drivers/infiniband/hw/bnxt_re/main.c

   1 /*
   2  * Broadcom NetXtreme-E RoCE driver.
   3  *
   4  * Copyright (c) 2016 - 2017, Broadcom. All rights reserved.  The term
   5  * Broadcom refers to Broadcom Limited and/or its subsidiaries.
   6  *
   7  * This software is available to you under a choice of one of two
   8  * licenses.  You may choose to be licensed under the terms of the GNU
   9  * General Public License (GPL) Version 2, available from the file
  10  * COPYING in the main directory of this source tree, or the
  11  * BSD license below:
  12  *
  13  * Redistribution and use in source and binary forms, with or without
  14  * modification, are permitted provided that the following conditions
  15  * are met:
  16  *
  17  * 1. Redistributions of source code must retain the above copyright
  18  *    notice, this list of conditions and the following disclaimer.
  19  * 2. Redistributions in binary form must reproduce the above copyright
  20  *    notice, this list of conditions and the following disclaimer in
  21  *    the documentation and/or other materials provided with the
  22  *    distribution.
  23  *
  24  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
  25  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  26  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  27  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
  28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
  29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
  30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
  31  * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  32  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
  33  * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
  34  * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  35  *
  36  * Description: Main component of the bnxt_re driver
  37  */
  38
  39 #include <linux/module.h>
  40 #include <linux/netdevice.h>
  41 #include <linux/ethtool.h>
  42 #include <linux/mutex.h>
  43 #include <linux/list.h>
  44 #include <linux/rculist.h>
  45 #include <linux/spinlock.h>
  46 #include <linux/pci.h>
  47 #include <net/dcbnl.h>
  48 #include <net/ipv6.h>
  49 #include <net/addrconf.h>
  50 #include <linux/if_ether.h>
  51
  52 #include <rdma/ib_verbs.h>
  53 #include <rdma/ib_user_verbs.h>
  54 #include <rdma/ib_umem.h>
  55 #include <rdma/ib_addr.h>
  56
  57 #include "bnxt_ulp.h"
  58 #include "roce_hsi.h"
  59 #include "qplib_res.h"
  60 #include "qplib_sp.h"
  61 #include "qplib_fp.h"
  62 #include "qplib_rcfw.h"
  63 #include "bnxt_re.h"
  64 #include "ib_verbs.h"
  65 #include <rdma/bnxt_re-abi.h>
  66 #include "bnxt.h"
  67 #include "hw_counters.h"
  68
  69 static char version[] =
  70                 BNXT_RE_DESC "\n";
  71
  72 MODULE_AUTHOR("Eddie Wai <eddie.wai@broadcom.com>");
  73 MODULE_DESCRIPTION(BNXT_RE_DESC " Driver");
  74 MODULE_LICENSE("Dual BSD/GPL");
  75
  76 /* globals */
  77 static struct list_head bnxt_re_dev_list = LIST_HEAD_INIT(bnxt_re_dev_list);
  78 /* Mutex to protect the list of bnxt_re devices added */
  79 static DEFINE_MUTEX(bnxt_re_dev_lock);
  80 static struct workqueue_struct *bnxt_re_wq;
  81 static void bnxt_re_remove_device(struct bnxt_re_dev *rdev);
  82 static void bnxt_re_dealloc_driver(struct ib_device *ib_dev);
  83 static void bnxt_re_stop_irq(void *handle);
  84 static void bnxt_re_dev_stop(struct bnxt_re_dev *rdev);
  85
  86 static void bnxt_re_set_drv_mode(struct bnxt_re_dev *rdev, u8 mode)
  87 {
  88         struct bnxt_qplib_chip_ctx *cctx;
  89
  90         cctx = rdev->chip_ctx;
  91         cctx->modes.wqe_mode = bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
  92                                mode : BNXT_QPLIB_WQE_MODE_STATIC;
  93 }
  94
  95 static void bnxt_re_destroy_chip_ctx(struct bnxt_re_dev *rdev)
  96 {
  97         struct bnxt_qplib_chip_ctx *chip_ctx;
  98
  99         if (!rdev->chip_ctx)
 100                 return;
 101         chip_ctx = rdev->chip_ctx;
 102         rdev->chip_ctx = NULL;
 103         rdev->rcfw.res = NULL;
 104         rdev->qplib_res.cctx = NULL;
 105         rdev->qplib_res.pdev = NULL;
 106         rdev->qplib_res.netdev = NULL;
 107         kfree(chip_ctx);
 108 }
 109
 110 static int bnxt_re_setup_chip_ctx(struct bnxt_re_dev *rdev, u8 wqe_mode)
 111 {
 112         struct bnxt_qplib_chip_ctx *chip_ctx;
 113         struct bnxt_en_dev *en_dev;
 114         struct bnxt *bp;
 115
 116         en_dev = rdev->en_dev;
 117         bp = netdev_priv(en_dev->net);
 118
 119         chip_ctx = kzalloc(sizeof(*chip_ctx), GFP_KERNEL);
 120         if (!chip_ctx)
 121                 return -ENOMEM;
 122         chip_ctx->chip_num = bp->chip_num;
 123
 124         rdev->chip_ctx = chip_ctx;
 125         /* rest members to follow eventually */
 126
 127         rdev->qplib_res.cctx = rdev->chip_ctx;
 128         rdev->rcfw.res = &rdev->qplib_res;
 129
 130         bnxt_re_set_drv_mode(rdev, wqe_mode);
 131         return 0;
 132 }
 133
 134 /* SR-IOV helper functions */
 135
 136 static void bnxt_re_get_sriov_func_type(struct bnxt_re_dev *rdev)
 137 {
 138         struct bnxt *bp;
 139
 140         bp = netdev_priv(rdev->en_dev->net);
 141         if (BNXT_VF(bp))
 142                 rdev->is_virtfn = 1;
 143 }
 144
 145 /* Set the maximum number of each resource that the driver actually wants
 146  * to allocate. This may be up to the maximum number the firmware has
 147  * reserved for the function. The driver may choose to allocate fewer
 148  * resources than the firmware maximum.
 149  */
 150 static void bnxt_re_limit_pf_res(struct bnxt_re_dev *rdev)
 151 {
 152         struct bnxt_qplib_dev_attr *attr;
 153         struct bnxt_qplib_ctx *ctx;
 154         int i;
 155
 156         attr = &rdev->dev_attr;
 157         ctx = &rdev->qplib_ctx;
 158
 159         ctx->qpc_count = min_t(u32, BNXT_RE_MAX_QPC_COUNT,
 160                                attr->max_qp);
 161         ctx->mrw_count = BNXT_RE_MAX_MRW_COUNT_256K;
 162         /* Use max_mr from fw since max_mrw does not get set */
 163         ctx->mrw_count = min_t(u32, ctx->mrw_count, attr->max_mr);
 164         ctx->srqc_count = min_t(u32, BNXT_RE_MAX_SRQC_COUNT,
 165                                 attr->max_srq);
 166         ctx->cq_count = min_t(u32, BNXT_RE_MAX_CQ_COUNT, attr->max_cq);
 167         if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx))
 168                 for (i = 0; i < MAX_TQM_ALLOC_REQ; i++)
 169                         rdev->qplib_ctx.tqm_ctx.qcount[i] =
 170                         rdev->dev_attr.tqm_alloc_reqs[i];
 171 }
 172
 173 static void bnxt_re_limit_vf_res(struct bnxt_qplib_ctx *qplib_ctx, u32 num_vf)
 174 {
 175         struct bnxt_qplib_vf_res *vf_res;
 176         u32 mrws = 0;
 177         u32 vf_pct;
 178         u32 nvfs;
 179
 180         vf_res = &qplib_ctx->vf_res;
 181         /*
 182          * Reserve a set of resources for the PF. Divide the remaining
 183          * resources among the VFs
 184          */
 185         vf_pct = 100 - BNXT_RE_PCT_RSVD_FOR_PF;
 186         nvfs = num_vf;
 187         num_vf = 100 * num_vf;
 188         vf_res->max_qp_per_vf = (qplib_ctx->qpc_count * vf_pct) / num_vf;
 189         vf_res->max_srq_per_vf = (qplib_ctx->srqc_count * vf_pct) / num_vf;
 190         vf_res->max_cq_per_vf = (qplib_ctx->cq_count * vf_pct) / num_vf;
 191         /*
 192          * The driver allows many more MRs than other resources. If the
 193          * firmware does also, then reserve a fixed amount for the PF and
 194          * divide the rest among VFs. VFs may use many MRs for NFS
 195          * mounts, ISER, NVME applications, etc. If the firmware severely
 196          * restricts the number of MRs, then let PF have half and divide
 197          * the rest among VFs, as for the other resource types.
 198          */
 199         if (qplib_ctx->mrw_count < BNXT_RE_MAX_MRW_COUNT_64K) {
 200                 mrws = qplib_ctx->mrw_count * vf_pct;
 201                 nvfs = num_vf;
 202         } else {
 203                 mrws = qplib_ctx->mrw_count - BNXT_RE_RESVD_MR_FOR_PF;
 204         }
 205         vf_res->max_mrw_per_vf = (mrws / nvfs);
 206         vf_res->max_gid_per_vf = BNXT_RE_MAX_GID_PER_VF;
 207 }
 208
 209 static void bnxt_re_set_resource_limits(struct bnxt_re_dev *rdev)
 210 {
 211         u32 num_vfs;
 212
 213         memset(&rdev->qplib_ctx.vf_res, 0, sizeof(struct bnxt_qplib_vf_res));
 214         bnxt_re_limit_pf_res(rdev);
 215
 216         num_vfs =  bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
 217                         BNXT_RE_GEN_P5_MAX_VF : rdev->num_vfs;
 218         if (num_vfs)
 219                 bnxt_re_limit_vf_res(&rdev->qplib_ctx, num_vfs);
 220 }
 221
 222 /* for handling bnxt_en callbacks later */
 223 static void bnxt_re_stop(void *p)
 224 {
 225         struct bnxt_re_dev *rdev = p;
 226         struct bnxt *bp;
 227
 228         if (!rdev)
 229                 return;
 230         ASSERT_RTNL();
 231
 232         /* L2 driver invokes this callback during device error/crash or device
 233          * reset. Current RoCE driver doesn't recover the device in case of
 234          * error. Handle the error by dispatching fatal events to all qps
 235          * ie. by calling bnxt_re_dev_stop and release the MSIx vectors as
 236          * L2 driver want to modify the MSIx table.
 237          */
 238         bp = netdev_priv(rdev->netdev);
 239
 240         ibdev_info(&rdev->ibdev, "Handle device stop call from L2 driver");
 241         /* Check the current device state from L2 structure and move the
 242          * device to detached state if FW_FATAL_COND is set.
 243          * This prevents more commands to HW during clean-up,
 244          * in case the device is already in error.
 245          */
 246         if (test_bit(BNXT_STATE_FW_FATAL_COND, &bp->state))
 247                 set_bit(ERR_DEVICE_DETACHED, &rdev->rcfw.cmdq.flags);
 248
 249         bnxt_re_dev_stop(rdev);
 250         bnxt_re_stop_irq(rdev);
 251         /* Move the device states to detached and  avoid sending any more
 252          * commands to HW
 253          */
 254         set_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags);
 255         set_bit(ERR_DEVICE_DETACHED, &rdev->rcfw.cmdq.flags);
 256 }
 257
 258 static void bnxt_re_start(void *p)
 259 {
 260 }
 261
 262 static void bnxt_re_sriov_config(void *p, int num_vfs)
 263 {
 264         struct bnxt_re_dev *rdev = p;
 265
 266         if (!rdev)
 267                 return;
 268
 269         if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags))
 270                 return;
 271         rdev->num_vfs = num_vfs;
 272         if (!bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx)) {
 273                 bnxt_re_set_resource_limits(rdev);
 274                 bnxt_qplib_set_func_resources(&rdev->qplib_res, &rdev->rcfw,
 275                                               &rdev->qplib_ctx);
 276         }
 277 }
 278
 279 static void bnxt_re_shutdown(void *p)
 280 {
 281         struct bnxt_re_dev *rdev = p;
 282
 283         if (!rdev)
 284                 return;
 285         ASSERT_RTNL();
 286         /* Release the MSIx vectors before queuing unregister */
 287         bnxt_re_stop_irq(rdev);
 288         ib_unregister_device_queued(&rdev->ibdev);
 289 }
 290
 291 static void bnxt_re_stop_irq(void *handle)
 292 {
 293         struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle;
 294         struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw;
 295         struct bnxt_qplib_nq *nq;
 296         int indx;
 297
 298         for (indx = BNXT_RE_NQ_IDX; indx < rdev->num_msix; indx++) {
 299                 nq = &rdev->nq[indx - 1];
 300                 bnxt_qplib_nq_stop_irq(nq, false);
 301         }
 302
 303         bnxt_qplib_rcfw_stop_irq(rcfw, false);
 304 }
 305
 306 static void bnxt_re_start_irq(void *handle, struct bnxt_msix_entry *ent)
 307 {
 308         struct bnxt_re_dev *rdev = (struct bnxt_re_dev *)handle;
 309         struct bnxt_msix_entry *msix_ent = rdev->msix_entries;
 310         struct bnxt_qplib_rcfw *rcfw = &rdev->rcfw;
 311         struct bnxt_qplib_nq *nq;
 312         int indx, rc;
 313
 314         if (!ent) {
 315                 /* Not setting the f/w timeout bit in rcfw.
 316                  * During the driver unload the first command
 317                  * to f/w will timeout and that will set the
 318                  * timeout bit.
 319                  */
 320                 ibdev_err(&rdev->ibdev, "Failed to re-start IRQs\n");
 321                 return;
 322         }
 323
 324         /* Vectors may change after restart, so update with new vectors
 325          * in device sctructure.
 326          */
 327         for (indx = 0; indx < rdev->num_msix; indx++)
 328                 rdev->msix_entries[indx].vector = ent[indx].vector;
 329
 330         bnxt_qplib_rcfw_start_irq(rcfw, msix_ent[BNXT_RE_AEQ_IDX].vector,
 331                                   false);
 332         for (indx = BNXT_RE_NQ_IDX ; indx < rdev->num_msix; indx++) {
 333                 nq = &rdev->nq[indx - 1];
 334                 rc = bnxt_qplib_nq_start_irq(nq, indx - 1,
 335                                              msix_ent[indx].vector, false);
 336                 if (rc)
 337                         ibdev_warn(&rdev->ibdev, "Failed to reinit NQ index %d\n",
 338                                    indx - 1);
 339         }
 340 }
 341
 342 static struct bnxt_ulp_ops bnxt_re_ulp_ops = {
 343         .ulp_async_notifier = NULL,
 344         .ulp_stop = bnxt_re_stop,
 345         .ulp_start = bnxt_re_start,
 346         .ulp_sriov_config = bnxt_re_sriov_config,
 347         .ulp_shutdown = bnxt_re_shutdown,
 348         .ulp_irq_stop = bnxt_re_stop_irq,
 349         .ulp_irq_restart = bnxt_re_start_irq
 350 };
 351
 352 /* RoCE -> Net driver */
 353
 354 /* Driver registration routines used to let the networking driver (bnxt_en)
 355  * to know that the RoCE driver is now installed
 356  */
 357 static int bnxt_re_unregister_netdev(struct bnxt_re_dev *rdev)
 358 {
 359         struct bnxt_en_dev *en_dev;
 360         int rc;
 361
 362         if (!rdev)
 363                 return -EINVAL;
 364
 365         en_dev = rdev->en_dev;
 366
 367         rc = en_dev->en_ops->bnxt_unregister_device(rdev->en_dev,
 368                                                     BNXT_ROCE_ULP);
 369         return rc;
 370 }
 371
 372 static int bnxt_re_register_netdev(struct bnxt_re_dev *rdev)
 373 {
 374         struct bnxt_en_dev *en_dev;
 375         int rc = 0;
 376
 377         if (!rdev)
 378                 return -EINVAL;
 379
 380         en_dev = rdev->en_dev;
 381
 382         rc = en_dev->en_ops->bnxt_register_device(en_dev, BNXT_ROCE_ULP,
 383                                                   &bnxt_re_ulp_ops, rdev);
 384         rdev->qplib_res.pdev = rdev->en_dev->pdev;
 385         return rc;
 386 }
 387
 388 static int bnxt_re_free_msix(struct bnxt_re_dev *rdev)
 389 {
 390         struct bnxt_en_dev *en_dev;
 391         int rc;
 392
 393         if (!rdev)
 394                 return -EINVAL;
 395
 396         en_dev = rdev->en_dev;
 397
 398
 399         rc = en_dev->en_ops->bnxt_free_msix(rdev->en_dev, BNXT_ROCE_ULP);
 400
 401         return rc;
 402 }
 403
 404 static int bnxt_re_request_msix(struct bnxt_re_dev *rdev)
 405 {
 406         int rc = 0, num_msix_want = BNXT_RE_MAX_MSIX, num_msix_got;
 407         struct bnxt_en_dev *en_dev;
 408
 409         if (!rdev)
 410                 return -EINVAL;
 411
 412         en_dev = rdev->en_dev;
 413
 414         num_msix_want = min_t(u32, BNXT_RE_MAX_MSIX, num_online_cpus());
 415
 416         num_msix_got = en_dev->en_ops->bnxt_request_msix(en_dev, BNXT_ROCE_ULP,
 417                                                          rdev->msix_entries,
 418                                                          num_msix_want);
 419         if (num_msix_got < BNXT_RE_MIN_MSIX) {
 420                 rc = -EINVAL;
 421                 goto done;
 422         }
 423         if (num_msix_got != num_msix_want) {
 424                 ibdev_warn(&rdev->ibdev,
 425                            "Requested %d MSI-X vectors, got %d\n",
 426                            num_msix_want, num_msix_got);
 427         }
 428         rdev->num_msix = num_msix_got;
 429 done:
 430         return rc;
 431 }
 432
 433 static void bnxt_re_init_hwrm_hdr(struct bnxt_re_dev *rdev, struct input *hdr,
 434                                   u16 opcd, u16 crid, u16 trid)
 435 {
 436         hdr->req_type = cpu_to_le16(opcd);
 437         hdr->cmpl_ring = cpu_to_le16(crid);
 438         hdr->target_id = cpu_to_le16(trid);
 439 }
 440
 441 static void bnxt_re_fill_fw_msg(struct bnxt_fw_msg *fw_msg, void *msg,
 442                                 int msg_len, void *resp, int resp_max_len,
 443                                 int timeout)
 444 {
 445         fw_msg->msg = msg;
 446         fw_msg->msg_len = msg_len;
 447         fw_msg->resp = resp;
 448         fw_msg->resp_max_len = resp_max_len;
 449         fw_msg->timeout = timeout;
 450 }
 451
 452 static int bnxt_re_net_ring_free(struct bnxt_re_dev *rdev,
 453                                  u16 fw_ring_id, int type)
 454 {
 455         struct bnxt_en_dev *en_dev = rdev->en_dev;
 456         struct hwrm_ring_free_input req = {0};
 457         struct hwrm_ring_free_output resp;
 458         struct bnxt_fw_msg fw_msg;
 459         int rc = -EINVAL;
 460
 461         if (!en_dev)
 462                 return rc;
 463
 464         if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags))
 465                 return 0;
 466
 467         memset(&fw_msg, 0, sizeof(fw_msg));
 468
 469         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_RING_FREE, -1, -1);
 470         req.ring_type = type;
 471         req.ring_id = cpu_to_le16(fw_ring_id);
 472         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 473                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 474         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 475         if (rc)
 476                 ibdev_err(&rdev->ibdev, "Failed to free HW ring:%d :%#x",
 477                           req.ring_id, rc);
 478         return rc;
 479 }
 480
 481 static int bnxt_re_net_ring_alloc(struct bnxt_re_dev *rdev,
 482                                   struct bnxt_re_ring_attr *ring_attr,
 483                                   u16 *fw_ring_id)
 484 {
 485         struct bnxt_en_dev *en_dev = rdev->en_dev;
 486         struct hwrm_ring_alloc_input req = {0};
 487         struct hwrm_ring_alloc_output resp;
 488         struct bnxt_fw_msg fw_msg;
 489         int rc = -EINVAL;
 490
 491         if (!en_dev)
 492                 return rc;
 493
 494         memset(&fw_msg, 0, sizeof(fw_msg));
 495         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_RING_ALLOC, -1, -1);
 496         req.enables = 0;
 497         req.page_tbl_addr =  cpu_to_le64(ring_attr->dma_arr[0]);
 498         if (ring_attr->pages > 1) {
 499                 /* Page size is in log2 units */
 500                 req.page_size = BNXT_PAGE_SHIFT;
 501                 req.page_tbl_depth = 1;
 502         }
 503         req.fbo = 0;
 504         /* Association of ring index with doorbell index and MSIX number */
 505         req.logical_id = cpu_to_le16(ring_attr->lrid);
 506         req.length = cpu_to_le32(ring_attr->depth + 1);
 507         req.ring_type = ring_attr->type;
 508         req.int_mode = ring_attr->mode;
 509         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 510                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 511         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 512         if (!rc)
 513                 *fw_ring_id = le16_to_cpu(resp.ring_id);
 514
 515         return rc;
 516 }
 517
 518 static int bnxt_re_net_stats_ctx_free(struct bnxt_re_dev *rdev,
 519                                       u32 fw_stats_ctx_id)
 520 {
 521         struct bnxt_en_dev *en_dev = rdev->en_dev;
 522         struct hwrm_stat_ctx_free_input req = {0};
 523         struct bnxt_fw_msg fw_msg;
 524         int rc = -EINVAL;
 525
 526         if (!en_dev)
 527                 return rc;
 528
 529         if (test_bit(BNXT_RE_FLAG_ERR_DEVICE_DETACHED, &rdev->flags))
 530                 return 0;
 531
 532         memset(&fw_msg, 0, sizeof(fw_msg));
 533
 534         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_STAT_CTX_FREE, -1, -1);
 535         req.stat_ctx_id = cpu_to_le32(fw_stats_ctx_id);
 536         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&req,
 537                             sizeof(req), DFLT_HWRM_CMD_TIMEOUT);
 538         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 539         if (rc)
 540                 ibdev_err(&rdev->ibdev, "Failed to free HW stats context %#x",
 541                           rc);
 542
 543         return rc;
 544 }
 545
 546 static int bnxt_re_net_stats_ctx_alloc(struct bnxt_re_dev *rdev,
 547                                        dma_addr_t dma_map,
 548                                        u32 *fw_stats_ctx_id)
 549 {
 550         struct hwrm_stat_ctx_alloc_output resp = {0};
 551         struct hwrm_stat_ctx_alloc_input req = {0};
 552         struct bnxt_en_dev *en_dev = rdev->en_dev;
 553         struct bnxt_fw_msg fw_msg;
 554         int rc = -EINVAL;
 555
 556         *fw_stats_ctx_id = INVALID_STATS_CTX_ID;
 557
 558         if (!en_dev)
 559                 return rc;
 560
 561         memset(&fw_msg, 0, sizeof(fw_msg));
 562
 563         bnxt_re_init_hwrm_hdr(rdev, (void *)&req, HWRM_STAT_CTX_ALLOC, -1, -1);
 564         req.update_period_ms = cpu_to_le32(1000);
 565         req.stats_dma_addr = cpu_to_le64(dma_map);
 566         req.stats_dma_length = cpu_to_le16(sizeof(struct ctx_hw_stats_ext));
 567         req.stat_ctx_flags = STAT_CTX_ALLOC_REQ_STAT_CTX_FLAGS_ROCE;
 568         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
 569                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
 570         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
 571         if (!rc)
 572                 *fw_stats_ctx_id = le32_to_cpu(resp.stat_ctx_id);
 573
 574         return rc;
 575 }
 576
 577 /* Device */
 578
 579 static bool is_bnxt_re_dev(struct net_device *netdev)
 580 {
 581         struct ethtool_drvinfo drvinfo;
 582
 583         if (netdev->ethtool_ops && netdev->ethtool_ops->get_drvinfo) {
 584                 memset(&drvinfo, 0, sizeof(drvinfo));
 585                 netdev->ethtool_ops->get_drvinfo(netdev, &drvinfo);
 586
 587                 if (strcmp(drvinfo.driver, "bnxt_en"))
 588                         return false;
 589                 return true;
 590         }
 591         return false;
 592 }
 593
 594 static struct bnxt_re_dev *bnxt_re_from_netdev(struct net_device *netdev)
 595 {
 596         struct ib_device *ibdev =
 597                 ib_device_get_by_netdev(netdev, RDMA_DRIVER_BNXT_RE);
 598         if (!ibdev)
 599                 return NULL;
 600
 601         return container_of(ibdev, struct bnxt_re_dev, ibdev);
 602 }
 603
 604 static void bnxt_re_dev_unprobe(struct net_device *netdev,
 605                                 struct bnxt_en_dev *en_dev)
 606 {
 607         dev_put(netdev);
 608         module_put(en_dev->pdev->driver->driver.owner);
 609 }
 610
 611 static struct bnxt_en_dev *bnxt_re_dev_probe(struct net_device *netdev)
 612 {
 613         struct bnxt *bp = netdev_priv(netdev);
 614         struct bnxt_en_dev *en_dev;
 615         struct pci_dev *pdev;
 616
 617         /* Call bnxt_en's RoCE probe via indirect API */
 618         if (!bp->ulp_probe)
 619                 return ERR_PTR(-EINVAL);
 620
 621         en_dev = bp->ulp_probe(netdev);
 622         if (IS_ERR(en_dev))
 623                 return en_dev;
 624
 625         pdev = en_dev->pdev;
 626         if (!pdev)
 627                 return ERR_PTR(-EINVAL);
 628
 629         if (!(en_dev->flags & BNXT_EN_FLAG_ROCE_CAP)) {
 630                 dev_info(&pdev->dev,
 631                         "%s: probe error: RoCE is not supported on this device",
 632                         ROCE_DRV_MODULE_NAME);
 633                 return ERR_PTR(-ENODEV);
 634         }
 635
 636         /* Bump net device reference count */
 637         if (!try_module_get(pdev->driver->driver.owner))
 638                 return ERR_PTR(-ENODEV);
 639
 640         dev_hold(netdev);
 641
 642         return en_dev;
 643 }
 644
 645 static ssize_t hw_rev_show(struct device *device, struct device_attribute *attr,
 646                            char *buf)
 647 {
 648         struct bnxt_re_dev *rdev =
 649                 rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev);
 650
 651         return sysfs_emit(buf, "0x%x\n", rdev->en_dev->pdev->vendor);
 652 }
 653 static DEVICE_ATTR_RO(hw_rev);
 654
 655 static ssize_t hca_type_show(struct device *device,
 656                              struct device_attribute *attr, char *buf)
 657 {
 658         struct bnxt_re_dev *rdev =
 659                 rdma_device_to_drv_device(device, struct bnxt_re_dev, ibdev);
 660
 661         return sysfs_emit(buf, "%s\n", rdev->ibdev.node_desc);
 662 }
 663 static DEVICE_ATTR_RO(hca_type);
 664
 665 static struct attribute *bnxt_re_attributes[] = {
 666         &dev_attr_hw_rev.attr,
 667         &dev_attr_hca_type.attr,
 668         NULL
 669 };
 670
 671 static const struct attribute_group bnxt_re_dev_attr_group = {
 672         .attrs = bnxt_re_attributes,
 673 };
 674
 675 static const struct ib_device_ops bnxt_re_dev_ops = {
 676         .owner = THIS_MODULE,
 677         .driver_id = RDMA_DRIVER_BNXT_RE,
 678         .uverbs_abi_ver = BNXT_RE_ABI_VERSION,
 679
 680         .add_gid = bnxt_re_add_gid,
 681         .alloc_hw_stats = bnxt_re_ib_alloc_hw_stats,
 682         .alloc_mr = bnxt_re_alloc_mr,
 683         .alloc_pd = bnxt_re_alloc_pd,
 684         .alloc_ucontext = bnxt_re_alloc_ucontext,
 685         .create_ah = bnxt_re_create_ah,
 686         .create_cq = bnxt_re_create_cq,
 687         .create_qp = bnxt_re_create_qp,
 688         .create_srq = bnxt_re_create_srq,
 689         .create_user_ah = bnxt_re_create_ah,
 690         .dealloc_driver = bnxt_re_dealloc_driver,
 691         .dealloc_pd = bnxt_re_dealloc_pd,
 692         .dealloc_ucontext = bnxt_re_dealloc_ucontext,
 693         .del_gid = bnxt_re_del_gid,
 694         .dereg_mr = bnxt_re_dereg_mr,
 695         .destroy_ah = bnxt_re_destroy_ah,
 696         .destroy_cq = bnxt_re_destroy_cq,
 697         .destroy_qp = bnxt_re_destroy_qp,
 698         .destroy_srq = bnxt_re_destroy_srq,
 699         .get_dev_fw_str = bnxt_re_query_fw_str,
 700         .get_dma_mr = bnxt_re_get_dma_mr,
 701         .get_hw_stats = bnxt_re_ib_get_hw_stats,
 702         .get_link_layer = bnxt_re_get_link_layer,
 703         .get_port_immutable = bnxt_re_get_port_immutable,
 704         .map_mr_sg = bnxt_re_map_mr_sg,
 705         .mmap = bnxt_re_mmap,
 706         .modify_ah = bnxt_re_modify_ah,
 707         .modify_qp = bnxt_re_modify_qp,
 708         .modify_srq = bnxt_re_modify_srq,
 709         .poll_cq = bnxt_re_poll_cq,
 710         .post_recv = bnxt_re_post_recv,
 711         .post_send = bnxt_re_post_send,
 712         .post_srq_recv = bnxt_re_post_srq_recv,
 713         .query_ah = bnxt_re_query_ah,
 714         .query_device = bnxt_re_query_device,
 715         .query_pkey = bnxt_re_query_pkey,
 716         .query_port = bnxt_re_query_port,
 717         .query_qp = bnxt_re_query_qp,
 718         .query_srq = bnxt_re_query_srq,
 719         .reg_user_mr = bnxt_re_reg_user_mr,
 720         .req_notify_cq = bnxt_re_req_notify_cq,
 721         INIT_RDMA_OBJ_SIZE(ib_ah, bnxt_re_ah, ib_ah),
 722         INIT_RDMA_OBJ_SIZE(ib_cq, bnxt_re_cq, ib_cq),
 723         INIT_RDMA_OBJ_SIZE(ib_pd, bnxt_re_pd, ib_pd),
 724         INIT_RDMA_OBJ_SIZE(ib_srq, bnxt_re_srq, ib_srq),
 725         INIT_RDMA_OBJ_SIZE(ib_ucontext, bnxt_re_ucontext, ib_uctx),
 726 };
 727
 728 static int bnxt_re_register_ib(struct bnxt_re_dev *rdev)
 729 {
 730         struct ib_device *ibdev = &rdev->ibdev;
 731         int ret;
 732
 733         /* ib device init */
 734         ibdev->node_type = RDMA_NODE_IB_CA;
 735         strlcpy(ibdev->node_desc, BNXT_RE_DESC " HCA",
 736                 strlen(BNXT_RE_DESC) + 5);
 737         ibdev->phys_port_cnt = 1;
 738
 739         bnxt_qplib_get_guid(rdev->netdev->dev_addr, (u8 *)&ibdev->node_guid);
 740
 741         ibdev->num_comp_vectors = rdev->num_msix - 1;
 742         ibdev->dev.parent = &rdev->en_dev->pdev->dev;
 743         ibdev->local_dma_lkey = BNXT_QPLIB_RSVD_LKEY;
 744
 745         rdma_set_device_sysfs_group(ibdev, &bnxt_re_dev_attr_group);
 746         ib_set_device_ops(ibdev, &bnxt_re_dev_ops);
 747         ret = ib_device_set_netdev(&rdev->ibdev, rdev->netdev, 1);
 748         if (ret)
 749                 return ret;
 750
 751         dma_set_max_seg_size(&rdev->en_dev->pdev->dev, UINT_MAX);
 752         return ib_register_device(ibdev, "bnxt_re%d", &rdev->en_dev->pdev->dev);
 753 }
 754
 755 static void bnxt_re_dev_remove(struct bnxt_re_dev *rdev)
 756 {
 757         dev_put(rdev->netdev);
 758         rdev->netdev = NULL;
 759         mutex_lock(&bnxt_re_dev_lock);
 760         list_del_rcu(&rdev->list);
 761         mutex_unlock(&bnxt_re_dev_lock);
 762
 763         synchronize_rcu();
 764 }
 765
 766 static struct bnxt_re_dev *bnxt_re_dev_add(struct net_device *netdev,
 767                                            struct bnxt_en_dev *en_dev)
 768 {
 769         struct bnxt_re_dev *rdev;
 770
 771         /* Allocate bnxt_re_dev instance here */
 772         rdev = ib_alloc_device(bnxt_re_dev, ibdev);
 773         if (!rdev) {
 774                 ibdev_err(NULL, "%s: bnxt_re_dev allocation failure!",
 775                           ROCE_DRV_MODULE_NAME);
 776                 return NULL;
 777         }
 778         /* Default values */
 779         rdev->netdev = netdev;
 780         dev_hold(rdev->netdev);
 781         rdev->en_dev = en_dev;
 782         rdev->id = rdev->en_dev->pdev->devfn;
 783         INIT_LIST_HEAD(&rdev->qp_list);
 784         mutex_init(&rdev->qp_lock);
 785         atomic_set(&rdev->qp_count, 0);
 786         atomic_set(&rdev->cq_count, 0);
 787         atomic_set(&rdev->srq_count, 0);
 788         atomic_set(&rdev->mr_count, 0);
 789         atomic_set(&rdev->mw_count, 0);
 790         rdev->cosq[0] = 0xFFFF;
 791         rdev->cosq[1] = 0xFFFF;
 792
 793         mutex_lock(&bnxt_re_dev_lock);
 794         list_add_tail_rcu(&rdev->list, &bnxt_re_dev_list);
 795         mutex_unlock(&bnxt_re_dev_lock);
 796         return rdev;
 797 }
 798
 799 static int bnxt_re_handle_unaffi_async_event(struct creq_func_event
 800                                              *unaffi_async)
 801 {
 802         switch (unaffi_async->event) {
 803         case CREQ_FUNC_EVENT_EVENT_TX_WQE_ERROR:
 804                 break;
 805         case CREQ_FUNC_EVENT_EVENT_TX_DATA_ERROR:
 806                 break;
 807         case CREQ_FUNC_EVENT_EVENT_RX_WQE_ERROR:
 808                 break;
 809         case CREQ_FUNC_EVENT_EVENT_RX_DATA_ERROR:
 810                 break;
 811         case CREQ_FUNC_EVENT_EVENT_CQ_ERROR:
 812                 break;
 813         case CREQ_FUNC_EVENT_EVENT_TQM_ERROR:
 814                 break;
 815         case CREQ_FUNC_EVENT_EVENT_CFCQ_ERROR:
 816                 break;
 817         case CREQ_FUNC_EVENT_EVENT_CFCS_ERROR:
 818                 break;
 819         case CREQ_FUNC_EVENT_EVENT_CFCC_ERROR:
 820                 break;
 821         case CREQ_FUNC_EVENT_EVENT_CFCM_ERROR:
 822                 break;
 823         case CREQ_FUNC_EVENT_EVENT_TIM_ERROR:
 824                 break;
 825         default:
 826                 return -EINVAL;
 827         }
 828         return 0;
 829 }
 830
 831 static int bnxt_re_handle_qp_async_event(struct creq_qp_event *qp_event,
 832                                          struct bnxt_re_qp *qp)
 833 {
 834         struct ib_event event;
 835         unsigned int flags;
 836
 837         if (qp->qplib_qp.state == CMDQ_MODIFY_QP_NEW_STATE_ERR &&
 838             rdma_is_kernel_res(&qp->ib_qp.res)) {
 839                 flags = bnxt_re_lock_cqs(qp);
 840                 bnxt_qplib_add_flush_qp(&qp->qplib_qp);
 841                 bnxt_re_unlock_cqs(qp, flags);
 842         }
 843
 844         memset(&event, 0, sizeof(event));
 845         if (qp->qplib_qp.srq) {
 846                 event.device = &qp->rdev->ibdev;
 847                 event.element.qp = &qp->ib_qp;
 848                 event.event = IB_EVENT_QP_LAST_WQE_REACHED;
 849         }
 850
 851         if (event.device && qp->ib_qp.event_handler)
 852                 qp->ib_qp.event_handler(&event, qp->ib_qp.qp_context);
 853
 854         return 0;
 855 }
 856
 857 static int bnxt_re_handle_affi_async_event(struct creq_qp_event *affi_async,
 858                                            void *obj)
 859 {
 860         int rc = 0;
 861         u8 event;
 862
 863         if (!obj)
 864                 return rc; /* QP was already dead, still return success */
 865
 866         event = affi_async->event;
 867         if (event == CREQ_QP_EVENT_EVENT_QP_ERROR_NOTIFICATION) {
 868                 struct bnxt_qplib_qp *lib_qp = obj;
 869                 struct bnxt_re_qp *qp = container_of(lib_qp, struct bnxt_re_qp,
 870                                                      qplib_qp);
 871                 rc = bnxt_re_handle_qp_async_event(affi_async, qp);
 872         }
 873         return rc;
 874 }
 875
 876 static int bnxt_re_aeq_handler(struct bnxt_qplib_rcfw *rcfw,
 877                                void *aeqe, void *obj)
 878 {
 879         struct creq_qp_event *affi_async;
 880         struct creq_func_event *unaffi_async;
 881         u8 type;
 882         int rc;
 883
 884         type = ((struct creq_base *)aeqe)->type;
 885         if (type == CREQ_BASE_TYPE_FUNC_EVENT) {
 886                 unaffi_async = aeqe;
 887                 rc = bnxt_re_handle_unaffi_async_event(unaffi_async);
 888         } else {
 889                 affi_async = aeqe;
 890                 rc = bnxt_re_handle_affi_async_event(affi_async, obj);
 891         }
 892
 893         return rc;
 894 }
 895
 896 static int bnxt_re_srqn_handler(struct bnxt_qplib_nq *nq,
 897                                 struct bnxt_qplib_srq *handle, u8 event)
 898 {
 899         struct bnxt_re_srq *srq = container_of(handle, struct bnxt_re_srq,
 900                                                qplib_srq);
 901         struct ib_event ib_event;
 902         int rc = 0;
 903
 904         if (!srq) {
 905                 ibdev_err(NULL, "%s: SRQ is NULL, SRQN not handled",
 906                           ROCE_DRV_MODULE_NAME);
 907                 rc = -EINVAL;
 908                 goto done;
 909         }
 910         ib_event.device = &srq->rdev->ibdev;
 911         ib_event.element.srq = &srq->ib_srq;
 912         if (event == NQ_SRQ_EVENT_EVENT_SRQ_THRESHOLD_EVENT)
 913                 ib_event.event = IB_EVENT_SRQ_LIMIT_REACHED;
 914         else
 915                 ib_event.event = IB_EVENT_SRQ_ERR;
 916
 917         if (srq->ib_srq.event_handler) {
 918                 /* Lock event_handler? */
 919                 (*srq->ib_srq.event_handler)(&ib_event,
 920                                              srq->ib_srq.srq_context);
 921         }
 922 done:
 923         return rc;
 924 }
 925
 926 static int bnxt_re_cqn_handler(struct bnxt_qplib_nq *nq,
 927                                struct bnxt_qplib_cq *handle)
 928 {
 929         struct bnxt_re_cq *cq = container_of(handle, struct bnxt_re_cq,
 930                                              qplib_cq);
 931
 932         if (!cq) {
 933                 ibdev_err(NULL, "%s: CQ is NULL, CQN not handled",
 934                           ROCE_DRV_MODULE_NAME);
 935                 return -EINVAL;
 936         }
 937         if (cq->ib_cq.comp_handler) {
 938                 /* Lock comp_handler? */
 939                 (*cq->ib_cq.comp_handler)(&cq->ib_cq, cq->ib_cq.cq_context);
 940         }
 941
 942         return 0;
 943 }
 944
 945 #define BNXT_RE_GEN_P5_PF_NQ_DB         0x10000
 946 #define BNXT_RE_GEN_P5_VF_NQ_DB         0x4000
 947 static u32 bnxt_re_get_nqdb_offset(struct bnxt_re_dev *rdev, u16 indx)
 948 {
 949         return bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx) ?
 950                 (rdev->is_virtfn ? BNXT_RE_GEN_P5_VF_NQ_DB :
 951                                    BNXT_RE_GEN_P5_PF_NQ_DB) :
 952                                    rdev->msix_entries[indx].db_offset;
 953 }
 954
 955 static void bnxt_re_cleanup_res(struct bnxt_re_dev *rdev)
 956 {
 957         int i;
 958
 959         for (i = 1; i < rdev->num_msix; i++)
 960                 bnxt_qplib_disable_nq(&rdev->nq[i - 1]);
 961
 962         if (rdev->qplib_res.rcfw)
 963                 bnxt_qplib_cleanup_res(&rdev->qplib_res);
 964 }
 965
 966 static int bnxt_re_init_res(struct bnxt_re_dev *rdev)
 967 {
 968         int num_vec_enabled = 0;
 969         int rc = 0, i;
 970         u32 db_offt;
 971
 972         bnxt_qplib_init_res(&rdev->qplib_res);
 973
 974         for (i = 1; i < rdev->num_msix ; i++) {
 975                 db_offt = bnxt_re_get_nqdb_offset(rdev, i);
 976                 rc = bnxt_qplib_enable_nq(rdev->en_dev->pdev, &rdev->nq[i - 1],
 977                                           i - 1, rdev->msix_entries[i].vector,
 978                                           db_offt, &bnxt_re_cqn_handler,
 979                                           &bnxt_re_srqn_handler);
 980                 if (rc) {
 981                         ibdev_err(&rdev->ibdev,
 982                                   "Failed to enable NQ with rc = 0x%x", rc);
 983                         goto fail;
 984                 }
 985                 num_vec_enabled++;
 986         }
 987         return 0;
 988 fail:
 989         for (i = num_vec_enabled; i >= 0; i--)
 990                 bnxt_qplib_disable_nq(&rdev->nq[i]);
 991         return rc;
 992 }
 993
 994 static void bnxt_re_free_nq_res(struct bnxt_re_dev *rdev)
 995 {
 996         u8 type;
 997         int i;
 998
 999         for (i = 0; i < rdev->num_msix - 1; i++) {
1000                 type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1001                 bnxt_re_net_ring_free(rdev, rdev->nq[i].ring_id, type);
1002                 bnxt_qplib_free_nq(&rdev->nq[i]);
1003                 rdev->nq[i].res = NULL;
1004         }
1005 }
1006
1007 static void bnxt_re_free_res(struct bnxt_re_dev *rdev)
1008 {
1009         bnxt_re_free_nq_res(rdev);
1010
1011         if (rdev->qplib_res.dpi_tbl.max) {
1012                 bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
1013                                        &rdev->qplib_res.dpi_tbl,
1014                                        &rdev->dpi_privileged);
1015         }
1016         if (rdev->qplib_res.rcfw) {
1017                 bnxt_qplib_free_res(&rdev->qplib_res);
1018                 rdev->qplib_res.rcfw = NULL;
1019         }
1020 }
1021
1022 static int bnxt_re_alloc_res(struct bnxt_re_dev *rdev)
1023 {
1024         struct bnxt_re_ring_attr rattr = {};
1025         int num_vec_created = 0;
1026         int rc = 0, i;
1027         u8 type;
1028
1029         /* Configure and allocate resources for qplib */
1030         rdev->qplib_res.rcfw = &rdev->rcfw;
1031         rc = bnxt_qplib_get_dev_attr(&rdev->rcfw, &rdev->dev_attr,
1032                                      rdev->is_virtfn);
1033         if (rc)
1034                 goto fail;
1035
1036         rc = bnxt_qplib_alloc_res(&rdev->qplib_res, rdev->en_dev->pdev,
1037                                   rdev->netdev, &rdev->dev_attr);
1038         if (rc)
1039                 goto fail;
1040
1041         rc = bnxt_qplib_alloc_dpi(&rdev->qplib_res.dpi_tbl,
1042                                   &rdev->dpi_privileged,
1043                                   rdev);
1044         if (rc)
1045                 goto dealloc_res;
1046
1047         for (i = 0; i < rdev->num_msix - 1; i++) {
1048                 struct bnxt_qplib_nq *nq;
1049
1050                 nq = &rdev->nq[i];
1051                 nq->hwq.max_elements = BNXT_QPLIB_NQE_MAX_CNT;
1052                 rc = bnxt_qplib_alloc_nq(&rdev->qplib_res, &rdev->nq[i]);
1053                 if (rc) {
1054                         ibdev_err(&rdev->ibdev, "Alloc Failed NQ%d rc:%#x",
1055                                   i, rc);
1056                         goto free_nq;
1057                 }
1058                 type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1059                 rattr.dma_arr = nq->hwq.pbl[PBL_LVL_0].pg_map_arr;
1060                 rattr.pages = nq->hwq.pbl[rdev->nq[i].hwq.level].pg_count;
1061                 rattr.type = type;
1062                 rattr.mode = RING_ALLOC_REQ_INT_MODE_MSIX;
1063                 rattr.depth = BNXT_QPLIB_NQE_MAX_CNT - 1;
1064                 rattr.lrid = rdev->msix_entries[i + 1].ring_idx;
1065                 rc = bnxt_re_net_ring_alloc(rdev, &rattr, &nq->ring_id);
1066                 if (rc) {
1067                         ibdev_err(&rdev->ibdev,
1068                                   "Failed to allocate NQ fw id with rc = 0x%x",
1069                                   rc);
1070                         bnxt_qplib_free_nq(&rdev->nq[i]);
1071                         goto free_nq;
1072                 }
1073                 num_vec_created++;
1074         }
1075         return 0;
1076 free_nq:
1077         for (i = num_vec_created - 1; i >= 0; i--) {
1078                 type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1079                 bnxt_re_net_ring_free(rdev, rdev->nq[i].ring_id, type);
1080                 bnxt_qplib_free_nq(&rdev->nq[i]);
1081         }
1082         bnxt_qplib_dealloc_dpi(&rdev->qplib_res,
1083                                &rdev->qplib_res.dpi_tbl,
1084                                &rdev->dpi_privileged);
1085 dealloc_res:
1086         bnxt_qplib_free_res(&rdev->qplib_res);
1087
1088 fail:
1089         rdev->qplib_res.rcfw = NULL;
1090         return rc;
1091 }
1092
1093 static void bnxt_re_dispatch_event(struct ib_device *ibdev, struct ib_qp *qp,
1094                                    u8 port_num, enum ib_event_type event)
1095 {
1096         struct ib_event ib_event;
1097
1098         ib_event.device = ibdev;
1099         if (qp) {
1100                 ib_event.element.qp = qp;
1101                 ib_event.event = event;
1102                 if (qp->event_handler)
1103                         qp->event_handler(&ib_event, qp->qp_context);
1104
1105         } else {
1106                 ib_event.element.port_num = port_num;
1107                 ib_event.event = event;
1108                 ib_dispatch_event(&ib_event);
1109         }
1110 }
1111
1112 #define HWRM_QUEUE_PRI2COS_QCFG_INPUT_FLAGS_IVLAN      0x02
1113 static int bnxt_re_query_hwrm_pri2cos(struct bnxt_re_dev *rdev, u8 dir,
1114                                       u64 *cid_map)
1115 {
1116         struct hwrm_queue_pri2cos_qcfg_input req = {0};
1117         struct bnxt *bp = netdev_priv(rdev->netdev);
1118         struct hwrm_queue_pri2cos_qcfg_output resp;
1119         struct bnxt_en_dev *en_dev = rdev->en_dev;
1120         struct bnxt_fw_msg fw_msg;
1121         u32 flags = 0;
1122         u8 *qcfgmap, *tmp_map;
1123         int rc = 0, i;
1124
1125         if (!cid_map)
1126                 return -EINVAL;
1127
1128         memset(&fw_msg, 0, sizeof(fw_msg));
1129         bnxt_re_init_hwrm_hdr(rdev, (void *)&req,
1130                               HWRM_QUEUE_PRI2COS_QCFG, -1, -1);
1131         flags |= (dir & 0x01);
1132         flags |= HWRM_QUEUE_PRI2COS_QCFG_INPUT_FLAGS_IVLAN;
1133         req.flags = cpu_to_le32(flags);
1134         req.port_id = bp->pf.port_id;
1135
1136         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
1137                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
1138         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
1139         if (rc)
1140                 return rc;
1141
1142         if (resp.queue_cfg_info) {
1143                 ibdev_warn(&rdev->ibdev,
1144                            "Asymmetric cos queue configuration detected");
1145                 ibdev_warn(&rdev->ibdev,
1146                            " on device, QoS may not be fully functional\n");
1147         }
1148         qcfgmap = &resp.pri0_cos_queue_id;
1149         tmp_map = (u8 *)cid_map;
1150         for (i = 0; i < IEEE_8021QAZ_MAX_TCS; i++)
1151                 tmp_map[i] = qcfgmap[i];
1152
1153         return rc;
1154 }
1155
1156 static bool bnxt_re_is_qp1_or_shadow_qp(struct bnxt_re_dev *rdev,
1157                                         struct bnxt_re_qp *qp)
1158 {
1159         return (qp->ib_qp.qp_type == IB_QPT_GSI) ||
1160                (qp == rdev->gsi_ctx.gsi_sqp);
1161 }
1162
1163 static void bnxt_re_dev_stop(struct bnxt_re_dev *rdev)
1164 {
1165         int mask = IB_QP_STATE;
1166         struct ib_qp_attr qp_attr;
1167         struct bnxt_re_qp *qp;
1168
1169         qp_attr.qp_state = IB_QPS_ERR;
1170         mutex_lock(&rdev->qp_lock);
1171         list_for_each_entry(qp, &rdev->qp_list, list) {
1172                 /* Modify the state of all QPs except QP1/Shadow QP */
1173                 if (!bnxt_re_is_qp1_or_shadow_qp(rdev, qp)) {
1174                         if (qp->qplib_qp.state !=
1175                             CMDQ_MODIFY_QP_NEW_STATE_RESET &&
1176                             qp->qplib_qp.state !=
1177                             CMDQ_MODIFY_QP_NEW_STATE_ERR) {
1178                                 bnxt_re_dispatch_event(&rdev->ibdev, &qp->ib_qp,
1179                                                        1, IB_EVENT_QP_FATAL);
1180                                 bnxt_re_modify_qp(&qp->ib_qp, &qp_attr, mask,
1181                                                   NULL);
1182                         }
1183                 }
1184         }
1185         mutex_unlock(&rdev->qp_lock);
1186 }
1187
1188 static int bnxt_re_update_gid(struct bnxt_re_dev *rdev)
1189 {
1190         struct bnxt_qplib_sgid_tbl *sgid_tbl = &rdev->qplib_res.sgid_tbl;
1191         struct bnxt_qplib_gid gid;
1192         u16 gid_idx, index;
1193         int rc = 0;
1194
1195         if (!ib_device_try_get(&rdev->ibdev))
1196                 return 0;
1197
1198         if (!sgid_tbl) {
1199                 ibdev_err(&rdev->ibdev, "QPLIB: SGID table not allocated");
1200                 rc = -EINVAL;
1201                 goto out;
1202         }
1203
1204         for (index = 0; index < sgid_tbl->active; index++) {
1205                 gid_idx = sgid_tbl->hw_id[index];
1206
1207                 if (!memcmp(&sgid_tbl->tbl[index], &bnxt_qplib_gid_zero,
1208                             sizeof(bnxt_qplib_gid_zero)))
1209                         continue;
1210                 /* need to modify the VLAN enable setting of non VLAN GID only
1211                  * as setting is done for VLAN GID while adding GID
1212                  */
1213                 if (sgid_tbl->vlan[index])
1214                         continue;
1215
1216                 memcpy(&gid, &sgid_tbl->tbl[index], sizeof(gid));
1217
1218                 rc = bnxt_qplib_update_sgid(sgid_tbl, &gid, gid_idx,
1219                                             rdev->qplib_res.netdev->dev_addr);
1220         }
1221 out:
1222         ib_device_put(&rdev->ibdev);
1223         return rc;
1224 }
1225
1226 static u32 bnxt_re_get_priority_mask(struct bnxt_re_dev *rdev)
1227 {
1228         u32 prio_map = 0, tmp_map = 0;
1229         struct net_device *netdev;
1230         struct dcb_app app;
1231
1232         netdev = rdev->netdev;
1233
1234         memset(&app, 0, sizeof(app));
1235         app.selector = IEEE_8021QAZ_APP_SEL_ETHERTYPE;
1236         app.protocol = ETH_P_IBOE;
1237         tmp_map = dcb_ieee_getapp_mask(netdev, &app);
1238         prio_map = tmp_map;
1239
1240         app.selector = IEEE_8021QAZ_APP_SEL_DGRAM;
1241         app.protocol = ROCE_V2_UDP_DPORT;
1242         tmp_map = dcb_ieee_getapp_mask(netdev, &app);
1243         prio_map |= tmp_map;
1244
1245         return prio_map;
1246 }
1247
1248 static void bnxt_re_parse_cid_map(u8 prio_map, u8 *cid_map, u16 *cosq)
1249 {
1250         u16 prio;
1251         u8 id;
1252
1253         for (prio = 0, id = 0; prio < 8; prio++) {
1254                 if (prio_map & (1 << prio)) {
1255                         cosq[id] = cid_map[prio];
1256                         id++;
1257                         if (id == 2) /* Max 2 tcs supported */
1258                                 break;
1259                 }
1260         }
1261 }
1262
1263 static int bnxt_re_setup_qos(struct bnxt_re_dev *rdev)
1264 {
1265         u8 prio_map = 0;
1266         u64 cid_map;
1267         int rc;
1268
1269         /* Get priority for roce */
1270         prio_map = bnxt_re_get_priority_mask(rdev);
1271
1272         if (prio_map == rdev->cur_prio_map)
1273                 return 0;
1274         rdev->cur_prio_map = prio_map;
1275         /* Get cosq id for this priority */
1276         rc = bnxt_re_query_hwrm_pri2cos(rdev, 0, &cid_map);
1277         if (rc) {
1278                 ibdev_warn(&rdev->ibdev, "no cos for p_mask %x\n", prio_map);
1279                 return rc;
1280         }
1281         /* Parse CoS IDs for app priority */
1282         bnxt_re_parse_cid_map(prio_map, (u8 *)&cid_map, rdev->cosq);
1283
1284         /* Config BONO. */
1285         rc = bnxt_qplib_map_tc2cos(&rdev->qplib_res, rdev->cosq);
1286         if (rc) {
1287                 ibdev_warn(&rdev->ibdev, "no tc for cos{%x, %x}\n",
1288                            rdev->cosq[0], rdev->cosq[1]);
1289                 return rc;
1290         }
1291
1292         /* Actual priorities are not programmed as they are already
1293          * done by L2 driver; just enable or disable priority vlan tagging
1294          */
1295         if ((prio_map == 0 && rdev->qplib_res.prio) ||
1296             (prio_map != 0 && !rdev->qplib_res.prio)) {
1297                 rdev->qplib_res.prio = prio_map ? true : false;
1298
1299                 bnxt_re_update_gid(rdev);
1300         }
1301
1302         return 0;
1303 }
1304
1305 static void bnxt_re_query_hwrm_intf_version(struct bnxt_re_dev *rdev)
1306 {
1307         struct bnxt_en_dev *en_dev = rdev->en_dev;
1308         struct hwrm_ver_get_output resp = {0};
1309         struct hwrm_ver_get_input req = {0};
1310         struct bnxt_fw_msg fw_msg;
1311         int rc = 0;
1312
1313         memset(&fw_msg, 0, sizeof(fw_msg));
1314         bnxt_re_init_hwrm_hdr(rdev, (void *)&req,
1315                               HWRM_VER_GET, -1, -1);
1316         req.hwrm_intf_maj = HWRM_VERSION_MAJOR;
1317         req.hwrm_intf_min = HWRM_VERSION_MINOR;
1318         req.hwrm_intf_upd = HWRM_VERSION_UPDATE;
1319         bnxt_re_fill_fw_msg(&fw_msg, (void *)&req, sizeof(req), (void *)&resp,
1320                             sizeof(resp), DFLT_HWRM_CMD_TIMEOUT);
1321         rc = en_dev->en_ops->bnxt_send_fw_msg(en_dev, BNXT_ROCE_ULP, &fw_msg);
1322         if (rc) {
1323                 ibdev_err(&rdev->ibdev, "Failed to query HW version, rc = 0x%x",
1324                           rc);
1325                 return;
1326         }
1327         rdev->qplib_ctx.hwrm_intf_ver =
1328                 (u64)le16_to_cpu(resp.hwrm_intf_major) << 48 |
1329                 (u64)le16_to_cpu(resp.hwrm_intf_minor) << 32 |
1330                 (u64)le16_to_cpu(resp.hwrm_intf_build) << 16 |
1331                 le16_to_cpu(resp.hwrm_intf_patch);
1332 }
1333
1334 static int bnxt_re_ib_init(struct bnxt_re_dev *rdev)
1335 {
1336         int rc = 0;
1337         u32 event;
1338
1339         /* Register ib dev */
1340         rc = bnxt_re_register_ib(rdev);
1341         if (rc) {
1342                 pr_err("Failed to register with IB: %#x\n", rc);
1343                 return rc;
1344         }
1345         dev_info(rdev_to_dev(rdev), "Device registered successfully");
1346         ib_get_eth_speed(&rdev->ibdev, 1, &rdev->active_speed,
1347                          &rdev->active_width);
1348         set_bit(BNXT_RE_FLAG_ISSUE_ROCE_STATS, &rdev->flags);
1349
1350         event = netif_running(rdev->netdev) && netif_carrier_ok(rdev->netdev) ?
1351                 IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
1352
1353         bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1, event);
1354
1355         return rc;
1356 }
1357
1358 static void bnxt_re_dev_uninit(struct bnxt_re_dev *rdev)
1359 {
1360         u8 type;
1361         int rc;
1362
1363         if (test_and_clear_bit(BNXT_RE_FLAG_QOS_WORK_REG, &rdev->flags))
1364                 cancel_delayed_work_sync(&rdev->worker);
1365
1366         if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED,
1367                                &rdev->flags))
1368                 bnxt_re_cleanup_res(rdev);
1369         if (test_and_clear_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, &rdev->flags))
1370                 bnxt_re_free_res(rdev);
1371
1372         if (test_and_clear_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, &rdev->flags)) {
1373                 rc = bnxt_qplib_deinit_rcfw(&rdev->rcfw);
1374                 if (rc)
1375                         ibdev_warn(&rdev->ibdev,
1376                                    "Failed to deinitialize RCFW: %#x", rc);
1377                 bnxt_re_net_stats_ctx_free(rdev, rdev->qplib_ctx.stats.fw_id);
1378                 bnxt_qplib_free_ctx(&rdev->qplib_res, &rdev->qplib_ctx);
1379                 bnxt_qplib_disable_rcfw_channel(&rdev->rcfw);
1380                 type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1381                 bnxt_re_net_ring_free(rdev, rdev->rcfw.creq.ring_id, type);
1382                 bnxt_qplib_free_rcfw_channel(&rdev->rcfw);
1383         }
1384         if (test_and_clear_bit(BNXT_RE_FLAG_GOT_MSIX, &rdev->flags)) {
1385                 rc = bnxt_re_free_msix(rdev);
1386                 if (rc)
1387                         ibdev_warn(&rdev->ibdev,
1388                                    "Failed to free MSI-X vectors: %#x", rc);
1389         }
1390
1391         bnxt_re_destroy_chip_ctx(rdev);
1392         if (test_and_clear_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, &rdev->flags)) {
1393                 rc = bnxt_re_unregister_netdev(rdev);
1394                 if (rc)
1395                         ibdev_warn(&rdev->ibdev,
1396                                    "Failed to unregister with netdev: %#x", rc);
1397         }
1398 }
1399
1400 /* worker thread for polling periodic events. Now used for QoS programming*/
1401 static void bnxt_re_worker(struct work_struct *work)
1402 {
1403         struct bnxt_re_dev *rdev = container_of(work, struct bnxt_re_dev,
1404                                                 worker.work);
1405
1406         bnxt_re_setup_qos(rdev);
1407         schedule_delayed_work(&rdev->worker, msecs_to_jiffies(30000));
1408 }
1409
1410 static int bnxt_re_dev_init(struct bnxt_re_dev *rdev, u8 wqe_mode)
1411 {
1412         struct bnxt_qplib_creq_ctx *creq;
1413         struct bnxt_re_ring_attr rattr;
1414         u32 db_offt;
1415         int vid;
1416         u8 type;
1417         int rc;
1418
1419         /* Registered a new RoCE device instance to netdev */
1420         memset(&rattr, 0, sizeof(rattr));
1421         rc = bnxt_re_register_netdev(rdev);
1422         if (rc) {
1423                 rtnl_unlock();
1424                 ibdev_err(&rdev->ibdev,
1425                           "Failed to register with netedev: %#x\n", rc);
1426                 return -EINVAL;
1427         }
1428         set_bit(BNXT_RE_FLAG_NETDEV_REGISTERED, &rdev->flags);
1429
1430         rc = bnxt_re_setup_chip_ctx(rdev, wqe_mode);
1431         if (rc) {
1432                 ibdev_err(&rdev->ibdev, "Failed to get chip context\n");
1433                 return -EINVAL;
1434         }
1435
1436         /* Check whether VF or PF */
1437         bnxt_re_get_sriov_func_type(rdev);
1438
1439         rc = bnxt_re_request_msix(rdev);
1440         if (rc) {
1441                 ibdev_err(&rdev->ibdev,
1442                           "Failed to get MSI-X vectors: %#x\n", rc);
1443                 rc = -EINVAL;
1444                 goto fail;
1445         }
1446         set_bit(BNXT_RE_FLAG_GOT_MSIX, &rdev->flags);
1447
1448         bnxt_re_query_hwrm_intf_version(rdev);
1449
1450         /* Establish RCFW Communication Channel to initialize the context
1451          * memory for the function and all child VFs
1452          */
1453         rc = bnxt_qplib_alloc_rcfw_channel(&rdev->qplib_res, &rdev->rcfw,
1454                                            &rdev->qplib_ctx,
1455                                            BNXT_RE_MAX_QPC_COUNT);
1456         if (rc) {
1457                 ibdev_err(&rdev->ibdev,
1458                           "Failed to allocate RCFW Channel: %#x\n", rc);
1459                 goto fail;
1460         }
1461
1462         type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1463         creq = &rdev->rcfw.creq;
1464         rattr.dma_arr = creq->hwq.pbl[PBL_LVL_0].pg_map_arr;
1465         rattr.pages = creq->hwq.pbl[creq->hwq.level].pg_count;
1466         rattr.type = type;
1467         rattr.mode = RING_ALLOC_REQ_INT_MODE_MSIX;
1468         rattr.depth = BNXT_QPLIB_CREQE_MAX_CNT - 1;
1469         rattr.lrid = rdev->msix_entries[BNXT_RE_AEQ_IDX].ring_idx;
1470         rc = bnxt_re_net_ring_alloc(rdev, &rattr, &creq->ring_id);
1471         if (rc) {
1472                 ibdev_err(&rdev->ibdev, "Failed to allocate CREQ: %#x\n", rc);
1473                 goto free_rcfw;
1474         }
1475         db_offt = bnxt_re_get_nqdb_offset(rdev, BNXT_RE_AEQ_IDX);
1476         vid = rdev->msix_entries[BNXT_RE_AEQ_IDX].vector;
1477         rc = bnxt_qplib_enable_rcfw_channel(&rdev->rcfw,
1478                                             vid, db_offt, rdev->is_virtfn,
1479                                             &bnxt_re_aeq_handler);
1480         if (rc) {
1481                 ibdev_err(&rdev->ibdev, "Failed to enable RCFW channel: %#x\n",
1482                           rc);
1483                 goto free_ring;
1484         }
1485
1486         rc = bnxt_qplib_get_dev_attr(&rdev->rcfw, &rdev->dev_attr,
1487                                      rdev->is_virtfn);
1488         if (rc)
1489                 goto disable_rcfw;
1490
1491         bnxt_re_set_resource_limits(rdev);
1492
1493         rc = bnxt_qplib_alloc_ctx(&rdev->qplib_res, &rdev->qplib_ctx, 0,
1494                                   bnxt_qplib_is_chip_gen_p5(rdev->chip_ctx));
1495         if (rc) {
1496                 ibdev_err(&rdev->ibdev,
1497                           "Failed to allocate QPLIB context: %#x\n", rc);
1498                 goto disable_rcfw;
1499         }
1500         rc = bnxt_re_net_stats_ctx_alloc(rdev,
1501                                          rdev->qplib_ctx.stats.dma_map,
1502                                          &rdev->qplib_ctx.stats.fw_id);
1503         if (rc) {
1504                 ibdev_err(&rdev->ibdev,
1505                           "Failed to allocate stats context: %#x\n", rc);
1506                 goto free_ctx;
1507         }
1508
1509         rc = bnxt_qplib_init_rcfw(&rdev->rcfw, &rdev->qplib_ctx,
1510                                   rdev->is_virtfn);
1511         if (rc) {
1512                 ibdev_err(&rdev->ibdev,
1513                           "Failed to initialize RCFW: %#x\n", rc);
1514                 goto free_sctx;
1515         }
1516         set_bit(BNXT_RE_FLAG_RCFW_CHANNEL_EN, &rdev->flags);
1517
1518         /* Resources based on the 'new' device caps */
1519         rc = bnxt_re_alloc_res(rdev);
1520         if (rc) {
1521                 ibdev_err(&rdev->ibdev,
1522                           "Failed to allocate resources: %#x\n", rc);
1523                 goto fail;
1524         }
1525         set_bit(BNXT_RE_FLAG_RESOURCES_ALLOCATED, &rdev->flags);
1526         rc = bnxt_re_init_res(rdev);
1527         if (rc) {
1528                 ibdev_err(&rdev->ibdev,
1529                           "Failed to initialize resources: %#x\n", rc);
1530                 goto fail;
1531         }
1532
1533         set_bit(BNXT_RE_FLAG_RESOURCES_INITIALIZED, &rdev->flags);
1534
1535         if (!rdev->is_virtfn) {
1536                 rc = bnxt_re_setup_qos(rdev);
1537                 if (rc)
1538                         ibdev_info(&rdev->ibdev,
1539                                    "RoCE priority not yet configured\n");
1540
1541                 INIT_DELAYED_WORK(&rdev->worker, bnxt_re_worker);
1542                 set_bit(BNXT_RE_FLAG_QOS_WORK_REG, &rdev->flags);
1543                 schedule_delayed_work(&rdev->worker, msecs_to_jiffies(30000));
1544         }
1545
1546         return 0;
1547 free_sctx:
1548         bnxt_re_net_stats_ctx_free(rdev, rdev->qplib_ctx.stats.fw_id);
1549 free_ctx:
1550         bnxt_qplib_free_ctx(&rdev->qplib_res, &rdev->qplib_ctx);
1551 disable_rcfw:
1552         bnxt_qplib_disable_rcfw_channel(&rdev->rcfw);
1553 free_ring:
1554         type = bnxt_qplib_get_ring_type(rdev->chip_ctx);
1555         bnxt_re_net_ring_free(rdev, rdev->rcfw.creq.ring_id, type);
1556 free_rcfw:
1557         bnxt_qplib_free_rcfw_channel(&rdev->rcfw);
1558 fail:
1559         bnxt_re_dev_uninit(rdev);
1560
1561         return rc;
1562 }
1563
1564 static void bnxt_re_dev_unreg(struct bnxt_re_dev *rdev)
1565 {
1566         struct bnxt_en_dev *en_dev = rdev->en_dev;
1567         struct net_device *netdev = rdev->netdev;
1568
1569         bnxt_re_dev_remove(rdev);
1570
1571         if (netdev)
1572                 bnxt_re_dev_unprobe(netdev, en_dev);
1573 }
1574
1575 static int bnxt_re_dev_reg(struct bnxt_re_dev **rdev, struct net_device *netdev)
1576 {
1577         struct bnxt_en_dev *en_dev;
1578         int rc = 0;
1579
1580         if (!is_bnxt_re_dev(netdev))
1581                 return -ENODEV;
1582
1583         en_dev = bnxt_re_dev_probe(netdev);
1584         if (IS_ERR(en_dev)) {
1585                 if (en_dev != ERR_PTR(-ENODEV))
1586                         ibdev_err(&(*rdev)->ibdev, "%s: Failed to probe\n",
1587                                   ROCE_DRV_MODULE_NAME);
1588                 rc = PTR_ERR(en_dev);
1589                 goto exit;
1590         }
1591         *rdev = bnxt_re_dev_add(netdev, en_dev);
1592         if (!*rdev) {
1593                 rc = -ENOMEM;
1594                 bnxt_re_dev_unprobe(netdev, en_dev);
1595                 goto exit;
1596         }
1597 exit:
1598         return rc;
1599 }
1600
1601 static void bnxt_re_remove_device(struct bnxt_re_dev *rdev)
1602 {
1603         bnxt_re_dev_uninit(rdev);
1604         pci_dev_put(rdev->en_dev->pdev);
1605         bnxt_re_dev_unreg(rdev);
1606 }
1607
1608 static int bnxt_re_add_device(struct bnxt_re_dev **rdev,
1609                               struct net_device *netdev, u8 wqe_mode)
1610 {
1611         int rc;
1612
1613         rc = bnxt_re_dev_reg(rdev, netdev);
1614         if (rc == -ENODEV)
1615                 return rc;
1616         if (rc) {
1617                 pr_err("Failed to register with the device %s: %#x\n",
1618                        netdev->name, rc);
1619                 return rc;
1620         }
1621
1622         pci_dev_get((*rdev)->en_dev->pdev);
1623         rc = bnxt_re_dev_init(*rdev, wqe_mode);
1624         if (rc) {
1625                 pci_dev_put((*rdev)->en_dev->pdev);
1626                 bnxt_re_dev_unreg(*rdev);
1627         }
1628
1629         return rc;
1630 }
1631
1632 static void bnxt_re_dealloc_driver(struct ib_device *ib_dev)
1633 {
1634         struct bnxt_re_dev *rdev =
1635                 container_of(ib_dev, struct bnxt_re_dev, ibdev);
1636
1637         dev_info(rdev_to_dev(rdev), "Unregistering Device");
1638
1639         rtnl_lock();
1640         bnxt_re_remove_device(rdev);
1641         rtnl_unlock();
1642 }
1643
1644 /* Handle all deferred netevents tasks */
1645 static void bnxt_re_task(struct work_struct *work)
1646 {
1647         struct bnxt_re_work *re_work;
1648         struct bnxt_re_dev *rdev;
1649         int rc = 0;
1650
1651         re_work = container_of(work, struct bnxt_re_work, work);
1652         rdev = re_work->rdev;
1653
1654         if (re_work->event == NETDEV_REGISTER) {
1655                 rc = bnxt_re_ib_init(rdev);
1656                 if (rc) {
1657                         ibdev_err(&rdev->ibdev,
1658                                   "Failed to register with IB: %#x", rc);
1659                         rtnl_lock();
1660                         bnxt_re_remove_device(rdev);
1661                         rtnl_unlock();
1662                         goto exit;
1663                 }
1664                 goto exit;
1665         }
1666
1667         if (!ib_device_try_get(&rdev->ibdev))
1668                 goto exit;
1669
1670         switch (re_work->event) {
1671         case NETDEV_UP:
1672                 bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1,
1673                                        IB_EVENT_PORT_ACTIVE);
1674                 break;
1675         case NETDEV_DOWN:
1676                 bnxt_re_dev_stop(rdev);
1677                 break;
1678         case NETDEV_CHANGE:
1679                 if (!netif_carrier_ok(rdev->netdev))
1680                         bnxt_re_dev_stop(rdev);
1681                 else if (netif_carrier_ok(rdev->netdev))
1682                         bnxt_re_dispatch_event(&rdev->ibdev, NULL, 1,
1683                                                IB_EVENT_PORT_ACTIVE);
1684                 ib_get_eth_speed(&rdev->ibdev, 1, &rdev->active_speed,
1685                                  &rdev->active_width);
1686                 break;
1687         default:
1688                 break;
1689         }
1690         ib_device_put(&rdev->ibdev);
1691 exit:
1692         put_device(&rdev->ibdev.dev);
1693         kfree(re_work);
1694 }
1695
1696 /*
1697  * "Notifier chain callback can be invoked for the same chain from
1698  * different CPUs at the same time".
1699  *
1700  * For cases when the netdev is already present, our call to the
1701  * register_netdevice_notifier() will actually get the rtnl_lock()
1702  * before sending NETDEV_REGISTER and (if up) NETDEV_UP
1703  * events.
1704  *
1705  * But for cases when the netdev is not already present, the notifier
1706  * chain is subjected to be invoked from different CPUs simultaneously.
1707  *
1708  * This is protected by the netdev_mutex.
1709  */
1710 static int bnxt_re_netdev_event(struct notifier_block *notifier,
1711                                 unsigned long event, void *ptr)
1712 {
1713         struct net_device *real_dev, *netdev = netdev_notifier_info_to_dev(ptr);
1714         struct bnxt_re_work *re_work;
1715         struct bnxt_re_dev *rdev;
1716         int rc = 0;
1717         bool sch_work = false;
1718         bool release = true;
1719
1720         real_dev = rdma_vlan_dev_real_dev(netdev);
1721         if (!real_dev)
1722                 real_dev = netdev;
1723
1724         rdev = bnxt_re_from_netdev(real_dev);
1725         if (!rdev && event != NETDEV_REGISTER)
1726                 return NOTIFY_OK;
1727
1728         if (real_dev != netdev)
1729                 goto exit;
1730
1731         switch (event) {
1732         case NETDEV_REGISTER:
1733                 if (rdev)
1734                         break;
1735                 rc = bnxt_re_add_device(&rdev, real_dev,
1736                                         BNXT_QPLIB_WQE_MODE_STATIC);
1737                 if (!rc)
1738                         sch_work = true;
1739                 release = false;
1740                 break;
1741
1742         case NETDEV_UNREGISTER:
1743                 ib_unregister_device_queued(&rdev->ibdev);
1744                 break;
1745
1746         default:
1747                 sch_work = true;
1748                 break;
1749         }
1750         if (sch_work) {
1751                 /* Allocate for the deferred task */
1752                 re_work = kzalloc(sizeof(*re_work), GFP_ATOMIC);
1753                 if (re_work) {
1754                         get_device(&rdev->ibdev.dev);
1755                         re_work->rdev = rdev;
1756                         re_work->event = event;
1757                         re_work->vlan_dev = (real_dev == netdev ?
1758                                              NULL : netdev);
1759                         INIT_WORK(&re_work->work, bnxt_re_task);
1760                         queue_work(bnxt_re_wq, &re_work->work);
1761                 }
1762         }
1763
1764 exit:
1765         if (rdev && release)
1766                 ib_device_put(&rdev->ibdev);
1767         return NOTIFY_DONE;
1768 }
1769
1770 static struct notifier_block bnxt_re_netdev_notifier = {
1771         .notifier_call = bnxt_re_netdev_event
1772 };
1773
1774 static int __init bnxt_re_mod_init(void)
1775 {
1776         int rc = 0;
1777
1778         pr_info("%s: %s", ROCE_DRV_MODULE_NAME, version);
1779
1780         bnxt_re_wq = create_singlethread_workqueue("bnxt_re");
1781         if (!bnxt_re_wq)
1782                 return -ENOMEM;
1783
1784         INIT_LIST_HEAD(&bnxt_re_dev_list);
1785
1786         rc = register_netdevice_notifier(&bnxt_re_netdev_notifier);
1787         if (rc) {
1788                 pr_err("%s: Cannot register to netdevice_notifier",
1789                        ROCE_DRV_MODULE_NAME);
1790                 goto err_netdev;
1791         }
1792         return 0;
1793
1794 err_netdev:
1795         destroy_workqueue(bnxt_re_wq);
1796
1797         return rc;
1798 }
1799
1800 static void __exit bnxt_re_mod_exit(void)
1801 {
1802         struct bnxt_re_dev *rdev;
1803
1804         unregister_netdevice_notifier(&bnxt_re_netdev_notifier);
1805         if (bnxt_re_wq)
1806                 destroy_workqueue(bnxt_re_wq);
1807         list_for_each_entry(rdev, &bnxt_re_dev_list, list) {
1808                 /* VF device removal should be called before the removal
1809                  * of PF device. Queue VFs unregister first, so that VFs
1810                  * shall be removed before the PF during the call of
1811                  * ib_unregister_driver.
1812                  */
1813                 if (rdev->is_virtfn)
1814                         ib_unregister_device(&rdev->ibdev);
1815         }
1816         ib_unregister_driver(RDMA_DRIVER_BNXT_RE);
1817 }
1818
1819 module_init(bnxt_re_mod_init);
1820 module_exit(bnxt_re_mod_exit);