sys/dev/mlx5/mlx5_en/mlx5_en_main.c

   1 /*-
   2  * Copyright (c) 2015-2018 Mellanox Technologies. All rights reserved.
   3  *
   4  * Redistribution and use in source and binary forms, with or without
   5  * modification, are permitted provided that the following conditions
   6  * are met:
   7  * 1. Redistributions of source code must retain the above copyright
   8  *    notice, this list of conditions and the following disclaimer.
   9  * 2. Redistributions in binary form must reproduce the above copyright
  10  *    notice, this list of conditions and the following disclaimer in the
  11  *    documentation and/or other materials provided with the distribution.
  12  *
  13  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS `AS IS' AND
  14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  16  * ARE DISCLAIMED.  IN NO EVENT SHALL AUTHOR OR CONTRIBUTORS BE LIABLE
  17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  23  * SUCH DAMAGE.
  24  *
  25  * $FreeBSD$
  26  */
  27
  28 #include "en.h"
  29
  30 #include <sys/eventhandler.h>
  31 #include <sys/sockio.h>
  32 #include <machine/atomic.h>
  33
  34 #ifndef ETH_DRIVER_VERSION
  35 #define ETH_DRIVER_VERSION      "3.5.1"
  36 #endif
  37 #define DRIVER_RELDATE  "April 2019"
  38
  39 static const char mlx5e_version[] = "mlx5en: Mellanox Ethernet driver "
  40         ETH_DRIVER_VERSION " (" DRIVER_RELDATE ")\n";
  41
  42 static int mlx5e_get_wqe_sz(struct mlx5e_priv *priv, u32 *wqe_sz, u32 *nsegs);
  43
  44 struct mlx5e_channel_param {
  45         struct mlx5e_rq_param rq;
  46         struct mlx5e_sq_param sq;
  47         struct mlx5e_cq_param rx_cq;
  48         struct mlx5e_cq_param tx_cq;
  49 };
  50
  51 struct media {
  52         u32     subtype;
  53         u64     baudrate;
  54 };
  55
  56 static const struct media mlx5e_mode_table[MLX5E_LINK_SPEEDS_NUMBER][MLX5E_LINK_MODES_NUMBER] = {
  57
  58         [MLX5E_1000BASE_CX_SGMII][MLX5E_SGMII] = {
  59                 .subtype = IFM_1000_CX_SGMII,
  60                 .baudrate = IF_Mbps(1000ULL),
  61         },
  62         [MLX5E_1000BASE_KX][MLX5E_KX] = {
  63                 .subtype = IFM_1000_KX,
  64                 .baudrate = IF_Mbps(1000ULL),
  65         },
  66         [MLX5E_10GBASE_CX4][MLX5E_CX4] = {
  67                 .subtype = IFM_10G_CX4,
  68                 .baudrate = IF_Gbps(10ULL),
  69         },
  70         [MLX5E_10GBASE_KX4][MLX5E_KX4] = {
  71                 .subtype = IFM_10G_KX4,
  72                 .baudrate = IF_Gbps(10ULL),
  73         },
  74         [MLX5E_10GBASE_KR][MLX5E_KR] = {
  75                 .subtype = IFM_10G_KR,
  76                 .baudrate = IF_Gbps(10ULL),
  77         },
  78         [MLX5E_20GBASE_KR2][MLX5E_KR2] = {
  79                 .subtype = IFM_20G_KR2,
  80                 .baudrate = IF_Gbps(20ULL),
  81         },
  82         [MLX5E_40GBASE_CR4][MLX5E_CR4] = {
  83                 .subtype = IFM_40G_CR4,
  84                 .baudrate = IF_Gbps(40ULL),
  85         },
  86         [MLX5E_40GBASE_KR4][MLX5E_KR4] = {
  87                 .subtype = IFM_40G_KR4,
  88                 .baudrate = IF_Gbps(40ULL),
  89         },
  90         [MLX5E_56GBASE_R4][MLX5E_R] = {
  91                 .subtype = IFM_56G_R4,
  92                 .baudrate = IF_Gbps(56ULL),
  93         },
  94         [MLX5E_10GBASE_CR][MLX5E_CR1] = {
  95                 .subtype = IFM_10G_CR1,
  96                 .baudrate = IF_Gbps(10ULL),
  97         },
  98         [MLX5E_10GBASE_SR][MLX5E_SR] = {
  99                 .subtype = IFM_10G_SR,
 100                 .baudrate = IF_Gbps(10ULL),
 101         },
 102         [MLX5E_10GBASE_ER_LR][MLX5E_ER] = {
 103                 .subtype = IFM_10G_ER,
 104                 .baudrate = IF_Gbps(10ULL),
 105         },
 106         [MLX5E_10GBASE_ER_LR][MLX5E_LR] = {
 107                 .subtype = IFM_10G_LR,
 108                 .baudrate = IF_Gbps(10ULL),
 109         },
 110         [MLX5E_40GBASE_SR4][MLX5E_SR4] = {
 111                 .subtype = IFM_40G_SR4,
 112                 .baudrate = IF_Gbps(40ULL),
 113         },
 114         [MLX5E_40GBASE_LR4_ER4][MLX5E_LR4] = {
 115                 .subtype = IFM_40G_LR4,
 116                 .baudrate = IF_Gbps(40ULL),
 117         },
 118         [MLX5E_40GBASE_LR4_ER4][MLX5E_ER4] = {
 119                 .subtype = IFM_40G_ER4,
 120                 .baudrate = IF_Gbps(40ULL),
 121         },
 122         [MLX5E_100GBASE_CR4][MLX5E_CR4] = {
 123                 .subtype = IFM_100G_CR4,
 124                 .baudrate = IF_Gbps(100ULL),
 125         },
 126         [MLX5E_100GBASE_SR4][MLX5E_SR4] = {
 127                 .subtype = IFM_100G_SR4,
 128                 .baudrate = IF_Gbps(100ULL),
 129         },
 130         [MLX5E_100GBASE_KR4][MLX5E_KR4] = {
 131                 .subtype = IFM_100G_KR4,
 132                 .baudrate = IF_Gbps(100ULL),
 133         },
 134         [MLX5E_100GBASE_LR4][MLX5E_LR4] = {
 135                 .subtype = IFM_100G_LR4,
 136                 .baudrate = IF_Gbps(100ULL),
 137         },
 138         [MLX5E_100BASE_TX][MLX5E_TX] = {
 139                 .subtype = IFM_100_TX,
 140                 .baudrate = IF_Mbps(100ULL),
 141         },
 142         [MLX5E_1000BASE_T][MLX5E_T] = {
 143                 .subtype = IFM_1000_T,
 144                 .baudrate = IF_Mbps(1000ULL),
 145         },
 146         [MLX5E_10GBASE_T][MLX5E_T] = {
 147                 .subtype = IFM_10G_T,
 148                 .baudrate = IF_Gbps(10ULL),
 149         },
 150         [MLX5E_25GBASE_CR][MLX5E_CR] = {
 151                 .subtype = IFM_25G_CR,
 152                 .baudrate = IF_Gbps(25ULL),
 153         },
 154         [MLX5E_25GBASE_KR][MLX5E_KR] = {
 155                 .subtype = IFM_25G_KR,
 156                 .baudrate = IF_Gbps(25ULL),
 157         },
 158         [MLX5E_25GBASE_SR][MLX5E_SR] = {
 159                 .subtype = IFM_25G_SR,
 160                 .baudrate = IF_Gbps(25ULL),
 161         },
 162         [MLX5E_50GBASE_CR2][MLX5E_CR2] = {
 163                 .subtype = IFM_50G_CR2,
 164                 .baudrate = IF_Gbps(50ULL),
 165         },
 166         [MLX5E_50GBASE_KR2][MLX5E_KR2] = {
 167                 .subtype = IFM_50G_KR2,
 168                 .baudrate = IF_Gbps(50ULL),
 169         },
 170 };
 171
 172 static const struct media mlx5e_ext_mode_table[MLX5E_EXT_LINK_SPEEDS_NUMBER][MLX5E_LINK_MODES_NUMBER] = {
 173         [MLX5E_SGMII_100M][MLX5E_SGMII] = {
 174                 .subtype = IFM_100_SGMII,
 175                 .baudrate = IF_Mbps(100),
 176         },
 177         [MLX5E_1000BASE_X_SGMII][MLX5E_KX] = {
 178                 .subtype = IFM_1000_KX,
 179                 .baudrate = IF_Mbps(1000),
 180         },
 181         [MLX5E_1000BASE_X_SGMII][MLX5E_CX_SGMII] = {
 182                 .subtype = IFM_1000_CX_SGMII,
 183                 .baudrate = IF_Mbps(1000),
 184         },
 185         [MLX5E_1000BASE_X_SGMII][MLX5E_CX] = {
 186                 .subtype = IFM_1000_CX,
 187                 .baudrate = IF_Mbps(1000),
 188         },
 189         [MLX5E_1000BASE_X_SGMII][MLX5E_LX] = {
 190                 .subtype = IFM_1000_LX,
 191                 .baudrate = IF_Mbps(1000),
 192         },
 193         [MLX5E_1000BASE_X_SGMII][MLX5E_SX] = {
 194                 .subtype = IFM_1000_SX,
 195                 .baudrate = IF_Mbps(1000),
 196         },
 197         [MLX5E_1000BASE_X_SGMII][MLX5E_T] = {
 198                 .subtype = IFM_1000_T,
 199                 .baudrate = IF_Mbps(1000),
 200         },
 201         [MLX5E_5GBASE_R][MLX5E_T] = {
 202                 .subtype = IFM_5000_T,
 203                 .baudrate = IF_Mbps(5000),
 204         },
 205         [MLX5E_5GBASE_R][MLX5E_KR] = {
 206                 .subtype = IFM_5000_KR,
 207                 .baudrate = IF_Mbps(5000),
 208         },
 209         [MLX5E_5GBASE_R][MLX5E_KR1] = {
 210                 .subtype = IFM_5000_KR1,
 211                 .baudrate = IF_Mbps(5000),
 212         },
 213         [MLX5E_5GBASE_R][MLX5E_KR_S] = {
 214                 .subtype = IFM_5000_KR_S,
 215                 .baudrate = IF_Mbps(5000),
 216         },
 217         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_ER] = {
 218                 .subtype = IFM_10G_ER,
 219                 .baudrate = IF_Gbps(10ULL),
 220         },
 221         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_KR] = {
 222                 .subtype = IFM_10G_KR,
 223                 .baudrate = IF_Gbps(10ULL),
 224         },
 225         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_LR] = {
 226                 .subtype = IFM_10G_LR,
 227                 .baudrate = IF_Gbps(10ULL),
 228         },
 229         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_SR] = {
 230                 .subtype = IFM_10G_SR,
 231                 .baudrate = IF_Gbps(10ULL),
 232         },
 233         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_T] = {
 234                 .subtype = IFM_10G_T,
 235                 .baudrate = IF_Gbps(10ULL),
 236         },
 237         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_AOC] = {
 238                 .subtype = IFM_10G_AOC,
 239                 .baudrate = IF_Gbps(10ULL),
 240         },
 241         [MLX5E_10GBASE_XFI_XAUI_1][MLX5E_CR1] = {
 242                 .subtype = IFM_10G_CR1,
 243                 .baudrate = IF_Gbps(10ULL),
 244         },
 245         [MLX5E_40GBASE_XLAUI_4_XLPPI_4][MLX5E_CR4] = {
 246                 .subtype = IFM_40G_CR4,
 247                 .baudrate = IF_Gbps(40ULL),
 248         },
 249         [MLX5E_40GBASE_XLAUI_4_XLPPI_4][MLX5E_KR4] = {
 250                 .subtype = IFM_40G_KR4,
 251                 .baudrate = IF_Gbps(40ULL),
 252         },
 253         [MLX5E_40GBASE_XLAUI_4_XLPPI_4][MLX5E_LR4] = {
 254                 .subtype = IFM_40G_LR4,
 255                 .baudrate = IF_Gbps(40ULL),
 256         },
 257         [MLX5E_40GBASE_XLAUI_4_XLPPI_4][MLX5E_SR4] = {
 258                 .subtype = IFM_40G_SR4,
 259                 .baudrate = IF_Gbps(40ULL),
 260         },
 261         [MLX5E_40GBASE_XLAUI_4_XLPPI_4][MLX5E_ER4] = {
 262                 .subtype = IFM_40G_ER4,
 263                 .baudrate = IF_Gbps(40ULL),
 264         },
 265
 266         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_CR] = {
 267                 .subtype = IFM_25G_CR,
 268                 .baudrate = IF_Gbps(25ULL),
 269         },
 270         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_KR] = {
 271                 .subtype = IFM_25G_KR,
 272                 .baudrate = IF_Gbps(25ULL),
 273         },
 274         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_SR] = {
 275                 .subtype = IFM_25G_SR,
 276                 .baudrate = IF_Gbps(25ULL),
 277         },
 278         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_ACC] = {
 279                 .subtype = IFM_25G_ACC,
 280                 .baudrate = IF_Gbps(25ULL),
 281         },
 282         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_AOC] = {
 283                 .subtype = IFM_25G_AOC,
 284                 .baudrate = IF_Gbps(25ULL),
 285         },
 286         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_CR1] = {
 287                 .subtype = IFM_25G_CR1,
 288                 .baudrate = IF_Gbps(25ULL),
 289         },
 290         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_CR_S] = {
 291                 .subtype = IFM_25G_CR_S,
 292                 .baudrate = IF_Gbps(25ULL),
 293         },
 294         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_KR1] = {
 295                 .subtype = IFM_5000_KR1,
 296                 .baudrate = IF_Gbps(25ULL),
 297         },
 298         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_KR_S] = {
 299                 .subtype = IFM_25G_KR_S,
 300                 .baudrate = IF_Gbps(25ULL),
 301         },
 302         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_LR] = {
 303                 .subtype = IFM_25G_LR,
 304                 .baudrate = IF_Gbps(25ULL),
 305         },
 306         [MLX5E_25GAUI_1_25GBASE_CR_KR][MLX5E_T] = {
 307                 .subtype = IFM_25G_T,
 308                 .baudrate = IF_Gbps(25ULL),
 309         },
 310         [MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2][MLX5E_CR2] = {
 311                 .subtype = IFM_50G_CR2,
 312                 .baudrate = IF_Gbps(50ULL),
 313         },
 314         [MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2][MLX5E_KR2] = {
 315                 .subtype = IFM_50G_KR2,
 316                 .baudrate = IF_Gbps(50ULL),
 317         },
 318         [MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2][MLX5E_SR2] = {
 319                 .subtype = IFM_50G_SR2,
 320                 .baudrate = IF_Gbps(50ULL),
 321         },
 322         [MLX5E_50GAUI_2_LAUI_2_50GBASE_CR2_KR2][MLX5E_LR2] = {
 323                 .subtype = IFM_50G_LR2,
 324                 .baudrate = IF_Gbps(50ULL),
 325         },
 326         [MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR][MLX5E_LR] = {
 327                 .subtype = IFM_50G_LR,
 328                 .baudrate = IF_Gbps(50ULL),
 329         },
 330         [MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR][MLX5E_SR] = {
 331                 .subtype = IFM_50G_SR,
 332                 .baudrate = IF_Gbps(50ULL),
 333         },
 334         [MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR][MLX5E_CP] = {
 335                 .subtype = IFM_50G_CP,
 336                 .baudrate = IF_Gbps(50ULL),
 337         },
 338         [MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR][MLX5E_FR] = {
 339                 .subtype = IFM_50G_FR,
 340                 .baudrate = IF_Gbps(50ULL),
 341         },
 342         [MLX5E_50GAUI_1_LAUI_1_50GBASE_CR_KR][MLX5E_KR_PAM4] = {
 343                 .subtype = IFM_50G_KR_PAM4,
 344                 .baudrate = IF_Gbps(50ULL),
 345         },
 346         [MLX5E_CAUI_4_100GBASE_CR4_KR4][MLX5E_CR4] = {
 347                 .subtype = IFM_100G_CR4,
 348                 .baudrate = IF_Gbps(100ULL),
 349         },
 350         [MLX5E_CAUI_4_100GBASE_CR4_KR4][MLX5E_KR4] = {
 351                 .subtype = IFM_100G_KR4,
 352                 .baudrate = IF_Gbps(100ULL),
 353         },
 354         [MLX5E_CAUI_4_100GBASE_CR4_KR4][MLX5E_LR4] = {
 355                 .subtype = IFM_100G_LR4,
 356                 .baudrate = IF_Gbps(100ULL),
 357         },
 358         [MLX5E_CAUI_4_100GBASE_CR4_KR4][MLX5E_SR4] = {
 359                 .subtype = IFM_100G_SR4,
 360                 .baudrate = IF_Gbps(100ULL),
 361         },
 362         [MLX5E_100GAUI_2_100GBASE_CR2_KR2][MLX5E_SR2] = {
 363                 .subtype = IFM_100G_SR2,
 364                 .baudrate = IF_Gbps(100ULL),
 365         },
 366         [MLX5E_100GAUI_2_100GBASE_CR2_KR2][MLX5E_CP2] = {
 367                 .subtype = IFM_100G_CP2,
 368                 .baudrate = IF_Gbps(100ULL),
 369         },
 370         [MLX5E_100GAUI_2_100GBASE_CR2_KR2][MLX5E_KR2_PAM4] = {
 371                 .subtype = IFM_100G_KR2_PAM4,
 372                 .baudrate = IF_Gbps(100ULL),
 373         },
 374         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_DR4] = {
 375                 .subtype = IFM_200G_DR4,
 376                 .baudrate = IF_Gbps(200ULL),
 377         },
 378         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_LR4] = {
 379                 .subtype = IFM_200G_LR4,
 380                 .baudrate = IF_Gbps(200ULL),
 381         },
 382         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_SR4] = {
 383                 .subtype = IFM_200G_SR4,
 384                 .baudrate = IF_Gbps(200ULL),
 385         },
 386         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_FR4] = {
 387                 .subtype = IFM_200G_FR4,
 388                 .baudrate = IF_Gbps(200ULL),
 389         },
 390         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_CR4_PAM4] = {
 391                 .subtype = IFM_200G_CR4_PAM4,
 392                 .baudrate = IF_Gbps(200ULL),
 393         },
 394         [MLX5E_200GAUI_4_200GBASE_CR4_KR4][MLX5E_KR4_PAM4] = {
 395                 .subtype = IFM_200G_KR4_PAM4,
 396                 .baudrate = IF_Gbps(200ULL),
 397         },
 398 };
 399
 400 MALLOC_DEFINE(M_MLX5EN, "MLX5EN", "MLX5 Ethernet");
 401
 402 static void
 403 mlx5e_update_carrier(struct mlx5e_priv *priv)
 404 {
 405         struct mlx5_core_dev *mdev = priv->mdev;
 406         u32 out[MLX5_ST_SZ_DW(ptys_reg)];
 407         u32 eth_proto_oper;
 408         int error;
 409         u8 port_state;
 410         u8 is_er_type;
 411         u8 i, j;
 412         bool ext;
 413         struct media media_entry = {};
 414
 415         port_state = mlx5_query_vport_state(mdev,
 416             MLX5_QUERY_VPORT_STATE_IN_OP_MOD_VNIC_VPORT, 0);
 417
 418         if (port_state == VPORT_STATE_UP) {
 419                 priv->media_status_last |= IFM_ACTIVE;
 420         } else {
 421                 priv->media_status_last &= ~IFM_ACTIVE;
 422                 priv->media_active_last = IFM_ETHER;
 423                 if_link_state_change(priv->ifp, LINK_STATE_DOWN);
 424                 return;
 425         }
 426
 427         error = mlx5_query_port_ptys(mdev, out, sizeof(out),
 428             MLX5_PTYS_EN, 1);
 429         if (error) {
 430                 priv->media_active_last = IFM_ETHER;
 431                 priv->ifp->if_baudrate = 1;
 432                 if_printf(priv->ifp, "%s: query port ptys failed: "
 433                     "0x%x\n", __func__, error);
 434                 return;
 435         }
 436
 437         ext = MLX5_CAP_PCAM_FEATURE(mdev, ptys_extended_ethernet);
 438         eth_proto_oper = MLX5_GET_ETH_PROTO(ptys_reg, out, ext,
 439             eth_proto_oper);
 440
 441         i = ilog2(eth_proto_oper);
 442
 443         for (j = 0; j != MLX5E_LINK_MODES_NUMBER; j++) {
 444                 media_entry = ext ? mlx5e_ext_mode_table[i][j] :
 445                     mlx5e_mode_table[i][j];
 446                 if (media_entry.baudrate != 0)
 447                         break;
 448         }
 449
 450         if (media_entry.subtype == 0) {
 451                 if_printf(priv->ifp, "%s: Could not find operational "
 452                     "media subtype\n", __func__);
 453                 return;
 454         }
 455
 456         switch (media_entry.subtype) {
 457         case IFM_10G_ER:
 458                 error = mlx5_query_pddr_range_info(mdev, 1, &is_er_type);
 459                 if (error != 0) {
 460                         if_printf(priv->ifp, "%s: query port pddr failed: %d\n",
 461                                   __func__, error);
 462                 }
 463                 if (error != 0 || is_er_type == 0)
 464                         media_entry.subtype = IFM_10G_LR;
 465                 break;
 466         case IFM_40G_LR4:
 467                 error = mlx5_query_pddr_range_info(mdev, 1, &is_er_type);
 468                 if (error != 0) {
 469                         if_printf(priv->ifp, "%s: query port pddr failed: %d\n",
 470                                   __func__, error);
 471                 }
 472                 if (error == 0 && is_er_type != 0)
 473                         media_entry.subtype = IFM_40G_ER4;
 474                 break;
 475         }
 476         priv->media_active_last = media_entry.subtype | IFM_ETHER | IFM_FDX;
 477         priv->ifp->if_baudrate = media_entry.baudrate;
 478
 479         if_link_state_change(priv->ifp, LINK_STATE_UP);
 480 }
 481
 482 static void
 483 mlx5e_media_status(struct ifnet *dev, struct ifmediareq *ifmr)
 484 {
 485         struct mlx5e_priv *priv = dev->if_softc;
 486
 487         ifmr->ifm_status = priv->media_status_last;
 488         ifmr->ifm_active = priv->media_active_last |
 489             (priv->params.rx_pauseframe_control ? IFM_ETH_RXPAUSE : 0) |
 490             (priv->params.tx_pauseframe_control ? IFM_ETH_TXPAUSE : 0);
 491
 492 }
 493
 494 static u32
 495 mlx5e_find_link_mode(u32 subtype, bool ext)
 496 {
 497         u32 i;
 498         u32 j;
 499         u32 link_mode = 0;
 500         u32 speeds_num = 0;
 501         struct media media_entry = {};
 502
 503         switch (subtype) {
 504         case IFM_10G_LR:
 505                 subtype = IFM_10G_ER;
 506                 break;
 507         case IFM_40G_ER4:
 508                 subtype = IFM_40G_LR4;
 509                 break;
 510         }
 511
 512         speeds_num = ext ? MLX5E_EXT_LINK_SPEEDS_NUMBER :
 513             MLX5E_LINK_SPEEDS_NUMBER;
 514
 515         for (i = 0; i != speeds_num; i++) {
 516                 for (j = 0; j < MLX5E_LINK_MODES_NUMBER ; ++j) {
 517                         media_entry = ext ? mlx5e_ext_mode_table[i][j] :
 518                             mlx5e_mode_table[i][j];
 519                         if (media_entry.baudrate == 0)
 520                                 continue;
 521                         if (media_entry.subtype == subtype) {
 522                                 link_mode |= MLX5E_PROT_MASK(i);
 523                         }
 524                 }
 525         }
 526
 527         return (link_mode);
 528 }
 529
 530 static int
 531 mlx5e_set_port_pause_and_pfc(struct mlx5e_priv *priv)
 532 {
 533         return (mlx5_set_port_pause_and_pfc(priv->mdev, 1,
 534             priv->params.rx_pauseframe_control,
 535             priv->params.tx_pauseframe_control,
 536             priv->params.rx_priority_flow_control,
 537             priv->params.tx_priority_flow_control));
 538 }
 539
 540 static int
 541 mlx5e_set_port_pfc(struct mlx5e_priv *priv)
 542 {
 543         int error;
 544
 545         if (priv->gone != 0) {
 546                 error = -ENXIO;
 547         } else if (priv->params.rx_pauseframe_control ||
 548             priv->params.tx_pauseframe_control) {
 549                 if_printf(priv->ifp,
 550                     "Global pauseframes must be disabled before "
 551                     "enabling PFC.\n");
 552                 error = -EINVAL;
 553         } else {
 554                 error = mlx5e_set_port_pause_and_pfc(priv);
 555         }
 556         return (error);
 557 }
 558
 559 static int
 560 mlx5e_media_change(struct ifnet *dev)
 561 {
 562         struct mlx5e_priv *priv = dev->if_softc;
 563         struct mlx5_core_dev *mdev = priv->mdev;
 564         u32 eth_proto_cap;
 565         u32 link_mode;
 566         u32 out[MLX5_ST_SZ_DW(ptys_reg)];
 567         int was_opened;
 568         int locked;
 569         int error;
 570         bool ext;
 571
 572         locked = PRIV_LOCKED(priv);
 573         if (!locked)
 574                 PRIV_LOCK(priv);
 575
 576         if (IFM_TYPE(priv->media.ifm_media) != IFM_ETHER) {
 577                 error = EINVAL;
 578                 goto done;
 579         }
 580
 581         error = mlx5_query_port_ptys(mdev, out, sizeof(out),
 582             MLX5_PTYS_EN, 1);
 583         if (error != 0) {
 584                 if_printf(dev, "Query port media capability failed\n");
 585                 goto done;
 586         }
 587
 588         ext = MLX5_CAP_PCAM_FEATURE(mdev, ptys_extended_ethernet);
 589         link_mode = mlx5e_find_link_mode(IFM_SUBTYPE(priv->media.ifm_media), ext);
 590
 591         /* query supported capabilities */
 592         eth_proto_cap = MLX5_GET_ETH_PROTO(ptys_reg, out, ext,
 593             eth_proto_capability);
 594
 595         /* check for autoselect */
 596         if (IFM_SUBTYPE(priv->media.ifm_media) == IFM_AUTO) {
 597                 link_mode = eth_proto_cap;
 598                 if (link_mode == 0) {
 599                         if_printf(dev, "Port media capability is zero\n");
 600                         error = EINVAL;
 601                         goto done;
 602                 }
 603         } else {
 604                 link_mode = link_mode & eth_proto_cap;
 605                 if (link_mode == 0) {
 606                         if_printf(dev, "Not supported link mode requested\n");
 607                         error = EINVAL;
 608                         goto done;
 609                 }
 610         }
 611         if (priv->media.ifm_media & (IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE)) {
 612                 /* check if PFC is enabled */
 613                 if (priv->params.rx_priority_flow_control ||
 614                     priv->params.tx_priority_flow_control) {
 615                         if_printf(dev, "PFC must be disabled before enabling global pauseframes.\n");
 616                         error = EINVAL;
 617                         goto done;
 618                 }
 619         }
 620         /* update pauseframe control bits */
 621         priv->params.rx_pauseframe_control =
 622             (priv->media.ifm_media & IFM_ETH_RXPAUSE) ? 1 : 0;
 623         priv->params.tx_pauseframe_control =
 624             (priv->media.ifm_media & IFM_ETH_TXPAUSE) ? 1 : 0;
 625
 626         /* check if device is opened */
 627         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
 628
 629         /* reconfigure the hardware */
 630         mlx5_set_port_status(mdev, MLX5_PORT_DOWN);
 631         mlx5_set_port_proto(mdev, link_mode, MLX5_PTYS_EN, ext);
 632         error = -mlx5e_set_port_pause_and_pfc(priv);
 633         if (was_opened)
 634                 mlx5_set_port_status(mdev, MLX5_PORT_UP);
 635
 636 done:
 637         if (!locked)
 638                 PRIV_UNLOCK(priv);
 639         return (error);
 640 }
 641
 642 static void
 643 mlx5e_update_carrier_work(struct work_struct *work)
 644 {
 645         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 646             update_carrier_work);
 647
 648         PRIV_LOCK(priv);
 649         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 650                 mlx5e_update_carrier(priv);
 651         PRIV_UNLOCK(priv);
 652 }
 653
 654 #define MLX5E_PCIE_PERF_GET_64(a,b,c,d,e,f)    \
 655         s_debug->c = MLX5_GET64(mpcnt_reg, out, counter_set.f.c);
 656
 657 #define MLX5E_PCIE_PERF_GET_32(a,b,c,d,e,f)    \
 658         s_debug->c = MLX5_GET(mpcnt_reg, out, counter_set.f.c);
 659
 660 static void
 661 mlx5e_update_pcie_counters(struct mlx5e_priv *priv)
 662 {
 663         struct mlx5_core_dev *mdev = priv->mdev;
 664         struct mlx5e_port_stats_debug *s_debug = &priv->stats.port_stats_debug;
 665         const unsigned sz = MLX5_ST_SZ_BYTES(mpcnt_reg);
 666         void *out;
 667         void *in;
 668         int err;
 669
 670         /* allocate firmware request structures */
 671         in = mlx5_vzalloc(sz);
 672         out = mlx5_vzalloc(sz);
 673         if (in == NULL || out == NULL)
 674                 goto free_out;
 675
 676         MLX5_SET(mpcnt_reg, in, grp, MLX5_PCIE_PERFORMANCE_COUNTERS_GROUP);
 677         err = mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_MPCNT, 0, 0);
 678         if (err != 0)
 679                 goto free_out;
 680
 681         MLX5E_PCIE_PERFORMANCE_COUNTERS_64(MLX5E_PCIE_PERF_GET_64)
 682         MLX5E_PCIE_PERFORMANCE_COUNTERS_32(MLX5E_PCIE_PERF_GET_32)
 683
 684         MLX5_SET(mpcnt_reg, in, grp, MLX5_PCIE_TIMERS_AND_STATES_COUNTERS_GROUP);
 685         err = mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_MPCNT, 0, 0);
 686         if (err != 0)
 687                 goto free_out;
 688
 689         MLX5E_PCIE_TIMERS_AND_STATES_COUNTERS_32(MLX5E_PCIE_PERF_GET_32)
 690
 691         MLX5_SET(mpcnt_reg, in, grp, MLX5_PCIE_LANE_COUNTERS_GROUP);
 692         err = mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_MPCNT, 0, 0);
 693         if (err != 0)
 694                 goto free_out;
 695
 696         MLX5E_PCIE_LANE_COUNTERS_32(MLX5E_PCIE_PERF_GET_32)
 697
 698 free_out:
 699         /* free firmware request structures */
 700         kvfree(in);
 701         kvfree(out);
 702 }
 703
 704 /*
 705  * This function reads the physical port counters from the firmware
 706  * using a pre-defined layout defined by various MLX5E_PPORT_XXX()
 707  * macros. The output is converted from big-endian 64-bit values into
 708  * host endian ones and stored in the "priv->stats.pport" structure.
 709  */
 710 static void
 711 mlx5e_update_pport_counters(struct mlx5e_priv *priv)
 712 {
 713         struct mlx5_core_dev *mdev = priv->mdev;
 714         struct mlx5e_pport_stats *s = &priv->stats.pport;
 715         struct mlx5e_port_stats_debug *s_debug = &priv->stats.port_stats_debug;
 716         u32 *in;
 717         u32 *out;
 718         const u64 *ptr;
 719         unsigned sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
 720         unsigned x;
 721         unsigned y;
 722         unsigned z;
 723
 724         /* allocate firmware request structures */
 725         in = mlx5_vzalloc(sz);
 726         out = mlx5_vzalloc(sz);
 727         if (in == NULL || out == NULL)
 728                 goto free_out;
 729
 730         /*
 731          * Get pointer to the 64-bit counter set which is located at a
 732          * fixed offset in the output firmware request structure:
 733          */
 734         ptr = (const uint64_t *)MLX5_ADDR_OF(ppcnt_reg, out, counter_set);
 735
 736         MLX5_SET(ppcnt_reg, in, local_port, 1);
 737
 738         /* read IEEE802_3 counter group using predefined counter layout */
 739         MLX5_SET(ppcnt_reg, in, grp, MLX5_IEEE_802_3_COUNTERS_GROUP);
 740         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 741         for (x = 0, y = MLX5E_PPORT_PER_PRIO_STATS_NUM;
 742              x != MLX5E_PPORT_IEEE802_3_STATS_NUM; x++, y++)
 743                 s->arg[y] = be64toh(ptr[x]);
 744
 745         /* read RFC2819 counter group using predefined counter layout */
 746         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2819_COUNTERS_GROUP);
 747         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 748         for (x = 0; x != MLX5E_PPORT_RFC2819_STATS_NUM; x++, y++)
 749                 s->arg[y] = be64toh(ptr[x]);
 750
 751         for (y = 0; x != MLX5E_PPORT_RFC2819_STATS_NUM +
 752             MLX5E_PPORT_RFC2819_STATS_DEBUG_NUM; x++, y++)
 753                 s_debug->arg[y] = be64toh(ptr[x]);
 754
 755         /* read RFC2863 counter group using predefined counter layout */
 756         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2863_COUNTERS_GROUP);
 757         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 758         for (x = 0; x != MLX5E_PPORT_RFC2863_STATS_DEBUG_NUM; x++, y++)
 759                 s_debug->arg[y] = be64toh(ptr[x]);
 760
 761         /* read physical layer stats counter group using predefined counter layout */
 762         MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_COUNTERS_GROUP);
 763         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 764         for (x = 0; x != MLX5E_PPORT_PHYSICAL_LAYER_STATS_DEBUG_NUM; x++, y++)
 765                 s_debug->arg[y] = be64toh(ptr[x]);
 766
 767         /* read Extended Ethernet counter group using predefined counter layout */
 768         MLX5_SET(ppcnt_reg, in, grp, MLX5_ETHERNET_EXTENDED_COUNTERS_GROUP);
 769         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 770         for (x = 0; x != MLX5E_PPORT_ETHERNET_EXTENDED_STATS_DEBUG_NUM; x++, y++)
 771                 s_debug->arg[y] = be64toh(ptr[x]);
 772
 773         /* read Extended Statistical Group */
 774         if (MLX5_CAP_GEN(mdev, pcam_reg) &&
 775             MLX5_CAP_PCAM_FEATURE(mdev, ppcnt_statistical_group) &&
 776             MLX5_CAP_PCAM_FEATURE(mdev, per_lane_error_counters)) {
 777                 /* read Extended Statistical counter group using predefined counter layout */
 778                 MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_STATISTICAL_GROUP);
 779                 mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 780
 781                 for (x = 0; x != MLX5E_PPORT_STATISTICAL_DEBUG_NUM; x++, y++)
 782                         s_debug->arg[y] = be64toh(ptr[x]);
 783         }
 784
 785         /* read PCIE counters */
 786         mlx5e_update_pcie_counters(priv);
 787
 788         /* read per-priority counters */
 789         MLX5_SET(ppcnt_reg, in, grp, MLX5_PER_PRIORITY_COUNTERS_GROUP);
 790
 791         /* iterate all the priorities */
 792         for (y = z = 0; z != MLX5E_PPORT_PER_PRIO_STATS_NUM_PRIO; z++) {
 793                 MLX5_SET(ppcnt_reg, in, prio_tc, z);
 794                 mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 795
 796                 /* read per priority stats counter group using predefined counter layout */
 797                 for (x = 0; x != (MLX5E_PPORT_PER_PRIO_STATS_NUM /
 798                     MLX5E_PPORT_PER_PRIO_STATS_NUM_PRIO); x++, y++)
 799                         s->arg[y] = be64toh(ptr[x]);
 800         }
 801
 802 free_out:
 803         /* free firmware request structures */
 804         kvfree(in);
 805         kvfree(out);
 806 }
 807
 808 static void
 809 mlx5e_grp_vnic_env_update_stats(struct mlx5e_priv *priv)
 810 {
 811         u32 out[MLX5_ST_SZ_DW(query_vnic_env_out)] = {};
 812         u32 in[MLX5_ST_SZ_DW(query_vnic_env_in)] = {};
 813
 814         if (!MLX5_CAP_GEN(priv->mdev, nic_receive_steering_discard))
 815                 return;
 816
 817         MLX5_SET(query_vnic_env_in, in, opcode,
 818             MLX5_CMD_OP_QUERY_VNIC_ENV);
 819         MLX5_SET(query_vnic_env_in, in, op_mod, 0);
 820         MLX5_SET(query_vnic_env_in, in, other_vport, 0);
 821
 822         if (mlx5_cmd_exec(priv->mdev, in, sizeof(in), out, sizeof(out)) != 0)
 823                 return;
 824
 825         priv->stats.vport.rx_steer_missed_packets =
 826             MLX5_GET64(query_vnic_env_out, out,
 827             vport_env.nic_receive_steering_discard);
 828 }
 829
 830 /*
 831  * This function is called regularly to collect all statistics
 832  * counters from the firmware. The values can be viewed through the
 833  * sysctl interface. Execution is serialized using the priv's global
 834  * configuration lock.
 835  */
 836 static void
 837 mlx5e_update_stats_locked(struct mlx5e_priv *priv)
 838 {
 839         struct mlx5_core_dev *mdev = priv->mdev;
 840         struct mlx5e_vport_stats *s = &priv->stats.vport;
 841         struct mlx5e_sq_stats *sq_stats;
 842         struct buf_ring *sq_br;
 843 #if (__FreeBSD_version < 1100000)
 844         struct ifnet *ifp = priv->ifp;
 845 #endif
 846
 847         u32 in[MLX5_ST_SZ_DW(query_vport_counter_in)];
 848         u32 *out;
 849         int outlen = MLX5_ST_SZ_BYTES(query_vport_counter_out);
 850         u64 tso_packets = 0;
 851         u64 tso_bytes = 0;
 852         u64 tx_queue_dropped = 0;
 853         u64 tx_defragged = 0;
 854         u64 tx_offload_none = 0;
 855         u64 lro_packets = 0;
 856         u64 lro_bytes = 0;
 857         u64 sw_lro_queued = 0;
 858         u64 sw_lro_flushed = 0;
 859         u64 rx_csum_none = 0;
 860         u64 rx_wqe_err = 0;
 861         u64 rx_packets = 0;
 862         u64 rx_bytes = 0;
 863         u32 rx_out_of_buffer = 0;
 864         int i;
 865         int j;
 866
 867         out = mlx5_vzalloc(outlen);
 868         if (out == NULL)
 869                 goto free_out;
 870
 871         /* Collect firts the SW counters and then HW for consistency */
 872         for (i = 0; i < priv->params.num_channels; i++) {
 873                 struct mlx5e_channel *pch = priv->channel + i;
 874                 struct mlx5e_rq *rq = &pch->rq;
 875                 struct mlx5e_rq_stats *rq_stats = &pch->rq.stats;
 876
 877                 /* collect stats from LRO */
 878                 rq_stats->sw_lro_queued = rq->lro.lro_queued;
 879                 rq_stats->sw_lro_flushed = rq->lro.lro_flushed;
 880                 sw_lro_queued += rq_stats->sw_lro_queued;
 881                 sw_lro_flushed += rq_stats->sw_lro_flushed;
 882                 lro_packets += rq_stats->lro_packets;
 883                 lro_bytes += rq_stats->lro_bytes;
 884                 rx_csum_none += rq_stats->csum_none;
 885                 rx_wqe_err += rq_stats->wqe_err;
 886                 rx_packets += rq_stats->packets;
 887                 rx_bytes += rq_stats->bytes;
 888
 889                 for (j = 0; j < priv->num_tc; j++) {
 890                         sq_stats = &pch->sq[j].stats;
 891                         sq_br = pch->sq[j].br;
 892
 893                         tso_packets += sq_stats->tso_packets;
 894                         tso_bytes += sq_stats->tso_bytes;
 895                         tx_queue_dropped += sq_stats->dropped;
 896                         if (sq_br != NULL)
 897                                 tx_queue_dropped += sq_br->br_drops;
 898                         tx_defragged += sq_stats->defragged;
 899                         tx_offload_none += sq_stats->csum_offload_none;
 900                 }
 901         }
 902
 903         /* update counters */
 904         s->tso_packets = tso_packets;
 905         s->tso_bytes = tso_bytes;
 906         s->tx_queue_dropped = tx_queue_dropped;
 907         s->tx_defragged = tx_defragged;
 908         s->lro_packets = lro_packets;
 909         s->lro_bytes = lro_bytes;
 910         s->sw_lro_queued = sw_lro_queued;
 911         s->sw_lro_flushed = sw_lro_flushed;
 912         s->rx_csum_none = rx_csum_none;
 913         s->rx_wqe_err = rx_wqe_err;
 914         s->rx_packets = rx_packets;
 915         s->rx_bytes = rx_bytes;
 916
 917         mlx5e_grp_vnic_env_update_stats(priv);
 918
 919         /* HW counters */
 920         memset(in, 0, sizeof(in));
 921
 922         MLX5_SET(query_vport_counter_in, in, opcode,
 923             MLX5_CMD_OP_QUERY_VPORT_COUNTER);
 924         MLX5_SET(query_vport_counter_in, in, op_mod, 0);
 925         MLX5_SET(query_vport_counter_in, in, other_vport, 0);
 926
 927         memset(out, 0, outlen);
 928
 929         /* get number of out-of-buffer drops first */
 930         if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0 &&
 931             mlx5_vport_query_out_of_rx_buffer(mdev, priv->counter_set_id,
 932             &rx_out_of_buffer) == 0) {
 933                 s->rx_out_of_buffer = rx_out_of_buffer;
 934         }
 935
 936         /* get port statistics */
 937         if (mlx5_cmd_exec(mdev, in, sizeof(in), out, outlen) == 0) {
 938 #define MLX5_GET_CTR(out, x) \
 939         MLX5_GET64(query_vport_counter_out, out, x)
 940
 941                 s->rx_error_packets =
 942                     MLX5_GET_CTR(out, received_errors.packets);
 943                 s->rx_error_bytes =
 944                     MLX5_GET_CTR(out, received_errors.octets);
 945                 s->tx_error_packets =
 946                     MLX5_GET_CTR(out, transmit_errors.packets);
 947                 s->tx_error_bytes =
 948                     MLX5_GET_CTR(out, transmit_errors.octets);
 949
 950                 s->rx_unicast_packets =
 951                     MLX5_GET_CTR(out, received_eth_unicast.packets);
 952                 s->rx_unicast_bytes =
 953                     MLX5_GET_CTR(out, received_eth_unicast.octets);
 954                 s->tx_unicast_packets =
 955                     MLX5_GET_CTR(out, transmitted_eth_unicast.packets);
 956                 s->tx_unicast_bytes =
 957                     MLX5_GET_CTR(out, transmitted_eth_unicast.octets);
 958
 959                 s->rx_multicast_packets =
 960                     MLX5_GET_CTR(out, received_eth_multicast.packets);
 961                 s->rx_multicast_bytes =
 962                     MLX5_GET_CTR(out, received_eth_multicast.octets);
 963                 s->tx_multicast_packets =
 964                     MLX5_GET_CTR(out, transmitted_eth_multicast.packets);
 965                 s->tx_multicast_bytes =
 966                     MLX5_GET_CTR(out, transmitted_eth_multicast.octets);
 967
 968                 s->rx_broadcast_packets =
 969                     MLX5_GET_CTR(out, received_eth_broadcast.packets);
 970                 s->rx_broadcast_bytes =
 971                     MLX5_GET_CTR(out, received_eth_broadcast.octets);
 972                 s->tx_broadcast_packets =
 973                     MLX5_GET_CTR(out, transmitted_eth_broadcast.packets);
 974                 s->tx_broadcast_bytes =
 975                     MLX5_GET_CTR(out, transmitted_eth_broadcast.octets);
 976
 977                 s->tx_packets = s->tx_unicast_packets +
 978                     s->tx_multicast_packets + s->tx_broadcast_packets;
 979                 s->tx_bytes = s->tx_unicast_bytes + s->tx_multicast_bytes +
 980                     s->tx_broadcast_bytes;
 981
 982                 /* Update calculated offload counters */
 983                 s->tx_csum_offload = s->tx_packets - tx_offload_none;
 984                 s->rx_csum_good = s->rx_packets - s->rx_csum_none;
 985         }
 986
 987         /* Get physical port counters */
 988         mlx5e_update_pport_counters(priv);
 989
 990         s->tx_jumbo_packets =
 991             priv->stats.port_stats_debug.tx_stat_p1519to2047octets +
 992             priv->stats.port_stats_debug.tx_stat_p2048to4095octets +
 993             priv->stats.port_stats_debug.tx_stat_p4096to8191octets +
 994             priv->stats.port_stats_debug.tx_stat_p8192to10239octets;
 995
 996 #if (__FreeBSD_version < 1100000)
 997         /* no get_counters interface in fbsd 10 */
 998         ifp->if_ipackets = s->rx_packets;
 999         ifp->if_ierrors = priv->stats.pport.in_range_len_errors +
1000             priv->stats.pport.out_of_range_len +
1001             priv->stats.pport.too_long_errors +
1002             priv->stats.pport.check_seq_err +
1003             priv->stats.pport.alignment_err;
1004         ifp->if_iqdrops = s->rx_out_of_buffer;
1005         ifp->if_opackets = s->tx_packets;
1006         ifp->if_oerrors = priv->stats.port_stats_debug.out_discards;
1007         ifp->if_snd.ifq_drops = s->tx_queue_dropped;
1008         ifp->if_ibytes = s->rx_bytes;
1009         ifp->if_obytes = s->tx_bytes;
1010         ifp->if_collisions =
1011             priv->stats.pport.collisions;
1012 #endif
1013
1014 free_out:
1015         kvfree(out);
1016
1017         /* Update diagnostics, if any */
1018         if (priv->params_ethtool.diag_pci_enable ||
1019             priv->params_ethtool.diag_general_enable) {
1020                 int error = mlx5_core_get_diagnostics_full(mdev,
1021                     priv->params_ethtool.diag_pci_enable ? &priv->params_pci : NULL,
1022                     priv->params_ethtool.diag_general_enable ? &priv->params_general : NULL);
1023                 if (error != 0)
1024                         if_printf(priv->ifp, "Failed reading diagnostics: %d\n", error);
1025         }
1026 }
1027
1028 static void
1029 mlx5e_update_stats_work(struct work_struct *work)
1030 {
1031         struct mlx5e_priv *priv;
1032
1033         priv  = container_of(work, struct mlx5e_priv, update_stats_work);
1034         PRIV_LOCK(priv);
1035         if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0)
1036                 mlx5e_update_stats_locked(priv);
1037         PRIV_UNLOCK(priv);
1038 }
1039
1040 static void
1041 mlx5e_update_stats(void *arg)
1042 {
1043         struct mlx5e_priv *priv = arg;
1044
1045         queue_work(priv->wq, &priv->update_stats_work);
1046
1047         callout_reset(&priv->watchdog, hz, &mlx5e_update_stats, priv);
1048 }
1049
1050 static void
1051 mlx5e_async_event_sub(struct mlx5e_priv *priv,
1052     enum mlx5_dev_event event)
1053 {
1054         switch (event) {
1055         case MLX5_DEV_EVENT_PORT_UP:
1056         case MLX5_DEV_EVENT_PORT_DOWN:
1057                 queue_work(priv->wq, &priv->update_carrier_work);
1058                 break;
1059
1060         default:
1061                 break;
1062         }
1063 }
1064
1065 static void
1066 mlx5e_async_event(struct mlx5_core_dev *mdev, void *vpriv,
1067     enum mlx5_dev_event event, unsigned long param)
1068 {
1069         struct mlx5e_priv *priv = vpriv;
1070
1071         mtx_lock(&priv->async_events_mtx);
1072         if (test_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state))
1073                 mlx5e_async_event_sub(priv, event);
1074         mtx_unlock(&priv->async_events_mtx);
1075 }
1076
1077 static void
1078 mlx5e_enable_async_events(struct mlx5e_priv *priv)
1079 {
1080         set_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
1081 }
1082
1083 static void
1084 mlx5e_disable_async_events(struct mlx5e_priv *priv)
1085 {
1086         mtx_lock(&priv->async_events_mtx);
1087         clear_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
1088         mtx_unlock(&priv->async_events_mtx);
1089 }
1090
1091 static void mlx5e_calibration_callout(void *arg);
1092 static int mlx5e_calibration_duration = 20;
1093 static int mlx5e_fast_calibration = 1;
1094 static int mlx5e_normal_calibration = 30;
1095
1096 static SYSCTL_NODE(_hw_mlx5, OID_AUTO, calibr, CTLFLAG_RW, 0,
1097     "MLX5 timestamp calibration parameteres");
1098
1099 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, duration, CTLFLAG_RWTUN,
1100     &mlx5e_calibration_duration, 0,
1101     "Duration of initial calibration");
1102 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, fast, CTLFLAG_RWTUN,
1103     &mlx5e_fast_calibration, 0,
1104     "Recalibration interval during initial calibration");
1105 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, normal, CTLFLAG_RWTUN,
1106     &mlx5e_normal_calibration, 0,
1107     "Recalibration interval during normal operations");
1108
1109 /*
1110  * Ignites the calibration process.
1111  */
1112 static void
1113 mlx5e_reset_calibration_callout(struct mlx5e_priv *priv)
1114 {
1115
1116         if (priv->clbr_done == 0)
1117                 mlx5e_calibration_callout(priv);
1118         else
1119                 callout_reset_curcpu(&priv->tstmp_clbr, (priv->clbr_done <
1120                     mlx5e_calibration_duration ? mlx5e_fast_calibration :
1121                     mlx5e_normal_calibration) * hz, mlx5e_calibration_callout,
1122                     priv);
1123 }
1124
1125 static uint64_t
1126 mlx5e_timespec2usec(const struct timespec *ts)
1127 {
1128
1129         return ((uint64_t)ts->tv_sec * 1000000000 + ts->tv_nsec);
1130 }
1131
1132 static uint64_t
1133 mlx5e_hw_clock(struct mlx5e_priv *priv)
1134 {
1135         struct mlx5_init_seg *iseg;
1136         uint32_t hw_h, hw_h1, hw_l;
1137
1138         iseg = priv->mdev->iseg;
1139         do {
1140                 hw_h = ioread32be(&iseg->internal_timer_h);
1141                 hw_l = ioread32be(&iseg->internal_timer_l);
1142                 hw_h1 = ioread32be(&iseg->internal_timer_h);
1143         } while (hw_h1 != hw_h);
1144         return (((uint64_t)hw_h << 32) | hw_l);
1145 }
1146
1147 /*
1148  * The calibration callout, it runs either in the context of the
1149  * thread which enables calibration, or in callout.  It takes the
1150  * snapshot of system and adapter clocks, then advances the pointers to
1151  * the calibration point to allow rx path to read the consistent data
1152  * lockless.
1153  */
1154 static void
1155 mlx5e_calibration_callout(void *arg)
1156 {
1157         struct mlx5e_priv *priv;
1158         struct mlx5e_clbr_point *next, *curr;
1159         struct timespec ts;
1160         int clbr_curr_next;
1161
1162         priv = arg;
1163         curr = &priv->clbr_points[priv->clbr_curr];
1164         clbr_curr_next = priv->clbr_curr + 1;
1165         if (clbr_curr_next >= nitems(priv->clbr_points))
1166                 clbr_curr_next = 0;
1167         next = &priv->clbr_points[clbr_curr_next];
1168
1169         next->base_prev = curr->base_curr;
1170         next->clbr_hw_prev = curr->clbr_hw_curr;
1171
1172         next->clbr_hw_curr = mlx5e_hw_clock(priv);
1173         if (((next->clbr_hw_curr - curr->clbr_hw_curr) >> MLX5E_TSTMP_PREC) ==
1174             0) {
1175                 if (priv->clbr_done != 0) {
1176                         if_printf(priv->ifp, "HW failed tstmp frozen %#jx %#jx,"
1177                             "disabling\n",
1178                              next->clbr_hw_curr, curr->clbr_hw_prev);
1179                         priv->clbr_done = 0;
1180                 }
1181                 atomic_store_rel_int(&curr->clbr_gen, 0);
1182                 return;
1183         }
1184
1185         nanouptime(&ts);
1186         next->base_curr = mlx5e_timespec2usec(&ts);
1187
1188         curr->clbr_gen = 0;
1189         atomic_thread_fence_rel();
1190         priv->clbr_curr = clbr_curr_next;
1191         atomic_store_rel_int(&next->clbr_gen, ++(priv->clbr_gen));
1192
1193         if (priv->clbr_done < mlx5e_calibration_duration)
1194                 priv->clbr_done++;
1195         mlx5e_reset_calibration_callout(priv);
1196 }
1197
1198 static const char *mlx5e_rq_stats_desc[] = {
1199         MLX5E_RQ_STATS(MLX5E_STATS_DESC)
1200 };
1201
1202 static int
1203 mlx5e_create_rq(struct mlx5e_channel *c,
1204     struct mlx5e_rq_param *param,
1205     struct mlx5e_rq *rq)
1206 {
1207         struct mlx5e_priv *priv = c->priv;
1208         struct mlx5_core_dev *mdev = priv->mdev;
1209         char buffer[16];
1210         void *rqc = param->rqc;
1211         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
1212         int wq_sz;
1213         int err;
1214         int i;
1215         u32 nsegs, wqe_sz;
1216
1217         err = mlx5e_get_wqe_sz(priv, &wqe_sz, &nsegs);
1218         if (err != 0)
1219                 goto done;
1220
1221         /* Create DMA descriptor TAG */
1222         if ((err = -bus_dma_tag_create(
1223             bus_get_dma_tag(mdev->pdev->dev.bsddev),
1224             1,                          /* any alignment */
1225             0,                          /* no boundary */
1226             BUS_SPACE_MAXADDR,          /* lowaddr */
1227             BUS_SPACE_MAXADDR,          /* highaddr */
1228             NULL, NULL,                 /* filter, filterarg */
1229             nsegs * MLX5E_MAX_RX_BYTES, /* maxsize */
1230             nsegs,                      /* nsegments */
1231             nsegs * MLX5E_MAX_RX_BYTES, /* maxsegsize */
1232             0,                          /* flags */
1233             NULL, NULL,                 /* lockfunc, lockfuncarg */
1234             &rq->dma_tag)))
1235                 goto done;
1236
1237         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
1238             &rq->wq_ctrl);
1239         if (err)
1240                 goto err_free_dma_tag;
1241
1242         rq->wq.db = &rq->wq.db[MLX5_RCV_DBR];
1243
1244         err = mlx5e_get_wqe_sz(priv, &rq->wqe_sz, &rq->nsegs);
1245         if (err != 0)
1246                 goto err_rq_wq_destroy;
1247
1248         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
1249
1250         err = -tcp_lro_init_args(&rq->lro, priv->ifp, TCP_LRO_ENTRIES, wq_sz);
1251         if (err)
1252                 goto err_rq_wq_destroy;
1253
1254         rq->mbuf = malloc(wq_sz * sizeof(rq->mbuf[0]), M_MLX5EN, M_WAITOK | M_ZERO);
1255         for (i = 0; i != wq_sz; i++) {
1256                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
1257                 int j;
1258
1259                 err = -bus_dmamap_create(rq->dma_tag, 0, &rq->mbuf[i].dma_map);
1260                 if (err != 0) {
1261                         while (i--)
1262                                 bus_dmamap_destroy(rq->dma_tag, rq->mbuf[i].dma_map);
1263                         goto err_rq_mbuf_free;
1264                 }
1265
1266                 /* set value for constant fields */
1267                 for (j = 0; j < rq->nsegs; j++)
1268                         wqe->data[j].lkey = c->mkey_be;
1269         }
1270
1271         INIT_WORK(&rq->dim.work, mlx5e_dim_work);
1272         if (priv->params.rx_cq_moderation_mode < 2) {
1273                 rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_DISABLED;
1274         } else {
1275                 void *cqc = container_of(param,
1276                     struct mlx5e_channel_param, rq)->rx_cq.cqc;
1277
1278                 switch (MLX5_GET(cqc, cqc, cq_period_mode)) {
1279                 case MLX5_CQ_PERIOD_MODE_START_FROM_EQE:
1280                         rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
1281                         break;
1282                 case MLX5_CQ_PERIOD_MODE_START_FROM_CQE:
1283                         rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE;
1284                         break;
1285                 default:
1286                         rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_DISABLED;
1287                         break;
1288                 }
1289         }
1290
1291         rq->ifp = priv->ifp;
1292         rq->channel = c;
1293         rq->ix = c->ix;
1294
1295         snprintf(buffer, sizeof(buffer), "rxstat%d", c->ix);
1296         mlx5e_create_stats(&rq->stats.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
1297             buffer, mlx5e_rq_stats_desc, MLX5E_RQ_STATS_NUM,
1298             rq->stats.arg);
1299         return (0);
1300
1301 err_rq_mbuf_free:
1302         free(rq->mbuf, M_MLX5EN);
1303         tcp_lro_free(&rq->lro);
1304 err_rq_wq_destroy:
1305         mlx5_wq_destroy(&rq->wq_ctrl);
1306 err_free_dma_tag:
1307         bus_dma_tag_destroy(rq->dma_tag);
1308 done:
1309         return (err);
1310 }
1311
1312 static void
1313 mlx5e_destroy_rq(struct mlx5e_rq *rq)
1314 {
1315         int wq_sz;
1316         int i;
1317
1318         /* destroy all sysctl nodes */
1319         sysctl_ctx_free(&rq->stats.ctx);
1320
1321         /* free leftover LRO packets, if any */
1322         tcp_lro_free(&rq->lro);
1323
1324         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
1325         for (i = 0; i != wq_sz; i++) {
1326                 if (rq->mbuf[i].mbuf != NULL) {
1327                         bus_dmamap_unload(rq->dma_tag, rq->mbuf[i].dma_map);
1328                         m_freem(rq->mbuf[i].mbuf);
1329                 }
1330                 bus_dmamap_destroy(rq->dma_tag, rq->mbuf[i].dma_map);
1331         }
1332         free(rq->mbuf, M_MLX5EN);
1333         mlx5_wq_destroy(&rq->wq_ctrl);
1334 }
1335
1336 static int
1337 mlx5e_enable_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
1338 {
1339         struct mlx5e_channel *c = rq->channel;
1340         struct mlx5e_priv *priv = c->priv;
1341         struct mlx5_core_dev *mdev = priv->mdev;
1342
1343         void *in;
1344         void *rqc;
1345         void *wq;
1346         int inlen;
1347         int err;
1348
1349         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
1350             sizeof(u64) * rq->wq_ctrl.buf.npages;
1351         in = mlx5_vzalloc(inlen);
1352         if (in == NULL)
1353                 return (-ENOMEM);
1354
1355         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
1356         wq = MLX5_ADDR_OF(rqc, rqc, wq);
1357
1358         memcpy(rqc, param->rqc, sizeof(param->rqc));
1359
1360         MLX5_SET(rqc, rqc, cqn, c->rq.cq.mcq.cqn);
1361         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
1362         MLX5_SET(rqc, rqc, flush_in_error_en, 1);
1363         if (priv->counter_set_id >= 0)
1364                 MLX5_SET(rqc, rqc, counter_set_id, priv->counter_set_id);
1365         MLX5_SET(wq, wq, log_wq_pg_sz, rq->wq_ctrl.buf.page_shift -
1366             PAGE_SHIFT);
1367         MLX5_SET64(wq, wq, dbr_addr, rq->wq_ctrl.db.dma);
1368
1369         mlx5_fill_page_array(&rq->wq_ctrl.buf,
1370             (__be64 *) MLX5_ADDR_OF(wq, wq, pas));
1371
1372         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
1373
1374         kvfree(in);
1375
1376         return (err);
1377 }
1378
1379 static int
1380 mlx5e_modify_rq(struct mlx5e_rq *rq, int curr_state, int next_state)
1381 {
1382         struct mlx5e_channel *c = rq->channel;
1383         struct mlx5e_priv *priv = c->priv;
1384         struct mlx5_core_dev *mdev = priv->mdev;
1385
1386         void *in;
1387         void *rqc;
1388         int inlen;
1389         int err;
1390
1391         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
1392         in = mlx5_vzalloc(inlen);
1393         if (in == NULL)
1394                 return (-ENOMEM);
1395
1396         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
1397
1398         MLX5_SET(modify_rq_in, in, rqn, rq->rqn);
1399         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
1400         MLX5_SET(rqc, rqc, state, next_state);
1401
1402         err = mlx5_core_modify_rq(mdev, in, inlen);
1403
1404         kvfree(in);
1405
1406         return (err);
1407 }
1408
1409 static void
1410 mlx5e_disable_rq(struct mlx5e_rq *rq)
1411 {
1412         struct mlx5e_channel *c = rq->channel;
1413         struct mlx5e_priv *priv = c->priv;
1414         struct mlx5_core_dev *mdev = priv->mdev;
1415
1416         mlx5_core_destroy_rq(mdev, rq->rqn);
1417 }
1418
1419 static int
1420 mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
1421 {
1422         struct mlx5e_channel *c = rq->channel;
1423         struct mlx5e_priv *priv = c->priv;
1424         struct mlx5_wq_ll *wq = &rq->wq;
1425         int i;
1426
1427         for (i = 0; i < 1000; i++) {
1428                 if (wq->cur_sz >= priv->params.min_rx_wqes)
1429                         return (0);
1430
1431                 msleep(4);
1432         }
1433         return (-ETIMEDOUT);
1434 }
1435
1436 static int
1437 mlx5e_open_rq(struct mlx5e_channel *c,
1438     struct mlx5e_rq_param *param,
1439     struct mlx5e_rq *rq)
1440 {
1441         int err;
1442
1443         err = mlx5e_create_rq(c, param, rq);
1444         if (err)
1445                 return (err);
1446
1447         err = mlx5e_enable_rq(rq, param);
1448         if (err)
1449                 goto err_destroy_rq;
1450
1451         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
1452         if (err)
1453                 goto err_disable_rq;
1454
1455         c->rq.enabled = 1;
1456
1457         return (0);
1458
1459 err_disable_rq:
1460         mlx5e_disable_rq(rq);
1461 err_destroy_rq:
1462         mlx5e_destroy_rq(rq);
1463
1464         return (err);
1465 }
1466
1467 static void
1468 mlx5e_close_rq(struct mlx5e_rq *rq)
1469 {
1470         mtx_lock(&rq->mtx);
1471         rq->enabled = 0;
1472         callout_stop(&rq->watchdog);
1473         mtx_unlock(&rq->mtx);
1474
1475         callout_drain(&rq->watchdog);
1476
1477         mlx5e_modify_rq(rq, MLX5_RQC_STATE_RDY, MLX5_RQC_STATE_ERR);
1478 }
1479
1480 static void
1481 mlx5e_close_rq_wait(struct mlx5e_rq *rq)
1482 {
1483
1484         mlx5e_disable_rq(rq);
1485         mlx5e_close_cq(&rq->cq);
1486         cancel_work_sync(&rq->dim.work);
1487         mlx5e_destroy_rq(rq);
1488 }
1489
1490 void
1491 mlx5e_free_sq_db(struct mlx5e_sq *sq)
1492 {
1493         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1494         int x;
1495
1496         for (x = 0; x != wq_sz; x++) {
1497                 if (sq->mbuf[x].mbuf != NULL) {
1498                         bus_dmamap_unload(sq->dma_tag, sq->mbuf[x].dma_map);
1499                         m_freem(sq->mbuf[x].mbuf);
1500                 }
1501                 bus_dmamap_destroy(sq->dma_tag, sq->mbuf[x].dma_map);
1502         }
1503         free(sq->mbuf, M_MLX5EN);
1504 }
1505
1506 int
1507 mlx5e_alloc_sq_db(struct mlx5e_sq *sq)
1508 {
1509         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1510         int err;
1511         int x;
1512
1513         sq->mbuf = malloc(wq_sz * sizeof(sq->mbuf[0]), M_MLX5EN, M_WAITOK | M_ZERO);
1514
1515         /* Create DMA descriptor MAPs */
1516         for (x = 0; x != wq_sz; x++) {
1517                 err = -bus_dmamap_create(sq->dma_tag, 0, &sq->mbuf[x].dma_map);
1518                 if (err != 0) {
1519                         while (x--)
1520                                 bus_dmamap_destroy(sq->dma_tag, sq->mbuf[x].dma_map);
1521                         free(sq->mbuf, M_MLX5EN);
1522                         return (err);
1523                 }
1524         }
1525         return (0);
1526 }
1527
1528 static const char *mlx5e_sq_stats_desc[] = {
1529         MLX5E_SQ_STATS(MLX5E_STATS_DESC)
1530 };
1531
1532 void
1533 mlx5e_update_sq_inline(struct mlx5e_sq *sq)
1534 {
1535         sq->max_inline = sq->priv->params.tx_max_inline;
1536         sq->min_inline_mode = sq->priv->params.tx_min_inline_mode;
1537
1538         /*
1539          * Check if trust state is DSCP or if inline mode is NONE which
1540          * indicates CX-5 or newer hardware.
1541          */
1542         if (sq->priv->params_ethtool.trust_state != MLX5_QPTS_TRUST_PCP ||
1543             sq->min_inline_mode == MLX5_INLINE_MODE_NONE) {
1544                 if (MLX5_CAP_ETH(sq->priv->mdev, wqe_vlan_insert))
1545                         sq->min_insert_caps = MLX5E_INSERT_VLAN | MLX5E_INSERT_NON_VLAN;
1546                 else
1547                         sq->min_insert_caps = MLX5E_INSERT_NON_VLAN;
1548         } else {
1549                 sq->min_insert_caps = 0;
1550         }
1551 }
1552
1553 static void
1554 mlx5e_refresh_sq_inline_sub(struct mlx5e_priv *priv, struct mlx5e_channel *c)
1555 {
1556         int i;
1557
1558         for (i = 0; i != c->num_tc; i++) {
1559                 mtx_lock(&c->sq[i].lock);
1560                 mlx5e_update_sq_inline(&c->sq[i]);
1561                 mtx_unlock(&c->sq[i].lock);
1562         }
1563 }
1564
1565 void
1566 mlx5e_refresh_sq_inline(struct mlx5e_priv *priv)
1567 {
1568         int i;
1569
1570         /* check if channels are closed */
1571         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
1572                 return;
1573
1574         for (i = 0; i < priv->params.num_channels; i++)
1575                 mlx5e_refresh_sq_inline_sub(priv, &priv->channel[i]);
1576 }
1577
1578 static int
1579 mlx5e_create_sq(struct mlx5e_channel *c,
1580     int tc,
1581     struct mlx5e_sq_param *param,
1582     struct mlx5e_sq *sq)
1583 {
1584         struct mlx5e_priv *priv = c->priv;
1585         struct mlx5_core_dev *mdev = priv->mdev;
1586         char buffer[16];
1587         void *sqc = param->sqc;
1588         void *sqc_wq = MLX5_ADDR_OF(sqc, sqc, wq);
1589         int err;
1590
1591         /* Create DMA descriptor TAG */
1592         if ((err = -bus_dma_tag_create(
1593             bus_get_dma_tag(mdev->pdev->dev.bsddev),
1594             1,                          /* any alignment */
1595             0,                          /* no boundary */
1596             BUS_SPACE_MAXADDR,          /* lowaddr */
1597             BUS_SPACE_MAXADDR,          /* highaddr */
1598             NULL, NULL,                 /* filter, filterarg */
1599             MLX5E_MAX_TX_PAYLOAD_SIZE,  /* maxsize */
1600             MLX5E_MAX_TX_MBUF_FRAGS,    /* nsegments */
1601             MLX5E_MAX_TX_MBUF_SIZE,     /* maxsegsize */
1602             0,                          /* flags */
1603             NULL, NULL,                 /* lockfunc, lockfuncarg */
1604             &sq->dma_tag)))
1605                 goto done;
1606
1607         err = mlx5_alloc_map_uar(mdev, &sq->uar);
1608         if (err)
1609                 goto err_free_dma_tag;
1610
1611         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq,
1612             &sq->wq_ctrl);
1613         if (err)
1614                 goto err_unmap_free_uar;
1615
1616         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1617         sq->bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
1618
1619         err = mlx5e_alloc_sq_db(sq);
1620         if (err)
1621                 goto err_sq_wq_destroy;
1622
1623         sq->mkey_be = c->mkey_be;
1624         sq->ifp = priv->ifp;
1625         sq->priv = priv;
1626         sq->tc = tc;
1627
1628         mlx5e_update_sq_inline(sq);
1629
1630         snprintf(buffer, sizeof(buffer), "txstat%dtc%d", c->ix, tc);
1631         mlx5e_create_stats(&sq->stats.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
1632             buffer, mlx5e_sq_stats_desc, MLX5E_SQ_STATS_NUM,
1633             sq->stats.arg);
1634
1635         return (0);
1636
1637 err_sq_wq_destroy:
1638         mlx5_wq_destroy(&sq->wq_ctrl);
1639
1640 err_unmap_free_uar:
1641         mlx5_unmap_free_uar(mdev, &sq->uar);
1642
1643 err_free_dma_tag:
1644         bus_dma_tag_destroy(sq->dma_tag);
1645 done:
1646         return (err);
1647 }
1648
1649 static void
1650 mlx5e_destroy_sq(struct mlx5e_sq *sq)
1651 {
1652         /* destroy all sysctl nodes */
1653         sysctl_ctx_free(&sq->stats.ctx);
1654
1655         mlx5e_free_sq_db(sq);
1656         mlx5_wq_destroy(&sq->wq_ctrl);
1657         mlx5_unmap_free_uar(sq->priv->mdev, &sq->uar);
1658 }
1659
1660 int
1661 mlx5e_enable_sq(struct mlx5e_sq *sq, struct mlx5e_sq_param *param,
1662     int tis_num)
1663 {
1664         void *in;
1665         void *sqc;
1666         void *wq;
1667         int inlen;
1668         int err;
1669
1670         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1671             sizeof(u64) * sq->wq_ctrl.buf.npages;
1672         in = mlx5_vzalloc(inlen);
1673         if (in == NULL)
1674                 return (-ENOMEM);
1675
1676         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1677         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1678
1679         memcpy(sqc, param->sqc, sizeof(param->sqc));
1680
1681         MLX5_SET(sqc, sqc, tis_num_0, tis_num);
1682         MLX5_SET(sqc, sqc, cqn, sq->cq.mcq.cqn);
1683         MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1684         MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1685         MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1686
1687         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1688         MLX5_SET(wq, wq, uar_page, sq->uar.index);
1689         MLX5_SET(wq, wq, log_wq_pg_sz, sq->wq_ctrl.buf.page_shift -
1690             PAGE_SHIFT);
1691         MLX5_SET64(wq, wq, dbr_addr, sq->wq_ctrl.db.dma);
1692
1693         mlx5_fill_page_array(&sq->wq_ctrl.buf,
1694             (__be64 *) MLX5_ADDR_OF(wq, wq, pas));
1695
1696         err = mlx5_core_create_sq(sq->priv->mdev, in, inlen, &sq->sqn);
1697
1698         kvfree(in);
1699
1700         return (err);
1701 }
1702
1703 int
1704 mlx5e_modify_sq(struct mlx5e_sq *sq, int curr_state, int next_state)
1705 {
1706         void *in;
1707         void *sqc;
1708         int inlen;
1709         int err;
1710
1711         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1712         in = mlx5_vzalloc(inlen);
1713         if (in == NULL)
1714                 return (-ENOMEM);
1715
1716         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1717
1718         MLX5_SET(modify_sq_in, in, sqn, sq->sqn);
1719         MLX5_SET(modify_sq_in, in, sq_state, curr_state);
1720         MLX5_SET(sqc, sqc, state, next_state);
1721
1722         err = mlx5_core_modify_sq(sq->priv->mdev, in, inlen);
1723
1724         kvfree(in);
1725
1726         return (err);
1727 }
1728
1729 void
1730 mlx5e_disable_sq(struct mlx5e_sq *sq)
1731 {
1732
1733         mlx5_core_destroy_sq(sq->priv->mdev, sq->sqn);
1734 }
1735
1736 static int
1737 mlx5e_open_sq(struct mlx5e_channel *c,
1738     int tc,
1739     struct mlx5e_sq_param *param,
1740     struct mlx5e_sq *sq)
1741 {
1742         int err;
1743
1744         err = mlx5e_create_sq(c, tc, param, sq);
1745         if (err)
1746                 return (err);
1747
1748         err = mlx5e_enable_sq(sq, param, c->priv->tisn[tc]);
1749         if (err)
1750                 goto err_destroy_sq;
1751
1752         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST, MLX5_SQC_STATE_RDY);
1753         if (err)
1754                 goto err_disable_sq;
1755
1756         WRITE_ONCE(sq->running, 1);
1757
1758         return (0);
1759
1760 err_disable_sq:
1761         mlx5e_disable_sq(sq);
1762 err_destroy_sq:
1763         mlx5e_destroy_sq(sq);
1764
1765         return (err);
1766 }
1767
1768 static void
1769 mlx5e_sq_send_nops_locked(struct mlx5e_sq *sq, int can_sleep)
1770 {
1771         /* fill up remainder with NOPs */
1772         while (sq->cev_counter != 0) {
1773                 while (!mlx5e_sq_has_room_for(sq, 1)) {
1774                         if (can_sleep != 0) {
1775                                 mtx_unlock(&sq->lock);
1776                                 msleep(4);
1777                                 mtx_lock(&sq->lock);
1778                         } else {
1779                                 goto done;
1780                         }
1781                 }
1782                 /* send a single NOP */
1783                 mlx5e_send_nop(sq, 1);
1784                 atomic_thread_fence_rel();
1785         }
1786 done:
1787         /* Check if we need to write the doorbell */
1788         if (likely(sq->doorbell.d64 != 0)) {
1789                 mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
1790                 sq->doorbell.d64 = 0;
1791         }
1792 }
1793
1794 void
1795 mlx5e_sq_cev_timeout(void *arg)
1796 {
1797         struct mlx5e_sq *sq = arg;
1798
1799         mtx_assert(&sq->lock, MA_OWNED);
1800
1801         /* check next state */
1802         switch (sq->cev_next_state) {
1803         case MLX5E_CEV_STATE_SEND_NOPS:
1804                 /* fill TX ring with NOPs, if any */
1805                 mlx5e_sq_send_nops_locked(sq, 0);
1806
1807                 /* check if completed */
1808                 if (sq->cev_counter == 0) {
1809                         sq->cev_next_state = MLX5E_CEV_STATE_INITIAL;
1810                         return;
1811                 }
1812                 break;
1813         default:
1814                 /* send NOPs on next timeout */
1815                 sq->cev_next_state = MLX5E_CEV_STATE_SEND_NOPS;
1816                 break;
1817         }
1818
1819         /* restart timer */
1820         callout_reset_curcpu(&sq->cev_callout, hz, mlx5e_sq_cev_timeout, sq);
1821 }
1822
1823 void
1824 mlx5e_drain_sq(struct mlx5e_sq *sq)
1825 {
1826         int error;
1827         struct mlx5_core_dev *mdev= sq->priv->mdev;
1828
1829         /*
1830          * Check if already stopped.
1831          *
1832          * NOTE: Serialization of this function is managed by the
1833          * caller ensuring the priv's state lock is locked or in case
1834          * of rate limit support, a single thread manages drain and
1835          * resume of SQs. The "running" variable can therefore safely
1836          * be read without any locks.
1837          */
1838         if (READ_ONCE(sq->running) == 0)
1839                 return;
1840
1841         /* don't put more packets into the SQ */
1842         WRITE_ONCE(sq->running, 0);
1843
1844         /* serialize access to DMA rings */
1845         mtx_lock(&sq->lock);
1846
1847         /* teardown event factor timer, if any */
1848         sq->cev_next_state = MLX5E_CEV_STATE_HOLD_NOPS;
1849         callout_stop(&sq->cev_callout);
1850
1851         /* send dummy NOPs in order to flush the transmit ring */
1852         mlx5e_sq_send_nops_locked(sq, 1);
1853         mtx_unlock(&sq->lock);
1854
1855         /* make sure it is safe to free the callout */
1856         callout_drain(&sq->cev_callout);
1857
1858         /* wait till SQ is empty or link is down */
1859         mtx_lock(&sq->lock);
1860         while (sq->cc != sq->pc &&
1861             (sq->priv->media_status_last & IFM_ACTIVE) != 0 &&
1862             mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
1863                 mtx_unlock(&sq->lock);
1864                 msleep(1);
1865                 sq->cq.mcq.comp(&sq->cq.mcq);
1866                 mtx_lock(&sq->lock);
1867         }
1868         mtx_unlock(&sq->lock);
1869
1870         /* error out remaining requests */
1871         error = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RDY, MLX5_SQC_STATE_ERR);
1872         if (error != 0) {
1873                 if_printf(sq->ifp,
1874                     "mlx5e_modify_sq() from RDY to ERR failed: %d\n", error);
1875         }
1876
1877         /* wait till SQ is empty */
1878         mtx_lock(&sq->lock);
1879         while (sq->cc != sq->pc &&
1880                mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
1881                 mtx_unlock(&sq->lock);
1882                 msleep(1);
1883                 sq->cq.mcq.comp(&sq->cq.mcq);
1884                 mtx_lock(&sq->lock);
1885         }
1886         mtx_unlock(&sq->lock);
1887 }
1888
1889 static void
1890 mlx5e_close_sq_wait(struct mlx5e_sq *sq)
1891 {
1892
1893         mlx5e_drain_sq(sq);
1894         mlx5e_disable_sq(sq);
1895         mlx5e_destroy_sq(sq);
1896 }
1897
1898 static int
1899 mlx5e_create_cq(struct mlx5e_priv *priv,
1900     struct mlx5e_cq_param *param,
1901     struct mlx5e_cq *cq,
1902     mlx5e_cq_comp_t *comp,
1903     int eq_ix)
1904 {
1905         struct mlx5_core_dev *mdev = priv->mdev;
1906         struct mlx5_core_cq *mcq = &cq->mcq;
1907         int eqn_not_used;
1908         int irqn;
1909         int err;
1910         u32 i;
1911
1912         param->wq.buf_numa_node = 0;
1913         param->wq.db_numa_node = 0;
1914
1915         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1916             &cq->wq_ctrl);
1917         if (err)
1918                 return (err);
1919
1920         mlx5_vector2eqn(mdev, eq_ix, &eqn_not_used, &irqn);
1921
1922         mcq->cqe_sz = 64;
1923         mcq->set_ci_db = cq->wq_ctrl.db.db;
1924         mcq->arm_db = cq->wq_ctrl.db.db + 1;
1925         *mcq->set_ci_db = 0;
1926         *mcq->arm_db = 0;
1927         mcq->vector = eq_ix;
1928         mcq->comp = comp;
1929         mcq->event = mlx5e_cq_error_event;
1930         mcq->irqn = irqn;
1931         mcq->uar = &priv->cq_uar;
1932
1933         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1934                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1935
1936                 cqe->op_own = 0xf1;
1937         }
1938
1939         cq->priv = priv;
1940
1941         return (0);
1942 }
1943
1944 static void
1945 mlx5e_destroy_cq(struct mlx5e_cq *cq)
1946 {
1947         mlx5_wq_destroy(&cq->wq_ctrl);
1948 }
1949
1950 static int
1951 mlx5e_enable_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param, int eq_ix)
1952 {
1953         struct mlx5_core_cq *mcq = &cq->mcq;
1954         void *in;
1955         void *cqc;
1956         int inlen;
1957         int irqn_not_used;
1958         int eqn;
1959         int err;
1960
1961         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1962             sizeof(u64) * cq->wq_ctrl.buf.npages;
1963         in = mlx5_vzalloc(inlen);
1964         if (in == NULL)
1965                 return (-ENOMEM);
1966
1967         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1968
1969         memcpy(cqc, param->cqc, sizeof(param->cqc));
1970
1971         mlx5_fill_page_array(&cq->wq_ctrl.buf,
1972             (__be64 *) MLX5_ADDR_OF(create_cq_in, in, pas));
1973
1974         mlx5_vector2eqn(cq->priv->mdev, eq_ix, &eqn, &irqn_not_used);
1975
1976         MLX5_SET(cqc, cqc, c_eqn, eqn);
1977         MLX5_SET(cqc, cqc, uar_page, mcq->uar->index);
1978         MLX5_SET(cqc, cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
1979             PAGE_SHIFT);
1980         MLX5_SET64(cqc, cqc, dbr_addr, cq->wq_ctrl.db.dma);
1981
1982         err = mlx5_core_create_cq(cq->priv->mdev, mcq, in, inlen);
1983
1984         kvfree(in);
1985
1986         if (err)
1987                 return (err);
1988
1989         mlx5e_cq_arm(cq, MLX5_GET_DOORBELL_LOCK(&cq->priv->doorbell_lock));
1990
1991         return (0);
1992 }
1993
1994 static void
1995 mlx5e_disable_cq(struct mlx5e_cq *cq)
1996 {
1997
1998         mlx5_core_destroy_cq(cq->priv->mdev, &cq->mcq);
1999 }
2000
2001 int
2002 mlx5e_open_cq(struct mlx5e_priv *priv,
2003     struct mlx5e_cq_param *param,
2004     struct mlx5e_cq *cq,
2005     mlx5e_cq_comp_t *comp,
2006     int eq_ix)
2007 {
2008         int err;
2009
2010         err = mlx5e_create_cq(priv, param, cq, comp, eq_ix);
2011         if (err)
2012                 return (err);
2013
2014         err = mlx5e_enable_cq(cq, param, eq_ix);
2015         if (err)
2016                 goto err_destroy_cq;
2017
2018         return (0);
2019
2020 err_destroy_cq:
2021         mlx5e_destroy_cq(cq);
2022
2023         return (err);
2024 }
2025
2026 void
2027 mlx5e_close_cq(struct mlx5e_cq *cq)
2028 {
2029         mlx5e_disable_cq(cq);
2030         mlx5e_destroy_cq(cq);
2031 }
2032
2033 static int
2034 mlx5e_open_tx_cqs(struct mlx5e_channel *c,
2035     struct mlx5e_channel_param *cparam)
2036 {
2037         int err;
2038         int tc;
2039
2040         for (tc = 0; tc < c->num_tc; tc++) {
2041                 /* open completion queue */
2042                 err = mlx5e_open_cq(c->priv, &cparam->tx_cq, &c->sq[tc].cq,
2043                     &mlx5e_tx_cq_comp, c->ix);
2044                 if (err)
2045                         goto err_close_tx_cqs;
2046         }
2047         return (0);
2048
2049 err_close_tx_cqs:
2050         for (tc--; tc >= 0; tc--)
2051                 mlx5e_close_cq(&c->sq[tc].cq);
2052
2053         return (err);
2054 }
2055
2056 static void
2057 mlx5e_close_tx_cqs(struct mlx5e_channel *c)
2058 {
2059         int tc;
2060
2061         for (tc = 0; tc < c->num_tc; tc++)
2062                 mlx5e_close_cq(&c->sq[tc].cq);
2063 }
2064
2065 static int
2066 mlx5e_open_sqs(struct mlx5e_channel *c,
2067     struct mlx5e_channel_param *cparam)
2068 {
2069         int err;
2070         int tc;
2071
2072         for (tc = 0; tc < c->num_tc; tc++) {
2073                 err = mlx5e_open_sq(c, tc, &cparam->sq, &c->sq[tc]);
2074                 if (err)
2075                         goto err_close_sqs;
2076         }
2077
2078         return (0);
2079
2080 err_close_sqs:
2081         for (tc--; tc >= 0; tc--)
2082                 mlx5e_close_sq_wait(&c->sq[tc]);
2083
2084         return (err);
2085 }
2086
2087 static void
2088 mlx5e_close_sqs_wait(struct mlx5e_channel *c)
2089 {
2090         int tc;
2091
2092         for (tc = 0; tc < c->num_tc; tc++)
2093                 mlx5e_close_sq_wait(&c->sq[tc]);
2094 }
2095
2096 static void
2097 mlx5e_chan_mtx_init(struct mlx5e_channel *c)
2098 {
2099         int tc;
2100
2101         mtx_init(&c->rq.mtx, "mlx5rx", MTX_NETWORK_LOCK, MTX_DEF);
2102
2103         callout_init_mtx(&c->rq.watchdog, &c->rq.mtx, 0);
2104
2105         for (tc = 0; tc < c->num_tc; tc++) {
2106                 struct mlx5e_sq *sq = c->sq + tc;
2107
2108                 mtx_init(&sq->lock, "mlx5tx",
2109                     MTX_NETWORK_LOCK " TX", MTX_DEF);
2110                 mtx_init(&sq->comp_lock, "mlx5comp",
2111                     MTX_NETWORK_LOCK " TX", MTX_DEF);
2112
2113                 callout_init_mtx(&sq->cev_callout, &sq->lock, 0);
2114
2115                 sq->cev_factor = c->priv->params_ethtool.tx_completion_fact;
2116
2117                 /* ensure the TX completion event factor is not zero */
2118                 if (sq->cev_factor == 0)
2119                         sq->cev_factor = 1;
2120         }
2121 }
2122
2123 static void
2124 mlx5e_chan_mtx_destroy(struct mlx5e_channel *c)
2125 {
2126         int tc;
2127
2128         mtx_destroy(&c->rq.mtx);
2129
2130         for (tc = 0; tc < c->num_tc; tc++) {
2131                 mtx_destroy(&c->sq[tc].lock);
2132                 mtx_destroy(&c->sq[tc].comp_lock);
2133         }
2134 }
2135
2136 static int
2137 mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
2138     struct mlx5e_channel_param *cparam,
2139     struct mlx5e_channel *c)
2140 {
2141         int err;
2142
2143         memset(c, 0, sizeof(*c));
2144
2145         c->priv = priv;
2146         c->ix = ix;
2147         /* setup send tag */
2148         c->tag.type = IF_SND_TAG_TYPE_UNLIMITED;
2149         c->mkey_be = cpu_to_be32(priv->mr.key);
2150         c->num_tc = priv->num_tc;
2151
2152         /* init mutexes */
2153         mlx5e_chan_mtx_init(c);
2154
2155         /* open transmit completion queue */
2156         err = mlx5e_open_tx_cqs(c, cparam);
2157         if (err)
2158                 goto err_free;
2159
2160         /* open receive completion queue */
2161         err = mlx5e_open_cq(c->priv, &cparam->rx_cq, &c->rq.cq,
2162             &mlx5e_rx_cq_comp, c->ix);
2163         if (err)
2164                 goto err_close_tx_cqs;
2165
2166         err = mlx5e_open_sqs(c, cparam);
2167         if (err)
2168                 goto err_close_rx_cq;
2169
2170         err = mlx5e_open_rq(c, &cparam->rq, &c->rq);
2171         if (err)
2172                 goto err_close_sqs;
2173
2174         /* poll receive queue initially */
2175         c->rq.cq.mcq.comp(&c->rq.cq.mcq);
2176
2177         return (0);
2178
2179 err_close_sqs:
2180         mlx5e_close_sqs_wait(c);
2181
2182 err_close_rx_cq:
2183         mlx5e_close_cq(&c->rq.cq);
2184
2185 err_close_tx_cqs:
2186         mlx5e_close_tx_cqs(c);
2187
2188 err_free:
2189         /* destroy mutexes */
2190         mlx5e_chan_mtx_destroy(c);
2191         return (err);
2192 }
2193
2194 static void
2195 mlx5e_close_channel(struct mlx5e_channel *c)
2196 {
2197         mlx5e_close_rq(&c->rq);
2198 }
2199
2200 static void
2201 mlx5e_close_channel_wait(struct mlx5e_channel *c)
2202 {
2203         mlx5e_close_rq_wait(&c->rq);
2204         mlx5e_close_sqs_wait(c);
2205         mlx5e_close_tx_cqs(c);
2206         /* destroy mutexes */
2207         mlx5e_chan_mtx_destroy(c);
2208 }
2209
2210 static int
2211 mlx5e_get_wqe_sz(struct mlx5e_priv *priv, u32 *wqe_sz, u32 *nsegs)
2212 {
2213         u32 r, n;
2214
2215         r = priv->params.hw_lro_en ? priv->params.lro_wqe_sz :
2216             MLX5E_SW2MB_MTU(priv->ifp->if_mtu);
2217         if (r > MJUM16BYTES)
2218                 return (-ENOMEM);
2219
2220         if (r > MJUM9BYTES)
2221                 r = MJUM16BYTES;
2222         else if (r > MJUMPAGESIZE)
2223                 r = MJUM9BYTES;
2224         else if (r > MCLBYTES)
2225                 r = MJUMPAGESIZE;
2226         else
2227                 r = MCLBYTES;
2228
2229         /*
2230          * n + 1 must be a power of two, because stride size must be.
2231          * Stride size is 16 * (n + 1), as the first segment is
2232          * control.
2233          */
2234         for (n = howmany(r, MLX5E_MAX_RX_BYTES); !powerof2(n + 1); n++)
2235                 ;
2236
2237         if (n > MLX5E_MAX_BUSDMA_RX_SEGS)
2238                 return (-ENOMEM);
2239
2240         *wqe_sz = r;
2241         *nsegs = n;
2242         return (0);
2243 }
2244
2245 static void
2246 mlx5e_build_rq_param(struct mlx5e_priv *priv,
2247     struct mlx5e_rq_param *param)
2248 {
2249         void *rqc = param->rqc;
2250         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
2251         u32 wqe_sz, nsegs;
2252
2253         mlx5e_get_wqe_sz(priv, &wqe_sz, &nsegs);
2254         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
2255         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
2256         MLX5_SET(wq, wq, log_wq_stride, ilog2(sizeof(struct mlx5e_rx_wqe) +
2257             nsegs * sizeof(struct mlx5_wqe_data_seg)));
2258         MLX5_SET(wq, wq, log_wq_sz, priv->params.log_rq_size);
2259         MLX5_SET(wq, wq, pd, priv->pdn);
2260
2261         param->wq.buf_numa_node = 0;
2262         param->wq.db_numa_node = 0;
2263         param->wq.linear = 1;
2264 }
2265
2266 static void
2267 mlx5e_build_sq_param(struct mlx5e_priv *priv,
2268     struct mlx5e_sq_param *param)
2269 {
2270         void *sqc = param->sqc;
2271         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
2272
2273         MLX5_SET(wq, wq, log_wq_sz, priv->params.log_sq_size);
2274         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
2275         MLX5_SET(wq, wq, pd, priv->pdn);
2276
2277         param->wq.buf_numa_node = 0;
2278         param->wq.db_numa_node = 0;
2279         param->wq.linear = 1;
2280 }
2281
2282 static void
2283 mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
2284     struct mlx5e_cq_param *param)
2285 {
2286         void *cqc = param->cqc;
2287
2288         MLX5_SET(cqc, cqc, uar_page, priv->cq_uar.index);
2289 }
2290
2291 static void
2292 mlx5e_get_default_profile(struct mlx5e_priv *priv, int mode, struct net_dim_cq_moder *ptr)
2293 {
2294
2295         *ptr = net_dim_get_profile(mode, MLX5E_DIM_DEFAULT_PROFILE);
2296
2297         /* apply LRO restrictions */
2298         if (priv->params.hw_lro_en &&
2299             ptr->pkts > MLX5E_DIM_MAX_RX_CQ_MODERATION_PKTS_WITH_LRO) {
2300                 ptr->pkts = MLX5E_DIM_MAX_RX_CQ_MODERATION_PKTS_WITH_LRO;
2301         }
2302 }
2303
2304 static void
2305 mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
2306     struct mlx5e_cq_param *param)
2307 {
2308         struct net_dim_cq_moder curr;
2309         void *cqc = param->cqc;
2310
2311         /*
2312          * We use MLX5_CQE_FORMAT_HASH because the RX hash mini CQE
2313          * format is more beneficial for FreeBSD use case.
2314          *
2315          * Adding support for MLX5_CQE_FORMAT_CSUM will require changes
2316          * in mlx5e_decompress_cqe.
2317          */
2318         if (priv->params.cqe_zipping_en) {
2319                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_HASH);
2320                 MLX5_SET(cqc, cqc, cqe_compression_en, 1);
2321         }
2322
2323         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_rq_size);
2324
2325         switch (priv->params.rx_cq_moderation_mode) {
2326         case 0:
2327                 MLX5_SET(cqc, cqc, cq_period, priv->params.rx_cq_moderation_usec);
2328                 MLX5_SET(cqc, cqc, cq_max_count, priv->params.rx_cq_moderation_pkts);
2329                 MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2330                 break;
2331         case 1:
2332                 MLX5_SET(cqc, cqc, cq_period, priv->params.rx_cq_moderation_usec);
2333                 MLX5_SET(cqc, cqc, cq_max_count, priv->params.rx_cq_moderation_pkts);
2334                 if (MLX5_CAP_GEN(priv->mdev, cq_period_start_from_cqe))
2335                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
2336                 else
2337                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2338                 break;
2339         case 2:
2340                 mlx5e_get_default_profile(priv, NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE, &curr);
2341                 MLX5_SET(cqc, cqc, cq_period, curr.usec);
2342                 MLX5_SET(cqc, cqc, cq_max_count, curr.pkts);
2343                 MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2344                 break;
2345         case 3:
2346                 mlx5e_get_default_profile(priv, NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE, &curr);
2347                 MLX5_SET(cqc, cqc, cq_period, curr.usec);
2348                 MLX5_SET(cqc, cqc, cq_max_count, curr.pkts);
2349                 if (MLX5_CAP_GEN(priv->mdev, cq_period_start_from_cqe))
2350                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
2351                 else
2352                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2353                 break;
2354         default:
2355                 break;
2356         }
2357
2358         mlx5e_dim_build_cq_param(priv, param);
2359
2360         mlx5e_build_common_cq_param(priv, param);
2361 }
2362
2363 static void
2364 mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
2365     struct mlx5e_cq_param *param)
2366 {
2367         void *cqc = param->cqc;
2368
2369         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_sq_size);
2370         MLX5_SET(cqc, cqc, cq_period, priv->params.tx_cq_moderation_usec);
2371         MLX5_SET(cqc, cqc, cq_max_count, priv->params.tx_cq_moderation_pkts);
2372
2373         switch (priv->params.tx_cq_moderation_mode) {
2374         case 0:
2375                 MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2376                 break;
2377         default:
2378                 if (MLX5_CAP_GEN(priv->mdev, cq_period_start_from_cqe))
2379                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
2380                 else
2381                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
2382                 break;
2383         }
2384
2385         mlx5e_build_common_cq_param(priv, param);
2386 }
2387
2388 static void
2389 mlx5e_build_channel_param(struct mlx5e_priv *priv,
2390     struct mlx5e_channel_param *cparam)
2391 {
2392         memset(cparam, 0, sizeof(*cparam));
2393
2394         mlx5e_build_rq_param(priv, &cparam->rq);
2395         mlx5e_build_sq_param(priv, &cparam->sq);
2396         mlx5e_build_rx_cq_param(priv, &cparam->rx_cq);
2397         mlx5e_build_tx_cq_param(priv, &cparam->tx_cq);
2398 }
2399
2400 static int
2401 mlx5e_open_channels(struct mlx5e_priv *priv)
2402 {
2403         struct mlx5e_channel_param cparam;
2404         int err;
2405         int i;
2406         int j;
2407
2408         mlx5e_build_channel_param(priv, &cparam);
2409         for (i = 0; i < priv->params.num_channels; i++) {
2410                 err = mlx5e_open_channel(priv, i, &cparam, &priv->channel[i]);
2411                 if (err)
2412                         goto err_close_channels;
2413         }
2414
2415         for (j = 0; j < priv->params.num_channels; j++) {
2416                 err = mlx5e_wait_for_min_rx_wqes(&priv->channel[j].rq);
2417                 if (err)
2418                         goto err_close_channels;
2419         }
2420         return (0);
2421
2422 err_close_channels:
2423         while (i--) {
2424                 mlx5e_close_channel(&priv->channel[i]);
2425                 mlx5e_close_channel_wait(&priv->channel[i]);
2426         }
2427         return (err);
2428 }
2429
2430 static void
2431 mlx5e_close_channels(struct mlx5e_priv *priv)
2432 {
2433         int i;
2434
2435         for (i = 0; i < priv->params.num_channels; i++)
2436                 mlx5e_close_channel(&priv->channel[i]);
2437         for (i = 0; i < priv->params.num_channels; i++)
2438                 mlx5e_close_channel_wait(&priv->channel[i]);
2439 }
2440
2441 static int
2442 mlx5e_refresh_sq_params(struct mlx5e_priv *priv, struct mlx5e_sq *sq)
2443 {
2444
2445         if (MLX5_CAP_GEN(priv->mdev, cq_period_mode_modify)) {
2446                 uint8_t cq_mode;
2447
2448                 switch (priv->params.tx_cq_moderation_mode) {
2449                 case 0:
2450                 case 2:
2451                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
2452                         break;
2453                 default:
2454                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_CQE;
2455                         break;
2456                 }
2457
2458                 return (mlx5_core_modify_cq_moderation_mode(priv->mdev, &sq->cq.mcq,
2459                     priv->params.tx_cq_moderation_usec,
2460                     priv->params.tx_cq_moderation_pkts,
2461                     cq_mode));
2462         }
2463
2464         return (mlx5_core_modify_cq_moderation(priv->mdev, &sq->cq.mcq,
2465             priv->params.tx_cq_moderation_usec,
2466             priv->params.tx_cq_moderation_pkts));
2467 }
2468
2469 static int
2470 mlx5e_refresh_rq_params(struct mlx5e_priv *priv, struct mlx5e_rq *rq)
2471 {
2472
2473         if (MLX5_CAP_GEN(priv->mdev, cq_period_mode_modify)) {
2474                 uint8_t cq_mode;
2475                 uint8_t dim_mode;
2476                 int retval;
2477
2478                 switch (priv->params.rx_cq_moderation_mode) {
2479                 case 0:
2480                 case 2:
2481                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
2482                         dim_mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_EQE;
2483                         break;
2484                 default:
2485                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_CQE;
2486                         dim_mode = NET_DIM_CQ_PERIOD_MODE_START_FROM_CQE;
2487                         break;
2488                 }
2489
2490                 /* tear down dynamic interrupt moderation */
2491                 mtx_lock(&rq->mtx);
2492                 rq->dim.mode = NET_DIM_CQ_PERIOD_MODE_DISABLED;
2493                 mtx_unlock(&rq->mtx);
2494
2495                 /* wait for dynamic interrupt moderation work task, if any */
2496                 cancel_work_sync(&rq->dim.work);
2497
2498                 if (priv->params.rx_cq_moderation_mode >= 2) {
2499                         struct net_dim_cq_moder curr;
2500
2501                         mlx5e_get_default_profile(priv, dim_mode, &curr);
2502
2503                         retval = mlx5_core_modify_cq_moderation_mode(priv->mdev, &rq->cq.mcq,
2504                             curr.usec, curr.pkts, cq_mode);
2505
2506                         /* set dynamic interrupt moderation mode and zero defaults */
2507                         mtx_lock(&rq->mtx);
2508                         rq->dim.mode = dim_mode;
2509                         rq->dim.state = 0;
2510                         rq->dim.profile_ix = MLX5E_DIM_DEFAULT_PROFILE;
2511                         mtx_unlock(&rq->mtx);
2512                 } else {
2513                         retval = mlx5_core_modify_cq_moderation_mode(priv->mdev, &rq->cq.mcq,
2514                             priv->params.rx_cq_moderation_usec,
2515                             priv->params.rx_cq_moderation_pkts,
2516                             cq_mode);
2517                 }
2518                 return (retval);
2519         }
2520
2521         return (mlx5_core_modify_cq_moderation(priv->mdev, &rq->cq.mcq,
2522             priv->params.rx_cq_moderation_usec,
2523             priv->params.rx_cq_moderation_pkts));
2524 }
2525
2526 static int
2527 mlx5e_refresh_channel_params_sub(struct mlx5e_priv *priv, struct mlx5e_channel *c)
2528 {
2529         int err;
2530         int i;
2531
2532         err = mlx5e_refresh_rq_params(priv, &c->rq);
2533         if (err)
2534                 goto done;
2535
2536         for (i = 0; i != c->num_tc; i++) {
2537                 err = mlx5e_refresh_sq_params(priv, &c->sq[i]);
2538                 if (err)
2539                         goto done;
2540         }
2541 done:
2542         return (err);
2543 }
2544
2545 int
2546 mlx5e_refresh_channel_params(struct mlx5e_priv *priv)
2547 {
2548         int i;
2549
2550         /* check if channels are closed */
2551         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
2552                 return (EINVAL);
2553
2554         for (i = 0; i < priv->params.num_channels; i++) {
2555                 int err;
2556
2557                 err = mlx5e_refresh_channel_params_sub(priv, &priv->channel[i]);
2558                 if (err)
2559                         return (err);
2560         }
2561         return (0);
2562 }
2563
2564 static int
2565 mlx5e_open_tis(struct mlx5e_priv *priv, int tc)
2566 {
2567         struct mlx5_core_dev *mdev = priv->mdev;
2568         u32 in[MLX5_ST_SZ_DW(create_tis_in)];
2569         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
2570
2571         memset(in, 0, sizeof(in));
2572
2573         MLX5_SET(tisc, tisc, prio, tc);
2574         MLX5_SET(tisc, tisc, transport_domain, priv->tdn);
2575
2576         return (mlx5_core_create_tis(mdev, in, sizeof(in), &priv->tisn[tc]));
2577 }
2578
2579 static void
2580 mlx5e_close_tis(struct mlx5e_priv *priv, int tc)
2581 {
2582         mlx5_core_destroy_tis(priv->mdev, priv->tisn[tc]);
2583 }
2584
2585 static int
2586 mlx5e_open_tises(struct mlx5e_priv *priv)
2587 {
2588         int num_tc = priv->num_tc;
2589         int err;
2590         int tc;
2591
2592         for (tc = 0; tc < num_tc; tc++) {
2593                 err = mlx5e_open_tis(priv, tc);
2594                 if (err)
2595                         goto err_close_tises;
2596         }
2597
2598         return (0);
2599
2600 err_close_tises:
2601         for (tc--; tc >= 0; tc--)
2602                 mlx5e_close_tis(priv, tc);
2603
2604         return (err);
2605 }
2606
2607 static void
2608 mlx5e_close_tises(struct mlx5e_priv *priv)
2609 {
2610         int num_tc = priv->num_tc;
2611         int tc;
2612
2613         for (tc = 0; tc < num_tc; tc++)
2614                 mlx5e_close_tis(priv, tc);
2615 }
2616
2617 static int
2618 mlx5e_open_rqt(struct mlx5e_priv *priv)
2619 {
2620         struct mlx5_core_dev *mdev = priv->mdev;
2621         u32 *in;
2622         u32 out[MLX5_ST_SZ_DW(create_rqt_out)] = {0};
2623         void *rqtc;
2624         int inlen;
2625         int err;
2626         int sz;
2627         int i;
2628
2629         sz = 1 << priv->params.rx_hash_log_tbl_sz;
2630
2631         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2632         in = mlx5_vzalloc(inlen);
2633         if (in == NULL)
2634                 return (-ENOMEM);
2635         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2636
2637         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2638         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2639
2640         for (i = 0; i < sz; i++) {
2641                 int ix = i;
2642 #ifdef RSS
2643                 ix = rss_get_indirection_to_bucket(ix);
2644 #endif
2645                 /* ensure we don't overflow */
2646                 ix %= priv->params.num_channels;
2647
2648                 /* apply receive side scaling stride, if any */
2649                 ix -= ix % (int)priv->params.channels_rsss;
2650
2651                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->channel[ix].rq.rqn);
2652         }
2653
2654         MLX5_SET(create_rqt_in, in, opcode, MLX5_CMD_OP_CREATE_RQT);
2655
2656         err = mlx5_cmd_exec(mdev, in, inlen, out, sizeof(out));
2657         if (!err)
2658                 priv->rqtn = MLX5_GET(create_rqt_out, out, rqtn);
2659
2660         kvfree(in);
2661
2662         return (err);
2663 }
2664
2665 static void
2666 mlx5e_close_rqt(struct mlx5e_priv *priv)
2667 {
2668         u32 in[MLX5_ST_SZ_DW(destroy_rqt_in)] = {0};
2669         u32 out[MLX5_ST_SZ_DW(destroy_rqt_out)] = {0};
2670
2671         MLX5_SET(destroy_rqt_in, in, opcode, MLX5_CMD_OP_DESTROY_RQT);
2672         MLX5_SET(destroy_rqt_in, in, rqtn, priv->rqtn);
2673
2674         mlx5_cmd_exec(priv->mdev, in, sizeof(in), out, sizeof(out));
2675 }
2676
2677 static void
2678 mlx5e_build_tir_ctx(struct mlx5e_priv *priv, u32 * tirc, int tt)
2679 {
2680         void *hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2681         __be32 *hkey;
2682
2683         MLX5_SET(tirc, tirc, transport_domain, priv->tdn);
2684
2685 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2686
2687 #define MLX5_HASH_IP     (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2688                           MLX5_HASH_FIELD_SEL_DST_IP)
2689
2690 #define MLX5_HASH_ALL    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2691                           MLX5_HASH_FIELD_SEL_DST_IP   |\
2692                           MLX5_HASH_FIELD_SEL_L4_SPORT |\
2693                           MLX5_HASH_FIELD_SEL_L4_DPORT)
2694
2695 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2696                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2697                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
2698
2699         if (priv->params.hw_lro_en) {
2700                 MLX5_SET(tirc, tirc, lro_enable_mask,
2701                     MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2702                     MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2703                 MLX5_SET(tirc, tirc, lro_max_msg_sz,
2704                     (priv->params.lro_wqe_sz -
2705                     ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2706                 /* TODO: add the option to choose timer value dynamically */
2707                 MLX5_SET(tirc, tirc, lro_timeout_period_usecs,
2708                     MLX5_CAP_ETH(priv->mdev,
2709                     lro_timer_supported_periods[2]));
2710         }
2711
2712         /* setup parameters for hashing TIR type, if any */
2713         switch (tt) {
2714         case MLX5E_TT_ANY:
2715                 MLX5_SET(tirc, tirc, disp_type,
2716                     MLX5_TIRC_DISP_TYPE_DIRECT);
2717                 MLX5_SET(tirc, tirc, inline_rqn,
2718                     priv->channel[0].rq.rqn);
2719                 break;
2720         default:
2721                 MLX5_SET(tirc, tirc, disp_type,
2722                     MLX5_TIRC_DISP_TYPE_INDIRECT);
2723                 MLX5_SET(tirc, tirc, indirect_table,
2724                     priv->rqtn);
2725                 MLX5_SET(tirc, tirc, rx_hash_fn,
2726                     MLX5_TIRC_RX_HASH_FN_HASH_TOEPLITZ);
2727                 hkey = (__be32 *) MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
2728 #ifdef RSS
2729                 /*
2730                  * The FreeBSD RSS implementation does currently not
2731                  * support symmetric Toeplitz hashes:
2732                  */
2733                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 0);
2734                 rss_getkey((uint8_t *)hkey);
2735 #else
2736                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2737                 hkey[0] = cpu_to_be32(0xD181C62C);
2738                 hkey[1] = cpu_to_be32(0xF7F4DB5B);
2739                 hkey[2] = cpu_to_be32(0x1983A2FC);
2740                 hkey[3] = cpu_to_be32(0x943E1ADB);
2741                 hkey[4] = cpu_to_be32(0xD9389E6B);
2742                 hkey[5] = cpu_to_be32(0xD1039C2C);
2743                 hkey[6] = cpu_to_be32(0xA74499AD);
2744                 hkey[7] = cpu_to_be32(0x593D56D9);
2745                 hkey[8] = cpu_to_be32(0xF3253C06);
2746                 hkey[9] = cpu_to_be32(0x2ADC1FFC);
2747 #endif
2748                 break;
2749         }
2750
2751         switch (tt) {
2752         case MLX5E_TT_IPV4_TCP:
2753                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2754                     MLX5_L3_PROT_TYPE_IPV4);
2755                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2756                     MLX5_L4_PROT_TYPE_TCP);
2757 #ifdef RSS
2758                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4)) {
2759                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2760                             MLX5_HASH_IP);
2761                 } else
2762 #endif
2763                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2764                     MLX5_HASH_ALL);
2765                 break;
2766
2767         case MLX5E_TT_IPV6_TCP:
2768                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2769                     MLX5_L3_PROT_TYPE_IPV6);
2770                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2771                     MLX5_L4_PROT_TYPE_TCP);
2772 #ifdef RSS
2773                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6)) {
2774                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2775                             MLX5_HASH_IP);
2776                 } else
2777 #endif
2778                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2779                     MLX5_HASH_ALL);
2780                 break;
2781
2782         case MLX5E_TT_IPV4_UDP:
2783                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2784                     MLX5_L3_PROT_TYPE_IPV4);
2785                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2786                     MLX5_L4_PROT_TYPE_UDP);
2787 #ifdef RSS
2788                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) {
2789                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2790                             MLX5_HASH_IP);
2791                 } else
2792 #endif
2793                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2794                     MLX5_HASH_ALL);
2795                 break;
2796
2797         case MLX5E_TT_IPV6_UDP:
2798                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2799                     MLX5_L3_PROT_TYPE_IPV6);
2800                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2801                     MLX5_L4_PROT_TYPE_UDP);
2802 #ifdef RSS
2803                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6)) {
2804                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2805                             MLX5_HASH_IP);
2806                 } else
2807 #endif
2808                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2809                     MLX5_HASH_ALL);
2810                 break;
2811
2812         case MLX5E_TT_IPV4_IPSEC_AH:
2813                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2814                     MLX5_L3_PROT_TYPE_IPV4);
2815                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2816                     MLX5_HASH_IP_IPSEC_SPI);
2817                 break;
2818
2819         case MLX5E_TT_IPV6_IPSEC_AH:
2820                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2821                     MLX5_L3_PROT_TYPE_IPV6);
2822                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2823                     MLX5_HASH_IP_IPSEC_SPI);
2824                 break;
2825
2826         case MLX5E_TT_IPV4_IPSEC_ESP:
2827                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2828                     MLX5_L3_PROT_TYPE_IPV4);
2829                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2830                     MLX5_HASH_IP_IPSEC_SPI);
2831                 break;
2832
2833         case MLX5E_TT_IPV6_IPSEC_ESP:
2834                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2835                     MLX5_L3_PROT_TYPE_IPV6);
2836                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2837                     MLX5_HASH_IP_IPSEC_SPI);
2838                 break;
2839
2840         case MLX5E_TT_IPV4:
2841                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2842                     MLX5_L3_PROT_TYPE_IPV4);
2843                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2844                     MLX5_HASH_IP);
2845                 break;
2846
2847         case MLX5E_TT_IPV6:
2848                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2849                     MLX5_L3_PROT_TYPE_IPV6);
2850                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2851                     MLX5_HASH_IP);
2852                 break;
2853
2854         default:
2855                 break;
2856         }
2857 }
2858
2859 static int
2860 mlx5e_open_tir(struct mlx5e_priv *priv, int tt)
2861 {
2862         struct mlx5_core_dev *mdev = priv->mdev;
2863         u32 *in;
2864         void *tirc;
2865         int inlen;
2866         int err;
2867
2868         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2869         in = mlx5_vzalloc(inlen);
2870         if (in == NULL)
2871                 return (-ENOMEM);
2872         tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
2873
2874         mlx5e_build_tir_ctx(priv, tirc, tt);
2875
2876         err = mlx5_core_create_tir(mdev, in, inlen, &priv->tirn[tt]);
2877
2878         kvfree(in);
2879
2880         return (err);
2881 }
2882
2883 static void
2884 mlx5e_close_tir(struct mlx5e_priv *priv, int tt)
2885 {
2886         mlx5_core_destroy_tir(priv->mdev, priv->tirn[tt]);
2887 }
2888
2889 static int
2890 mlx5e_open_tirs(struct mlx5e_priv *priv)
2891 {
2892         int err;
2893         int i;
2894
2895         for (i = 0; i < MLX5E_NUM_TT; i++) {
2896                 err = mlx5e_open_tir(priv, i);
2897                 if (err)
2898                         goto err_close_tirs;
2899         }
2900
2901         return (0);
2902
2903 err_close_tirs:
2904         for (i--; i >= 0; i--)
2905                 mlx5e_close_tir(priv, i);
2906
2907         return (err);
2908 }
2909
2910 static void
2911 mlx5e_close_tirs(struct mlx5e_priv *priv)
2912 {
2913         int i;
2914
2915         for (i = 0; i < MLX5E_NUM_TT; i++)
2916                 mlx5e_close_tir(priv, i);
2917 }
2918
2919 /*
2920  * SW MTU does not include headers,
2921  * HW MTU includes all headers and checksums.
2922  */
2923 static int
2924 mlx5e_set_dev_port_mtu(struct ifnet *ifp, int sw_mtu)
2925 {
2926         struct mlx5e_priv *priv = ifp->if_softc;
2927         struct mlx5_core_dev *mdev = priv->mdev;
2928         int hw_mtu;
2929         int err;
2930
2931         hw_mtu = MLX5E_SW2HW_MTU(sw_mtu);
2932
2933         err = mlx5_set_port_mtu(mdev, hw_mtu);
2934         if (err) {
2935                 if_printf(ifp, "%s: mlx5_set_port_mtu failed setting %d, err=%d\n",
2936                     __func__, sw_mtu, err);
2937                 return (err);
2938         }
2939
2940         /* Update vport context MTU */
2941         err = mlx5_set_vport_mtu(mdev, hw_mtu);
2942         if (err) {
2943                 if_printf(ifp, "%s: Failed updating vport context with MTU size, err=%d\n",
2944                     __func__, err);
2945         }
2946
2947         ifp->if_mtu = sw_mtu;
2948
2949         err = mlx5_query_vport_mtu(mdev, &hw_mtu);
2950         if (err || !hw_mtu) {
2951                 /* fallback to port oper mtu */
2952                 err = mlx5_query_port_oper_mtu(mdev, &hw_mtu);
2953         }
2954         if (err) {
2955                 if_printf(ifp, "Query port MTU, after setting new "
2956                     "MTU value, failed\n");
2957                 return (err);
2958         } else if (MLX5E_HW2SW_MTU(hw_mtu) < sw_mtu) {
2959                 err = -E2BIG,
2960                 if_printf(ifp, "Port MTU %d is smaller than "
2961                     "ifp mtu %d\n", hw_mtu, sw_mtu);
2962         } else if (MLX5E_HW2SW_MTU(hw_mtu) > sw_mtu) {
2963                 err = -EINVAL;
2964                 if_printf(ifp, "Port MTU %d is bigger than "
2965                     "ifp mtu %d\n", hw_mtu, sw_mtu);
2966         }
2967         priv->params_ethtool.hw_mtu = hw_mtu;
2968
2969         return (err);
2970 }
2971
2972 int
2973 mlx5e_open_locked(struct ifnet *ifp)
2974 {
2975         struct mlx5e_priv *priv = ifp->if_softc;
2976         int err;
2977         u16 set_id;
2978
2979         /* check if already opened */
2980         if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0)
2981                 return (0);
2982
2983 #ifdef RSS
2984         if (rss_getnumbuckets() > priv->params.num_channels) {
2985                 if_printf(ifp, "NOTE: There are more RSS buckets(%u) than "
2986                     "channels(%u) available\n", rss_getnumbuckets(),
2987                     priv->params.num_channels);
2988         }
2989 #endif
2990         err = mlx5e_open_tises(priv);
2991         if (err) {
2992                 if_printf(ifp, "%s: mlx5e_open_tises failed, %d\n",
2993                     __func__, err);
2994                 return (err);
2995         }
2996         err = mlx5_vport_alloc_q_counter(priv->mdev,
2997             MLX5_INTERFACE_PROTOCOL_ETH, &set_id);
2998         if (err) {
2999                 if_printf(priv->ifp,
3000                     "%s: mlx5_vport_alloc_q_counter failed: %d\n",
3001                     __func__, err);
3002                 goto err_close_tises;
3003         }
3004         /* store counter set ID */
3005         priv->counter_set_id = set_id;
3006
3007         err = mlx5e_open_channels(priv);
3008         if (err) {
3009                 if_printf(ifp, "%s: mlx5e_open_channels failed, %d\n",
3010                     __func__, err);
3011                 goto err_dalloc_q_counter;
3012         }
3013         err = mlx5e_open_rqt(priv);
3014         if (err) {
3015                 if_printf(ifp, "%s: mlx5e_open_rqt failed, %d\n",
3016                     __func__, err);
3017                 goto err_close_channels;
3018         }
3019         err = mlx5e_open_tirs(priv);
3020         if (err) {
3021                 if_printf(ifp, "%s: mlx5e_open_tir failed, %d\n",
3022                     __func__, err);
3023                 goto err_close_rqls;
3024         }
3025         err = mlx5e_open_flow_table(priv);
3026         if (err) {
3027                 if_printf(ifp, "%s: mlx5e_open_flow_table failed, %d\n",
3028                     __func__, err);
3029                 goto err_close_tirs;
3030         }
3031         err = mlx5e_add_all_vlan_rules(priv);
3032         if (err) {
3033                 if_printf(ifp, "%s: mlx5e_add_all_vlan_rules failed, %d\n",
3034                     __func__, err);
3035                 goto err_close_flow_table;
3036         }
3037         set_bit(MLX5E_STATE_OPENED, &priv->state);
3038
3039         mlx5e_update_carrier(priv);
3040         mlx5e_set_rx_mode_core(priv);
3041
3042         return (0);
3043
3044 err_close_flow_table:
3045         mlx5e_close_flow_table(priv);
3046
3047 err_close_tirs:
3048         mlx5e_close_tirs(priv);
3049
3050 err_close_rqls:
3051         mlx5e_close_rqt(priv);
3052
3053 err_close_channels:
3054         mlx5e_close_channels(priv);
3055
3056 err_dalloc_q_counter:
3057         mlx5_vport_dealloc_q_counter(priv->mdev,
3058             MLX5_INTERFACE_PROTOCOL_ETH, priv->counter_set_id);
3059
3060 err_close_tises:
3061         mlx5e_close_tises(priv);
3062
3063         return (err);
3064 }
3065
3066 static void
3067 mlx5e_open(void *arg)
3068 {
3069         struct mlx5e_priv *priv = arg;
3070
3071         PRIV_LOCK(priv);
3072         if (mlx5_set_port_status(priv->mdev, MLX5_PORT_UP))
3073                 if_printf(priv->ifp,
3074                     "%s: Setting port status to up failed\n",
3075                     __func__);
3076
3077         mlx5e_open_locked(priv->ifp);
3078         priv->ifp->if_drv_flags |= IFF_DRV_RUNNING;
3079         PRIV_UNLOCK(priv);
3080 }
3081
3082 int
3083 mlx5e_close_locked(struct ifnet *ifp)
3084 {
3085         struct mlx5e_priv *priv = ifp->if_softc;
3086
3087         /* check if already closed */
3088         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
3089                 return (0);
3090
3091         clear_bit(MLX5E_STATE_OPENED, &priv->state);
3092
3093         mlx5e_set_rx_mode_core(priv);
3094         mlx5e_del_all_vlan_rules(priv);
3095         if_link_state_change(priv->ifp, LINK_STATE_DOWN);
3096         mlx5e_close_flow_table(priv);
3097         mlx5e_close_tirs(priv);
3098         mlx5e_close_rqt(priv);
3099         mlx5e_close_channels(priv);
3100         mlx5_vport_dealloc_q_counter(priv->mdev,
3101             MLX5_INTERFACE_PROTOCOL_ETH, priv->counter_set_id);
3102         mlx5e_close_tises(priv);
3103
3104         return (0);
3105 }
3106
3107 #if (__FreeBSD_version >= 1100000)
3108 static uint64_t
3109 mlx5e_get_counter(struct ifnet *ifp, ift_counter cnt)
3110 {
3111         struct mlx5e_priv *priv = ifp->if_softc;
3112         u64 retval;
3113
3114         /* PRIV_LOCK(priv); XXX not allowed */
3115         switch (cnt) {
3116         case IFCOUNTER_IPACKETS:
3117                 retval = priv->stats.vport.rx_packets;
3118                 break;
3119         case IFCOUNTER_IERRORS:
3120                 retval = priv->stats.pport.in_range_len_errors +
3121                     priv->stats.pport.out_of_range_len +
3122                     priv->stats.pport.too_long_errors +
3123                     priv->stats.pport.check_seq_err +
3124                     priv->stats.pport.alignment_err;
3125                 break;
3126         case IFCOUNTER_IQDROPS:
3127                 retval = priv->stats.vport.rx_out_of_buffer;
3128                 break;
3129         case IFCOUNTER_OPACKETS:
3130                 retval = priv->stats.vport.tx_packets;
3131                 break;
3132         case IFCOUNTER_OERRORS:
3133                 retval = priv->stats.port_stats_debug.out_discards;
3134                 break;
3135         case IFCOUNTER_IBYTES:
3136                 retval = priv->stats.vport.rx_bytes;
3137                 break;
3138         case IFCOUNTER_OBYTES:
3139                 retval = priv->stats.vport.tx_bytes;
3140                 break;
3141         case IFCOUNTER_IMCASTS:
3142                 retval = priv->stats.vport.rx_multicast_packets;
3143                 break;
3144         case IFCOUNTER_OMCASTS:
3145                 retval = priv->stats.vport.tx_multicast_packets;
3146                 break;
3147         case IFCOUNTER_OQDROPS:
3148                 retval = priv->stats.vport.tx_queue_dropped;
3149                 break;
3150         case IFCOUNTER_COLLISIONS:
3151                 retval = priv->stats.pport.collisions;
3152                 break;
3153         default:
3154                 retval = if_get_counter_default(ifp, cnt);
3155                 break;
3156         }
3157         /* PRIV_UNLOCK(priv); XXX not allowed */
3158         return (retval);
3159 }
3160 #endif
3161
3162 static void
3163 mlx5e_set_rx_mode(struct ifnet *ifp)
3164 {
3165         struct mlx5e_priv *priv = ifp->if_softc;
3166
3167         queue_work(priv->wq, &priv->set_rx_mode_work);
3168 }
3169
3170 static int
3171 mlx5e_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
3172 {
3173         struct mlx5e_priv *priv;
3174         struct ifreq *ifr;
3175         struct ifi2creq i2c;
3176         int error = 0;
3177         int mask = 0;
3178         int size_read = 0;
3179         int module_status;
3180         int module_num;
3181         int max_mtu;
3182         uint8_t read_addr;
3183
3184         priv = ifp->if_softc;
3185
3186         /* check if detaching */
3187         if (priv == NULL || priv->gone != 0)
3188                 return (ENXIO);
3189
3190         switch (command) {
3191         case SIOCSIFMTU:
3192                 ifr = (struct ifreq *)data;
3193
3194                 PRIV_LOCK(priv);
3195                 mlx5_query_port_max_mtu(priv->mdev, &max_mtu);
3196
3197                 if (ifr->ifr_mtu >= MLX5E_MTU_MIN &&
3198                     ifr->ifr_mtu <= MIN(MLX5E_MTU_MAX, max_mtu)) {
3199                         int was_opened;
3200
3201                         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
3202                         if (was_opened)
3203                                 mlx5e_close_locked(ifp);
3204
3205                         /* set new MTU */
3206                         mlx5e_set_dev_port_mtu(ifp, ifr->ifr_mtu);
3207
3208                         if (was_opened)
3209                                 mlx5e_open_locked(ifp);
3210                 } else {
3211                         error = EINVAL;
3212                         if_printf(ifp, "Invalid MTU value. Min val: %d, Max val: %d\n",
3213                             MLX5E_MTU_MIN, MIN(MLX5E_MTU_MAX, max_mtu));
3214                 }
3215                 PRIV_UNLOCK(priv);
3216                 break;
3217         case SIOCSIFFLAGS:
3218                 if ((ifp->if_flags & IFF_UP) &&
3219                     (ifp->if_drv_flags & IFF_DRV_RUNNING)) {
3220                         mlx5e_set_rx_mode(ifp);
3221                         break;
3222                 }
3223                 PRIV_LOCK(priv);
3224                 if (ifp->if_flags & IFF_UP) {
3225                         if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
3226                                 if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
3227                                         mlx5e_open_locked(ifp);
3228                                 ifp->if_drv_flags |= IFF_DRV_RUNNING;
3229                                 mlx5_set_port_status(priv->mdev, MLX5_PORT_UP);
3230                         }
3231                 } else {
3232                         if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
3233                                 mlx5_set_port_status(priv->mdev,
3234                                     MLX5_PORT_DOWN);
3235                                 if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0)
3236                                         mlx5e_close_locked(ifp);
3237                                 mlx5e_update_carrier(priv);
3238                                 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
3239                         }
3240                 }
3241                 PRIV_UNLOCK(priv);
3242                 break;
3243         case SIOCADDMULTI:
3244         case SIOCDELMULTI:
3245                 mlx5e_set_rx_mode(ifp);
3246                 break;
3247         case SIOCSIFMEDIA:
3248         case SIOCGIFMEDIA:
3249         case SIOCGIFXMEDIA:
3250                 ifr = (struct ifreq *)data;
3251                 error = ifmedia_ioctl(ifp, ifr, &priv->media, command);
3252                 break;
3253         case SIOCSIFCAP:
3254                 ifr = (struct ifreq *)data;
3255                 PRIV_LOCK(priv);
3256                 mask = ifr->ifr_reqcap ^ ifp->if_capenable;
3257
3258                 if (mask & IFCAP_TXCSUM) {
3259                         ifp->if_capenable ^= IFCAP_TXCSUM;
3260                         ifp->if_hwassist ^= (CSUM_TCP | CSUM_UDP | CSUM_IP);
3261
3262                         if (IFCAP_TSO4 & ifp->if_capenable &&
3263                             !(IFCAP_TXCSUM & ifp->if_capenable)) {
3264                                 ifp->if_capenable &= ~IFCAP_TSO4;
3265                                 ifp->if_hwassist &= ~CSUM_IP_TSO;
3266                                 if_printf(ifp,
3267                                     "tso4 disabled due to -txcsum.\n");
3268                         }
3269                 }
3270                 if (mask & IFCAP_TXCSUM_IPV6) {
3271                         ifp->if_capenable ^= IFCAP_TXCSUM_IPV6;
3272                         ifp->if_hwassist ^= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6);
3273
3274                         if (IFCAP_TSO6 & ifp->if_capenable &&
3275                             !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) {
3276                                 ifp->if_capenable &= ~IFCAP_TSO6;
3277                                 ifp->if_hwassist &= ~CSUM_IP6_TSO;
3278                                 if_printf(ifp,
3279                                     "tso6 disabled due to -txcsum6.\n");
3280                         }
3281                 }
3282                 if (mask & IFCAP_RXCSUM)
3283                         ifp->if_capenable ^= IFCAP_RXCSUM;
3284                 if (mask & IFCAP_RXCSUM_IPV6)
3285                         ifp->if_capenable ^= IFCAP_RXCSUM_IPV6;
3286                 if (mask & IFCAP_TSO4) {
3287                         if (!(IFCAP_TSO4 & ifp->if_capenable) &&
3288                             !(IFCAP_TXCSUM & ifp->if_capenable)) {
3289                                 if_printf(ifp, "enable txcsum first.\n");
3290                                 error = EAGAIN;
3291                                 goto out;
3292                         }
3293                         ifp->if_capenable ^= IFCAP_TSO4;
3294                         ifp->if_hwassist ^= CSUM_IP_TSO;
3295                 }
3296                 if (mask & IFCAP_TSO6) {
3297                         if (!(IFCAP_TSO6 & ifp->if_capenable) &&
3298                             !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) {
3299                                 if_printf(ifp, "enable txcsum6 first.\n");
3300                                 error = EAGAIN;
3301                                 goto out;
3302                         }
3303                         ifp->if_capenable ^= IFCAP_TSO6;
3304                         ifp->if_hwassist ^= CSUM_IP6_TSO;
3305                 }
3306                 if (mask & IFCAP_VLAN_HWFILTER) {
3307                         if (ifp->if_capenable & IFCAP_VLAN_HWFILTER)
3308                                 mlx5e_disable_vlan_filter(priv);
3309                         else
3310                                 mlx5e_enable_vlan_filter(priv);
3311
3312                         ifp->if_capenable ^= IFCAP_VLAN_HWFILTER;
3313                 }
3314                 if (mask & IFCAP_VLAN_HWTAGGING)
3315                         ifp->if_capenable ^= IFCAP_VLAN_HWTAGGING;
3316                 if (mask & IFCAP_WOL_MAGIC)
3317                         ifp->if_capenable ^= IFCAP_WOL_MAGIC;
3318
3319                 VLAN_CAPABILITIES(ifp);
3320                 /* turn off LRO means also turn of HW LRO - if it's on */
3321                 if (mask & IFCAP_LRO) {
3322                         int was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
3323                         bool need_restart = false;
3324
3325                         ifp->if_capenable ^= IFCAP_LRO;
3326
3327                         /* figure out if updating HW LRO is needed */
3328                         if (!(ifp->if_capenable & IFCAP_LRO)) {
3329                                 if (priv->params.hw_lro_en) {
3330                                         priv->params.hw_lro_en = false;
3331                                         need_restart = true;
3332                                 }
3333                         } else {
3334                                 if (priv->params.hw_lro_en == false &&
3335                                     priv->params_ethtool.hw_lro != 0) {
3336                                         priv->params.hw_lro_en = true;
3337                                         need_restart = true;
3338                                 }
3339                         }
3340                         if (was_opened && need_restart) {
3341                                 mlx5e_close_locked(ifp);
3342                                 mlx5e_open_locked(ifp);
3343                         }
3344                 }
3345                 if (mask & IFCAP_HWRXTSTMP) {
3346                         ifp->if_capenable ^= IFCAP_HWRXTSTMP;
3347                         if (ifp->if_capenable & IFCAP_HWRXTSTMP) {
3348                                 if (priv->clbr_done == 0)
3349                                         mlx5e_reset_calibration_callout(priv);
3350                         } else {
3351                                 callout_drain(&priv->tstmp_clbr);
3352                                 priv->clbr_done = 0;
3353                         }
3354                 }
3355 out:
3356                 PRIV_UNLOCK(priv);
3357                 break;
3358
3359         case SIOCGI2C:
3360                 ifr = (struct ifreq *)data;
3361
3362                 /*
3363                  * Copy from the user-space address ifr_data to the
3364                  * kernel-space address i2c
3365                  */
3366                 error = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
3367                 if (error)
3368                         break;
3369
3370                 if (i2c.len > sizeof(i2c.data)) {
3371                         error = EINVAL;
3372                         break;
3373                 }
3374
3375                 PRIV_LOCK(priv);
3376                 /* Get module_num which is required for the query_eeprom */
3377                 error = mlx5_query_module_num(priv->mdev, &module_num);
3378                 if (error) {
3379                         if_printf(ifp, "Query module num failed, eeprom "
3380                             "reading is not supported\n");
3381                         error = EINVAL;
3382                         goto err_i2c;
3383                 }
3384                 /* Check if module is present before doing an access */
3385                 module_status = mlx5_query_module_status(priv->mdev, module_num);
3386                 if (module_status != MLX5_MODULE_STATUS_PLUGGED_ENABLED &&
3387                     module_status != MLX5_MODULE_STATUS_PLUGGED_DISABLED) {
3388                         error = EINVAL;
3389                         goto err_i2c;
3390                 }
3391                 /*
3392                  * Currently 0XA0 and 0xA2 are the only addresses permitted.
3393                  * The internal conversion is as follows:
3394                  */
3395                 if (i2c.dev_addr == 0xA0)
3396                         read_addr = MLX5E_I2C_ADDR_LOW;
3397                 else if (i2c.dev_addr == 0xA2)
3398                         read_addr = MLX5E_I2C_ADDR_HIGH;
3399                 else {
3400                         if_printf(ifp, "Query eeprom failed, "
3401                             "Invalid Address: %X\n", i2c.dev_addr);
3402                         error = EINVAL;
3403                         goto err_i2c;
3404                 }
3405                 error = mlx5_query_eeprom(priv->mdev,
3406                     read_addr, MLX5E_EEPROM_LOW_PAGE,
3407                     (uint32_t)i2c.offset, (uint32_t)i2c.len, module_num,
3408                     (uint32_t *)i2c.data, &size_read);
3409                 if (error) {
3410                         if_printf(ifp, "Query eeprom failed, eeprom "
3411                             "reading is not supported\n");
3412                         error = EINVAL;
3413                         goto err_i2c;
3414                 }
3415
3416                 if (i2c.len > MLX5_EEPROM_MAX_BYTES) {
3417                         error = mlx5_query_eeprom(priv->mdev,
3418                             read_addr, MLX5E_EEPROM_LOW_PAGE,
3419                             (uint32_t)(i2c.offset + size_read),
3420                             (uint32_t)(i2c.len - size_read), module_num,
3421                             (uint32_t *)(i2c.data + size_read), &size_read);
3422                 }
3423                 if (error) {
3424                         if_printf(ifp, "Query eeprom failed, eeprom "
3425                             "reading is not supported\n");
3426                         error = EINVAL;
3427                         goto err_i2c;
3428                 }
3429
3430                 error = copyout(&i2c, ifr_data_get_ptr(ifr), sizeof(i2c));
3431 err_i2c:
3432                 PRIV_UNLOCK(priv);
3433                 break;
3434
3435         default:
3436                 error = ether_ioctl(ifp, command, data);
3437                 break;
3438         }
3439         return (error);
3440 }
3441
3442 static int
3443 mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
3444 {
3445         /*
3446          * TODO: uncoment once FW really sets all these bits if
3447          * (!mdev->caps.eth.rss_ind_tbl_cap || !mdev->caps.eth.csum_cap ||
3448          * !mdev->caps.eth.max_lso_cap || !mdev->caps.eth.vlan_cap ||
3449          * !(mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_SCQE_BRK_MOD)) return
3450          * -ENOTSUPP;
3451          */
3452
3453         /* TODO: add more must-to-have features */
3454
3455         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
3456                 return (-ENODEV);
3457
3458         return (0);
3459 }
3460
3461 static u16
3462 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
3463 {
3464         uint32_t bf_buf_size = (1U << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2U;
3465
3466         bf_buf_size -= sizeof(struct mlx5e_tx_wqe) - 2;
3467
3468         /* verify against driver hardware limit */
3469         if (bf_buf_size > MLX5E_MAX_TX_INLINE)
3470                 bf_buf_size = MLX5E_MAX_TX_INLINE;
3471
3472         return (bf_buf_size);
3473 }
3474
3475 static int
3476 mlx5e_build_ifp_priv(struct mlx5_core_dev *mdev,
3477     struct mlx5e_priv *priv,
3478     int num_comp_vectors)
3479 {
3480         int err;
3481
3482         /*
3483          * TODO: Consider link speed for setting "log_sq_size",
3484          * "log_rq_size" and "cq_moderation_xxx":
3485          */
3486         priv->params.log_sq_size =
3487             MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
3488         priv->params.log_rq_size =
3489             MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
3490         priv->params.rx_cq_moderation_usec =
3491             MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
3492             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE :
3493             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
3494         priv->params.rx_cq_moderation_mode =
3495             MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ? 1 : 0;
3496         priv->params.rx_cq_moderation_pkts =
3497             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
3498         priv->params.tx_cq_moderation_usec =
3499             MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
3500         priv->params.tx_cq_moderation_pkts =
3501             MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
3502         priv->params.min_rx_wqes =
3503             MLX5E_PARAMS_DEFAULT_MIN_RX_WQES;
3504         priv->params.rx_hash_log_tbl_sz =
3505             (order_base_2(num_comp_vectors) >
3506             MLX5E_PARAMS_DEFAULT_RX_HASH_LOG_TBL_SZ) ?
3507             order_base_2(num_comp_vectors) :
3508             MLX5E_PARAMS_DEFAULT_RX_HASH_LOG_TBL_SZ;
3509         priv->params.num_tc = 1;
3510         priv->params.default_vlan_prio = 0;
3511         priv->counter_set_id = -1;
3512         priv->params.tx_max_inline = mlx5e_get_max_inline_cap(mdev);
3513
3514         err = mlx5_query_min_inline(mdev, &priv->params.tx_min_inline_mode);
3515         if (err)
3516                 return (err);
3517
3518         /*
3519          * hw lro is currently defaulted to off. when it won't anymore we
3520          * will consider the HW capability: "!!MLX5_CAP_ETH(mdev, lro_cap)"
3521          */
3522         priv->params.hw_lro_en = false;
3523         priv->params.lro_wqe_sz = MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ;
3524
3525         /*
3526          * CQE zipping is currently defaulted to off. when it won't
3527          * anymore we will consider the HW capability:
3528          * "!!MLX5_CAP_GEN(mdev, cqe_compression)"
3529          */
3530         priv->params.cqe_zipping_en = false;
3531
3532         priv->mdev = mdev;
3533         priv->params.num_channels = num_comp_vectors;
3534         priv->params.channels_rsss = 1;
3535         priv->order_base_2_num_channels = order_base_2(num_comp_vectors);
3536         priv->queue_mapping_channel_mask =
3537             roundup_pow_of_two(num_comp_vectors) - 1;
3538         priv->num_tc = priv->params.num_tc;
3539         priv->default_vlan_prio = priv->params.default_vlan_prio;
3540
3541         INIT_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
3542         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
3543         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
3544
3545         return (0);
3546 }
3547
3548 static int
3549 mlx5e_create_mkey(struct mlx5e_priv *priv, u32 pdn,
3550                   struct mlx5_core_mr *mkey)
3551 {
3552         struct ifnet *ifp = priv->ifp;
3553         struct mlx5_core_dev *mdev = priv->mdev;
3554         int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
3555         void *mkc;
3556         u32 *in;
3557         int err;
3558
3559         in = mlx5_vzalloc(inlen);
3560         if (in == NULL) {
3561                 if_printf(ifp, "%s: failed to allocate inbox\n", __func__);
3562                 return (-ENOMEM);
3563         }
3564
3565         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
3566         MLX5_SET(mkc, mkc, access_mode, MLX5_ACCESS_MODE_PA);
3567         MLX5_SET(mkc, mkc, lw, 1);
3568         MLX5_SET(mkc, mkc, lr, 1);
3569
3570         MLX5_SET(mkc, mkc, pd, pdn);
3571         MLX5_SET(mkc, mkc, length64, 1);
3572         MLX5_SET(mkc, mkc, qpn, 0xffffff);
3573
3574         err = mlx5_core_create_mkey(mdev, mkey, in, inlen);
3575         if (err)
3576                 if_printf(ifp, "%s: mlx5_core_create_mkey failed, %d\n",
3577                     __func__, err);
3578
3579         kvfree(in);
3580         return (err);
3581 }
3582
3583 static const char *mlx5e_vport_stats_desc[] = {
3584         MLX5E_VPORT_STATS(MLX5E_STATS_DESC)
3585 };
3586
3587 static const char *mlx5e_pport_stats_desc[] = {
3588         MLX5E_PPORT_STATS(MLX5E_STATS_DESC)
3589 };
3590
3591 static void
3592 mlx5e_priv_mtx_init(struct mlx5e_priv *priv)
3593 {
3594         mtx_init(&priv->async_events_mtx, "mlx5async", MTX_NETWORK_LOCK, MTX_DEF);
3595         sx_init(&priv->state_lock, "mlx5state");
3596         callout_init_mtx(&priv->watchdog, &priv->async_events_mtx, 0);
3597         MLX5_INIT_DOORBELL_LOCK(&priv->doorbell_lock);
3598 }
3599
3600 static void
3601 mlx5e_priv_mtx_destroy(struct mlx5e_priv *priv)
3602 {
3603         mtx_destroy(&priv->async_events_mtx);
3604         sx_destroy(&priv->state_lock);
3605 }
3606
3607 static int
3608 sysctl_firmware(SYSCTL_HANDLER_ARGS)
3609 {
3610         /*
3611          * %d.%d%.d the string format.
3612          * fw_rev_{maj,min,sub} return u16, 2^16 = 65536.
3613          * We need at most 5 chars to store that.
3614          * It also has: two "." and NULL at the end, which means we need 18
3615          * (5*3 + 3) chars at most.
3616          */
3617         char fw[18];
3618         struct mlx5e_priv *priv = arg1;
3619         int error;
3620
3621         snprintf(fw, sizeof(fw), "%d.%d.%d", fw_rev_maj(priv->mdev), fw_rev_min(priv->mdev),
3622             fw_rev_sub(priv->mdev));
3623         error = sysctl_handle_string(oidp, fw, sizeof(fw), req);
3624         return (error);
3625 }
3626
3627 static void
3628 mlx5e_disable_tx_dma(struct mlx5e_channel *ch)
3629 {
3630         int i;
3631
3632         for (i = 0; i < ch->num_tc; i++)
3633                 mlx5e_drain_sq(&ch->sq[i]);
3634 }
3635
3636 static void
3637 mlx5e_reset_sq_doorbell_record(struct mlx5e_sq *sq)
3638 {
3639
3640         sq->doorbell.d32[0] = cpu_to_be32(MLX5_OPCODE_NOP);
3641         sq->doorbell.d32[1] = cpu_to_be32(sq->sqn << 8);
3642         mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
3643         sq->doorbell.d64 = 0;
3644 }
3645
3646 void
3647 mlx5e_resume_sq(struct mlx5e_sq *sq)
3648 {
3649         int err;
3650
3651         /* check if already enabled */
3652         if (READ_ONCE(sq->running) != 0)
3653                 return;
3654
3655         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_ERR,
3656             MLX5_SQC_STATE_RST);
3657         if (err != 0) {
3658                 if_printf(sq->ifp,
3659                     "mlx5e_modify_sq() from ERR to RST failed: %d\n", err);
3660         }
3661
3662         sq->cc = 0;
3663         sq->pc = 0;
3664
3665         /* reset doorbell prior to moving from RST to RDY */
3666         mlx5e_reset_sq_doorbell_record(sq);
3667
3668         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST,
3669             MLX5_SQC_STATE_RDY);
3670         if (err != 0) {
3671                 if_printf(sq->ifp,
3672                     "mlx5e_modify_sq() from RST to RDY failed: %d\n", err);
3673         }
3674
3675         sq->cev_next_state = MLX5E_CEV_STATE_INITIAL;
3676         WRITE_ONCE(sq->running, 1);
3677 }
3678
3679 static void
3680 mlx5e_enable_tx_dma(struct mlx5e_channel *ch)
3681 {
3682         int i;
3683
3684         for (i = 0; i < ch->num_tc; i++)
3685                 mlx5e_resume_sq(&ch->sq[i]);
3686 }
3687
3688 static void
3689 mlx5e_disable_rx_dma(struct mlx5e_channel *ch)
3690 {
3691         struct mlx5e_rq *rq = &ch->rq;
3692         int err;
3693
3694         mtx_lock(&rq->mtx);
3695         rq->enabled = 0;
3696         callout_stop(&rq->watchdog);
3697         mtx_unlock(&rq->mtx);
3698
3699         callout_drain(&rq->watchdog);
3700
3701         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RDY, MLX5_RQC_STATE_ERR);
3702         if (err != 0) {
3703                 if_printf(rq->ifp,
3704                     "mlx5e_modify_rq() from RDY to RST failed: %d\n", err);
3705         }
3706
3707         while (!mlx5_wq_ll_is_empty(&rq->wq)) {
3708                 msleep(1);
3709                 rq->cq.mcq.comp(&rq->cq.mcq);
3710         }
3711
3712         /*
3713          * Transitioning into RST state will allow the FW to track less ERR state queues,
3714          * thus reducing the recv queue flushing time
3715          */
3716         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_ERR, MLX5_RQC_STATE_RST);
3717         if (err != 0) {
3718                 if_printf(rq->ifp,
3719                     "mlx5e_modify_rq() from ERR to RST failed: %d\n", err);
3720         }
3721 }
3722
3723 static void
3724 mlx5e_enable_rx_dma(struct mlx5e_channel *ch)
3725 {
3726         struct mlx5e_rq *rq = &ch->rq;
3727         int err;
3728
3729         rq->wq.wqe_ctr = 0;
3730         mlx5_wq_ll_update_db_record(&rq->wq);
3731         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
3732         if (err != 0) {
3733                 if_printf(rq->ifp,
3734                     "mlx5e_modify_rq() from RST to RDY failed: %d\n", err);
3735         }
3736
3737         rq->enabled = 1;
3738
3739         rq->cq.mcq.comp(&rq->cq.mcq);
3740 }
3741
3742 void
3743 mlx5e_modify_tx_dma(struct mlx5e_priv *priv, uint8_t value)
3744 {
3745         int i;
3746
3747         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
3748                 return;
3749
3750         for (i = 0; i < priv->params.num_channels; i++) {
3751                 if (value)
3752                         mlx5e_disable_tx_dma(&priv->channel[i]);
3753                 else
3754                         mlx5e_enable_tx_dma(&priv->channel[i]);
3755         }
3756 }
3757
3758 void
3759 mlx5e_modify_rx_dma(struct mlx5e_priv *priv, uint8_t value)
3760 {
3761         int i;
3762
3763         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
3764                 return;
3765
3766         for (i = 0; i < priv->params.num_channels; i++) {
3767                 if (value)
3768                         mlx5e_disable_rx_dma(&priv->channel[i]);
3769                 else
3770                         mlx5e_enable_rx_dma(&priv->channel[i]);
3771         }
3772 }
3773
3774 static void
3775 mlx5e_add_hw_stats(struct mlx5e_priv *priv)
3776 {
3777         SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_hw),
3778             OID_AUTO, "fw_version", CTLTYPE_STRING | CTLFLAG_RD, priv, 0,
3779             sysctl_firmware, "A", "HCA firmware version");
3780
3781         SYSCTL_ADD_STRING(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_hw),
3782             OID_AUTO, "board_id", CTLFLAG_RD, priv->mdev->board_id, 0,
3783             "Board ID");
3784 }
3785
3786 static int
3787 mlx5e_sysctl_tx_priority_flow_control(SYSCTL_HANDLER_ARGS)
3788 {
3789         struct mlx5e_priv *priv = arg1;
3790         uint8_t temp[MLX5E_MAX_PRIORITY];
3791         uint32_t tx_pfc;
3792         int err;
3793         int i;
3794
3795         PRIV_LOCK(priv);
3796
3797         tx_pfc = priv->params.tx_priority_flow_control;
3798
3799         for (i = 0; i != MLX5E_MAX_PRIORITY; i++)
3800                 temp[i] = (tx_pfc >> i) & 1;
3801
3802         err = SYSCTL_OUT(req, temp, MLX5E_MAX_PRIORITY);
3803         if (err || !req->newptr)
3804                 goto done;
3805         err = SYSCTL_IN(req, temp, MLX5E_MAX_PRIORITY);
3806         if (err)
3807                 goto done;
3808
3809         priv->params.tx_priority_flow_control = 0;
3810
3811         /* range check input value */
3812         for (i = 0; i != MLX5E_MAX_PRIORITY; i++) {
3813                 if (temp[i] > 1) {
3814                         err = ERANGE;
3815                         goto done;
3816                 }
3817                 priv->params.tx_priority_flow_control |= (temp[i] << i);
3818         }
3819
3820         /* check if update is required */
3821         if (tx_pfc != priv->params.tx_priority_flow_control)
3822                 err = -mlx5e_set_port_pfc(priv);
3823 done:
3824         if (err != 0)
3825                 priv->params.tx_priority_flow_control= tx_pfc;
3826         PRIV_UNLOCK(priv);
3827
3828         return (err);
3829 }
3830
3831 static int
3832 mlx5e_sysctl_rx_priority_flow_control(SYSCTL_HANDLER_ARGS)
3833 {
3834         struct mlx5e_priv *priv = arg1;
3835         uint8_t temp[MLX5E_MAX_PRIORITY];
3836         uint32_t rx_pfc;
3837         int err;
3838         int i;
3839
3840         PRIV_LOCK(priv);
3841
3842         rx_pfc = priv->params.rx_priority_flow_control;
3843
3844         for (i = 0; i != MLX5E_MAX_PRIORITY; i++)
3845                 temp[i] = (rx_pfc >> i) & 1;
3846
3847         err = SYSCTL_OUT(req, temp, MLX5E_MAX_PRIORITY);
3848         if (err || !req->newptr)
3849                 goto done;
3850         err = SYSCTL_IN(req, temp, MLX5E_MAX_PRIORITY);
3851         if (err)
3852                 goto done;
3853
3854         priv->params.rx_priority_flow_control = 0;
3855
3856         /* range check input value */
3857         for (i = 0; i != MLX5E_MAX_PRIORITY; i++) {
3858                 if (temp[i] > 1) {
3859                         err = ERANGE;
3860                         goto done;
3861                 }
3862                 priv->params.rx_priority_flow_control |= (temp[i] << i);
3863         }
3864
3865         /* check if update is required */
3866         if (rx_pfc != priv->params.rx_priority_flow_control)
3867                 err = -mlx5e_set_port_pfc(priv);
3868 done:
3869         if (err != 0)
3870                 priv->params.rx_priority_flow_control= rx_pfc;
3871         PRIV_UNLOCK(priv);
3872
3873         return (err);
3874 }
3875
3876 static void
3877 mlx5e_setup_pauseframes(struct mlx5e_priv *priv)
3878 {
3879 #if (__FreeBSD_version < 1100000)
3880         char path[96];
3881 #endif
3882         int error;
3883
3884         /* enable pauseframes by default */
3885         priv->params.tx_pauseframe_control = 1;
3886         priv->params.rx_pauseframe_control = 1;
3887
3888         /* disable ports flow control, PFC, by default */
3889         priv->params.tx_priority_flow_control = 0;
3890         priv->params.rx_priority_flow_control = 0;
3891
3892 #if (__FreeBSD_version < 1100000)
3893         /* compute path for sysctl */
3894         snprintf(path, sizeof(path), "dev.mce.%d.tx_pauseframe_control",
3895             device_get_unit(priv->mdev->pdev->dev.bsddev));
3896
3897         /* try to fetch tunable, if any */
3898         TUNABLE_INT_FETCH(path, &priv->params.tx_pauseframe_control);
3899
3900         /* compute path for sysctl */
3901         snprintf(path, sizeof(path), "dev.mce.%d.rx_pauseframe_control",
3902             device_get_unit(priv->mdev->pdev->dev.bsddev));
3903
3904         /* try to fetch tunable, if any */
3905         TUNABLE_INT_FETCH(path, &priv->params.rx_pauseframe_control);
3906 #endif
3907
3908         /* register pauseframe SYSCTLs */
3909         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3910             OID_AUTO, "tx_pauseframe_control", CTLFLAG_RDTUN,
3911             &priv->params.tx_pauseframe_control, 0,
3912             "Set to enable TX pause frames. Clear to disable.");
3913
3914         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3915             OID_AUTO, "rx_pauseframe_control", CTLFLAG_RDTUN,
3916             &priv->params.rx_pauseframe_control, 0,
3917             "Set to enable RX pause frames. Clear to disable.");
3918
3919         /* register priority flow control, PFC, SYSCTLs */
3920         SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3921             OID_AUTO, "tx_priority_flow_control", CTLTYPE_U8 | CTLFLAG_RWTUN |
3922             CTLFLAG_MPSAFE, priv, 0, &mlx5e_sysctl_tx_priority_flow_control, "CU",
3923             "Set to enable TX ports flow control frames for priorities 0..7. Clear to disable.");
3924
3925         SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3926             OID_AUTO, "rx_priority_flow_control", CTLTYPE_U8 | CTLFLAG_RWTUN |
3927             CTLFLAG_MPSAFE, priv, 0, &mlx5e_sysctl_rx_priority_flow_control, "CU",
3928             "Set to enable RX ports flow control frames for priorities 0..7. Clear to disable.");
3929
3930         PRIV_LOCK(priv);
3931
3932         /* range check */
3933         priv->params.tx_pauseframe_control =
3934             priv->params.tx_pauseframe_control ? 1 : 0;
3935         priv->params.rx_pauseframe_control =
3936             priv->params.rx_pauseframe_control ? 1 : 0;
3937
3938         /* update firmware */
3939         error = mlx5e_set_port_pause_and_pfc(priv);
3940         if (error == -EINVAL) {
3941                 if_printf(priv->ifp,
3942                     "Global pauseframes must be disabled before enabling PFC.\n");
3943                 priv->params.rx_priority_flow_control = 0;
3944                 priv->params.tx_priority_flow_control = 0;
3945
3946                 /* update firmware */
3947                 (void) mlx5e_set_port_pause_and_pfc(priv);
3948         }
3949         PRIV_UNLOCK(priv);
3950 }
3951
3952 static int
3953 mlx5e_ul_snd_tag_alloc(struct ifnet *ifp,
3954     union if_snd_tag_alloc_params *params,
3955     struct m_snd_tag **ppmt)
3956 {
3957         struct mlx5e_priv *priv;
3958         struct mlx5e_channel *pch;
3959
3960         priv = ifp->if_softc;
3961
3962         if (unlikely(priv->gone || params->hdr.flowtype == M_HASHTYPE_NONE)) {
3963                 return (EOPNOTSUPP);
3964         } else {
3965                 /* keep this code synced with mlx5e_select_queue() */
3966                 u32 ch = priv->params.num_channels;
3967 #ifdef RSS
3968                 u32 temp;
3969
3970                 if (rss_hash2bucket(params->hdr.flowid,
3971                     params->hdr.flowtype, &temp) == 0)
3972                         ch = temp % ch;
3973                 else
3974 #endif
3975                         ch = (params->hdr.flowid % 128) % ch;
3976
3977                 /*
3978                  * NOTE: The channels array is only freed at detach
3979                  * and it safe to return a pointer to the send tag
3980                  * inside the channels structure as long as we
3981                  * reference the priv.
3982                  */
3983                 pch = priv->channel + ch;
3984
3985                 /* check if send queue is not running */
3986                 if (unlikely(pch->sq[0].running == 0))
3987                         return (ENXIO);
3988                 mlx5e_ref_channel(priv);
3989                 MPASS(pch->tag.m_snd_tag.refcount == 0);
3990                 m_snd_tag_init(&pch->tag.m_snd_tag, ifp);
3991                 *ppmt = &pch->tag.m_snd_tag;
3992                 return (0);
3993         }
3994 }
3995
3996 static int
3997 mlx5e_ul_snd_tag_query(struct m_snd_tag *pmt, union if_snd_tag_query_params *params)
3998 {
3999         struct mlx5e_channel *pch =
4000             container_of(pmt, struct mlx5e_channel, tag.m_snd_tag);
4001
4002         params->unlimited.max_rate = -1ULL;
4003         params->unlimited.queue_level = mlx5e_sq_queue_level(&pch->sq[0]);
4004         return (0);
4005 }
4006
4007 static void
4008 mlx5e_ul_snd_tag_free(struct m_snd_tag *pmt)
4009 {
4010         struct mlx5e_channel *pch =
4011             container_of(pmt, struct mlx5e_channel, tag.m_snd_tag);
4012
4013         mlx5e_unref_channel(pch->priv);
4014 }
4015
4016 static int
4017 mlx5e_snd_tag_alloc(struct ifnet *ifp,
4018     union if_snd_tag_alloc_params *params,
4019     struct m_snd_tag **ppmt)
4020 {
4021
4022         switch (params->hdr.type) {
4023 #ifdef RATELIMIT
4024         case IF_SND_TAG_TYPE_RATE_LIMIT:
4025                 return (mlx5e_rl_snd_tag_alloc(ifp, params, ppmt));
4026 #endif
4027         case IF_SND_TAG_TYPE_UNLIMITED:
4028                 return (mlx5e_ul_snd_tag_alloc(ifp, params, ppmt));
4029         default:
4030                 return (EOPNOTSUPP);
4031         }
4032 }
4033
4034 static int
4035 mlx5e_snd_tag_modify(struct m_snd_tag *pmt, union if_snd_tag_modify_params *params)
4036 {
4037         struct mlx5e_snd_tag *tag =
4038             container_of(pmt, struct mlx5e_snd_tag, m_snd_tag);
4039
4040         switch (tag->type) {
4041 #ifdef RATELIMIT
4042         case IF_SND_TAG_TYPE_RATE_LIMIT:
4043                 return (mlx5e_rl_snd_tag_modify(pmt, params));
4044 #endif
4045         case IF_SND_TAG_TYPE_UNLIMITED:
4046         default:
4047                 return (EOPNOTSUPP);
4048         }
4049 }
4050
4051 static int
4052 mlx5e_snd_tag_query(struct m_snd_tag *pmt, union if_snd_tag_query_params *params)
4053 {
4054         struct mlx5e_snd_tag *tag =
4055             container_of(pmt, struct mlx5e_snd_tag, m_snd_tag);
4056
4057         switch (tag->type) {
4058 #ifdef RATELIMIT
4059         case IF_SND_TAG_TYPE_RATE_LIMIT:
4060                 return (mlx5e_rl_snd_tag_query(pmt, params));
4061 #endif
4062         case IF_SND_TAG_TYPE_UNLIMITED:
4063                 return (mlx5e_ul_snd_tag_query(pmt, params));
4064         default:
4065                 return (EOPNOTSUPP);
4066         }
4067 }
4068
4069 static void
4070 mlx5e_snd_tag_free(struct m_snd_tag *pmt)
4071 {
4072         struct mlx5e_snd_tag *tag =
4073             container_of(pmt, struct mlx5e_snd_tag, m_snd_tag);
4074
4075         switch (tag->type) {
4076 #ifdef RATELIMIT
4077         case IF_SND_TAG_TYPE_RATE_LIMIT:
4078                 mlx5e_rl_snd_tag_free(pmt);
4079                 break;
4080 #endif
4081         case IF_SND_TAG_TYPE_UNLIMITED:
4082                 mlx5e_ul_snd_tag_free(pmt);
4083                 break;
4084         default:
4085                 break;
4086         }
4087 }
4088
4089 static void *
4090 mlx5e_create_ifp(struct mlx5_core_dev *mdev)
4091 {
4092         struct ifnet *ifp;
4093         struct mlx5e_priv *priv;
4094         u8 dev_addr[ETHER_ADDR_LEN] __aligned(4);
4095         u8 connector_type;
4096         struct sysctl_oid_list *child;
4097         int ncv = mdev->priv.eq_table.num_comp_vectors;
4098         char unit[16];
4099         struct pfil_head_args pa;
4100         int err;
4101         int i,j;
4102         u32 eth_proto_cap;
4103         u32 out[MLX5_ST_SZ_DW(ptys_reg)];
4104         bool ext = 0;
4105         u32 speeds_num;
4106         struct media media_entry = {};
4107
4108         if (mlx5e_check_required_hca_cap(mdev)) {
4109                 mlx5_core_dbg(mdev, "mlx5e_check_required_hca_cap() failed\n");
4110                 return (NULL);
4111         }
4112         /*
4113          * Try to allocate the priv and make room for worst-case
4114          * number of channel structures:
4115          */
4116         priv = malloc(sizeof(*priv) +
4117             (sizeof(priv->channel[0]) * mdev->priv.eq_table.num_comp_vectors),
4118             M_MLX5EN, M_WAITOK | M_ZERO);
4119         mlx5e_priv_mtx_init(priv);
4120
4121         ifp = priv->ifp = if_alloc_dev(IFT_ETHER, mdev->pdev->dev.bsddev);
4122         if (ifp == NULL) {
4123                 mlx5_core_err(mdev, "if_alloc() failed\n");
4124                 goto err_free_priv;
4125         }
4126         ifp->if_softc = priv;
4127         if_initname(ifp, "mce", device_get_unit(mdev->pdev->dev.bsddev));
4128         ifp->if_mtu = ETHERMTU;
4129         ifp->if_init = mlx5e_open;
4130         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
4131         ifp->if_ioctl = mlx5e_ioctl;
4132         ifp->if_transmit = mlx5e_xmit;
4133         ifp->if_qflush = if_qflush;
4134 #if (__FreeBSD_version >= 1100000)
4135         ifp->if_get_counter = mlx5e_get_counter;
4136 #endif
4137         ifp->if_snd.ifq_maxlen = ifqmaxlen;
4138         /*
4139          * Set driver features
4140          */
4141         ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6;
4142         ifp->if_capabilities |= IFCAP_VLAN_MTU | IFCAP_VLAN_HWTAGGING;
4143         ifp->if_capabilities |= IFCAP_VLAN_HWCSUM | IFCAP_VLAN_HWFILTER;
4144         ifp->if_capabilities |= IFCAP_LINKSTATE | IFCAP_JUMBO_MTU;
4145         ifp->if_capabilities |= IFCAP_LRO;
4146         ifp->if_capabilities |= IFCAP_TSO | IFCAP_VLAN_HWTSO;
4147         ifp->if_capabilities |= IFCAP_HWSTATS | IFCAP_HWRXTSTMP;
4148         ifp->if_capabilities |= IFCAP_TXRTLMT;
4149         ifp->if_snd_tag_alloc = mlx5e_snd_tag_alloc;
4150         ifp->if_snd_tag_free = mlx5e_snd_tag_free;
4151         ifp->if_snd_tag_modify = mlx5e_snd_tag_modify;
4152         ifp->if_snd_tag_query = mlx5e_snd_tag_query;
4153
4154         /* set TSO limits so that we don't have to drop TX packets */
4155         ifp->if_hw_tsomax = MLX5E_MAX_TX_PAYLOAD_SIZE - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
4156         ifp->if_hw_tsomaxsegcount = MLX5E_MAX_TX_MBUF_FRAGS - 1 /* hdr */;
4157         ifp->if_hw_tsomaxsegsize = MLX5E_MAX_TX_MBUF_SIZE;
4158
4159         ifp->if_capenable = ifp->if_capabilities;
4160         ifp->if_hwassist = 0;
4161         if (ifp->if_capenable & IFCAP_TSO)
4162                 ifp->if_hwassist |= CSUM_TSO;
4163         if (ifp->if_capenable & IFCAP_TXCSUM)
4164                 ifp->if_hwassist |= (CSUM_TCP | CSUM_UDP | CSUM_IP);
4165         if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
4166                 ifp->if_hwassist |= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6);
4167
4168         /* ifnet sysctl tree */
4169         sysctl_ctx_init(&priv->sysctl_ctx);
4170         priv->sysctl_ifnet = SYSCTL_ADD_NODE(&priv->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_dev),
4171             OID_AUTO, ifp->if_dname, CTLFLAG_RD, 0, "MLX5 ethernet - interface name");
4172         if (priv->sysctl_ifnet == NULL) {
4173                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
4174                 goto err_free_sysctl;
4175         }
4176         snprintf(unit, sizeof(unit), "%d", ifp->if_dunit);
4177         priv->sysctl_ifnet = SYSCTL_ADD_NODE(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
4178             OID_AUTO, unit, CTLFLAG_RD, 0, "MLX5 ethernet - interface unit");
4179         if (priv->sysctl_ifnet == NULL) {
4180                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
4181                 goto err_free_sysctl;
4182         }
4183
4184         /* HW sysctl tree */
4185         child = SYSCTL_CHILDREN(device_get_sysctl_tree(mdev->pdev->dev.bsddev));
4186         priv->sysctl_hw = SYSCTL_ADD_NODE(&priv->sysctl_ctx, child,
4187             OID_AUTO, "hw", CTLFLAG_RD, 0, "MLX5 ethernet dev hw");
4188         if (priv->sysctl_hw == NULL) {
4189                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
4190                 goto err_free_sysctl;
4191         }
4192
4193         err = mlx5e_build_ifp_priv(mdev, priv, ncv);
4194         if (err) {
4195                 mlx5_core_err(mdev, "mlx5e_build_ifp_priv() failed (%d)\n", err);
4196                 goto err_free_sysctl;
4197         }
4198
4199         /* reuse mlx5core's watchdog workqueue */
4200         priv->wq = mdev->priv.health.wq_watchdog;
4201
4202         err = mlx5_alloc_map_uar(mdev, &priv->cq_uar);
4203         if (err) {
4204                 if_printf(ifp, "%s: mlx5_alloc_map_uar failed, %d\n",
4205                     __func__, err);
4206                 goto err_free_wq;
4207         }
4208         err = mlx5_core_alloc_pd(mdev, &priv->pdn);
4209         if (err) {
4210                 if_printf(ifp, "%s: mlx5_core_alloc_pd failed, %d\n",
4211                     __func__, err);
4212                 goto err_unmap_free_uar;
4213         }
4214         err = mlx5_alloc_transport_domain(mdev, &priv->tdn);
4215         if (err) {
4216                 if_printf(ifp, "%s: mlx5_alloc_transport_domain failed, %d\n",
4217                     __func__, err);
4218                 goto err_dealloc_pd;
4219         }
4220         err = mlx5e_create_mkey(priv, priv->pdn, &priv->mr);
4221         if (err) {
4222                 if_printf(ifp, "%s: mlx5e_create_mkey failed, %d\n",
4223                     __func__, err);
4224                 goto err_dealloc_transport_domain;
4225         }
4226         mlx5_query_nic_vport_mac_address(priv->mdev, 0, dev_addr);
4227
4228         /* check if we should generate a random MAC address */
4229         if (MLX5_CAP_GEN(priv->mdev, vport_group_manager) == 0 &&
4230             is_zero_ether_addr(dev_addr)) {
4231                 random_ether_addr(dev_addr);
4232                 if_printf(ifp, "Assigned random MAC address\n");
4233         }
4234 #ifdef RATELIMIT
4235         err = mlx5e_rl_init(priv);
4236         if (err) {
4237                 if_printf(ifp, "%s: mlx5e_rl_init failed, %d\n",
4238                     __func__, err);
4239                 goto err_create_mkey;
4240         }
4241 #endif
4242
4243         /* set default MTU */
4244         mlx5e_set_dev_port_mtu(ifp, ifp->if_mtu);
4245
4246         /* Set default media status */
4247         priv->media_status_last = IFM_AVALID;
4248         priv->media_active_last = IFM_ETHER | IFM_AUTO |
4249             IFM_ETH_RXPAUSE | IFM_FDX;
4250
4251         /* setup default pauseframes configuration */
4252         mlx5e_setup_pauseframes(priv);
4253
4254         /* Setup supported medias */
4255         //TODO: If we failed to query ptys is it ok to proceed??
4256         if (!mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN, 1)) {
4257                 ext = MLX5_CAP_PCAM_FEATURE(mdev,
4258                     ptys_extended_ethernet);
4259                 eth_proto_cap = MLX5_GET_ETH_PROTO(ptys_reg, out, ext,
4260                     eth_proto_capability);
4261                 if (MLX5_CAP_PCAM_FEATURE(mdev, ptys_connector_type))
4262                         connector_type = MLX5_GET(ptys_reg, out,
4263                             connector_type);
4264         } else {
4265                 eth_proto_cap = 0;
4266                 if_printf(ifp, "%s: Query port media capability failed,"
4267                     " %d\n", __func__, err);
4268         }
4269
4270         ifmedia_init(&priv->media, IFM_IMASK | IFM_ETH_FMASK,
4271             mlx5e_media_change, mlx5e_media_status);
4272
4273         speeds_num = ext ? MLX5E_EXT_LINK_SPEEDS_NUMBER : MLX5E_LINK_SPEEDS_NUMBER;
4274         for (i = 0; i != speeds_num; i++) {
4275                 for (j = 0; j < MLX5E_LINK_MODES_NUMBER ; ++j) {
4276                         media_entry = ext ? mlx5e_ext_mode_table[i][j] :
4277                             mlx5e_mode_table[i][j];
4278                         if (media_entry.baudrate == 0)
4279                                 continue;
4280                         if (MLX5E_PROT_MASK(i) & eth_proto_cap) {
4281                                 ifmedia_add(&priv->media,
4282                                     media_entry.subtype |
4283                                     IFM_ETHER, 0, NULL);
4284                                 ifmedia_add(&priv->media,
4285                                     media_entry.subtype |
4286                                     IFM_ETHER | IFM_FDX |
4287                                     IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE, 0, NULL);
4288                         }
4289                 }
4290         }
4291
4292         ifmedia_add(&priv->media, IFM_ETHER | IFM_AUTO, 0, NULL);
4293         ifmedia_add(&priv->media, IFM_ETHER | IFM_AUTO | IFM_FDX |
4294             IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE, 0, NULL);
4295
4296         /* Set autoselect by default */
4297         ifmedia_set(&priv->media, IFM_ETHER | IFM_AUTO | IFM_FDX |
4298             IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE);
4299         ether_ifattach(ifp, dev_addr);
4300
4301         /* Register for VLAN events */
4302         priv->vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
4303             mlx5e_vlan_rx_add_vid, priv, EVENTHANDLER_PRI_FIRST);
4304         priv->vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
4305             mlx5e_vlan_rx_kill_vid, priv, EVENTHANDLER_PRI_FIRST);
4306
4307         /* Link is down by default */
4308         if_link_state_change(ifp, LINK_STATE_DOWN);
4309
4310         mlx5e_enable_async_events(priv);
4311
4312         mlx5e_add_hw_stats(priv);
4313
4314         mlx5e_create_stats(&priv->stats.vport.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
4315             "vstats", mlx5e_vport_stats_desc, MLX5E_VPORT_STATS_NUM,
4316             priv->stats.vport.arg);
4317
4318         mlx5e_create_stats(&priv->stats.pport.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
4319             "pstats", mlx5e_pport_stats_desc, MLX5E_PPORT_STATS_NUM,
4320             priv->stats.pport.arg);
4321
4322         mlx5e_create_ethtool(priv);
4323
4324         mtx_lock(&priv->async_events_mtx);
4325         mlx5e_update_stats(priv);
4326         mtx_unlock(&priv->async_events_mtx);
4327
4328         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
4329             OID_AUTO, "rx_clbr_done", CTLFLAG_RD,
4330             &priv->clbr_done, 0,
4331             "RX timestamps calibration state");
4332         callout_init(&priv->tstmp_clbr, CALLOUT_DIRECT);
4333         mlx5e_reset_calibration_callout(priv);
4334
4335         pa.pa_version = PFIL_VERSION;
4336         pa.pa_flags = PFIL_IN;
4337         pa.pa_type = PFIL_TYPE_ETHERNET;
4338         pa.pa_headname = ifp->if_xname;
4339         priv->pfil = pfil_head_register(&pa);
4340
4341         return (priv);
4342
4343 #ifdef RATELIMIT
4344 err_create_mkey:
4345         mlx5_core_destroy_mkey(priv->mdev, &priv->mr);
4346 #endif
4347 err_dealloc_transport_domain:
4348         mlx5_dealloc_transport_domain(mdev, priv->tdn);
4349
4350 err_dealloc_pd:
4351         mlx5_core_dealloc_pd(mdev, priv->pdn);
4352
4353 err_unmap_free_uar:
4354         mlx5_unmap_free_uar(mdev, &priv->cq_uar);
4355
4356 err_free_wq:
4357         flush_workqueue(priv->wq);
4358
4359 err_free_sysctl:
4360         sysctl_ctx_free(&priv->sysctl_ctx);
4361         if (priv->sysctl_debug)
4362                 sysctl_ctx_free(&priv->stats.port_stats_debug.ctx);
4363         if_free(ifp);
4364
4365 err_free_priv:
4366         mlx5e_priv_mtx_destroy(priv);
4367         free(priv, M_MLX5EN);
4368         return (NULL);
4369 }
4370
4371 static void
4372 mlx5e_destroy_ifp(struct mlx5_core_dev *mdev, void *vpriv)
4373 {
4374         struct mlx5e_priv *priv = vpriv;
4375         struct ifnet *ifp = priv->ifp;
4376
4377         /* don't allow more IOCTLs */
4378         priv->gone = 1;
4379
4380         /* XXX wait a bit to allow IOCTL handlers to complete */
4381         pause("W", hz);
4382
4383 #ifdef RATELIMIT
4384         /*
4385          * The kernel can have reference(s) via the m_snd_tag's into
4386          * the ratelimit channels, and these must go away before
4387          * detaching:
4388          */
4389         while (READ_ONCE(priv->rl.stats.tx_active_connections) != 0) {
4390                 if_printf(priv->ifp, "Waiting for all ratelimit connections "
4391                     "to terminate\n");
4392                 pause("W", hz);
4393         }
4394 #endif
4395         /* stop watchdog timer */
4396         callout_drain(&priv->watchdog);
4397
4398         callout_drain(&priv->tstmp_clbr);
4399
4400         if (priv->vlan_attach != NULL)
4401                 EVENTHANDLER_DEREGISTER(vlan_config, priv->vlan_attach);
4402         if (priv->vlan_detach != NULL)
4403                 EVENTHANDLER_DEREGISTER(vlan_unconfig, priv->vlan_detach);
4404
4405         /* make sure device gets closed */
4406         PRIV_LOCK(priv);
4407         mlx5e_close_locked(ifp);
4408         PRIV_UNLOCK(priv);
4409
4410         /* wait for all unlimited send tags to go away */
4411         while (priv->channel_refs != 0) {
4412                 if_printf(priv->ifp, "Waiting for all unlimited connections "
4413                     "to terminate\n");
4414                 pause("W", hz);
4415         }
4416
4417         /* deregister pfil */
4418         if (priv->pfil != NULL) {
4419                 pfil_head_unregister(priv->pfil);
4420                 priv->pfil = NULL;
4421         }
4422
4423         /* unregister device */
4424         ifmedia_removeall(&priv->media);
4425         ether_ifdetach(ifp);
4426         if_free(ifp);
4427
4428 #ifdef RATELIMIT
4429         mlx5e_rl_cleanup(priv);
4430 #endif
4431         /* destroy all remaining sysctl nodes */
4432         sysctl_ctx_free(&priv->stats.vport.ctx);
4433         sysctl_ctx_free(&priv->stats.pport.ctx);
4434         if (priv->sysctl_debug)
4435                 sysctl_ctx_free(&priv->stats.port_stats_debug.ctx);
4436         sysctl_ctx_free(&priv->sysctl_ctx);
4437
4438         mlx5_core_destroy_mkey(priv->mdev, &priv->mr);
4439         mlx5_dealloc_transport_domain(priv->mdev, priv->tdn);
4440         mlx5_core_dealloc_pd(priv->mdev, priv->pdn);
4441         mlx5_unmap_free_uar(priv->mdev, &priv->cq_uar);
4442         mlx5e_disable_async_events(priv);
4443         flush_workqueue(priv->wq);
4444         mlx5e_priv_mtx_destroy(priv);
4445         free(priv, M_MLX5EN);
4446 }
4447
4448 static void *
4449 mlx5e_get_ifp(void *vpriv)
4450 {
4451         struct mlx5e_priv *priv = vpriv;
4452
4453         return (priv->ifp);
4454 }
4455
4456 static struct mlx5_interface mlx5e_interface = {
4457         .add = mlx5e_create_ifp,
4458         .remove = mlx5e_destroy_ifp,
4459         .event = mlx5e_async_event,
4460         .protocol = MLX5_INTERFACE_PROTOCOL_ETH,
4461         .get_dev = mlx5e_get_ifp,
4462 };
4463
4464 void
4465 mlx5e_init(void)
4466 {
4467         mlx5_register_interface(&mlx5e_interface);
4468 }
4469
4470 void
4471 mlx5e_cleanup(void)
4472 {
4473         mlx5_unregister_interface(&mlx5e_interface);
4474 }
4475
4476 static void
4477 mlx5e_show_version(void __unused *arg)
4478 {
4479
4480         printf("%s", mlx5e_version);
4481 }
4482 SYSINIT(mlx5e_show_version, SI_SUB_DRIVERS, SI_ORDER_ANY, mlx5e_show_version, NULL);
4483
4484 module_init_order(mlx5e_init, SI_ORDER_THIRD);
4485 module_exit_order(mlx5e_cleanup, SI_ORDER_THIRD);
4486
4487 #if (__FreeBSD_version >= 1100000)
4488 MODULE_DEPEND(mlx5en, linuxkpi, 1, 1, 1);
4489 #endif
4490 MODULE_DEPEND(mlx5en, mlx5, 1, 1, 1);
4491 MODULE_VERSION(mlx5en, 1);