drivers/infiniband/hw/cxgb4/qp.c

   1 /*
   2  * Copyright (c) 2009-2010 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
  33 #include <linux/module.h>
  34
  35 #include "iw_cxgb4.h"
  36
  37 static int db_delay_usecs = 1;
  38 module_param(db_delay_usecs, int, 0644);
  39 MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
  40
  41 static int ocqp_support = 1;
  42 module_param(ocqp_support, int, 0644);
  43 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
  44
  45 int db_fc_threshold = 1000;
  46 module_param(db_fc_threshold, int, 0644);
  47 MODULE_PARM_DESC(db_fc_threshold,
  48                  "QP count/threshold that triggers"
  49                  " automatic db flow control mode (default = 1000)");
  50
  51 int db_coalescing_threshold;
  52 module_param(db_coalescing_threshold, int, 0644);
  53 MODULE_PARM_DESC(db_coalescing_threshold,
  54                  "QP count/threshold that triggers"
  55                  " disabling db coalescing (default = 0)");
  56
  57 static int max_fr_immd = T4_MAX_FR_IMMD;
  58 module_param(max_fr_immd, int, 0644);
  59 MODULE_PARM_DESC(max_fr_immd, "fastreg threshold for using DSGL instead of immedate");
  60
  61 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
  62 {
  63         unsigned long flag;
  64         spin_lock_irqsave(&qhp->lock, flag);
  65         qhp->attr.state = state;
  66         spin_unlock_irqrestore(&qhp->lock, flag);
  67 }
  68
  69 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
  70 {
  71         c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
  72 }
  73
  74 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
  75 {
  76         dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
  77                           pci_unmap_addr(sq, mapping));
  78 }
  79
  80 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
  81 {
  82         if (t4_sq_onchip(sq))
  83                 dealloc_oc_sq(rdev, sq);
  84         else
  85                 dealloc_host_sq(rdev, sq);
  86 }
  87
  88 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
  89 {
  90         if (!ocqp_support || !ocqp_supported(&rdev->lldi))
  91                 return -ENOSYS;
  92         sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
  93         if (!sq->dma_addr)
  94                 return -ENOMEM;
  95         sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
  96                         rdev->lldi.vr->ocq.start;
  97         sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
  98                                             rdev->lldi.vr->ocq.start);
  99         sq->flags |= T4_SQ_ONCHIP;
 100         return 0;
 101 }
 102
 103 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
 104 {
 105         sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
 106                                        &(sq->dma_addr), GFP_KERNEL);
 107         if (!sq->queue)
 108                 return -ENOMEM;
 109         sq->phys_addr = virt_to_phys(sq->queue);
 110         pci_unmap_addr_set(sq, mapping, sq->dma_addr);
 111         return 0;
 112 }
 113
 114 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
 115 {
 116         int ret = -ENOSYS;
 117         if (user)
 118                 ret = alloc_oc_sq(rdev, sq);
 119         if (ret)
 120                 ret = alloc_host_sq(rdev, sq);
 121         return ret;
 122 }
 123
 124 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
 125                       struct c4iw_dev_ucontext *uctx)
 126 {
 127         /*
 128          * uP clears EQ contexts when the connection exits rdma mode,
 129          * so no need to post a RESET WR for these EQs.
 130          */
 131         dma_free_coherent(&(rdev->lldi.pdev->dev),
 132                           wq->rq.memsize, wq->rq.queue,
 133                           dma_unmap_addr(&wq->rq, mapping));
 134         dealloc_sq(rdev, &wq->sq);
 135         c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
 136         kfree(wq->rq.sw_rq);
 137         kfree(wq->sq.sw_sq);
 138         c4iw_put_qpid(rdev, wq->rq.qid, uctx);
 139         c4iw_put_qpid(rdev, wq->sq.qid, uctx);
 140         return 0;
 141 }
 142
 143 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
 144                      struct t4_cq *rcq, struct t4_cq *scq,
 145                      struct c4iw_dev_ucontext *uctx)
 146 {
 147         int user = (uctx != &rdev->uctx);
 148         struct fw_ri_res_wr *res_wr;
 149         struct fw_ri_res *res;
 150         int wr_len;
 151         struct c4iw_wr_wait wr_wait;
 152         struct sk_buff *skb;
 153         int ret = 0;
 154         int eqsize;
 155
 156         wq->sq.qid = c4iw_get_qpid(rdev, uctx);
 157         if (!wq->sq.qid)
 158                 return -ENOMEM;
 159
 160         wq->rq.qid = c4iw_get_qpid(rdev, uctx);
 161         if (!wq->rq.qid) {
 162                 ret = -ENOMEM;
 163                 goto free_sq_qid;
 164         }
 165
 166         if (!user) {
 167                 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
 168                                  GFP_KERNEL);
 169                 if (!wq->sq.sw_sq) {
 170                         ret = -ENOMEM;
 171                         goto free_rq_qid;
 172                 }
 173
 174                 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
 175                                  GFP_KERNEL);
 176                 if (!wq->rq.sw_rq) {
 177                         ret = -ENOMEM;
 178                         goto free_sw_sq;
 179                 }
 180         }
 181
 182         /*
 183          * RQT must be a power of 2.
 184          */
 185         wq->rq.rqt_size = roundup_pow_of_two(wq->rq.size);
 186         wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
 187         if (!wq->rq.rqt_hwaddr) {
 188                 ret = -ENOMEM;
 189                 goto free_sw_rq;
 190         }
 191
 192         ret = alloc_sq(rdev, &wq->sq, user);
 193         if (ret)
 194                 goto free_hwaddr;
 195         memset(wq->sq.queue, 0, wq->sq.memsize);
 196         dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
 197
 198         wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
 199                                           wq->rq.memsize, &(wq->rq.dma_addr),
 200                                           GFP_KERNEL);
 201         if (!wq->rq.queue) {
 202                 ret = -ENOMEM;
 203                 goto free_sq;
 204         }
 205         PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
 206                 __func__, wq->sq.queue,
 207                 (unsigned long long)virt_to_phys(wq->sq.queue),
 208                 wq->rq.queue,
 209                 (unsigned long long)virt_to_phys(wq->rq.queue));
 210         memset(wq->rq.queue, 0, wq->rq.memsize);
 211         dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
 212
 213         wq->db = rdev->lldi.db_reg;
 214         wq->gts = rdev->lldi.gts_reg;
 215         if (user) {
 216                 wq->sq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
 217                                         (wq->sq.qid << rdev->qpshift);
 218                 wq->sq.udb &= PAGE_MASK;
 219                 wq->rq.udb = (u64)pci_resource_start(rdev->lldi.pdev, 2) +
 220                                         (wq->rq.qid << rdev->qpshift);
 221                 wq->rq.udb &= PAGE_MASK;
 222         }
 223         wq->rdev = rdev;
 224         wq->rq.msn = 1;
 225
 226         /* build fw_ri_res_wr */
 227         wr_len = sizeof *res_wr + 2 * sizeof *res;
 228
 229         skb = alloc_skb(wr_len, GFP_KERNEL);
 230         if (!skb) {
 231                 ret = -ENOMEM;
 232                 goto free_dma;
 233         }
 234         set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
 235
 236         res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
 237         memset(res_wr, 0, wr_len);
 238         res_wr->op_nres = cpu_to_be32(
 239                         FW_WR_OP(FW_RI_RES_WR) |
 240                         V_FW_RI_RES_WR_NRES(2) |
 241                         FW_WR_COMPL(1));
 242         res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
 243         res_wr->cookie = (unsigned long) &wr_wait;
 244         res = res_wr->res;
 245         res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
 246         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 247
 248         /*
 249          * eqsize is the number of 64B entries plus the status page size.
 250          */
 251         eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
 252
 253         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 254                 V_FW_RI_RES_WR_HOSTFCMODE(0) |  /* no host cidx updates */
 255                 V_FW_RI_RES_WR_CPRIO(0) |       /* don't keep in chip cache */
 256                 V_FW_RI_RES_WR_PCIECHN(0) |     /* set by uP at ri_init time */
 257                 (t4_sq_onchip(&wq->sq) ? F_FW_RI_RES_WR_ONCHIP : 0) |
 258                 V_FW_RI_RES_WR_IQID(scq->cqid));
 259         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 260                 V_FW_RI_RES_WR_DCAEN(0) |
 261                 V_FW_RI_RES_WR_DCACPU(0) |
 262                 V_FW_RI_RES_WR_FBMIN(2) |
 263                 V_FW_RI_RES_WR_FBMAX(2) |
 264                 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
 265                 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
 266                 V_FW_RI_RES_WR_EQSIZE(eqsize));
 267         res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
 268         res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
 269         res++;
 270         res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
 271         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 272
 273         /*
 274          * eqsize is the number of 64B entries plus the status page size.
 275          */
 276         eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + T4_EQ_STATUS_ENTRIES;
 277         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 278                 V_FW_RI_RES_WR_HOSTFCMODE(0) |  /* no host cidx updates */
 279                 V_FW_RI_RES_WR_CPRIO(0) |       /* don't keep in chip cache */
 280                 V_FW_RI_RES_WR_PCIECHN(0) |     /* set by uP at ri_init time */
 281                 V_FW_RI_RES_WR_IQID(rcq->cqid));
 282         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 283                 V_FW_RI_RES_WR_DCAEN(0) |
 284                 V_FW_RI_RES_WR_DCACPU(0) |
 285                 V_FW_RI_RES_WR_FBMIN(2) |
 286                 V_FW_RI_RES_WR_FBMAX(2) |
 287                 V_FW_RI_RES_WR_CIDXFTHRESHO(0) |
 288                 V_FW_RI_RES_WR_CIDXFTHRESH(0) |
 289                 V_FW_RI_RES_WR_EQSIZE(eqsize));
 290         res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
 291         res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
 292
 293         c4iw_init_wr_wait(&wr_wait);
 294
 295         ret = c4iw_ofld_send(rdev, skb);
 296         if (ret)
 297                 goto free_dma;
 298         ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
 299         if (ret)
 300                 goto free_dma;
 301
 302         PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx\n",
 303              __func__, wq->sq.qid, wq->rq.qid, wq->db,
 304              (unsigned long long)wq->sq.udb, (unsigned long long)wq->rq.udb);
 305
 306         return 0;
 307 free_dma:
 308         dma_free_coherent(&(rdev->lldi.pdev->dev),
 309                           wq->rq.memsize, wq->rq.queue,
 310                           dma_unmap_addr(&wq->rq, mapping));
 311 free_sq:
 312         dealloc_sq(rdev, &wq->sq);
 313 free_hwaddr:
 314         c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
 315 free_sw_rq:
 316         kfree(wq->rq.sw_rq);
 317 free_sw_sq:
 318         kfree(wq->sq.sw_sq);
 319 free_rq_qid:
 320         c4iw_put_qpid(rdev, wq->rq.qid, uctx);
 321 free_sq_qid:
 322         c4iw_put_qpid(rdev, wq->sq.qid, uctx);
 323         return ret;
 324 }
 325
 326 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
 327                       struct ib_send_wr *wr, int max, u32 *plenp)
 328 {
 329         u8 *dstp, *srcp;
 330         u32 plen = 0;
 331         int i;
 332         int rem, len;
 333
 334         dstp = (u8 *)immdp->data;
 335         for (i = 0; i < wr->num_sge; i++) {
 336                 if ((plen + wr->sg_list[i].length) > max)
 337                         return -EMSGSIZE;
 338                 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
 339                 plen += wr->sg_list[i].length;
 340                 rem = wr->sg_list[i].length;
 341                 while (rem) {
 342                         if (dstp == (u8 *)&sq->queue[sq->size])
 343                                 dstp = (u8 *)sq->queue;
 344                         if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
 345                                 len = rem;
 346                         else
 347                                 len = (u8 *)&sq->queue[sq->size] - dstp;
 348                         memcpy(dstp, srcp, len);
 349                         dstp += len;
 350                         srcp += len;
 351                         rem -= len;
 352                 }
 353         }
 354         len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
 355         if (len)
 356                 memset(dstp, 0, len);
 357         immdp->op = FW_RI_DATA_IMMD;
 358         immdp->r1 = 0;
 359         immdp->r2 = 0;
 360         immdp->immdlen = cpu_to_be32(plen);
 361         *plenp = plen;
 362         return 0;
 363 }
 364
 365 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
 366                       struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
 367                       int num_sge, u32 *plenp)
 368
 369 {
 370         int i;
 371         u32 plen = 0;
 372         __be64 *flitp = (__be64 *)isglp->sge;
 373
 374         for (i = 0; i < num_sge; i++) {
 375                 if ((plen + sg_list[i].length) < plen)
 376                         return -EMSGSIZE;
 377                 plen += sg_list[i].length;
 378                 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
 379                                      sg_list[i].length);
 380                 if (++flitp == queue_end)
 381                         flitp = queue_start;
 382                 *flitp = cpu_to_be64(sg_list[i].addr);
 383                 if (++flitp == queue_end)
 384                         flitp = queue_start;
 385         }
 386         *flitp = (__force __be64)0;
 387         isglp->op = FW_RI_DATA_ISGL;
 388         isglp->r1 = 0;
 389         isglp->nsge = cpu_to_be16(num_sge);
 390         isglp->r2 = 0;
 391         if (plenp)
 392                 *plenp = plen;
 393         return 0;
 394 }
 395
 396 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
 397                            struct ib_send_wr *wr, u8 *len16)
 398 {
 399         u32 plen;
 400         int size;
 401         int ret;
 402
 403         if (wr->num_sge > T4_MAX_SEND_SGE)
 404                 return -EINVAL;
 405         switch (wr->opcode) {
 406         case IB_WR_SEND:
 407                 if (wr->send_flags & IB_SEND_SOLICITED)
 408                         wqe->send.sendop_pkd = cpu_to_be32(
 409                                 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE));
 410                 else
 411                         wqe->send.sendop_pkd = cpu_to_be32(
 412                                 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND));
 413                 wqe->send.stag_inv = 0;
 414                 break;
 415         case IB_WR_SEND_WITH_INV:
 416                 if (wr->send_flags & IB_SEND_SOLICITED)
 417                         wqe->send.sendop_pkd = cpu_to_be32(
 418                                 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV));
 419                 else
 420                         wqe->send.sendop_pkd = cpu_to_be32(
 421                                 V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV));
 422                 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 423                 break;
 424
 425         default:
 426                 return -EINVAL;
 427         }
 428
 429         plen = 0;
 430         if (wr->num_sge) {
 431                 if (wr->send_flags & IB_SEND_INLINE) {
 432                         ret = build_immd(sq, wqe->send.u.immd_src, wr,
 433                                          T4_MAX_SEND_INLINE, &plen);
 434                         if (ret)
 435                                 return ret;
 436                         size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
 437                                plen;
 438                 } else {
 439                         ret = build_isgl((__be64 *)sq->queue,
 440                                          (__be64 *)&sq->queue[sq->size],
 441                                          wqe->send.u.isgl_src,
 442                                          wr->sg_list, wr->num_sge, &plen);
 443                         if (ret)
 444                                 return ret;
 445                         size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
 446                                wr->num_sge * sizeof(struct fw_ri_sge);
 447                 }
 448         } else {
 449                 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
 450                 wqe->send.u.immd_src[0].r1 = 0;
 451                 wqe->send.u.immd_src[0].r2 = 0;
 452                 wqe->send.u.immd_src[0].immdlen = 0;
 453                 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
 454                 plen = 0;
 455         }
 456         *len16 = DIV_ROUND_UP(size, 16);
 457         wqe->send.plen = cpu_to_be32(plen);
 458         return 0;
 459 }
 460
 461 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
 462                             struct ib_send_wr *wr, u8 *len16)
 463 {
 464         u32 plen;
 465         int size;
 466         int ret;
 467
 468         if (wr->num_sge > T4_MAX_SEND_SGE)
 469                 return -EINVAL;
 470         wqe->write.r2 = 0;
 471         wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
 472         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
 473         if (wr->num_sge) {
 474                 if (wr->send_flags & IB_SEND_INLINE) {
 475                         ret = build_immd(sq, wqe->write.u.immd_src, wr,
 476                                          T4_MAX_WRITE_INLINE, &plen);
 477                         if (ret)
 478                                 return ret;
 479                         size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
 480                                plen;
 481                 } else {
 482                         ret = build_isgl((__be64 *)sq->queue,
 483                                          (__be64 *)&sq->queue[sq->size],
 484                                          wqe->write.u.isgl_src,
 485                                          wr->sg_list, wr->num_sge, &plen);
 486                         if (ret)
 487                                 return ret;
 488                         size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
 489                                wr->num_sge * sizeof(struct fw_ri_sge);
 490                 }
 491         } else {
 492                 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
 493                 wqe->write.u.immd_src[0].r1 = 0;
 494                 wqe->write.u.immd_src[0].r2 = 0;
 495                 wqe->write.u.immd_src[0].immdlen = 0;
 496                 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
 497                 plen = 0;
 498         }
 499         *len16 = DIV_ROUND_UP(size, 16);
 500         wqe->write.plen = cpu_to_be32(plen);
 501         return 0;
 502 }
 503
 504 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
 505 {
 506         if (wr->num_sge > 1)
 507                 return -EINVAL;
 508         if (wr->num_sge) {
 509                 wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
 510                 wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
 511                                                         >> 32));
 512                 wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
 513                 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
 514                 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
 515                 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
 516                                                          >> 32));
 517                 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
 518         } else {
 519                 wqe->read.stag_src = cpu_to_be32(2);
 520                 wqe->read.to_src_hi = 0;
 521                 wqe->read.to_src_lo = 0;
 522                 wqe->read.stag_sink = cpu_to_be32(2);
 523                 wqe->read.plen = 0;
 524                 wqe->read.to_sink_hi = 0;
 525                 wqe->read.to_sink_lo = 0;
 526         }
 527         wqe->read.r2 = 0;
 528         wqe->read.r5 = 0;
 529         *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
 530         return 0;
 531 }
 532
 533 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
 534                            struct ib_recv_wr *wr, u8 *len16)
 535 {
 536         int ret;
 537
 538         ret = build_isgl((__be64 *)qhp->wq.rq.queue,
 539                          (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
 540                          &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
 541         if (ret)
 542                 return ret;
 543         *len16 = DIV_ROUND_UP(sizeof wqe->recv +
 544                               wr->num_sge * sizeof(struct fw_ri_sge), 16);
 545         return 0;
 546 }
 547
 548 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
 549                          struct ib_send_wr *wr, u8 *len16, u8 t5dev)
 550 {
 551
 552         struct fw_ri_immd *imdp;
 553         __be64 *p;
 554         int i;
 555         int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
 556         int rem;
 557
 558         if (wr->wr.fast_reg.page_list_len > T4_MAX_FR_DEPTH)
 559                 return -EINVAL;
 560
 561         wqe->fr.qpbinde_to_dcacpu = 0;
 562         wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
 563         wqe->fr.addr_type = FW_RI_VA_BASED_TO;
 564         wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
 565         wqe->fr.len_hi = 0;
 566         wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
 567         wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
 568         wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
 569         wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
 570                                         0xffffffff);
 571
 572         if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
 573                 struct c4iw_fr_page_list *c4pl =
 574                         to_c4iw_fr_page_list(wr->wr.fast_reg.page_list);
 575                 struct fw_ri_dsgl *sglp;
 576
 577                 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 578                         wr->wr.fast_reg.page_list->page_list[i] = (__force u64)
 579                                 cpu_to_be64((u64)
 580                                 wr->wr.fast_reg.page_list->page_list[i]);
 581                 }
 582
 583                 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
 584                 sglp->op = FW_RI_DATA_DSGL;
 585                 sglp->r1 = 0;
 586                 sglp->nsge = cpu_to_be16(1);
 587                 sglp->addr0 = cpu_to_be64(c4pl->dma_addr);
 588                 sglp->len0 = cpu_to_be32(pbllen);
 589
 590                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
 591         } else {
 592                 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
 593                 imdp->op = FW_RI_DATA_IMMD;
 594                 imdp->r1 = 0;
 595                 imdp->r2 = 0;
 596                 imdp->immdlen = cpu_to_be32(pbllen);
 597                 p = (__be64 *)(imdp + 1);
 598                 rem = pbllen;
 599                 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 600                         *p = cpu_to_be64(
 601                                 (u64)wr->wr.fast_reg.page_list->page_list[i]);
 602                         rem -= sizeof(*p);
 603                         if (++p == (__be64 *)&sq->queue[sq->size])
 604                                 p = (__be64 *)sq->queue;
 605                 }
 606                 BUG_ON(rem < 0);
 607                 while (rem) {
 608                         *p = 0;
 609                         rem -= sizeof(*p);
 610                         if (++p == (__be64 *)&sq->queue[sq->size])
 611                                 p = (__be64 *)sq->queue;
 612                 }
 613                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
 614                                       + pbllen, 16);
 615         }
 616         return 0;
 617 }
 618
 619 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
 620                           u8 *len16)
 621 {
 622         wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 623         wqe->inv.r2 = 0;
 624         *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
 625         return 0;
 626 }
 627
 628 void c4iw_qp_add_ref(struct ib_qp *qp)
 629 {
 630         PDBG("%s ib_qp %p\n", __func__, qp);
 631         atomic_inc(&(to_c4iw_qp(qp)->refcnt));
 632 }
 633
 634 void c4iw_qp_rem_ref(struct ib_qp *qp)
 635 {
 636         PDBG("%s ib_qp %p\n", __func__, qp);
 637         if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
 638                 wake_up(&(to_c4iw_qp(qp)->wait));
 639 }
 640
 641 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 642                    struct ib_send_wr **bad_wr)
 643 {
 644         int err = 0;
 645         u8 len16 = 0;
 646         enum fw_wr_opcodes fw_opcode = 0;
 647         enum fw_ri_wr_flags fw_flags;
 648         struct c4iw_qp *qhp;
 649         union t4_wr *wqe;
 650         u32 num_wrs;
 651         struct t4_swsqe *swsqe;
 652         unsigned long flag;
 653         u16 idx = 0;
 654
 655         qhp = to_c4iw_qp(ibqp);
 656         spin_lock_irqsave(&qhp->lock, flag);
 657         if (t4_wq_in_error(&qhp->wq)) {
 658                 spin_unlock_irqrestore(&qhp->lock, flag);
 659                 return -EINVAL;
 660         }
 661         num_wrs = t4_sq_avail(&qhp->wq);
 662         if (num_wrs == 0) {
 663                 spin_unlock_irqrestore(&qhp->lock, flag);
 664                 return -ENOMEM;
 665         }
 666         while (wr) {
 667                 if (num_wrs == 0) {
 668                         err = -ENOMEM;
 669                         *bad_wr = wr;
 670                         break;
 671                 }
 672                 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
 673                       qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
 674
 675                 fw_flags = 0;
 676                 if (wr->send_flags & IB_SEND_SOLICITED)
 677                         fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
 678                 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
 679                         fw_flags |= FW_RI_COMPLETION_FLAG;
 680                 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
 681                 switch (wr->opcode) {
 682                 case IB_WR_SEND_WITH_INV:
 683                 case IB_WR_SEND:
 684                         if (wr->send_flags & IB_SEND_FENCE)
 685                                 fw_flags |= FW_RI_READ_FENCE_FLAG;
 686                         fw_opcode = FW_RI_SEND_WR;
 687                         if (wr->opcode == IB_WR_SEND)
 688                                 swsqe->opcode = FW_RI_SEND;
 689                         else
 690                                 swsqe->opcode = FW_RI_SEND_WITH_INV;
 691                         err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
 692                         break;
 693                 case IB_WR_RDMA_WRITE:
 694                         fw_opcode = FW_RI_RDMA_WRITE_WR;
 695                         swsqe->opcode = FW_RI_RDMA_WRITE;
 696                         err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
 697                         break;
 698                 case IB_WR_RDMA_READ:
 699                 case IB_WR_RDMA_READ_WITH_INV:
 700                         fw_opcode = FW_RI_RDMA_READ_WR;
 701                         swsqe->opcode = FW_RI_READ_REQ;
 702                         if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
 703                                 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
 704                         else
 705                                 fw_flags = 0;
 706                         err = build_rdma_read(wqe, wr, &len16);
 707                         if (err)
 708                                 break;
 709                         swsqe->read_len = wr->sg_list[0].length;
 710                         if (!qhp->wq.sq.oldest_read)
 711                                 qhp->wq.sq.oldest_read = swsqe;
 712                         break;
 713                 case IB_WR_FAST_REG_MR:
 714                         fw_opcode = FW_RI_FR_NSMR_WR;
 715                         swsqe->opcode = FW_RI_FAST_REGISTER;
 716                         err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16,
 717                                             is_t5(
 718                                             qhp->rhp->rdev.lldi.adapter_type) ?
 719                                             1 : 0);
 720                         break;
 721                 case IB_WR_LOCAL_INV:
 722                         if (wr->send_flags & IB_SEND_FENCE)
 723                                 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
 724                         fw_opcode = FW_RI_INV_LSTAG_WR;
 725                         swsqe->opcode = FW_RI_LOCAL_INV;
 726                         err = build_inv_stag(wqe, wr, &len16);
 727                         break;
 728                 default:
 729                         PDBG("%s post of type=%d TBD!\n", __func__,
 730                              wr->opcode);
 731                         err = -EINVAL;
 732                 }
 733                 if (err) {
 734                         *bad_wr = wr;
 735                         break;
 736                 }
 737                 swsqe->idx = qhp->wq.sq.pidx;
 738                 swsqe->complete = 0;
 739                 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
 740                                   qhp->sq_sig_all;
 741                 swsqe->flushed = 0;
 742                 swsqe->wr_id = wr->wr_id;
 743
 744                 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
 745
 746                 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
 747                      __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
 748                      swsqe->opcode, swsqe->read_len);
 749                 wr = wr->next;
 750                 num_wrs--;
 751                 t4_sq_produce(&qhp->wq, len16);
 752                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 753         }
 754         if (t4_wq_db_enabled(&qhp->wq))
 755                 t4_ring_sq_db(&qhp->wq, idx);
 756         spin_unlock_irqrestore(&qhp->lock, flag);
 757         return err;
 758 }
 759
 760 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 761                       struct ib_recv_wr **bad_wr)
 762 {
 763         int err = 0;
 764         struct c4iw_qp *qhp;
 765         union t4_recv_wr *wqe;
 766         u32 num_wrs;
 767         u8 len16 = 0;
 768         unsigned long flag;
 769         u16 idx = 0;
 770
 771         qhp = to_c4iw_qp(ibqp);
 772         spin_lock_irqsave(&qhp->lock, flag);
 773         if (t4_wq_in_error(&qhp->wq)) {
 774                 spin_unlock_irqrestore(&qhp->lock, flag);
 775                 return -EINVAL;
 776         }
 777         num_wrs = t4_rq_avail(&qhp->wq);
 778         if (num_wrs == 0) {
 779                 spin_unlock_irqrestore(&qhp->lock, flag);
 780                 return -ENOMEM;
 781         }
 782         while (wr) {
 783                 if (wr->num_sge > T4_MAX_RECV_SGE) {
 784                         err = -EINVAL;
 785                         *bad_wr = wr;
 786                         break;
 787                 }
 788                 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
 789                                            qhp->wq.rq.wq_pidx *
 790                                            T4_EQ_ENTRY_SIZE);
 791                 if (num_wrs)
 792                         err = build_rdma_recv(qhp, wqe, wr, &len16);
 793                 else
 794                         err = -ENOMEM;
 795                 if (err) {
 796                         *bad_wr = wr;
 797                         break;
 798                 }
 799
 800                 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
 801
 802                 wqe->recv.opcode = FW_RI_RECV_WR;
 803                 wqe->recv.r1 = 0;
 804                 wqe->recv.wrid = qhp->wq.rq.pidx;
 805                 wqe->recv.r2[0] = 0;
 806                 wqe->recv.r2[1] = 0;
 807                 wqe->recv.r2[2] = 0;
 808                 wqe->recv.len16 = len16;
 809                 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
 810                      (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
 811                 t4_rq_produce(&qhp->wq, len16);
 812                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 813                 wr = wr->next;
 814                 num_wrs--;
 815         }
 816         if (t4_wq_db_enabled(&qhp->wq))
 817                 t4_ring_rq_db(&qhp->wq, idx);
 818         spin_unlock_irqrestore(&qhp->lock, flag);
 819         return err;
 820 }
 821
 822 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
 823 {
 824         return -ENOSYS;
 825 }
 826
 827 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
 828                                     u8 *ecode)
 829 {
 830         int status;
 831         int tagged;
 832         int opcode;
 833         int rqtype;
 834         int send_inv;
 835
 836         if (!err_cqe) {
 837                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 838                 *ecode = 0;
 839                 return;
 840         }
 841
 842         status = CQE_STATUS(err_cqe);
 843         opcode = CQE_OPCODE(err_cqe);
 844         rqtype = RQ_TYPE(err_cqe);
 845         send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
 846                    (opcode == FW_RI_SEND_WITH_SE_INV);
 847         tagged = (opcode == FW_RI_RDMA_WRITE) ||
 848                  (rqtype && (opcode == FW_RI_READ_RESP));
 849
 850         switch (status) {
 851         case T4_ERR_STAG:
 852                 if (send_inv) {
 853                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 854                         *ecode = RDMAP_CANT_INV_STAG;
 855                 } else {
 856                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 857                         *ecode = RDMAP_INV_STAG;
 858                 }
 859                 break;
 860         case T4_ERR_PDID:
 861                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 862                 if ((opcode == FW_RI_SEND_WITH_INV) ||
 863                     (opcode == FW_RI_SEND_WITH_SE_INV))
 864                         *ecode = RDMAP_CANT_INV_STAG;
 865                 else
 866                         *ecode = RDMAP_STAG_NOT_ASSOC;
 867                 break;
 868         case T4_ERR_QPID:
 869                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 870                 *ecode = RDMAP_STAG_NOT_ASSOC;
 871                 break;
 872         case T4_ERR_ACCESS:
 873                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 874                 *ecode = RDMAP_ACC_VIOL;
 875                 break;
 876         case T4_ERR_WRAP:
 877                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 878                 *ecode = RDMAP_TO_WRAP;
 879                 break;
 880         case T4_ERR_BOUND:
 881                 if (tagged) {
 882                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 883                         *ecode = DDPT_BASE_BOUNDS;
 884                 } else {
 885                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 886                         *ecode = RDMAP_BASE_BOUNDS;
 887                 }
 888                 break;
 889         case T4_ERR_INVALIDATE_SHARED_MR:
 890         case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
 891                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 892                 *ecode = RDMAP_CANT_INV_STAG;
 893                 break;
 894         case T4_ERR_ECC:
 895         case T4_ERR_ECC_PSTAG:
 896         case T4_ERR_INTERNAL_ERR:
 897                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
 898                 *ecode = 0;
 899                 break;
 900         case T4_ERR_OUT_OF_RQE:
 901                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 902                 *ecode = DDPU_INV_MSN_NOBUF;
 903                 break;
 904         case T4_ERR_PBL_ADDR_BOUND:
 905                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 906                 *ecode = DDPT_BASE_BOUNDS;
 907                 break;
 908         case T4_ERR_CRC:
 909                 *layer_type = LAYER_MPA|DDP_LLP;
 910                 *ecode = MPA_CRC_ERR;
 911                 break;
 912         case T4_ERR_MARKER:
 913                 *layer_type = LAYER_MPA|DDP_LLP;
 914                 *ecode = MPA_MARKER_ERR;
 915                 break;
 916         case T4_ERR_PDU_LEN_ERR:
 917                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 918                 *ecode = DDPU_MSG_TOOBIG;
 919                 break;
 920         case T4_ERR_DDP_VERSION:
 921                 if (tagged) {
 922                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
 923                         *ecode = DDPT_INV_VERS;
 924                 } else {
 925                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 926                         *ecode = DDPU_INV_VERS;
 927                 }
 928                 break;
 929         case T4_ERR_RDMA_VERSION:
 930                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 931                 *ecode = RDMAP_INV_VERS;
 932                 break;
 933         case T4_ERR_OPCODE:
 934                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 935                 *ecode = RDMAP_INV_OPCODE;
 936                 break;
 937         case T4_ERR_DDP_QUEUE_NUM:
 938                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 939                 *ecode = DDPU_INV_QN;
 940                 break;
 941         case T4_ERR_MSN:
 942         case T4_ERR_MSN_GAP:
 943         case T4_ERR_MSN_RANGE:
 944         case T4_ERR_IRD_OVERFLOW:
 945                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 946                 *ecode = DDPU_INV_MSN_RANGE;
 947                 break;
 948         case T4_ERR_TBIT:
 949                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
 950                 *ecode = 0;
 951                 break;
 952         case T4_ERR_MO:
 953                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
 954                 *ecode = DDPU_INV_MO;
 955                 break;
 956         default:
 957                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 958                 *ecode = 0;
 959                 break;
 960         }
 961 }
 962
 963 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
 964                            gfp_t gfp)
 965 {
 966         struct fw_ri_wr *wqe;
 967         struct sk_buff *skb;
 968         struct terminate_message *term;
 969
 970         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
 971              qhp->ep->hwtid);
 972
 973         skb = alloc_skb(sizeof *wqe, gfp);
 974         if (!skb)
 975                 return;
 976         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
 977
 978         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
 979         memset(wqe, 0, sizeof *wqe);
 980         wqe->op_compl = cpu_to_be32(FW_WR_OP(FW_RI_INIT_WR));
 981         wqe->flowid_len16 = cpu_to_be32(
 982                 FW_WR_FLOWID(qhp->ep->hwtid) |
 983                 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
 984
 985         wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
 986         wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
 987         term = (struct terminate_message *)wqe->u.terminate.termmsg;
 988         if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
 989                 term->layer_etype = qhp->attr.layer_etype;
 990                 term->ecode = qhp->attr.ecode;
 991         } else
 992                 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
 993         c4iw_ofld_send(&qhp->rhp->rdev, skb);
 994 }
 995
 996 /*
 997  * Assumes qhp lock is held.
 998  */
 999 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1000                        struct c4iw_cq *schp)
1001 {
1002         int count;
1003         int flushed;
1004         unsigned long flag;
1005
1006         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1007
1008         /* locking hierarchy: cq lock first, then qp lock. */
1009         spin_lock_irqsave(&rchp->lock, flag);
1010         spin_lock(&qhp->lock);
1011
1012         if (qhp->wq.flushed) {
1013                 spin_unlock(&qhp->lock);
1014                 spin_unlock_irqrestore(&rchp->lock, flag);
1015                 return;
1016         }
1017         qhp->wq.flushed = 1;
1018
1019         c4iw_flush_hw_cq(rchp);
1020         c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1021         flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1022         spin_unlock(&qhp->lock);
1023         spin_unlock_irqrestore(&rchp->lock, flag);
1024         if (flushed) {
1025                 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1026                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1027                 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1028         }
1029
1030         /* locking hierarchy: cq lock first, then qp lock. */
1031         spin_lock_irqsave(&schp->lock, flag);
1032         spin_lock(&qhp->lock);
1033         if (schp != rchp)
1034                 c4iw_flush_hw_cq(schp);
1035         flushed = c4iw_flush_sq(qhp);
1036         spin_unlock(&qhp->lock);
1037         spin_unlock_irqrestore(&schp->lock, flag);
1038         if (flushed) {
1039                 spin_lock_irqsave(&schp->comp_handler_lock, flag);
1040                 (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context);
1041                 spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1042         }
1043 }
1044
1045 static void flush_qp(struct c4iw_qp *qhp)
1046 {
1047         struct c4iw_cq *rchp, *schp;
1048         unsigned long flag;
1049
1050         rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1051         schp = to_c4iw_cq(qhp->ibqp.send_cq);
1052
1053         t4_set_wq_in_error(&qhp->wq);
1054         if (qhp->ibqp.uobject) {
1055                 t4_set_cq_in_error(&rchp->cq);
1056                 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1057                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1058                 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1059                 if (schp != rchp) {
1060                         t4_set_cq_in_error(&schp->cq);
1061                         spin_lock_irqsave(&schp->comp_handler_lock, flag);
1062                         (*schp->ibcq.comp_handler)(&schp->ibcq,
1063                                         schp->ibcq.cq_context);
1064                         spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1065                 }
1066                 return;
1067         }
1068         __flush_qp(qhp, rchp, schp);
1069 }
1070
1071 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1072                      struct c4iw_ep *ep)
1073 {
1074         struct fw_ri_wr *wqe;
1075         int ret;
1076         struct sk_buff *skb;
1077
1078         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1079              ep->hwtid);
1080
1081         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1082         if (!skb)
1083                 return -ENOMEM;
1084         set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1085
1086         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1087         memset(wqe, 0, sizeof *wqe);
1088         wqe->op_compl = cpu_to_be32(
1089                 FW_WR_OP(FW_RI_INIT_WR) |
1090                 FW_WR_COMPL(1));
1091         wqe->flowid_len16 = cpu_to_be32(
1092                 FW_WR_FLOWID(ep->hwtid) |
1093                 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1094         wqe->cookie = (unsigned long) &ep->com.wr_wait;
1095
1096         wqe->u.fini.type = FW_RI_TYPE_FINI;
1097         ret = c4iw_ofld_send(&rhp->rdev, skb);
1098         if (ret)
1099                 goto out;
1100
1101         ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1102                              qhp->wq.sq.qid, __func__);
1103 out:
1104         PDBG("%s ret %d\n", __func__, ret);
1105         return ret;
1106 }
1107
1108 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1109 {
1110         PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1111         memset(&init->u, 0, sizeof init->u);
1112         switch (p2p_type) {
1113         case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1114                 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1115                 init->u.write.stag_sink = cpu_to_be32(1);
1116                 init->u.write.to_sink = cpu_to_be64(1);
1117                 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1118                 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1119                                                    sizeof(struct fw_ri_immd),
1120                                                    16);
1121                 break;
1122         case FW_RI_INIT_P2PTYPE_READ_REQ:
1123                 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1124                 init->u.read.stag_src = cpu_to_be32(1);
1125                 init->u.read.to_src_lo = cpu_to_be32(1);
1126                 init->u.read.stag_sink = cpu_to_be32(1);
1127                 init->u.read.to_sink_lo = cpu_to_be32(1);
1128                 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1129                 break;
1130         }
1131 }
1132
1133 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1134 {
1135         struct fw_ri_wr *wqe;
1136         int ret;
1137         struct sk_buff *skb;
1138
1139         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1140              qhp->ep->hwtid);
1141
1142         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1143         if (!skb)
1144                 return -ENOMEM;
1145         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1146
1147         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1148         memset(wqe, 0, sizeof *wqe);
1149         wqe->op_compl = cpu_to_be32(
1150                 FW_WR_OP(FW_RI_INIT_WR) |
1151                 FW_WR_COMPL(1));
1152         wqe->flowid_len16 = cpu_to_be32(
1153                 FW_WR_FLOWID(qhp->ep->hwtid) |
1154                 FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16)));
1155
1156         wqe->cookie = (unsigned long) &qhp->ep->com.wr_wait;
1157
1158         wqe->u.init.type = FW_RI_TYPE_INIT;
1159         wqe->u.init.mpareqbit_p2ptype =
1160                 V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) |
1161                 V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type);
1162         wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1163         if (qhp->attr.mpa_attr.recv_marker_enabled)
1164                 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1165         if (qhp->attr.mpa_attr.xmit_marker_enabled)
1166                 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1167         if (qhp->attr.mpa_attr.crc_enabled)
1168                 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1169
1170         wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1171                             FW_RI_QP_RDMA_WRITE_ENABLE |
1172                             FW_RI_QP_BIND_ENABLE;
1173         if (!qhp->ibqp.uobject)
1174                 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1175                                      FW_RI_QP_STAG0_ENABLE;
1176         wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1177         wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1178         wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1179         wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1180         wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1181         wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1182         wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1183         wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1184         wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1185         wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1186         wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1187         wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1188         wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1189                                          rhp->rdev.lldi.vr->rq.start);
1190         if (qhp->attr.mpa_attr.initiator)
1191                 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1192
1193         ret = c4iw_ofld_send(&rhp->rdev, skb);
1194         if (ret)
1195                 goto out;
1196
1197         ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1198                                   qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1199 out:
1200         PDBG("%s ret %d\n", __func__, ret);
1201         return ret;
1202 }
1203
1204 /*
1205  * Called by the library when the qp has user dbs disabled due to
1206  * a DB_FULL condition.  This function will single-thread all user
1207  * DB rings to avoid overflowing the hw db-fifo.
1208  */
1209 static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
1210 {
1211         int delay = db_delay_usecs;
1212
1213         mutex_lock(&qhp->rhp->db_mutex);
1214         do {
1215
1216                 /*
1217                  * The interrupt threshold is dbfifo_int_thresh << 6. So
1218                  * make sure we don't cross that and generate an interrupt.
1219                  */
1220                 if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
1221                     (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
1222                         writel(QID(qid) | PIDX(inc), qhp->wq.db);
1223                         break;
1224                 }
1225                 set_current_state(TASK_UNINTERRUPTIBLE);
1226                 schedule_timeout(usecs_to_jiffies(delay));
1227                 delay = min(delay << 1, 2000);
1228         } while (1);
1229         mutex_unlock(&qhp->rhp->db_mutex);
1230         return 0;
1231 }
1232
1233 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1234                    enum c4iw_qp_attr_mask mask,
1235                    struct c4iw_qp_attributes *attrs,
1236                    int internal)
1237 {
1238         int ret = 0;
1239         struct c4iw_qp_attributes newattr = qhp->attr;
1240         int disconnect = 0;
1241         int terminate = 0;
1242         int abort = 0;
1243         int free = 0;
1244         struct c4iw_ep *ep = NULL;
1245
1246         PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1247              qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1248              (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1249
1250         mutex_lock(&qhp->mutex);
1251
1252         /* Process attr changes if in IDLE */
1253         if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1254                 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1255                         ret = -EIO;
1256                         goto out;
1257                 }
1258                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1259                         newattr.enable_rdma_read = attrs->enable_rdma_read;
1260                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1261                         newattr.enable_rdma_write = attrs->enable_rdma_write;
1262                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1263                         newattr.enable_bind = attrs->enable_bind;
1264                 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1265                         if (attrs->max_ord > c4iw_max_read_depth) {
1266                                 ret = -EINVAL;
1267                                 goto out;
1268                         }
1269                         newattr.max_ord = attrs->max_ord;
1270                 }
1271                 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1272                         if (attrs->max_ird > c4iw_max_read_depth) {
1273                                 ret = -EINVAL;
1274                                 goto out;
1275                         }
1276                         newattr.max_ird = attrs->max_ird;
1277                 }
1278                 qhp->attr = newattr;
1279         }
1280
1281         if (mask & C4IW_QP_ATTR_SQ_DB) {
1282                 ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
1283                 goto out;
1284         }
1285         if (mask & C4IW_QP_ATTR_RQ_DB) {
1286                 ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
1287                 goto out;
1288         }
1289
1290         if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1291                 goto out;
1292         if (qhp->attr.state == attrs->next_state)
1293                 goto out;
1294
1295         switch (qhp->attr.state) {
1296         case C4IW_QP_STATE_IDLE:
1297                 switch (attrs->next_state) {
1298                 case C4IW_QP_STATE_RTS:
1299                         if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1300                                 ret = -EINVAL;
1301                                 goto out;
1302                         }
1303                         if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1304                                 ret = -EINVAL;
1305                                 goto out;
1306                         }
1307                         qhp->attr.mpa_attr = attrs->mpa_attr;
1308                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1309                         qhp->ep = qhp->attr.llp_stream_handle;
1310                         set_state(qhp, C4IW_QP_STATE_RTS);
1311
1312                         /*
1313                          * Ref the endpoint here and deref when we
1314                          * disassociate the endpoint from the QP.  This
1315                          * happens in CLOSING->IDLE transition or *->ERROR
1316                          * transition.
1317                          */
1318                         c4iw_get_ep(&qhp->ep->com);
1319                         ret = rdma_init(rhp, qhp);
1320                         if (ret)
1321                                 goto err;
1322                         break;
1323                 case C4IW_QP_STATE_ERROR:
1324                         set_state(qhp, C4IW_QP_STATE_ERROR);
1325                         flush_qp(qhp);
1326                         break;
1327                 default:
1328                         ret = -EINVAL;
1329                         goto out;
1330                 }
1331                 break;
1332         case C4IW_QP_STATE_RTS:
1333                 switch (attrs->next_state) {
1334                 case C4IW_QP_STATE_CLOSING:
1335                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1336                         set_state(qhp, C4IW_QP_STATE_CLOSING);
1337                         ep = qhp->ep;
1338                         if (!internal) {
1339                                 abort = 0;
1340                                 disconnect = 1;
1341                                 c4iw_get_ep(&qhp->ep->com);
1342                         }
1343                         t4_set_wq_in_error(&qhp->wq);
1344                         ret = rdma_fini(rhp, qhp, ep);
1345                         if (ret)
1346                                 goto err;
1347                         break;
1348                 case C4IW_QP_STATE_TERMINATE:
1349                         set_state(qhp, C4IW_QP_STATE_TERMINATE);
1350                         qhp->attr.layer_etype = attrs->layer_etype;
1351                         qhp->attr.ecode = attrs->ecode;
1352                         t4_set_wq_in_error(&qhp->wq);
1353                         ep = qhp->ep;
1354                         disconnect = 1;
1355                         if (!internal)
1356                                 terminate = 1;
1357                         else {
1358                                 ret = rdma_fini(rhp, qhp, ep);
1359                                 if (ret)
1360                                         goto err;
1361                         }
1362                         c4iw_get_ep(&qhp->ep->com);
1363                         break;
1364                 case C4IW_QP_STATE_ERROR:
1365                         set_state(qhp, C4IW_QP_STATE_ERROR);
1366                         t4_set_wq_in_error(&qhp->wq);
1367                         if (!internal) {
1368                                 abort = 1;
1369                                 disconnect = 1;
1370                                 ep = qhp->ep;
1371                                 c4iw_get_ep(&qhp->ep->com);
1372                         }
1373                         goto err;
1374                         break;
1375                 default:
1376                         ret = -EINVAL;
1377                         goto out;
1378                 }
1379                 break;
1380         case C4IW_QP_STATE_CLOSING:
1381                 if (!internal) {
1382                         ret = -EINVAL;
1383                         goto out;
1384                 }
1385                 switch (attrs->next_state) {
1386                 case C4IW_QP_STATE_IDLE:
1387                         flush_qp(qhp);
1388                         set_state(qhp, C4IW_QP_STATE_IDLE);
1389                         qhp->attr.llp_stream_handle = NULL;
1390                         c4iw_put_ep(&qhp->ep->com);
1391                         qhp->ep = NULL;
1392                         wake_up(&qhp->wait);
1393                         break;
1394                 case C4IW_QP_STATE_ERROR:
1395                         goto err;
1396                 default:
1397                         ret = -EINVAL;
1398                         goto err;
1399                 }
1400                 break;
1401         case C4IW_QP_STATE_ERROR:
1402                 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1403                         ret = -EINVAL;
1404                         goto out;
1405                 }
1406                 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1407                         ret = -EINVAL;
1408                         goto out;
1409                 }
1410                 set_state(qhp, C4IW_QP_STATE_IDLE);
1411                 break;
1412         case C4IW_QP_STATE_TERMINATE:
1413                 if (!internal) {
1414                         ret = -EINVAL;
1415                         goto out;
1416                 }
1417                 goto err;
1418                 break;
1419         default:
1420                 printk(KERN_ERR "%s in a bad state %d\n",
1421                        __func__, qhp->attr.state);
1422                 ret = -EINVAL;
1423                 goto err;
1424                 break;
1425         }
1426         goto out;
1427 err:
1428         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1429              qhp->wq.sq.qid);
1430
1431         /* disassociate the LLP connection */
1432         qhp->attr.llp_stream_handle = NULL;
1433         if (!ep)
1434                 ep = qhp->ep;
1435         qhp->ep = NULL;
1436         set_state(qhp, C4IW_QP_STATE_ERROR);
1437         free = 1;
1438         abort = 1;
1439         wake_up(&qhp->wait);
1440         BUG_ON(!ep);
1441         flush_qp(qhp);
1442 out:
1443         mutex_unlock(&qhp->mutex);
1444
1445         if (terminate)
1446                 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1447
1448         /*
1449          * If disconnect is 1, then we need to initiate a disconnect
1450          * on the EP.  This can be a normal close (RTS->CLOSING) or
1451          * an abnormal close (RTS/CLOSING->ERROR).
1452          */
1453         if (disconnect) {
1454                 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1455                                                          GFP_KERNEL);
1456                 c4iw_put_ep(&ep->com);
1457         }
1458
1459         /*
1460          * If free is 1, then we've disassociated the EP from the QP
1461          * and we need to dereference the EP.
1462          */
1463         if (free)
1464                 c4iw_put_ep(&ep->com);
1465         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1466         return ret;
1467 }
1468
1469 static int enable_qp_db(int id, void *p, void *data)
1470 {
1471         struct c4iw_qp *qp = p;
1472
1473         t4_enable_wq_db(&qp->wq);
1474         return 0;
1475 }
1476
1477 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1478 {
1479         struct c4iw_dev *rhp;
1480         struct c4iw_qp *qhp;
1481         struct c4iw_qp_attributes attrs;
1482         struct c4iw_ucontext *ucontext;
1483
1484         qhp = to_c4iw_qp(ib_qp);
1485         rhp = qhp->rhp;
1486
1487         attrs.next_state = C4IW_QP_STATE_ERROR;
1488         if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1489                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1490         else
1491                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1492         wait_event(qhp->wait, !qhp->ep);
1493
1494         spin_lock_irq(&rhp->lock);
1495         remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1496         rhp->qpcnt--;
1497         BUG_ON(rhp->qpcnt < 0);
1498         if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
1499                 rhp->rdev.stats.db_state_transitions++;
1500                 rhp->db_state = NORMAL;
1501                 idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
1502         }
1503         if (db_coalescing_threshold >= 0)
1504                 if (rhp->qpcnt <= db_coalescing_threshold)
1505                         cxgb4_enable_db_coalescing(rhp->rdev.lldi.ports[0]);
1506         spin_unlock_irq(&rhp->lock);
1507         atomic_dec(&qhp->refcnt);
1508         wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1509
1510         ucontext = ib_qp->uobject ?
1511                    to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1512         destroy_qp(&rhp->rdev, &qhp->wq,
1513                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1514
1515         PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1516         kfree(qhp);
1517         return 0;
1518 }
1519
1520 static int disable_qp_db(int id, void *p, void *data)
1521 {
1522         struct c4iw_qp *qp = p;
1523
1524         t4_disable_wq_db(&qp->wq);
1525         return 0;
1526 }
1527
1528 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1529                              struct ib_udata *udata)
1530 {
1531         struct c4iw_dev *rhp;
1532         struct c4iw_qp *qhp;
1533         struct c4iw_pd *php;
1534         struct c4iw_cq *schp;
1535         struct c4iw_cq *rchp;
1536         struct c4iw_create_qp_resp uresp;
1537         unsigned int sqsize, rqsize;
1538         struct c4iw_ucontext *ucontext;
1539         int ret;
1540         struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1541
1542         PDBG("%s ib_pd %p\n", __func__, pd);
1543
1544         if (attrs->qp_type != IB_QPT_RC)
1545                 return ERR_PTR(-EINVAL);
1546
1547         php = to_c4iw_pd(pd);
1548         rhp = php->rhp;
1549         schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1550         rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1551         if (!schp || !rchp)
1552                 return ERR_PTR(-EINVAL);
1553
1554         if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1555                 return ERR_PTR(-EINVAL);
1556
1557         rqsize = roundup(attrs->cap.max_recv_wr + 1, 16);
1558         if (rqsize > T4_MAX_RQ_SIZE)
1559                 return ERR_PTR(-E2BIG);
1560
1561         sqsize = roundup(attrs->cap.max_send_wr + 1, 16);
1562         if (sqsize > T4_MAX_SQ_SIZE)
1563                 return ERR_PTR(-E2BIG);
1564
1565         ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1566
1567         qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1568         if (!qhp)
1569                 return ERR_PTR(-ENOMEM);
1570         qhp->wq.sq.size = sqsize;
1571         qhp->wq.sq.memsize = (sqsize + 1) * sizeof *qhp->wq.sq.queue;
1572         qhp->wq.sq.flush_cidx = -1;
1573         qhp->wq.rq.size = rqsize;
1574         qhp->wq.rq.memsize = (rqsize + 1) * sizeof *qhp->wq.rq.queue;
1575
1576         if (ucontext) {
1577                 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1578                 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1579         }
1580
1581         PDBG("%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu\n",
1582              __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize);
1583
1584         ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1585                         ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1586         if (ret)
1587                 goto err1;
1588
1589         attrs->cap.max_recv_wr = rqsize - 1;
1590         attrs->cap.max_send_wr = sqsize - 1;
1591         attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1592
1593         qhp->rhp = rhp;
1594         qhp->attr.pd = php->pdid;
1595         qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1596         qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1597         qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1598         qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1599         qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1600         qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1601         qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1602         qhp->attr.state = C4IW_QP_STATE_IDLE;
1603         qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1604         qhp->attr.enable_rdma_read = 1;
1605         qhp->attr.enable_rdma_write = 1;
1606         qhp->attr.enable_bind = 1;
1607         qhp->attr.max_ord = 1;
1608         qhp->attr.max_ird = 1;
1609         qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1610         spin_lock_init(&qhp->lock);
1611         mutex_init(&qhp->mutex);
1612         init_waitqueue_head(&qhp->wait);
1613         atomic_set(&qhp->refcnt, 1);
1614
1615         spin_lock_irq(&rhp->lock);
1616         if (rhp->db_state != NORMAL)
1617                 t4_disable_wq_db(&qhp->wq);
1618         rhp->qpcnt++;
1619         if (rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
1620                 rhp->rdev.stats.db_state_transitions++;
1621                 rhp->db_state = FLOW_CONTROL;
1622                 idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
1623         }
1624         if (db_coalescing_threshold >= 0)
1625                 if (rhp->qpcnt > db_coalescing_threshold)
1626                         cxgb4_disable_db_coalescing(rhp->rdev.lldi.ports[0]);
1627         ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1628         spin_unlock_irq(&rhp->lock);
1629         if (ret)
1630                 goto err2;
1631
1632         if (udata) {
1633                 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1634                 if (!mm1) {
1635                         ret = -ENOMEM;
1636                         goto err3;
1637                 }
1638                 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1639                 if (!mm2) {
1640                         ret = -ENOMEM;
1641                         goto err4;
1642                 }
1643                 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1644                 if (!mm3) {
1645                         ret = -ENOMEM;
1646                         goto err5;
1647                 }
1648                 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1649                 if (!mm4) {
1650                         ret = -ENOMEM;
1651                         goto err6;
1652                 }
1653                 if (t4_sq_onchip(&qhp->wq.sq)) {
1654                         mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1655                         if (!mm5) {
1656                                 ret = -ENOMEM;
1657                                 goto err7;
1658                         }
1659                         uresp.flags = C4IW_QPF_ONCHIP;
1660                 } else
1661                         uresp.flags = 0;
1662                 uresp.qid_mask = rhp->rdev.qpmask;
1663                 uresp.sqid = qhp->wq.sq.qid;
1664                 uresp.sq_size = qhp->wq.sq.size;
1665                 uresp.sq_memsize = qhp->wq.sq.memsize;
1666                 uresp.rqid = qhp->wq.rq.qid;
1667                 uresp.rq_size = qhp->wq.rq.size;
1668                 uresp.rq_memsize = qhp->wq.rq.memsize;
1669                 spin_lock(&ucontext->mmap_lock);
1670                 if (mm5) {
1671                         uresp.ma_sync_key = ucontext->key;
1672                         ucontext->key += PAGE_SIZE;
1673                 } else {
1674                         uresp.ma_sync_key =  0;
1675                 }
1676                 uresp.sq_key = ucontext->key;
1677                 ucontext->key += PAGE_SIZE;
1678                 uresp.rq_key = ucontext->key;
1679                 ucontext->key += PAGE_SIZE;
1680                 uresp.sq_db_gts_key = ucontext->key;
1681                 ucontext->key += PAGE_SIZE;
1682                 uresp.rq_db_gts_key = ucontext->key;
1683                 ucontext->key += PAGE_SIZE;
1684                 spin_unlock(&ucontext->mmap_lock);
1685                 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1686                 if (ret)
1687                         goto err8;
1688                 mm1->key = uresp.sq_key;
1689                 mm1->addr = qhp->wq.sq.phys_addr;
1690                 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1691                 insert_mmap(ucontext, mm1);
1692                 mm2->key = uresp.rq_key;
1693                 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1694                 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1695                 insert_mmap(ucontext, mm2);
1696                 mm3->key = uresp.sq_db_gts_key;
1697                 mm3->addr = qhp->wq.sq.udb;
1698                 mm3->len = PAGE_SIZE;
1699                 insert_mmap(ucontext, mm3);
1700                 mm4->key = uresp.rq_db_gts_key;
1701                 mm4->addr = qhp->wq.rq.udb;
1702                 mm4->len = PAGE_SIZE;
1703                 insert_mmap(ucontext, mm4);
1704                 if (mm5) {
1705                         mm5->key = uresp.ma_sync_key;
1706                         mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1707                                     + A_PCIE_MA_SYNC) & PAGE_MASK;
1708                         mm5->len = PAGE_SIZE;
1709                         insert_mmap(ucontext, mm5);
1710                 }
1711         }
1712         qhp->ibqp.qp_num = qhp->wq.sq.qid;
1713         init_timer(&(qhp->timer));
1714         PDBG("%s qhp %p sq_num_entries %d, rq_num_entries %d qpid 0x%0x\n",
1715              __func__, qhp, qhp->attr.sq_num_entries, qhp->attr.rq_num_entries,
1716              qhp->wq.sq.qid);
1717         return &qhp->ibqp;
1718 err8:
1719         kfree(mm5);
1720 err7:
1721         kfree(mm4);
1722 err6:
1723         kfree(mm3);
1724 err5:
1725         kfree(mm2);
1726 err4:
1727         kfree(mm1);
1728 err3:
1729         remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1730 err2:
1731         destroy_qp(&rhp->rdev, &qhp->wq,
1732                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1733 err1:
1734         kfree(qhp);
1735         return ERR_PTR(ret);
1736 }
1737
1738 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1739                       int attr_mask, struct ib_udata *udata)
1740 {
1741         struct c4iw_dev *rhp;
1742         struct c4iw_qp *qhp;
1743         enum c4iw_qp_attr_mask mask = 0;
1744         struct c4iw_qp_attributes attrs;
1745
1746         PDBG("%s ib_qp %p\n", __func__, ibqp);
1747
1748         /* iwarp does not support the RTR state */
1749         if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1750                 attr_mask &= ~IB_QP_STATE;
1751
1752         /* Make sure we still have something left to do */
1753         if (!attr_mask)
1754                 return 0;
1755
1756         memset(&attrs, 0, sizeof attrs);
1757         qhp = to_c4iw_qp(ibqp);
1758         rhp = qhp->rhp;
1759
1760         attrs.next_state = c4iw_convert_state(attr->qp_state);
1761         attrs.enable_rdma_read = (attr->qp_access_flags &
1762                                IB_ACCESS_REMOTE_READ) ?  1 : 0;
1763         attrs.enable_rdma_write = (attr->qp_access_flags &
1764                                 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1765         attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1766
1767
1768         mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1769         mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1770                         (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1771                          C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1772                          C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1773
1774         /*
1775          * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1776          * ringing the queue db when we're in DB_FULL mode.
1777          */
1778         attrs.sq_db_inc = attr->sq_psn;
1779         attrs.rq_db_inc = attr->rq_psn;
1780         mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1781         mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1782
1783         return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1784 }
1785
1786 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1787 {
1788         PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1789         return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1790 }
1791
1792 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1793                      int attr_mask, struct ib_qp_init_attr *init_attr)
1794 {
1795         struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1796
1797         memset(attr, 0, sizeof *attr);
1798         memset(init_attr, 0, sizeof *init_attr);
1799         attr->qp_state = to_ib_qp_state(qhp->attr.state);
1800         return 0;
1801 }