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