdrsize + pad + 8))) {
1100                         /* XXX Need to generate an error CQ entry. */
1101                         goto ack_done;
1102                 }
1103                 tlen -= hdrsize + pad + 8;
1104                 if (unlikely(tlen != qp->s_len)) {
1105                         /* XXX Need to generate an error CQ entry. */
1106                         goto ack_done;
1107                 }
1108                 if (!header_in_data)
1109                         aeth = be32_to_cpu(ohdr->u.aeth);
1110                 else {
1111                         aeth = be32_to_cpu(((__be32 *) data)[0]);
1112                         data += sizeof(__be32);
1113                 }
1114                 ipath_copy_sge(&qp->s_sge, data, tlen);
1115                 if (do_rc_ack(qp, aeth, psn, OP(RDMA_READ_RESPONSE_LAST))) {
1116                         /*
1117                          * Change the state so we contimue
1118                          * processing new requests and wake up the
1119                          * tasklet if there are posted sends.
1120                          */
1121                         qp->s_state = OP(SEND_LAST);
1122                         if (qp->s_tail != qp->s_head)
1123                                 tasklet_hi_schedule(&qp->s_task);
1124                 }
1125                 goto ack_done;
1126         }
1127
1128 ack_done:
1129         spin_unlock_irqrestore(&qp->s_lock, flags);
1130 bail:
1131         return;
1132 }
1133
1134 /**
1135  * ipath_rc_rcv_error - process an incoming duplicate or error RC packet
1136  * @dev: the device this packet came in on
1137  * @ohdr: the other headers for this packet
1138  * @data: the packet data
1139  * @qp: the QP for this packet
1140  * @opcode: the opcode for this packet
1141  * @psn: the packet sequence number for this packet
1142  * @diff: the difference between the PSN and the expected PSN
1143  * @header_in_data: true if part of the header data is in the data buffer
1144  *
1145  * This is called from ipath_rc_rcv() to process an unexpected
1146  * incoming RC packet for the given QP.
1147  * Called at interrupt level.
1148  * Return 1 if no more processing is needed; otherwise return 0 to
1149  * schedule a response to be sent and the s_lock unlocked.
1150  */
1151 static inline int ipath_rc_rcv_error(struct ipath_ibdev *dev,
1152                                      struct ipath_other_headers *ohdr,
1153                                      void *data,
1154                                      struct ipath_qp *qp,
1155                                      u32 opcode,
1156                                      u32 psn,
1157                                      int diff,
1158                                      int header_in_data)
1159 {
1160         struct ib_reth *reth;
1161
1162         if (diff > 0) {
1163                 /*
1164                  * Packet sequence error.
1165                  * A NAK will ACK earlier sends and RDMA writes.
1166                  * Don't queue the NAK if a RDMA read, atomic, or
1167                  * NAK is pending though.
1168                  */
1169                 if (qp->s_ack_state != OP(ACKNOWLEDGE) ||
1170                     qp->r_nak_state != 0)
1171                         goto done;
1172                 if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1173                         qp->r_ack_state = OP(SEND_ONLY);
1174                         qp->r_nak_state = IB_NAK_PSN_ERROR;
1175                         /* Use the expected PSN. */
1176                         qp->r_ack_psn = qp->r_psn;
1177                 }
1178                 goto send_ack;
1179         }
1180
1181         /*
1182          * Handle a duplicate request.  Don't re-execute SEND, RDMA
1183          * write or atomic op.  Don't NAK errors, just silently drop
1184          * the duplicate request.  Note that r_sge, r_len, and
1185          * r_rcv_len may be in use so don't modify them.
1186          *
1187          * We are supposed to ACK the earliest duplicate PSN but we
1188          * can coalesce an outstanding duplicate ACK.  We have to
1189          * send the earliest so that RDMA reads can be restarted at
1190          * the requester's expected PSN.
1191          */
1192         if (opcode == OP(RDMA_READ_REQUEST)) {
1193                 /* RETH comes after BTH */
1194                 if (!header_in_data)
1195                         reth = &ohdr->u.rc.reth;
1196                 else {
1197                         reth = (struct ib_reth *)data;
1198                         data += sizeof(*reth);
1199                 }
1200                 /*
1201                  * If we receive a duplicate RDMA request, it means the
1202                  * requester saw a sequence error and needs to restart
1203                  * from an earlier point.  We can abort the current
1204                  * RDMA read send in that case.
1205                  */
1206                 spin_lock_irq(&qp->s_lock);
1207                 if (qp->s_ack_state != OP(ACKNOWLEDGE) &&
1208                     (qp->s_hdrwords || ipath_cmp24(psn, qp->s_ack_psn) >= 0)) {
1209                         /*
1210                          * We are already sending earlier requested data.
1211                          * Don't abort it to send later out of sequence data.
1212                          */
1213                         spin_unlock_irq(&qp->s_lock);
1214                         goto done;
1215                 }
1216                 qp->s_rdma_len = be32_to_cpu(reth->length);
1217                 if (qp->s_rdma_len != 0) {
1218                         u32 rkey = be32_to_cpu(reth->rkey);
1219                         u64 vaddr = be64_to_cpu(reth->vaddr);
1220                         int ok;
1221
1222                         /*
1223                          * Address range must be a subset of the original
1224                          * request and start on pmtu boundaries.
1225                          */
1226                         ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
1227                                            qp->s_rdma_len, vaddr, rkey,
1228                                            IB_ACCESS_REMOTE_READ);
1229                         if (unlikely(!ok)) {
1230                                 spin_unlock_irq(&qp->s_lock);
1231                                 goto done;
1232                         }
1233                 } else {
1234                         qp->s_rdma_sge.sg_list = NULL;
1235                         qp->s_rdma_sge.num_sge = 0;
1236                         qp->s_rdma_sge.sge.mr = NULL;
1237                         qp->s_rdma_sge.sge.vaddr = NULL;
1238                         qp->s_rdma_sge.sge.length = 0;
1239                         qp->s_rdma_sge.sge.sge_length = 0;
1240                 }
1241                 qp->s_ack_state = opcode;
1242                 qp->s_ack_psn = psn;
1243                 spin_unlock_irq(&qp->s_lock);
1244                 tasklet_hi_schedule(&qp->s_task);
1245                 goto send_ack;
1246         }
1247
1248         /*
1249          * A pending RDMA read will ACK anything before it so
1250          * ignore earlier duplicate requests.
1251          */
1252         if (qp->s_ack_state != OP(ACKNOWLEDGE))
1253                 goto done;
1254
1255         /*
1256          * If an ACK is pending, don't replace the pending ACK
1257          * with an earlier one since the later one will ACK the earlier.
1258          * Also, if we already have a pending atomic, send it.
1259          */
1260         if (qp->r_ack_state != OP(ACKNOWLEDGE) &&
1261             (ipath_cmp24(psn, qp->r_ack_psn) <= 0 ||
1262              qp->r_ack_state >= OP(COMPARE_SWAP)))
1263                 goto send_ack;
1264         switch (opcode) {
1265         case OP(COMPARE_SWAP):
1266         case OP(FETCH_ADD):
1267                 /*
1268                  * Check for the PSN of the last atomic operation
1269                  * performed and resend the result if found.
1270                  */
1271                 if ((psn & IPATH_PSN_MASK) != qp->r_atomic_psn)
1272                         goto done;
1273                 break;
1274         }
1275         qp->r_ack_state = opcode;
1276         qp->r_nak_state = 0;
1277         qp->r_ack_psn = psn;
1278 send_ack:
1279         return 0;
1280
1281 done:
1282         return 1;
1283 }
1284
1285 /**
1286  * ipath_rc_rcv - process an incoming RC packet
1287  * @dev: the device this packet came in on
1288  * @hdr: the header of this packet
1289  * @has_grh: true if the header has a GRH
1290  * @data: the packet data
1291  * @tlen: the packet length
1292  * @qp: the QP for this packet
1293  *
1294  * This is called from ipath_qp_rcv() to process an incoming RC packet
1295  * for the given QP.
1296  * Called at interrupt level.
1297  */
1298 void ipath_rc_rcv(struct ipath_ibdev *dev, struct ipath_ib_header *hdr,
1299                   int has_grh, void *data, u32 tlen, struct ipath_qp *qp)
1300 {
1301         struct ipath_other_headers *ohdr;
1302         u32 opcode;
1303         u32 hdrsize;
1304         u32 psn;
1305         u32 pad;
1306         struct ib_wc wc;
1307         u32 pmtu = ib_mtu_enum_to_int(qp->path_mtu);
1308         int diff;
1309         struct ib_reth *reth;
1310         int header_in_data;
1311
1312         /* Check for GRH */
1313         if (!has_grh) {
1314                 ohdr = &hdr->u.oth;
1315                 hdrsize = 8 + 12;       /* LRH + BTH */
1316                 psn = be32_to_cpu(ohdr->bth[2]);
1317                 header_in_data = 0;
1318         } else {
1319                 ohdr = &hdr->u.l.oth;
1320                 hdrsize = 8 + 40 + 12;  /* LRH + GRH + BTH */
1321                 /*
1322                  * The header with GRH is 60 bytes and the core driver sets
1323                  * the eager header buffer size to 56 bytes so the last 4
1324                  * bytes of the BTH header (PSN) is in the data buffer.
1325                  */
1326                 header_in_data =
1327                         ipath_layer_get_rcvhdrentsize(dev->dd) == 16;
1328                 if (header_in_data) {
1329                         psn = be32_to_cpu(((__be32 *) data)[0]);
1330                         data += sizeof(__be32);
1331                 } else
1332                         psn = be32_to_cpu(ohdr->bth[2]);
1333         }
1334
1335         /*
1336          * Process responses (ACKs) before anything else.  Note that the
1337          * packet sequence number will be for something in the send work
1338          * queue rather than the expected receive packet sequence number.
1339          * In other words, this QP is the requester.
1340          */
1341         opcode = be32_to_cpu(ohdr->bth[0]) >> 24;
1342         if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
1343             opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
1344                 ipath_rc_rcv_resp(dev, ohdr, data, tlen, qp, opcode, psn,
1345                                   hdrsize, pmtu, header_in_data);
1346                 goto done;
1347         }
1348
1349         /* Compute 24 bits worth of difference. */
1350         diff = ipath_cmp24(psn, qp->r_psn);
1351         if (unlikely(diff)) {
1352                 if (ipath_rc_rcv_error(dev, ohdr, data, qp, opcode,
1353                                        psn, diff, header_in_data))
1354                         goto done;
1355                 goto send_ack;
1356         }
1357
1358         /* Check for opcode sequence errors. */
1359         switch (qp->r_state) {
1360         case OP(SEND_FIRST):
1361         case OP(SEND_MIDDLE):
1362                 if (opcode == OP(SEND_MIDDLE) ||
1363                     opcode == OP(SEND_LAST) ||
1364                     opcode == OP(SEND_LAST_WITH_IMMEDIATE))
1365                         break;
1366         nack_inv:
1367                 /*
1368                  * A NAK will ACK earlier sends and RDMA writes.
1369                  * Don't queue the NAK if a RDMA read, atomic, or NAK
1370                  * is pending though.
1371                  */
1372                 if (qp->r_ack_state >= OP(COMPARE_SWAP))
1373                         goto send_ack;
1374                 /* XXX Flush WQEs */
1375                 qp->state = IB_QPS_ERR;
1376                 qp->r_ack_state = OP(SEND_ONLY);
1377                 qp->r_nak_state = IB_NAK_INVALID_REQUEST;
1378                 qp->r_ack_psn = qp->r_psn;
1379                 goto send_ack;
1380
1381         case OP(RDMA_WRITE_FIRST):
1382         case OP(RDMA_WRITE_MIDDLE):
1383                 if (opcode == OP(RDMA_WRITE_MIDDLE) ||
1384                     opcode == OP(RDMA_WRITE_LAST) ||
1385                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1386                         break;
1387                 goto nack_inv;
1388
1389         default:
1390                 if (opcode == OP(SEND_MIDDLE) ||
1391                     opcode == OP(SEND_LAST) ||
1392                     opcode == OP(SEND_LAST_WITH_IMMEDIATE) ||
1393                     opcode == OP(RDMA_WRITE_MIDDLE) ||
1394                     opcode == OP(RDMA_WRITE_LAST) ||
1395                     opcode == OP(RDMA_WRITE_LAST_WITH_IMMEDIATE))
1396                         goto nack_inv;
1397                 /*
1398                  * Note that it is up to the requester to not send a new
1399                  * RDMA read or atomic operation before receiving an ACK
1400                  * for the previous operation.
1401                  */
1402                 break;
1403         }
1404
1405         wc.imm_data = 0;
1406         wc.wc_flags = 0;
1407
1408         /* OK, process the packet. */
1409         switch (opcode) {
1410         case OP(SEND_FIRST):
1411                 if (!ipath_get_rwqe(qp, 0)) {
1412                 rnr_nak:
1413                         /*
1414                          * A RNR NAK will ACK earlier sends and RDMA writes.
1415                          * Don't queue the NAK if a RDMA read or atomic
1416                          * is pending though.
1417                          */
1418                         if (qp->r_ack_state >= OP(COMPARE_SWAP))
1419                                 goto send_ack;
1420                         qp->r_ack_state = OP(SEND_ONLY);
1421                         qp->r_nak_state = IB_RNR_NAK | qp->r_min_rnr_timer;
1422                         qp->r_ack_psn = qp->r_psn;
1423                         goto send_ack;
1424                 }
1425                 qp->r_rcv_len = 0;
1426                 /* FALLTHROUGH */
1427         case OP(SEND_MIDDLE):
1428         case OP(RDMA_WRITE_MIDDLE):
1429         send_middle:
1430                 /* Check for invalid length PMTU or posted rwqe len. */
1431                 if (unlikely(tlen != (hdrsize + pmtu + 4)))
1432                         goto nack_inv;
1433                 qp->r_rcv_len += pmtu;
1434                 if (unlikely(qp->r_rcv_len > qp->r_len))
1435                         goto nack_inv;
1436                 ipath_copy_sge(&qp->r_sge, data, pmtu);
1437                 break;
1438
1439         case OP(RDMA_WRITE_LAST_WITH_IMMEDIATE):
1440                 /* consume RWQE */
1441                 if (!ipath_get_rwqe(qp, 1))
1442                         goto rnr_nak;
1443                 goto send_last_imm;
1444
1445         case OP(SEND_ONLY):
1446         case OP(SEND_ONLY_WITH_IMMEDIATE):
1447                 if (!ipath_get_rwqe(qp, 0))
1448                         goto rnr_nak;
1449                 qp->r_rcv_len = 0;
1450                 if (opcode == OP(SEND_ONLY))
1451                         goto send_last;
1452                 /* FALLTHROUGH */
1453         case OP(SEND_LAST_WITH_IMMEDIATE):
1454         send_last_imm:
1455                 if (header_in_data) {
1456                         wc.imm_data = *(__be32 *) data;
1457                         data += sizeof(__be32);
1458                 } else {
1459                         /* Immediate data comes after BTH */
1460                         wc.imm_data = ohdr->u.imm_data;
1461                 }
1462                 hdrsize += 4;
1463                 wc.wc_flags = IB_WC_WITH_IMM;
1464                 /* FALLTHROUGH */
1465         case OP(SEND_LAST):
1466         case OP(RDMA_WRITE_LAST):
1467         send_last:
1468                 /* Get the number of bytes the message was padded by. */
1469                 pad = (be32_to_cpu(ohdr->bth[0]) >> 20) & 3;
1470                 /* Check for invalid length. */
1471                 /* XXX LAST len should be >= 1 */
1472                 if (unlikely(tlen < (hdrsize + pad + 4)))
1473                         goto nack_inv;
1474                 /* Don't count the CRC. */
1475                 tlen -= (hdrsize + pad + 4);
1476                 wc.byte_len = tlen + qp->r_rcv_len;
1477                 if (unlikely(wc.byte_len > qp->r_len))
1478                         goto nack_inv;
1479                 ipath_copy_sge(&qp->r_sge, data, tlen);
1480                 qp->r_msn++;
1481                 if (opcode == OP(RDMA_WRITE_LAST) ||
1482                     opcode == OP(RDMA_WRITE_ONLY))
1483                         break;
1484                 wc.wr_id = qp->r_wr_id;
1485                 wc.status = IB_WC_SUCCESS;
1486                 wc.opcode = IB_WC_RECV;
1487                 wc.vendor_err = 0;
1488                 wc.qp_num = qp->ibqp.qp_num;
1489                 wc.src_qp = qp->remote_qpn;
1490                 wc.pkey_index = 0;
1491                 wc.slid = qp->remote_ah_attr.dlid;
1492                 wc.sl = qp->remote_ah_attr.sl;
1493                 wc.dlid_path_bits = 0;
1494                 wc.port_num = 0;
1495                 /* Signal completion event if the solicited bit is set. */
1496                 ipath_cq_enter(to_icq(qp->ibqp.recv_cq), &wc,
1497                                (ohdr->bth[0] &
1498                                 __constant_cpu_to_be32(1 << 23)) != 0);
1499                 break;
1500
1501         case OP(RDMA_WRITE_FIRST):
1502         case OP(RDMA_WRITE_ONLY):
1503         case OP(RDMA_WRITE_ONLY_WITH_IMMEDIATE):
1504                 /* consume RWQE */
1505                 /* RETH comes after BTH */
1506                 if (!header_in_data)
1507                         reth = &ohdr->u.rc.reth;
1508                 else {
1509                         reth = (struct ib_reth *)data;
1510                         data += sizeof(*reth);
1511                 }
1512                 hdrsize += sizeof(*reth);
1513                 qp->r_len = be32_to_cpu(reth->length);
1514                 qp->r_rcv_len = 0;
1515                 if (qp->r_len != 0) {
1516                         u32 rkey = be32_to_cpu(reth->rkey);
1517                         u64 vaddr = be64_to_cpu(reth->vaddr);
1518                         int ok;
1519
1520                         /* Check rkey & NAK */
1521                         ok = ipath_rkey_ok(dev, &qp->r_sge,
1522                                            qp->r_len, vaddr, rkey,
1523                                            IB_ACCESS_REMOTE_WRITE);
1524                         if (unlikely(!ok))
1525                                 goto nack_acc;
1526                 } else {
1527                         qp->r_sge.sg_list = NULL;
1528                         qp->r_sge.sge.mr = NULL;
1529                         qp->r_sge.sge.vaddr = NULL;
1530                         qp->r_sge.sge.length = 0;
1531                         qp->r_sge.sge.sge_length = 0;
1532                 }
1533                 if (unlikely(!(qp->qp_access_flags &
1534                                IB_ACCESS_REMOTE_WRITE)))
1535                         goto nack_acc;
1536                 if (opcode == OP(RDMA_WRITE_FIRST))
1537                         goto send_middle;
1538                 else if (opcode == OP(RDMA_WRITE_ONLY))
1539                         goto send_last;
1540                 if (!ipath_get_rwqe(qp, 1))
1541                         goto rnr_nak;
1542                 goto send_last_imm;
1543
1544         case OP(RDMA_READ_REQUEST):
1545                 /* RETH comes after BTH */
1546                 if (!header_in_data)
1547                         reth = &ohdr->u.rc.reth;
1548                 else {
1549                         reth = (struct ib_reth *)data;
1550                         data += sizeof(*reth);
1551                 }
1552                 if (unlikely(!(qp->qp_access_flags &
1553                                IB_ACCESS_REMOTE_READ)))
1554                         goto nack_acc;
1555                 spin_lock_irq(&qp->s_lock);
1556                 qp->s_rdma_len = be32_to_cpu(reth->length);
1557                 if (qp->s_rdma_len != 0) {
1558                         u32 rkey = be32_to_cpu(reth->rkey);
1559                         u64 vaddr = be64_to_cpu(reth->vaddr);
1560                         int ok;
1561
1562                         /* Check rkey & NAK */
1563                         ok = ipath_rkey_ok(dev, &qp->s_rdma_sge,
1564                                            qp->s_rdma_len, vaddr, rkey,
1565                                            IB_ACCESS_REMOTE_READ);
1566                         if (unlikely(!ok)) {
1567                                 spin_unlock_irq(&qp->s_lock);
1568                                 goto nack_acc;
1569                         }
1570                         /*
1571                          * Update the next expected PSN.  We add 1 later
1572                          * below, so only add the remainder here.
1573                          */
1574                         if (qp->s_rdma_len > pmtu)
1575                                 qp->r_psn += (qp->s_rdma_len - 1) / pmtu;
1576                 } else {
1577                         qp->s_rdma_sge.sg_list = NULL;
1578                         qp->s_rdma_sge.num_sge = 0;
1579                         qp->s_rdma_sge.sge.mr = NULL;
1580                         qp->s_rdma_sge.sge.vaddr = NULL;
1581                         qp->s_rdma_sge.sge.length = 0;
1582                         qp->s_rdma_sge.sge.sge_length = 0;
1583                 }
1584                 /*
1585                  * We need to increment the MSN here instead of when we
1586                  * finish sending the result since a duplicate request would
1587                  * increment it more than once.
1588                  */
1589                 qp->r_msn++;
1590
1591                 qp->s_ack_state = opcode;
1592                 qp->s_ack_psn = psn;
1593                 spin_unlock_irq(&qp->s_lock);
1594
1595                 qp->r_psn++;
1596                 qp->r_state = opcode;
1597                 qp->r_nak_state = 0;
1598
1599                 /* Call ipath_do_rc_send() in another thread. */
1600                 tasklet_hi_schedule(&qp->s_task);
1601
1602                 goto done;
1603
1604         case OP(COMPARE_SWAP):
1605         case OP(FETCH_ADD): {
1606                 struct ib_atomic_eth *ateth;
1607                 u64 vaddr;
1608                 u64 sdata;
1609                 u32 rkey;
1610
1611                 if (!header_in_data)
1612                         ateth = &ohdr->u.atomic_eth;
1613                 else {
1614                         ateth = (struct ib_atomic_eth *)data;
1615                         data += sizeof(*ateth);
1616                 }
1617                 vaddr = be64_to_cpu(ateth->vaddr);
1618                 if (unlikely(vaddr & (sizeof(u64) - 1)))
1619                         goto nack_inv;
1620                 rkey = be32_to_cpu(ateth->rkey);
1621                 /* Check rkey & NAK */
1622                 if (unlikely(!ipath_rkey_ok(dev, &qp->r_sge,
1623                                             sizeof(u64), vaddr, rkey,
1624                                             IB_ACCESS_REMOTE_ATOMIC)))
1625                         goto nack_acc;
1626                 if (unlikely(!(qp->qp_access_flags &
1627                                IB_ACCESS_REMOTE_ATOMIC)))
1628                         goto nack_acc;
1629                 /* Perform atomic OP and save result. */
1630                 sdata = be64_to_cpu(ateth->swap_data);
1631                 spin_lock_irq(&dev->pending_lock);
1632                 qp->r_atomic_data = *(u64 *) qp->r_sge.sge.vaddr;
1633                 if (opcode == OP(FETCH_ADD))
1634                         *(u64 *) qp->r_sge.sge.vaddr =
1635                                 qp->r_atomic_data + sdata;
1636                 else if (qp->r_atomic_data ==
1637                          be64_to_cpu(ateth->compare_data))
1638                         *(u64 *) qp->r_sge.sge.vaddr = sdata;
1639                 spin_unlock_irq(&dev->pending_lock);
1640                 qp->r_msn++;
1641                 qp->r_atomic_psn = psn & IPATH_PSN_MASK;
1642                 psn |= 1 << 31;
1643                 break;
1644         }
1645
1646         default:
1647                 /* Drop packet for unknown opcodes. */
1648                 goto done;
1649         }
1650         qp->r_psn++;
1651         qp->r_state = opcode;
1652         qp->r_nak_state = 0;
1653         /* Send an ACK if requested or required. */
1654         if (psn & (1 << 31)) {
1655                 /*
1656                  * Coalesce ACKs unless there is a RDMA READ or
1657                  * ATOMIC pending.
1658                  */
1659                 if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1660                         qp->r_ack_state = opcode;
1661                         qp->r_ack_psn = psn;
1662                 }
1663                 goto send_ack;
1664         }
1665         goto done;
1666
1667 nack_acc:
1668         /*
1669          * A NAK will ACK earlier sends and RDMA writes.
1670          * Don't queue the NAK if a RDMA read, atomic, or NAK
1671          * is pending though.
1672          */
1673         if (qp->r_ack_state < OP(COMPARE_SWAP)) {
1674                 /* XXX Flush WQEs */
1675                 qp->state = IB_QPS_ERR;
1676                 qp->r_ack_state = OP(RDMA_WRITE_ONLY);
1677                 qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR;
1678                 qp->r_ack_psn = qp->r_psn;
1679         }
1680 send_ack:
1681         /* Send ACK right away unless the send tasklet has a pending ACK. */
1682         if (qp->s_ack_state == OP(ACKNOWLEDGE))
1683                 send_rc_ack(qp);
1684
1685 done:
1686         return;
1687 }