linux/drivers/infiniband/hw/cxgb3/iwch_qp.c
<<
>>
Prefs
   1/*
   2 * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
   3 *
   4 * This software is available to you under a choice of one of two
   5 * licenses.  You may choose to be licensed under the terms of the GNU
   6 * General Public License (GPL) Version 2, available from the file
   7 * COPYING in the main directory of this source tree, or the
   8 * OpenIB.org BSD license below:
   9 *
  10 *     Redistribution and use in source and binary forms, with or
  11 *     without modification, are permitted provided that the following
  12 *     conditions are met:
  13 *
  14 *      - Redistributions of source code must retain the above
  15 *        copyright notice, this list of conditions and the following
  16 *        disclaimer.
  17 *
  18 *      - Redistributions in binary form must reproduce the above
  19 *        copyright notice, this list of conditions and the following
  20 *        disclaimer in the documentation and/or other materials
  21 *        provided with the distribution.
  22 *
  23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30 * SOFTWARE.
  31 */
  32#include <linux/sched.h>
  33#include <linux/gfp.h>
  34#include "iwch_provider.h"
  35#include "iwch.h"
  36#include "iwch_cm.h"
  37#include "cxio_hal.h"
  38#include "cxio_resource.h"
  39
  40#define NO_SUPPORT -1
  41
  42static int build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
  43                                u8 * flit_cnt)
  44{
  45        int i;
  46        u32 plen;
  47
  48        switch (wr->opcode) {
  49        case IB_WR_SEND:
  50                if (wr->send_flags & IB_SEND_SOLICITED)
  51                        wqe->send.rdmaop = T3_SEND_WITH_SE;
  52                else
  53                        wqe->send.rdmaop = T3_SEND;
  54                wqe->send.rem_stag = 0;
  55                break;
  56        case IB_WR_SEND_WITH_INV:
  57                if (wr->send_flags & IB_SEND_SOLICITED)
  58                        wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
  59                else
  60                        wqe->send.rdmaop = T3_SEND_WITH_INV;
  61                wqe->send.rem_stag = cpu_to_be32(wr->ex.invalidate_rkey);
  62                break;
  63        default:
  64                return -EINVAL;
  65        }
  66        if (wr->num_sge > T3_MAX_SGE)
  67                return -EINVAL;
  68        wqe->send.reserved[0] = 0;
  69        wqe->send.reserved[1] = 0;
  70        wqe->send.reserved[2] = 0;
  71        plen = 0;
  72        for (i = 0; i < wr->num_sge; i++) {
  73                if ((plen + wr->sg_list[i].length) < plen)
  74                        return -EMSGSIZE;
  75
  76                plen += wr->sg_list[i].length;
  77                wqe->send.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
  78                wqe->send.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
  79                wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
  80        }
  81        wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
  82        *flit_cnt = 4 + ((wr->num_sge) << 1);
  83        wqe->send.plen = cpu_to_be32(plen);
  84        return 0;
  85}
  86
  87static int build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
  88                                 u8 *flit_cnt)
  89{
  90        int i;
  91        u32 plen;
  92        if (wr->num_sge > T3_MAX_SGE)
  93                return -EINVAL;
  94        wqe->write.rdmaop = T3_RDMA_WRITE;
  95        wqe->write.reserved[0] = 0;
  96        wqe->write.reserved[1] = 0;
  97        wqe->write.reserved[2] = 0;
  98        wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
  99        wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
 100
 101        if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
 102                plen = 4;
 103                wqe->write.sgl[0].stag = wr->ex.imm_data;
 104                wqe->write.sgl[0].len = cpu_to_be32(0);
 105                wqe->write.num_sgle = cpu_to_be32(0);
 106                *flit_cnt = 6;
 107        } else {
 108                plen = 0;
 109                for (i = 0; i < wr->num_sge; i++) {
 110                        if ((plen + wr->sg_list[i].length) < plen) {
 111                                return -EMSGSIZE;
 112                        }
 113                        plen += wr->sg_list[i].length;
 114                        wqe->write.sgl[i].stag =
 115                            cpu_to_be32(wr->sg_list[i].lkey);
 116                        wqe->write.sgl[i].len =
 117                            cpu_to_be32(wr->sg_list[i].length);
 118                        wqe->write.sgl[i].to =
 119                            cpu_to_be64(wr->sg_list[i].addr);
 120                }
 121                wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
 122                *flit_cnt = 5 + ((wr->num_sge) << 1);
 123        }
 124        wqe->write.plen = cpu_to_be32(plen);
 125        return 0;
 126}
 127
 128static int build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
 129                                u8 *flit_cnt)
 130{
 131        if (wr->num_sge > 1)
 132                return -EINVAL;
 133        wqe->read.rdmaop = T3_READ_REQ;
 134        if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
 135                wqe->read.local_inv = 1;
 136        else
 137                wqe->read.local_inv = 0;
 138        wqe->read.reserved[0] = 0;
 139        wqe->read.reserved[1] = 0;
 140        wqe->read.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
 141        wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.remote_addr);
 142        wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
 143        wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
 144        wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
 145        *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
 146        return 0;
 147}
 148
 149static int build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
 150                                u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
 151{
 152        int i;
 153        __be64 *p;
 154
 155        if (wr->wr.fast_reg.page_list_len > T3_MAX_FASTREG_DEPTH)
 156                return -EINVAL;
 157        *wr_cnt = 1;
 158        wqe->fastreg.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
 159        wqe->fastreg.len = cpu_to_be32(wr->wr.fast_reg.length);
 160        wqe->fastreg.va_base_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
 161        wqe->fastreg.va_base_lo_fbo =
 162                                cpu_to_be32(wr->wr.fast_reg.iova_start & 0xffffffff);
 163        wqe->fastreg.page_type_perms = cpu_to_be32(
 164                V_FR_PAGE_COUNT(wr->wr.fast_reg.page_list_len) |
 165                V_FR_PAGE_SIZE(wr->wr.fast_reg.page_shift-12) |
 166                V_FR_TYPE(TPT_VATO) |
 167                V_FR_PERMS(iwch_ib_to_tpt_access(wr->wr.fast_reg.access_flags)));
 168        p = &wqe->fastreg.pbl_addrs[0];
 169        for (i = 0; i < wr->wr.fast_reg.page_list_len; i++, p++) {
 170
 171                /* If we need a 2nd WR, then set it up */
 172                if (i == T3_MAX_FASTREG_FRAG) {
 173                        *wr_cnt = 2;
 174                        wqe = (union t3_wr *)(wq->queue +
 175                                Q_PTR2IDX((wq->wptr+1), wq->size_log2));
 176                        build_fw_riwrh((void *)wqe, T3_WR_FASTREG, 0,
 177                               Q_GENBIT(wq->wptr + 1, wq->size_log2),
 178                               0, 1 + wr->wr.fast_reg.page_list_len - T3_MAX_FASTREG_FRAG,
 179                               T3_EOP);
 180
 181                        p = &wqe->pbl_frag.pbl_addrs[0];
 182                }
 183                *p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
 184        }
 185        *flit_cnt = 5 + wr->wr.fast_reg.page_list_len;
 186        if (*flit_cnt > 15)
 187                *flit_cnt = 15;
 188        return 0;
 189}
 190
 191static int build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
 192                                u8 *flit_cnt)
 193{
 194        wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
 195        wqe->local_inv.reserved = 0;
 196        *flit_cnt = sizeof(struct t3_local_inv_wr) >> 3;
 197        return 0;
 198}
 199
 200static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
 201                            u32 num_sgle, u32 * pbl_addr, u8 * page_size)
 202{
 203        int i;
 204        struct iwch_mr *mhp;
 205        u64 offset;
 206        for (i = 0; i < num_sgle; i++) {
 207
 208                mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
 209                if (!mhp) {
 210                        PDBG("%s %d\n", __func__, __LINE__);
 211                        return -EIO;
 212                }
 213                if (!mhp->attr.state) {
 214                        PDBG("%s %d\n", __func__, __LINE__);
 215                        return -EIO;
 216                }
 217                if (mhp->attr.zbva) {
 218                        PDBG("%s %d\n", __func__, __LINE__);
 219                        return -EIO;
 220                }
 221
 222                if (sg_list[i].addr < mhp->attr.va_fbo) {
 223                        PDBG("%s %d\n", __func__, __LINE__);
 224                        return -EINVAL;
 225                }
 226                if (sg_list[i].addr + ((u64) sg_list[i].length) <
 227                    sg_list[i].addr) {
 228                        PDBG("%s %d\n", __func__, __LINE__);
 229                        return -EINVAL;
 230                }
 231                if (sg_list[i].addr + ((u64) sg_list[i].length) >
 232                    mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
 233                        PDBG("%s %d\n", __func__, __LINE__);
 234                        return -EINVAL;
 235                }
 236                offset = sg_list[i].addr - mhp->attr.va_fbo;
 237                offset += mhp->attr.va_fbo &
 238                          ((1UL << (12 + mhp->attr.page_size)) - 1);
 239                pbl_addr[i] = ((mhp->attr.pbl_addr -
 240                                rhp->rdev.rnic_info.pbl_base) >> 3) +
 241                              (offset >> (12 + mhp->attr.page_size));
 242                page_size[i] = mhp->attr.page_size;
 243        }
 244        return 0;
 245}
 246
 247static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
 248                                struct ib_recv_wr *wr)
 249{
 250        int i, err = 0;
 251        u32 pbl_addr[T3_MAX_SGE];
 252        u8 page_size[T3_MAX_SGE];
 253
 254        err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
 255                               page_size);
 256        if (err)
 257                return err;
 258        wqe->recv.pagesz[0] = page_size[0];
 259        wqe->recv.pagesz[1] = page_size[1];
 260        wqe->recv.pagesz[2] = page_size[2];
 261        wqe->recv.pagesz[3] = page_size[3];
 262        wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 263        for (i = 0; i < wr->num_sge; i++) {
 264                wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
 265                wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 266
 267                /* to in the WQE == the offset into the page */
 268                wqe->recv.sgl[i].to = cpu_to_be64(((u32)wr->sg_list[i].addr) &
 269                                ((1UL << (12 + page_size[i])) - 1));
 270
 271                /* pbl_addr is the adapters address in the PBL */
 272                wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
 273        }
 274        for (; i < T3_MAX_SGE; i++) {
 275                wqe->recv.sgl[i].stag = 0;
 276                wqe->recv.sgl[i].len = 0;
 277                wqe->recv.sgl[i].to = 0;
 278                wqe->recv.pbl_addr[i] = 0;
 279        }
 280        qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 281                             qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
 282        qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 283                             qhp->wq.rq_size_log2)].pbl_addr = 0;
 284        return 0;
 285}
 286
 287static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
 288                                struct ib_recv_wr *wr)
 289{
 290        int i;
 291        u32 pbl_addr;
 292        u32 pbl_offset;
 293
 294
 295        /*
 296         * The T3 HW requires the PBL in the HW recv descriptor to reference
 297         * a PBL entry.  So we allocate the max needed PBL memory here and pass
 298         * it to the uP in the recv WR.  The uP will build the PBL and setup
 299         * the HW recv descriptor.
 300         */
 301        pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
 302        if (!pbl_addr)
 303                return -ENOMEM;
 304
 305        /*
 306         * Compute the 8B aligned offset.
 307         */
 308        pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
 309
 310        wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 311
 312        for (i = 0; i < wr->num_sge; i++) {
 313
 314                /*
 315                 * Use a 128MB page size. This and an imposed 128MB
 316                 * sge length limit allows us to require only a 2-entry HW
 317                 * PBL for each SGE.  This restriction is acceptable since
 318                 * since it is not possible to allocate 128MB of contiguous
 319                 * DMA coherent memory!
 320                 */
 321                if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
 322                        return -EINVAL;
 323                wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
 324
 325                /*
 326                 * T3 restricts a recv to all zero-stag or all non-zero-stag.
 327                 */
 328                if (wr->sg_list[i].lkey != 0)
 329                        return -EINVAL;
 330                wqe->recv.sgl[i].stag = 0;
 331                wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 332                wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
 333                wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
 334                pbl_offset += 2;
 335        }
 336        for (; i < T3_MAX_SGE; i++) {
 337                wqe->recv.pagesz[i] = 0;
 338                wqe->recv.sgl[i].stag = 0;
 339                wqe->recv.sgl[i].len = 0;
 340                wqe->recv.sgl[i].to = 0;
 341                wqe->recv.pbl_addr[i] = 0;
 342        }
 343        qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 344                             qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
 345        qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 346                             qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
 347        return 0;
 348}
 349
 350int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 351                      struct ib_send_wr **bad_wr)
 352{
 353        int err = 0;
 354        u8 uninitialized_var(t3_wr_flit_cnt);
 355        enum t3_wr_opcode t3_wr_opcode = 0;
 356        enum t3_wr_flags t3_wr_flags;
 357        struct iwch_qp *qhp;
 358        u32 idx;
 359        union t3_wr *wqe;
 360        u32 num_wrs;
 361        unsigned long flag;
 362        struct t3_swsq *sqp;
 363        int wr_cnt = 1;
 364
 365        qhp = to_iwch_qp(ibqp);
 366        spin_lock_irqsave(&qhp->lock, flag);
 367        if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 368                spin_unlock_irqrestore(&qhp->lock, flag);
 369                err = -EINVAL;
 370                goto out;
 371        }
 372        num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
 373                  qhp->wq.sq_size_log2);
 374        if (num_wrs == 0) {
 375                spin_unlock_irqrestore(&qhp->lock, flag);
 376                err = -ENOMEM;
 377                goto out;
 378        }
 379        while (wr) {
 380                if (num_wrs == 0) {
 381                        err = -ENOMEM;
 382                        break;
 383                }
 384                idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 385                wqe = (union t3_wr *) (qhp->wq.queue + idx);
 386                t3_wr_flags = 0;
 387                if (wr->send_flags & IB_SEND_SOLICITED)
 388                        t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
 389                if (wr->send_flags & IB_SEND_SIGNALED)
 390                        t3_wr_flags |= T3_COMPLETION_FLAG;
 391                sqp = qhp->wq.sq +
 392                      Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
 393                switch (wr->opcode) {
 394                case IB_WR_SEND:
 395                case IB_WR_SEND_WITH_INV:
 396                        if (wr->send_flags & IB_SEND_FENCE)
 397                                t3_wr_flags |= T3_READ_FENCE_FLAG;
 398                        t3_wr_opcode = T3_WR_SEND;
 399                        err = build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
 400                        break;
 401                case IB_WR_RDMA_WRITE:
 402                case IB_WR_RDMA_WRITE_WITH_IMM:
 403                        t3_wr_opcode = T3_WR_WRITE;
 404                        err = build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
 405                        break;
 406                case IB_WR_RDMA_READ:
 407                case IB_WR_RDMA_READ_WITH_INV:
 408                        t3_wr_opcode = T3_WR_READ;
 409                        t3_wr_flags = 0; /* T3 reads are always signaled */
 410                        err = build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
 411                        if (err)
 412                                break;
 413                        sqp->read_len = wqe->read.local_len;
 414                        if (!qhp->wq.oldest_read)
 415                                qhp->wq.oldest_read = sqp;
 416                        break;
 417                case IB_WR_FAST_REG_MR:
 418                        t3_wr_opcode = T3_WR_FASTREG;
 419                        err = build_fastreg(wqe, wr, &t3_wr_flit_cnt,
 420                                                 &wr_cnt, &qhp->wq);
 421                        break;
 422                case IB_WR_LOCAL_INV:
 423                        if (wr->send_flags & IB_SEND_FENCE)
 424                                t3_wr_flags |= T3_LOCAL_FENCE_FLAG;
 425                        t3_wr_opcode = T3_WR_INV_STAG;
 426                        err = build_inv_stag(wqe, wr, &t3_wr_flit_cnt);
 427                        break;
 428                default:
 429                        PDBG("%s post of type=%d TBD!\n", __func__,
 430                             wr->opcode);
 431                        err = -EINVAL;
 432                }
 433                if (err)
 434                        break;
 435                wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 436                sqp->wr_id = wr->wr_id;
 437                sqp->opcode = wr2opcode(t3_wr_opcode);
 438                sqp->sq_wptr = qhp->wq.sq_wptr;
 439                sqp->complete = 0;
 440                sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
 441
 442                build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
 443                               Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 444                               0, t3_wr_flit_cnt,
 445                               (wr_cnt == 1) ? T3_SOPEOP : T3_SOP);
 446                PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
 447                     __func__, (unsigned long long) wr->wr_id, idx,
 448                     Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
 449                     sqp->opcode);
 450                wr = wr->next;
 451                num_wrs--;
 452                qhp->wq.wptr += wr_cnt;
 453                ++(qhp->wq.sq_wptr);
 454        }
 455        spin_unlock_irqrestore(&qhp->lock, flag);
 456        if (cxio_wq_db_enabled(&qhp->wq))
 457                ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 458
 459out:
 460        if (err)
 461                *bad_wr = wr;
 462        return err;
 463}
 464
 465int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 466                      struct ib_recv_wr **bad_wr)
 467{
 468        int err = 0;
 469        struct iwch_qp *qhp;
 470        u32 idx;
 471        union t3_wr *wqe;
 472        u32 num_wrs;
 473        unsigned long flag;
 474
 475        qhp = to_iwch_qp(ibqp);
 476        spin_lock_irqsave(&qhp->lock, flag);
 477        if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 478                spin_unlock_irqrestore(&qhp->lock, flag);
 479                err = -EINVAL;
 480                goto out;
 481        }
 482        num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
 483                            qhp->wq.rq_size_log2) - 1;
 484        if (!wr) {
 485                spin_unlock_irqrestore(&qhp->lock, flag);
 486                err = -ENOMEM;
 487                goto out;
 488        }
 489        while (wr) {
 490                if (wr->num_sge > T3_MAX_SGE) {
 491                        err = -EINVAL;
 492                        break;
 493                }
 494                idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 495                wqe = (union t3_wr *) (qhp->wq.queue + idx);
 496                if (num_wrs)
 497                        if (wr->sg_list[0].lkey)
 498                                err = build_rdma_recv(qhp, wqe, wr);
 499                        else
 500                                err = build_zero_stag_recv(qhp, wqe, wr);
 501                else
 502                        err = -ENOMEM;
 503
 504                if (err)
 505                        break;
 506
 507                build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
 508                               Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 509                               0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
 510                PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
 511                     "wqe %p \n", __func__, (unsigned long long) wr->wr_id,
 512                     idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
 513                ++(qhp->wq.rq_wptr);
 514                ++(qhp->wq.wptr);
 515                wr = wr->next;
 516                num_wrs--;
 517        }
 518        spin_unlock_irqrestore(&qhp->lock, flag);
 519        if (cxio_wq_db_enabled(&qhp->wq))
 520                ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 521
 522out:
 523        if (err)
 524                *bad_wr = wr;
 525        return err;
 526}
 527
 528int iwch_bind_mw(struct ib_qp *qp,
 529                             struct ib_mw *mw,
 530                             struct ib_mw_bind *mw_bind)
 531{
 532        struct iwch_dev *rhp;
 533        struct iwch_mw *mhp;
 534        struct iwch_qp *qhp;
 535        union t3_wr *wqe;
 536        u32 pbl_addr;
 537        u8 page_size;
 538        u32 num_wrs;
 539        unsigned long flag;
 540        struct ib_sge sgl;
 541        int err=0;
 542        enum t3_wr_flags t3_wr_flags;
 543        u32 idx;
 544        struct t3_swsq *sqp;
 545
 546        qhp = to_iwch_qp(qp);
 547        mhp = to_iwch_mw(mw);
 548        rhp = qhp->rhp;
 549
 550        spin_lock_irqsave(&qhp->lock, flag);
 551        if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 552                spin_unlock_irqrestore(&qhp->lock, flag);
 553                return -EINVAL;
 554        }
 555        num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
 556                            qhp->wq.sq_size_log2);
 557        if (num_wrs == 0) {
 558                spin_unlock_irqrestore(&qhp->lock, flag);
 559                return -ENOMEM;
 560        }
 561        idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 562        PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__, idx,
 563             mw, mw_bind);
 564        wqe = (union t3_wr *) (qhp->wq.queue + idx);
 565
 566        t3_wr_flags = 0;
 567        if (mw_bind->send_flags & IB_SEND_SIGNALED)
 568                t3_wr_flags = T3_COMPLETION_FLAG;
 569
 570        sgl.addr = mw_bind->addr;
 571        sgl.lkey = mw_bind->mr->lkey;
 572        sgl.length = mw_bind->length;
 573        wqe->bind.reserved = 0;
 574        wqe->bind.type = TPT_VATO;
 575
 576        /* TBD: check perms */
 577        wqe->bind.perms = iwch_ib_to_tpt_bind_access(mw_bind->mw_access_flags);
 578        wqe->bind.mr_stag = cpu_to_be32(mw_bind->mr->lkey);
 579        wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
 580        wqe->bind.mw_len = cpu_to_be32(mw_bind->length);
 581        wqe->bind.mw_va = cpu_to_be64(mw_bind->addr);
 582        err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
 583        if (err) {
 584                spin_unlock_irqrestore(&qhp->lock, flag);
 585                return err;
 586        }
 587        wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 588        sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
 589        sqp->wr_id = mw_bind->wr_id;
 590        sqp->opcode = T3_BIND_MW;
 591        sqp->sq_wptr = qhp->wq.sq_wptr;
 592        sqp->complete = 0;
 593        sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
 594        wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
 595        wqe->bind.mr_pagesz = page_size;
 596        build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
 597                       Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
 598                       sizeof(struct t3_bind_mw_wr) >> 3, T3_SOPEOP);
 599        ++(qhp->wq.wptr);
 600        ++(qhp->wq.sq_wptr);
 601        spin_unlock_irqrestore(&qhp->lock, flag);
 602
 603        if (cxio_wq_db_enabled(&qhp->wq))
 604                ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 605
 606        return err;
 607}
 608
 609static inline void build_term_codes(struct respQ_msg_t *rsp_msg,
 610                                    u8 *layer_type, u8 *ecode)
 611{
 612        int status = TPT_ERR_INTERNAL_ERR;
 613        int tagged = 0;
 614        int opcode = -1;
 615        int rqtype = 0;
 616        int send_inv = 0;
 617
 618        if (rsp_msg) {
 619                status = CQE_STATUS(rsp_msg->cqe);
 620                opcode = CQE_OPCODE(rsp_msg->cqe);
 621                rqtype = RQ_TYPE(rsp_msg->cqe);
 622                send_inv = (opcode == T3_SEND_WITH_INV) ||
 623                           (opcode == T3_SEND_WITH_SE_INV);
 624                tagged = (opcode == T3_RDMA_WRITE) ||
 625                         (rqtype && (opcode == T3_READ_RESP));
 626        }
 627
 628        switch (status) {
 629        case TPT_ERR_STAG:
 630                if (send_inv) {
 631                        *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 632                        *ecode = RDMAP_CANT_INV_STAG;
 633                } else {
 634                        *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 635                        *ecode = RDMAP_INV_STAG;
 636                }
 637                break;
 638        case TPT_ERR_PDID:
 639                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 640                if ((opcode == T3_SEND_WITH_INV) ||
 641                    (opcode == T3_SEND_WITH_SE_INV))
 642                        *ecode = RDMAP_CANT_INV_STAG;
 643                else
 644                        *ecode = RDMAP_STAG_NOT_ASSOC;
 645                break;
 646        case TPT_ERR_QPID:
 647                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 648                *ecode = RDMAP_STAG_NOT_ASSOC;
 649                break;
 650        case TPT_ERR_ACCESS:
 651                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 652                *ecode = RDMAP_ACC_VIOL;
 653                break;
 654        case TPT_ERR_WRAP:
 655                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 656                *ecode = RDMAP_TO_WRAP;
 657                break;
 658        case TPT_ERR_BOUND:
 659                if (tagged) {
 660                        *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 661                        *ecode = DDPT_BASE_BOUNDS;
 662                } else {
 663                        *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 664                        *ecode = RDMAP_BASE_BOUNDS;
 665                }
 666                break;
 667        case TPT_ERR_INVALIDATE_SHARED_MR:
 668        case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
 669                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 670                *ecode = RDMAP_CANT_INV_STAG;
 671                break;
 672        case TPT_ERR_ECC:
 673        case TPT_ERR_ECC_PSTAG:
 674        case TPT_ERR_INTERNAL_ERR:
 675                *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
 676                *ecode = 0;
 677                break;
 678        case TPT_ERR_OUT_OF_RQE:
 679                *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 680                *ecode = DDPU_INV_MSN_NOBUF;
 681                break;
 682        case TPT_ERR_PBL_ADDR_BOUND:
 683                *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 684                *ecode = DDPT_BASE_BOUNDS;
 685                break;
 686        case TPT_ERR_CRC:
 687                *layer_type = LAYER_MPA|DDP_LLP;
 688                *ecode = MPA_CRC_ERR;
 689                break;
 690        case TPT_ERR_MARKER:
 691                *layer_type = LAYER_MPA|DDP_LLP;
 692                *ecode = MPA_MARKER_ERR;
 693                break;
 694        case TPT_ERR_PDU_LEN_ERR:
 695                *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 696                *ecode = DDPU_MSG_TOOBIG;
 697                break;
 698        case TPT_ERR_DDP_VERSION:
 699                if (tagged) {
 700                        *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 701                        *ecode = DDPT_INV_VERS;
 702                } else {
 703                        *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 704                        *ecode = DDPU_INV_VERS;
 705                }
 706                break;
 707        case TPT_ERR_RDMA_VERSION:
 708                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 709                *ecode = RDMAP_INV_VERS;
 710                break;
 711        case TPT_ERR_OPCODE:
 712                *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 713                *ecode = RDMAP_INV_OPCODE;
 714                break;
 715        case TPT_ERR_DDP_QUEUE_NUM:
 716                *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 717                *ecode = DDPU_INV_QN;
 718                break;
 719        case TPT_ERR_MSN:
 720        case TPT_ERR_MSN_GAP:
 721        case TPT_ERR_MSN_RANGE:
 722        case TPT_ERR_IRD_OVERFLOW:
 723                *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 724                *ecode = DDPU_INV_MSN_RANGE;
 725                break;
 726        case TPT_ERR_TBIT:
 727                *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
 728                *ecode = 0;
 729                break;
 730        case TPT_ERR_MO:
 731                *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 732                *ecode = DDPU_INV_MO;
 733                break;
 734        default:
 735                *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 736                *ecode = 0;
 737                break;
 738        }
 739}
 740
 741int iwch_post_zb_read(struct iwch_ep *ep)
 742{
 743        union t3_wr *wqe;
 744        struct sk_buff *skb;
 745        u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
 746
 747        PDBG("%s enter\n", __func__);
 748        skb = alloc_skb(40, GFP_KERNEL);
 749        if (!skb) {
 750                printk(KERN_ERR "%s cannot send zb_read!!\n", __func__);
 751                return -ENOMEM;
 752        }
 753        wqe = (union t3_wr *)skb_put(skb, sizeof(struct t3_rdma_read_wr));
 754        memset(wqe, 0, sizeof(struct t3_rdma_read_wr));
 755        wqe->read.rdmaop = T3_READ_REQ;
 756        wqe->read.reserved[0] = 0;
 757        wqe->read.reserved[1] = 0;
 758        wqe->read.rem_stag = cpu_to_be32(1);
 759        wqe->read.rem_to = cpu_to_be64(1);
 760        wqe->read.local_stag = cpu_to_be32(1);
 761        wqe->read.local_len = cpu_to_be32(0);
 762        wqe->read.local_to = cpu_to_be64(1);
 763        wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
 764        wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(ep->hwtid)|
 765                                                V_FW_RIWR_LEN(flit_cnt));
 766        skb->priority = CPL_PRIORITY_DATA;
 767        return iwch_cxgb3_ofld_send(ep->com.qp->rhp->rdev.t3cdev_p, skb);
 768}
 769
 770/*
 771 * This posts a TERMINATE with layer=RDMA, type=catastrophic.
 772 */
 773int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
 774{
 775        union t3_wr *wqe;
 776        struct terminate_message *term;
 777        struct sk_buff *skb;
 778
 779        PDBG("%s %d\n", __func__, __LINE__);
 780        skb = alloc_skb(40, GFP_ATOMIC);
 781        if (!skb) {
 782                printk(KERN_ERR "%s cannot send TERMINATE!\n", __func__);
 783                return -ENOMEM;
 784        }
 785        wqe = (union t3_wr *)skb_put(skb, 40);
 786        memset(wqe, 0, 40);
 787        wqe->send.rdmaop = T3_TERMINATE;
 788
 789        /* immediate data length */
 790        wqe->send.plen = htonl(4);
 791
 792        /* immediate data starts here. */
 793        term = (struct terminate_message *)wqe->send.sgl;
 794        build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
 795        wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
 796                         V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
 797        wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
 798        skb->priority = CPL_PRIORITY_DATA;
 799        return iwch_cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
 800}
 801
 802/*
 803 * Assumes qhp lock is held.
 804 */
 805static void __flush_qp(struct iwch_qp *qhp, struct iwch_cq *rchp,
 806                                struct iwch_cq *schp, unsigned long *flag)
 807{
 808        int count;
 809        int flushed;
 810
 811
 812        PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
 813        /* take a ref on the qhp since we must release the lock */
 814        atomic_inc(&qhp->refcnt);
 815        spin_unlock_irqrestore(&qhp->lock, *flag);
 816
 817        /* locking hierarchy: cq lock first, then qp lock. */
 818        spin_lock_irqsave(&rchp->lock, *flag);
 819        spin_lock(&qhp->lock);
 820        cxio_flush_hw_cq(&rchp->cq);
 821        cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
 822        flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
 823        spin_unlock(&qhp->lock);
 824        spin_unlock_irqrestore(&rchp->lock, *flag);
 825        if (flushed) {
 826                spin_lock_irqsave(&rchp->comp_handler_lock, *flag);
 827                (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
 828                spin_unlock_irqrestore(&rchp->comp_handler_lock, *flag);
 829        }
 830
 831        /* locking hierarchy: cq lock first, then qp lock. */
 832        spin_lock_irqsave(&schp->lock, *flag);
 833        spin_lock(&qhp->lock);
 834        cxio_flush_hw_cq(&schp->cq);
 835        cxio_count_scqes(&schp->cq, &qhp->wq, &count);
 836        flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
 837        spin_unlock(&qhp->lock);
 838        spin_unlock_irqrestore(&schp->lock, *flag);
 839        if (flushed) {
 840                spin_lock_irqsave(&schp->comp_handler_lock, *flag);
 841                (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
 842                spin_unlock_irqrestore(&schp->comp_handler_lock, *flag);
 843        }
 844
 845        /* deref */
 846        if (atomic_dec_and_test(&qhp->refcnt))
 847                wake_up(&qhp->wait);
 848
 849        spin_lock_irqsave(&qhp->lock, *flag);
 850}
 851
 852static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
 853{
 854        struct iwch_cq *rchp, *schp;
 855
 856        rchp = get_chp(qhp->rhp, qhp->attr.rcq);
 857        schp = get_chp(qhp->rhp, qhp->attr.scq);
 858
 859        if (qhp->ibqp.uobject) {
 860                cxio_set_wq_in_error(&qhp->wq);
 861                cxio_set_cq_in_error(&rchp->cq);
 862                spin_lock_irqsave(&rchp->comp_handler_lock, *flag);
 863                (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
 864                spin_unlock_irqrestore(&rchp->comp_handler_lock, *flag);
 865                if (schp != rchp) {
 866                        cxio_set_cq_in_error(&schp->cq);
 867                        spin_lock_irqsave(&schp->comp_handler_lock, *flag);
 868                        (*schp->ibcq.comp_handler)(&schp->ibcq,
 869                                                   schp->ibcq.cq_context);
 870                        spin_unlock_irqrestore(&schp->comp_handler_lock, *flag);
 871                }
 872                return;
 873        }
 874        __flush_qp(qhp, rchp, schp, flag);
 875}
 876
 877
 878/*
 879 * Return count of RECV WRs posted
 880 */
 881u16 iwch_rqes_posted(struct iwch_qp *qhp)
 882{
 883        union t3_wr *wqe = qhp->wq.queue;
 884        u16 count = 0;
 885        while ((count+1) != 0 && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
 886                count++;
 887                wqe++;
 888        }
 889        PDBG("%s qhp %p count %u\n", __func__, qhp, count);
 890        return count;
 891}
 892
 893static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
 894                                enum iwch_qp_attr_mask mask,
 895                                struct iwch_qp_attributes *attrs)
 896{
 897        struct t3_rdma_init_attr init_attr;
 898        int ret;
 899
 900        init_attr.tid = qhp->ep->hwtid;
 901        init_attr.qpid = qhp->wq.qpid;
 902        init_attr.pdid = qhp->attr.pd;
 903        init_attr.scqid = qhp->attr.scq;
 904        init_attr.rcqid = qhp->attr.rcq;
 905        init_attr.rq_addr = qhp->wq.rq_addr;
 906        init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
 907        init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
 908                qhp->attr.mpa_attr.recv_marker_enabled |
 909                (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
 910                (qhp->attr.mpa_attr.crc_enabled << 2);
 911
 912        init_attr.qpcaps = uP_RI_QP_RDMA_READ_ENABLE |
 913                           uP_RI_QP_RDMA_WRITE_ENABLE |
 914                           uP_RI_QP_BIND_ENABLE;
 915        if (!qhp->ibqp.uobject)
 916                init_attr.qpcaps |= uP_RI_QP_STAG0_ENABLE |
 917                                    uP_RI_QP_FAST_REGISTER_ENABLE;
 918
 919        init_attr.tcp_emss = qhp->ep->emss;
 920        init_attr.ord = qhp->attr.max_ord;
 921        init_attr.ird = qhp->attr.max_ird;
 922        init_attr.qp_dma_addr = qhp->wq.dma_addr;
 923        init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
 924        init_attr.rqe_count = iwch_rqes_posted(qhp);
 925        init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
 926        init_attr.chan = qhp->ep->l2t->smt_idx;
 927        if (peer2peer) {
 928                init_attr.rtr_type = RTR_READ;
 929                if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
 930                        init_attr.ord = 1;
 931                if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
 932                        init_attr.ird = 1;
 933        } else
 934                init_attr.rtr_type = 0;
 935        init_attr.irs = qhp->ep->rcv_seq;
 936        PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
 937             "flags 0x%x qpcaps 0x%x\n", __func__,
 938             init_attr.rq_addr, init_attr.rq_size,
 939             init_attr.flags, init_attr.qpcaps);
 940        ret = cxio_rdma_init(&rhp->rdev, &init_attr);
 941        PDBG("%s ret %d\n", __func__, ret);
 942        return ret;
 943}
 944
 945int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
 946                                enum iwch_qp_attr_mask mask,
 947                                struct iwch_qp_attributes *attrs,
 948                                int internal)
 949{
 950        int ret = 0;
 951        struct iwch_qp_attributes newattr = qhp->attr;
 952        unsigned long flag;
 953        int disconnect = 0;
 954        int terminate = 0;
 955        int abort = 0;
 956        int free = 0;
 957        struct iwch_ep *ep = NULL;
 958
 959        PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__,
 960             qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
 961             (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
 962
 963        spin_lock_irqsave(&qhp->lock, flag);
 964
 965        /* Process attr changes if in IDLE */
 966        if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
 967                if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
 968                        ret = -EIO;
 969                        goto out;
 970                }
 971                if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
 972                        newattr.enable_rdma_read = attrs->enable_rdma_read;
 973                if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
 974                        newattr.enable_rdma_write = attrs->enable_rdma_write;
 975                if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
 976                        newattr.enable_bind = attrs->enable_bind;
 977                if (mask & IWCH_QP_ATTR_MAX_ORD) {
 978                        if (attrs->max_ord >
 979                            rhp->attr.max_rdma_read_qp_depth) {
 980                                ret = -EINVAL;
 981                                goto out;
 982                        }
 983                        newattr.max_ord = attrs->max_ord;
 984                }
 985                if (mask & IWCH_QP_ATTR_MAX_IRD) {
 986                        if (attrs->max_ird >
 987                            rhp->attr.max_rdma_reads_per_qp) {
 988                                ret = -EINVAL;
 989                                goto out;
 990                        }
 991                        newattr.max_ird = attrs->max_ird;
 992                }
 993                qhp->attr = newattr;
 994        }
 995
 996        if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
 997                goto out;
 998        if (qhp->attr.state == attrs->next_state)
 999                goto out;
1000
1001        switch (qhp->attr.state) {
1002        case IWCH_QP_STATE_IDLE:
1003                switch (attrs->next_state) {
1004                case IWCH_QP_STATE_RTS:
1005                        if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
1006                                ret = -EINVAL;
1007                                goto out;
1008                        }
1009                        if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
1010                                ret = -EINVAL;
1011                                goto out;
1012                        }
1013                        qhp->attr.mpa_attr = attrs->mpa_attr;
1014                        qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1015                        qhp->ep = qhp->attr.llp_stream_handle;
1016                        qhp->attr.state = IWCH_QP_STATE_RTS;
1017
1018                        /*
1019                         * Ref the endpoint here and deref when we
1020                         * disassociate the endpoint from the QP.  This
1021                         * happens in CLOSING->IDLE transition or *->ERROR
1022                         * transition.
1023                         */
1024                        get_ep(&qhp->ep->com);
1025                        spin_unlock_irqrestore(&qhp->lock, flag);
1026                        ret = rdma_init(rhp, qhp, mask, attrs);
1027                        spin_lock_irqsave(&qhp->lock, flag);
1028                        if (ret)
1029                                goto err;
1030                        break;
1031                case IWCH_QP_STATE_ERROR:
1032                        qhp->attr.state = IWCH_QP_STATE_ERROR;
1033                        flush_qp(qhp, &flag);
1034                        break;
1035                default:
1036                        ret = -EINVAL;
1037                        goto out;
1038                }
1039                break;
1040        case IWCH_QP_STATE_RTS:
1041                switch (attrs->next_state) {
1042                case IWCH_QP_STATE_CLOSING:
1043                        BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1044                        qhp->attr.state = IWCH_QP_STATE_CLOSING;
1045                        if (!internal) {
1046                                abort=0;
1047                                disconnect = 1;
1048                                ep = qhp->ep;
1049                                get_ep(&ep->com);
1050                        }
1051                        break;
1052                case IWCH_QP_STATE_TERMINATE:
1053                        qhp->attr.state = IWCH_QP_STATE_TERMINATE;
1054                        if (qhp->ibqp.uobject)
1055                                cxio_set_wq_in_error(&qhp->wq);
1056                        if (!internal)
1057                                terminate = 1;
1058                        break;
1059                case IWCH_QP_STATE_ERROR:
1060                        qhp->attr.state = IWCH_QP_STATE_ERROR;
1061                        if (!internal) {
1062                                abort=1;
1063                                disconnect = 1;
1064                                ep = qhp->ep;
1065                                get_ep(&ep->com);
1066                        }
1067                        goto err;
1068                        break;
1069                default:
1070                        ret = -EINVAL;
1071                        goto out;
1072                }
1073                break;
1074        case IWCH_QP_STATE_CLOSING:
1075                if (!internal) {
1076                        ret = -EINVAL;
1077                        goto out;
1078                }
1079                switch (attrs->next_state) {
1080                        case IWCH_QP_STATE_IDLE:
1081                                flush_qp(qhp, &flag);
1082                                qhp->attr.state = IWCH_QP_STATE_IDLE;
1083                                qhp->attr.llp_stream_handle = NULL;
1084                                put_ep(&qhp->ep->com);
1085                                qhp->ep = NULL;
1086                                wake_up(&qhp->wait);
1087                                break;
1088                        case IWCH_QP_STATE_ERROR:
1089                                goto err;
1090                        default:
1091                                ret = -EINVAL;
1092                                goto err;
1093                }
1094                break;
1095        case IWCH_QP_STATE_ERROR:
1096                if (attrs->next_state != IWCH_QP_STATE_IDLE) {
1097                        ret = -EINVAL;
1098                        goto out;
1099                }
1100
1101                if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
1102                    !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
1103                        ret = -EINVAL;
1104                        goto out;
1105                }
1106                qhp->attr.state = IWCH_QP_STATE_IDLE;
1107                break;
1108        case IWCH_QP_STATE_TERMINATE:
1109                if (!internal) {
1110                        ret = -EINVAL;
1111                        goto out;
1112                }
1113                goto err;
1114                break;
1115        default:
1116                printk(KERN_ERR "%s in a bad state %d\n",
1117                       __func__, qhp->attr.state);
1118                ret = -EINVAL;
1119                goto err;
1120                break;
1121        }
1122        goto out;
1123err:
1124        PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1125             qhp->wq.qpid);
1126
1127        /* disassociate the LLP connection */
1128        qhp->attr.llp_stream_handle = NULL;
1129        ep = qhp->ep;
1130        qhp->ep = NULL;
1131        qhp->attr.state = IWCH_QP_STATE_ERROR;
1132        free=1;
1133        wake_up(&qhp->wait);
1134        BUG_ON(!ep);
1135        flush_qp(qhp, &flag);
1136out:
1137        spin_unlock_irqrestore(&qhp->lock, flag);
1138
1139        if (terminate)
1140                iwch_post_terminate(qhp, NULL);
1141
1142        /*
1143         * If disconnect is 1, then we need to initiate a disconnect
1144         * on the EP.  This can be a normal close (RTS->CLOSING) or
1145         * an abnormal close (RTS/CLOSING->ERROR).
1146         */
1147        if (disconnect) {
1148                iwch_ep_disconnect(ep, abort, GFP_KERNEL);
1149                put_ep(&ep->com);
1150        }
1151
1152        /*
1153         * If free is 1, then we've disassociated the EP from the QP
1154         * and we need to dereference the EP.
1155         */
1156        if (free)
1157                put_ep(&ep->com);
1158
1159        PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1160        return ret;
1161}
1162