git.oblomov.eu Git - linux-2.6/blob - 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 #include <net/net_namespace.h>
  41
  42 #include "cxio_resource.h"
  43 #include "cxio_hal.h"
  44 #include "cxgb3_offload.h"
  45 #include "sge_defs.h"
  46
  47 static LIST_HEAD(rdev_list);
  48 static cxio_hal_ev_callback_func_t cxio_ev_cb = NULL;
  49
  50 static struct cxio_rdev *cxio_hal_find_rdev_by_name(char *dev_name)
  51 {
  52         struct cxio_rdev *rdev;
  53
  54         list_for_each_entry(rdev, &rdev_list, entry)
  55                 if (!strcmp(rdev->dev_name, dev_name))
  56                         return rdev;
  57         return NULL;
  58 }
  59
  60 static struct cxio_rdev *cxio_hal_find_rdev_by_t3cdev(struct t3cdev *tdev)
  61 {
  62         struct cxio_rdev *rdev;
  63
  64         list_for_each_entry(rdev, &rdev_list, entry)
  65                 if (rdev->t3cdev_p == tdev)
  66                         return rdev;
  67         return NULL;
  68 }
  69
  70 int cxio_hal_cq_op(struct cxio_rdev *rdev_p, struct t3_cq *cq,
  71                    enum t3_cq_opcode op, u32 credit)
  72 {
  73         int ret;
  74         struct t3_cqe *cqe;
  75         u32 rptr;
  76
  77         struct rdma_cq_op setup;
  78         setup.id = cq->cqid;
  79         setup.credits = (op == CQ_CREDIT_UPDATE) ? credit : 0;
  80         setup.op = op;
  81         ret = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_OP, &setup);
  82
  83         if ((ret < 0) || (op == CQ_CREDIT_UPDATE))
  84                 return ret;
  85
  86         /*
  87          * If the rearm returned an index other than our current index,
  88          * then there might be CQE's in flight (being DMA'd).  We must wait
  89          * here for them to complete or the consumer can miss a notification.
  90          */
  91         if (Q_PTR2IDX((cq->rptr), cq->size_log2) != ret) {
  92                 int i=0;
  93
  94                 rptr = cq->rptr;
  95
  96                 /*
  97                  * Keep the generation correct by bumping rptr until it
  98                  * matches the index returned by the rearm - 1.
  99                  */
 100                 while (Q_PTR2IDX((rptr+1), cq->size_log2) != ret)
 101                         rptr++;
 102
 103                 /*
 104                  * Now rptr is the index for the (last) cqe that was
 105                  * in-flight at the time the HW rearmed the CQ.  We
 106                  * spin until that CQE is valid.
 107                  */
 108                 cqe = cq->queue + Q_PTR2IDX(rptr, cq->size_log2);
 109                 while (!CQ_VLD_ENTRY(rptr, cq->size_log2, cqe)) {
 110                         udelay(1);
 111                         if (i++ > 1000000) {
 112                                 BUG_ON(1);
 113                                 printk(KERN_ERR "%s: stalled rnic\n",
 114                                        rdev_p->dev_name);
 115                                 return -EIO;
 116                         }
 117                 }
 118
 119                 return 1;
 120         }
 121
 122         return 0;
 123 }
 124
 125 static int cxio_hal_clear_cq_ctx(struct cxio_rdev *rdev_p, u32 cqid)
 126 {
 127         struct rdma_cq_setup setup;
 128         setup.id = cqid;
 129         setup.base_addr = 0;    /* NULL address */
 130         setup.size = 0;         /* disaable the CQ */
 131         setup.credits = 0;
 132         setup.credit_thres = 0;
 133         setup.ovfl_mode = 0;
 134         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 135 }
 136
 137 static int cxio_hal_clear_qp_ctx(struct cxio_rdev *rdev_p, u32 qpid)
 138 {
 139         u64 sge_cmd;
 140         struct t3_modify_qp_wr *wqe;
 141         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
 142         if (!skb) {
 143                 PDBG("%s alloc_skb failed\n", __func__);
 144                 return -ENOMEM;
 145         }
 146         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
 147         memset(wqe, 0, sizeof(*wqe));
 148         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD,
 149                        T3_COMPLETION_FLAG | T3_NOTIFY_FLAG, 0, qpid, 7,
 150                        T3_SOPEOP);
 151         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
 152         sge_cmd = qpid << 8 | 3;
 153         wqe->sge_cmd = cpu_to_be64(sge_cmd);
 154         skb->priority = CPL_PRIORITY_CONTROL;
 155         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 156 }
 157
 158 int cxio_create_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 159 {
 160         struct rdma_cq_setup setup;
 161         int size = (1UL << (cq->size_log2)) * sizeof(struct t3_cqe);
 162
 163         cq->cqid = cxio_hal_get_cqid(rdev_p->rscp);
 164         if (!cq->cqid)
 165                 return -ENOMEM;
 166         cq->sw_queue = kzalloc(size, GFP_KERNEL);
 167         if (!cq->sw_queue)
 168                 return -ENOMEM;
 169         cq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
 170                                              (1UL << (cq->size_log2)) *
 171                                              sizeof(struct t3_cqe),
 172                                              &(cq->dma_addr), GFP_KERNEL);
 173         if (!cq->queue) {
 174                 kfree(cq->sw_queue);
 175                 return -ENOMEM;
 176         }
 177         pci_unmap_addr_set(cq, mapping, cq->dma_addr);
 178         memset(cq->queue, 0, size);
 179         setup.id = cq->cqid;
 180         setup.base_addr = (u64) (cq->dma_addr);
 181         setup.size = 1UL << cq->size_log2;
 182         setup.credits = 65535;
 183         setup.credit_thres = 1;
 184         if (rdev_p->t3cdev_p->type != T3A)
 185                 setup.ovfl_mode = 0;
 186         else
 187                 setup.ovfl_mode = 1;
 188         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 189 }
 190
 191 int cxio_resize_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 192 {
 193         struct rdma_cq_setup setup;
 194         setup.id = cq->cqid;
 195         setup.base_addr = (u64) (cq->dma_addr);
 196         setup.size = 1UL << cq->size_log2;
 197         setup.credits = setup.size;
 198         setup.credit_thres = setup.size;        /* TBD: overflow recovery */
 199         setup.ovfl_mode = 1;
 200         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 201 }
 202
 203 static u32 get_qpid(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 204 {
 205         struct cxio_qpid_list *entry;
 206         u32 qpid;
 207         int i;
 208
 209         mutex_lock(&uctx->lock);
 210         if (!list_empty(&uctx->qpids)) {
 211                 entry = list_entry(uctx->qpids.next, struct cxio_qpid_list,
 212                                    entry);
 213                 list_del(&entry->entry);
 214                 qpid = entry->qpid;
 215                 kfree(entry);
 216         } else {
 217                 qpid = cxio_hal_get_qpid(rdev_p->rscp);
 218                 if (!qpid)
 219                         goto out;
 220                 for (i = qpid+1; i & rdev_p->qpmask; i++) {
 221                         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 222                         if (!entry)
 223                                 break;
 224                         entry->qpid = i;
 225                         list_add_tail(&entry->entry, &uctx->qpids);
 226                 }
 227         }
 228 out:
 229         mutex_unlock(&uctx->lock);
 230         PDBG("%s qpid 0x%x\n", __func__, qpid);
 231         return qpid;
 232 }
 233
 234 static void put_qpid(struct cxio_rdev *rdev_p, u32 qpid,
 235                      struct cxio_ucontext *uctx)
 236 {
 237         struct cxio_qpid_list *entry;
 238
 239         entry = kmalloc(sizeof *entry, GFP_KERNEL);
 240         if (!entry)
 241                 return;
 242         PDBG("%s qpid 0x%x\n", __func__, qpid);
 243         entry->qpid = qpid;
 244         mutex_lock(&uctx->lock);
 245         list_add_tail(&entry->entry, &uctx->qpids);
 246         mutex_unlock(&uctx->lock);
 247 }
 248
 249 void cxio_release_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 250 {
 251         struct list_head *pos, *nxt;
 252         struct cxio_qpid_list *entry;
 253
 254         mutex_lock(&uctx->lock);
 255         list_for_each_safe(pos, nxt, &uctx->qpids) {
 256                 entry = list_entry(pos, struct cxio_qpid_list, entry);
 257                 list_del_init(&entry->entry);
 258                 if (!(entry->qpid & rdev_p->qpmask))
 259                         cxio_hal_put_qpid(rdev_p->rscp, entry->qpid);
 260                 kfree(entry);
 261         }
 262         mutex_unlock(&uctx->lock);
 263 }
 264
 265 void cxio_init_ucontext(struct cxio_rdev *rdev_p, struct cxio_ucontext *uctx)
 266 {
 267         INIT_LIST_HEAD(&uctx->qpids);
 268         mutex_init(&uctx->lock);
 269 }
 270
 271 int cxio_create_qp(struct cxio_rdev *rdev_p, u32 kernel_domain,
 272                    struct t3_wq *wq, struct cxio_ucontext *uctx)
 273 {
 274         int depth = 1UL << wq->size_log2;
 275         int rqsize = 1UL << wq->rq_size_log2;
 276
 277         wq->qpid = get_qpid(rdev_p, uctx);
 278         if (!wq->qpid)
 279                 return -ENOMEM;
 280
 281         wq->rq = kzalloc(depth * sizeof(struct t3_swrq), GFP_KERNEL);
 282         if (!wq->rq)
 283                 goto err1;
 284
 285         wq->rq_addr = cxio_hal_rqtpool_alloc(rdev_p, rqsize);
 286         if (!wq->rq_addr)
 287                 goto err2;
 288
 289         wq->sq = kzalloc(depth * sizeof(struct t3_swsq), GFP_KERNEL);
 290         if (!wq->sq)
 291                 goto err3;
 292
 293         wq->queue = dma_alloc_coherent(&(rdev_p->rnic_info.pdev->dev),
 294                                              depth * sizeof(union t3_wr),
 295                                              &(wq->dma_addr), GFP_KERNEL);
 296         if (!wq->queue)
 297                 goto err4;
 298
 299         memset(wq->queue, 0, depth * sizeof(union t3_wr));
 300         pci_unmap_addr_set(wq, mapping, wq->dma_addr);
 301         wq->doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
 302         if (!kernel_domain)
 303                 wq->udb = (u64)rdev_p->rnic_info.udbell_physbase +
 304                                         (wq->qpid << rdev_p->qpshift);
 305         wq->rdev = rdev_p;
 306         PDBG("%s qpid 0x%x doorbell 0x%p udb 0x%llx\n", __func__,
 307              wq->qpid, wq->doorbell, (unsigned long long) wq->udb);
 308         return 0;
 309 err4:
 310         kfree(wq->sq);
 311 err3:
 312         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, rqsize);
 313 err2:
 314         kfree(wq->rq);
 315 err1:
 316         put_qpid(rdev_p, wq->qpid, uctx);
 317         return -ENOMEM;
 318 }
 319
 320 int cxio_destroy_cq(struct cxio_rdev *rdev_p, struct t3_cq *cq)
 321 {
 322         int err;
 323         err = cxio_hal_clear_cq_ctx(rdev_p, cq->cqid);
 324         kfree(cq->sw_queue);
 325         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 326                           (1UL << (cq->size_log2))
 327                           * sizeof(struct t3_cqe), cq->queue,
 328                           pci_unmap_addr(cq, mapping));
 329         cxio_hal_put_cqid(rdev_p->rscp, cq->cqid);
 330         return err;
 331 }
 332
 333 int cxio_destroy_qp(struct cxio_rdev *rdev_p, struct t3_wq *wq,
 334                     struct cxio_ucontext *uctx)
 335 {
 336         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 337                           (1UL << (wq->size_log2))
 338                           * sizeof(union t3_wr), wq->queue,
 339                           pci_unmap_addr(wq, mapping));
 340         kfree(wq->sq);
 341         cxio_hal_rqtpool_free(rdev_p, wq->rq_addr, (1UL << wq->rq_size_log2));
 342         kfree(wq->rq);
 343         put_qpid(rdev_p, wq->qpid, uctx);
 344         return 0;
 345 }
 346
 347 static void insert_recv_cqe(struct t3_wq *wq, struct t3_cq *cq)
 348 {
 349         struct t3_cqe cqe;
 350
 351         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
 352              wq, cq, cq->sw_rptr, cq->sw_wptr);
 353         memset(&cqe, 0, sizeof(cqe));
 354         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
 355                                  V_CQE_OPCODE(T3_SEND) |
 356                                  V_CQE_TYPE(0) |
 357                                  V_CQE_SWCQE(1) |
 358                                  V_CQE_QPID(wq->qpid) |
 359                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
 360                                                        cq->size_log2)));
 361         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
 362         cq->sw_wptr++;
 363 }
 364
 365 int cxio_flush_rq(struct t3_wq *wq, struct t3_cq *cq, int count)
 366 {
 367         u32 ptr;
 368         int flushed = 0;
 369
 370         PDBG("%s wq %p cq %p\n", __func__, wq, cq);
 371
 372         /* flush RQ */
 373         PDBG("%s rq_rptr %u rq_wptr %u skip count %u\n", __func__,
 374             wq->rq_rptr, wq->rq_wptr, count);
 375         ptr = wq->rq_rptr + count;
 376         while (ptr++ != wq->rq_wptr) {
 377                 insert_recv_cqe(wq, cq);
 378                 flushed++;
 379         }
 380         return flushed;
 381 }
 382
 383 static void insert_sq_cqe(struct t3_wq *wq, struct t3_cq *cq,
 384                           struct t3_swsq *sqp)
 385 {
 386         struct t3_cqe cqe;
 387
 388         PDBG("%s wq %p cq %p sw_rptr 0x%x sw_wptr 0x%x\n", __func__,
 389              wq, cq, cq->sw_rptr, cq->sw_wptr);
 390         memset(&cqe, 0, sizeof(cqe));
 391         cqe.header = cpu_to_be32(V_CQE_STATUS(TPT_ERR_SWFLUSH) |
 392                                  V_CQE_OPCODE(sqp->opcode) |
 393                                  V_CQE_TYPE(1) |
 394                                  V_CQE_SWCQE(1) |
 395                                  V_CQE_QPID(wq->qpid) |
 396                                  V_CQE_GENBIT(Q_GENBIT(cq->sw_wptr,
 397                                                        cq->size_log2)));
 398         cqe.u.scqe.wrid_hi = sqp->sq_wptr;
 399
 400         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2)) = cqe;
 401         cq->sw_wptr++;
 402 }
 403
 404 int cxio_flush_sq(struct t3_wq *wq, struct t3_cq *cq, int count)
 405 {
 406         __u32 ptr;
 407         int flushed = 0;
 408         struct t3_swsq *sqp = wq->sq + Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2);
 409
 410         ptr = wq->sq_rptr + count;
 411         sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
 412         while (ptr != wq->sq_wptr) {
 413                 insert_sq_cqe(wq, cq, sqp);
 414                 ptr++;
 415                 sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
 416                 flushed++;
 417         }
 418         return flushed;
 419 }
 420
 421 /*
 422  * Move all CQEs from the HWCQ into the SWCQ.
 423  */
 424 void cxio_flush_hw_cq(struct t3_cq *cq)
 425 {
 426         struct t3_cqe *cqe, *swcqe;
 427
 428         PDBG("%s cq %p cqid 0x%x\n", __func__, cq, cq->cqid);
 429         cqe = cxio_next_hw_cqe(cq);
 430         while (cqe) {
 431                 PDBG("%s flushing hwcq rptr 0x%x to swcq wptr 0x%x\n",
 432                      __func__, cq->rptr, cq->sw_wptr);
 433                 swcqe = cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2);
 434                 *swcqe = *cqe;
 435                 swcqe->header |= cpu_to_be32(V_CQE_SWCQE(1));
 436                 cq->sw_wptr++;
 437                 cq->rptr++;
 438                 cqe = cxio_next_hw_cqe(cq);
 439         }
 440 }
 441
 442 static int cqe_completes_wr(struct t3_cqe *cqe, struct t3_wq *wq)
 443 {
 444         if (CQE_OPCODE(*cqe) == T3_TERMINATE)
 445                 return 0;
 446
 447         if ((CQE_OPCODE(*cqe) == T3_RDMA_WRITE) && RQ_TYPE(*cqe))
 448                 return 0;
 449
 450         if ((CQE_OPCODE(*cqe) == T3_READ_RESP) && SQ_TYPE(*cqe))
 451                 return 0;
 452
 453         if ((CQE_OPCODE(*cqe) == T3_SEND) && RQ_TYPE(*cqe) &&
 454             Q_EMPTY(wq->rq_rptr, wq->rq_wptr))
 455                 return 0;
 456
 457         return 1;
 458 }
 459
 460 void cxio_count_scqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
 461 {
 462         struct t3_cqe *cqe;
 463         u32 ptr;
 464
 465         *count = 0;
 466         ptr = cq->sw_rptr;
 467         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
 468                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
 469                 if ((SQ_TYPE(*cqe) ||
 470                      ((CQE_OPCODE(*cqe) == T3_READ_RESP) && wq->oldest_read)) &&
 471                     (CQE_QPID(*cqe) == wq->qpid))
 472                         (*count)++;
 473                 ptr++;
 474         }
 475         PDBG("%s cq %p count %d\n", __func__, cq, *count);
 476 }
 477
 478 void cxio_count_rcqes(struct t3_cq *cq, struct t3_wq *wq, int *count)
 479 {
 480         struct t3_cqe *cqe;
 481         u32 ptr;
 482
 483         *count = 0;
 484         PDBG("%s count zero %d\n", __func__, *count);
 485         ptr = cq->sw_rptr;
 486         while (!Q_EMPTY(ptr, cq->sw_wptr)) {
 487                 cqe = cq->sw_queue + (Q_PTR2IDX(ptr, cq->size_log2));
 488                 if (RQ_TYPE(*cqe) && (CQE_OPCODE(*cqe) != T3_READ_RESP) &&
 489                     (CQE_QPID(*cqe) == wq->qpid) && cqe_completes_wr(cqe, wq))
 490                         (*count)++;
 491                 ptr++;
 492         }
 493         PDBG("%s cq %p count %d\n", __func__, cq, *count);
 494 }
 495
 496 static int cxio_hal_init_ctrl_cq(struct cxio_rdev *rdev_p)
 497 {
 498         struct rdma_cq_setup setup;
 499         setup.id = 0;
 500         setup.base_addr = 0;    /* NULL address */
 501         setup.size = 1;         /* enable the CQ */
 502         setup.credits = 0;
 503
 504         /* force SGE to redirect to RspQ and interrupt */
 505         setup.credit_thres = 0;
 506         setup.ovfl_mode = 1;
 507         return (rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_CQ_SETUP, &setup));
 508 }
 509
 510 static int cxio_hal_init_ctrl_qp(struct cxio_rdev *rdev_p)
 511 {
 512         int err;
 513         u64 sge_cmd, ctx0, ctx1;
 514         u64 base_addr;
 515         struct t3_modify_qp_wr *wqe;
 516         struct sk_buff *skb;
 517
 518         skb = alloc_skb(sizeof(*wqe), GFP_KERNEL);
 519         if (!skb) {
 520                 PDBG("%s alloc_skb failed\n", __func__);
 521                 return -ENOMEM;
 522         }
 523         err = cxio_hal_init_ctrl_cq(rdev_p);
 524         if (err) {
 525                 PDBG("%s err %d initializing ctrl_cq\n", __func__, err);
 526                 goto err;
 527         }
 528         rdev_p->ctrl_qp.workq = dma_alloc_coherent(
 529                                         &(rdev_p->rnic_info.pdev->dev),
 530                                         (1 << T3_CTRL_QP_SIZE_LOG2) *
 531                                         sizeof(union t3_wr),
 532                                         &(rdev_p->ctrl_qp.dma_addr),
 533                                         GFP_KERNEL);
 534         if (!rdev_p->ctrl_qp.workq) {
 535                 PDBG("%s dma_alloc_coherent failed\n", __func__);
 536                 err = -ENOMEM;
 537                 goto err;
 538         }
 539         pci_unmap_addr_set(&rdev_p->ctrl_qp, mapping,
 540                            rdev_p->ctrl_qp.dma_addr);
 541         rdev_p->ctrl_qp.doorbell = (void __iomem *)rdev_p->rnic_info.kdb_addr;
 542         memset(rdev_p->ctrl_qp.workq, 0,
 543                (1 << T3_CTRL_QP_SIZE_LOG2) * sizeof(union t3_wr));
 544
 545         mutex_init(&rdev_p->ctrl_qp.lock);
 546         init_waitqueue_head(&rdev_p->ctrl_qp.waitq);
 547
 548         /* update HW Ctrl QP context */
 549         base_addr = rdev_p->ctrl_qp.dma_addr;
 550         base_addr >>= 12;
 551         ctx0 = (V_EC_SIZE((1 << T3_CTRL_QP_SIZE_LOG2)) |
 552                 V_EC_BASE_LO((u32) base_addr & 0xffff));
 553         ctx0 <<= 32;
 554         ctx0 |= V_EC_CREDITS(FW_WR_NUM);
 555         base_addr >>= 16;
 556         ctx1 = (u32) base_addr;
 557         base_addr >>= 32;
 558         ctx1 |= ((u64) (V_EC_BASE_HI((u32) base_addr & 0xf) | V_EC_RESPQ(0) |
 559                         V_EC_TYPE(0) | V_EC_GEN(1) |
 560                         V_EC_UP_TOKEN(T3_CTL_QP_TID) | F_EC_VALID)) << 32;
 561         wqe = (struct t3_modify_qp_wr *) skb_put(skb, sizeof(*wqe));
 562         memset(wqe, 0, sizeof(*wqe));
 563         build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_QP_MOD, 0, 0,
 564                        T3_CTL_QP_TID, 7, T3_SOPEOP);
 565         wqe->flags = cpu_to_be32(MODQP_WRITE_EC);
 566         sge_cmd = (3ULL << 56) | FW_RI_SGEEC_START << 8 | 3;
 567         wqe->sge_cmd = cpu_to_be64(sge_cmd);
 568         wqe->ctx1 = cpu_to_be64(ctx1);
 569         wqe->ctx0 = cpu_to_be64(ctx0);
 570         PDBG("CtrlQP dma_addr 0x%llx workq %p size %d\n",
 571              (unsigned long long) rdev_p->ctrl_qp.dma_addr,
 572              rdev_p->ctrl_qp.workq, 1 << T3_CTRL_QP_SIZE_LOG2);
 573         skb->priority = CPL_PRIORITY_CONTROL;
 574         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 575 err:
 576         kfree_skb(skb);
 577         return err;
 578 }
 579
 580 static int cxio_hal_destroy_ctrl_qp(struct cxio_rdev *rdev_p)
 581 {
 582         dma_free_coherent(&(rdev_p->rnic_info.pdev->dev),
 583                           (1UL << T3_CTRL_QP_SIZE_LOG2)
 584                           * sizeof(union t3_wr), rdev_p->ctrl_qp.workq,
 585                           pci_unmap_addr(&rdev_p->ctrl_qp, mapping));
 586         return cxio_hal_clear_qp_ctx(rdev_p, T3_CTRL_QP_ID);
 587 }
 588
 589 /* write len bytes of data into addr (32B aligned address)
 590  * If data is NULL, clear len byte of memory to zero.
 591  * caller aquires the ctrl_qp lock before the call
 592  */
 593 static int cxio_hal_ctrl_qp_write_mem(struct cxio_rdev *rdev_p, u32 addr,
 594                                       u32 len, void *data)
 595 {
 596         u32 i, nr_wqe, copy_len;
 597         u8 *copy_data;
 598         u8 wr_len, utx_len;     /* length in 8 byte flit */
 599         enum t3_wr_flags flag;
 600         __be64 *wqe;
 601         u64 utx_cmd;
 602         addr &= 0x7FFFFFF;
 603         nr_wqe = len % 96 ? len / 96 + 1 : len / 96;    /* 96B max per WQE */
 604         PDBG("%s wptr 0x%x rptr 0x%x len %d, nr_wqe %d data %p addr 0x%0x\n",
 605              __func__, rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, len,
 606              nr_wqe, data, addr);
 607         utx_len = 3;            /* in 32B unit */
 608         for (i = 0; i < nr_wqe; i++) {
 609                 if (Q_FULL(rdev_p->ctrl_qp.rptr, rdev_p->ctrl_qp.wptr,
 610                            T3_CTRL_QP_SIZE_LOG2)) {
 611                         PDBG("%s ctrl_qp full wtpr 0x%0x rptr 0x%0x, "
 612                              "wait for more space i %d\n", __func__,
 613                              rdev_p->ctrl_qp.wptr, rdev_p->ctrl_qp.rptr, i);
 614                         if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
 615                                              !Q_FULL(rdev_p->ctrl_qp.rptr,
 616                                                      rdev_p->ctrl_qp.wptr,
 617                                                      T3_CTRL_QP_SIZE_LOG2))) {
 618                                 PDBG("%s ctrl_qp workq interrupted\n",
 619                                      __func__);
 620                                 return -ERESTARTSYS;
 621                         }
 622                         PDBG("%s ctrl_qp wakeup, continue posting work request "
 623                              "i %d\n", __func__, i);
 624                 }
 625                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
 626                                                 (1 << T3_CTRL_QP_SIZE_LOG2)));
 627                 flag = 0;
 628                 if (i == (nr_wqe - 1)) {
 629                         /* last WQE */
 630                         flag = T3_COMPLETION_FLAG;
 631                         if (len % 32)
 632                                 utx_len = len / 32 + 1;
 633                         else
 634                                 utx_len = len / 32;
 635                 }
 636
 637                 /*
 638                  * Force a CQE to return the credit to the workq in case
 639                  * we posted more than half the max QP size of WRs
 640                  */
 641                 if ((i != 0) &&
 642                     (i % (((1 << T3_CTRL_QP_SIZE_LOG2)) >> 1) == 0)) {
 643                         flag = T3_COMPLETION_FLAG;
 644                         PDBG("%s force completion at i %d\n", __func__, i);
 645                 }
 646
 647                 /* build the utx mem command */
 648                 wqe += (sizeof(struct t3_bypass_wr) >> 3);
 649                 utx_cmd = (T3_UTX_MEM_WRITE << 28) | (addr + i * 3);
 650                 utx_cmd <<= 32;
 651                 utx_cmd |= (utx_len << 28) | ((utx_len << 2) + 1);
 652                 *wqe = cpu_to_be64(utx_cmd);
 653                 wqe++;
 654                 copy_data = (u8 *) data + i * 96;
 655                 copy_len = len > 96 ? 96 : len;
 656
 657                 /* clear memory content if data is NULL */
 658                 if (data)
 659                         memcpy(wqe, copy_data, copy_len);
 660                 else
 661                         memset(wqe, 0, copy_len);
 662                 if (copy_len % 32)
 663                         memset(((u8 *) wqe) + copy_len, 0,
 664                                32 - (copy_len % 32));
 665                 wr_len = ((sizeof(struct t3_bypass_wr)) >> 3) + 1 +
 666                          (utx_len << 2);
 667                 wqe = (__be64 *)(rdev_p->ctrl_qp.workq + (rdev_p->ctrl_qp.wptr %
 668                               (1 << T3_CTRL_QP_SIZE_LOG2)));
 669
 670                 /* wptr in the WRID[31:0] */
 671                 ((union t3_wrid *)(wqe+1))->id0.low = rdev_p->ctrl_qp.wptr;
 672
 673                 /*
 674                  * This must be the last write with a memory barrier
 675                  * for the genbit
 676                  */
 677                 build_fw_riwrh((struct fw_riwrh *) wqe, T3_WR_BP, flag,
 678                                Q_GENBIT(rdev_p->ctrl_qp.wptr,
 679                                         T3_CTRL_QP_SIZE_LOG2), T3_CTRL_QP_ID,
 680                                wr_len, T3_SOPEOP);
 681                 if (flag == T3_COMPLETION_FLAG)
 682                         ring_doorbell(rdev_p->ctrl_qp.doorbell, T3_CTRL_QP_ID);
 683                 len -= 96;
 684                 rdev_p->ctrl_qp.wptr++;
 685         }
 686         return 0;
 687 }
 688
 689 /* IN: stag key, pdid, perm, zbva, to, len, page_size, pbl_size and pbl_addr
 690  * OUT: stag index
 691  * TBD: shared memory region support
 692  */
 693 static int __cxio_tpt_op(struct cxio_rdev *rdev_p, u32 reset_tpt_entry,
 694                          u32 *stag, u8 stag_state, u32 pdid,
 695                          enum tpt_mem_type type, enum tpt_mem_perm perm,
 696                          u32 zbva, u64 to, u32 len, u8 page_size,
 697                          u32 pbl_size, u32 pbl_addr)
 698 {
 699         int err;
 700         struct tpt_entry tpt;
 701         u32 stag_idx;
 702         u32 wptr;
 703
 704         stag_state = stag_state > 0;
 705         stag_idx = (*stag) >> 8;
 706
 707         if ((!reset_tpt_entry) && !(*stag != T3_STAG_UNSET)) {
 708                 stag_idx = cxio_hal_get_stag(rdev_p->rscp);
 709                 if (!stag_idx)
 710                         return -ENOMEM;
 711                 *stag = (stag_idx << 8) | ((*stag) & 0xFF);
 712         }
 713         PDBG("%s stag_state 0x%0x type 0x%0x pdid 0x%0x, stag_idx 0x%x\n",
 714              __func__, stag_state, type, pdid, stag_idx);
 715
 716         mutex_lock(&rdev_p->ctrl_qp.lock);
 717
 718         /* write TPT entry */
 719         if (reset_tpt_entry)
 720                 memset(&tpt, 0, sizeof(tpt));
 721         else {
 722                 tpt.valid_stag_pdid = cpu_to_be32(F_TPT_VALID |
 723                                 V_TPT_STAG_KEY((*stag) & M_TPT_STAG_KEY) |
 724                                 V_TPT_STAG_STATE(stag_state) |
 725                                 V_TPT_STAG_TYPE(type) | V_TPT_PDID(pdid));
 726                 BUG_ON(page_size >= 28);
 727                 tpt.flags_pagesize_qpid = cpu_to_be32(V_TPT_PERM(perm) |
 728                                 F_TPT_MW_BIND_ENABLE |
 729                                 V_TPT_ADDR_TYPE((zbva ? TPT_ZBTO : TPT_VATO)) |
 730                                 V_TPT_PAGE_SIZE(page_size));
 731                 tpt.rsvd_pbl_addr = reset_tpt_entry ? 0 :
 732                                     cpu_to_be32(V_TPT_PBL_ADDR(PBL_OFF(rdev_p, pbl_addr)>>3));
 733                 tpt.len = cpu_to_be32(len);
 734                 tpt.va_hi = cpu_to_be32((u32) (to >> 32));
 735                 tpt.va_low_or_fbo = cpu_to_be32((u32) (to & 0xFFFFFFFFULL));
 736                 tpt.rsvd_bind_cnt_or_pstag = 0;
 737                 tpt.rsvd_pbl_size = reset_tpt_entry ? 0 :
 738                                   cpu_to_be32(V_TPT_PBL_SIZE(pbl_size >> 2));
 739         }
 740         err = cxio_hal_ctrl_qp_write_mem(rdev_p,
 741                                        stag_idx +
 742                                        (rdev_p->rnic_info.tpt_base >> 5),
 743                                        sizeof(tpt), &tpt);
 744
 745         /* release the stag index to free pool */
 746         if (reset_tpt_entry)
 747                 cxio_hal_put_stag(rdev_p->rscp, stag_idx);
 748
 749         wptr = rdev_p->ctrl_qp.wptr;
 750         mutex_unlock(&rdev_p->ctrl_qp.lock);
 751         if (!err)
 752                 if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
 753                                              SEQ32_GE(rdev_p->ctrl_qp.rptr,
 754                                                       wptr)))
 755                         return -ERESTARTSYS;
 756         return err;
 757 }
 758
 759 int cxio_write_pbl(struct cxio_rdev *rdev_p, __be64 *pbl,
 760                    u32 pbl_addr, u32 pbl_size)
 761 {
 762         u32 wptr;
 763         int err;
 764
 765         PDBG("%s *pdb_addr 0x%x, pbl_base 0x%x, pbl_size %d\n",
 766              __func__, pbl_addr, rdev_p->rnic_info.pbl_base,
 767              pbl_size);
 768
 769         mutex_lock(&rdev_p->ctrl_qp.lock);
 770         err = cxio_hal_ctrl_qp_write_mem(rdev_p, pbl_addr >> 5, pbl_size << 3,
 771                                          pbl);
 772         wptr = rdev_p->ctrl_qp.wptr;
 773         mutex_unlock(&rdev_p->ctrl_qp.lock);
 774         if (err)
 775                 return err;
 776
 777         if (wait_event_interruptible(rdev_p->ctrl_qp.waitq,
 778                                      SEQ32_GE(rdev_p->ctrl_qp.rptr,
 779                                               wptr)))
 780                 return -ERESTARTSYS;
 781
 782         return 0;
 783 }
 784
 785 int cxio_register_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 786                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
 787                            u8 page_size, u32 pbl_size, u32 pbl_addr)
 788 {
 789         *stag = T3_STAG_UNSET;
 790         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
 791                              zbva, to, len, page_size, pbl_size, pbl_addr);
 792 }
 793
 794 int cxio_reregister_phys_mem(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid,
 795                            enum tpt_mem_perm perm, u32 zbva, u64 to, u32 len,
 796                            u8 page_size, u32 pbl_size, u32 pbl_addr)
 797 {
 798         return __cxio_tpt_op(rdev_p, 0, stag, 1, pdid, TPT_NON_SHARED_MR, perm,
 799                              zbva, to, len, page_size, pbl_size, pbl_addr);
 800 }
 801
 802 int cxio_dereg_mem(struct cxio_rdev *rdev_p, u32 stag, u32 pbl_size,
 803                    u32 pbl_addr)
 804 {
 805         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
 806                              pbl_size, pbl_addr);
 807 }
 808
 809 int cxio_allocate_window(struct cxio_rdev *rdev_p, u32 * stag, u32 pdid)
 810 {
 811         *stag = T3_STAG_UNSET;
 812         return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_MW, 0, 0, 0ULL, 0, 0,
 813                              0, 0);
 814 }
 815
 816 int cxio_deallocate_window(struct cxio_rdev *rdev_p, u32 stag)
 817 {
 818         return __cxio_tpt_op(rdev_p, 1, &stag, 0, 0, 0, 0, 0, 0ULL, 0, 0,
 819                              0, 0);
 820 }
 821
 822 int cxio_allocate_stag(struct cxio_rdev *rdev_p, u32 *stag, u32 pdid, u32 pbl_size, u32 pbl_addr)
 823 {
 824         *stag = T3_STAG_UNSET;
 825         return __cxio_tpt_op(rdev_p, 0, stag, 0, pdid, TPT_NON_SHARED_MR,
 826                              0, 0, 0ULL, 0, 0, pbl_size, pbl_addr);
 827 }
 828
 829 int cxio_rdma_init(struct cxio_rdev *rdev_p, struct t3_rdma_init_attr *attr)
 830 {
 831         struct t3_rdma_init_wr *wqe;
 832         struct sk_buff *skb = alloc_skb(sizeof(*wqe), GFP_ATOMIC);
 833         if (!skb)
 834                 return -ENOMEM;
 835         PDBG("%s rdev_p %p\n", __func__, rdev_p);
 836         wqe = (struct t3_rdma_init_wr *) __skb_put(skb, sizeof(*wqe));
 837         wqe->wrh.op_seop_flags = cpu_to_be32(V_FW_RIWR_OP(T3_WR_INIT));
 838         wqe->wrh.gen_tid_len = cpu_to_be32(V_FW_RIWR_TID(attr->tid) |
 839                                            V_FW_RIWR_LEN(sizeof(*wqe) >> 3));
 840         wqe->wrid.id1 = 0;
 841         wqe->qpid = cpu_to_be32(attr->qpid);
 842         wqe->pdid = cpu_to_be32(attr->pdid);
 843         wqe->scqid = cpu_to_be32(attr->scqid);
 844         wqe->rcqid = cpu_to_be32(attr->rcqid);
 845         wqe->rq_addr = cpu_to_be32(attr->rq_addr - rdev_p->rnic_info.rqt_base);
 846         wqe->rq_size = cpu_to_be32(attr->rq_size);
 847         wqe->mpaattrs = attr->mpaattrs;
 848         wqe->qpcaps = attr->qpcaps;
 849         wqe->ulpdu_size = cpu_to_be16(attr->tcp_emss);
 850         wqe->rqe_count = cpu_to_be16(attr->rqe_count);
 851         wqe->flags_rtr_type = cpu_to_be16(attr->flags|V_RTR_TYPE(attr->rtr_type));
 852         wqe->ord = cpu_to_be32(attr->ord);
 853         wqe->ird = cpu_to_be32(attr->ird);
 854         wqe->qp_dma_addr = cpu_to_be64(attr->qp_dma_addr);
 855         wqe->qp_dma_size = cpu_to_be32(attr->qp_dma_size);
 856         wqe->irs = cpu_to_be32(attr->irs);
 857         skb->priority = 0;      /* 0=>ToeQ; 1=>CtrlQ */
 858         return (cxgb3_ofld_send(rdev_p->t3cdev_p, skb));
 859 }
 860
 861 void cxio_register_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
 862 {
 863         cxio_ev_cb = ev_cb;
 864 }
 865
 866 void cxio_unregister_ev_cb(cxio_hal_ev_callback_func_t ev_cb)
 867 {
 868         cxio_ev_cb = NULL;
 869 }
 870
 871 static int cxio_hal_ev_handler(struct t3cdev *t3cdev_p, struct sk_buff *skb)
 872 {
 873         static int cnt;
 874         struct cxio_rdev *rdev_p = NULL;
 875         struct respQ_msg_t *rsp_msg = (struct respQ_msg_t *) skb->data;
 876         PDBG("%d: %s cq_id 0x%x cq_ptr 0x%x genbit %0x overflow %0x an %0x"
 877              " se %0x notify %0x cqbranch %0x creditth %0x\n",
 878              cnt, __func__, RSPQ_CQID(rsp_msg), RSPQ_CQPTR(rsp_msg),
 879              RSPQ_GENBIT(rsp_msg), RSPQ_OVERFLOW(rsp_msg), RSPQ_AN(rsp_msg),
 880              RSPQ_SE(rsp_msg), RSPQ_NOTIFY(rsp_msg), RSPQ_CQBRANCH(rsp_msg),
 881              RSPQ_CREDIT_THRESH(rsp_msg));
 882         PDBG("CQE: QPID 0x%0x genbit %0x type 0x%0x status 0x%0x opcode %d "
 883              "len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
 884              CQE_QPID(rsp_msg->cqe), CQE_GENBIT(rsp_msg->cqe),
 885              CQE_TYPE(rsp_msg->cqe), CQE_STATUS(rsp_msg->cqe),
 886              CQE_OPCODE(rsp_msg->cqe), CQE_LEN(rsp_msg->cqe),
 887              CQE_WRID_HI(rsp_msg->cqe), CQE_WRID_LOW(rsp_msg->cqe));
 888         rdev_p = (struct cxio_rdev *)t3cdev_p->ulp;
 889         if (!rdev_p) {
 890                 PDBG("%s called by t3cdev %p with null ulp\n", __func__,
 891                      t3cdev_p);
 892                 return 0;
 893         }
 894         if (CQE_QPID(rsp_msg->cqe) == T3_CTRL_QP_ID) {
 895                 rdev_p->ctrl_qp.rptr = CQE_WRID_LOW(rsp_msg->cqe) + 1;
 896                 wake_up_interruptible(&rdev_p->ctrl_qp.waitq);
 897                 dev_kfree_skb_irq(skb);
 898         } else if (CQE_QPID(rsp_msg->cqe) == 0xfff8)
 899                 dev_kfree_skb_irq(skb);
 900         else if (cxio_ev_cb)
 901                 (*cxio_ev_cb) (rdev_p, skb);
 902         else
 903                 dev_kfree_skb_irq(skb);
 904         cnt++;
 905         return 0;
 906 }
 907
 908 /* Caller takes care of locking if needed */
 909 int cxio_rdev_open(struct cxio_rdev *rdev_p)
 910 {
 911         struct net_device *netdev_p = NULL;
 912         int err = 0;
 913         if (strlen(rdev_p->dev_name)) {
 914                 if (cxio_hal_find_rdev_by_name(rdev_p->dev_name)) {
 915                         return -EBUSY;
 916                 }
 917                 netdev_p = dev_get_by_name(&init_net, rdev_p->dev_name);
 918                 if (!netdev_p) {
 919                         return -EINVAL;
 920                 }
 921                 dev_put(netdev_p);
 922         } else if (rdev_p->t3cdev_p) {
 923                 if (cxio_hal_find_rdev_by_t3cdev(rdev_p->t3cdev_p)) {
 924                         return -EBUSY;
 925                 }
 926                 netdev_p = rdev_p->t3cdev_p->lldev;
 927                 strncpy(rdev_p->dev_name, rdev_p->t3cdev_p->name,
 928                         T3_MAX_DEV_NAME_LEN);
 929         } else {
 930                 PDBG("%s t3cdev_p or dev_name must be set\n", __func__);
 931                 return -EINVAL;
 932         }
 933
 934         list_add_tail(&rdev_p->entry, &rdev_list);
 935
 936         PDBG("%s opening rnic dev %s\n", __func__, rdev_p->dev_name);
 937         memset(&rdev_p->ctrl_qp, 0, sizeof(rdev_p->ctrl_qp));
 938         if (!rdev_p->t3cdev_p)
 939                 rdev_p->t3cdev_p = dev2t3cdev(netdev_p);
 940         rdev_p->t3cdev_p->ulp = (void *) rdev_p;
 941         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, RDMA_GET_PARAMS,
 942                                          &(rdev_p->rnic_info));
 943         if (err) {
 944                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
 945                      __func__, rdev_p->t3cdev_p, err);
 946                 goto err1;
 947         }
 948         err = rdev_p->t3cdev_p->ctl(rdev_p->t3cdev_p, GET_PORTS,
 949                                     &(rdev_p->port_info));
 950         if (err) {
 951                 printk(KERN_ERR "%s t3cdev_p(%p)->ctl returned error %d.\n",
 952                      __func__, rdev_p->t3cdev_p, err);
 953                 goto err1;
 954         }
 955
 956         /*
 957          * qpshift is the number of bits to shift the qpid left in order
 958          * to get the correct address of the doorbell for that qp.
 959          */
 960         cxio_init_ucontext(rdev_p, &rdev_p->uctx);
 961         rdev_p->qpshift = PAGE_SHIFT -
 962                           ilog2(65536 >>
 963                                     ilog2(rdev_p->rnic_info.udbell_len >>
 964                                               PAGE_SHIFT));
 965         rdev_p->qpnr = rdev_p->rnic_info.udbell_len >> PAGE_SHIFT;
 966         rdev_p->qpmask = (65536 >> ilog2(rdev_p->qpnr)) - 1;
 967         PDBG("%s rnic %s info: tpt_base 0x%0x tpt_top 0x%0x num stags %d "
 968              "pbl_base 0x%0x pbl_top 0x%0x rqt_base 0x%0x, rqt_top 0x%0x\n",
 969              __func__, rdev_p->dev_name, rdev_p->rnic_info.tpt_base,
 970              rdev_p->rnic_info.tpt_top, cxio_num_stags(rdev_p),
 971              rdev_p->rnic_info.pbl_base,
 972              rdev_p->rnic_info.pbl_top, rdev_p->rnic_info.rqt_base,
 973              rdev_p->rnic_info.rqt_top);
 974         PDBG("udbell_len 0x%0x udbell_physbase 0x%lx kdb_addr %p qpshift %lu "
 975              "qpnr %d qpmask 0x%x\n",
 976              rdev_p->rnic_info.udbell_len,
 977              rdev_p->rnic_info.udbell_physbase, rdev_p->rnic_info.kdb_addr,
 978              rdev_p->qpshift, rdev_p->qpnr, rdev_p->qpmask);
 979
 980         err = cxio_hal_init_ctrl_qp(rdev_p);
 981         if (err) {
 982                 printk(KERN_ERR "%s error %d initializing ctrl_qp.\n",
 983                        __func__, err);
 984                 goto err1;
 985         }
 986         err = cxio_hal_init_resource(rdev_p, cxio_num_stags(rdev_p), 0,
 987                                      0, T3_MAX_NUM_QP, T3_MAX_NUM_CQ,
 988                                      T3_MAX_NUM_PD);
 989         if (err) {
 990                 printk(KERN_ERR "%s error %d initializing hal resources.\n",
 991                        __func__, err);
 992                 goto err2;
 993         }
 994         err = cxio_hal_pblpool_create(rdev_p);
 995         if (err) {
 996                 printk(KERN_ERR "%s error %d initializing pbl mem pool.\n",
 997                        __func__, err);
 998                 goto err3;
 999         }
1000         err = cxio_hal_rqtpool_create(rdev_p);
1001         if (err) {
1002                 printk(KERN_ERR "%s error %d initializing rqt mem pool.\n",
1003                        __func__, err);
1004                 goto err4;
1005         }
1006         return 0;
1007 err4:
1008         cxio_hal_pblpool_destroy(rdev_p);
1009 err3:
1010         cxio_hal_destroy_resource(rdev_p->rscp);
1011 err2:
1012         cxio_hal_destroy_ctrl_qp(rdev_p);
1013 err1:
1014         list_del(&rdev_p->entry);
1015         return err;
1016 }
1017
1018 void cxio_rdev_close(struct cxio_rdev *rdev_p)
1019 {
1020         if (rdev_p) {
1021                 cxio_hal_pblpool_destroy(rdev_p);
1022                 cxio_hal_rqtpool_destroy(rdev_p);
1023                 list_del(&rdev_p->entry);
1024                 rdev_p->t3cdev_p->ulp = NULL;
1025                 cxio_hal_destroy_ctrl_qp(rdev_p);
1026                 cxio_hal_destroy_resource(rdev_p->rscp);
1027         }
1028 }
1029
1030 int __init cxio_hal_init(void)
1031 {
1032         if (cxio_hal_init_rhdl_resource(T3_MAX_NUM_RI))
1033                 return -ENOMEM;
1034         t3_register_cpl_handler(CPL_ASYNC_NOTIF, cxio_hal_ev_handler);
1035         return 0;
1036 }
1037
1038 void __exit cxio_hal_exit(void)
1039 {
1040         struct cxio_rdev *rdev, *tmp;
1041
1042         t3_register_cpl_handler(CPL_ASYNC_NOTIF, NULL);
1043         list_for_each_entry_safe(rdev, tmp, &rdev_list, entry)
1044                 cxio_rdev_close(rdev);
1045         cxio_hal_destroy_rhdl_resource();
1046 }
1047
1048 static void flush_completed_wrs(struct t3_wq *wq, struct t3_cq *cq)
1049 {
1050         struct t3_swsq *sqp;
1051         __u32 ptr = wq->sq_rptr;
1052         int count = Q_COUNT(wq->sq_rptr, wq->sq_wptr);
1053
1054         sqp = wq->sq + Q_PTR2IDX(ptr, wq->sq_size_log2);
1055         while (count--)
1056                 if (!sqp->signaled) {
1057                         ptr++;
1058                         sqp = wq->sq + Q_PTR2IDX(ptr,  wq->sq_size_log2);
1059                 } else if (sqp->complete) {
1060
1061                         /*
1062                          * Insert this completed cqe into the swcq.
1063                          */
1064                         PDBG("%s moving cqe into swcq sq idx %ld cq idx %ld\n",
1065                              __func__, Q_PTR2IDX(ptr,  wq->sq_size_log2),
1066                              Q_PTR2IDX(cq->sw_wptr, cq->size_log2));
1067                         sqp->cqe.header |= htonl(V_CQE_SWCQE(1));
1068                         *(cq->sw_queue + Q_PTR2IDX(cq->sw_wptr, cq->size_log2))
1069                                 = sqp->cqe;
1070                         cq->sw_wptr++;
1071                         sqp->signaled = 0;
1072                         break;
1073                 } else
1074                         break;
1075 }
1076
1077 static void create_read_req_cqe(struct t3_wq *wq, struct t3_cqe *hw_cqe,
1078                                 struct t3_cqe *read_cqe)
1079 {
1080         read_cqe->u.scqe.wrid_hi = wq->oldest_read->sq_wptr;
1081         read_cqe->len = wq->oldest_read->read_len;
1082         read_cqe->header = htonl(V_CQE_QPID(CQE_QPID(*hw_cqe)) |
1083                                  V_CQE_SWCQE(SW_CQE(*hw_cqe)) |
1084                                  V_CQE_OPCODE(T3_READ_REQ) |
1085                                  V_CQE_TYPE(1));
1086 }
1087
1088 /*
1089  * Return a ptr to the next read wr in the SWSQ or NULL.
1090  */
1091 static void advance_oldest_read(struct t3_wq *wq)
1092 {
1093
1094         u32 rptr = wq->oldest_read - wq->sq + 1;
1095         u32 wptr = Q_PTR2IDX(wq->sq_wptr, wq->sq_size_log2);
1096
1097         while (Q_PTR2IDX(rptr, wq->sq_size_log2) != wptr) {
1098                 wq->oldest_read = wq->sq + Q_PTR2IDX(rptr, wq->sq_size_log2);
1099
1100                 if (wq->oldest_read->opcode == T3_READ_REQ)
1101                         return;
1102                 rptr++;
1103         }
1104         wq->oldest_read = NULL;
1105 }
1106
1107 /*
1108  * cxio_poll_cq
1109  *
1110  * Caller must:
1111  *     check the validity of the first CQE,
1112  *     supply the wq assicated with the qpid.
1113  *
1114  * credit: cq credit to return to sge.
1115  * cqe_flushed: 1 iff the CQE is flushed.
1116  * cqe: copy of the polled CQE.
1117  *
1118  * return value:
1119  *     0       CQE returned,
1120  *    -1       CQE skipped, try again.
1121  */
1122 int cxio_poll_cq(struct t3_wq *wq, struct t3_cq *cq, struct t3_cqe *cqe,
1123                      u8 *cqe_flushed, u64 *cookie, u32 *credit)
1124 {
1125         int ret = 0;
1126         struct t3_cqe *hw_cqe, read_cqe;
1127
1128         *cqe_flushed = 0;
1129         *credit = 0;
1130         hw_cqe = cxio_next_cqe(cq);
1131
1132         PDBG("%s CQE OOO %d qpid 0x%0x genbit %d type %d status 0x%0x"
1133              " opcode 0x%0x len 0x%0x wrid_hi_stag 0x%x wrid_low_msn 0x%x\n",
1134              __func__, CQE_OOO(*hw_cqe), CQE_QPID(*hw_cqe),
1135              CQE_GENBIT(*hw_cqe), CQE_TYPE(*hw_cqe), CQE_STATUS(*hw_cqe),
1136              CQE_OPCODE(*hw_cqe), CQE_LEN(*hw_cqe), CQE_WRID_HI(*hw_cqe),
1137              CQE_WRID_LOW(*hw_cqe));
1138
1139         /*
1140          * skip cqe's not affiliated with a QP.
1141          */
1142         if (wq == NULL) {
1143                 ret = -1;
1144                 goto skip_cqe;
1145         }
1146
1147         /*
1148          * Gotta tweak READ completions:
1149          *      1) the cqe doesn't contain the sq_wptr from the wr.
1150          *      2) opcode not reflected from the wr.
1151          *      3) read_len not reflected from the wr.
1152          *      4) cq_type is RQ_TYPE not SQ_TYPE.
1153          */
1154         if (RQ_TYPE(*hw_cqe) && (CQE_OPCODE(*hw_cqe) == T3_READ_RESP)) {
1155
1156                 /*
1157                  * If this is an unsolicited read response, then the read
1158                  * was generated by the kernel driver as part of peer-2-peer
1159                  * connection setup.  So ignore the completion.
1160                  */
1161                 if (!wq->oldest_read) {
1162                         if (CQE_STATUS(*hw_cqe))
1163                                 wq->error = 1;
1164                         ret = -1;
1165                         goto skip_cqe;
1166                 }
1167
1168                 /*
1169                  * Don't write to the HWCQ, so create a new read req CQE
1170                  * in local memory.
1171                  */
1172                 create_read_req_cqe(wq, hw_cqe, &read_cqe);
1173                 hw_cqe = &read_cqe;
1174                 advance_oldest_read(wq);
1175         }
1176
1177         /*
1178          * T3A: Discard TERMINATE CQEs.
1179          */
1180         if (CQE_OPCODE(*hw_cqe) == T3_TERMINATE) {
1181                 ret = -1;
1182                 wq->error = 1;
1183                 goto skip_cqe;
1184         }
1185
1186         if (CQE_STATUS(*hw_cqe) || wq->error) {
1187                 *cqe_flushed = wq->error;
1188                 wq->error = 1;
1189
1190                 /*
1191                  * T3A inserts errors into the CQE.  We cannot return
1192                  * these as work completions.
1193                  */
1194                 /* incoming write failures */
1195                 if ((CQE_OPCODE(*hw_cqe) == T3_RDMA_WRITE)
1196                      && RQ_TYPE(*hw_cqe)) {
1197                         ret = -1;
1198                         goto skip_cqe;
1199                 }
1200                 /* incoming read request failures */
1201                 if ((CQE_OPCODE(*hw_cqe) == T3_READ_RESP) && SQ_TYPE(*hw_cqe)) {
1202                         ret = -1;
1203                         goto skip_cqe;
1204                 }
1205
1206                 /* incoming SEND with no receive posted failures */
1207                 if ((CQE_OPCODE(*hw_cqe) == T3_SEND) && RQ_TYPE(*hw_cqe) &&
1208                     Q_EMPTY(wq->rq_rptr, wq->rq_wptr)) {
1209                         ret = -1;
1210                         goto skip_cqe;
1211                 }
1212                 goto proc_cqe;
1213         }
1214
1215         /*
1216          * RECV completion.
1217          */
1218         if (RQ_TYPE(*hw_cqe)) {
1219
1220                 /*
1221                  * HW only validates 4 bits of MSN.  So we must validate that
1222                  * the MSN in the SEND is the next expected MSN.  If its not,
1223                  * then we complete this with TPT_ERR_MSN and mark the wq in
1224                  * error.
1225                  */
1226                 if (unlikely((CQE_WRID_MSN(*hw_cqe) != (wq->rq_rptr + 1)))) {
1227                         wq->error = 1;
1228                         hw_cqe->header |= htonl(V_CQE_STATUS(TPT_ERR_MSN));
1229                         goto proc_cqe;
1230                 }
1231                 goto proc_cqe;
1232         }
1233
1234         /*
1235          * If we get here its a send completion.
1236          *
1237          * Handle out of order completion. These get stuffed
1238          * in the SW SQ. Then the SW SQ is walked to move any
1239          * now in-order completions into the SW CQ.  This handles
1240          * 2 cases:
1241          *      1) reaping unsignaled WRs when the first subsequent
1242          *         signaled WR is completed.
1243          *      2) out of order read completions.
1244          */
1245         if (!SW_CQE(*hw_cqe) && (CQE_WRID_SQ_WPTR(*hw_cqe) != wq->sq_rptr)) {
1246                 struct t3_swsq *sqp;
1247
1248                 PDBG("%s out of order completion going in swsq at idx %ld\n",
1249                      __func__,
1250                      Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2));
1251                 sqp = wq->sq +
1252                       Q_PTR2IDX(CQE_WRID_SQ_WPTR(*hw_cqe), wq->sq_size_log2);
1253                 sqp->cqe = *hw_cqe;
1254                 sqp->complete = 1;
1255                 ret = -1;
1256                 goto flush_wq;
1257         }
1258
1259 proc_cqe:
1260         *cqe = *hw_cqe;
1261
1262         /*
1263          * Reap the associated WR(s) that are freed up with this
1264          * completion.
1265          */
1266         if (SQ_TYPE(*hw_cqe)) {
1267                 wq->sq_rptr = CQE_WRID_SQ_WPTR(*hw_cqe);
1268                 PDBG("%s completing sq idx %ld\n", __func__,
1269                      Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2));
1270                 *cookie = wq->sq[Q_PTR2IDX(wq->sq_rptr, wq->sq_size_log2)].wr_id;
1271                 wq->sq_rptr++;
1272         } else {
1273                 PDBG("%s completing rq idx %ld\n", __func__,
1274                      Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2));
1275                 *cookie = wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].wr_id;
1276                 if (wq->rq[Q_PTR2IDX(wq->rq_rptr, wq->rq_size_log2)].pbl_addr)
1277                         cxio_hal_pblpool_free(wq->rdev,
1278                                 wq->rq[Q_PTR2IDX(wq->rq_rptr,
1279                                 wq->rq_size_log2)].pbl_addr, T3_STAG0_PBL_SIZE);
1280                 wq->rq_rptr++;
1281         }
1282
1283 flush_wq:
1284         /*
1285          * Flush any completed cqes that are now in-order.
1286          */
1287         flush_completed_wrs(wq, cq);
1288
1289 skip_cqe:
1290         if (SW_CQE(*hw_cqe)) {
1291                 PDBG("%s cq %p cqid 0x%x skip sw cqe sw_rptr 0x%x\n",
1292                      __func__, cq, cq->cqid, cq->sw_rptr);
1293                 ++cq->sw_rptr;
1294         } else {
1295                 PDBG("%s cq %p cqid 0x%x skip hw cqe rptr 0x%x\n",
1296                      __func__, cq, cq->cqid, cq->rptr);
1297                 ++cq->rptr;
1298
1299                 /*
1300                  * T3A: compute credits.
1301                  */
1302                 if (((cq->rptr - cq->wptr) > (1 << (cq->size_log2 - 1)))
1303                     || ((cq->rptr - cq->wptr) >= 128)) {
1304                         *credit = cq->rptr - cq->wptr;
1305                         cq->wptr = cq->rptr;
1306                 }
1307         }
1308         return ret;
1309 }