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
  37 #define NO_SUPPORT -1
  38
  39 static int iwch_build_rdma_send(union t3_wr *wqe, struct ib_send_wr *wr,
  40                                 u8 * flit_cnt)
  41 {
  42         int i;
  43         u32 plen;
  44
  45         switch (wr->opcode) {
  46         case IB_WR_SEND:
  47         case IB_WR_SEND_WITH_IMM:
  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 #if 0                           /* Not currently supported */
  55         case TYPE_SEND_INVALIDATE:
  56         case TYPE_SEND_INVALIDATE_IMMEDIATE:
  57                 wqe->send.rdmaop = T3_SEND_WITH_INV;
  58                 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
  59                 break;
  60         case TYPE_SEND_SE_INVALIDATE:
  61                 wqe->send.rdmaop = T3_SEND_WITH_SE_INV;
  62                 wqe->send.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
  63                 break;
  64 #endif
  65         default:
  66                 break;
  67         }
  68         if (wr->num_sge > T3_MAX_SGE)
  69                 return -EINVAL;
  70         wqe->send.reserved[0] = 0;
  71         wqe->send.reserved[1] = 0;
  72         wqe->send.reserved[2] = 0;
  73         if (wr->opcode == IB_WR_SEND_WITH_IMM) {
  74                 plen = 4;
  75                 wqe->send.sgl[0].stag = wr->imm_data;
  76                 wqe->send.sgl[0].len = __constant_cpu_to_be32(0);
  77                 wqe->send.num_sgle = __constant_cpu_to_be32(0);
  78                 *flit_cnt = 5;
  79         } else {
  80                 plen = 0;
  81                 for (i = 0; i < wr->num_sge; i++) {
  82                         if ((plen + wr->sg_list[i].length) < plen) {
  83                                 return -EMSGSIZE;
  84                         }
  85                         plen += wr->sg_list[i].length;
  86                         wqe->send.sgl[i].stag =
  87                             cpu_to_be32(wr->sg_list[i].lkey);
  88                         wqe->send.sgl[i].len =
  89                             cpu_to_be32(wr->sg_list[i].length);
  90                         wqe->send.sgl[i].to = cpu_to_be64(wr->sg_list[i].addr);
  91                 }
  92                 wqe->send.num_sgle = cpu_to_be32(wr->num_sge);
  93                 *flit_cnt = 4 + ((wr->num_sge) << 1);
  94         }
  95         wqe->send.plen = cpu_to_be32(plen);
  96         return 0;
  97 }
  98
  99 static int iwch_build_rdma_write(union t3_wr *wqe, struct ib_send_wr *wr,
 100                                  u8 *flit_cnt)
 101 {
 102         int i;
 103         u32 plen;
 104         if (wr->num_sge > T3_MAX_SGE)
 105                 return -EINVAL;
 106         wqe->write.rdmaop = T3_RDMA_WRITE;
 107         wqe->write.reserved[0] = 0;
 108         wqe->write.reserved[1] = 0;
 109         wqe->write.reserved[2] = 0;
 110         wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
 111         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
 112
 113         if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM) {
 114                 plen = 4;
 115                 wqe->write.sgl[0].stag = wr->imm_data;
 116                 wqe->write.sgl[0].len = __constant_cpu_to_be32(0);
 117                 wqe->write.num_sgle = __constant_cpu_to_be32(0);
 118                 *flit_cnt = 6;
 119         } else {
 120                 plen = 0;
 121                 for (i = 0; i < wr->num_sge; i++) {
 122                         if ((plen + wr->sg_list[i].length) < plen) {
 123                                 return -EMSGSIZE;
 124                         }
 125                         plen += wr->sg_list[i].length;
 126                         wqe->write.sgl[i].stag =
 127                             cpu_to_be32(wr->sg_list[i].lkey);
 128                         wqe->write.sgl[i].len =
 129                             cpu_to_be32(wr->sg_list[i].length);
 130                         wqe->write.sgl[i].to =
 131                             cpu_to_be64(wr->sg_list[i].addr);
 132                 }
 133                 wqe->write.num_sgle = cpu_to_be32(wr->num_sge);
 134                 *flit_cnt = 5 + ((wr->num_sge) << 1);
 135         }
 136         wqe->write.plen = cpu_to_be32(plen);
 137         return 0;
 138 }
 139
 140 static int iwch_build_rdma_read(union t3_wr *wqe, struct ib_send_wr *wr,
 141                                 u8 *flit_cnt)
 142 {
 143         if (wr->num_sge > 1)
 144                 return -EINVAL;
 145         wqe->read.rdmaop = T3_READ_REQ;
 146         wqe->read.reserved[0] = 0;
 147         wqe->read.reserved[1] = 0;
 148         wqe->read.reserved[2] = 0;
 149         wqe->read.rem_stag = cpu_to_be32(wr->wr.rdma.rkey);
 150         wqe->read.rem_to = cpu_to_be64(wr->wr.rdma.remote_addr);
 151         wqe->read.local_stag = cpu_to_be32(wr->sg_list[0].lkey);
 152         wqe->read.local_len = cpu_to_be32(wr->sg_list[0].length);
 153         wqe->read.local_to = cpu_to_be64(wr->sg_list[0].addr);
 154         *flit_cnt = sizeof(struct t3_rdma_read_wr) >> 3;
 155         return 0;
 156 }
 157
 158 /*
 159  * TBD: this is going to be moved to firmware. Missing pdid/qpid check for now.
 160  */
 161 static int iwch_sgl2pbl_map(struct iwch_dev *rhp, struct ib_sge *sg_list,
 162                             u32 num_sgle, u32 * pbl_addr, u8 * page_size)
 163 {
 164         int i;
 165         struct iwch_mr *mhp;
 166         u32 offset;
 167         for (i = 0; i < num_sgle; i++) {
 168
 169                 mhp = get_mhp(rhp, (sg_list[i].lkey) >> 8);
 170                 if (!mhp) {
 171                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 172                         return -EIO;
 173                 }
 174                 if (!mhp->attr.state) {
 175                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 176                         return -EIO;
 177                 }
 178                 if (mhp->attr.zbva) {
 179                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 180                         return -EIO;
 181                 }
 182
 183                 if (sg_list[i].addr < mhp->attr.va_fbo) {
 184                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 185                         return -EINVAL;
 186                 }
 187                 if (sg_list[i].addr + ((u64) sg_list[i].length) <
 188                     sg_list[i].addr) {
 189                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 190                         return -EINVAL;
 191                 }
 192                 if (sg_list[i].addr + ((u64) sg_list[i].length) >
 193                     mhp->attr.va_fbo + ((u64) mhp->attr.len)) {
 194                         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 195                         return -EINVAL;
 196                 }
 197                 offset = sg_list[i].addr - mhp->attr.va_fbo;
 198                 offset += ((u32) mhp->attr.va_fbo) %
 199                           (1UL << (12 + mhp->attr.page_size));
 200                 pbl_addr[i] = ((mhp->attr.pbl_addr -
 201                                 rhp->rdev.rnic_info.pbl_base) >> 3) +
 202                               (offset >> (12 + mhp->attr.page_size));
 203                 page_size[i] = mhp->attr.page_size;
 204         }
 205         return 0;
 206 }
 207
 208 static int iwch_build_rdma_recv(struct iwch_dev *rhp, union t3_wr *wqe,
 209                                 struct ib_recv_wr *wr)
 210 {
 211         int i, err = 0;
 212         u32 pbl_addr[4];
 213         u8 page_size[4];
 214         if (wr->num_sge > T3_MAX_SGE)
 215                 return -EINVAL;
 216         err = iwch_sgl2pbl_map(rhp, wr->sg_list, wr->num_sge, pbl_addr,
 217                                page_size);
 218         if (err)
 219                 return err;
 220         wqe->recv.pagesz[0] = page_size[0];
 221         wqe->recv.pagesz[1] = page_size[1];
 222         wqe->recv.pagesz[2] = page_size[2];
 223         wqe->recv.pagesz[3] = page_size[3];
 224         wqe->recv.num_sgle = cpu_to_be32(wr->num_sge);
 225         for (i = 0; i < wr->num_sge; i++) {
 226                 wqe->recv.sgl[i].stag = cpu_to_be32(wr->sg_list[i].lkey);
 227                 wqe->recv.sgl[i].len = cpu_to_be32(wr->sg_list[i].length);
 228
 229                 /* to in the WQE == the offset into the page */
 230                 wqe->recv.sgl[i].to = cpu_to_be64(((u32) wr->sg_list[i].addr) %
 231                                 (1UL << (12 + page_size[i])));
 232
 233                 /* pbl_addr is the adapters address in the PBL */
 234                 wqe->recv.pbl_addr[i] = cpu_to_be32(pbl_addr[i]);
 235         }
 236         for (; i < T3_MAX_SGE; i++) {
 237                 wqe->recv.sgl[i].stag = 0;
 238                 wqe->recv.sgl[i].len = 0;
 239                 wqe->recv.sgl[i].to = 0;
 240                 wqe->recv.pbl_addr[i] = 0;
 241         }
 242         return 0;
 243 }
 244
 245 int iwch_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 246                       struct ib_send_wr **bad_wr)
 247 {
 248         int err = 0;
 249         u8 t3_wr_flit_cnt;
 250         enum t3_wr_opcode t3_wr_opcode = 0;
 251         enum t3_wr_flags t3_wr_flags;
 252         struct iwch_qp *qhp;
 253         u32 idx;
 254         union t3_wr *wqe;
 255         u32 num_wrs;
 256         unsigned long flag;
 257         struct t3_swsq *sqp;
 258
 259         qhp = to_iwch_qp(ibqp);
 260         spin_lock_irqsave(&qhp->lock, flag);
 261         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 262                 spin_unlock_irqrestore(&qhp->lock, flag);
 263                 return -EINVAL;
 264         }
 265         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
 266                   qhp->wq.sq_size_log2);
 267         if (num_wrs <= 0) {
 268                 spin_unlock_irqrestore(&qhp->lock, flag);
 269                 return -ENOMEM;
 270         }
 271         while (wr) {
 272                 if (num_wrs == 0) {
 273                         err = -ENOMEM;
 274                         *bad_wr = wr;
 275                         break;
 276                 }
 277                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 278                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
 279                 t3_wr_flags = 0;
 280                 if (wr->send_flags & IB_SEND_SOLICITED)
 281                         t3_wr_flags |= T3_SOLICITED_EVENT_FLAG;
 282                 if (wr->send_flags & IB_SEND_FENCE)
 283                         t3_wr_flags |= T3_READ_FENCE_FLAG;
 284                 if (wr->send_flags & IB_SEND_SIGNALED)
 285                         t3_wr_flags |= T3_COMPLETION_FLAG;
 286                 sqp = qhp->wq.sq +
 287                       Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
 288                 switch (wr->opcode) {
 289                 case IB_WR_SEND:
 290                 case IB_WR_SEND_WITH_IMM:
 291                         t3_wr_opcode = T3_WR_SEND;
 292                         err = iwch_build_rdma_send(wqe, wr, &t3_wr_flit_cnt);
 293                         break;
 294                 case IB_WR_RDMA_WRITE:
 295                 case IB_WR_RDMA_WRITE_WITH_IMM:
 296                         t3_wr_opcode = T3_WR_WRITE;
 297                         err = iwch_build_rdma_write(wqe, wr, &t3_wr_flit_cnt);
 298                         break;
 299                 case IB_WR_RDMA_READ:
 300                         t3_wr_opcode = T3_WR_READ;
 301                         t3_wr_flags = 0; /* T3 reads are always signaled */
 302                         err = iwch_build_rdma_read(wqe, wr, &t3_wr_flit_cnt);
 303                         if (err)
 304                                 break;
 305                         sqp->read_len = wqe->read.local_len;
 306                         if (!qhp->wq.oldest_read)
 307                                 qhp->wq.oldest_read = sqp;
 308                         break;
 309                 default:
 310                         PDBG("%s post of type=%d TBD!\n", __FUNCTION__,
 311                              wr->opcode);
 312                         err = -EINVAL;
 313                 }
 314                 if (err) {
 315                         *bad_wr = wr;
 316                         break;
 317                 }
 318                 wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 319                 sqp->wr_id = wr->wr_id;
 320                 sqp->opcode = wr2opcode(t3_wr_opcode);
 321                 sqp->sq_wptr = qhp->wq.sq_wptr;
 322                 sqp->complete = 0;
 323                 sqp->signaled = (wr->send_flags & IB_SEND_SIGNALED);
 324
 325                 build_fw_riwrh((void *) wqe, t3_wr_opcode, t3_wr_flags,
 326                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 327                                0, t3_wr_flit_cnt);
 328                 PDBG("%s cookie 0x%llx wq idx 0x%x swsq idx %ld opcode %d\n",
 329                      __FUNCTION__, (unsigned long long) wr->wr_id, idx,
 330                      Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2),
 331                      sqp->opcode);
 332                 wr = wr->next;
 333                 num_wrs--;
 334                 ++(qhp->wq.wptr);
 335                 ++(qhp->wq.sq_wptr);
 336         }
 337         spin_unlock_irqrestore(&qhp->lock, flag);
 338         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 339         return err;
 340 }
 341
 342 int iwch_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 343                       struct ib_recv_wr **bad_wr)
 344 {
 345         int err = 0;
 346         struct iwch_qp *qhp;
 347         u32 idx;
 348         union t3_wr *wqe;
 349         u32 num_wrs;
 350         unsigned long flag;
 351
 352         qhp = to_iwch_qp(ibqp);
 353         spin_lock_irqsave(&qhp->lock, flag);
 354         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 355                 spin_unlock_irqrestore(&qhp->lock, flag);
 356                 return -EINVAL;
 357         }
 358         num_wrs = Q_FREECNT(qhp->wq.rq_rptr, qhp->wq.rq_wptr,
 359                             qhp->wq.rq_size_log2) - 1;
 360         if (!wr) {
 361                 spin_unlock_irqrestore(&qhp->lock, flag);
 362                 return -EINVAL;
 363         }
 364         while (wr) {
 365                 idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 366                 wqe = (union t3_wr *) (qhp->wq.queue + idx);
 367                 if (num_wrs)
 368                         err = iwch_build_rdma_recv(qhp->rhp, wqe, wr);
 369                 else
 370                         err = -ENOMEM;
 371                 if (err) {
 372                         *bad_wr = wr;
 373                         break;
 374                 }
 375                 qhp->wq.rq[Q_PTR2IDX(qhp->wq.rq_wptr, qhp->wq.rq_size_log2)] =
 376                         wr->wr_id;
 377                 build_fw_riwrh((void *) wqe, T3_WR_RCV, T3_COMPLETION_FLAG,
 378                                Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2),
 379                                0, sizeof(struct t3_receive_wr) >> 3);
 380                 PDBG("%s cookie 0x%llx idx 0x%x rq_wptr 0x%x rw_rptr 0x%x "
 381                      "wqe %p \n", __FUNCTION__, (unsigned long long) wr->wr_id,
 382                      idx, qhp->wq.rq_wptr, qhp->wq.rq_rptr, wqe);
 383                 ++(qhp->wq.rq_wptr);
 384                 ++(qhp->wq.wptr);
 385                 wr = wr->next;
 386                 num_wrs--;
 387         }
 388         spin_unlock_irqrestore(&qhp->lock, flag);
 389         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 390         return err;
 391 }
 392
 393 int iwch_bind_mw(struct ib_qp *qp,
 394                              struct ib_mw *mw,
 395                              struct ib_mw_bind *mw_bind)
 396 {
 397         struct iwch_dev *rhp;
 398         struct iwch_mw *mhp;
 399         struct iwch_qp *qhp;
 400         union t3_wr *wqe;
 401         u32 pbl_addr;
 402         u8 page_size;
 403         u32 num_wrs;
 404         unsigned long flag;
 405         struct ib_sge sgl;
 406         int err=0;
 407         enum t3_wr_flags t3_wr_flags;
 408         u32 idx;
 409         struct t3_swsq *sqp;
 410
 411         qhp = to_iwch_qp(qp);
 412         mhp = to_iwch_mw(mw);
 413         rhp = qhp->rhp;
 414
 415         spin_lock_irqsave(&qhp->lock, flag);
 416         if (qhp->attr.state > IWCH_QP_STATE_RTS) {
 417                 spin_unlock_irqrestore(&qhp->lock, flag);
 418                 return -EINVAL;
 419         }
 420         num_wrs = Q_FREECNT(qhp->wq.sq_rptr, qhp->wq.sq_wptr,
 421                             qhp->wq.sq_size_log2);
 422         if ((num_wrs) <= 0) {
 423                 spin_unlock_irqrestore(&qhp->lock, flag);
 424                 return -ENOMEM;
 425         }
 426         idx = Q_PTR2IDX(qhp->wq.wptr, qhp->wq.size_log2);
 427         PDBG("%s: idx 0x%0x, mw 0x%p, mw_bind 0x%p\n", __FUNCTION__, idx,
 428              mw, mw_bind);
 429         wqe = (union t3_wr *) (qhp->wq.queue + idx);
 430
 431         t3_wr_flags = 0;
 432         if (mw_bind->send_flags & IB_SEND_SIGNALED)
 433                 t3_wr_flags = T3_COMPLETION_FLAG;
 434
 435         sgl.addr = mw_bind->addr;
 436         sgl.lkey = mw_bind->mr->lkey;
 437         sgl.length = mw_bind->length;
 438         wqe->bind.reserved = 0;
 439         wqe->bind.type = T3_VA_BASED_TO;
 440
 441         /* TBD: check perms */
 442         wqe->bind.perms = iwch_convert_access(mw_bind->mw_access_flags);
 443         wqe->bind.mr_stag = cpu_to_be32(mw_bind->mr->lkey);
 444         wqe->bind.mw_stag = cpu_to_be32(mw->rkey);
 445         wqe->bind.mw_len = cpu_to_be32(mw_bind->length);
 446         wqe->bind.mw_va = cpu_to_be64(mw_bind->addr);
 447         err = iwch_sgl2pbl_map(rhp, &sgl, 1, &pbl_addr, &page_size);
 448         if (err) {
 449                 spin_unlock_irqrestore(&qhp->lock, flag);
 450                 return err;
 451         }
 452         wqe->send.wrid.id0.hi = qhp->wq.sq_wptr;
 453         sqp = qhp->wq.sq + Q_PTR2IDX(qhp->wq.sq_wptr, qhp->wq.sq_size_log2);
 454         sqp->wr_id = mw_bind->wr_id;
 455         sqp->opcode = T3_BIND_MW;
 456         sqp->sq_wptr = qhp->wq.sq_wptr;
 457         sqp->complete = 0;
 458         sqp->signaled = (mw_bind->send_flags & IB_SEND_SIGNALED);
 459         wqe->bind.mr_pbl_addr = cpu_to_be32(pbl_addr);
 460         wqe->bind.mr_pagesz = page_size;
 461         wqe->flit[T3_SQ_COOKIE_FLIT] = mw_bind->wr_id;
 462         build_fw_riwrh((void *)wqe, T3_WR_BIND, t3_wr_flags,
 463                        Q_GENBIT(qhp->wq.wptr, qhp->wq.size_log2), 0,
 464                                 sizeof(struct t3_bind_mw_wr) >> 3);
 465         ++(qhp->wq.wptr);
 466         ++(qhp->wq.sq_wptr);
 467         spin_unlock_irqrestore(&qhp->lock, flag);
 468
 469         ring_doorbell(qhp->wq.doorbell, qhp->wq.qpid);
 470
 471         return err;
 472 }
 473
 474 static void build_term_codes(int t3err, u8 *layer_type, u8 *ecode, int tagged)
 475 {
 476         switch (t3err) {
 477         case TPT_ERR_STAG:
 478                 if (tagged == 1) {
 479                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 480                         *ecode = DDPT_INV_STAG;
 481                 } else if (tagged == 2) {
 482                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 483                         *ecode = RDMAP_INV_STAG;
 484                 }
 485                 break;
 486         case TPT_ERR_PDID:
 487         case TPT_ERR_QPID:
 488         case TPT_ERR_ACCESS:
 489                 if (tagged == 1) {
 490                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 491                         *ecode = DDPT_STAG_NOT_ASSOC;
 492                 } else if (tagged == 2) {
 493                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 494                         *ecode = RDMAP_STAG_NOT_ASSOC;
 495                 }
 496                 break;
 497         case TPT_ERR_WRAP:
 498                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 499                 *ecode = RDMAP_TO_WRAP;
 500                 break;
 501         case TPT_ERR_BOUND:
 502                 if (tagged == 1) {
 503                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 504                         *ecode = DDPT_BASE_BOUNDS;
 505                 } else if (tagged == 2) {
 506                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 507                         *ecode = RDMAP_BASE_BOUNDS;
 508                 } else {
 509                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 510                         *ecode = DDPU_MSG_TOOBIG;
 511                 }
 512                 break;
 513         case TPT_ERR_INVALIDATE_SHARED_MR:
 514         case TPT_ERR_INVALIDATE_MR_WITH_MW_BOUND:
 515                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 516                 *ecode = RDMAP_CANT_INV_STAG;
 517                 break;
 518         case TPT_ERR_ECC:
 519         case TPT_ERR_ECC_PSTAG:
 520         case TPT_ERR_INTERNAL_ERR:
 521                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
 522                 *ecode = 0;
 523                 break;
 524         case TPT_ERR_OUT_OF_RQE:
 525                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 526                 *ecode = DDPU_INV_MSN_NOBUF;
 527                 break;
 528         case TPT_ERR_PBL_ADDR_BOUND:
 529                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 530                 *ecode = DDPT_BASE_BOUNDS;
 531                 break;
 532         case TPT_ERR_CRC:
 533                 *layer_type = LAYER_MPA|DDP_LLP;
 534                 *ecode = MPA_CRC_ERR;
 535                 break;
 536         case TPT_ERR_MARKER:
 537                 *layer_type = LAYER_MPA|DDP_LLP;
 538                 *ecode = MPA_MARKER_ERR;
 539                 break;
 540         case TPT_ERR_PDU_LEN_ERR:
 541                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 542                 *ecode = DDPU_MSG_TOOBIG;
 543                 break;
 544         case TPT_ERR_DDP_VERSION:
 545                 if (tagged) {
 546                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 547                         *ecode = DDPT_INV_VERS;
 548                 } else {
 549                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 550                         *ecode = DDPU_INV_VERS;
 551                 }
 552                 break;
 553         case TPT_ERR_RDMA_VERSION:
 554                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 555                 *ecode = RDMAP_INV_VERS;
 556                 break;
 557         case TPT_ERR_OPCODE:
 558                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 559                 *ecode = RDMAP_INV_OPCODE;
 560                 break;
 561         case TPT_ERR_DDP_QUEUE_NUM:
 562                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 563                 *ecode = DDPU_INV_QN;
 564                 break;
 565         case TPT_ERR_MSN:
 566         case TPT_ERR_MSN_GAP:
 567         case TPT_ERR_MSN_RANGE:
 568         case TPT_ERR_IRD_OVERFLOW:
 569                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 570                 *ecode = DDPU_INV_MSN_RANGE;
 571                 break;
 572         case TPT_ERR_TBIT:
 573                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
 574                 *ecode = 0;
 575                 break;
 576         case TPT_ERR_MO:
 577                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 578                 *ecode = DDPU_INV_MO;
 579                 break;
 580         default:
 581                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 582                 *ecode = 0;
 583                 break;
 584         }
 585 }
 586
 587 /*
 588  * This posts a TERMINATE with layer=RDMA, type=catastrophic.
 589  */
 590 int iwch_post_terminate(struct iwch_qp *qhp, struct respQ_msg_t *rsp_msg)
 591 {
 592         union t3_wr *wqe;
 593         struct terminate_message *term;
 594         int status;
 595         int tagged = 0;
 596         struct sk_buff *skb;
 597
 598         PDBG("%s %d\n", __FUNCTION__, __LINE__);
 599         skb = alloc_skb(40, GFP_ATOMIC);
 600         if (!skb) {
 601                 printk(KERN_ERR "%s cannot send TERMINATE!\n", __FUNCTION__);
 602                 return -ENOMEM;
 603         }
 604         wqe = (union t3_wr *)skb_put(skb, 40);
 605         memset(wqe, 0, 40);
 606         wqe->send.rdmaop = T3_TERMINATE;
 607
 608         /* immediate data length */
 609         wqe->send.plen = htonl(4);
 610
 611         /* immediate data starts here. */
 612         term = (struct terminate_message *)wqe->send.sgl;
 613         if (rsp_msg) {
 614                 status = CQE_STATUS(rsp_msg->cqe);
 615                 if (CQE_OPCODE(rsp_msg->cqe) == T3_RDMA_WRITE)
 616                         tagged = 1;
 617                 if ((CQE_OPCODE(rsp_msg->cqe) == T3_READ_REQ) ||
 618                     (CQE_OPCODE(rsp_msg->cqe) == T3_READ_RESP))
 619                         tagged = 2;
 620         } else {
 621                 status = TPT_ERR_INTERNAL_ERR;
 622         }
 623         build_term_codes(status, &term->layer_etype, &term->ecode, tagged);
 624         build_fw_riwrh((void *)wqe, T3_WR_SEND,
 625                        T3_COMPLETION_FLAG | T3_NOTIFY_FLAG, 1,
 626                        qhp->ep->hwtid, 5);
 627         skb->priority = CPL_PRIORITY_DATA;
 628         return cxgb3_ofld_send(qhp->rhp->rdev.t3cdev_p, skb);
 629 }
 630
 631 /*
 632  * Assumes qhp lock is held.
 633  */
 634 static void __flush_qp(struct iwch_qp *qhp, unsigned long *flag)
 635 {
 636         struct iwch_cq *rchp, *schp;
 637         int count;
 638
 639         rchp = get_chp(qhp->rhp, qhp->attr.rcq);
 640         schp = get_chp(qhp->rhp, qhp->attr.scq);
 641
 642         PDBG("%s qhp %p rchp %p schp %p\n", __FUNCTION__, qhp, rchp, schp);
 643         /* take a ref on the qhp since we must release the lock */
 644         atomic_inc(&qhp->refcnt);
 645         spin_unlock_irqrestore(&qhp->lock, *flag);
 646
 647         /* locking heirarchy: cq lock first, then qp lock. */
 648         spin_lock_irqsave(&rchp->lock, *flag);
 649         spin_lock(&qhp->lock);
 650         cxio_flush_hw_cq(&rchp->cq);
 651         cxio_count_rcqes(&rchp->cq, &qhp->wq, &count);
 652         cxio_flush_rq(&qhp->wq, &rchp->cq, count);
 653         spin_unlock(&qhp->lock);
 654         spin_unlock_irqrestore(&rchp->lock, *flag);
 655
 656         /* locking heirarchy: cq lock first, then qp lock. */
 657         spin_lock_irqsave(&schp->lock, *flag);
 658         spin_lock(&qhp->lock);
 659         cxio_flush_hw_cq(&schp->cq);
 660         cxio_count_scqes(&schp->cq, &qhp->wq, &count);
 661         cxio_flush_sq(&qhp->wq, &schp->cq, count);
 662         spin_unlock(&qhp->lock);
 663         spin_unlock_irqrestore(&schp->lock, *flag);
 664
 665         /* deref */
 666         if (atomic_dec_and_test(&qhp->refcnt))
 667                 wake_up(&qhp->wait);
 668
 669         spin_lock_irqsave(&qhp->lock, *flag);
 670 }
 671
 672 static void flush_qp(struct iwch_qp *qhp, unsigned long *flag)
 673 {
 674         if (t3b_device(qhp->rhp))
 675                 cxio_set_wq_in_error(&qhp->wq);
 676         else
 677                 __flush_qp(qhp, flag);
 678 }
 679
 680
 681 /*
 682  * Return non zero if at least one RECV was pre-posted.
 683  */
 684 static int rqes_posted(struct iwch_qp *qhp)
 685 {
 686         return fw_riwrh_opcode((struct fw_riwrh *)qhp->wq.queue) == T3_WR_RCV;
 687 }
 688
 689 static int rdma_init(struct iwch_dev *rhp, struct iwch_qp *qhp,
 690                                 enum iwch_qp_attr_mask mask,
 691                                 struct iwch_qp_attributes *attrs)
 692 {
 693         struct t3_rdma_init_attr init_attr;
 694         int ret;
 695
 696         init_attr.tid = qhp->ep->hwtid;
 697         init_attr.qpid = qhp->wq.qpid;
 698         init_attr.pdid = qhp->attr.pd;
 699         init_attr.scqid = qhp->attr.scq;
 700         init_attr.rcqid = qhp->attr.rcq;
 701         init_attr.rq_addr = qhp->wq.rq_addr;
 702         init_attr.rq_size = 1 << qhp->wq.rq_size_log2;
 703         init_attr.mpaattrs = uP_RI_MPA_IETF_ENABLE |
 704                 qhp->attr.mpa_attr.recv_marker_enabled |
 705                 (qhp->attr.mpa_attr.xmit_marker_enabled << 1) |
 706                 (qhp->attr.mpa_attr.crc_enabled << 2);
 707
 708         /*
 709          * XXX - The IWCM doesn't quite handle getting these
 710          * attrs set before going into RTS.  For now, just turn
 711          * them on always...
 712          */
 713 #if 0
 714         init_attr.qpcaps = qhp->attr.enableRdmaRead |
 715                 (qhp->attr.enableRdmaWrite << 1) |
 716                 (qhp->attr.enableBind << 2) |
 717                 (qhp->attr.enable_stag0_fastreg << 3) |
 718                 (qhp->attr.enable_stag0_fastreg << 4);
 719 #else
 720         init_attr.qpcaps = 0x1f;
 721 #endif
 722         init_attr.tcp_emss = qhp->ep->emss;
 723         init_attr.ord = qhp->attr.max_ord;
 724         init_attr.ird = qhp->attr.max_ird;
 725         init_attr.qp_dma_addr = qhp->wq.dma_addr;
 726         init_attr.qp_dma_size = (1UL << qhp->wq.size_log2);
 727         init_attr.flags = rqes_posted(qhp) ? RECVS_POSTED : 0;
 728         PDBG("%s init_attr.rq_addr 0x%x init_attr.rq_size = %d "
 729              "flags 0x%x qpcaps 0x%x\n", __FUNCTION__,
 730              init_attr.rq_addr, init_attr.rq_size,
 731              init_attr.flags, init_attr.qpcaps);
 732         ret = cxio_rdma_init(&rhp->rdev, &init_attr);
 733         PDBG("%s ret %d\n", __FUNCTION__, ret);
 734         return ret;
 735 }
 736
 737 int iwch_modify_qp(struct iwch_dev *rhp, struct iwch_qp *qhp,
 738                                 enum iwch_qp_attr_mask mask,
 739                                 struct iwch_qp_attributes *attrs,
 740                                 int internal)
 741 {
 742         int ret = 0;
 743         struct iwch_qp_attributes newattr = qhp->attr;
 744         unsigned long flag;
 745         int disconnect = 0;
 746         int terminate = 0;
 747         int abort = 0;
 748         int free = 0;
 749         struct iwch_ep *ep = NULL;
 750
 751         PDBG("%s qhp %p qpid 0x%x ep %p state %d -> %d\n", __FUNCTION__,
 752              qhp, qhp->wq.qpid, qhp->ep, qhp->attr.state,
 753              (mask & IWCH_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
 754
 755         spin_lock_irqsave(&qhp->lock, flag);
 756
 757         /* Process attr changes if in IDLE */
 758         if (mask & IWCH_QP_ATTR_VALID_MODIFY) {
 759                 if (qhp->attr.state != IWCH_QP_STATE_IDLE) {
 760                         ret = -EIO;
 761                         goto out;
 762                 }
 763                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_READ)
 764                         newattr.enable_rdma_read = attrs->enable_rdma_read;
 765                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_WRITE)
 766                         newattr.enable_rdma_write = attrs->enable_rdma_write;
 767                 if (mask & IWCH_QP_ATTR_ENABLE_RDMA_BIND)
 768                         newattr.enable_bind = attrs->enable_bind;
 769                 if (mask & IWCH_QP_ATTR_MAX_ORD) {
 770                         if (attrs->max_ord >
 771                             rhp->attr.max_rdma_read_qp_depth) {
 772                                 ret = -EINVAL;
 773                                 goto out;
 774                         }
 775                         newattr.max_ord = attrs->max_ord;
 776                 }
 777                 if (mask & IWCH_QP_ATTR_MAX_IRD) {
 778                         if (attrs->max_ird >
 779                             rhp->attr.max_rdma_reads_per_qp) {
 780                                 ret = -EINVAL;
 781                                 goto out;
 782                         }
 783                         newattr.max_ird = attrs->max_ird;
 784                 }
 785                 qhp->attr = newattr;
 786         }
 787
 788         if (!(mask & IWCH_QP_ATTR_NEXT_STATE))
 789                 goto out;
 790         if (qhp->attr.state == attrs->next_state)
 791                 goto out;
 792
 793         switch (qhp->attr.state) {
 794         case IWCH_QP_STATE_IDLE:
 795                 switch (attrs->next_state) {
 796                 case IWCH_QP_STATE_RTS:
 797                         if (!(mask & IWCH_QP_ATTR_LLP_STREAM_HANDLE)) {
 798                                 ret = -EINVAL;
 799                                 goto out;
 800                         }
 801                         if (!(mask & IWCH_QP_ATTR_MPA_ATTR)) {
 802                                 ret = -EINVAL;
 803                                 goto out;
 804                         }
 805                         qhp->attr.mpa_attr = attrs->mpa_attr;
 806                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
 807                         qhp->ep = qhp->attr.llp_stream_handle;
 808                         qhp->attr.state = IWCH_QP_STATE_RTS;
 809
 810                         /*
 811                          * Ref the endpoint here and deref when we
 812                          * disassociate the endpoint from the QP.  This
 813                          * happens in CLOSING->IDLE transition or *->ERROR
 814                          * transition.
 815                          */
 816                         get_ep(&qhp->ep->com);
 817                         spin_unlock_irqrestore(&qhp->lock, flag);
 818                         ret = rdma_init(rhp, qhp, mask, attrs);
 819                         spin_lock_irqsave(&qhp->lock, flag);
 820                         if (ret)
 821                                 goto err;
 822                         break;
 823                 case IWCH_QP_STATE_ERROR:
 824                         qhp->attr.state = IWCH_QP_STATE_ERROR;
 825                         flush_qp(qhp, &flag);
 826                         break;
 827                 default:
 828                         ret = -EINVAL;
 829                         goto out;
 830                 }
 831                 break;
 832         case IWCH_QP_STATE_RTS:
 833                 switch (attrs->next_state) {
 834                 case IWCH_QP_STATE_CLOSING:
 835                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
 836                         qhp->attr.state = IWCH_QP_STATE_CLOSING;
 837                         if (!internal) {
 838                                 abort=0;
 839                                 disconnect = 1;
 840                                 ep = qhp->ep;
 841                         }
 842                         break;
 843                 case IWCH_QP_STATE_TERMINATE:
 844                         qhp->attr.state = IWCH_QP_STATE_TERMINATE;
 845                         if (t3b_device(qhp->rhp))
 846                                 cxio_set_wq_in_error(&qhp->wq);
 847                         if (!internal)
 848                                 terminate = 1;
 849                         break;
 850                 case IWCH_QP_STATE_ERROR:
 851                         qhp->attr.state = IWCH_QP_STATE_ERROR;
 852                         if (!internal) {
 853                                 abort=1;
 854                                 disconnect = 1;
 855                                 ep = qhp->ep;
 856                         }
 857                         goto err;
 858                         break;
 859                 default:
 860                         ret = -EINVAL;
 861                         goto out;
 862                 }
 863                 break;
 864         case IWCH_QP_STATE_CLOSING:
 865                 if (!internal) {
 866                         ret = -EINVAL;
 867                         goto out;
 868                 }
 869                 switch (attrs->next_state) {
 870                         case IWCH_QP_STATE_IDLE:
 871                                 qhp->attr.state = IWCH_QP_STATE_IDLE;
 872                                 qhp->attr.llp_stream_handle = NULL;
 873                                 put_ep(&qhp->ep->com);
 874                                 qhp->ep = NULL;
 875                                 wake_up(&qhp->wait);
 876                                 break;
 877                         case IWCH_QP_STATE_ERROR:
 878                                 goto err;
 879                         default:
 880                                 ret = -EINVAL;
 881                                 goto err;
 882                 }
 883                 break;
 884         case IWCH_QP_STATE_ERROR:
 885                 if (attrs->next_state != IWCH_QP_STATE_IDLE) {
 886                         ret = -EINVAL;
 887                         goto out;
 888                 }
 889
 890                 if (!Q_EMPTY(qhp->wq.sq_rptr, qhp->wq.sq_wptr) ||
 891                     !Q_EMPTY(qhp->wq.rq_rptr, qhp->wq.rq_wptr)) {
 892                         ret = -EINVAL;
 893                         goto out;
 894                 }
 895                 qhp->attr.state = IWCH_QP_STATE_IDLE;
 896                 memset(&qhp->attr, 0, sizeof(qhp->attr));
 897                 break;
 898         case IWCH_QP_STATE_TERMINATE:
 899                 if (!internal) {
 900                         ret = -EINVAL;
 901                         goto out;
 902                 }
 903                 goto err;
 904                 break;
 905         default:
 906                 printk(KERN_ERR "%s in a bad state %d\n",
 907                        __FUNCTION__, qhp->attr.state);
 908                 ret = -EINVAL;
 909                 goto err;
 910                 break;
 911         }
 912         goto out;
 913 err:
 914         PDBG("%s disassociating ep %p qpid 0x%x\n", __FUNCTION__, qhp->ep,
 915              qhp->wq.qpid);
 916
 917         /* disassociate the LLP connection */
 918         qhp->attr.llp_stream_handle = NULL;
 919         ep = qhp->ep;
 920         qhp->ep = NULL;
 921         qhp->attr.state = IWCH_QP_STATE_ERROR;
 922         free=1;
 923         wake_up(&qhp->wait);
 924         BUG_ON(!ep);
 925         flush_qp(qhp, &flag);
 926 out:
 927         spin_unlock_irqrestore(&qhp->lock, flag);
 928
 929         if (terminate)
 930                 iwch_post_terminate(qhp, NULL);
 931
 932         /*
 933          * If disconnect is 1, then we need to initiate a disconnect
 934          * on the EP.  This can be a normal close (RTS->CLOSING) or
 935          * an abnormal close (RTS/CLOSING->ERROR).
 936          */
 937         if (disconnect)
 938                 iwch_ep_disconnect(ep, abort, GFP_KERNEL);
 939
 940         /*
 941          * If free is 1, then we've disassociated the EP from the QP
 942          * and we need to dereference the EP.
 943          */
 944         if (free)
 945                 put_ep(&ep->com);
 946
 947         PDBG("%s exit state %d\n", __FUNCTION__, qhp->attr.state);
 948         return ret;
 949 }
 950
 951 static int quiesce_qp(struct iwch_qp *qhp)
 952 {
 953         spin_lock_irq(&qhp->lock);
 954         iwch_quiesce_tid(qhp->ep);
 955         qhp->flags |= QP_QUIESCED;
 956         spin_unlock_irq(&qhp->lock);
 957         return 0;
 958 }
 959
 960 static int resume_qp(struct iwch_qp *qhp)
 961 {
 962         spin_lock_irq(&qhp->lock);
 963         iwch_resume_tid(qhp->ep);
 964         qhp->flags &= ~QP_QUIESCED;
 965         spin_unlock_irq(&qhp->lock);
 966         return 0;
 967 }
 968
 969 int iwch_quiesce_qps(struct iwch_cq *chp)
 970 {
 971         int i;
 972         struct iwch_qp *qhp;
 973
 974         for (i=0; i < T3_MAX_NUM_QP; i++) {
 975                 qhp = get_qhp(chp->rhp, i);
 976                 if (!qhp)
 977                         continue;
 978                 if ((qhp->attr.rcq == chp->cq.cqid) && !qp_quiesced(qhp)) {
 979                         quiesce_qp(qhp);
 980                         continue;
 981                 }
 982                 if ((qhp->attr.scq == chp->cq.cqid) && !qp_quiesced(qhp))
 983                         quiesce_qp(qhp);
 984         }
 985         return 0;
 986 }
 987
 988 int iwch_resume_qps(struct iwch_cq *chp)
 989 {
 990         int i;
 991         struct iwch_qp *qhp;
 992
 993         for (i=0; i < T3_MAX_NUM_QP; i++) {
 994                 qhp = get_qhp(chp->rhp, i);
 995                 if (!qhp)
 996                         continue;
 997                 if ((qhp->attr.rcq == chp->cq.cqid) && qp_quiesced(qhp)) {
 998                         resume_qp(qhp);
 999                         continue;
1000                 }
1001                 if ((qhp->attr.scq == chp->cq.cqid) && qp_quiesced(qhp))
1002                         resume_qp(qhp);
1003         }
1004         return 0;
1005 }