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