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         return rdev->bar2_kva + bar2_qoffset;
 189 }
 190
 191 static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq,
 192                      struct t4_cq *rcq, struct t4_cq *scq,
 193                      struct c4iw_dev_ucontext *uctx)
 194 {
 195         int user = (uctx != &rdev->uctx);
 196         struct fw_ri_res_wr *res_wr;
 197         struct fw_ri_res *res;
 198         int wr_len;
 199         struct c4iw_wr_wait wr_wait;
 200         struct sk_buff *skb;
 201         int ret = 0;
 202         int eqsize;
 203
 204         wq->sq.qid = c4iw_get_qpid(rdev, uctx);
 205         if (!wq->sq.qid)
 206                 return -ENOMEM;
 207
 208         wq->rq.qid = c4iw_get_qpid(rdev, uctx);
 209         if (!wq->rq.qid) {
 210                 ret = -ENOMEM;
 211                 goto free_sq_qid;
 212         }
 213
 214         if (!user) {
 215                 wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq,
 216                                  GFP_KERNEL);
 217                 if (!wq->sq.sw_sq) {
 218                         ret = -ENOMEM;
 219                         goto free_rq_qid;
 220                 }
 221
 222                 wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq,
 223                                  GFP_KERNEL);
 224                 if (!wq->rq.sw_rq) {
 225                         ret = -ENOMEM;
 226                         goto free_sw_sq;
 227                 }
 228         }
 229
 230         /*
 231          * RQT must be a power of 2 and at least 16 deep.
 232          */
 233         wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16));
 234         wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size);
 235         if (!wq->rq.rqt_hwaddr) {
 236                 ret = -ENOMEM;
 237                 goto free_sw_rq;
 238         }
 239
 240         ret = alloc_sq(rdev, &wq->sq, user);
 241         if (ret)
 242                 goto free_hwaddr;
 243         memset(wq->sq.queue, 0, wq->sq.memsize);
 244         dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr);
 245
 246         wq->rq.queue = dma_alloc_coherent(&(rdev->lldi.pdev->dev),
 247                                           wq->rq.memsize, &(wq->rq.dma_addr),
 248                                           GFP_KERNEL);
 249         if (!wq->rq.queue) {
 250                 ret = -ENOMEM;
 251                 goto free_sq;
 252         }
 253         PDBG("%s sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx\n",
 254                 __func__, wq->sq.queue,
 255                 (unsigned long long)virt_to_phys(wq->sq.queue),
 256                 wq->rq.queue,
 257                 (unsigned long long)virt_to_phys(wq->rq.queue));
 258         memset(wq->rq.queue, 0, wq->rq.memsize);
 259         dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr);
 260
 261         wq->db = rdev->lldi.db_reg;
 262
 263         wq->sq.bar2_va = c4iw_bar2_addrs(rdev, wq->sq.qid, T4_BAR2_QTYPE_EGRESS,
 264                                          &wq->sq.bar2_qid,
 265                                          user ? &wq->sq.bar2_pa : NULL);
 266         wq->rq.bar2_va = c4iw_bar2_addrs(rdev, wq->rq.qid, T4_BAR2_QTYPE_EGRESS,
 267                                          &wq->rq.bar2_qid,
 268                                          user ? &wq->rq.bar2_pa : NULL);
 269
 270         /*
 271          * User mode must have bar2 access.
 272          */
 273         if (user && (!wq->sq.bar2_va || !wq->rq.bar2_va)) {
 274                 pr_warn(MOD "%s: sqid %u or rqid %u not in BAR2 range.\n",
 275                         pci_name(rdev->lldi.pdev), wq->sq.qid, wq->rq.qid);
 276                 goto free_dma;
 277         }
 278
 279         wq->rdev = rdev;
 280         wq->rq.msn = 1;
 281
 282         /* build fw_ri_res_wr */
 283         wr_len = sizeof *res_wr + 2 * sizeof *res;
 284
 285         skb = alloc_skb(wr_len, GFP_KERNEL);
 286         if (!skb) {
 287                 ret = -ENOMEM;
 288                 goto free_dma;
 289         }
 290         set_wr_txq(skb, CPL_PRIORITY_CONTROL, 0);
 291
 292         res_wr = (struct fw_ri_res_wr *)__skb_put(skb, wr_len);
 293         memset(res_wr, 0, wr_len);
 294         res_wr->op_nres = cpu_to_be32(
 295                         FW_WR_OP_V(FW_RI_RES_WR) |
 296                         FW_RI_RES_WR_NRES_V(2) |
 297                         FW_WR_COMPL_F);
 298         res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16));
 299         res_wr->cookie = (uintptr_t)&wr_wait;
 300         res = res_wr->res;
 301         res->u.sqrq.restype = FW_RI_RES_TYPE_SQ;
 302         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 303
 304         /*
 305          * eqsize is the number of 64B entries plus the status page size.
 306          */
 307         eqsize = wq->sq.size * T4_SQ_NUM_SLOTS +
 308                 rdev->hw_queue.t4_eq_status_entries;
 309
 310         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 311                 FW_RI_RES_WR_HOSTFCMODE_V(0) |  /* no host cidx updates */
 312                 FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
 313                 FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
 314                 (t4_sq_onchip(&wq->sq) ? FW_RI_RES_WR_ONCHIP_F : 0) |
 315                 FW_RI_RES_WR_IQID_V(scq->cqid));
 316         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 317                 FW_RI_RES_WR_DCAEN_V(0) |
 318                 FW_RI_RES_WR_DCACPU_V(0) |
 319                 FW_RI_RES_WR_FBMIN_V(2) |
 320                 FW_RI_RES_WR_FBMAX_V(2) |
 321                 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
 322                 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
 323                 FW_RI_RES_WR_EQSIZE_V(eqsize));
 324         res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid);
 325         res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr);
 326         res++;
 327         res->u.sqrq.restype = FW_RI_RES_TYPE_RQ;
 328         res->u.sqrq.op = FW_RI_RES_OP_WRITE;
 329
 330         /*
 331          * eqsize is the number of 64B entries plus the status page size.
 332          */
 333         eqsize = wq->rq.size * T4_RQ_NUM_SLOTS +
 334                 rdev->hw_queue.t4_eq_status_entries;
 335         res->u.sqrq.fetchszm_to_iqid = cpu_to_be32(
 336                 FW_RI_RES_WR_HOSTFCMODE_V(0) |  /* no host cidx updates */
 337                 FW_RI_RES_WR_CPRIO_V(0) |       /* don't keep in chip cache */
 338                 FW_RI_RES_WR_PCIECHN_V(0) |     /* set by uP at ri_init time */
 339                 FW_RI_RES_WR_IQID_V(rcq->cqid));
 340         res->u.sqrq.dcaen_to_eqsize = cpu_to_be32(
 341                 FW_RI_RES_WR_DCAEN_V(0) |
 342                 FW_RI_RES_WR_DCACPU_V(0) |
 343                 FW_RI_RES_WR_FBMIN_V(2) |
 344                 FW_RI_RES_WR_FBMAX_V(2) |
 345                 FW_RI_RES_WR_CIDXFTHRESHO_V(0) |
 346                 FW_RI_RES_WR_CIDXFTHRESH_V(0) |
 347                 FW_RI_RES_WR_EQSIZE_V(eqsize));
 348         res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid);
 349         res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr);
 350
 351         c4iw_init_wr_wait(&wr_wait);
 352
 353         ret = c4iw_ofld_send(rdev, skb);
 354         if (ret)
 355                 goto free_dma;
 356         ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, __func__);
 357         if (ret)
 358                 goto free_dma;
 359
 360         PDBG("%s sqid 0x%x rqid 0x%x kdb 0x%p sq_bar2_addr %p rq_bar2_addr %p\n",
 361              __func__, wq->sq.qid, wq->rq.qid, wq->db,
 362              wq->sq.bar2_va, wq->rq.bar2_va);
 363
 364         return 0;
 365 free_dma:
 366         dma_free_coherent(&(rdev->lldi.pdev->dev),
 367                           wq->rq.memsize, wq->rq.queue,
 368                           dma_unmap_addr(&wq->rq, mapping));
 369 free_sq:
 370         dealloc_sq(rdev, &wq->sq);
 371 free_hwaddr:
 372         c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size);
 373 free_sw_rq:
 374         kfree(wq->rq.sw_rq);
 375 free_sw_sq:
 376         kfree(wq->sq.sw_sq);
 377 free_rq_qid:
 378         c4iw_put_qpid(rdev, wq->rq.qid, uctx);
 379 free_sq_qid:
 380         c4iw_put_qpid(rdev, wq->sq.qid, uctx);
 381         return ret;
 382 }
 383
 384 static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp,
 385                       struct ib_send_wr *wr, int max, u32 *plenp)
 386 {
 387         u8 *dstp, *srcp;
 388         u32 plen = 0;
 389         int i;
 390         int rem, len;
 391
 392         dstp = (u8 *)immdp->data;
 393         for (i = 0; i < wr->num_sge; i++) {
 394                 if ((plen + wr->sg_list[i].length) > max)
 395                         return -EMSGSIZE;
 396                 srcp = (u8 *)(unsigned long)wr->sg_list[i].addr;
 397                 plen += wr->sg_list[i].length;
 398                 rem = wr->sg_list[i].length;
 399                 while (rem) {
 400                         if (dstp == (u8 *)&sq->queue[sq->size])
 401                                 dstp = (u8 *)sq->queue;
 402                         if (rem <= (u8 *)&sq->queue[sq->size] - dstp)
 403                                 len = rem;
 404                         else
 405                                 len = (u8 *)&sq->queue[sq->size] - dstp;
 406                         memcpy(dstp, srcp, len);
 407                         dstp += len;
 408                         srcp += len;
 409                         rem -= len;
 410                 }
 411         }
 412         len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp);
 413         if (len)
 414                 memset(dstp, 0, len);
 415         immdp->op = FW_RI_DATA_IMMD;
 416         immdp->r1 = 0;
 417         immdp->r2 = 0;
 418         immdp->immdlen = cpu_to_be32(plen);
 419         *plenp = plen;
 420         return 0;
 421 }
 422
 423 static int build_isgl(__be64 *queue_start, __be64 *queue_end,
 424                       struct fw_ri_isgl *isglp, struct ib_sge *sg_list,
 425                       int num_sge, u32 *plenp)
 426
 427 {
 428         int i;
 429         u32 plen = 0;
 430         __be64 *flitp = (__be64 *)isglp->sge;
 431
 432         for (i = 0; i < num_sge; i++) {
 433                 if ((plen + sg_list[i].length) < plen)
 434                         return -EMSGSIZE;
 435                 plen += sg_list[i].length;
 436                 *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) |
 437                                      sg_list[i].length);
 438                 if (++flitp == queue_end)
 439                         flitp = queue_start;
 440                 *flitp = cpu_to_be64(sg_list[i].addr);
 441                 if (++flitp == queue_end)
 442                         flitp = queue_start;
 443         }
 444         *flitp = (__force __be64)0;
 445         isglp->op = FW_RI_DATA_ISGL;
 446         isglp->r1 = 0;
 447         isglp->nsge = cpu_to_be16(num_sge);
 448         isglp->r2 = 0;
 449         if (plenp)
 450                 *plenp = plen;
 451         return 0;
 452 }
 453
 454 static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe,
 455                            struct ib_send_wr *wr, u8 *len16)
 456 {
 457         u32 plen;
 458         int size;
 459         int ret;
 460
 461         if (wr->num_sge > T4_MAX_SEND_SGE)
 462                 return -EINVAL;
 463         switch (wr->opcode) {
 464         case IB_WR_SEND:
 465                 if (wr->send_flags & IB_SEND_SOLICITED)
 466                         wqe->send.sendop_pkd = cpu_to_be32(
 467                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE));
 468                 else
 469                         wqe->send.sendop_pkd = cpu_to_be32(
 470                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND));
 471                 wqe->send.stag_inv = 0;
 472                 break;
 473         case IB_WR_SEND_WITH_INV:
 474                 if (wr->send_flags & IB_SEND_SOLICITED)
 475                         wqe->send.sendop_pkd = cpu_to_be32(
 476                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_SE_INV));
 477                 else
 478                         wqe->send.sendop_pkd = cpu_to_be32(
 479                                 FW_RI_SEND_WR_SENDOP_V(FW_RI_SEND_WITH_INV));
 480                 wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 481                 break;
 482
 483         default:
 484                 return -EINVAL;
 485         }
 486         wqe->send.r3 = 0;
 487         wqe->send.r4 = 0;
 488
 489         plen = 0;
 490         if (wr->num_sge) {
 491                 if (wr->send_flags & IB_SEND_INLINE) {
 492                         ret = build_immd(sq, wqe->send.u.immd_src, wr,
 493                                          T4_MAX_SEND_INLINE, &plen);
 494                         if (ret)
 495                                 return ret;
 496                         size = sizeof wqe->send + sizeof(struct fw_ri_immd) +
 497                                plen;
 498                 } else {
 499                         ret = build_isgl((__be64 *)sq->queue,
 500                                          (__be64 *)&sq->queue[sq->size],
 501                                          wqe->send.u.isgl_src,
 502                                          wr->sg_list, wr->num_sge, &plen);
 503                         if (ret)
 504                                 return ret;
 505                         size = sizeof wqe->send + sizeof(struct fw_ri_isgl) +
 506                                wr->num_sge * sizeof(struct fw_ri_sge);
 507                 }
 508         } else {
 509                 wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD;
 510                 wqe->send.u.immd_src[0].r1 = 0;
 511                 wqe->send.u.immd_src[0].r2 = 0;
 512                 wqe->send.u.immd_src[0].immdlen = 0;
 513                 size = sizeof wqe->send + sizeof(struct fw_ri_immd);
 514                 plen = 0;
 515         }
 516         *len16 = DIV_ROUND_UP(size, 16);
 517         wqe->send.plen = cpu_to_be32(plen);
 518         return 0;
 519 }
 520
 521 static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe,
 522                             struct ib_send_wr *wr, u8 *len16)
 523 {
 524         u32 plen;
 525         int size;
 526         int ret;
 527
 528         if (wr->num_sge > T4_MAX_SEND_SGE)
 529                 return -EINVAL;
 530         wqe->write.r2 = 0;
 531         wqe->write.stag_sink = cpu_to_be32(wr->wr.rdma.rkey);
 532         wqe->write.to_sink = cpu_to_be64(wr->wr.rdma.remote_addr);
 533         if (wr->num_sge) {
 534                 if (wr->send_flags & IB_SEND_INLINE) {
 535                         ret = build_immd(sq, wqe->write.u.immd_src, wr,
 536                                          T4_MAX_WRITE_INLINE, &plen);
 537                         if (ret)
 538                                 return ret;
 539                         size = sizeof wqe->write + sizeof(struct fw_ri_immd) +
 540                                plen;
 541                 } else {
 542                         ret = build_isgl((__be64 *)sq->queue,
 543                                          (__be64 *)&sq->queue[sq->size],
 544                                          wqe->write.u.isgl_src,
 545                                          wr->sg_list, wr->num_sge, &plen);
 546                         if (ret)
 547                                 return ret;
 548                         size = sizeof wqe->write + sizeof(struct fw_ri_isgl) +
 549                                wr->num_sge * sizeof(struct fw_ri_sge);
 550                 }
 551         } else {
 552                 wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD;
 553                 wqe->write.u.immd_src[0].r1 = 0;
 554                 wqe->write.u.immd_src[0].r2 = 0;
 555                 wqe->write.u.immd_src[0].immdlen = 0;
 556                 size = sizeof wqe->write + sizeof(struct fw_ri_immd);
 557                 plen = 0;
 558         }
 559         *len16 = DIV_ROUND_UP(size, 16);
 560         wqe->write.plen = cpu_to_be32(plen);
 561         return 0;
 562 }
 563
 564 static int build_rdma_read(union t4_wr *wqe, struct ib_send_wr *wr, u8 *len16)
 565 {
 566         if (wr->num_sge > 1)
 567                 return -EINVAL;
 568         if (wr->num_sge) {
 569                 wqe->read.stag_src = cpu_to_be32(wr->wr.rdma.rkey);
 570                 wqe->read.to_src_hi = cpu_to_be32((u32)(wr->wr.rdma.remote_addr
 571                                                         >> 32));
 572                 wqe->read.to_src_lo = cpu_to_be32((u32)wr->wr.rdma.remote_addr);
 573                 wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey);
 574                 wqe->read.plen = cpu_to_be32(wr->sg_list[0].length);
 575                 wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr
 576                                                          >> 32));
 577                 wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr));
 578         } else {
 579                 wqe->read.stag_src = cpu_to_be32(2);
 580                 wqe->read.to_src_hi = 0;
 581                 wqe->read.to_src_lo = 0;
 582                 wqe->read.stag_sink = cpu_to_be32(2);
 583                 wqe->read.plen = 0;
 584                 wqe->read.to_sink_hi = 0;
 585                 wqe->read.to_sink_lo = 0;
 586         }
 587         wqe->read.r2 = 0;
 588         wqe->read.r5 = 0;
 589         *len16 = DIV_ROUND_UP(sizeof wqe->read, 16);
 590         return 0;
 591 }
 592
 593 static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe,
 594                            struct ib_recv_wr *wr, u8 *len16)
 595 {
 596         int ret;
 597
 598         ret = build_isgl((__be64 *)qhp->wq.rq.queue,
 599                          (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size],
 600                          &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL);
 601         if (ret)
 602                 return ret;
 603         *len16 = DIV_ROUND_UP(sizeof wqe->recv +
 604                               wr->num_sge * sizeof(struct fw_ri_sge), 16);
 605         return 0;
 606 }
 607
 608 static int build_fastreg(struct t4_sq *sq, union t4_wr *wqe,
 609                          struct ib_send_wr *wr, u8 *len16, u8 t5dev)
 610 {
 611
 612         struct fw_ri_immd *imdp;
 613         __be64 *p;
 614         int i;
 615         int pbllen = roundup(wr->wr.fast_reg.page_list_len * sizeof(u64), 32);
 616         int rem;
 617
 618         if (wr->wr.fast_reg.page_list_len >
 619             t4_max_fr_depth(use_dsgl))
 620                 return -EINVAL;
 621
 622         wqe->fr.qpbinde_to_dcacpu = 0;
 623         wqe->fr.pgsz_shift = wr->wr.fast_reg.page_shift - 12;
 624         wqe->fr.addr_type = FW_RI_VA_BASED_TO;
 625         wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->wr.fast_reg.access_flags);
 626         wqe->fr.len_hi = 0;
 627         wqe->fr.len_lo = cpu_to_be32(wr->wr.fast_reg.length);
 628         wqe->fr.stag = cpu_to_be32(wr->wr.fast_reg.rkey);
 629         wqe->fr.va_hi = cpu_to_be32(wr->wr.fast_reg.iova_start >> 32);
 630         wqe->fr.va_lo_fbo = cpu_to_be32(wr->wr.fast_reg.iova_start &
 631                                         0xffffffff);
 632
 633         if (t5dev && use_dsgl && (pbllen > max_fr_immd)) {
 634                 struct c4iw_fr_page_list *c4pl =
 635                         to_c4iw_fr_page_list(wr->wr.fast_reg.page_list);
 636                 struct fw_ri_dsgl *sglp;
 637
 638                 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 639                         wr->wr.fast_reg.page_list->page_list[i] = (__force u64)
 640                                 cpu_to_be64((u64)
 641                                 wr->wr.fast_reg.page_list->page_list[i]);
 642                 }
 643
 644                 sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1);
 645                 sglp->op = FW_RI_DATA_DSGL;
 646                 sglp->r1 = 0;
 647                 sglp->nsge = cpu_to_be16(1);
 648                 sglp->addr0 = cpu_to_be64(c4pl->dma_addr);
 649                 sglp->len0 = cpu_to_be32(pbllen);
 650
 651                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16);
 652         } else {
 653                 imdp = (struct fw_ri_immd *)(&wqe->fr + 1);
 654                 imdp->op = FW_RI_DATA_IMMD;
 655                 imdp->r1 = 0;
 656                 imdp->r2 = 0;
 657                 imdp->immdlen = cpu_to_be32(pbllen);
 658                 p = (__be64 *)(imdp + 1);
 659                 rem = pbllen;
 660                 for (i = 0; i < wr->wr.fast_reg.page_list_len; i++) {
 661                         *p = cpu_to_be64(
 662                                 (u64)wr->wr.fast_reg.page_list->page_list[i]);
 663                         rem -= sizeof(*p);
 664                         if (++p == (__be64 *)&sq->queue[sq->size])
 665                                 p = (__be64 *)sq->queue;
 666                 }
 667                 BUG_ON(rem < 0);
 668                 while (rem) {
 669                         *p = 0;
 670                         rem -= sizeof(*p);
 671                         if (++p == (__be64 *)&sq->queue[sq->size])
 672                                 p = (__be64 *)sq->queue;
 673                 }
 674                 *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp)
 675                                       + pbllen, 16);
 676         }
 677         return 0;
 678 }
 679
 680 static int build_inv_stag(union t4_wr *wqe, struct ib_send_wr *wr,
 681                           u8 *len16)
 682 {
 683         wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey);
 684         wqe->inv.r2 = 0;
 685         *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16);
 686         return 0;
 687 }
 688
 689 void c4iw_qp_add_ref(struct ib_qp *qp)
 690 {
 691         PDBG("%s ib_qp %p\n", __func__, qp);
 692         atomic_inc(&(to_c4iw_qp(qp)->refcnt));
 693 }
 694
 695 void c4iw_qp_rem_ref(struct ib_qp *qp)
 696 {
 697         PDBG("%s ib_qp %p\n", __func__, qp);
 698         if (atomic_dec_and_test(&(to_c4iw_qp(qp)->refcnt)))
 699                 wake_up(&(to_c4iw_qp(qp)->wait));
 700 }
 701
 702 static void add_to_fc_list(struct list_head *head, struct list_head *entry)
 703 {
 704         if (list_empty(entry))
 705                 list_add_tail(entry, head);
 706 }
 707
 708 static int ring_kernel_sq_db(struct c4iw_qp *qhp, u16 inc)
 709 {
 710         unsigned long flags;
 711
 712         spin_lock_irqsave(&qhp->rhp->lock, flags);
 713         spin_lock(&qhp->lock);
 714         if (qhp->rhp->db_state == NORMAL)
 715                 t4_ring_sq_db(&qhp->wq, inc,
 716                               is_t5(qhp->rhp->rdev.lldi.adapter_type), NULL);
 717         else {
 718                 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
 719                 qhp->wq.sq.wq_pidx_inc += inc;
 720         }
 721         spin_unlock(&qhp->lock);
 722         spin_unlock_irqrestore(&qhp->rhp->lock, flags);
 723         return 0;
 724 }
 725
 726 static int ring_kernel_rq_db(struct c4iw_qp *qhp, u16 inc)
 727 {
 728         unsigned long flags;
 729
 730         spin_lock_irqsave(&qhp->rhp->lock, flags);
 731         spin_lock(&qhp->lock);
 732         if (qhp->rhp->db_state == NORMAL)
 733                 t4_ring_rq_db(&qhp->wq, inc,
 734                               is_t5(qhp->rhp->rdev.lldi.adapter_type), NULL);
 735         else {
 736                 add_to_fc_list(&qhp->rhp->db_fc_list, &qhp->db_fc_entry);
 737                 qhp->wq.rq.wq_pidx_inc += inc;
 738         }
 739         spin_unlock(&qhp->lock);
 740         spin_unlock_irqrestore(&qhp->rhp->lock, flags);
 741         return 0;
 742 }
 743
 744 int c4iw_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
 745                    struct ib_send_wr **bad_wr)
 746 {
 747         int err = 0;
 748         u8 len16 = 0;
 749         enum fw_wr_opcodes fw_opcode = 0;
 750         enum fw_ri_wr_flags fw_flags;
 751         struct c4iw_qp *qhp;
 752         union t4_wr *wqe = NULL;
 753         u32 num_wrs;
 754         struct t4_swsqe *swsqe;
 755         unsigned long flag;
 756         u16 idx = 0;
 757
 758         qhp = to_c4iw_qp(ibqp);
 759         spin_lock_irqsave(&qhp->lock, flag);
 760         if (t4_wq_in_error(&qhp->wq)) {
 761                 spin_unlock_irqrestore(&qhp->lock, flag);
 762                 return -EINVAL;
 763         }
 764         num_wrs = t4_sq_avail(&qhp->wq);
 765         if (num_wrs == 0) {
 766                 spin_unlock_irqrestore(&qhp->lock, flag);
 767                 return -ENOMEM;
 768         }
 769         while (wr) {
 770                 if (num_wrs == 0) {
 771                         err = -ENOMEM;
 772                         *bad_wr = wr;
 773                         break;
 774                 }
 775                 wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue +
 776                       qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE);
 777
 778                 fw_flags = 0;
 779                 if (wr->send_flags & IB_SEND_SOLICITED)
 780                         fw_flags |= FW_RI_SOLICITED_EVENT_FLAG;
 781                 if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all)
 782                         fw_flags |= FW_RI_COMPLETION_FLAG;
 783                 swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
 784                 switch (wr->opcode) {
 785                 case IB_WR_SEND_WITH_INV:
 786                 case IB_WR_SEND:
 787                         if (wr->send_flags & IB_SEND_FENCE)
 788                                 fw_flags |= FW_RI_READ_FENCE_FLAG;
 789                         fw_opcode = FW_RI_SEND_WR;
 790                         if (wr->opcode == IB_WR_SEND)
 791                                 swsqe->opcode = FW_RI_SEND;
 792                         else
 793                                 swsqe->opcode = FW_RI_SEND_WITH_INV;
 794                         err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16);
 795                         break;
 796                 case IB_WR_RDMA_WRITE:
 797                         fw_opcode = FW_RI_RDMA_WRITE_WR;
 798                         swsqe->opcode = FW_RI_RDMA_WRITE;
 799                         err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16);
 800                         break;
 801                 case IB_WR_RDMA_READ:
 802                 case IB_WR_RDMA_READ_WITH_INV:
 803                         fw_opcode = FW_RI_RDMA_READ_WR;
 804                         swsqe->opcode = FW_RI_READ_REQ;
 805                         if (wr->opcode == IB_WR_RDMA_READ_WITH_INV)
 806                                 fw_flags = FW_RI_RDMA_READ_INVALIDATE;
 807                         else
 808                                 fw_flags = 0;
 809                         err = build_rdma_read(wqe, wr, &len16);
 810                         if (err)
 811                                 break;
 812                         swsqe->read_len = wr->sg_list[0].length;
 813                         if (!qhp->wq.sq.oldest_read)
 814                                 qhp->wq.sq.oldest_read = swsqe;
 815                         break;
 816                 case IB_WR_FAST_REG_MR:
 817                         fw_opcode = FW_RI_FR_NSMR_WR;
 818                         swsqe->opcode = FW_RI_FAST_REGISTER;
 819                         err = build_fastreg(&qhp->wq.sq, wqe, wr, &len16,
 820                                             is_t5(
 821                                             qhp->rhp->rdev.lldi.adapter_type) ?
 822                                             1 : 0);
 823                         break;
 824                 case IB_WR_LOCAL_INV:
 825                         if (wr->send_flags & IB_SEND_FENCE)
 826                                 fw_flags |= FW_RI_LOCAL_FENCE_FLAG;
 827                         fw_opcode = FW_RI_INV_LSTAG_WR;
 828                         swsqe->opcode = FW_RI_LOCAL_INV;
 829                         err = build_inv_stag(wqe, wr, &len16);
 830                         break;
 831                 default:
 832                         PDBG("%s post of type=%d TBD!\n", __func__,
 833                              wr->opcode);
 834                         err = -EINVAL;
 835                 }
 836                 if (err) {
 837                         *bad_wr = wr;
 838                         break;
 839                 }
 840                 swsqe->idx = qhp->wq.sq.pidx;
 841                 swsqe->complete = 0;
 842                 swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) ||
 843                                   qhp->sq_sig_all;
 844                 swsqe->flushed = 0;
 845                 swsqe->wr_id = wr->wr_id;
 846                 if (c4iw_wr_log) {
 847                         swsqe->sge_ts = cxgb4_read_sge_timestamp(
 848                                         qhp->rhp->rdev.lldi.ports[0]);
 849                         getnstimeofday(&swsqe->host_ts);
 850                 }
 851
 852                 init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16);
 853
 854                 PDBG("%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u\n",
 855                      __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx,
 856                      swsqe->opcode, swsqe->read_len);
 857                 wr = wr->next;
 858                 num_wrs--;
 859                 t4_sq_produce(&qhp->wq, len16);
 860                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 861         }
 862         if (!qhp->rhp->rdev.status_page->db_off) {
 863                 t4_ring_sq_db(&qhp->wq, idx,
 864                               is_t5(qhp->rhp->rdev.lldi.adapter_type), wqe);
 865                 spin_unlock_irqrestore(&qhp->lock, flag);
 866         } else {
 867                 spin_unlock_irqrestore(&qhp->lock, flag);
 868                 ring_kernel_sq_db(qhp, idx);
 869         }
 870         return err;
 871 }
 872
 873 int c4iw_post_receive(struct ib_qp *ibqp, struct ib_recv_wr *wr,
 874                       struct ib_recv_wr **bad_wr)
 875 {
 876         int err = 0;
 877         struct c4iw_qp *qhp;
 878         union t4_recv_wr *wqe = NULL;
 879         u32 num_wrs;
 880         u8 len16 = 0;
 881         unsigned long flag;
 882         u16 idx = 0;
 883
 884         qhp = to_c4iw_qp(ibqp);
 885         spin_lock_irqsave(&qhp->lock, flag);
 886         if (t4_wq_in_error(&qhp->wq)) {
 887                 spin_unlock_irqrestore(&qhp->lock, flag);
 888                 return -EINVAL;
 889         }
 890         num_wrs = t4_rq_avail(&qhp->wq);
 891         if (num_wrs == 0) {
 892                 spin_unlock_irqrestore(&qhp->lock, flag);
 893                 return -ENOMEM;
 894         }
 895         while (wr) {
 896                 if (wr->num_sge > T4_MAX_RECV_SGE) {
 897                         err = -EINVAL;
 898                         *bad_wr = wr;
 899                         break;
 900                 }
 901                 wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue +
 902                                            qhp->wq.rq.wq_pidx *
 903                                            T4_EQ_ENTRY_SIZE);
 904                 if (num_wrs)
 905                         err = build_rdma_recv(qhp, wqe, wr, &len16);
 906                 else
 907                         err = -ENOMEM;
 908                 if (err) {
 909                         *bad_wr = wr;
 910                         break;
 911                 }
 912
 913                 qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id;
 914                 if (c4iw_wr_log) {
 915                         qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].sge_ts =
 916                                 cxgb4_read_sge_timestamp(
 917                                                 qhp->rhp->rdev.lldi.ports[0]);
 918                         getnstimeofday(
 919                                 &qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].host_ts);
 920                 }
 921
 922                 wqe->recv.opcode = FW_RI_RECV_WR;
 923                 wqe->recv.r1 = 0;
 924                 wqe->recv.wrid = qhp->wq.rq.pidx;
 925                 wqe->recv.r2[0] = 0;
 926                 wqe->recv.r2[1] = 0;
 927                 wqe->recv.r2[2] = 0;
 928                 wqe->recv.len16 = len16;
 929                 PDBG("%s cookie 0x%llx pidx %u\n", __func__,
 930                      (unsigned long long) wr->wr_id, qhp->wq.rq.pidx);
 931                 t4_rq_produce(&qhp->wq, len16);
 932                 idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE);
 933                 wr = wr->next;
 934                 num_wrs--;
 935         }
 936         if (!qhp->rhp->rdev.status_page->db_off) {
 937                 t4_ring_rq_db(&qhp->wq, idx,
 938                               is_t5(qhp->rhp->rdev.lldi.adapter_type), wqe);
 939                 spin_unlock_irqrestore(&qhp->lock, flag);
 940         } else {
 941                 spin_unlock_irqrestore(&qhp->lock, flag);
 942                 ring_kernel_rq_db(qhp, idx);
 943         }
 944         return err;
 945 }
 946
 947 int c4iw_bind_mw(struct ib_qp *qp, struct ib_mw *mw, struct ib_mw_bind *mw_bind)
 948 {
 949         return -ENOSYS;
 950 }
 951
 952 static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type,
 953                                     u8 *ecode)
 954 {
 955         int status;
 956         int tagged;
 957         int opcode;
 958         int rqtype;
 959         int send_inv;
 960
 961         if (!err_cqe) {
 962                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
 963                 *ecode = 0;
 964                 return;
 965         }
 966
 967         status = CQE_STATUS(err_cqe);
 968         opcode = CQE_OPCODE(err_cqe);
 969         rqtype = RQ_TYPE(err_cqe);
 970         send_inv = (opcode == FW_RI_SEND_WITH_INV) ||
 971                    (opcode == FW_RI_SEND_WITH_SE_INV);
 972         tagged = (opcode == FW_RI_RDMA_WRITE) ||
 973                  (rqtype && (opcode == FW_RI_READ_RESP));
 974
 975         switch (status) {
 976         case T4_ERR_STAG:
 977                 if (send_inv) {
 978                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
 979                         *ecode = RDMAP_CANT_INV_STAG;
 980                 } else {
 981                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 982                         *ecode = RDMAP_INV_STAG;
 983                 }
 984                 break;
 985         case T4_ERR_PDID:
 986                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 987                 if ((opcode == FW_RI_SEND_WITH_INV) ||
 988                     (opcode == FW_RI_SEND_WITH_SE_INV))
 989                         *ecode = RDMAP_CANT_INV_STAG;
 990                 else
 991                         *ecode = RDMAP_STAG_NOT_ASSOC;
 992                 break;
 993         case T4_ERR_QPID:
 994                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 995                 *ecode = RDMAP_STAG_NOT_ASSOC;
 996                 break;
 997         case T4_ERR_ACCESS:
 998                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
 999                 *ecode = RDMAP_ACC_VIOL;
1000                 break;
1001         case T4_ERR_WRAP:
1002                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1003                 *ecode = RDMAP_TO_WRAP;
1004                 break;
1005         case T4_ERR_BOUND:
1006                 if (tagged) {
1007                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1008                         *ecode = DDPT_BASE_BOUNDS;
1009                 } else {
1010                         *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT;
1011                         *ecode = RDMAP_BASE_BOUNDS;
1012                 }
1013                 break;
1014         case T4_ERR_INVALIDATE_SHARED_MR:
1015         case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND:
1016                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1017                 *ecode = RDMAP_CANT_INV_STAG;
1018                 break;
1019         case T4_ERR_ECC:
1020         case T4_ERR_ECC_PSTAG:
1021         case T4_ERR_INTERNAL_ERR:
1022                 *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA;
1023                 *ecode = 0;
1024                 break;
1025         case T4_ERR_OUT_OF_RQE:
1026                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1027                 *ecode = DDPU_INV_MSN_NOBUF;
1028                 break;
1029         case T4_ERR_PBL_ADDR_BOUND:
1030                 *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1031                 *ecode = DDPT_BASE_BOUNDS;
1032                 break;
1033         case T4_ERR_CRC:
1034                 *layer_type = LAYER_MPA|DDP_LLP;
1035                 *ecode = MPA_CRC_ERR;
1036                 break;
1037         case T4_ERR_MARKER:
1038                 *layer_type = LAYER_MPA|DDP_LLP;
1039                 *ecode = MPA_MARKER_ERR;
1040                 break;
1041         case T4_ERR_PDU_LEN_ERR:
1042                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1043                 *ecode = DDPU_MSG_TOOBIG;
1044                 break;
1045         case T4_ERR_DDP_VERSION:
1046                 if (tagged) {
1047                         *layer_type = LAYER_DDP|DDP_TAGGED_ERR;
1048                         *ecode = DDPT_INV_VERS;
1049                 } else {
1050                         *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1051                         *ecode = DDPU_INV_VERS;
1052                 }
1053                 break;
1054         case T4_ERR_RDMA_VERSION:
1055                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1056                 *ecode = RDMAP_INV_VERS;
1057                 break;
1058         case T4_ERR_OPCODE:
1059                 *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP;
1060                 *ecode = RDMAP_INV_OPCODE;
1061                 break;
1062         case T4_ERR_DDP_QUEUE_NUM:
1063                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1064                 *ecode = DDPU_INV_QN;
1065                 break;
1066         case T4_ERR_MSN:
1067         case T4_ERR_MSN_GAP:
1068         case T4_ERR_MSN_RANGE:
1069         case T4_ERR_IRD_OVERFLOW:
1070                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1071                 *ecode = DDPU_INV_MSN_RANGE;
1072                 break;
1073         case T4_ERR_TBIT:
1074                 *layer_type = LAYER_DDP|DDP_LOCAL_CATA;
1075                 *ecode = 0;
1076                 break;
1077         case T4_ERR_MO:
1078                 *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR;
1079                 *ecode = DDPU_INV_MO;
1080                 break;
1081         default:
1082                 *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA;
1083                 *ecode = 0;
1084                 break;
1085         }
1086 }
1087
1088 static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe,
1089                            gfp_t gfp)
1090 {
1091         struct fw_ri_wr *wqe;
1092         struct sk_buff *skb;
1093         struct terminate_message *term;
1094
1095         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1096              qhp->ep->hwtid);
1097
1098         skb = alloc_skb(sizeof *wqe, gfp);
1099         if (!skb)
1100                 return;
1101         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1102
1103         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1104         memset(wqe, 0, sizeof *wqe);
1105         wqe->op_compl = cpu_to_be32(FW_WR_OP_V(FW_RI_INIT_WR));
1106         wqe->flowid_len16 = cpu_to_be32(
1107                 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1108                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1109
1110         wqe->u.terminate.type = FW_RI_TYPE_TERMINATE;
1111         wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term);
1112         term = (struct terminate_message *)wqe->u.terminate.termmsg;
1113         if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) {
1114                 term->layer_etype = qhp->attr.layer_etype;
1115                 term->ecode = qhp->attr.ecode;
1116         } else
1117                 build_term_codes(err_cqe, &term->layer_etype, &term->ecode);
1118         c4iw_ofld_send(&qhp->rhp->rdev, skb);
1119 }
1120
1121 /*
1122  * Assumes qhp lock is held.
1123  */
1124 static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp,
1125                        struct c4iw_cq *schp)
1126 {
1127         int count;
1128         int rq_flushed, sq_flushed;
1129         unsigned long flag;
1130
1131         PDBG("%s qhp %p rchp %p schp %p\n", __func__, qhp, rchp, schp);
1132
1133         /* locking hierarchy: cq lock first, then qp lock. */
1134         spin_lock_irqsave(&rchp->lock, flag);
1135         spin_lock(&qhp->lock);
1136
1137         if (qhp->wq.flushed) {
1138                 spin_unlock(&qhp->lock);
1139                 spin_unlock_irqrestore(&rchp->lock, flag);
1140                 return;
1141         }
1142         qhp->wq.flushed = 1;
1143
1144         c4iw_flush_hw_cq(rchp);
1145         c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count);
1146         rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count);
1147         spin_unlock(&qhp->lock);
1148         spin_unlock_irqrestore(&rchp->lock, flag);
1149
1150         /* locking hierarchy: cq lock first, then qp lock. */
1151         spin_lock_irqsave(&schp->lock, flag);
1152         spin_lock(&qhp->lock);
1153         if (schp != rchp)
1154                 c4iw_flush_hw_cq(schp);
1155         sq_flushed = c4iw_flush_sq(qhp);
1156         spin_unlock(&qhp->lock);
1157         spin_unlock_irqrestore(&schp->lock, flag);
1158
1159         if (schp == rchp) {
1160                 if (t4_clear_cq_armed(&rchp->cq) &&
1161                     (rq_flushed || sq_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         } else {
1168                 if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) {
1169                         spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1170                         (*rchp->ibcq.comp_handler)(&rchp->ibcq,
1171                                                    rchp->ibcq.cq_context);
1172                         spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1173                 }
1174                 if (t4_clear_cq_armed(&schp->cq) && sq_flushed) {
1175                         spin_lock_irqsave(&schp->comp_handler_lock, flag);
1176                         (*schp->ibcq.comp_handler)(&schp->ibcq,
1177                                                    schp->ibcq.cq_context);
1178                         spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1179                 }
1180         }
1181 }
1182
1183 static void flush_qp(struct c4iw_qp *qhp)
1184 {
1185         struct c4iw_cq *rchp, *schp;
1186         unsigned long flag;
1187
1188         rchp = to_c4iw_cq(qhp->ibqp.recv_cq);
1189         schp = to_c4iw_cq(qhp->ibqp.send_cq);
1190
1191         t4_set_wq_in_error(&qhp->wq);
1192         if (qhp->ibqp.uobject) {
1193                 t4_set_cq_in_error(&rchp->cq);
1194                 spin_lock_irqsave(&rchp->comp_handler_lock, flag);
1195                 (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context);
1196                 spin_unlock_irqrestore(&rchp->comp_handler_lock, flag);
1197                 if (schp != rchp) {
1198                         t4_set_cq_in_error(&schp->cq);
1199                         spin_lock_irqsave(&schp->comp_handler_lock, flag);
1200                         (*schp->ibcq.comp_handler)(&schp->ibcq,
1201                                         schp->ibcq.cq_context);
1202                         spin_unlock_irqrestore(&schp->comp_handler_lock, flag);
1203                 }
1204                 return;
1205         }
1206         __flush_qp(qhp, rchp, schp);
1207 }
1208
1209 static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1210                      struct c4iw_ep *ep)
1211 {
1212         struct fw_ri_wr *wqe;
1213         int ret;
1214         struct sk_buff *skb;
1215
1216         PDBG("%s qhp %p qid 0x%x tid %u\n", __func__, qhp, qhp->wq.sq.qid,
1217              ep->hwtid);
1218
1219         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1220         if (!skb)
1221                 return -ENOMEM;
1222         set_wr_txq(skb, CPL_PRIORITY_DATA, ep->txq_idx);
1223
1224         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1225         memset(wqe, 0, sizeof *wqe);
1226         wqe->op_compl = cpu_to_be32(
1227                 FW_WR_OP_V(FW_RI_INIT_WR) |
1228                 FW_WR_COMPL_F);
1229         wqe->flowid_len16 = cpu_to_be32(
1230                 FW_WR_FLOWID_V(ep->hwtid) |
1231                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1232         wqe->cookie = (uintptr_t)&ep->com.wr_wait;
1233
1234         wqe->u.fini.type = FW_RI_TYPE_FINI;
1235         ret = c4iw_ofld_send(&rhp->rdev, skb);
1236         if (ret)
1237                 goto out;
1238
1239         ret = c4iw_wait_for_reply(&rhp->rdev, &ep->com.wr_wait, qhp->ep->hwtid,
1240                              qhp->wq.sq.qid, __func__);
1241 out:
1242         PDBG("%s ret %d\n", __func__, ret);
1243         return ret;
1244 }
1245
1246 static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init)
1247 {
1248         PDBG("%s p2p_type = %d\n", __func__, p2p_type);
1249         memset(&init->u, 0, sizeof init->u);
1250         switch (p2p_type) {
1251         case FW_RI_INIT_P2PTYPE_RDMA_WRITE:
1252                 init->u.write.opcode = FW_RI_RDMA_WRITE_WR;
1253                 init->u.write.stag_sink = cpu_to_be32(1);
1254                 init->u.write.to_sink = cpu_to_be64(1);
1255                 init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD;
1256                 init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write +
1257                                                    sizeof(struct fw_ri_immd),
1258                                                    16);
1259                 break;
1260         case FW_RI_INIT_P2PTYPE_READ_REQ:
1261                 init->u.write.opcode = FW_RI_RDMA_READ_WR;
1262                 init->u.read.stag_src = cpu_to_be32(1);
1263                 init->u.read.to_src_lo = cpu_to_be32(1);
1264                 init->u.read.stag_sink = cpu_to_be32(1);
1265                 init->u.read.to_sink_lo = cpu_to_be32(1);
1266                 init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16);
1267                 break;
1268         }
1269 }
1270
1271 static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp)
1272 {
1273         struct fw_ri_wr *wqe;
1274         int ret;
1275         struct sk_buff *skb;
1276
1277         PDBG("%s qhp %p qid 0x%x tid %u ird %u ord %u\n", __func__, qhp,
1278              qhp->wq.sq.qid, qhp->ep->hwtid, qhp->ep->ird, qhp->ep->ord);
1279
1280         skb = alloc_skb(sizeof *wqe, GFP_KERNEL);
1281         if (!skb) {
1282                 ret = -ENOMEM;
1283                 goto out;
1284         }
1285         ret = alloc_ird(rhp, qhp->attr.max_ird);
1286         if (ret) {
1287                 qhp->attr.max_ird = 0;
1288                 kfree_skb(skb);
1289                 goto out;
1290         }
1291         set_wr_txq(skb, CPL_PRIORITY_DATA, qhp->ep->txq_idx);
1292
1293         wqe = (struct fw_ri_wr *)__skb_put(skb, sizeof(*wqe));
1294         memset(wqe, 0, sizeof *wqe);
1295         wqe->op_compl = cpu_to_be32(
1296                 FW_WR_OP_V(FW_RI_INIT_WR) |
1297                 FW_WR_COMPL_F);
1298         wqe->flowid_len16 = cpu_to_be32(
1299                 FW_WR_FLOWID_V(qhp->ep->hwtid) |
1300                 FW_WR_LEN16_V(DIV_ROUND_UP(sizeof(*wqe), 16)));
1301
1302         wqe->cookie = (uintptr_t)&qhp->ep->com.wr_wait;
1303
1304         wqe->u.init.type = FW_RI_TYPE_INIT;
1305         wqe->u.init.mpareqbit_p2ptype =
1306                 FW_RI_WR_MPAREQBIT_V(qhp->attr.mpa_attr.initiator) |
1307                 FW_RI_WR_P2PTYPE_V(qhp->attr.mpa_attr.p2p_type);
1308         wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE;
1309         if (qhp->attr.mpa_attr.recv_marker_enabled)
1310                 wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE;
1311         if (qhp->attr.mpa_attr.xmit_marker_enabled)
1312                 wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE;
1313         if (qhp->attr.mpa_attr.crc_enabled)
1314                 wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE;
1315
1316         wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE |
1317                             FW_RI_QP_RDMA_WRITE_ENABLE |
1318                             FW_RI_QP_BIND_ENABLE;
1319         if (!qhp->ibqp.uobject)
1320                 wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE |
1321                                      FW_RI_QP_STAG0_ENABLE;
1322         wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq));
1323         wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd);
1324         wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid);
1325         wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid);
1326         wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid);
1327         wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq);
1328         wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq);
1329         wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord);
1330         wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird);
1331         wqe->u.init.iss = cpu_to_be32(qhp->ep->snd_seq);
1332         wqe->u.init.irs = cpu_to_be32(qhp->ep->rcv_seq);
1333         wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size);
1334         wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr -
1335                                          rhp->rdev.lldi.vr->rq.start);
1336         if (qhp->attr.mpa_attr.initiator)
1337                 build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init);
1338
1339         ret = c4iw_ofld_send(&rhp->rdev, skb);
1340         if (ret)
1341                 goto err1;
1342
1343         ret = c4iw_wait_for_reply(&rhp->rdev, &qhp->ep->com.wr_wait,
1344                                   qhp->ep->hwtid, qhp->wq.sq.qid, __func__);
1345         if (!ret)
1346                 goto out;
1347 err1:
1348         free_ird(rhp, qhp->attr.max_ird);
1349 out:
1350         PDBG("%s ret %d\n", __func__, ret);
1351         return ret;
1352 }
1353
1354 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
1355                    enum c4iw_qp_attr_mask mask,
1356                    struct c4iw_qp_attributes *attrs,
1357                    int internal)
1358 {
1359         int ret = 0;
1360         struct c4iw_qp_attributes newattr = qhp->attr;
1361         int disconnect = 0;
1362         int terminate = 0;
1363         int abort = 0;
1364         int free = 0;
1365         struct c4iw_ep *ep = NULL;
1366
1367         PDBG("%s qhp %p sqid 0x%x rqid 0x%x ep %p state %d -> %d\n", __func__,
1368              qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep, qhp->attr.state,
1369              (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1);
1370
1371         mutex_lock(&qhp->mutex);
1372
1373         /* Process attr changes if in IDLE */
1374         if (mask & C4IW_QP_ATTR_VALID_MODIFY) {
1375                 if (qhp->attr.state != C4IW_QP_STATE_IDLE) {
1376                         ret = -EIO;
1377                         goto out;
1378                 }
1379                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ)
1380                         newattr.enable_rdma_read = attrs->enable_rdma_read;
1381                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE)
1382                         newattr.enable_rdma_write = attrs->enable_rdma_write;
1383                 if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND)
1384                         newattr.enable_bind = attrs->enable_bind;
1385                 if (mask & C4IW_QP_ATTR_MAX_ORD) {
1386                         if (attrs->max_ord > c4iw_max_read_depth) {
1387                                 ret = -EINVAL;
1388                                 goto out;
1389                         }
1390                         newattr.max_ord = attrs->max_ord;
1391                 }
1392                 if (mask & C4IW_QP_ATTR_MAX_IRD) {
1393                         if (attrs->max_ird > cur_max_read_depth(rhp)) {
1394                                 ret = -EINVAL;
1395                                 goto out;
1396                         }
1397                         newattr.max_ird = attrs->max_ird;
1398                 }
1399                 qhp->attr = newattr;
1400         }
1401
1402         if (mask & C4IW_QP_ATTR_SQ_DB) {
1403                 ret = ring_kernel_sq_db(qhp, attrs->sq_db_inc);
1404                 goto out;
1405         }
1406         if (mask & C4IW_QP_ATTR_RQ_DB) {
1407                 ret = ring_kernel_rq_db(qhp, attrs->rq_db_inc);
1408                 goto out;
1409         }
1410
1411         if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
1412                 goto out;
1413         if (qhp->attr.state == attrs->next_state)
1414                 goto out;
1415
1416         switch (qhp->attr.state) {
1417         case C4IW_QP_STATE_IDLE:
1418                 switch (attrs->next_state) {
1419                 case C4IW_QP_STATE_RTS:
1420                         if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) {
1421                                 ret = -EINVAL;
1422                                 goto out;
1423                         }
1424                         if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) {
1425                                 ret = -EINVAL;
1426                                 goto out;
1427                         }
1428                         qhp->attr.mpa_attr = attrs->mpa_attr;
1429                         qhp->attr.llp_stream_handle = attrs->llp_stream_handle;
1430                         qhp->ep = qhp->attr.llp_stream_handle;
1431                         set_state(qhp, C4IW_QP_STATE_RTS);
1432
1433                         /*
1434                          * Ref the endpoint here and deref when we
1435                          * disassociate the endpoint from the QP.  This
1436                          * happens in CLOSING->IDLE transition or *->ERROR
1437                          * transition.
1438                          */
1439                         c4iw_get_ep(&qhp->ep->com);
1440                         ret = rdma_init(rhp, qhp);
1441                         if (ret)
1442                                 goto err;
1443                         break;
1444                 case C4IW_QP_STATE_ERROR:
1445                         set_state(qhp, C4IW_QP_STATE_ERROR);
1446                         flush_qp(qhp);
1447                         break;
1448                 default:
1449                         ret = -EINVAL;
1450                         goto out;
1451                 }
1452                 break;
1453         case C4IW_QP_STATE_RTS:
1454                 switch (attrs->next_state) {
1455                 case C4IW_QP_STATE_CLOSING:
1456                         BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2);
1457                         t4_set_wq_in_error(&qhp->wq);
1458                         set_state(qhp, C4IW_QP_STATE_CLOSING);
1459                         ep = qhp->ep;
1460                         if (!internal) {
1461                                 abort = 0;
1462                                 disconnect = 1;
1463                                 c4iw_get_ep(&qhp->ep->com);
1464                         }
1465                         ret = rdma_fini(rhp, qhp, ep);
1466                         if (ret)
1467                                 goto err;
1468                         break;
1469                 case C4IW_QP_STATE_TERMINATE:
1470                         t4_set_wq_in_error(&qhp->wq);
1471                         set_state(qhp, C4IW_QP_STATE_TERMINATE);
1472                         qhp->attr.layer_etype = attrs->layer_etype;
1473                         qhp->attr.ecode = attrs->ecode;
1474                         ep = qhp->ep;
1475                         if (!internal) {
1476                                 c4iw_get_ep(&qhp->ep->com);
1477                                 terminate = 1;
1478                                 disconnect = 1;
1479                         } else {
1480                                 terminate = qhp->attr.send_term;
1481                                 ret = rdma_fini(rhp, qhp, ep);
1482                                 if (ret)
1483                                         goto err;
1484                         }
1485                         break;
1486                 case C4IW_QP_STATE_ERROR:
1487                         t4_set_wq_in_error(&qhp->wq);
1488                         set_state(qhp, C4IW_QP_STATE_ERROR);
1489                         if (!internal) {
1490                                 abort = 1;
1491                                 disconnect = 1;
1492                                 ep = qhp->ep;
1493                                 c4iw_get_ep(&qhp->ep->com);
1494                         }
1495                         goto err;
1496                         break;
1497                 default:
1498                         ret = -EINVAL;
1499                         goto out;
1500                 }
1501                 break;
1502         case C4IW_QP_STATE_CLOSING:
1503                 if (!internal) {
1504                         ret = -EINVAL;
1505                         goto out;
1506                 }
1507                 switch (attrs->next_state) {
1508                 case C4IW_QP_STATE_IDLE:
1509                         flush_qp(qhp);
1510                         set_state(qhp, C4IW_QP_STATE_IDLE);
1511                         qhp->attr.llp_stream_handle = NULL;
1512                         c4iw_put_ep(&qhp->ep->com);
1513                         qhp->ep = NULL;
1514                         wake_up(&qhp->wait);
1515                         break;
1516                 case C4IW_QP_STATE_ERROR:
1517                         goto err;
1518                 default:
1519                         ret = -EINVAL;
1520                         goto err;
1521                 }
1522                 break;
1523         case C4IW_QP_STATE_ERROR:
1524                 if (attrs->next_state != C4IW_QP_STATE_IDLE) {
1525                         ret = -EINVAL;
1526                         goto out;
1527                 }
1528                 if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) {
1529                         ret = -EINVAL;
1530                         goto out;
1531                 }
1532                 set_state(qhp, C4IW_QP_STATE_IDLE);
1533                 break;
1534         case C4IW_QP_STATE_TERMINATE:
1535                 if (!internal) {
1536                         ret = -EINVAL;
1537                         goto out;
1538                 }
1539                 goto err;
1540                 break;
1541         default:
1542                 printk(KERN_ERR "%s in a bad state %d\n",
1543                        __func__, qhp->attr.state);
1544                 ret = -EINVAL;
1545                 goto err;
1546                 break;
1547         }
1548         goto out;
1549 err:
1550         PDBG("%s disassociating ep %p qpid 0x%x\n", __func__, qhp->ep,
1551              qhp->wq.sq.qid);
1552
1553         /* disassociate the LLP connection */
1554         qhp->attr.llp_stream_handle = NULL;
1555         if (!ep)
1556                 ep = qhp->ep;
1557         qhp->ep = NULL;
1558         set_state(qhp, C4IW_QP_STATE_ERROR);
1559         free = 1;
1560         abort = 1;
1561         BUG_ON(!ep);
1562         flush_qp(qhp);
1563         wake_up(&qhp->wait);
1564 out:
1565         mutex_unlock(&qhp->mutex);
1566
1567         if (terminate)
1568                 post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL);
1569
1570         /*
1571          * If disconnect is 1, then we need to initiate a disconnect
1572          * on the EP.  This can be a normal close (RTS->CLOSING) or
1573          * an abnormal close (RTS/CLOSING->ERROR).
1574          */
1575         if (disconnect) {
1576                 c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC :
1577                                                          GFP_KERNEL);
1578                 c4iw_put_ep(&ep->com);
1579         }
1580
1581         /*
1582          * If free is 1, then we've disassociated the EP from the QP
1583          * and we need to dereference the EP.
1584          */
1585         if (free)
1586                 c4iw_put_ep(&ep->com);
1587         PDBG("%s exit state %d\n", __func__, qhp->attr.state);
1588         return ret;
1589 }
1590
1591 int c4iw_destroy_qp(struct ib_qp *ib_qp)
1592 {
1593         struct c4iw_dev *rhp;
1594         struct c4iw_qp *qhp;
1595         struct c4iw_qp_attributes attrs;
1596         struct c4iw_ucontext *ucontext;
1597
1598         qhp = to_c4iw_qp(ib_qp);
1599         rhp = qhp->rhp;
1600
1601         attrs.next_state = C4IW_QP_STATE_ERROR;
1602         if (qhp->attr.state == C4IW_QP_STATE_TERMINATE)
1603                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1);
1604         else
1605                 c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
1606         wait_event(qhp->wait, !qhp->ep);
1607
1608         remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1609         atomic_dec(&qhp->refcnt);
1610         wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
1611
1612         spin_lock_irq(&rhp->lock);
1613         if (!list_empty(&qhp->db_fc_entry))
1614                 list_del_init(&qhp->db_fc_entry);
1615         spin_unlock_irq(&rhp->lock);
1616         free_ird(rhp, qhp->attr.max_ird);
1617
1618         ucontext = ib_qp->uobject ?
1619                    to_c4iw_ucontext(ib_qp->uobject->context) : NULL;
1620         destroy_qp(&rhp->rdev, &qhp->wq,
1621                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1622
1623         PDBG("%s ib_qp %p qpid 0x%0x\n", __func__, ib_qp, qhp->wq.sq.qid);
1624         kfree(qhp);
1625         return 0;
1626 }
1627
1628 struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
1629                              struct ib_udata *udata)
1630 {
1631         struct c4iw_dev *rhp;
1632         struct c4iw_qp *qhp;
1633         struct c4iw_pd *php;
1634         struct c4iw_cq *schp;
1635         struct c4iw_cq *rchp;
1636         struct c4iw_create_qp_resp uresp;
1637         unsigned int sqsize, rqsize;
1638         struct c4iw_ucontext *ucontext;
1639         int ret;
1640         struct c4iw_mm_entry *mm1, *mm2, *mm3, *mm4, *mm5 = NULL;
1641
1642         PDBG("%s ib_pd %p\n", __func__, pd);
1643
1644         if (attrs->qp_type != IB_QPT_RC)
1645                 return ERR_PTR(-EINVAL);
1646
1647         php = to_c4iw_pd(pd);
1648         rhp = php->rhp;
1649         schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid);
1650         rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid);
1651         if (!schp || !rchp)
1652                 return ERR_PTR(-EINVAL);
1653
1654         if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE)
1655                 return ERR_PTR(-EINVAL);
1656
1657         if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size)
1658                 return ERR_PTR(-E2BIG);
1659         rqsize = attrs->cap.max_recv_wr + 1;
1660         if (rqsize < 8)
1661                 rqsize = 8;
1662
1663         if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size)
1664                 return ERR_PTR(-E2BIG);
1665         sqsize = attrs->cap.max_send_wr + 1;
1666         if (sqsize < 8)
1667                 sqsize = 8;
1668
1669         ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL;
1670
1671         qhp = kzalloc(sizeof(*qhp), GFP_KERNEL);
1672         if (!qhp)
1673                 return ERR_PTR(-ENOMEM);
1674         qhp->wq.sq.size = sqsize;
1675         qhp->wq.sq.memsize =
1676                 (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1677                 sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64);
1678         qhp->wq.sq.flush_cidx = -1;
1679         qhp->wq.rq.size = rqsize;
1680         qhp->wq.rq.memsize =
1681                 (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) *
1682                 sizeof(*qhp->wq.rq.queue);
1683
1684         if (ucontext) {
1685                 qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE);
1686                 qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE);
1687         }
1688
1689         ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq,
1690                         ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1691         if (ret)
1692                 goto err1;
1693
1694         attrs->cap.max_recv_wr = rqsize - 1;
1695         attrs->cap.max_send_wr = sqsize - 1;
1696         attrs->cap.max_inline_data = T4_MAX_SEND_INLINE;
1697
1698         qhp->rhp = rhp;
1699         qhp->attr.pd = php->pdid;
1700         qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid;
1701         qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid;
1702         qhp->attr.sq_num_entries = attrs->cap.max_send_wr;
1703         qhp->attr.rq_num_entries = attrs->cap.max_recv_wr;
1704         qhp->attr.sq_max_sges = attrs->cap.max_send_sge;
1705         qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge;
1706         qhp->attr.rq_max_sges = attrs->cap.max_recv_sge;
1707         qhp->attr.state = C4IW_QP_STATE_IDLE;
1708         qhp->attr.next_state = C4IW_QP_STATE_IDLE;
1709         qhp->attr.enable_rdma_read = 1;
1710         qhp->attr.enable_rdma_write = 1;
1711         qhp->attr.enable_bind = 1;
1712         qhp->attr.max_ord = 0;
1713         qhp->attr.max_ird = 0;
1714         qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR;
1715         spin_lock_init(&qhp->lock);
1716         mutex_init(&qhp->mutex);
1717         init_waitqueue_head(&qhp->wait);
1718         atomic_set(&qhp->refcnt, 1);
1719
1720         ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
1721         if (ret)
1722                 goto err2;
1723
1724         if (udata) {
1725                 mm1 = kmalloc(sizeof *mm1, GFP_KERNEL);
1726                 if (!mm1) {
1727                         ret = -ENOMEM;
1728                         goto err3;
1729                 }
1730                 mm2 = kmalloc(sizeof *mm2, GFP_KERNEL);
1731                 if (!mm2) {
1732                         ret = -ENOMEM;
1733                         goto err4;
1734                 }
1735                 mm3 = kmalloc(sizeof *mm3, GFP_KERNEL);
1736                 if (!mm3) {
1737                         ret = -ENOMEM;
1738                         goto err5;
1739                 }
1740                 mm4 = kmalloc(sizeof *mm4, GFP_KERNEL);
1741                 if (!mm4) {
1742                         ret = -ENOMEM;
1743                         goto err6;
1744                 }
1745                 if (t4_sq_onchip(&qhp->wq.sq)) {
1746                         mm5 = kmalloc(sizeof *mm5, GFP_KERNEL);
1747                         if (!mm5) {
1748                                 ret = -ENOMEM;
1749                                 goto err7;
1750                         }
1751                         uresp.flags = C4IW_QPF_ONCHIP;
1752                 } else
1753                         uresp.flags = 0;
1754                 uresp.qid_mask = rhp->rdev.qpmask;
1755                 uresp.sqid = qhp->wq.sq.qid;
1756                 uresp.sq_size = qhp->wq.sq.size;
1757                 uresp.sq_memsize = qhp->wq.sq.memsize;
1758                 uresp.rqid = qhp->wq.rq.qid;
1759                 uresp.rq_size = qhp->wq.rq.size;
1760                 uresp.rq_memsize = qhp->wq.rq.memsize;
1761                 spin_lock(&ucontext->mmap_lock);
1762                 if (mm5) {
1763                         uresp.ma_sync_key = ucontext->key;
1764                         ucontext->key += PAGE_SIZE;
1765                 } else {
1766                         uresp.ma_sync_key =  0;
1767                 }
1768                 uresp.sq_key = ucontext->key;
1769                 ucontext->key += PAGE_SIZE;
1770                 uresp.rq_key = ucontext->key;
1771                 ucontext->key += PAGE_SIZE;
1772                 uresp.sq_db_gts_key = ucontext->key;
1773                 ucontext->key += PAGE_SIZE;
1774                 uresp.rq_db_gts_key = ucontext->key;
1775                 ucontext->key += PAGE_SIZE;
1776                 spin_unlock(&ucontext->mmap_lock);
1777                 ret = ib_copy_to_udata(udata, &uresp, sizeof uresp);
1778                 if (ret)
1779                         goto err8;
1780                 mm1->key = uresp.sq_key;
1781                 mm1->addr = qhp->wq.sq.phys_addr;
1782                 mm1->len = PAGE_ALIGN(qhp->wq.sq.memsize);
1783                 insert_mmap(ucontext, mm1);
1784                 mm2->key = uresp.rq_key;
1785                 mm2->addr = virt_to_phys(qhp->wq.rq.queue);
1786                 mm2->len = PAGE_ALIGN(qhp->wq.rq.memsize);
1787                 insert_mmap(ucontext, mm2);
1788                 mm3->key = uresp.sq_db_gts_key;
1789                 mm3->addr = (__force unsigned long)qhp->wq.sq.bar2_pa;
1790                 mm3->len = PAGE_SIZE;
1791                 insert_mmap(ucontext, mm3);
1792                 mm4->key = uresp.rq_db_gts_key;
1793                 mm4->addr = (__force unsigned long)qhp->wq.rq.bar2_pa;
1794                 mm4->len = PAGE_SIZE;
1795                 insert_mmap(ucontext, mm4);
1796                 if (mm5) {
1797                         mm5->key = uresp.ma_sync_key;
1798                         mm5->addr = (pci_resource_start(rhp->rdev.lldi.pdev, 0)
1799                                     + PCIE_MA_SYNC_A) & PAGE_MASK;
1800                         mm5->len = PAGE_SIZE;
1801                         insert_mmap(ucontext, mm5);
1802                 }
1803         }
1804         qhp->ibqp.qp_num = qhp->wq.sq.qid;
1805         init_timer(&(qhp->timer));
1806         INIT_LIST_HEAD(&qhp->db_fc_entry);
1807         PDBG("%s sq id %u size %u memsize %zu num_entries %u "
1808              "rq id %u size %u memsize %zu num_entries %u\n", __func__,
1809              qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize,
1810              attrs->cap.max_send_wr, qhp->wq.rq.qid, qhp->wq.rq.size,
1811              qhp->wq.rq.memsize, attrs->cap.max_recv_wr);
1812         return &qhp->ibqp;
1813 err8:
1814         kfree(mm5);
1815 err7:
1816         kfree(mm4);
1817 err6:
1818         kfree(mm3);
1819 err5:
1820         kfree(mm2);
1821 err4:
1822         kfree(mm1);
1823 err3:
1824         remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
1825 err2:
1826         destroy_qp(&rhp->rdev, &qhp->wq,
1827                    ucontext ? &ucontext->uctx : &rhp->rdev.uctx);
1828 err1:
1829         kfree(qhp);
1830         return ERR_PTR(ret);
1831 }
1832
1833 int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1834                       int attr_mask, struct ib_udata *udata)
1835 {
1836         struct c4iw_dev *rhp;
1837         struct c4iw_qp *qhp;
1838         enum c4iw_qp_attr_mask mask = 0;
1839         struct c4iw_qp_attributes attrs;
1840
1841         PDBG("%s ib_qp %p\n", __func__, ibqp);
1842
1843         /* iwarp does not support the RTR state */
1844         if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR))
1845                 attr_mask &= ~IB_QP_STATE;
1846
1847         /* Make sure we still have something left to do */
1848         if (!attr_mask)
1849                 return 0;
1850
1851         memset(&attrs, 0, sizeof attrs);
1852         qhp = to_c4iw_qp(ibqp);
1853         rhp = qhp->rhp;
1854
1855         attrs.next_state = c4iw_convert_state(attr->qp_state);
1856         attrs.enable_rdma_read = (attr->qp_access_flags &
1857                                IB_ACCESS_REMOTE_READ) ?  1 : 0;
1858         attrs.enable_rdma_write = (attr->qp_access_flags &
1859                                 IB_ACCESS_REMOTE_WRITE) ? 1 : 0;
1860         attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0;
1861
1862
1863         mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0;
1864         mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ?
1865                         (C4IW_QP_ATTR_ENABLE_RDMA_READ |
1866                          C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
1867                          C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
1868
1869         /*
1870          * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
1871          * ringing the queue db when we're in DB_FULL mode.
1872          * Only allow this on T4 devices.
1873          */
1874         attrs.sq_db_inc = attr->sq_psn;
1875         attrs.rq_db_inc = attr->rq_psn;
1876         mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
1877         mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
1878         if (is_t5(to_c4iw_qp(ibqp)->rhp->rdev.lldi.adapter_type) &&
1879             (mask & (C4IW_QP_ATTR_SQ_DB|C4IW_QP_ATTR_RQ_DB)))
1880                 return -EINVAL;
1881
1882         return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
1883 }
1884
1885 struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn)
1886 {
1887         PDBG("%s ib_dev %p qpn 0x%x\n", __func__, dev, qpn);
1888         return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn);
1889 }
1890
1891 int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1892                      int attr_mask, struct ib_qp_init_attr *init_attr)
1893 {
1894         struct c4iw_qp *qhp = to_c4iw_qp(ibqp);
1895
1896         memset(attr, 0, sizeof *attr);
1897         memset(init_attr, 0, sizeof *init_attr);
1898         attr->qp_state = to_ib_qp_state(qhp->attr.state);
1899         init_attr->cap.max_send_wr = qhp->attr.sq_num_entries;
1900         init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries;
1901         init_attr->cap.max_send_sge = qhp->attr.sq_max_sges;
1902         init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges;
1903         init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE;
1904         init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0;
1905         return 0;
1906 }