drivers/infiniband/hw/cxgb3/cxio_hal.c

   1 /*
   2  * Copyright (c) 2006 Chelsio, Inc. All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  */
  32 #include <asm/delay.h>
  33
  34 #include <linux/mutex.h>
  35 #include <linux/netdevice.h>
  36 #include <linux/sched.h>
  37 #include <linux/spinlock.h>
  38 #include <linux/pci.h>
  39 #include <linux/dma-mapping.h>
  40
  41 #include "cxio_resource.h"
  42 #include "cxio_hal.h"
  43 #include "cxgb3_offload.h"
  44 #include "sge_defs.h"
  45
  46 static LIST_HEAD(rdev_list);
  47 static cxio_hal_ev_callback_func_t cxio_ev_cb = NULL;
  48
  49 static struct cxio_rdev *cxio_hal_find_rdev_by_name(char *dev_name)
  50 {
  51         struct cxio_rdev *rdev;
  52
  53         list_for_each_entry(rdev, &rdev_list, entry)
  54                 if (!strcmp(rdev->dev_name, dev_name))
  55                         return rdev;
  56         return NULL;
  57 }
  58
  59 static struct cxio_rdev *cxio_hal_find_rdev_by_t3cdev(struct t3cdev *tdev)
  60 {
  61         struct cxio_rdev *rdev;
  62
  63         list_for_each_entry(rdev, &rdev_list, entry)
  64                 if (rdev->t3cdev_p == tdev)
  65                         return rdev;
  66         return NULL;
  67 }
  68
  69 int cxio_hal_cq_op(struct cxio_rdev *rdev_p, struct t3_cq *cq,
  70                    enum t3_cq_opcode op, u32 credit)
  71 {
  72         int ret;
  73         struct t3_cqe *cqe;
  74         u32 rptr;
  75
  76         struct rdma_cq_op setup;
  77         setup.id = cq->cqid;
  78         setup.credits = (op == CQ_CREDIT_UPDATE) ? credit : 0;
  79         setup.op = op;
  80         ret = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_OP, &setup);
  81
  82         if ((ret < 0) || (op == CQ_CREDIT_UPDATE))
  83                 return ret;
  84
  85         /*
  86          * If the rearm returned an index other than our current index,
  87          * then there might be CQE's in flight (being DMA'd).  We must wait
  88          * here for them to complete or the consumer can miss a notification.
  89          */
  90         if (Q_PTR2IDX((cq->rptr), cq->size_log2) != ret) {
  91                 int i=0;
  92
  93                 rptr = cq->rptr;
  94
  95                 /*
  96                  * Keep the generation correct by bumping rptr until it
  97                  * matches the index returned by the rearm - 1.
  98                  */
  99                 while (Q_PTR2IDX((rptr+1), cq->size_log2) != ret)
 100                         rptr++;
 101
 102                 /*
 103                  * Now rptr is the index for the (last) cqe that was
 104                  * in-flight at the time the HW rearmed the CQ.  We
 105                  * spin until that CQE is valid.
 106                  */
 107                 cqe = cq->queue + Q_PTR2IDX(rptr, cq->size_log2);
 108                 while (!CQ_VLD_ENTRY(rptr, cq->size_log2, cqe)) {
 109                         udelay(1);
 110                         if (i++ > 1000000) {
 111                                 BUG_ON(1);
 112                                 printk(KERN_ERR "%s: stalled rnic\n",
 113                                        rdev_p->dev_name);
 114                                 return -EIO;
 115                         }
 116                 }
 117         }
 118         return 0;
 119 }
 120
 121 static int cxio_hal_clear_cq_ctx(struct cxio_rdev *rdev_p, u32 cqid)
 122 {
 123         struct rdma_cq_setup setup;
 124         setup.id = cqid;
 125         setup.base_addr = 0;    /* NULL address */
 126         setup.size = 0;         /* disaable the CQ */
 127         setup.credits = 0;
 128         setup.credit_thres = 0;
 129         setup.ovfl_mode = 0;
 130         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 131 }
 132
 133 static int cxio_hal_clear_qp_ctx(struct cxio_rdev *rdev_p, u32 qpid)
 134 {
 135         u64 sge_cmd;
 136         struct t3_modify_qp_wr *wqe;
 137         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
 138         if (!skb) {
 139                 PDBG("%s alloc_skb failed\n", __FUNCTION__);
 140                 return -ENOMEM;
 141         }
 142         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
 143         memset(wqe, 0, sizeof(*wqe));
 144         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 3, 1, qpid, 7);
 145         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
 146         sge_cmd = qpid << 8 | 3;
 147         wqe->sge_cmd = cpu_to_be64(sge_cmd);
 148         skb->priority = CPL_PRIORITY_CONTROL;
 149         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 150 }
 151
 152 int cxio_create_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 153 {
 154         struct rdma_cq_setup setup;
 155         int size = (1UL << (cq->size_log2)) * sizeof(struct t3_cqe);
 156
 157         cq->cqid = cxio_hal_get_cqid(rdev_p->rscp);
 158         if (!cq->cqid)
 159                 return -ENOMEM;
 160         cq->sw_queue = kzalloc(size, GFP_KERNEL);
 161         if (!cq->sw_queue)
 162                 return -ENOMEM;
 163         cq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
 164                                              (1UL << (cq->size_log2)) *
 165                                              sizeof(struct t3_cqe),
 166                                              &(cq->dma_addr), GFP_KERNEL);
 167         if (!cq->queue) {
 168                 kfree(cq->sw_queue);
 169                 return -ENOMEM;
 170         }
 171         pci_unmap_addr_set(cq, mapping, cq->dma_addr);
 172         memset(cq->queue, 0, size);
 173         setup.id = cq->cqid;
 174         setup.base_addr = (u64) (cq->dma_addr);
 175         setup.size = 1UL << cq->size_log2;
 176         setup.credits = 65535;
 177         setup.credit_thres = 1;
 178         if (rdev_p->t3cdev_p->type == T3B)
 179                 setup.ovfl_mode = 0;
 180         else
 181                 setup.ovfl_mode = 1;
 182         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 183 }
 184
 185 int cxio_resize_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 186 {
 187         struct rdma_cq_setup setup;
 188         setup.id = cq->cqid;
 189         setup.base_addr = (u64) (cq->dma_addr);
 190         setup.size = 1UL << cq->size_log2;
 191         setup.credits = setup.size;
 192         setup.credit_thres = setup.size;        /* TBD: overflow recovery */
 193         setup.ovfl_mode = 1;
 194         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 195 }
 196
 197 static u32 get_qpid(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 198 {
 199         struct cxio_qpid_list *entry;
 200         u32 qpid;
 201         int i;
 202
 203         mutex_lock(&uctx->lock);
 204         if (!list_empty(&uctx->qpids)) {
 205                 entry = list_entry(uctx->qpids.next, struct cxio_qpid_list,
 206                                    entry);
 207                 list_del(&entry->entry);
 208                 qpid = entry->qpid;
 209                 kfree(entry);
 210         } else {
 211                 qpid = cxio_hal_get_qpid(rdev_p->rscp);
 212                 if (!qpid)
 213                         goto out;
 214                 for (i = qpid+1; i & rdev_p->qpmask; i++) {
 215                         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 216                         if (!entry)
 217                                 break;
 218                         entry->qpid = i;
 219                         list_add_tail(&entry->entry, &uctx->qpids);
 220                 }
 221         }
 222 out:
 223         mutex_unlock(&uctx->lock);
 224         PDBG("%s qpid 0x%x\n", __FUNCTION__, qpid);
 225         return qpid;
 226 }
 227
 228 static void put_qpid(struct cxio_rdev *rdev_p, u32 qpid,
 229                      struct cxio_ucontext *uctx)
 230 {
 231         struct cxio_qpid_list *entry;
 232
 233         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 234         if (!entry)
 235                 return;
 236         PDBG("%s qpid 0x%x\n", __FUNCTION__, qpid);
 237         entry->qpid = qpid;
 238         mutex_lock(&uctx->lock);
 239         list_add_tail(&entry->entry, &uctx->qpids);
 240         mutex_unlock(&uctx->lock);
 241 }
 242
 243 void cxio_release_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 244 {
 245         struct list_head *pos, *nxt;
 246         struct cxio_qpid_list *entry;
 247
 248         mutex_lock(&uctx->lock);
 249         list_for_each_safe(pos, nxt, &uctx->qpids) {
 250                 entry = list_entry(pos, struct cxio_qpid_list, entry);
 251                 list_del_init(&entry->entry);
 252                 if (!(entry->qpid & rdev_p->qpmask))
 253                         cxio_hal_put_qpid(rdev_p->rscp, entry->qpid);
 254                 kfree(entry);
 255         }
 256         mutex_unlock(&uctx->lock);
 257 }
 258
 259 void cxio_init_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 260 {
 261         INIT_LIST_HEAD(&uctx->qpids);
 262         mutex_init(&uctx->lock);
 263 }
 264
 265 int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
 266                    struct t3_wq *wq, struct cxio_ucontext *uctx)
 267 {
 268         int depth = 1UL << wq->size_log2;
 269         int rqsize = 1UL << wq->rq_size_log2;
 270
 271         wq->qpid = get_qpid(rdev_p, uctx);
 272         if (!wq->qpid)
 273                 return -ENOMEM;
 274
 275         wq->rq = kzalloc(depth * sizeof(u64), GFP_KERNEL);
 276         if (!wq->rq)
 277                 goto err1;
 278
 279         wq->rq_addr = cxio_hal_rqtpool_alloc(rdev_p, rqsize);
 280         if (!wq->rq_addr)
 281                 goto err2;
 282
 283         wq->sq = kzalloc(depth * sizeof(struct t3_swsq), GFP_KERNEL);
 284         if (!wq->sq)
 285                 goto err3;
 286
 287         wq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
 288                                              depth * sizeof(union t3_wr),
 289                                              &(wq->dma_addr), GFP_KERNEL);
 290         if (!wq->queue)
 291                 goto err4;
 292
 293         memset(wq->queue, 0, depth * sizeof(union t3_wr));
 294         pci_unmap_addr_set(wq, mapping, wq->dma_addr);
 295         wq->doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
 296         if (!kernel_domain)
 297                 wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
 298                                         (wq->qpid << rdev_p->qpshift);
 299         PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __FUNCTION__,
 300              wq->qpid, wq->doorbell, (unsigned long long) wq->udb);
 301         return 0;
 302 err4:
 303         kfree(wq->sq);
 304 err3:
 305         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, rqsize);
 306 err2:
 307         kfree(wq->rq);
 308 err1:
 309         put_qpid(rdev_p, wq->qpid, uctx);
 310         return -ENOMEM;
 311 }
 312
 313 int cxio_destroy_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 314 {
 315         int err;
 316         err = cxio_hal_clear_cq_ctx(rdev_p, cq->cqid);
 317         kfree(cq->sw_queue);
 318         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 319                           (1UL << (cq->size_log2))
 320                           * sizeof(struct t3_cqe), cq->queue,
 321                           pci_unmap_addr(cq, mapping));
 322         cxio_hal_put_cqid(rdev_p->rscp, cq->cqid);
 323         return err;
 324 }
 325
 326 int cxio_destroy_qp(struct cxio_rdev *rdev_p, struct t3_wq *wq,
 327                     struct cxio_ucontext *uctx)
 328 {
 329         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 330                           (1UL << (wq->size_log2))
 331                           * sizeof(union t3_wr), wq->queue,
 332                           pci_unmap_addr(wq, mapping));
 333         kfree(wq->sq);
 334         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, (1UL << wq->rq_size_log2));
 335         kfree(wq->rq);
 336         put_qpid(rdev_p, wq->qpid, uctx);
 337         return 0;
 338 }
 339
 340 static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq)
 341 {
 342         struct t3_cqe cqe;
 343
 344         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __FUNCTION__,
 345              wq, cq, cq->sw_rptr, cq->sw_wptr);
 346         memset(&cqe, 0, sizeof(cqe));
 347         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
 348                                  V_CQE_OPCODE(T3_SEND) |
 349                                  V_CQE_TYPE(0) |
 350                                  V_CQE_SWCQE(1) |
 351                                  V_CQE_QPID(wq->qpid) |
 352                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
 353                                                        cq->size_log2)));
 354         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
 355         cq->sw_wptr++;
 356 }
 357
 358 void cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count)
 359 {
 360         u32 ptr;
 361
 362         PDBG("%s wq %p cq %p\n", __FUNCTION__, wq, cq);
 363
 364         /* flush RQ */
 365         PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __FUNCTION__,
 366             wq->rq_rptr, wq->rq_wptr, count);
 367         ptr = wq->rq_rptr + count;
 368         while (ptr++ != wq->rq_wptr)
 369                 insert_recv_cqe(wq, cq);
 370 }
 371
 372 static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq,
 373                           struct t3_swsq *sqp)
 374 {
 375         struct t3_cqe cqe;
 376
 377         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __FUNCTION__,
 378              wq, cq, cq->sw_rptr, cq->sw_wptr);
 379         memset(&cqe, 0, sizeof(cqe));
 380         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
 381                                  V_CQE_OPCODE(sqp->opcode) |
 382                                  V_CQE_TYPE(1) |
 383                                  V_CQE_SWCQE(1) |
 384                                  V_CQE_QPID(wq->qpid) |
 385                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
 386                                                        cq->size_log2)));
 387         cqe.u.scqe.wrid_hi = sqp->sq_wptr;
 388
 389         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
 390         cq->sw_wptr++;
 391 }
 392
 393 void cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count)
 394 {
 395         __u32 ptr;
 396         struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2);
 397
 398         ptr = wq->sq_rptr + count;
 399         sqp += count;
 400         while (ptr != wq->sq_wptr) {
 401                 insert_sq_cqe(wq, cq, sqp);
 402                 sqp++;
 403                 ptr++;
 404         }
 405 }
 406
 407 /*
 408  * Move all CQEs from the HWCQ into the SWCQ.
 409  */
 410 void cxio_flush_hw_cq(struct t3_cq *cq)
 411 {
 412         struct t3_cqe *cqe, *swcqe;
 413
 414         PDBG("%s cq %p cqid 0x%x\n", __FUNCTION__, cq, cq->cqid);
 415         cqe = cxio_next_hw_cqe(cq);
 416         while (cqe) {
 417                 PDBG("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
 418                      __FUNCTION__, cq->rptr, cq->sw_wptr);
 419                 swcqe = cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2);
 420                 *swcqe = *cqe;
 421                 swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
 422                 cq->sw_wptr++;
 423                 cq->rptr++;
 424                 cqe = cxio_next_hw_cqe(cq);
 425         }
 426 }
 427
 428 static int cqe_completes_wr(struct t3_cqe *cqe, struct t3_wq *wq)
 429 {
 430         if (CQE_OPCODE(*cqe) == T3_TERMINATE)
 431                 return 0;
 432
 433         if ((CQE_OPCODE(*cqe) == T3_RDMA_WRITE) && RQ_TYPE(*cqe))
 434                 return 0;
 435
 436         if ((CQE_OPCODE(*cqe) == T3_READ_RESP) && SQ_TYPE(*cqe))
 437                 return 0;
 438
 439         if ((CQE_OPCODE(*cqe) == T3_SEND) && RQ_TYPE(*cqe) &&
 440             Q_EMPTY(wq->rq_rptr, wq->rq_wptr))
 441                 return 0;
 442
 443         return 1;
 444 }
 445
 446 void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
 447 {
 448         struct t3_cqe *cqe;
 449         u32 ptr;
 450
 451         *count = 0;
 452         ptr = cq->sw_rptr;
 453         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
 454                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
 455                 if ((SQ_TYPE(*cqe) || (CQE_OPCODE(*cqe) == T3_READ_RESP)) &&
 456                     (CQE_QPID(*cqe) == wq->qpid))
 457                         (*count)++;
 458                 ptr++;
 459         }
 460         PDBG("%s cq %p count %d\n", __FUNCTION__, cq, *count);
 461 }
 462
 463 void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
 464 {
 465         struct t3_cqe *cqe;
 466         u32 ptr;
 467
 468         *count = 0;
 469         PDBG("%s count zero %d\n", __FUNCTION__, *count);
 470         ptr = cq->sw_rptr;
 471         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
 472                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
 473                 if (RQ_TYPE(*cqe) && (CQE_OPCODE(*cqe) != T3_READ_RESP) &&
 474                     (CQE_QPID(*cqe) == wq->qpid) && cqe_completes_wr(cqe, wq))
 475                         (*count)++;
 476                 ptr++;
 477         }
 478         PDBG("%s cq %p count %d\n", __FUNCTION__, cq, *count);
 479 }
 480
 481 static int cxio_hal_init_ctrl_cq(struct cxio_rdev *rdev_p)
 482 {
 483         struct rdma_cq_setup setup;
 484         setup.id = 0;
 485         setup.base_addr = 0;    /* NULL address */
 486         setup.size = 1;         /* enable the CQ */
 487         setup.credits = 0;
 488
 489         /* force SGE to redirect to RspQ and interrupt */
 490         setup.credit_thres = 0;
 491         setup.ovfl_mode = 1;
 492         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 493 }
 494
 495 static int cxio_hal_init_ctrl_qp(struct cxio_rdev *rdev_p)
 496 {
 497         int err;
 498         u64 sge_cmd, ctx0, ctx1;
 499         u64 base_addr;
 500         struct t3_modify_qp_wr *wqe;
 501         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
 502
 503
 504         if (!skb) {
 505                 PDBG("%s alloc_skb failed\n", __FUNCTION__);
 506                 return -ENOMEM;
 507         }
 508         err = cxio_hal_init_ctrl_cq(rdev_p);
 509         if (err) {
 510                 PDBG("%s err %d initializing ctrl_cq\n", __FUNCTION__, err);
 511                 return err;
 512         }
 513         rdev_p->ctrl_qp.workq = dma_alloc_coherent(
 514                                         &(rdev_p->rnic_info.pdev->dev),
 515                                         (1 << T3_CTRL_QP_SIZE_LOG2) *
 516                                         sizeof(union t3_wr),
 517                                         &(rdev_p->ctrl_qp.dma_addr),
 518                                         GFP_KERNEL);
 519         if (!rdev_p->ctrl_qp.workq) {
 520                 PDBG("%s dma_alloc_coherent failed\n", __FUNCTION__);
 521                 return -ENOMEM;
 522         }
 523         pci_unmap_addr_set(&rdev_p->ctrl_qp, mapping,
 524                            rdev_p->ctrl_qp.dma_addr);
 525         rdev_p->ctrl_qp.doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
 526         memset(rdev_p->ctrl_qp.workq, 0,
 527                (1 << T3_CTRL_QP_SIZE_LOG2) * sizeof(union t3_wr));
 528
 529         mutex_init(&rdev_p->ctrl_qp.lock);
 530         init_waitqueue_head(&rdev_p->ctrl_qp.waitq);
 531
 532         /* update HW Ctrl QP context */
 533         base_addr = rdev_p->ctrl_qp.dma_addr;
 534         base_addr >>= 12;
 535         ctx0 = (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2)) |
 536                 V_EC_BASE_LO((u32) base_addr & 0xffff));
 537         ctx0 <<= 32;
 538         ctx0 |= V_EC_CREDITS(FW_WR_NUM);
 539         base_addr >>= 16;
 540         ctx1 = (u32) base_addr;
 541         base_addr >>= 32;
 542         ctx1 |= ((u64) (V_EC_BASE_HI((u32) base_addr & 0xf) | V_EC_RESPQ(0) |
 543                         V_EC_TYPE(0) | V_EC_GEN(1) |
 544                         V_EC_UP_TOKEN(T3_CTL_QP_TID) | F_EC_VALID)) << 32;
 545         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
 546         memset(wqe, 0, sizeof(*wqe));
 547         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 0, 1,
 548                        T3_CTL_QP_TID, 7);
 549         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
 550         sge_cmd = (3ULL << 56) | FW_RI_SGEEC_START << 8 | 3;
 551         wqe->sge_cmd = cpu_to_be64(sge_cmd);
 552         wqe->ctx1 = cpu_to_be64(ctx1);
 553         wqe->ctx0 = cpu_to_be64(ctx0);
 554         PDBG("CtrlQP dma_addr 0x%llx workq %p size %d\n",
 555              (unsigned long long) rdev_p->ctrl_qp.dma_addr,
 556              rdev_p->ctrl_qp.workq, 1 << T3_CTRL_QP_SIZE_LOG2);
 557         skb->priority = CPL_PRIORITY_CONTROL;
 558         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 559 }
 560
 561 static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
 562 {
 563         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 564                           (1UL << T3_CTRL_QP_SIZE_LOG2)
 565                           * sizeof(union t3_wr), rdev_p->ctrl_qp.workq,
 566                           pci_unmap_addr(&rdev_p->ctrl_qp, mapping));
 567         return cxio_hal_clear_qp_ctx(rdev_p, T3_CTRL_QP_ID);
 568 }
 569
 570 /* write len bytes of data into addr (32B aligned address)
 571  * If data is NULL, clear len byte of memory to zero.
 572  * caller aquires the ctrl_qp lock before the call
 573  */
 574 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
 575                                       u32 len, void *data, int completion)
 576 {
 577         u32 i, nr_wqe, copy_len;
 578         u8 *copy_data;
 579         u8 wr_len, utx_len;     /* lenght in 8 byte flit */
 580         enum t3_wr_flags flag;
 581         __be64 *wqe;
 582         u64 utx_cmd;
 583         addr &= 0x7FFFFFF;
 584         nr_wqe = len % 96 ? len / 96 + 1 : len / 96;    /* 96B max per WQE */
 585         PDBG("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
 586              __FUNCTION__, rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, len,
 587              nr_wqe, data, addr);
 588         utx_len = 3;            /* in 32B unit */
 589         for (i = 0; i < nr_wqe; i++) {
 590                 if (Q_FULL(rdev_p->ctrl_qp.rptr, rdev_p->ctrl_qp.wptr,
 591                            T3_CTRL_QP_SIZE_LOG2)) {
 592                         PDBG("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, "
 593                              "wait for more space i %d\n", __FUNCTION__,
 594                              rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, i);
 595                         if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
 596                                              !Q_FULL(rdev_p->ctrl_qp.rptr,
 597                                                      rdev_p->ctrl_qp.wptr,
 598                                                      T3_CTRL_QP_SIZE_LOG2))) {
 599                                 PDBG("%s ctrl_qp workq interrupted\n",
 600                                      __FUNCTION__);
 601                                 return -ERESTARTSYS;
 602                         }
 603                         PDBG("%s ctrl_qp wakeup, continue posting work request "
 604                              "i %d\n", __FUNCTION__, i);
 605                 }
 606                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
 607                                                 (1 << T3_CTRL_QP_SIZE_LOG2)));
 608                 flag = 0;
 609                 if (i == (nr_wqe - 1)) {
 610                         /* last WQE */
 611                         flag = completion ? T3_COMPLETION_FLAG : 0;
 612                         if (len % 32)
 613                                 utx_len = len / 32 + 1;
 614                         else
 615                                 utx_len = len / 32;
 616                 }
 617
 618                 /*
 619                  * Force a CQE to return the credit to the workq in case
 620                  * we posted more than half the max QP size of WRs
 621                  */
 622                 if ((i != 0) &&
 623                     (i % (((1 << T3_CTRL_QP_SIZE_LOG2)) >> 1) == 0)) {
 624                         flag = T3_COMPLETION_FLAG;
 625                         PDBG("%s force completion at i %d\n", __FUNCTION__, i);
 626                 }
 627
 628                 /* build the utx mem command */
 629                 wqe += (sizeof(struct t3_bypass_wr) >> 3);
 630                 utx_cmd = (T3_UTX_MEM_WRITE << 28) | (addr + i * 3);
 631                 utx_cmd <<= 32;
 632                 utx_cmd |= (utx_len << 28) | ((utx_len << 2) + 1);
 633                 *wqe = cpu_to_be64(utx_cmd);
 634                 wqe++;
 635                 copy_data = (u8 *) data + i * 96;
 636                 copy_len = len > 96 ? 96 : len;
 637
 638                 /* clear memory content if data is NULL */
 639                 if (data)
 640                         memcpy(wqe, copy_data, copy_len);
 641                 else
 642                         memset(wqe, 0, copy_len);
 643                 if (copy_len % 32)
 644                         memset(((u8 *) wqe) + copy_len, 0,
 645                                32 - (copy_len % 32));
 646                 wr_len = ((sizeof(struct t3_bypass_wr)) >> 3) + 1 +
 647                          (utx_len << 2);
 648                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
 649                               (1 << T3_CTRL_QP_SIZE_LOG2)));
 650
 651                 /* wptr in the WRID[31:0] */
 652                 ((union t3_wrid *)(wqe+1))->id0.low = rdev_p->ctrl_qp.wptr;
 653
 654                 /*
 655                  * This must be the last write with a memory barrier
 656                  * for the genbit
 657                  */
 658                 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_BP, flag,
 659                                Q_GENBIT(rdev_p->ctrl_qp.wptr,
 660                                         T3_CTRL_QP_SIZE_LOG2), T3_CTRL_QP_ID,
 661                                wr_len);
 662                 if (flag == T3_COMPLETION_FLAG)
 663                         ring_doorbell(rdev_p->ctrl_qp.doorbell, T3_CTRL_QP_ID);
 664                 len -= 96;
 665                 rdev_p->ctrl_qp.wptr++;
 666         }
 667         return 0;
 668 }
 669
 670 /* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl, and pbl_size
 671  * OUT: stag index, actual pbl_size, pbl_addr allocated.
 672  * TBD: shared memory region support
 673  */
 674 static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
 675                          u32 *stag, u8 stag_state, u32 pdid,
 676                          enum tpt_mem_type type, enum tpt_mem_perm perm,
 677                          u32 zbva, u64 to, u32 len, u8 page_size, __be64 *pbl,
 678                          u32 *pbl_size, u32 *pbl_addr)
 679 {
 680         int err;
 681         struct tpt_entry tpt;
 682         u32 stag_idx;
 683         u32 wptr;
 684         int rereg = (*stag != T3_STAG_UNSET);
 685
 686         stag_state = stag_state > 0;
 687         stag_idx = (*stag) >> 8;
 688
 689         if ((!reset_tpt_entry) && !(*stag != T3_STAG_UNSET)) {
 690                 stag_idx = cxio_hal_get_stag(rdev_p->rscp);
 691                 if (!stag_idx)
 692                         return -ENOMEM;
 693                 *stag = (stag_idx << 8) | ((*stag) & 0xFF);
 694         }
 695         PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
 696              __FUNCTION__, stag_state, type, pdid, stag_idx);
 697
 698         if (reset_tpt_entry)
 699                 cxio_hal_pblpool_free(rdev_p, *pbl_addr, *pbl_size << 3);
 700         else if (!rereg) {
 701                 *pbl_addr = cxio_hal_pblpool_alloc(rdev_p, *pbl_size << 3);
 702                 if (!*pbl_addr) {
 703                         return -ENOMEM;
 704                 }
 705         }
 706
 707         mutex_lock(&rdev_p->ctrl_qp.lock);
 708
 709         /* write PBL first if any - update pbl only if pbl list exist */
 710         if (pbl) {
 711
 712                 PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
 713                      __FUNCTION__, *pbl_addr, rdev_p->rnic_info.pbl_base,
 714                      *pbl_size);
 715                 err = cxio_hal_ctrl_qp_write_mem(rdev_p,
 716                                 (*pbl_addr >> 5),
 717                                 (*pbl_size << 3), pbl, 0);
 718                 if (err)
 719                         goto ret;
 720         }
 721
 722         /* write TPT entry */
 723         if (reset_tpt_entry)
 724                 memset(&tpt, 0, sizeof(tpt));
 725         else {
 726                 tpt.valid_stag_pdid = cpu_to_be32(F_TPT_VALID |
 727                                 V_TPT_STAG_KEY((*stag) & M_TPT_STAG_KEY) |
 728                                 V_TPT_STAG_STATE(stag_state) |
 729                                 V_TPT_STAG_TYPE(type) | V_TPT_PDID(pdid));
 730                 BUG_ON(page_size >= 28);
 731                 tpt.flags_pagesize_qpid = cpu_to_be32(V_TPT_PERM(perm) |
 732                                 F_TPT_MW_BIND_ENABLE |
 733                                 V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
 734                                 V_TPT_PAGE_SIZE(page_size));
 735                 tpt.rsvd_pbl_addr = reset_tpt_entry ? 0 :
 736                                     cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, *pbl_addr)>>3));
 737                 tpt.len = cpu_to_be32(len);
 738                 tpt.va_hi = cpu_to_be32((u32) (to >> 32));
 739                 tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
 740                 tpt.rsvd_bind_cnt_or_pstag = 0;
 741                 tpt.rsvd_pbl_size = reset_tpt_entry ? 0 :
 742                                   cpu_to_be32(V_TPT_PBL_SIZE((*pbl_size) >> 2));
 743         }
 744         err = cxio_hal_ctrl_qp_write_mem(rdev_p,
 745                                        stag_idx +
 746                                        (rdev_p->rnic_info.tpt_base >> 5),
 747                                        sizeof(tpt), &tpt, 1);
 748
 749         /* release the stag index to free pool */
 750         if (reset_tpt_entry)
 751                 cxio_hal_put_stag(rdev_p->rscp, stag_idx);
 752 ret:
 753         wptr = rdev_p->ctrl_qp.wptr;
 754         mutex_unlock(&rdev_p->ctrl_qp.lock);
 755         if (!err)
 756                 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
 757                                              SEQ32_GE(rdev_p->ctrl_qp.rptr,
 758                                                       wptr)))
 759                         return -ERESTARTSYS;
 760         return err;
 761 }
 762
 763 int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 764                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
 765                            u8 page_size, __be64 *pbl, u32 *pbl_size,
 766                            u32 *pbl_addr)
 767 {
 768         *stag = T3_STAG_UNSET;
 769         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
 770                              zbva, to, len, page_size, pbl, pbl_size, pbl_addr);
 771 }
 772
 773 int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 774                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
 775                            u8 page_size, __be64 *pbl, u32 *pbl_size,
 776                            u32 *pbl_addr)
 777 {
 778         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
 779                              zbva, to, len, page_size, pbl, pbl_size, pbl_addr);
 780 }
 781
 782 int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
 783                    u32 pbl_addr)
 784 {
 785         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL,
 786                              &pbl_size, &pbl_addr);
 787 }
 788
 789 int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
 790 {
 791         u32 pbl_size = 0;
 792         *stag = T3_STAG_UNSET;
 793         return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
 794                              NULL, &pbl_size, NULL);
 795 }
 796
 797 int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
 798 {
 799         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0, NULL,
 800                              NULL, NULL);
 801 }
 802
 803 int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
 804 {
 805         struct t3_rdma_init_wr *wqe;
 806         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_ATOMIC);
 807         if (!skb)
 808                 return -ENOMEM;
 809         PDBG("%s rdev_p %p\n", __FUNCTION__, rdev_p);
 810         wqe = (struct t3_rdma_init_wr *) __skb_put(skb, sizeof(*wqe));
 811         wqe->wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT));
 812         wqe->wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(attr->tid) |
 813                                            V_FW_RIWR_LEN(sizeof(*wqe) >> 3));
 814         wqe->wrid.id1 = 0;
 815         wqe->qpid = cpu_to_be32(attr->qpid);
 816         wqe->pdid = cpu_to_be32(attr->pdid);
 817         wqe->scqid = cpu_to_be32(attr->scqid);
 818         wqe->rcqid = cpu_to_be32(attr->rcqid);
 819         wqe->rq_addr = cpu_to_be32(attr->rq_addr - rdev_p->rnic_info.rqt_base);
 820         wqe->rq_size = cpu_to_be32(attr->rq_size);
 821         wqe->mpaattrs = attr->mpaattrs;
 822         wqe->qpcaps = attr->qpcaps;
 823         wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
 824         wqe->flags = cpu_to_be32(attr->flags);
 825         wqe->ord = cpu_to_be32(attr->ord);
 826         wqe->ird = cpu_to_be32(attr->ird);
 827         wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
 828         wqe->qp_dma_size = cpu_to_be32(attr->qp_dma_size);
 829         wqe->rsvd = 0;
 830         skb->priority = 0;      /* 0=>ToeQ; 1=>CtrlQ */
 831         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 832 }
 833
 834 void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
 835 {
 836         cxio_ev_cb = ev_cb;
 837 }
 838
 839 void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
 840 {
 841         cxio_ev_cb = NULL;
 842 }
 843
 844 static int cxio_hal_ev_handler(struct t3cdev *t3cdev_p, struct sk_buff *skb)
 845 {
 846         static int cnt;
 847         struct cxio_rdev *rdev_p = NULL;
 848         struct respQ_msg_t *rsp_msg = (struct respQ_msg_t *) skb->data;
 849         PDBG("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x"
 850              " se %0x notify %0x cqbranch %0x creditth %0x\n",
 851              cnt, __FUNCTION__, RSPQ_CQID(rsp_msg), RSPQ_CQPTR(rsp_msg),
 852              RSPQ_GENBIT(rsp_msg), RSPQ_OVERFLOW(rsp_msg), RSPQ_AN(rsp_msg),
 853              RSPQ_SE(rsp_msg), RSPQ_NOTIFY(rsp_msg), RSPQ_CQBRANCH(rsp_msg),
 854              RSPQ_CREDIT_THRESH(rsp_msg));
 855         PDBG("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d "
 856              "len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
 857              CQE_QPID(rsp_msg->cqe), CQE_GENBIT(rsp_msg->cqe),
 858              CQE_TYPE(rsp_msg->cqe), CQE_STATUS(rsp_msg->cqe),
 859              CQE_OPCODE(rsp_msg->cqe), CQE_LEN(rsp_msg->cqe),
 860              CQE_WRID_HI(rsp_msg->cqe), CQE_WRID_LOW(rsp_msg->cqe));
 861         rdev_p = (struct cxio_rdev *)t3cdev_p->ulp;
 862         if (!rdev_p) {
 863                 PDBG("%s called by t3cdev %p with null ulp\n", __FUNCTION__,
 864                      t3cdev_p);
 865                 return 0;
 866         }
 867         if (CQE_QPID(rsp_msg->cqe) == T3_CTRL_QP_ID) {
 868                 rdev_p->ctrl_qp.rptr = CQE_WRID_LOW(rsp_msg->cqe) + 1;
 869                 wake_up_interruptible(&rdev_p->ctrl_qp.waitq);
 870                 dev_kfree_skb_irq(skb);
 871         } else if (CQE_QPID(rsp_msg->cqe) == 0xfff8)
 872                 dev_kfree_skb_irq(skb);
 873         else if (cxio_ev_cb)
 874                 (*cxio_ev_cb) (rdev_p, skb);
 875         else
 876                 dev_kfree_skb_irq(skb);
 877         cnt++;
 878         return 0;
 879 }
 880
 881 /* Caller takes care of locking if needed */
 882 int cxio_rdev_open(struct cxio_rdev *rdev_p)
 883 {
 884         struct net_device *netdev_p = NULL;
 885         int err = 0;
 886         if (strlen(rdev_p->dev_name)) {
 887                 if (cxio_hal_find_rdev_by_name(rdev_p->dev_name)) {
 888                         return -EBUSY;
 889                 }
 890                 netdev_p = dev_get_by_name(rdev_p->dev_name);
 891                 if (!netdev_p) {
 892                         return -EINVAL;
 893                 }
 894                 dev_put(netdev_p);
 895         } else if (rdev_p->t3cdev_p) {
 896                 if (cxio_hal_find_rdev_by_t3cdev(rdev_p->t3cdev_p)) {
 897                         return -EBUSY;
 898                 }
 899                 netdev_p = rdev_p->t3cdev_p->lldev;
 900                 strncpy(rdev_p->dev_name, rdev_p->t3cdev_p->name,
 901                         T3_MAX_DEV_NAME_LEN);
 902         } else {
 903                 PDBG("%s t3cdev_p or dev_name must be set\n", __FUNCTION__);
 904                 return -EINVAL;
 905         }
 906
 907         list_add_tail(&rdev_p->entry, &rdev_list);
 908
 909         PDBG("%s opening rnic dev %s\n", __FUNCTION__, rdev_p->dev_name);
 910         memset(&rdev_p->ctrl_qp, 0, sizeof(rdev_p->ctrl_qp));
 911         if (!rdev_p->t3cdev_p)
 912                 rdev_p->t3cdev_p = T3CDEV(netdev_p);
 913         rdev_p->t3cdev_p->ulp = (void *) rdev_p;
 914         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_GET_PARAMS,
 915                                          &(rdev_p->rnic_info));
 916         if (err) {
 917                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
 918                      __FUNCTION__, rdev_p->t3cdev_p, err);
 919                 goto err1;
 920         }
 921         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_PORTS,
 922                                     &(rdev_p->port_info));
 923         if (err) {
 924                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
 925                      __FUNCTION__, rdev_p->t3cdev_p, err);
 926                 goto err1;
 927         }
 928
 929         /*
 930          * qpshift is the number of bits to shift the qpid left in order
 931          * to get the correct address of the doorbell for that qp.
 932          */
 933         cxio_init_ucontext(rdev_p, &rdev_p->uctx);
 934         rdev_p->qpshift = PAGE_SHIFT -
 935                           ilog2(65536 >>
 936                                     ilog2(rdev_p->rnic_info.udbell_len >>
 937                                               PAGE_SHIFT));
 938         rdev_p->qpnr = rdev_p->rnic_info.udbell_len >> PAGE_SHIFT;
 939         rdev_p->qpmask = (65536 >> ilog2(rdev_p->qpnr)) - 1;
 940         PDBG("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d "
 941              "pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
 942              __FUNCTION__, rdev_p->dev_name, rdev_p->rnic_info.tpt_base,
 943              rdev_p->rnic_info.tpt_top, cxio_num_stags(rdev_p),
 944              rdev_p->rnic_info.pbl_base,
 945              rdev_p->rnic_info.pbl_top, rdev_p->rnic_info.rqt_base,
 946              rdev_p->rnic_info.rqt_top);
 947         PDBG("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu "
 948              "qpnr %d qpmask 0x%x\n",
 949              rdev_p->rnic_info.udbell_len,
 950              rdev_p->rnic_info.udbell_physbase, rdev_p->rnic_info.kdb_addr,
 951              rdev_p->qpshift, rdev_p->qpnr, rdev_p->qpmask);
 952
 953         err = cxio_hal_init_ctrl_qp(rdev_p);
 954         if (err) {
 955                 printk(KERN_ERR "%s error %d initializing ctrl_qp.\n",
 956                        __FUNCTION__, err);
 957                 goto err1;
 958         }
 959         err = cxio_hal_init_resource(rdev_p, cxio_num_stags(rdev_p), 0,
 960                                      0, T3_MAX_NUM_QP, T3_MAX_NUM_CQ,
 961                                      T3_MAX_NUM_PD);
 962         if (err) {
 963                 printk(KERN_ERR "%s error %d initializing hal resources.\n",
 964                        __FUNCTION__, err);
 965                 goto err2;
 966         }
 967         err = cxio_hal_pblpool_create(rdev_p);
 968         if (err) {
 969                 printk(KERN_ERR "%s error %d initializing pbl mem pool.\n",
 970                        __FUNCTION__, err);
 971                 goto err3;
 972         }
 973         err = cxio_hal_rqtpool_create(rdev_p);
 974         if (err) {
 975                 printk(KERN_ERR "%s error %d initializing rqt mem pool.\n",
 976                        __FUNCTION__, err);
 977                 goto err4;
 978         }
 979         return 0;
 980 err4:
 981         cxio_hal_pblpool_destroy(rdev_p);
 982 err3:
 983         cxio_hal_destroy_resource(rdev_p->rscp);
 984 err2:
 985         cxio_hal_destroy_ctrl_qp(rdev_p);
 986 err1:
 987         list_del(&rdev_p->entry);
 988         return err;
 989 }
 990
 991 void cxio_rdev_close(struct cxio_rdev *rdev_p)
 992 {
 993         if (rdev_p) {
 994                 cxio_hal_pblpool_destroy(rdev_p);
 995                 cxio_hal_rqtpool_destroy(rdev_p);
 996                 list_del(&rdev_p->entry);
 997                 rdev_p->t3cdev_p->ulp = NULL;
 998                 cxio_hal_destroy_ctrl_qp(rdev_p);
 999                 cxio_hal_destroy_resource(rdev_p->rscp);
1000         }
1001 }
1002
1003 int __init cxio_hal_init(void)
1004 {
1005         if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI))
1006                 return -ENOMEM;
1007         t3_register_cpl_handler(CPL_ASYNC_NOTIF, cxio_hal_ev_handler);
1008         return 0;
1009 }
1010
1011 void __exit cxio_hal_exit(void)
1012 {
1013         struct cxio_rdev *rdev, *tmp;
1014
1015         t3_register_cpl_handler(CPL_ASYNC_NOTIF, NULL);
1016         list_for_each_entry_safe(rdev, tmp, &rdev_list, entry)
1017                 cxio_rdev_close(rdev);
1018         cxio_hal_destroy_rhdl_resource();
1019 }
1020
1021 static void flush_completed_wrs(struct t3_wq *wq, struct t3_cq *cq)
1022 {
1023         struct t3_swsq *sqp;
1024         __u32 ptr = wq->sq_rptr;
1025         int count = Q_COUNT(wq->sq_rptr, wq->sq_wptr);
1026
1027         sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1028         while (count--)
1029                 if (!sqp->signaled) {
1030                         ptr++;
1031                         sqp = wq->sq + Q_PTR2IDX(ptr,  wq->sq_size_log2);
1032                 } else if (sqp->complete) {
1033
1034                         /*
1035                          * Insert this completed cqe into the swcq.
1036                          */
1037                         PDBG("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1038                              __FUNCTION__, Q_PTR2IDX(ptr,  wq->sq_size_log2),
1039                              Q_PTR2IDX(cq->sw_wptr, cq->size_log2));
1040                         sqp->cqe.header |= htonl(V_CQE_SWCQE(1));
1041                         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2))
1042                                 = sqp->cqe;
1043                         cq->sw_wptr++;
1044                         sqp->signaled = 0;
1045                         break;
1046                 } else
1047                         break;
1048 }
1049
1050 static void create_read_req_cqe(struct t3_wq *wq, struct t3_cqe *hw_cqe,
1051                                 struct t3_cqe *read_cqe)
1052 {
1053         read_cqe->u.scqe.wrid_hi = wq->oldest_read->sq_wptr;
1054         read_cqe->len = wq->oldest_read->read_len;
1055         read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(*hw_cqe)) |
1056                                  V_CQE_SWCQE(SW_CQE(*hw_cqe)) |
1057                                  V_CQE_OPCODE(T3_READ_REQ) |
1058                                  V_CQE_TYPE(1));
1059 }
1060
1061 /*
1062  * Return a ptr to the next read wr in the SWSQ or NULL.
1063  */
1064 static void advance_oldest_read(struct t3_wq *wq)
1065 {
1066
1067         u32 rptr = wq->oldest_read - wq->sq + 1;
1068         u32 wptr = Q_PTR2IDX(wq->sq_wptr, wq->sq_size_log2);
1069
1070         while (Q_PTR2IDX(rptr, wq->sq_size_log2) != wptr) {
1071                 wq->oldest_read = wq->sq + Q_PTR2IDX(rptr, wq->sq_size_log2);
1072
1073                 if (wq->oldest_read->opcode == T3_READ_REQ)
1074                         return;
1075                 rptr++;
1076         }
1077         wq->oldest_read = NULL;
1078 }
1079
1080 /*
1081  * cxio_poll_cq
1082  *
1083  * Caller must:
1084  *     check the validity of the first CQE,
1085  *     supply the wq assicated with the qpid.
1086  *
1087  * credit: cq credit to return to sge.
1088  * cqe_flushed: 1 iff the CQE is flushed.
1089  * cqe: copy of the polled CQE.
1090  *
1091  * return value:
1092  *     0       CQE returned,
1093  *    -1       CQE skipped, try again.
1094  */
1095 int cxio_poll_cq(struct t3_wq *wq, struct t3_cq *cq, struct t3_cqe *cqe,
1096                      u8 *cqe_flushed, u64 *cookie, u32 *credit)
1097 {
1098         int ret = 0;
1099         struct t3_cqe *hw_cqe, read_cqe;
1100
1101         *cqe_flushed = 0;
1102         *credit = 0;
1103         hw_cqe = cxio_next_cqe(cq);
1104
1105         PDBG("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x"
1106              " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1107              __FUNCTION__, CQE_OOO(*hw_cqe), CQE_QPID(*hw_cqe),
1108              CQE_GENBIT(*hw_cqe), CQE_TYPE(*hw_cqe), CQE_STATUS(*hw_cqe),
1109              CQE_OPCODE(*hw_cqe), CQE_LEN(*hw_cqe), CQE_WRID_HI(*hw_cqe),
1110              CQE_WRID_LOW(*hw_cqe));
1111
1112         /*
1113          * skip cqe's not affiliated with a QP.
1114          */
1115         if (wq == NULL) {
1116                 ret = -1;
1117                 goto skip_cqe;
1118         }
1119
1120         /*
1121          * Gotta tweak READ completions:
1122          *      1) the cqe doesn't contain the sq_wptr from the wr.
1123          *      2) opcode not reflected from the wr.
1124          *      3) read_len not reflected from the wr.
1125          *      4) cq_type is RQ_TYPE not SQ_TYPE.
1126          */
1127         if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
1128
1129                 /*
1130                  * Don't write to the HWCQ, so create a new read req CQE
1131                  * in local memory.
1132                  */
1133                 create_read_req_cqe(wq, hw_cqe, &read_cqe);
1134                 hw_cqe = &read_cqe;
1135                 advance_oldest_read(wq);
1136         }
1137
1138         /*
1139          * T3A: Discard TERMINATE CQEs.
1140          */
1141         if (CQE_OPCODE(*hw_cqe) == T3_TERMINATE) {
1142                 ret = -1;
1143                 wq->error = 1;
1144                 goto skip_cqe;
1145         }
1146
1147         if (CQE_STATUS(*hw_cqe) || wq->error) {
1148                 *cqe_flushed = wq->error;
1149                 wq->error = 1;
1150
1151                 /*
1152                  * T3A inserts errors into the CQE.  We cannot return
1153                  * these as work completions.
1154                  */
1155                 /* incoming write failures */
1156                 if ((CQE_OPCODE(*hw_cqe) == T3_RDMA_WRITE)
1157                      && RQ_TYPE(*hw_cqe)) {
1158                         ret = -1;
1159                         goto skip_cqe;
1160                 }
1161                 /* incoming read request failures */
1162                 if ((CQE_OPCODE(*hw_cqe) == T3_READ_RESP) && SQ_TYPE(*hw_cqe)) {
1163                         ret = -1;
1164                         goto skip_cqe;
1165                 }
1166
1167                 /* incoming SEND with no receive posted failures */
1168                 if ((CQE_OPCODE(*hw_cqe) == T3_SEND) && RQ_TYPE(*hw_cqe) &&
1169                     Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1170                         ret = -1;
1171                         goto skip_cqe;
1172                 }
1173                 goto proc_cqe;
1174         }
1175
1176         /*
1177          * RECV completion.
1178          */
1179         if (RQ_TYPE(*hw_cqe)) {
1180
1181                 /*
1182                  * HW only validates 4 bits of MSN.  So we must validate that
1183                  * the MSN in the SEND is the next expected MSN.  If its not,
1184                  * then we complete this with TPT_ERR_MSN and mark the wq in
1185                  * error.
1186                  */
1187                 if (unlikely((CQE_WRID_MSN(*hw_cqe) != (wq->rq_rptr + 1)))) {
1188                         wq->error = 1;
1189                         hw_cqe->header |= htonl(V_CQE_STATUS(TPT_ERR_MSN));
1190                         goto proc_cqe;
1191                 }
1192                 goto proc_cqe;
1193         }
1194
1195         /*
1196          * If we get here its a send completion.
1197          *
1198          * Handle out of order completion. These get stuffed
1199          * in the SW SQ. Then the SW SQ is walked to move any
1200          * now in-order completions into the SW CQ.  This handles
1201          * 2 cases:
1202          *      1) reaping unsignaled WRs when the first subsequent
1203          *         signaled WR is completed.
1204          *      2) out of order read completions.
1205          */
1206         if (!SW_CQE(*hw_cqe) && (CQE_WRID_SQ_WPTR(*hw_cqe) != wq->sq_rptr)) {
1207                 struct t3_swsq *sqp;
1208
1209                 PDBG("%s out of order completion going in swsq at idx %ld\n",
1210                      __FUNCTION__,
1211                      Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2));
1212                 sqp = wq->sq +
1213                       Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2);
1214                 sqp->cqe = *hw_cqe;
1215                 sqp->complete = 1;
1216                 ret = -1;
1217                 goto flush_wq;
1218         }
1219
1220 proc_cqe:
1221         *cqe = *hw_cqe;
1222
1223         /*
1224          * Reap the associated WR(s) that are freed up with this
1225          * completion.
1226          */
1227         if (SQ_TYPE(*hw_cqe)) {
1228                 wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
1229                 PDBG("%s completing sq idx %ld\n", __FUNCTION__,
1230                      Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
1231                 *cookie = (wq->sq +
1232                            Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2))->wr_id;
1233                 wq->sq_rptr++;
1234         } else {
1235                 PDBG("%s completing rq idx %ld\n", __FUNCTION__,
1236                      Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1237                 *cookie = *(wq->rq + Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1238                 wq->rq_rptr++;
1239         }
1240
1241 flush_wq:
1242         /*
1243          * Flush any completed cqes that are now in-order.
1244          */
1245         flush_completed_wrs(wq, cq);
1246
1247 skip_cqe:
1248         if (SW_CQE(*hw_cqe)) {
1249                 PDBG("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1250                      __FUNCTION__, cq, cq->cqid, cq->sw_rptr);
1251                 ++cq->sw_rptr;
1252         } else {
1253                 PDBG("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1254                      __FUNCTION__, cq, cq->cqid, cq->rptr);
1255                 ++cq->rptr;
1256
1257                 /*
1258                  * T3A: compute credits.
1259                  */
1260                 if (((cq->rptr - cq->wptr) > (1 << (cq->size_log2 - 1)))
1261                     || ((cq->rptr - cq->wptr) >= 128)) {
1262                         *credit = cq->rptr - cq->wptr;
1263                         cq->wptr = cq->rptr;
1264                 }
1265         }
1266         return ret;
1267 }