err.no Git - linux-2.6/blob - drivers/infiniband/hw/cxgb3/iwch_qp.c

   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 "iwch_provider.h"
  33 #include "iwch.h"
  34 #include "iwch_cm.h"
  35 #include "cxio_hal.h"
  36 #include "cxio_resource.h"
  37
  38 #define NO_SUPPORT -1
  39
  40 static int build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
  41                                 u8 * flit_cnt)
  42 {
  43         int i;
  44         u32 plen;
  45
  46         switch (wr->opcode) {
  47         case IB_WR_SEND:
  48                 if (wr->send_flags & IB_SEND_SOLICITED)
  49                         wqe->send.rdmaop = T3_SEND_WITH_SE;
  50                 else
  51                         wqe->send.rdmaop = T3_SEND;
  52                 wqe->send.rem_stag = 0;
  53                 break;
  54         case IB_WR_SEND_WITH_INV:
  55                 if (wr->send_flags & IB_SEND_SOLICITED)
  56                         wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
  57                 else
  58                         wqe->send.rdmaop = T3_SEND_WITH_INV;
  59                 wqe->send.rem_stag = cpu_to_be32(wr->ex.invalidate_rkey);
  60                 break;
  61         default:
  62                 return -EINVAL;
  63         }
  64         if (wr->num_sge > T3_MAX_SGE)
  65                 return -EINVAL;
  66         wqe->send.reserved[0] = 0;
  67         wqe->send.reserved[1] = 0;
  68         wqe->send.reserved[2] = 0;
  69         plen = 0;
  70         for (i = 0; i < wr->num_sge; i++) {
  71                 if ((plen + wr->sg_list[i].length) < plen)
  72                         return -EMSGSIZE;
  73
  74                 plen += wr->sg_list[i].length;
  75                 wqe->send.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
  76                 wqe->send.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
  77                 wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
  78         }
  79         wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
  80         *flit_cnt = 4 + ((wr->num_sge) << 1);
  81         wqe->send.plen = cpu_to_be32(plen);
  82         return 0;
  83 }
  84
  85 static int build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
  86                                  u8 *flit_cnt)
  87 {
  88         int i;
  89         u32 plen;
  90         if (wr->num_sge > T3_MAX_SGE)
  91                 return -EINVAL;
  92         wqe->write.rdmaop = T3_RDMA_WRITE;
  93         wqe->write.reserved[0] = 0;
  94         wqe->write.reserved[1] = 0;
  95         wqe->write.reserved[2] = 0;
  96         wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
  97         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
  98
  99         if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
 100                 plen = 4;
 101                 wqe->write.sgl[0].stag = wr->ex.imm_data;
 102                 wqe->write.sgl[0].len = __constant_cpu_to_be32(0);
 103                 wqe->write.num_sgle = __constant_cpu_to_be32(0);
 104                 *flit_cnt = 6;
 105         } else {
 106                 plen = 0;
 107                 for (i = 0; i < wr->num_sge; i++) {
 108                         if ((plen + wr->sg_list[i].length) < plen) {
 109                                 return -EMSGSIZE;
 110                         }
 111                         plen += wr->sg_list[i].length;
 112                         wqe->write.sgl[i].stag =
 113                             cpu_to_be32(wr->sg_list[i].lkey);
 114                         wqe->write.sgl[i].len =
 115                             cpu_to_be32(wr->sg_list[i].length);
 116                         wqe->write.sgl[i].to =
 117                             cpu_to_be64(wr->sg_list[i].addr);
 118                 }
 119                 wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
 120                 *flit_cnt = 5 + ((wr->num_sge) << 1);
 121         }
 122         wqe->write.plen = cpu_to_be32(plen);
 123         return 0;
 124 }
 125
 126 static int build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
 127                                 u8 *flit_cnt)
 128 {
 129         if (wr->num_sge > 1)
 130                 return -EINVAL;
 131         wqe->read.rdmaop = T3_READ_REQ;
 132         if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
 133                 wqe->read.local_inv = 1;
 134         else
 135                 wqe->read.local_inv = 0;
 136         wqe->read.reserved[0] = 0;
 137         wqe->read.reserved[1] = 0;
 138         wqe->read.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
 139         wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.remote_addr);
 140         wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
 141         wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
 142         wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
 143         *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
 144         return 0;
 145 }
 146
 147 static int build_fastreg(union t3_wr *wqe, struct ib_send_wr *wr,
 148                                 u8 *flit_cnt, int *wr_cnt, struct t3_wq *wq)
 149 {
 150         int i;
 151         __be64 *p;
 152
 153         if (wr->wr.fast_reg.page_list_len > T3_MAX_FASTREG_DEPTH)
 154                 return -EINVAL;
 155         *wr_cnt = 1;
 156         wqe->fastreg.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
 157         wqe->fastreg.len = cpu_to_be32(wr->wr.fast_reg.length);
 158         wqe->fastreg.va_base_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
 159         wqe->fastreg.va_base_lo_fbo =
 160                                 cpu_to_be32(wr->wr.fast_reg.iova_start & 0xffffffff);
 161         wqe->fastreg.page_type_perms = cpu_to_be32(
 162                 V_FR_PAGE_COUNT(wr->wr.fast_reg.page_list_len) |
 163                 V_FR_PAGE_SIZE(wr->wr.fast_reg.page_shift-12) |
 164                 V_FR_TYPE(TPT_VATO) |
 165                 V_FR_PERMS(iwch_ib_to_tpt_access(wr->wr.fast_reg.access_flags)));
 166         p = &wqe->fastreg.pbl_addrs[0];
 167         for (i = 0; i < wr->wr.fast_reg.page_list_len; i++, p++) {
 168
 169                 /* If we need a 2nd WR, then set it up */
 170                 if (i == T3_MAX_FASTREG_FRAG) {
 171                         *wr_cnt = 2;
 172                         wqe = (union t3_wr *)(wq->queue +
 173                                 Q_PTR2IDX((wq->wptr+1), wq->size_log2));
 174                         build_fw_riwrh((void *)wqe, T3_WR_FASTREG, 0,
 175                                Q_GENBIT(wq->wptr + 1, wq->size_log2),
 176                                0, 1 + wr->wr.fast_reg.page_list_len - T3_MAX_FASTREG_FRAG,
 177                                T3_EOP);
 178
 179                         p = &wqe->pbl_frag.pbl_addrs[0];
 180                 }
 181                 *p = cpu_to_be64((u64)wr->wr.fast_reg.page_list->page_list[i]);
 182         }
 183         *flit_cnt = 5 + wr->wr.fast_reg.page_list_len;
 184         if (*flit_cnt > 15)
 185                 *flit_cnt = 15;
 186         return 0;
 187 }
 188
 189 static int build_inv_stag(union t3_wr *wqe, struct ib_send_wr *wr,
 190                                 u8 *flit_cnt)
 191 {
 192         wqe->local_inv.stag = cpu_to_be32(wr->ex.invalidate_rkey);
 193         wqe->local_inv.reserved = 0;
 194         *flit_cnt = sizeof(struct t3_local_inv_wr) >> 3;
 195         return 0;
 196 }
 197
 198 /*
 199  * TBD: this is going to be moved to firmware. Missing pdid/qpid check for now.
 200  */
 201 static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
 202                             u32 num_sgle, u32 * pbl_addr, u8 * page_size)
 203 {
 204         int i;
 205         struct iwch_mr *mhp;
 206         u32 offset;
 207         for (i = 0; i < num_sgle; i++) {
 208
 209                 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
 210                 if (!mhp) {
 211                         PDBG("%s %d\n", __func__, __LINE__);
 212                         return -EIO;
 213                 }
 214                 if (!mhp->attr.state) {
 215                         PDBG("%s %d\n", __func__, __LINE__);
 216                         return -EIO;
 217                 }
 218                 if (mhp->attr.zbva) {
 219                         PDBG("%s %d\n", __func__, __LINE__);
 220                         return -EIO;
 221                 }
 222
 223                 if (sg_list[i].addr < mhp->attr.va_fbo) {
 224                         PDBG("%s %d\n", __func__, __LINE__);
 225                         return -EINVAL;
 226                 }
 227                 if (sg_list[i].addr + ((u64) sg_list[i].length) <
 228                     sg_list[i].addr) {
 229                         PDBG("%s %d\n", __func__, __LINE__);
 230                         return -EINVAL;
 231                 }
 232                 if (sg_list[i].addr + ((u64) sg_list[i].length) >
 233                     mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
 234                         PDBG("%s %d\n", __func__, __LINE__);
 235                         return -EINVAL;
 236                 }
 237                 offset = sg_list[i].addr - mhp->attr.va_fbo;
 238                 offset += ((u32) mhp->attr.va_fbo) %
 239                           (1UL << (12 + mhp->attr.page_size));
 240                 pbl_addr[i] = ((mhp->attr.pbl_addr -
 241                                 rhp->rdev.rnic_info.pbl_base) >> 3) +
 242                               (offset >> (12 + mhp->attr.page_size));
 243                 page_size[i] = mhp->attr.page_size;
 244         }
 245         return 0;
 246 }
 247
 248 static int build_rdma_recv(struct iwch_qp *qhp, union t3_wr *wqe,
 249                                 struct ib_recv_wr *wr)
 250 {
 251         int i, err = 0;
 252         u32 pbl_addr[T3_MAX_SGE];
 253         u8 page_size[T3_MAX_SGE];
 254
 255         err = iwch_sgl2pbl_map(qhp->rhp, wr->sg_list, wr->num_sge, pbl_addr,
 256                                page_size);
 257         if (err)
 258                 return err;
 259         wqe->recv.pagesz[0] = page_size[0];
 260         wqe->recv.pagesz[1] = page_size[1];
 261         wqe->recv.pagesz[2] = page_size[2];
 262         wqe->recv.pagesz[3] = page_size[3];
 263         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 264         for (i = 0; i < wr->num_sge; i++) {
 265                 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
 266                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 267
 268                 /* to in the WQE == the offset into the page */
 269                 wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
 270                                 (1UL << (12 + page_size[i])));
 271
 272                 /* pbl_addr is the adapters address in the PBL */
 273                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
 274         }
 275         for (; i < T3_MAX_SGE; i++) {
 276                 wqe->recv.sgl[i].stag = 0;
 277                 wqe->recv.sgl[i].len = 0;
 278                 wqe->recv.sgl[i].to = 0;
 279                 wqe->recv.pbl_addr[i] = 0;
 280         }
 281         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 282                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
 283         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 284                              qhp->wq.rq_size_log2)].pbl_addr = 0;
 285         return 0;
 286 }
 287
 288 static int build_zero_stag_recv(struct iwch_qp *qhp, union t3_wr *wqe,
 289                                 struct ib_recv_wr *wr)
 290 {
 291         int i;
 292         u32 pbl_addr;
 293         u32 pbl_offset;
 294
 295
 296         /*
 297          * The T3 HW requires the PBL in the HW recv descriptor to reference
 298          * a PBL entry.  So we allocate the max needed PBL memory here and pass
 299          * it to the uP in the recv WR.  The uP will build the PBL and setup
 300          * the HW recv descriptor.
 301          */
 302         pbl_addr = cxio_hal_pblpool_alloc(&qhp->rhp->rdev, T3_STAG0_PBL_SIZE);
 303         if (!pbl_addr)
 304                 return -ENOMEM;
 305
 306         /*
 307          * Compute the 8B aligned offset.
 308          */
 309         pbl_offset = (pbl_addr - qhp->rhp->rdev.rnic_info.pbl_base) >> 3;
 310
 311         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 312
 313         for (i = 0; i < wr->num_sge; i++) {
 314
 315                 /*
 316                  * Use a 128MB page size. This and an imposed 128MB
 317                  * sge length limit allows us to require only a 2-entry HW
 318                  * PBL for each SGE.  This restriction is acceptable since
 319                  * since it is not possible to allocate 128MB of contiguous
 320                  * DMA coherent memory!
 321                  */
 322                 if (wr->sg_list[i].length > T3_STAG0_MAX_PBE_LEN)
 323                         return -EINVAL;
 324                 wqe->recv.pagesz[i] = T3_STAG0_PAGE_SHIFT;
 325
 326                 /*
 327                  * T3 restricts a recv to all zero-stag or all non-zero-stag.
 328                  */
 329                 if (wr->sg_list[i].lkey != 0)
 330                         return -EINVAL;
 331                 wqe->recv.sgl[i].stag = 0;
 332                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 333                 wqe->recv.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
 334                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_offset);
 335                 pbl_offset += 2;
 336         }
 337         for (; i < T3_MAX_SGE; i++) {
 338                 wqe->recv.pagesz[i] = 0;
 339                 wqe->recv.sgl[i].stag = 0;
 340                 wqe->recv.sgl[i].len = 0;
 341                 wqe->recv.sgl[i].to = 0;
 342                 wqe->recv.pbl_addr[i] = 0;
 343         }
 344         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 345                              qhp->wq.rq_size_log2)].wr_id = wr->wr_id;
 346         qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr,
 347                              qhp->wq.rq_size_log2)].pbl_addr = pbl_addr;
 348         return 0;
 349 }
 350
 351 int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 352                       struct ib_send_wr **bad_wr)
 353 {
 354         int err = 0;
 355         u8 uninitialized_var(t3_wr_flit_cnt);
 356         enum t3_wr_opcode t3_wr_opcode = 0;
 357         enum t3_wr_flags t3_wr_flags;
 358         struct iwch_qp *qhp;
 359         u32 idx;
 360         union t3_wr *wqe;
 361         u32 num_wrs;
 362         unsigned long flag;
 363         struct t3_swsq *sqp;
 364         int wr_cnt = 1;
 365
 366         qhp = to_iwch_qp(ibqp);
 367         spin_lock_irqsave(&qhp->lock, flag);
 368         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 369                 spin_unlock_irqrestore(&qhp->lock, flag);
 370                 return -EINVAL;
 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                 return -ENOMEM;
 377         }
 378         while (wr) {
 379                 if (num_wrs == 0) {
 380                         err = -ENOMEM;
 381                         *bad_wr = wr;
 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                         *bad_wr = wr;
 435                         break;
 436                 }
 437                 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 438                 sqp->wr_id = wr->wr_id;
 439                 sqp->opcode = wr2opcode(t3_wr_opcode);
 440                 sqp->sq_wptr = qhp->wq.sq_wptr;
 441                 sqp->complete = 0;
 442                 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
 443
 444                 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
 445                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 446                                0, t3_wr_flit_cnt,
 447                                (wr_cnt == 1) ? T3_SOPEOP : T3_SOP);
 448                 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
 449                      __func__, (unsigned long long) wr->wr_id, idx,
 450                      Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
 451                      sqp->opcode);
 452                 wr = wr->next;
 453                 num_wrs--;
 454                 qhp->wq.wptr += wr_cnt;
 455                 ++(qhp->wq.sq_wptr);
 456         }
 457         spin_unlock_irqrestore(&qhp->lock, flag);
 458         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 459         return err;
 460 }
 461
 462 int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 463                       struct ib_recv_wr **bad_wr)
 464 {
 465         int err = 0;
 466         struct iwch_qp *qhp;
 467         u32 idx;
 468         union t3_wr *wqe;
 469         u32 num_wrs;
 470         unsigned long flag;
 471
 472         qhp = to_iwch_qp(ibqp);
 473         spin_lock_irqsave(&qhp->lock, flag);
 474         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 475                 spin_unlock_irqrestore(&qhp->lock, flag);
 476                 return -EINVAL;
 477         }
 478         num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
 479                             qhp->wq.rq_size_log2) - 1;
 480         if (!wr) {
 481                 spin_unlock_irqrestore(&qhp->lock, flag);
 482                 return -EINVAL;
 483         }
 484         while (wr) {
 485                 if (wr->num_sge > T3_MAX_SGE) {
 486                         err = -EINVAL;
 487                         *bad_wr = wr;
 488                         break;
 489                 }
 490                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 491                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
 492                 if (num_wrs)
 493                         if (wr->sg_list[0].lkey)
 494                                 err = build_rdma_recv(qhp, wqe, wr);
 495                         else
 496                                 err = build_zero_stag_recv(qhp, wqe, wr);
 497                 else
 498                         err = -ENOMEM;
 499                 if (err) {
 500                         *bad_wr = wr;
 501                         break;
 502                 }
 503                 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
 504                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 505                                0, sizeof(struct t3_receive_wr) >> 3, T3_SOPEOP);
 506                 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
 507                      "wqe %p \n", __func__, (unsigned long long) wr->wr_id,
 508                      idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
 509                 ++(qhp->wq.rq_wptr);
 510                 ++(qhp->wq.wptr);
 511                 wr = wr->next;
 512                 num_wrs--;
 513         }
 514         spin_unlock_irqrestore(&qhp->lock, flag);
 515         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 516         return err;
 517 }
 518
 519 int iwch_bind_mw(struct ib_qp *qp,
 520                              struct ib_mw *mw,
 521                              struct ib_mw_bind *mw_bind)
 522 {
 523         struct iwch_dev *rhp;
 524         struct iwch_mw *mhp;
 525         struct iwch_qp *qhp;
 526         union t3_wr *wqe;
 527         u32 pbl_addr;
 528         u8 page_size;
 529         u32 num_wrs;
 530         unsigned long flag;
 531         struct ib_sge sgl;
 532         int err=0;
 533         enum t3_wr_flags t3_wr_flags;
 534         u32 idx;
 535         struct t3_swsq *sqp;
 536
 537         qhp = to_iwch_qp(qp);
 538         mhp = to_iwch_mw(mw);
 539         rhp = qhp->rhp;
 540
 541         spin_lock_irqsave(&qhp->lock, flag);
 542         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 543                 spin_unlock_irqrestore(&qhp->lock, flag);
 544                 return -EINVAL;
 545         }
 546         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
 547                             qhp->wq.sq_size_log2);
 548         if ((num_wrs) <= 0) {
 549                 spin_unlock_irqrestore(&qhp->lock, flag);
 550                 return -ENOMEM;
 551         }
 552         idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 553         PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __func__, idx,
 554              mw, mw_bind);
 555         wqe = (union t3_wr *) (qhp->wq.queue + idx);
 556
 557         t3_wr_flags = 0;
 558         if (mw_bind->send_flags & IB_SEND_SIGNALED)
 559                 t3_wr_flags = T3_COMPLETION_FLAG;
 560
 561         sgl.addr = mw_bind->addr;
 562         sgl.lkey = mw_bind->mr->lkey;
 563         sgl.length = mw_bind->length;
 564         wqe->bind.reserved = 0;
 565         wqe->bind.type = TPT_VATO;
 566
 567         /* TBD: check perms */
 568         wqe->bind.perms = iwch_ib_to_tpt_access(mw_bind->mw_access_flags);
 569         wqe->bind.mr_stag = cpu_to_be32(mw_bind->mr->lkey);
 570         wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
 571         wqe->bind.mw_len = cpu_to_be32(mw_bind->length);
 572         wqe->bind.mw_va = cpu_to_be64(mw_bind->addr);
 573         err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
 574         if (err) {
 575                 spin_unlock_irqrestore(&qhp->lock, flag);
 576                 return err;
 577         }
 578         wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 579         sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
 580         sqp->wr_id = mw_bind->wr_id;
 581         sqp->opcode = T3_BIND_MW;
 582         sqp->sq_wptr = qhp->wq.sq_wptr;
 583         sqp->complete = 0;
 584         sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
 585         wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
 586         wqe->bind.mr_pagesz = page_size;
 587         build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
 588                        Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
 589                        sizeof(struct t3_bind_mw_wr) >> 3, T3_SOPEOP);
 590         ++(qhp->wq.wptr);
 591         ++(qhp->wq.sq_wptr);
 592         spin_unlock_irqrestore(&qhp->lock, flag);
 593
 594         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 595
 596         return err;
 597 }
 598
 599 static inline void build_term_codes(struct respQ_msg_t *rsp_msg,
 600                                     u8 *layer_type, u8 *ecode)
 601 {
 602         int status = TPT_ERR_INTERNAL_ERR;
 603         int tagged = 0;
 604         int opcode = -1;
 605         int rqtype = 0;
 606         int send_inv = 0;
 607
 608         if (rsp_msg) {
 609                 status = CQE_STATUS(rsp_msg->cqe);
 610                 opcode = CQE_OPCODE(rsp_msg->cqe);
 611                 rqtype = RQ_TYPE(rsp_msg->cqe);
 612                 send_inv = (opcode == T3_SEND_WITH_INV) ||
 613                            (opcode == T3_SEND_WITH_SE_INV);
 614                 tagged = (opcode == T3_RDMA_WRITE) ||
 615                          (rqtype && (opcode == T3_READ_RESP));
 616         }
 617
 618         switch (status) {
 619         case TPT_ERR_STAG:
 620                 if (send_inv) {
 621                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 622                         *ecode = RDMAP_CANT_INV_STAG;
 623                 } else {
 624                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 625                         *ecode = RDMAP_INV_STAG;
 626                 }
 627                 break;
 628         case TPT_ERR_PDID:
 629                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 630                 if ((opcode == T3_SEND_WITH_INV) ||
 631                     (opcode == T3_SEND_WITH_SE_INV))
 632                         *ecode = RDMAP_CANT_INV_STAG;
 633                 else
 634                         *ecode = RDMAP_STAG_NOT_ASSOC;
 635                 break;
 636         case TPT_ERR_QPID:
 637                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 638                 *ecode = RDMAP_STAG_NOT_ASSOC;
 639                 break;
 640         case TPT_ERR_ACCESS:
 641                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 642                 *ecode = RDMAP_ACC_VIOL;
 643                 break;
 644         case TPT_ERR_WRAP:
 645                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 646                 *ecode = RDMAP_TO_WRAP;
 647                 break;
 648         case TPT_ERR_BOUND:
 649                 if (tagged) {
 650                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 651                         *ecode = DDPT_BASE_BOUNDS;
 652                 } else {
 653                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 654                         *ecode = RDMAP_BASE_BOUNDS;
 655                 }
 656                 break;
 657         case TPT_ERR_INVALIDATE_SHARED_MR:
 658         case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
 659                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 660                 *ecode = RDMAP_CANT_INV_STAG;
 661                 break;
 662         case TPT_ERR_ECC:
 663         case TPT_ERR_ECC_PSTAG:
 664         case TPT_ERR_INTERNAL_ERR:
 665                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
 666                 *ecode = 0;
 667                 break;
 668         case TPT_ERR_OUT_OF_RQE:
 669                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 670                 *ecode = DDPU_INV_MSN_NOBUF;
 671                 break;
 672         case TPT_ERR_PBL_ADDR_BOUND:
 673                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 674                 *ecode = DDPT_BASE_BOUNDS;
 675                 break;
 676         case TPT_ERR_CRC:
 677                 *layer_type = LAYER_MPA|DDP_LLP;
 678                 *ecode = MPA_CRC_ERR;
 679                 break;
 680         case TPT_ERR_MARKER:
 681                 *layer_type = LAYER_MPA|DDP_LLP;
 682                 *ecode = MPA_MARKER_ERR;
 683                 break;
 684         case TPT_ERR_PDU_LEN_ERR:
 685                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 686                 *ecode = DDPU_MSG_TOOBIG;
 687                 break;
 688         case TPT_ERR_DDP_VERSION:
 689                 if (tagged) {
 690                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 691                         *ecode = DDPT_INV_VERS;
 692                 } else {
 693                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 694                         *ecode = DDPU_INV_VERS;
 695                 }
 696                 break;
 697         case TPT_ERR_RDMA_VERSION:
 698                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 699                 *ecode = RDMAP_INV_VERS;
 700                 break;
 701         case TPT_ERR_OPCODE:
 702                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 703                 *ecode = RDMAP_INV_OPCODE;
 704                 break;
 705         case TPT_ERR_DDP_QUEUE_NUM:
 706                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 707                 *ecode = DDPU_INV_QN;
 708                 break;
 709         case TPT_ERR_MSN:
 710         case TPT_ERR_MSN_GAP:
 711         case TPT_ERR_MSN_RANGE:
 712         case TPT_ERR_IRD_OVERFLOW:
 713                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 714                 *ecode = DDPU_INV_MSN_RANGE;
 715                 break;
 716         case TPT_ERR_TBIT:
 717                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
 718                 *ecode = 0;
 719                 break;
 720         case TPT_ERR_MO:
 721                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 722                 *ecode = DDPU_INV_MO;
 723                 break;
 724         default:
 725                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 726                 *ecode = 0;
 727                 break;
 728         }
 729 }
 730
 731 int iwch_post_zb_read(struct iwch_qp *qhp)
 732 {
 733         union t3_wr *wqe;
 734         struct sk_buff *skb;
 735         u8 flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
 736
 737         PDBG("%s enter\n", __func__);
 738         skb = alloc_skb(40, GFP_KERNEL);
 739         if (!skb) {
 740                 printk(KERN_ERR "%s cannot send zb_read!!\n", __func__);
 741                 return -ENOMEM;
 742         }
 743         wqe = (union t3_wr *)skb_put(skb, sizeof(struct t3_rdma_read_wr));
 744         memset(wqe, 0, sizeof(struct t3_rdma_read_wr));
 745         wqe->read.rdmaop = T3_READ_REQ;
 746         wqe->read.reserved[0] = 0;
 747         wqe->read.reserved[1] = 0;
 748         wqe->read.reserved[2] = 0;
 749         wqe->read.rem_stag = cpu_to_be32(1);
 750         wqe->read.rem_to = cpu_to_be64(1);
 751         wqe->read.local_stag = cpu_to_be32(1);
 752         wqe->read.local_len = cpu_to_be32(0);
 753         wqe->read.local_to = cpu_to_be64(1);
 754         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_READ));
 755         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid)|
 756                                                 V_FW_RIWR_LEN(flit_cnt));
 757         skb->priority = CPL_PRIORITY_DATA;
 758         return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
 759 }
 760
 761 /*
 762  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
 763  */
 764 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
 765 {
 766         union t3_wr *wqe;
 767         struct terminate_message *term;
 768         struct sk_buff *skb;
 769
 770         PDBG("%s %d\n", __func__, __LINE__);
 771         skb = alloc_skb(40, GFP_ATOMIC);
 772         if (!skb) {
 773                 printk(KERN_ERR "%s cannot send TERMINATE!\n", __func__);
 774                 return -ENOMEM;
 775         }
 776         wqe = (union t3_wr *)skb_put(skb, 40);
 777         memset(wqe, 0, 40);
 778         wqe->send.rdmaop = T3_TERMINATE;
 779
 780         /* immediate data length */
 781         wqe->send.plen = htonl(4);
 782
 783         /* immediate data starts here. */
 784         term = (struct terminate_message *)wqe->send.sgl;
 785         build_term_codes(rsp_msg, &term->layer_etype, &term->ecode);
 786         wqe->send.wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_SEND) |
 787                          V_FW_RIWR_FLAGS(T3_COMPLETION_FLAG | T3_NOTIFY_FLAG));
 788         wqe->send.wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(qhp->ep->hwtid));
 789         skb->priority = CPL_PRIORITY_DATA;
 790         return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
 791 }
 792
 793 /*
 794  * Assumes qhp lock is held.
 795  */
 796 static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag)
 797 {
 798         struct iwch_cq *rchp, *schp;
 799         int count;
 800         int flushed;
 801
 802         rchp = get_chp(qhp->rhp, qhp->attr.rcq);
 803         schp = get_chp(qhp->rhp, qhp->attr.scq);
 804
 805         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
 806         /* take a ref on the qhp since we must release the lock */
 807         atomic_inc(&qhp->refcnt);
 808         spin_unlock_irqrestore(&qhp->lock, *flag);
 809
 810         /* locking heirarchy: cq lock first, then qp lock. */
 811         spin_lock_irqsave(&rchp->lock, *flag);
 812         spin_lock(&qhp->lock);
 813         cxio_flush_hw_cq(&rchp->cq);
 814         cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
 815         flushed = cxio_flush_rq(&qhp->wq, &rchp->cq, count);
 816         spin_unlock(&qhp->lock);
 817         spin_unlock_irqrestore(&rchp->lock, *flag);
 818         if (flushed)
 819                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
 820
 821         /* locking heirarchy: cq lock first, then qp lock. */
 822         spin_lock_irqsave(&schp->lock, *flag);
 823         spin_lock(&qhp->lock);
 824         cxio_flush_hw_cq(&schp->cq);
 825         cxio_count_scqes(&schp->cq, &qhp->wq, &count);
 826         flushed = cxio_flush_sq(&qhp->wq, &schp->cq, count);
 827         spin_unlock(&qhp->lock);
 828         spin_unlock_irqrestore(&schp->lock, *flag);
 829         if (flushed)
 830                 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
 831
 832         /* deref */
 833         if (atomic_dec_and_test(&qhp->refcnt))
 834                 wake_up(&qhp->wait);
 835
 836         spin_lock_irqsave(&qhp->lock, *flag);
 837 }
 838
 839 static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
 840 {
 841         if (qhp->ibqp.uobject)
 842                 cxio_set_wq_in_error(&qhp->wq);
 843         else
 844                 __flush_qp(qhp, flag);
 845 }
 846
 847
 848 /*
 849  * Return count of RECV WRs posted
 850  */
 851 u16 iwch_rqes_posted(struct iwch_qp *qhp)
 852 {
 853         union t3_wr *wqe = qhp->wq.queue;
 854         u16 count = 0;
 855         while ((count+1) != 0 && fw_riwrh_opcode((struct fw_riwrh *)wqe) == T3_WR_RCV) {
 856                 count++;
 857                 wqe++;
 858         }
 859         PDBG("%s qhp %p count %u\n", __func__, qhp, count);
 860         return count;
 861 }
 862
 863 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
 864                                 enum iwch_qp_attr_mask mask,
 865                                 struct iwch_qp_attributes *attrs)
 866 {
 867         struct t3_rdma_init_attr init_attr;
 868         int ret;
 869
 870         init_attr.tid = qhp->ep->hwtid;
 871         init_attr.qpid = qhp->wq.qpid;
 872         init_attr.pdid = qhp->attr.pd;
 873         init_attr.scqid = qhp->attr.scq;
 874         init_attr.rcqid = qhp->attr.rcq;
 875         init_attr.rq_addr = qhp->wq.rq_addr;
 876         init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
 877         init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
 878                 qhp->attr.mpa_attr.recv_marker_enabled |
 879                 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
 880                 (qhp->attr.mpa_attr.crc_enabled << 2);
 881
 882         /*
 883          * XXX - The IWCM doesn't quite handle getting these
 884          * attrs set before going into RTS.  For now, just turn
 885          * them on always...
 886          */
 887 #if 0
 888         init_attr.qpcaps = qhp->attr.enableRdmaRead |
 889                 (qhp->attr.enableRdmaWrite << 1) |
 890                 (qhp->attr.enableBind << 2) |
 891                 (qhp->attr.enable_stag0_fastreg << 3) |
 892                 (qhp->attr.enable_stag0_fastreg << 4);
 893 #else
 894         init_attr.qpcaps = 0x1f;
 895 #endif
 896         init_attr.tcp_emss = qhp->ep->emss;
 897         init_attr.ord = qhp->attr.max_ord;
 898         init_attr.ird = qhp->attr.max_ird;
 899         init_attr.qp_dma_addr = qhp->wq.dma_addr;
 900         init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
 901         init_attr.rqe_count = iwch_rqes_posted(qhp);
 902         init_attr.flags = qhp->attr.mpa_attr.initiator ? MPA_INITIATOR : 0;
 903         if (!qhp->ibqp.uobject)
 904                 init_attr.flags |= PRIV_QP;
 905         if (peer2peer) {
 906                 init_attr.rtr_type = RTR_READ;
 907                 if (init_attr.ord == 0 && qhp->attr.mpa_attr.initiator)
 908                         init_attr.ord = 1;
 909                 if (init_attr.ird == 0 && !qhp->attr.mpa_attr.initiator)
 910                         init_attr.ird = 1;
 911         } else
 912                 init_attr.rtr_type = 0;
 913         init_attr.irs = qhp->ep->rcv_seq;
 914         PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
 915              "flags 0x%x qpcaps 0x%x\n", __func__,
 916              init_attr.rq_addr, init_attr.rq_size,
 917              init_attr.flags, init_attr.qpcaps);
 918         ret = cxio_rdma_init(&rhp->rdev, &init_attr);
 919         PDBG("%s ret %d\n", __func__, ret);
 920         return ret;
 921 }
 922
 923 int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
 924                                 enum iwch_qp_attr_mask mask,
 925                                 struct iwch_qp_attributes *attrs,
 926                                 int internal)
 927 {
 928         int ret = 0;
 929         struct iwch_qp_attributes newattr = qhp->attr;
 930         unsigned long flag;
 931         int disconnect = 0;
 932         int terminate = 0;
 933         int abort = 0;
 934         int free = 0;
 935         struct iwch_ep *ep = NULL;
 936
 937         PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __func__,
 938              qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
 939              (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
 940
 941         spin_lock_irqsave(&qhp->lock, flag);
 942
 943         /* Process attr changes if in IDLE */
 944         if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
 945                 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
 946                         ret = -EIO;
 947                         goto out;
 948                 }
 949                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
 950                         newattr.enable_rdma_read = attrs->enable_rdma_read;
 951                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
 952                         newattr.enable_rdma_write = attrs->enable_rdma_write;
 953                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
 954                         newattr.enable_bind = attrs->enable_bind;
 955                 if (mask & IWCH_QP_ATTR_MAX_ORD) {
 956                         if (attrs->max_ord >
 957                             rhp->attr.max_rdma_read_qp_depth) {
 958                                 ret = -EINVAL;
 959                                 goto out;
 960                         }
 961                         newattr.max_ord = attrs->max_ord;
 962                 }
 963                 if (mask & IWCH_QP_ATTR_MAX_IRD) {
 964                         if (attrs->max_ird >
 965                             rhp->attr.max_rdma_reads_per_qp) {
 966                                 ret = -EINVAL;
 967                                 goto out;
 968                         }
 969                         newattr.max_ird = attrs->max_ird;
 970                 }
 971                 qhp->attr = newattr;
 972         }
 973
 974         if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
 975                 goto out;
 976         if (qhp->attr.state == attrs->next_state)
 977                 goto out;
 978
 979         switch (qhp->attr.state) {
 980         case IWCH_QP_STATE_IDLE:
 981                 switch (attrs->next_state) {
 982                 case IWCH_QP_STATE_RTS:
 983                         if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
 984                                 ret = -EINVAL;
 985                                 goto out;
 986                         }
 987                         if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
 988                                 ret = -EINVAL;
 989                                 goto out;
 990                         }
 991                         qhp->attr.mpa_attr = attrs->mpa_attr;
 992                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
 993                         qhp->ep = qhp->attr.llp_stream_handle;
 994                         qhp->attr.state = IWCH_QP_STATE_RTS;
 995
 996                         /*
 997                          * Ref the endpoint here and deref when we
 998                          * disassociate the endpoint from the QP.  This
 999                          * happens in CLOSING->IDLE transition or *->ERROR
1000                          * transition.
1001                          */
1002                         get_ep(&qhp->ep->com);
1003                         spin_unlock_irqrestore(&qhp->lock, flag);
1004                         ret = rdma_init(rhp, qhp, mask, attrs);
1005                         spin_lock_irqsave(&qhp->lock, flag);
1006                         if (ret)
1007                                 goto err;
1008                         break;
1009                 case IWCH_QP_STATE_ERROR:
1010                         qhp->attr.state = IWCH_QP_STATE_ERROR;
1011                         flush_qp(qhp, &flag);
1012                         break;
1013                 default:
1014                         ret = -EINVAL;
1015                         goto out;
1016                 }
1017                 break;
1018         case IWCH_QP_STATE_RTS:
1019                 switch (attrs->next_state) {
1020                 case IWCH_QP_STATE_CLOSING:
1021                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1022                         qhp->attr.state = IWCH_QP_STATE_CLOSING;
1023                         if (!internal) {
1024                                 abort=0;
1025                                 disconnect = 1;
1026                                 ep = qhp->ep;
1027                                 get_ep(&ep->com);
1028                         }
1029                         break;
1030                 case IWCH_QP_STATE_TERMINATE:
1031                         qhp->attr.state = IWCH_QP_STATE_TERMINATE;
1032                         if (qhp->ibqp.uobject)
1033                                 cxio_set_wq_in_error(&qhp->wq);
1034                         if (!internal)
1035                                 terminate = 1;
1036                         break;
1037                 case IWCH_QP_STATE_ERROR:
1038                         qhp->attr.state = IWCH_QP_STATE_ERROR;
1039                         if (!internal) {
1040                                 abort=1;
1041                                 disconnect = 1;
1042                                 ep = qhp->ep;
1043                                 get_ep(&ep->com);
1044                         }
1045                         goto err;
1046                         break;
1047                 default:
1048                         ret = -EINVAL;
1049                         goto out;
1050                 }
1051                 break;
1052         case IWCH_QP_STATE_CLOSING:
1053                 if (!internal) {
1054                         ret = -EINVAL;
1055                         goto out;
1056                 }
1057                 switch (attrs->next_state) {
1058                         case IWCH_QP_STATE_IDLE:
1059                                 flush_qp(qhp, &flag);
1060                                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1061                                 qhp->attr.llp_stream_handle = NULL;
1062                                 put_ep(&qhp->ep->com);
1063                                 qhp->ep = NULL;
1064                                 wake_up(&qhp->wait);
1065                                 break;
1066                         case IWCH_QP_STATE_ERROR:
1067                                 goto err;
1068                         default:
1069                                 ret = -EINVAL;
1070                                 goto err;
1071                 }
1072                 break;
1073         case IWCH_QP_STATE_ERROR:
1074                 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
1075                         ret = -EINVAL;
1076                         goto out;
1077                 }
1078
1079                 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
1080                     !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
1081                         ret = -EINVAL;
1082                         goto out;
1083                 }
1084                 qhp->attr.state = IWCH_QP_STATE_IDLE;
1085                 memset(&qhp->attr, 0, sizeof(qhp->attr));
1086                 break;
1087         case IWCH_QP_STATE_TERMINATE:
1088                 if (!internal) {
1089                         ret = -EINVAL;
1090                         goto out;
1091                 }
1092                 goto err;
1093                 break;
1094         default:
1095                 printk(KERN_ERR "%s in a bad state %d\n",
1096                        __func__, qhp->attr.state);
1097                 ret = -EINVAL;
1098                 goto err;
1099                 break;
1100         }
1101         goto out;
1102 err:
1103         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1104              qhp->wq.qpid);
1105
1106         /* disassociate the LLP connection */
1107         qhp->attr.llp_stream_handle = NULL;
1108         ep = qhp->ep;
1109         qhp->ep = NULL;
1110         qhp->attr.state = IWCH_QP_STATE_ERROR;
1111         free=1;
1112         wake_up(&qhp->wait);
1113         BUG_ON(!ep);
1114         flush_qp(qhp, &flag);
1115 out:
1116         spin_unlock_irqrestore(&qhp->lock, flag);
1117
1118         if (terminate)
1119                 iwch_post_terminate(qhp, NULL);
1120
1121         /*
1122          * If disconnect is 1, then we need to initiate a disconnect
1123          * on the EP.  This can be a normal close (RTS->CLOSING) or
1124          * an abnormal close (RTS/CLOSING->ERROR).
1125          */
1126         if (disconnect) {
1127                 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
1128                 put_ep(&ep->com);
1129         }
1130
1131         /*
1132          * If free is 1, then we've disassociated the EP from the QP
1133          * and we need to dereference the EP.
1134          */
1135         if (free)
1136                 put_ep(&ep->com);
1137
1138         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1139         return ret;
1140 }
1141
1142 static int quiesce_qp(struct iwch_qp *qhp)
1143 {
1144         spin_lock_irq(&qhp->lock);
1145         iwch_quiesce_tid(qhp->ep);
1146         qhp->flags |= QP_QUIESCED;
1147         spin_unlock_irq(&qhp->lock);
1148         return 0;
1149 }
1150
1151 static int resume_qp(struct iwch_qp *qhp)
1152 {
1153         spin_lock_irq(&qhp->lock);
1154         iwch_resume_tid(qhp->ep);
1155         qhp->flags &= ~QP_QUIESCED;
1156         spin_unlock_irq(&qhp->lock);
1157         return 0;
1158 }
1159
1160 int iwch_quiesce_qps(struct iwch_cq *chp)
1161 {
1162         int i;
1163         struct iwch_qp *qhp;
1164
1165         for (i=0; i < T3_MAX_NUM_QP; i++) {
1166                 qhp = get_qhp(chp->rhp, i);
1167                 if (!qhp)
1168                         continue;
1169                 if ((qhp->attr.rcq == chp->cq.cqid) && !qp_quiesced(qhp)) {
1170                         quiesce_qp(qhp);
1171                         continue;
1172                 }
1173                 if ((qhp->attr.scq == chp->cq.cqid) && !qp_quiesced(qhp))
1174                         quiesce_qp(qhp);
1175         }
1176         return 0;
1177 }
1178
1179 int iwch_resume_qps(struct iwch_cq *chp)
1180 {
1181         int i;
1182         struct iwch_qp *qhp;
1183
1184         for (i=0; i < T3_MAX_NUM_QP; i++) {
1185                 qhp = get_qhp(chp->rhp, i);
1186                 if (!qhp)
1187                         continue;
1188                 if ((qhp->attr.rcq == chp->cq.cqid) && qp_quiesced(qhp)) {
1189                         resume_qp(qhp);
1190                         continue;
1191                 }
1192                 if ((qhp->attr.scq == chp->cq.cqid) && qp_quiesced(qhp))
1193                         resume_qp(qhp);
1194         }
1195         return 0;
1196 }