kernel/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 int alloc_ird(struct c4iw_dev *dev, u32 ird)
  62 {
  63         int ret = 0;
  64
  65         spin_lock_irq(&dev->lock);
  66         if (ird <= dev->avail_ird)
  67                 dev->avail_ird -= ird;
  68         else
  69                 ret = -ENOMEM;
  70         spin_unlock_irq(&dev->lock);
  71
  72         if (ret)
  73                 dev_warn(&dev->rdev.lldi.pdev->dev,
  74                          "device IRD resources exhausted\n");
  75
  76         return ret;
  77 }
  78
  79 static void free_ird(struct c4iw_dev *dev, int ird)
  80 {
  81         spin_lock_irq(&dev->lock);
  82         dev->avail_ird += ird;
  83         spin_unlock_irq(&dev->lock);
  84 }
  85
  86 static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
  87 {
  88         unsigned long flag;
  89         spin_lock_irqsave(&qhp->lock, flag);
  90         qhp->attr.state = state;
  91         spin_unlock_irqrestore(&qhp->lock, flag);
  92 }
  93
  94 static void dealloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
  95 {
  96         c4iw_ocqp_pool_free(rdev, sq->dma_addr, sq->memsize);
  97 }
  98
  99 static void dealloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
 100 {
 101         dma_free_coherent(&(rdev->lldi.pdev->dev), sq->memsize, sq->queue,
 102                           pci_unmap_addr(sq, mapping));
 103 }
 104
 105 static void dealloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
 106 {
 107         if (t4_sq_onchip(sq))
 108                 dealloc_oc_sq(rdev, sq);
 109         else
 110                 dealloc_host_sq(rdev, sq);
 111 }
 112
 113 static int alloc_oc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
 114 {
 115         if (!ocqp_support || !ocqp_supported(&rdev->lldi))
 116                 return -ENOSYS;
 117         sq->dma_addr = c4iw_ocqp_pool_alloc(rdev, sq->memsize);
 118         if (!sq->dma_addr)
 119                 return -ENOMEM;
 120         sq->phys_addr = rdev->oc_mw_pa + sq->dma_addr -
 121                         rdev->lldi.vr->ocq.start;
 122         sq->queue = (__force union t4_wr *)(rdev->oc_mw_kva + sq->dma_addr -
 123                                             rdev->lldi.vr->ocq.start);
 124         sq->flags |= T4_SQ_ONCHIP;
 125         return 0;
 126 }
 127
 128 static int alloc_host_sq(struct c4iw_rdev *rdev, struct t4_sq *sq)
 129 {
 130         sq->queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev), sq->memsize,
 131                                        &(sq->dma_addr), GFP_KERNEL);
 132         if (!sq->queue)
 133                 return -ENOMEM;
 134         sq->phys_addr = virt_to_phys(sq->queue);
 135         pci_unmap_addr_set(sq, mapping, sq->dma_addr);
 136         return 0;
 137 }
 138
 139 static int alloc_sq(struct c4iw_rdev *rdev, struct t4_sq *sq, int user)
 140 {
 141         int ret = -ENOSYS;
 142         if (user)
 143                 ret = alloc_oc_sq(rdev, sq);
 144         if (ret)
 145                 ret = alloc_host_sq(rdev, sq);
 146         return ret;
 147 }
 148
 149 static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
 150                       struct c4iw_dev_ucontext *uctx)
 151 {
 152         /*
 153          * uP clears EQ contexts when the connection exits rdma mode,
 154          * so no need to post a RESET WR for these EQs.
 155          */
 156         dma_free_coherent(&(rdev->lldi.pdev->dev),
 157                           wq->rq.memsize, wq->rq.queue,
 158                           dma_unmap_addr(&wq->rq, mapping));
 159         dealloc_sq(rdev, &wq->sq);
 160         c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
 161         kfree(wq->rq.sw_rq);
 162         kfree(wq->sq.sw_sq);
 163         c4iw_put_qpid(rdev, wq->rq.qid, uctx);
 164         c4iw_put_qpid(rdev, wq->sq.qid, uctx);
 165         return 0;
 166 }
 167
 168 /*
 169  * Determine the BAR2 virtual address and qid. If pbar2_pa is not NULL,
 170  * then this is a user mapping so compute the page-aligned physical address
 171  * for mapping.
 172  */
 173 void __iomem *c4iw_bar2_addrs(struct c4iw_rdev *rdev, unsigned int qid,
 174                               enum cxgb4_bar2_qtype qtype,
 175                               unsigned int *pbar2_qid, u64 *pbar2_pa)
 176 {
 177         u64 bar2_qoffset;
 178         int ret;
 179
 180         ret = cxgb4_bar2_sge_qregs(rdev->lldi.ports[0], qid, qtype,
 181                                    pbar2_pa ? 1 : 0,
 182                                    &bar2_qoffset, pbar2_qid);
 183         if (ret)
 184                 return NULL;
 185
 186         if (pbar2_pa)
 187                 *pbar2_pa = (rdev->bar2_pa + bar2_qoffset) & PAGE_MASK;
 188
 189         if (is_t4(rdev->lldi.adapter_type))
 190                 return NULL;
 191
 192         return rdev->bar2_kva + bar2_qoffset;
 193 }
 194
 195 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
 196                      struct t4_cq *rcq, struct t4_cq *scq,
 197                      struct c4iw_dev_ucontext *uctx)
 198 {
 199         int user = (uctx != &rdev->uctx);
 200         struct fw_ri_res_wr *res_wr;
 201         struct fw_ri_res *res;
 202         int wr_len;
 203         struct c4iw_wr_wait wr_wait;
 204         struct sk_buff *skb;
 205         int ret = 0;
 206         int eqsize;
 207
 208         wq->sq.qid = c4iw_get_qpid(rdev, uctx);
 209         if (!wq->sq.qid)
 210                 return -ENOMEM;
 211
 212         wq->rq.qid = c4iw_get_qpid(rdev, uctx);
 213         if (!wq->rq.qid) {
 214                 ret = -ENOMEM;
 215                 goto free_sq_qid;
 216         }
 217
 218         if (!user) {
 219                 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
 220                                  GFP_KERNEL);
 221                 if (!wq->sq.sw_sq) {
 222                         ret = -ENOMEM;
 223                         goto free_rq_qid;
 224                 }
 225
 226                 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
 227                                  GFP_KERNEL);
 228                 if (!wq->rq.sw_rq) {
 229                         ret = -ENOMEM;
 230                         goto free_sw_sq;
 231                 }
 232         }
 233
 234         /*
 235          * RQT must be a power of 2 and at least 16 deep.
 236          */
 237         wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
 238         wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
 239         if (!wq->rq.rqt_hwaddr) {
 240                 ret = -ENOMEM;
 241                 goto free_sw_rq;
 242         }
 243
 244         ret = alloc_sq(rdev, &wq->sq, user);
 245         if (ret)
 246                 goto free_hwaddr;
 247         memset(wq->sq.queue, 0, wq->sq.memsize);
 248         dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
 249
 250         wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
 251                                           wq->rq.memsize, &(wq->rq.dma_addr),
 252                                           GFP_KERNEL);
 253         if (!wq->rq.queue) {
 254                 ret = -ENOMEM;
 255                 goto free_sq;
 256         }
 257         PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
 258                 __func__, wq->sq.queue,
 259                 (unsigned long long)virt_to_phys(wq->sq.queue),
 260                 wq->rq.queue,
 261                 (unsigned long long)virt_to_phys(wq->rq.queue));
 262         memset(wq->rq.queue, 0, wq->rq.memsize);
 263         dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
 264
 265         wq->db = rdev->lldi.db_reg;
 266
 267         wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
 268                                          &wq->sq.bar2_qid,
 269                                          user ? &wq->sq.bar2_pa : NULL);
 270         wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
 271                                          &wq->rq.bar2_qid,
 272                                          user ? &wq->rq.bar2_pa : NULL);
 273
 274         /*
 275          * User mode must have bar2 access.
 276          */
 277         if (user && (!wq->sq.bar2_pa || !wq->rq.bar2_pa)) {
 278                 pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
 279                         pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
 280                 goto free_dma;
 281         }
 282
 283         wq->rdev = rdev;
 284         wq->rq.msn = 1;
 285
 286         /* build fw_ri_res_wr */
 287         wr_len = sizeof *res_wr + 2 * sizeof *res;
 288
 289         skb = alloc_skb(wr_len, GFP_KERNEL);
 290         if (!skb) {
 291                 ret = -ENOMEM;
 292                 goto free_dma;
 293         }
 294         set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
 295
 296         res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
 297         memset(res_wr, 0, wr_len);
 298         res_wr->op_nres = cpu_to_be32(
 299                         FW_WR_OP_V(FW_RI_RES_WR) |
 300                         FW_RI_RES_WR_NRES_V(2) |
 301                         FW_WR_COMPL_F);
 302         res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
 303         res_wr->cookie = (uintptr_t)&wr_wait;
 304         res = res_wr->res;
 305         res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
 306         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 307
 308         /*
 309          * eqsize is the number of 64B entries plus the status page size.
 310          */
 311         eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
 312                 rdev->hw_queue.t4_eq_status_entries;
 313
 314         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 315                 FW_RI_RES_WR_HOSTFCMODE_V(0) |  /* no host cidx updates */
 316                 FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
 317                 FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
 318                 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
 319                 FW_RI_RES_WR_IQID_V(scq->cqid));
 320         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 321                 FW_RI_RES_WR_DCAEN_V(0) |
 322                 FW_RI_RES_WR_DCACPU_V(0) |
 323                 FW_RI_RES_WR_FBMIN_V(2) |
 324                 FW_RI_RES_WR_FBMAX_V(2) |
 325                 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
 326                 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
 327                 FW_RI_RES_WR_EQSIZE_V(eqsize));
 328         res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
 329         res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
 330         res++;
 331         res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
 332         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 333
 334         /*
 335          * eqsize is the number of 64B entries plus the status page size.
 336          */
 337         eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
 338                 rdev->hw_queue.t4_eq_status_entries;
 339         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 340                 FW_RI_RES_WR_HOSTFCMODE_V(0) |  /* no host cidx updates */
 341                 FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
 342                 FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
 343                 FW_RI_RES_WR_IQID_V(rcq->cqid));
 344         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 345                 FW_RI_RES_WR_DCAEN_V(0) |
 346                 FW_RI_RES_WR_DCACPU_V(0) |
 347                 FW_RI_RES_WR_FBMIN_V(2) |
 348                 FW_RI_RES_WR_FBMAX_V(2) |
 349                 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
 350                 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
 351                 FW_RI_RES_WR_EQSIZE_V(eqsize));
 352         res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
 353         res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
 354
 355         c4iw_init_wr_wait(&wr_wait);
 356
 357         ret = c4iw_ofld_send(rdev, skb);
 358         if (ret)
 359                 goto free_dma;
 360         ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
 361         if (ret)
 362                 goto free_dma;
 363
 364         PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
 365              __func__, wq->sq.qid, wq->rq.qid, wq->db,
 366              wq->sq.bar2_va, wq->rq.bar2_va);
 367
 368         return 0;
 369 free_dma:
 370         dma_free_coherent(&(rdev->lldi.pdev->dev),
 371                           wq->rq.memsize, wq->rq.queue,
 372                           dma_unmap_addr(&wq->rq, mapping));
 373 free_sq:
 374         dealloc_sq(rdev, &wq->sq);
 375 free_hwaddr:
 376         c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
 377 free_sw_rq:
 378         kfree(wq->rq.sw_rq);
 379 free_sw_sq:
 380         kfree(wq->sq.sw_sq);
 381 free_rq_qid:
 382         c4iw_put_qpid(rdev, wq->rq.qid, uctx);
 383 free_sq_qid:
 384         c4iw_put_qpid(rdev, wq->sq.qid, uctx);
 385         return ret;
 386 }
 387
 388 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
 389                       struct ib_send_wr *wr, int max, u32 *plenp)
 390 {
 391         u8 *dstp, *srcp;
 392         u32 plen = 0;
 393         int i;
 394         int rem, len;
 395
 396         dstp = (u8 *)immdp->data;
 397         for (i = 0; i < wr->num_sge; i++) {
 398                 if ((plen + wr->sg_list[i].length) > max)
 399                         return -EMSGSIZE;
 400                 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
 401                 plen += wr->sg_list[i].length;
 402                 rem = wr->sg_list[i].length;
 403                 while (rem) {
 404                         if (dstp == (u8 *)&sq->queue[sq->size])
 405                                 dstp = (u8 *)sq->queue;
 406                         if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
 407                                 len = rem;
 408                         else
 409                                 len = (u8 *)&sq->queue[sq->size] - dstp;
 410                         memcpy(dstp, srcp, len);
 411                         dstp += len;
 412                         srcp += len;
 413                         rem -= len;
 414                 }
 415         }
 416         len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
 417         if (len)
 418                 memset(dstp, 0, len);
 419         immdp->op = FW_RI_DATA_IMMD;
 420         immdp->r1 = 0;
 421         immdp->r2 = 0;
 422         immdp->immdlen = cpu_to_be32(plen);
 423         *plenp = plen;
 424         return 0;
 425 }
 426
 427 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
 428                       struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
 429                       int num_sge, u32 *plenp)
 430
 431 {
 432         int i;
 433         u32 plen = 0;
 434         __be64 *flitp = (__be64 *)isglp->sge;
 435
 436         for (i = 0; i < num_sge; i++) {
 437                 if ((plen + sg_list[i].length) < plen)
 438                         return -EMSGSIZE;
 439                 plen += sg_list[i].length;
 440                 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
 441                                      sg_list[i].length);
 442                 if (++flitp == queue_end)
 443                         flitp = queue_start;
 444                 *flitp = cpu_to_be64(sg_list[i].addr);
 445                 if (++flitp == queue_end)
 446                         flitp = queue_start;
 447         }
 448         *flitp = (__force __be64)0;
 449         isglp->op = FW_RI_DATA_ISGL;
 450         isglp->r1 = 0;
 451         isglp->nsge = cpu_to_be16(num_sge);
 452         isglp->r2 = 0;
 453         if (plenp)
 454                 *plenp = plen;
 455         return 0;
 456 }
 457
 458 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
 459                            struct ib_send_wr *wr, u8 *len16)
 460 {
 461         u32 plen;
 462         int size;
 463         int ret;
 464
 465         if (wr->num_sge > T4_MAX_SEND_SGE)
 466                 return -EINVAL;
 467         switch (wr->opcode) {
 468         case IB_WR_SEND:
 469                 if (wr->send_flags & IB_SEND_SOLICITED)
 470                         wqe->send.sendop_pkd = cpu_to_be32(
 471                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
 472                 else
 473                         wqe->send.sendop_pkd = cpu_to_be32(
 474                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
 475                 wqe->send.stag_inv = 0;
 476                 break;
 477         case IB_WR_SEND_WITH_INV:
 478                 if (wr->send_flags & IB_SEND_SOLICITED)
 479                         wqe->send.sendop_pkd = cpu_to_be32(
 480                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
 481                 else
 482                         wqe->send.sendop_pkd = cpu_to_be32(
 483                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
 484                 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 485                 break;
 486
 487         default:
 488                 return -EINVAL;
 489         }
 490         wqe->send.r3 = 0;
 491         wqe->send.r4 = 0;
 492
 493         plen = 0;
 494         if (wr->num_sge) {
 495                 if (wr->send_flags & IB_SEND_INLINE) {
 496                         ret = build_immd(sq, wqe->send.u.immd_src, wr,
 497                                          T4_MAX_SEND_INLINE, &plen);
 498                         if (ret)
 499                                 return ret;
 500                         size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
 501                                plen;
 502                 } else {
 503                         ret = build_isgl((__be64 *)sq->queue,
 504                                          (__be64 *)&sq->queue[sq->size],
 505                                          wqe->send.u.isgl_src,
 506                                          wr->sg_list, wr->num_sge, &plen);
 507                         if (ret)
 508                                 return ret;
 509                         size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
 510                                wr->num_sge * sizeof(struct fw_ri_sge);
 511                 }
 512         } else {
 513                 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
 514                 wqe->send.u.immd_src[0].r1 = 0;
 515                 wqe->send.u.immd_src[0].r2 = 0;
 516                 wqe->send.u.immd_src[0].immdlen = 0;
 517                 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
 518                 plen = 0;
 519         }
 520         *len16 = DIV_ROUND_UP(size, 16);
 521         wqe->send.plen = cpu_to_be32(plen);
 522         return 0;
 523 }
 524
 525 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
 526                             struct ib_send_wr *wr, u8 *len16)
 527 {
 528         u32 plen;
 529         int size;
 530         int ret;
 531
 532         if (wr->num_sge > T4_MAX_SEND_SGE)
 533                 return -EINVAL;
 534         wqe->write.r2 = 0;
 535         wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
 536         wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
 537         if (wr->num_sge) {
 538                 if (wr->send_flags & IB_SEND_INLINE) {
 539                         ret = build_immd(sq, wqe->write.u.immd_src, wr,
 540                                          T4_MAX_WRITE_INLINE, &plen);
 541                         if (ret)
 542                                 return ret;
 543                         size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
 544                                plen;
 545                 } else {
 546                         ret = build_isgl((__be64 *)sq->queue,
 547                                          (__be64 *)&sq->queue[sq->size],
 548                                          wqe->write.u.isgl_src,
 549                                          wr->sg_list, wr->num_sge, &plen);
 550                         if (ret)
 551                                 return ret;
 552                         size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
 553                                wr->num_sge * sizeof(struct fw_ri_sge);
 554                 }
 555         } else {
 556                 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
 557                 wqe->write.u.immd_src[0].r1 = 0;
 558                 wqe->write.u.immd_src[0].r2 = 0;
 559                 wqe->write.u.immd_src[0].immdlen = 0;
 560                 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
 561                 plen = 0;
 562         }
 563         *len16 = DIV_ROUND_UP(size, 16);
 564         wqe->write.plen = cpu_to_be32(plen);
 565         return 0;
 566 }
 567
 568 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
 569 {
 570         if (wr->num_sge > 1)
 571                 return -EINVAL;
 572         if (wr->num_sge) {
 573                 wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey);
 574                 wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr
 575                                                         >> 32));
 576                 wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr);
 577                 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
 578                 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
 579                 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
 580                                                          >> 32));
 581                 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
 582         } else {
 583                 wqe->read.stag_src = cpu_to_be32(2);
 584                 wqe->read.to_src_hi = 0;
 585                 wqe->read.to_src_lo = 0;
 586                 wqe->read.stag_sink = cpu_to_be32(2);
 587                 wqe->read.plen = 0;
 588                 wqe->read.to_sink_hi = 0;
 589                 wqe->read.to_sink_lo = 0;
 590         }
 591         wqe->read.r2 = 0;
 592         wqe->read.r5 = 0;
 593         *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
 594         return 0;
 595 }
 596
 597 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
 598                            struct ib_recv_wr *wr, u8 *len16)
 599 {
 600         int ret;
 601
 602         ret = build_isgl((__be64 *)qhp->wq.rq.queue,
 603                          (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
 604                          &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
 605         if (ret)
 606                 return ret;
 607         *len16 = DIV_ROUND_UP(sizeof wqe->recv +
 608                               wr->num_sge * sizeof(struct fw_ri_sge), 16);
 609         return 0;
 610 }
 611
 612 static int build_memreg(struct t4_sq *sq, union t4_wr *wqe,
 613                         struct ib_reg_wr *wr, u8 *len16, u8 t5dev)
 614 {
 615         struct c4iw_mr *mhp = to_c4iw_mr(wr->mr);
 616         struct fw_ri_immd *imdp;
 617         __be64 *p;
 618         int i;
 619         int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32);
 620         int rem;
 621
 622         if (mhp->mpl_len > t4_max_fr_depth(use_dsgl))
 623                 return -EINVAL;
 624
 625         wqe->fr.qpbinde_to_dcacpu = 0;
 626         wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12;
 627         wqe->fr.addr_type = FW_RI_VA_BASED_TO;
 628         wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access);
 629         wqe->fr.len_hi = 0;
 630         wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length);
 631         wqe->fr.stag = cpu_to_be32(wr->key);
 632         wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32);
 633         wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova &
 634                                         0xffffffff);
 635
 636         if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
 637                 struct fw_ri_dsgl *sglp;
 638
 639                 for (i = 0; i < mhp->mpl_len; i++)
 640                         mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]);
 641
 642                 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
 643                 sglp->op = FW_RI_DATA_DSGL;
 644                 sglp->r1 = 0;
 645                 sglp->nsge = cpu_to_be16(1);
 646                 sglp->addr0 = cpu_to_be64(mhp->mpl_addr);
 647                 sglp->len0 = cpu_to_be32(pbllen);
 648
 649                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
 650         } else {
 651                 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
 652                 imdp->op = FW_RI_DATA_IMMD;
 653                 imdp->r1 = 0;
 654                 imdp->r2 = 0;
 655                 imdp->immdlen = cpu_to_be32(pbllen);
 656                 p = (__be64 *)(imdp + 1);
 657                 rem = pbllen;
 658                 for (i = 0; i < mhp->mpl_len; i++) {
 659                         *p = cpu_to_be64((u64)mhp->mpl[i]);
 660                         rem -= sizeof(*p);
 661                         if (++p == (__be64 *)&sq->queue[sq->size])
 662                                 p = (__be64 *)sq->queue;
 663                 }
 664                 BUG_ON(rem < 0);
 665                 while (rem) {
 666                         *p = 0;
 667                         rem -= sizeof(*p);
 668                         if (++p == (__be64 *)&sq->queue[sq->size])
 669                                 p = (__be64 *)sq->queue;
 670                 }
 671                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
 672                                       + pbllen, 16);
 673         }
 674         return 0;
 675 }
 676
 677 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
 678                           u8 *len16)
 679 {
 680         wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 681         wqe->inv.r2 = 0;
 682         *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
 683         return 0;
 684 }
 685
 686 void c4iw_qp_add_ref(struct ib_qp *qp)
 687 {
 688         PDBG("%s ib_qp %p\n", __func__, qp);
 689         atomic_inc(&(to_c4iw_qp(qp)->refcnt));
 690 }
 691
 692 void c4iw_qp_rem_ref(struct ib_qp *qp)
 693 {
 694         PDBG("%s ib_qp %p\n", __func__, qp);
 695         if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
 696                 wake_up(&(to_c4iw_qp(qp)->wait));
 697 }
 698
 699 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
 700 {
 701         if (list_empty(entry))
 702                 list_add_tail(entry, head);
 703 }
 704
 705 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
 706 {
 707         unsigned long flags;
 708
 709         spin_lock_irqsave(&qhp->rhp->lock, flags);
 710         spin_lock(&qhp->lock);
 711         if (qhp->rhp->db_state == NORMAL)
 712                 t4_ring_sq_db(&qhp->wq, inc, NULL);
 713         else {
 714                 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
 715                 qhp->wq.sq.wq_pidx_inc += inc;
 716         }
 717         spin_unlock(&qhp->lock);
 718         spin_unlock_irqrestore(&qhp->rhp->lock, flags);
 719         return 0;
 720 }
 721
 722 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
 723 {
 724         unsigned long flags;
 725
 726         spin_lock_irqsave(&qhp->rhp->lock, flags);
 727         spin_lock(&qhp->lock);
 728         if (qhp->rhp->db_state == NORMAL)
 729                 t4_ring_rq_db(&qhp->wq, inc, NULL);
 730         else {
 731                 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
 732                 qhp->wq.rq.wq_pidx_inc += inc;
 733         }
 734         spin_unlock(&qhp->lock);
 735         spin_unlock_irqrestore(&qhp->rhp->lock, flags);
 736         return 0;
 737 }
 738
 739 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 740                    struct ib_send_wr **bad_wr)
 741 {
 742         int err = 0;
 743         u8 len16 = 0;
 744         enum fw_wr_opcodes fw_opcode = 0;
 745         enum fw_ri_wr_flags fw_flags;
 746         struct c4iw_qp *qhp;
 747         union t4_wr *wqe = NULL;
 748         u32 num_wrs;
 749         struct t4_swsqe *swsqe;
 750         unsigned long flag;
 751         u16 idx = 0;
 752
 753         qhp = to_c4iw_qp(ibqp);
 754         spin_lock_irqsave(&qhp->lock, flag);
 755         if (t4_wq_in_error(&qhp->wq)) {
 756                 spin_unlock_irqrestore(&qhp->lock, flag);
 757                 return -EINVAL;
 758         }
 759         num_wrs = t4_sq_avail(&qhp->wq);
 760         if (num_wrs == 0) {
 761                 spin_unlock_irqrestore(&qhp->lock, flag);
 762                 return -ENOMEM;
 763         }
 764         while (wr) {
 765                 if (num_wrs == 0) {
 766                         err = -ENOMEM;
 767                         *bad_wr = wr;
 768                         break;
 769                 }
 770                 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
 771                       qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
 772
 773                 fw_flags = 0;
 774                 if (wr->send_flags & IB_SEND_SOLICITED)
 775                         fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
 776                 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
 777                         fw_flags |= FW_RI_COMPLETION_FLAG;
 778                 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
 779                 switch (wr->opcode) {
 780                 case IB_WR_SEND_WITH_INV:
 781                 case IB_WR_SEND:
 782                         if (wr->send_flags & IB_SEND_FENCE)
 783                                 fw_flags |= FW_RI_READ_FENCE_FLAG;
 784                         fw_opcode = FW_RI_SEND_WR;
 785                         if (wr->opcode == IB_WR_SEND)
 786                                 swsqe->opcode = FW_RI_SEND;
 787                         else
 788                                 swsqe->opcode = FW_RI_SEND_WITH_INV;
 789                         err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
 790                         break;
 791                 case IB_WR_RDMA_WRITE:
 792                         fw_opcode = FW_RI_RDMA_WRITE_WR;
 793                         swsqe->opcode = FW_RI_RDMA_WRITE;
 794                         err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
 795                         break;
 796                 case IB_WR_RDMA_READ:
 797                 case IB_WR_RDMA_READ_WITH_INV:
 798                         fw_opcode = FW_RI_RDMA_READ_WR;
 799                         swsqe->opcode = FW_RI_READ_REQ;
 800                         if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
 801                                 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
 802                         else
 803                                 fw_flags = 0;
 804                         err = build_rdma_read(wqe, wr, &len16);
 805                         if (err)
 806                                 break;
 807                         swsqe->read_len = wr->sg_list[0].length;
 808                         if (!qhp->wq.sq.oldest_read)
 809                                 qhp->wq.sq.oldest_read = swsqe;
 810                         break;
 811                 case IB_WR_REG_MR:
 812                         fw_opcode = FW_RI_FR_NSMR_WR;
 813                         swsqe->opcode = FW_RI_FAST_REGISTER;
 814                         err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr), &len16,
 815                                            is_t5(
 816                                            qhp->rhp->rdev.lldi.adapter_type) ?
 817                                            1 : 0);
 818                         break;
 819                 case IB_WR_LOCAL_INV:
 820                         if (wr->send_flags & IB_SEND_FENCE)
 821                                 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
 822                         fw_opcode = FW_RI_INV_LSTAG_WR;
 823                         swsqe->opcode = FW_RI_LOCAL_INV;
 824                         err = build_inv_stag(wqe, wr, &len16);
 825                         break;
 826                 default:
 827                         PDBG("%s post of type=%d TBD!\n", __func__,
 828                              wr->opcode);
 829                         err = -EINVAL;
 830                 }
 831                 if (err) {
 832                         *bad_wr = wr;
 833                         break;
 834                 }
 835                 swsqe->idx = qhp->wq.sq.pidx;
 836                 swsqe->complete = 0;
 837                 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
 838                                   qhp->sq_sig_all;
 839                 swsqe->flushed = 0;
 840                 swsqe->wr_id = wr->wr_id;
 841                 if (c4iw_wr_log) {
 842                         swsqe->sge_ts = cxgb4_read_sge_timestamp(
 843                                         qhp->rhp->rdev.lldi.ports[0]);
 844                         getnstimeofday(&swsqe->host_ts);
 845                 }
 846
 847                 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
 848
 849                 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
 850                      __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
 851                      swsqe->opcode, swsqe->read_len);
 852                 wr = wr->next;
 853                 num_wrs--;
 854                 t4_sq_produce(&qhp->wq, len16);
 855                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 856         }
 857         if (!qhp->rhp->rdev.status_page->db_off) {
 858                 t4_ring_sq_db(&qhp->wq, idx, wqe);
 859                 spin_unlock_irqrestore(&qhp->lock, flag);
 860         } else {
 861                 spin_unlock_irqrestore(&qhp->lock, flag);
 862                 ring_kernel_sq_db(qhp, idx);
 863         }
 864         return err;
 865 }
 866
 867 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 868                       struct ib_recv_wr **bad_wr)
 869 {
 870         int err = 0;
 871         struct c4iw_qp *qhp;
 872         union t4_recv_wr *wqe = NULL;
 873         u32 num_wrs;
 874         u8 len16 = 0;
 875         unsigned long flag;
 876         u16 idx = 0;
 877
 878         qhp = to_c4iw_qp(ibqp);
 879         spin_lock_irqsave(&qhp->lock, flag);
 880         if (t4_wq_in_error(&qhp->wq)) {
 881                 spin_unlock_irqrestore(&qhp->lock, flag);
 882                 return -EINVAL;
 883         }
 884         num_wrs = t4_rq_avail(&qhp->wq);
 885         if (num_wrs == 0) {
 886                 spin_unlock_irqrestore(&qhp->lock, flag);
 887                 return -ENOMEM;
 888         }
 889         while (wr) {
 890                 if (wr->num_sge > T4_MAX_RECV_SGE) {
 891                         err = -EINVAL;
 892                         *bad_wr = wr;
 893                         break;
 894                 }
 895                 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
 896                                            qhp->wq.rq.wq_pidx *
 897                                            T4_EQ_ENTRY_SIZE);
 898                 if (num_wrs)
 899                         err = build_rdma_recv(qhp, wqe, wr, &len16);
 900                 else
 901                         err = -ENOMEM;
 902                 if (err) {
 903                         *bad_wr = wr;
 904                         break;
 905                 }
 906
 907                 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
 908                 if (c4iw_wr_log) {
 909                         qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
 910                                 cxgb4_read_sge_timestamp(
 911                                                 qhp->rhp->rdev.lldi.ports[0]);
 912                         getnstimeofday(
 913                                 &qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
 914                 }
 915
 916                 wqe->recv.opcode = FW_RI_RECV_WR;
 917                 wqe->recv.r1 = 0;
 918                 wqe->recv.wrid = qhp->wq.rq.pidx;
 919                 wqe->recv.r2[0] = 0;
 920                 wqe->recv.r2[1] = 0;
 921                 wqe->recv.r2[2] = 0;
 922                 wqe->recv.len16 = len16;
 923                 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
 924                      (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
 925                 t4_rq_produce(&qhp->wq, len16);
 926                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 927                 wr = wr->next;
 928                 num_wrs--;
 929         }
 930         if (!qhp->rhp->rdev.status_page->db_off) {
 931                 t4_ring_rq_db(&qhp->wq, idx, wqe);
 932                 spin_unlock_irqrestore(&qhp->lock, flag);
 933         } else {
 934                 spin_unlock_irqrestore(&qhp->lock, flag);
 935                 ring_kernel_rq_db(qhp, idx);
 936         }
 937         return err;
 938 }
 939
 940 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
 941 {
 942         return -ENOSYS;
 943 }
 944
 945 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
 946                                     u8 *ecode)
 947 {
 948         int status;
 949         int tagged;
 950         int opcode;
 951         int rqtype;
 952         int send_inv;
 953
 954         if (!err_cqe) {
 955                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 956                 *ecode = 0;
 957                 return;
 958         }
 959
 960         status = CQE_STATUS(err_cqe);
 961         opcode = CQE_OPCODE(err_cqe);
 962         rqtype = RQ_TYPE(err_cqe);
 963         send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
 964                    (opcode == FW_RI_SEND_WITH_SE_INV);
 965         tagged = (opcode == FW_RI_RDMA_WRITE) ||
 966                  (rqtype && (opcode == FW_RI_READ_RESP));
 967
 968         switch (status) {
 969         case T4_ERR_STAG:
 970                 if (send_inv) {
 971                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 972                         *ecode = RDMAP_CANT_INV_STAG;
 973                 } else {
 974                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 975                         *ecode = RDMAP_INV_STAG;
 976                 }
 977                 break;
 978         case T4_ERR_PDID:
 979                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 980                 if ((opcode == FW_RI_SEND_WITH_INV) ||
 981                     (opcode == FW_RI_SEND_WITH_SE_INV))
 982                         *ecode = RDMAP_CANT_INV_STAG;
 983                 else
 984                         *ecode = RDMAP_STAG_NOT_ASSOC;
 985                 break;
 986         case T4_ERR_QPID:
 987                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 988                 *ecode = RDMAP_STAG_NOT_ASSOC;
 989                 break;
 990         case T4_ERR_ACCESS:
 991                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 992                 *ecode = RDMAP_ACC_VIOL;
 993                 break;
 994         case T4_ERR_WRAP:
 995                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 996                 *ecode = RDMAP_TO_WRAP;
 997                 break;
 998         case T4_ERR_BOUND:
 999                 if (tagged) {
1000                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1001                         *ecode = DDPT_BASE_BOUNDS;
1002                 } else {
1003                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1004                         *ecode = RDMAP_BASE_BOUNDS;
1005                 }
1006                 break;
1007         case T4_ERR_INVALIDATE_SHARED_MR:
1008         case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1009                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1010                 *ecode = RDMAP_CANT_INV_STAG;
1011                 break;
1012         case T4_ERR_ECC:
1013         case T4_ERR_ECC_PSTAG:
1014         case T4_ERR_INTERNAL_ERR:
1015                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1016                 *ecode = 0;
1017                 break;
1018         case T4_ERR_OUT_OF_RQE:
1019                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1020                 *ecode = DDPU_INV_MSN_NOBUF;
1021                 break;
1022         case T4_ERR_PBL_ADDR_BOUND:
1023                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1024                 *ecode = DDPT_BASE_BOUNDS;
1025                 break;
1026         case T4_ERR_CRC:
1027                 *layer_type = LAYER_MPA|DDP_LLP;
1028                 *ecode = MPA_CRC_ERR;
1029                 break;
1030         case T4_ERR_MARKER:
1031                 *layer_type = LAYER_MPA|DDP_LLP;
1032                 *ecode = MPA_MARKER_ERR;
1033                 break;
1034         case T4_ERR_PDU_LEN_ERR:
1035                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1036                 *ecode = DDPU_MSG_TOOBIG;
1037                 break;
1038         case T4_ERR_DDP_VERSION:
1039                 if (tagged) {
1040                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1041                         *ecode = DDPT_INV_VERS;
1042                 } else {
1043                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1044                         *ecode = DDPU_INV_VERS;
1045                 }
1046                 break;
1047         case T4_ERR_RDMA_VERSION:
1048                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1049                 *ecode = RDMAP_INV_VERS;
1050                 break;
1051         case T4_ERR_OPCODE:
1052                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1053                 *ecode = RDMAP_INV_OPCODE;
1054                 break;
1055         case T4_ERR_DDP_QUEUE_NUM:
1056                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1057                 *ecode = DDPU_INV_QN;
1058                 break;
1059         case T4_ERR_MSN:
1060         case T4_ERR_MSN_GAP:
1061         case T4_ERR_MSN_RANGE:
1062         case T4_ERR_IRD_OVERFLOW:
1063                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1064                 *ecode = DDPU_INV_MSN_RANGE;
1065                 break;
1066         case T4_ERR_TBIT:
1067                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1068                 *ecode = 0;
1069                 break;
1070         case T4_ERR_MO:
1071                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1072                 *ecode = DDPU_INV_MO;
1073                 break;
1074         default:
1075                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1076                 *ecode = 0;
1077                 break;
1078         }
1079 }
1080
1081 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1082                            gfp_t gfp)
1083 {
1084         struct fw_ri_wr *wqe;
1085         struct sk_buff *skb;
1086         struct terminate_message *term;
1087
1088         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1089              qhp->ep->hwtid);
1090
1091         skb = alloc_skb(sizeof *wqe, gfp);
1092         if (!skb)
1093                 return;
1094         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1095
1096         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1097         memset(wqe, 0, sizeof *wqe);
1098         wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1099         wqe->flowid_len16 = cpu_to_be32(
1100                 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1101                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1102
1103         wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1104         wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1105         term = (struct terminate_message *)wqe->u.terminate.termmsg;
1106         if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1107                 term->layer_etype = qhp->attr.layer_etype;
1108                 term->ecode = qhp->attr.ecode;
1109         } else
1110                 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1111         c4iw_ofld_send(&qhp->rhp->rdev, skb);
1112 }
1113
1114 /*
1115  * Assumes qhp lock is held.
1116  */
1117 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1118                        struct c4iw_cq *schp)
1119 {
1120         int count;
1121         int rq_flushed, sq_flushed;
1122         unsigned long flag;
1123
1124         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1125
1126         /* locking hierarchy: cq lock first, then qp lock. */
1127         spin_lock_irqsave(&rchp->lock, flag);
1128         spin_lock(&qhp->lock);
1129
1130         if (qhp->wq.flushed) {
1131                 spin_unlock(&qhp->lock);
1132                 spin_unlock_irqrestore(&rchp->lock, flag);
1133                 return;
1134         }
1135         qhp->wq.flushed = 1;
1136
1137         c4iw_flush_hw_cq(rchp);
1138         c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1139         rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1140         spin_unlock(&qhp->lock);
1141         spin_unlock_irqrestore(&rchp->lock, flag);
1142
1143         /* locking hierarchy: cq lock first, then qp lock. */
1144         spin_lock_irqsave(&schp->lock, flag);
1145         spin_lock(&qhp->lock);
1146         if (schp != rchp)
1147                 c4iw_flush_hw_cq(schp);
1148         sq_flushed = c4iw_flush_sq(qhp);
1149         spin_unlock(&qhp->lock);
1150         spin_unlock_irqrestore(&schp->lock, flag);
1151
1152         if (schp == rchp) {
1153                 if (t4_clear_cq_armed(&rchp->cq) &&
1154                     (rq_flushed || sq_flushed)) {
1155                         spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1156                         (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1157                                                    rchp->ibcq.cq_context);
1158                         spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1159                 }
1160         } else {
1161                 if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1162                         spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1163                         (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1164                                                    rchp->ibcq.cq_context);
1165                         spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1166                 }
1167                 if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1168                         spin_lock_irqsave(&schp->comp_handler_lock, flag);
1169                         (*schp->ibcq.comp_handler)(&schp->ibcq,
1170                                                    schp->ibcq.cq_context);
1171                         spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1172                 }
1173         }
1174 }
1175
1176 static void flush_qp(struct c4iw_qp *qhp)
1177 {
1178         struct c4iw_cq *rchp, *schp;
1179         unsigned long flag;
1180
1181         rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1182         schp = to_c4iw_cq(qhp->ibqp.send_cq);
1183
1184         t4_set_wq_in_error(&qhp->wq);
1185         if (qhp->ibqp.uobject) {
1186                 t4_set_cq_in_error(&rchp->cq);
1187                 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1188                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1189                 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1190                 if (schp != rchp) {
1191                         t4_set_cq_in_error(&schp->cq);
1192                         spin_lock_irqsave(&schp->comp_handler_lock, flag);
1193                         (*schp->ibcq.comp_handler)(&schp->ibcq,
1194                                         schp->ibcq.cq_context);
1195                         spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1196                 }
1197                 return;
1198         }
1199         __flush_qp(qhp, rchp, schp);
1200 }
1201
1202 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1203                      struct c4iw_ep *ep)
1204 {
1205         struct fw_ri_wr *wqe;
1206         int ret;
1207         struct sk_buff *skb;
1208
1209         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1210              ep->hwtid);
1211
1212         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1213         if (!skb)
1214                 return -ENOMEM;
1215         set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1216
1217         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1218         memset(wqe, 0, sizeof *wqe);
1219         wqe->op_compl = cpu_to_be32(
1220                 FW_WR_OP_V(FW_RI_INIT_WR) |
1221                 FW_WR_COMPL_F);
1222         wqe->flowid_len16 = cpu_to_be32(
1223                 FW_WR_FLOWID_V(ep->hwtid) |
1224                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1225         wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1226
1227         wqe->u.fini.type = FW_RI_TYPE_FINI;
1228         ret = c4iw_ofld_send(&rhp->rdev, skb);
1229         if (ret)
1230                 goto out;
1231
1232         ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1233                              qhp->wq.sq.qid, __func__);
1234 out:
1235         PDBG("%s ret %d\n", __func__, ret);
1236         return ret;
1237 }
1238
1239 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1240 {
1241         PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1242         memset(&init->u, 0, sizeof init->u);
1243         switch (p2p_type) {
1244         case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1245                 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1246                 init->u.write.stag_sink = cpu_to_be32(1);
1247                 init->u.write.to_sink = cpu_to_be64(1);
1248                 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1249                 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1250                                                    sizeof(struct fw_ri_immd),
1251                                                    16);
1252                 break;
1253         case FW_RI_INIT_P2PTYPE_READ_REQ:
1254                 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1255                 init->u.read.stag_src = cpu_to_be32(1);
1256                 init->u.read.to_src_lo = cpu_to_be32(1);
1257                 init->u.read.stag_sink = cpu_to_be32(1);
1258                 init->u.read.to_sink_lo = cpu_to_be32(1);
1259                 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1260                 break;
1261         }
1262 }
1263
1264 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1265 {
1266         struct fw_ri_wr *wqe;
1267         int ret;
1268         struct sk_buff *skb;
1269
1270         PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1271              qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1272
1273         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1274         if (!skb) {
1275                 ret = -ENOMEM;
1276                 goto out;
1277         }
1278         ret = alloc_ird(rhp, qhp->attr.max_ird);
1279         if (ret) {
1280                 qhp->attr.max_ird = 0;
1281                 kfree_skb(skb);
1282                 goto out;
1283         }
1284         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1285
1286         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1287         memset(wqe, 0, sizeof *wqe);
1288         wqe->op_compl = cpu_to_be32(
1289                 FW_WR_OP_V(FW_RI_INIT_WR) |
1290                 FW_WR_COMPL_F);
1291         wqe->flowid_len16 = cpu_to_be32(
1292                 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1293                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1294
1295         wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1296
1297         wqe->u.init.type = FW_RI_TYPE_INIT;
1298         wqe->u.init.mpareqbit_p2ptype =
1299                 FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1300                 FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1301         wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1302         if (qhp->attr.mpa_attr.recv_marker_enabled)
1303                 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1304         if (qhp->attr.mpa_attr.xmit_marker_enabled)
1305                 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1306         if (qhp->attr.mpa_attr.crc_enabled)
1307                 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1308
1309         wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1310                             FW_RI_QP_RDMA_WRITE_ENABLE |
1311                             FW_RI_QP_BIND_ENABLE;
1312         if (!qhp->ibqp.uobject)
1313                 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1314                                      FW_RI_QP_STAG0_ENABLE;
1315         wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1316         wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1317         wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1318         wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1319         wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1320         wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1321         wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1322         wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1323         wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1324         wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1325         wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1326         wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1327         wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1328                                          rhp->rdev.lldi.vr->rq.start);
1329         if (qhp->attr.mpa_attr.initiator)
1330                 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1331
1332         ret = c4iw_ofld_send(&rhp->rdev, skb);
1333         if (ret)
1334                 goto err1;
1335
1336         ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1337                                   qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1338         if (!ret)
1339                 goto out;
1340 err1:
1341         free_ird(rhp, qhp->attr.max_ird);
1342 out:
1343         PDBG("%s ret %d\n", __func__, ret);
1344         return ret;
1345 }
1346
1347 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1348                    enum c4iw_qp_attr_mask mask,
1349                    struct c4iw_qp_attributes *attrs,
1350                    int internal)
1351 {
1352         int ret = 0;
1353         struct c4iw_qp_attributes newattr = qhp->attr;
1354         int disconnect = 0;
1355         int terminate = 0;
1356         int abort = 0;
1357         int free = 0;
1358         struct c4iw_ep *ep = NULL;
1359
1360         PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1361              qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1362              (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1363
1364         mutex_lock(&qhp->mutex);
1365
1366         /* Process attr changes if in IDLE */
1367         if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1368                 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1369                         ret = -EIO;
1370                         goto out;
1371                 }
1372                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1373                         newattr.enable_rdma_read = attrs->enable_rdma_read;
1374                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1375                         newattr.enable_rdma_write = attrs->enable_rdma_write;
1376                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1377                         newattr.enable_bind = attrs->enable_bind;
1378                 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1379                         if (attrs->max_ord > c4iw_max_read_depth) {
1380                                 ret = -EINVAL;
1381                                 goto out;
1382                         }
1383                         newattr.max_ord = attrs->max_ord;
1384                 }
1385                 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1386                         if (attrs->max_ird > cur_max_read_depth(rhp)) {
1387                                 ret = -EINVAL;
1388                                 goto out;
1389                         }
1390                         newattr.max_ird = attrs->max_ird;
1391                 }
1392                 qhp->attr = newattr;
1393         }
1394
1395         if (mask & C4IW_QP_ATTR_SQ_DB) {
1396                 ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1397                 goto out;
1398         }
1399         if (mask & C4IW_QP_ATTR_RQ_DB) {
1400                 ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1401                 goto out;
1402         }
1403
1404         if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1405                 goto out;
1406         if (qhp->attr.state == attrs->next_state)
1407                 goto out;
1408
1409         switch (qhp->attr.state) {
1410         case C4IW_QP_STATE_IDLE:
1411                 switch (attrs->next_state) {
1412                 case C4IW_QP_STATE_RTS:
1413                         if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1414                                 ret = -EINVAL;
1415                                 goto out;
1416                         }
1417                         if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1418                                 ret = -EINVAL;
1419                                 goto out;
1420                         }
1421                         qhp->attr.mpa_attr = attrs->mpa_attr;
1422                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1423                         qhp->ep = qhp->attr.llp_stream_handle;
1424                         set_state(qhp, C4IW_QP_STATE_RTS);
1425
1426                         /*
1427                          * Ref the endpoint here and deref when we
1428                          * disassociate the endpoint from the QP.  This
1429                          * happens in CLOSING->IDLE transition or *->ERROR
1430                          * transition.
1431                          */
1432                         c4iw_get_ep(&qhp->ep->com);
1433                         ret = rdma_init(rhp, qhp);
1434                         if (ret)
1435                                 goto err;
1436                         break;
1437                 case C4IW_QP_STATE_ERROR:
1438                         set_state(qhp, C4IW_QP_STATE_ERROR);
1439                         flush_qp(qhp);
1440                         break;
1441                 default:
1442                         ret = -EINVAL;
1443                         goto out;
1444                 }
1445                 break;
1446         case C4IW_QP_STATE_RTS:
1447                 switch (attrs->next_state) {
1448                 case C4IW_QP_STATE_CLOSING:
1449                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1450                         t4_set_wq_in_error(&qhp->wq);
1451                         set_state(qhp, C4IW_QP_STATE_CLOSING);
1452                         ep = qhp->ep;
1453                         if (!internal) {
1454                                 abort = 0;
1455                                 disconnect = 1;
1456                                 c4iw_get_ep(&qhp->ep->com);
1457                         }
1458                         ret = rdma_fini(rhp, qhp, ep);
1459                         if (ret)
1460                                 goto err;
1461                         break;
1462                 case C4IW_QP_STATE_TERMINATE:
1463                         t4_set_wq_in_error(&qhp->wq);
1464                         set_state(qhp, C4IW_QP_STATE_TERMINATE);
1465                         qhp->attr.layer_etype = attrs->layer_etype;
1466                         qhp->attr.ecode = attrs->ecode;
1467                         ep = qhp->ep;
1468                         if (!internal) {
1469                                 c4iw_get_ep(&qhp->ep->com);
1470                                 terminate = 1;
1471                                 disconnect = 1;
1472                         } else {
1473                                 terminate = qhp->attr.send_term;
1474                                 ret = rdma_fini(rhp, qhp, ep);
1475                                 if (ret)
1476                                         goto err;
1477                         }
1478                         break;
1479                 case C4IW_QP_STATE_ERROR:
1480                         t4_set_wq_in_error(&qhp->wq);
1481                         set_state(qhp, C4IW_QP_STATE_ERROR);
1482                         if (!internal) {
1483                                 abort = 1;
1484                                 disconnect = 1;
1485                                 ep = qhp->ep;
1486                                 c4iw_get_ep(&qhp->ep->com);
1487                         }
1488                         goto err;
1489                         break;
1490                 default:
1491                         ret = -EINVAL;
1492                         goto out;
1493                 }
1494                 break;
1495         case C4IW_QP_STATE_CLOSING:
1496                 if (!internal) {
1497                         ret = -EINVAL;
1498                         goto out;
1499                 }
1500                 switch (attrs->next_state) {
1501                 case C4IW_QP_STATE_IDLE:
1502                         flush_qp(qhp);
1503                         set_state(qhp, C4IW_QP_STATE_IDLE);
1504                         qhp->attr.llp_stream_handle = NULL;
1505                         c4iw_put_ep(&qhp->ep->com);
1506                         qhp->ep = NULL;
1507                         wake_up(&qhp->wait);
1508                         break;
1509                 case C4IW_QP_STATE_ERROR:
1510                         goto err;
1511                 default:
1512                         ret = -EINVAL;
1513                         goto err;
1514                 }
1515                 break;
1516         case C4IW_QP_STATE_ERROR:
1517                 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1518                         ret = -EINVAL;
1519                         goto out;
1520                 }
1521                 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1522                         ret = -EINVAL;
1523                         goto out;
1524                 }
1525                 set_state(qhp, C4IW_QP_STATE_IDLE);
1526                 break;
1527         case C4IW_QP_STATE_TERMINATE:
1528                 if (!internal) {
1529                         ret = -EINVAL;
1530                         goto out;
1531                 }
1532                 goto err;
1533                 break;
1534         default:
1535                 printk(KERN_ERR "%s in a bad state %d\n",
1536                        __func__, qhp->attr.state);
1537                 ret = -EINVAL;
1538                 goto err;
1539                 break;
1540         }
1541         goto out;
1542 err:
1543         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1544              qhp->wq.sq.qid);
1545
1546         /* disassociate the LLP connection */
1547         qhp->attr.llp_stream_handle = NULL;
1548         if (!ep)
1549                 ep = qhp->ep;
1550         qhp->ep = NULL;
1551         set_state(qhp, C4IW_QP_STATE_ERROR);
1552         free = 1;
1553         abort = 1;
1554         BUG_ON(!ep);
1555         flush_qp(qhp);
1556         wake_up(&qhp->wait);
1557 out:
1558         mutex_unlock(&qhp->mutex);
1559
1560         if (terminate)
1561                 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1562
1563         /*
1564          * If disconnect is 1, then we need to initiate a disconnect
1565          * on the EP.  This can be a normal close (RTS->CLOSING) or
1566          * an abnormal close (RTS/CLOSING->ERROR).
1567          */
1568         if (disconnect) {
1569                 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1570                                                          GFP_KERNEL);
1571                 c4iw_put_ep(&ep->com);
1572         }
1573
1574         /*
1575          * If free is 1, then we've disassociated the EP from the QP
1576          * and we need to dereference the EP.
1577          */
1578         if (free)
1579                 c4iw_put_ep(&ep->com);
1580         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1581         return ret;
1582 }
1583
1584 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1585 {
1586         struct c4iw_dev *rhp;
1587         struct c4iw_qp *qhp;
1588         struct c4iw_qp_attributes attrs;
1589         struct c4iw_ucontext *ucontext;
1590
1591         qhp = to_c4iw_qp(ib_qp);
1592         rhp = qhp->rhp;
1593
1594         attrs.next_state = C4IW_QP_STATE_ERROR;
1595         if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1596                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1597         else
1598                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1599         wait_event(qhp->wait, !qhp->ep);
1600
1601         remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1602         atomic_dec(&qhp->refcnt);
1603         wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1604
1605         spin_lock_irq(&rhp->lock);
1606         if (!list_empty(&qhp->db_fc_entry))
1607                 list_del_init(&qhp->db_fc_entry);
1608         spin_unlock_irq(&rhp->lock);
1609         free_ird(rhp, qhp->attr.max_ird);
1610
1611         ucontext = ib_qp->uobject ?
1612                    to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1613         destroy_qp(&rhp->rdev, &qhp->wq,
1614                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1615
1616         PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1617         kfree(qhp);
1618         return 0;
1619 }
1620
1621 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1622                              struct ib_udata *udata)
1623 {
1624         struct c4iw_dev *rhp;
1625         struct c4iw_qp *qhp;
1626         struct c4iw_pd *php;
1627         struct c4iw_cq *schp;
1628         struct c4iw_cq *rchp;
1629         struct c4iw_create_qp_resp uresp;
1630         unsigned int sqsize, rqsize;
1631         struct c4iw_ucontext *ucontext;
1632         int ret;
1633         struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1634
1635         PDBG("%s ib_pd %p\n", __func__, pd);
1636
1637         if (attrs->qp_type != IB_QPT_RC)
1638                 return ERR_PTR(-EINVAL);
1639
1640         php = to_c4iw_pd(pd);
1641         rhp = php->rhp;
1642         schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1643         rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1644         if (!schp || !rchp)
1645                 return ERR_PTR(-EINVAL);
1646
1647         if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1648                 return ERR_PTR(-EINVAL);
1649
1650         if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1651                 return ERR_PTR(-E2BIG);
1652         rqsize = attrs->cap.max_recv_wr + 1;
1653         if (rqsize < 8)
1654                 rqsize = 8;
1655
1656         if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1657                 return ERR_PTR(-E2BIG);
1658         sqsize = attrs->cap.max_send_wr + 1;
1659         if (sqsize < 8)
1660                 sqsize = 8;
1661
1662         ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1663
1664         qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1665         if (!qhp)
1666                 return ERR_PTR(-ENOMEM);
1667         qhp->wq.sq.size = sqsize;
1668         qhp->wq.sq.memsize =
1669                 (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1670                 sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1671         qhp->wq.sq.flush_cidx = -1;
1672         qhp->wq.rq.size = rqsize;
1673         qhp->wq.rq.memsize =
1674                 (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1675                 sizeof(*qhp->wq.rq.queue);
1676
1677         if (ucontext) {
1678                 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1679                 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1680         }
1681
1682         ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1683                         ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1684         if (ret)
1685                 goto err1;
1686
1687         attrs->cap.max_recv_wr = rqsize - 1;
1688         attrs->cap.max_send_wr = sqsize - 1;
1689         attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1690
1691         qhp->rhp = rhp;
1692         qhp->attr.pd = php->pdid;
1693         qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1694         qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1695         qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1696         qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1697         qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1698         qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1699         qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1700         qhp->attr.state = C4IW_QP_STATE_IDLE;
1701         qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1702         qhp->attr.enable_rdma_read = 1;
1703         qhp->attr.enable_rdma_write = 1;
1704         qhp->attr.enable_bind = 1;
1705         qhp->attr.max_ord = 0;
1706         qhp->attr.max_ird = 0;
1707         qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1708         spin_lock_init(&qhp->lock);
1709         mutex_init(&qhp->mutex);
1710         init_waitqueue_head(&qhp->wait);
1711         atomic_set(&qhp->refcnt, 1);
1712
1713         ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1714         if (ret)
1715                 goto err2;
1716
1717         if (udata) {
1718                 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1719                 if (!mm1) {
1720                         ret = -ENOMEM;
1721                         goto err3;
1722                 }
1723                 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1724                 if (!mm2) {
1725                         ret = -ENOMEM;
1726                         goto err4;
1727                 }
1728                 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1729                 if (!mm3) {
1730                         ret = -ENOMEM;
1731                         goto err5;
1732                 }
1733                 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1734                 if (!mm4) {
1735                         ret = -ENOMEM;
1736                         goto err6;
1737                 }
1738                 if (t4_sq_onchip(&qhp->wq.sq)) {
1739                         mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1740                         if (!mm5) {
1741                                 ret = -ENOMEM;
1742                                 goto err7;
1743                         }
1744                         uresp.flags = C4IW_QPF_ONCHIP;
1745                 } else
1746                         uresp.flags = 0;
1747                 uresp.qid_mask = rhp->rdev.qpmask;
1748                 uresp.sqid = qhp->wq.sq.qid;
1749                 uresp.sq_size = qhp->wq.sq.size;
1750                 uresp.sq_memsize = qhp->wq.sq.memsize;
1751                 uresp.rqid = qhp->wq.rq.qid;
1752                 uresp.rq_size = qhp->wq.rq.size;
1753                 uresp.rq_memsize = qhp->wq.rq.memsize;
1754                 spin_lock(&ucontext->mmap_lock);
1755                 if (mm5) {
1756                         uresp.ma_sync_key = ucontext->key;
1757                         ucontext->key += PAGE_SIZE;
1758                 } else {
1759                         uresp.ma_sync_key =  0;
1760                 }
1761                 uresp.sq_key = ucontext->key;
1762                 ucontext->key += PAGE_SIZE;
1763                 uresp.rq_key = ucontext->key;
1764                 ucontext->key += PAGE_SIZE;
1765                 uresp.sq_db_gts_key = ucontext->key;
1766                 ucontext->key += PAGE_SIZE;
1767                 uresp.rq_db_gts_key = ucontext->key;
1768                 ucontext->key += PAGE_SIZE;
1769                 spin_unlock(&ucontext->mmap_lock);
1770                 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1771                 if (ret)
1772                         goto err8;
1773                 mm1->key = uresp.sq_key;
1774                 mm1->addr = qhp->wq.sq.phys_addr;
1775                 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1776                 insert_mmap(ucontext, mm1);
1777                 mm2->key = uresp.rq_key;
1778                 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1779                 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1780                 insert_mmap(ucontext, mm2);
1781                 mm3->key = uresp.sq_db_gts_key;
1782                 mm3->addr = (__force unsigned long)qhp->wq.sq.bar2_pa;
1783                 mm3->len = PAGE_SIZE;
1784                 insert_mmap(ucontext, mm3);
1785                 mm4->key = uresp.rq_db_gts_key;
1786                 mm4->addr = (__force unsigned long)qhp->wq.rq.bar2_pa;
1787                 mm4->len = PAGE_SIZE;
1788                 insert_mmap(ucontext, mm4);
1789                 if (mm5) {
1790                         mm5->key = uresp.ma_sync_key;
1791                         mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1792                                     + PCIE_MA_SYNC_A) & PAGE_MASK;
1793                         mm5->len = PAGE_SIZE;
1794                         insert_mmap(ucontext, mm5);
1795                 }
1796         }
1797         qhp->ibqp.qp_num = qhp->wq.sq.qid;
1798         init_timer(&(qhp->timer));
1799         INIT_LIST_HEAD(&qhp->db_fc_entry);
1800         PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1801              "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1802              qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1803              attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1804              qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1805         return &qhp->ibqp;
1806 err8:
1807         kfree(mm5);
1808 err7:
1809         kfree(mm4);
1810 err6:
1811         kfree(mm3);
1812 err5:
1813         kfree(mm2);
1814 err4:
1815         kfree(mm1);
1816 err3:
1817         remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1818 err2:
1819         destroy_qp(&rhp->rdev, &qhp->wq,
1820                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1821 err1:
1822         kfree(qhp);
1823         return ERR_PTR(ret);
1824 }
1825
1826 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1827                       int attr_mask, struct ib_udata *udata)
1828 {
1829         struct c4iw_dev *rhp;
1830         struct c4iw_qp *qhp;
1831         enum c4iw_qp_attr_mask mask = 0;
1832         struct c4iw_qp_attributes attrs;
1833
1834         PDBG("%s ib_qp %p\n", __func__, ibqp);
1835
1836         /* iwarp does not support the RTR state */
1837         if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1838                 attr_mask &= ~IB_QP_STATE;
1839
1840         /* Make sure we still have something left to do */
1841         if (!attr_mask)
1842                 return 0;
1843
1844         memset(&attrs, 0, sizeof attrs);
1845         qhp = to_c4iw_qp(ibqp);
1846         rhp = qhp->rhp;
1847
1848         attrs.next_state = c4iw_convert_state(attr->qp_state);
1849         attrs.enable_rdma_read = (attr->qp_access_flags &
1850                                IB_ACCESS_REMOTE_READ) ?  1 : 0;
1851         attrs.enable_rdma_write = (attr->qp_access_flags &
1852                                 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1853         attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1854
1855
1856         mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1857         mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1858                         (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1859                          C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1860                          C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1861
1862         /*
1863          * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1864          * ringing the queue db when we're in DB_FULL mode.
1865          * Only allow this on T4 devices.
1866          */
1867         attrs.sq_db_inc = attr->sq_psn;
1868         attrs.rq_db_inc = attr->rq_psn;
1869         mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1870         mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1871         if (!is_t4(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1872             (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1873                 return -EINVAL;
1874
1875         return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1876 }
1877
1878 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1879 {
1880         PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1881         return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1882 }
1883
1884 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1885                      int attr_mask, struct ib_qp_init_attr *init_attr)
1886 {
1887         struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1888
1889         memset(attr, 0, sizeof *attr);
1890         memset(init_attr, 0, sizeof *init_attr);
1891         attr->qp_state = to_ib_qp_state(qhp->attr.state);
1892         init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1893         init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1894         init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1895         init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1896         init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1897         init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1898         return 0;
1899 }