sys/dev/mlx5/mlx5_en/mlx5_en_main.c

   1 /*-
   2  * Copyright (c) 2015 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/sockio.h>
  31 #include <machine/atomic.h>
  32
  33 #ifndef ETH_DRIVER_VERSION
  34 #define ETH_DRIVER_VERSION      "3.4.2"
  35 #endif
  36
  37 char mlx5e_version[] = "Mellanox Ethernet driver"
  38     " (" ETH_DRIVER_VERSION ")";
  39
  40 static int mlx5e_get_wqe_sz(struct mlx5e_priv *priv, u32 *wqe_sz, u32 *nsegs);
  41
  42 struct mlx5e_channel_param {
  43         struct mlx5e_rq_param rq;
  44         struct mlx5e_sq_param sq;
  45         struct mlx5e_cq_param rx_cq;
  46         struct mlx5e_cq_param tx_cq;
  47 };
  48
  49 static const struct {
  50         u32     subtype;
  51         u64     baudrate;
  52 }       mlx5e_mode_table[MLX5E_LINK_MODES_NUMBER] = {
  53
  54         [MLX5E_1000BASE_CX_SGMII] = {
  55                 .subtype = IFM_1000_CX_SGMII,
  56                 .baudrate = IF_Mbps(1000ULL),
  57         },
  58         [MLX5E_1000BASE_KX] = {
  59                 .subtype = IFM_1000_KX,
  60                 .baudrate = IF_Mbps(1000ULL),
  61         },
  62         [MLX5E_10GBASE_CX4] = {
  63                 .subtype = IFM_10G_CX4,
  64                 .baudrate = IF_Gbps(10ULL),
  65         },
  66         [MLX5E_10GBASE_KX4] = {
  67                 .subtype = IFM_10G_KX4,
  68                 .baudrate = IF_Gbps(10ULL),
  69         },
  70         [MLX5E_10GBASE_KR] = {
  71                 .subtype = IFM_10G_KR,
  72                 .baudrate = IF_Gbps(10ULL),
  73         },
  74         [MLX5E_20GBASE_KR2] = {
  75                 .subtype = IFM_20G_KR2,
  76                 .baudrate = IF_Gbps(20ULL),
  77         },
  78         [MLX5E_40GBASE_CR4] = {
  79                 .subtype = IFM_40G_CR4,
  80                 .baudrate = IF_Gbps(40ULL),
  81         },
  82         [MLX5E_40GBASE_KR4] = {
  83                 .subtype = IFM_40G_KR4,
  84                 .baudrate = IF_Gbps(40ULL),
  85         },
  86         [MLX5E_56GBASE_R4] = {
  87                 .subtype = IFM_56G_R4,
  88                 .baudrate = IF_Gbps(56ULL),
  89         },
  90         [MLX5E_10GBASE_CR] = {
  91                 .subtype = IFM_10G_CR1,
  92                 .baudrate = IF_Gbps(10ULL),
  93         },
  94         [MLX5E_10GBASE_SR] = {
  95                 .subtype = IFM_10G_SR,
  96                 .baudrate = IF_Gbps(10ULL),
  97         },
  98         [MLX5E_10GBASE_ER] = {
  99                 .subtype = IFM_10G_ER,
 100                 .baudrate = IF_Gbps(10ULL),
 101         },
 102         [MLX5E_40GBASE_SR4] = {
 103                 .subtype = IFM_40G_SR4,
 104                 .baudrate = IF_Gbps(40ULL),
 105         },
 106         [MLX5E_40GBASE_LR4] = {
 107                 .subtype = IFM_40G_LR4,
 108                 .baudrate = IF_Gbps(40ULL),
 109         },
 110         [MLX5E_100GBASE_CR4] = {
 111                 .subtype = IFM_100G_CR4,
 112                 .baudrate = IF_Gbps(100ULL),
 113         },
 114         [MLX5E_100GBASE_SR4] = {
 115                 .subtype = IFM_100G_SR4,
 116                 .baudrate = IF_Gbps(100ULL),
 117         },
 118         [MLX5E_100GBASE_KR4] = {
 119                 .subtype = IFM_100G_KR4,
 120                 .baudrate = IF_Gbps(100ULL),
 121         },
 122         [MLX5E_100GBASE_LR4] = {
 123                 .subtype = IFM_100G_LR4,
 124                 .baudrate = IF_Gbps(100ULL),
 125         },
 126         [MLX5E_100BASE_TX] = {
 127                 .subtype = IFM_100_TX,
 128                 .baudrate = IF_Mbps(100ULL),
 129         },
 130         [MLX5E_1000BASE_T] = {
 131                 .subtype = IFM_1000_T,
 132                 .baudrate = IF_Mbps(1000ULL),
 133         },
 134         [MLX5E_10GBASE_T] = {
 135                 .subtype = IFM_10G_T,
 136                 .baudrate = IF_Gbps(10ULL),
 137         },
 138         [MLX5E_25GBASE_CR] = {
 139                 .subtype = IFM_25G_CR,
 140                 .baudrate = IF_Gbps(25ULL),
 141         },
 142         [MLX5E_25GBASE_KR] = {
 143                 .subtype = IFM_25G_KR,
 144                 .baudrate = IF_Gbps(25ULL),
 145         },
 146         [MLX5E_25GBASE_SR] = {
 147                 .subtype = IFM_25G_SR,
 148                 .baudrate = IF_Gbps(25ULL),
 149         },
 150         [MLX5E_50GBASE_CR2] = {
 151                 .subtype = IFM_50G_CR2,
 152                 .baudrate = IF_Gbps(50ULL),
 153         },
 154         [MLX5E_50GBASE_KR2] = {
 155                 .subtype = IFM_50G_KR2,
 156                 .baudrate = IF_Gbps(50ULL),
 157         },
 158 };
 159
 160 MALLOC_DEFINE(M_MLX5EN, "MLX5EN", "MLX5 Ethernet");
 161
 162 static SYSCTL_NODE(_hw, OID_AUTO, mlx5, CTLFLAG_RW, 0, "MLX5 driver parameters");
 163
 164 static void
 165 mlx5e_update_carrier(struct mlx5e_priv *priv)
 166 {
 167         struct mlx5_core_dev *mdev = priv->mdev;
 168         u32 out[MLX5_ST_SZ_DW(ptys_reg)];
 169         u32 eth_proto_oper;
 170         int error;
 171         u8 port_state;
 172         u8 i;
 173
 174         port_state = mlx5_query_vport_state(mdev,
 175             MLX5_QUERY_VPORT_STATE_IN_OP_MOD_VNIC_VPORT, 0);
 176
 177         if (port_state == VPORT_STATE_UP) {
 178                 priv->media_status_last |= IFM_ACTIVE;
 179         } else {
 180                 priv->media_status_last &= ~IFM_ACTIVE;
 181                 priv->media_active_last = IFM_ETHER;
 182                 if_link_state_change(priv->ifp, LINK_STATE_DOWN);
 183                 return;
 184         }
 185
 186         error = mlx5_query_port_ptys(mdev, out, sizeof(out), MLX5_PTYS_EN, 1);
 187         if (error) {
 188                 priv->media_active_last = IFM_ETHER;
 189                 priv->ifp->if_baudrate = 1;
 190                 if_printf(priv->ifp, "%s: query port ptys failed: 0x%x\n",
 191                     __func__, error);
 192                 return;
 193         }
 194         eth_proto_oper = MLX5_GET(ptys_reg, out, eth_proto_oper);
 195
 196         for (i = 0; i != MLX5E_LINK_MODES_NUMBER; i++) {
 197                 if (mlx5e_mode_table[i].baudrate == 0)
 198                         continue;
 199                 if (MLX5E_PROT_MASK(i) & eth_proto_oper) {
 200                         priv->ifp->if_baudrate =
 201                             mlx5e_mode_table[i].baudrate;
 202                         priv->media_active_last =
 203                             mlx5e_mode_table[i].subtype | IFM_ETHER | IFM_FDX;
 204                 }
 205         }
 206         if_link_state_change(priv->ifp, LINK_STATE_UP);
 207 }
 208
 209 static void
 210 mlx5e_media_status(struct ifnet *dev, struct ifmediareq *ifmr)
 211 {
 212         struct mlx5e_priv *priv = dev->if_softc;
 213
 214         ifmr->ifm_status = priv->media_status_last;
 215         ifmr->ifm_active = priv->media_active_last |
 216             (priv->params.rx_pauseframe_control ? IFM_ETH_RXPAUSE : 0) |
 217             (priv->params.tx_pauseframe_control ? IFM_ETH_TXPAUSE : 0);
 218
 219 }
 220
 221 static u32
 222 mlx5e_find_link_mode(u32 subtype)
 223 {
 224         u32 i;
 225         u32 link_mode = 0;
 226
 227         for (i = 0; i < MLX5E_LINK_MODES_NUMBER; ++i) {
 228                 if (mlx5e_mode_table[i].baudrate == 0)
 229                         continue;
 230                 if (mlx5e_mode_table[i].subtype == subtype)
 231                         link_mode |= MLX5E_PROT_MASK(i);
 232         }
 233
 234         return (link_mode);
 235 }
 236
 237 static int
 238 mlx5e_set_port_pause_and_pfc(struct mlx5e_priv *priv)
 239 {
 240         return (mlx5_set_port_pause_and_pfc(priv->mdev, 1,
 241             priv->params.rx_pauseframe_control,
 242             priv->params.tx_pauseframe_control,
 243             priv->params.rx_priority_flow_control,
 244             priv->params.tx_priority_flow_control));
 245 }
 246
 247 static int
 248 mlx5e_set_port_pfc(struct mlx5e_priv *priv)
 249 {
 250         int error;
 251
 252         if (priv->params.rx_pauseframe_control ||
 253             priv->params.tx_pauseframe_control) {
 254                 if_printf(priv->ifp,
 255                     "Global pauseframes must be disabled before enabling PFC.\n");
 256                 error = -EINVAL;
 257         } else {
 258                 error = mlx5e_set_port_pause_and_pfc(priv);
 259         }
 260         return (error);
 261 }
 262
 263 static int
 264 mlx5e_media_change(struct ifnet *dev)
 265 {
 266         struct mlx5e_priv *priv = dev->if_softc;
 267         struct mlx5_core_dev *mdev = priv->mdev;
 268         u32 eth_proto_cap;
 269         u32 link_mode;
 270         int was_opened;
 271         int locked;
 272         int error;
 273
 274         locked = PRIV_LOCKED(priv);
 275         if (!locked)
 276                 PRIV_LOCK(priv);
 277
 278         if (IFM_TYPE(priv->media.ifm_media) != IFM_ETHER) {
 279                 error = EINVAL;
 280                 goto done;
 281         }
 282         link_mode = mlx5e_find_link_mode(IFM_SUBTYPE(priv->media.ifm_media));
 283
 284         /* query supported capabilities */
 285         error = mlx5_query_port_proto_cap(mdev, &eth_proto_cap, MLX5_PTYS_EN);
 286         if (error != 0) {
 287                 if_printf(dev, "Query port media capability failed\n");
 288                 goto done;
 289         }
 290         /* check for autoselect */
 291         if (IFM_SUBTYPE(priv->media.ifm_media) == IFM_AUTO) {
 292                 link_mode = eth_proto_cap;
 293                 if (link_mode == 0) {
 294                         if_printf(dev, "Port media capability is zero\n");
 295                         error = EINVAL;
 296                         goto done;
 297                 }
 298         } else {
 299                 link_mode = link_mode & eth_proto_cap;
 300                 if (link_mode == 0) {
 301                         if_printf(dev, "Not supported link mode requested\n");
 302                         error = EINVAL;
 303                         goto done;
 304                 }
 305         }
 306         if (priv->media.ifm_media & (IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE)) {
 307                 /* check if PFC is enabled */
 308                 if (priv->params.rx_priority_flow_control ||
 309                     priv->params.tx_priority_flow_control) {
 310                         if_printf(dev, "PFC must be disabled before enabling global pauseframes.\n");
 311                         error = EINVAL;
 312                         goto done;
 313                 }
 314         }
 315         /* update pauseframe control bits */
 316         priv->params.rx_pauseframe_control =
 317             (priv->media.ifm_media & IFM_ETH_RXPAUSE) ? 1 : 0;
 318         priv->params.tx_pauseframe_control =
 319             (priv->media.ifm_media & IFM_ETH_TXPAUSE) ? 1 : 0;
 320
 321         /* check if device is opened */
 322         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
 323
 324         /* reconfigure the hardware */
 325         mlx5_set_port_status(mdev, MLX5_PORT_DOWN);
 326         mlx5_set_port_proto(mdev, link_mode, MLX5_PTYS_EN);
 327         error = -mlx5e_set_port_pause_and_pfc(priv);
 328         if (was_opened)
 329                 mlx5_set_port_status(mdev, MLX5_PORT_UP);
 330
 331 done:
 332         if (!locked)
 333                 PRIV_UNLOCK(priv);
 334         return (error);
 335 }
 336
 337 static void
 338 mlx5e_update_carrier_work(struct work_struct *work)
 339 {
 340         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 341             update_carrier_work);
 342
 343         PRIV_LOCK(priv);
 344         if (test_bit(MLX5E_STATE_OPENED, &priv->state))
 345                 mlx5e_update_carrier(priv);
 346         PRIV_UNLOCK(priv);
 347 }
 348
 349 /*
 350  * This function reads the physical port counters from the firmware
 351  * using a pre-defined layout defined by various MLX5E_PPORT_XXX()
 352  * macros. The output is converted from big-endian 64-bit values into
 353  * host endian ones and stored in the "priv->stats.pport" structure.
 354  */
 355 static void
 356 mlx5e_update_pport_counters(struct mlx5e_priv *priv)
 357 {
 358         struct mlx5_core_dev *mdev = priv->mdev;
 359         struct mlx5e_pport_stats *s = &priv->stats.pport;
 360         struct mlx5e_port_stats_debug *s_debug = &priv->stats.port_stats_debug;
 361         u32 *in;
 362         u32 *out;
 363         const u64 *ptr;
 364         unsigned sz = MLX5_ST_SZ_BYTES(ppcnt_reg);
 365         unsigned x;
 366         unsigned y;
 367         unsigned z;
 368
 369         /* allocate firmware request structures */
 370         in = mlx5_vzalloc(sz);
 371         out = mlx5_vzalloc(sz);
 372         if (in == NULL || out == NULL)
 373                 goto free_out;
 374
 375         /*
 376          * Get pointer to the 64-bit counter set which is located at a
 377          * fixed offset in the output firmware request structure:
 378          */
 379         ptr = (const uint64_t *)MLX5_ADDR_OF(ppcnt_reg, out, counter_set);
 380
 381         MLX5_SET(ppcnt_reg, in, local_port, 1);
 382
 383         /* read IEEE802_3 counter group using predefined counter layout */
 384         MLX5_SET(ppcnt_reg, in, grp, MLX5_IEEE_802_3_COUNTERS_GROUP);
 385         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 386         for (x = 0, y = MLX5E_PPORT_PER_PRIO_STATS_NUM;
 387              x != MLX5E_PPORT_IEEE802_3_STATS_NUM; x++, y++)
 388                 s->arg[y] = be64toh(ptr[x]);
 389
 390         /* read RFC2819 counter group using predefined counter layout */
 391         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2819_COUNTERS_GROUP);
 392         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 393         for (x = 0; x != MLX5E_PPORT_RFC2819_STATS_NUM; x++, y++)
 394                 s->arg[y] = be64toh(ptr[x]);
 395         for (y = 0; x != MLX5E_PPORT_RFC2819_STATS_NUM +
 396             MLX5E_PPORT_RFC2819_STATS_DEBUG_NUM; x++, y++)
 397                 s_debug->arg[y] = be64toh(ptr[x]);
 398
 399         /* read RFC2863 counter group using predefined counter layout */
 400         MLX5_SET(ppcnt_reg, in, grp, MLX5_RFC_2863_COUNTERS_GROUP);
 401         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 402         for (x = 0; x != MLX5E_PPORT_RFC2863_STATS_DEBUG_NUM; x++, y++)
 403                 s_debug->arg[y] = be64toh(ptr[x]);
 404
 405         /* read physical layer stats counter group using predefined counter layout */
 406         MLX5_SET(ppcnt_reg, in, grp, MLX5_PHYSICAL_LAYER_COUNTERS_GROUP);
 407         mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 408         for (x = 0; x != MLX5E_PPORT_PHYSICAL_LAYER_STATS_DEBUG_NUM; x++, y++)
 409                 s_debug->arg[y] = be64toh(ptr[x]);
 410
 411         /* read per-priority counters */
 412         MLX5_SET(ppcnt_reg, in, grp, MLX5_PER_PRIORITY_COUNTERS_GROUP);
 413
 414         /* iterate all the priorities */
 415         for (y = z = 0; z != MLX5E_PPORT_PER_PRIO_STATS_NUM_PRIO; z++) {
 416                 MLX5_SET(ppcnt_reg, in, prio_tc, z);
 417                 mlx5_core_access_reg(mdev, in, sz, out, sz, MLX5_REG_PPCNT, 0, 0);
 418
 419                 /* read per priority stats counter group using predefined counter layout */
 420                 for (x = 0; x != (MLX5E_PPORT_PER_PRIO_STATS_NUM /
 421                     MLX5E_PPORT_PER_PRIO_STATS_NUM_PRIO); x++, y++)
 422                         s->arg[y] = be64toh(ptr[x]);
 423         }
 424 free_out:
 425         /* free firmware request structures */
 426         kvfree(in);
 427         kvfree(out);
 428 }
 429
 430 /*
 431  * This function is called regularly to collect all statistics
 432  * counters from the firmware. The values can be viewed through the
 433  * sysctl interface. Execution is serialized using the priv's global
 434  * configuration lock.
 435  */
 436 static void
 437 mlx5e_update_stats_work(struct work_struct *work)
 438 {
 439         struct mlx5e_priv *priv = container_of(work, struct mlx5e_priv,
 440             update_stats_work);
 441         struct mlx5_core_dev *mdev = priv->mdev;
 442         struct mlx5e_vport_stats *s = &priv->stats.vport;
 443         struct mlx5e_rq_stats *rq_stats;
 444         struct mlx5e_sq_stats *sq_stats;
 445         struct buf_ring *sq_br;
 446 #if (__FreeBSD_version < 1100000)
 447         struct ifnet *ifp = priv->ifp;
 448 #endif
 449
 450         u32 in[MLX5_ST_SZ_DW(query_vport_counter_in)];
 451         u32 *out;
 452         int outlen = MLX5_ST_SZ_BYTES(query_vport_counter_out);
 453         u64 tso_packets = 0;
 454         u64 tso_bytes = 0;
 455         u64 tx_queue_dropped = 0;
 456         u64 tx_defragged = 0;
 457         u64 tx_offload_none = 0;
 458         u64 lro_packets = 0;
 459         u64 lro_bytes = 0;
 460         u64 sw_lro_queued = 0;
 461         u64 sw_lro_flushed = 0;
 462         u64 rx_csum_none = 0;
 463         u64 rx_wqe_err = 0;
 464         u32 rx_out_of_buffer = 0;
 465         int i;
 466         int j;
 467
 468         PRIV_LOCK(priv);
 469         out = mlx5_vzalloc(outlen);
 470         if (out == NULL)
 471                 goto free_out;
 472         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
 473                 goto free_out;
 474
 475         /* Collect firts the SW counters and then HW for consistency */
 476         for (i = 0; i < priv->params.num_channels; i++) {
 477                 struct mlx5e_rq *rq = &priv->channel[i]->rq;
 478
 479                 rq_stats = &priv->channel[i]->rq.stats;
 480
 481                 /* collect stats from LRO */
 482                 rq_stats->sw_lro_queued = rq->lro.lro_queued;
 483                 rq_stats->sw_lro_flushed = rq->lro.lro_flushed;
 484                 sw_lro_queued += rq_stats->sw_lro_queued;
 485                 sw_lro_flushed += rq_stats->sw_lro_flushed;
 486                 lro_packets += rq_stats->lro_packets;
 487                 lro_bytes += rq_stats->lro_bytes;
 488                 rx_csum_none += rq_stats->csum_none;
 489                 rx_wqe_err += rq_stats->wqe_err;
 490
 491                 for (j = 0; j < priv->num_tc; j++) {
 492                         sq_stats = &priv->channel[i]->sq[j].stats;
 493                         sq_br = priv->channel[i]->sq[j].br;
 494
 495                         tso_packets += sq_stats->tso_packets;
 496                         tso_bytes += sq_stats->tso_bytes;
 497                         tx_queue_dropped += sq_stats->dropped;
 498                         if (sq_br != NULL)
 499                                 tx_queue_dropped += sq_br->br_drops;
 500                         tx_defragged += sq_stats->defragged;
 501                         tx_offload_none += sq_stats->csum_offload_none;
 502                 }
 503         }
 504
 505         /* update counters */
 506         s->tso_packets = tso_packets;
 507         s->tso_bytes = tso_bytes;
 508         s->tx_queue_dropped = tx_queue_dropped;
 509         s->tx_defragged = tx_defragged;
 510         s->lro_packets = lro_packets;
 511         s->lro_bytes = lro_bytes;
 512         s->sw_lro_queued = sw_lro_queued;
 513         s->sw_lro_flushed = sw_lro_flushed;
 514         s->rx_csum_none = rx_csum_none;
 515         s->rx_wqe_err = rx_wqe_err;
 516
 517         /* HW counters */
 518         memset(in, 0, sizeof(in));
 519
 520         MLX5_SET(query_vport_counter_in, in, opcode,
 521             MLX5_CMD_OP_QUERY_VPORT_COUNTER);
 522         MLX5_SET(query_vport_counter_in, in, op_mod, 0);
 523         MLX5_SET(query_vport_counter_in, in, other_vport, 0);
 524
 525         memset(out, 0, outlen);
 526
 527         /* get number of out-of-buffer drops first */
 528         if (mlx5_vport_query_out_of_rx_buffer(mdev, priv->counter_set_id,
 529             &rx_out_of_buffer))
 530                 goto free_out;
 531
 532         /* accumulate difference into a 64-bit counter */
 533         s->rx_out_of_buffer += (u64)(u32)(rx_out_of_buffer - s->rx_out_of_buffer_prev);
 534         s->rx_out_of_buffer_prev = rx_out_of_buffer;
 535
 536         /* get port statistics */
 537         if (mlx5_cmd_exec(mdev, in, sizeof(in), out, outlen))
 538                 goto free_out;
 539
 540 #define MLX5_GET_CTR(out, x) \
 541         MLX5_GET64(query_vport_counter_out, out, x)
 542
 543         s->rx_error_packets =
 544             MLX5_GET_CTR(out, received_errors.packets);
 545         s->rx_error_bytes =
 546             MLX5_GET_CTR(out, received_errors.octets);
 547         s->tx_error_packets =
 548             MLX5_GET_CTR(out, transmit_errors.packets);
 549         s->tx_error_bytes =
 550             MLX5_GET_CTR(out, transmit_errors.octets);
 551
 552         s->rx_unicast_packets =
 553             MLX5_GET_CTR(out, received_eth_unicast.packets);
 554         s->rx_unicast_bytes =
 555             MLX5_GET_CTR(out, received_eth_unicast.octets);
 556         s->tx_unicast_packets =
 557             MLX5_GET_CTR(out, transmitted_eth_unicast.packets);
 558         s->tx_unicast_bytes =
 559             MLX5_GET_CTR(out, transmitted_eth_unicast.octets);
 560
 561         s->rx_multicast_packets =
 562             MLX5_GET_CTR(out, received_eth_multicast.packets);
 563         s->rx_multicast_bytes =
 564             MLX5_GET_CTR(out, received_eth_multicast.octets);
 565         s->tx_multicast_packets =
 566             MLX5_GET_CTR(out, transmitted_eth_multicast.packets);
 567         s->tx_multicast_bytes =
 568             MLX5_GET_CTR(out, transmitted_eth_multicast.octets);
 569
 570         s->rx_broadcast_packets =
 571             MLX5_GET_CTR(out, received_eth_broadcast.packets);
 572         s->rx_broadcast_bytes =
 573             MLX5_GET_CTR(out, received_eth_broadcast.octets);
 574         s->tx_broadcast_packets =
 575             MLX5_GET_CTR(out, transmitted_eth_broadcast.packets);
 576         s->tx_broadcast_bytes =
 577             MLX5_GET_CTR(out, transmitted_eth_broadcast.octets);
 578
 579         s->rx_packets =
 580             s->rx_unicast_packets +
 581             s->rx_multicast_packets +
 582             s->rx_broadcast_packets -
 583             s->rx_out_of_buffer;
 584         s->rx_bytes =
 585             s->rx_unicast_bytes +
 586             s->rx_multicast_bytes +
 587             s->rx_broadcast_bytes;
 588         s->tx_packets =
 589             s->tx_unicast_packets +
 590             s->tx_multicast_packets +
 591             s->tx_broadcast_packets;
 592         s->tx_bytes =
 593             s->tx_unicast_bytes +
 594             s->tx_multicast_bytes +
 595             s->tx_broadcast_bytes;
 596
 597         /* Update calculated offload counters */
 598         s->tx_csum_offload = s->tx_packets - tx_offload_none;
 599         s->rx_csum_good = s->rx_packets - s->rx_csum_none;
 600
 601         /* Get physical port counters */
 602         mlx5e_update_pport_counters(priv);
 603
 604 #if (__FreeBSD_version < 1100000)
 605         /* no get_counters interface in fbsd 10 */
 606         ifp->if_ipackets = s->rx_packets;
 607         ifp->if_ierrors = s->rx_error_packets +
 608             priv->stats.pport.alignment_err +
 609             priv->stats.pport.check_seq_err +
 610             priv->stats.pport.crc_align_errors +
 611             priv->stats.pport.in_range_len_errors +
 612             priv->stats.pport.jabbers +
 613             priv->stats.pport.out_of_range_len +
 614             priv->stats.pport.oversize_pkts +
 615             priv->stats.pport.symbol_err +
 616             priv->stats.pport.too_long_errors +
 617             priv->stats.pport.undersize_pkts +
 618             priv->stats.pport.unsupported_op_rx;
 619         ifp->if_iqdrops = s->rx_out_of_buffer +
 620             priv->stats.pport.drop_events;
 621         ifp->if_opackets = s->tx_packets;
 622         ifp->if_oerrors = s->tx_error_packets;
 623         ifp->if_snd.ifq_drops = s->tx_queue_dropped;
 624         ifp->if_ibytes = s->rx_bytes;
 625         ifp->if_obytes = s->tx_bytes;
 626         ifp->if_collisions =
 627             priv->stats.pport.collisions;
 628 #endif
 629
 630 free_out:
 631         kvfree(out);
 632
 633         /* Update diagnostics, if any */
 634         if (priv->params_ethtool.diag_pci_enable ||
 635             priv->params_ethtool.diag_general_enable) {
 636                 int error = mlx5_core_get_diagnostics_full(mdev,
 637                     priv->params_ethtool.diag_pci_enable ? &priv->params_pci : NULL,
 638                     priv->params_ethtool.diag_general_enable ? &priv->params_general : NULL);
 639                 if (error != 0)
 640                         if_printf(priv->ifp, "Failed reading diagnostics: %d\n", error);
 641         }
 642         PRIV_UNLOCK(priv);
 643 }
 644
 645 static void
 646 mlx5e_update_stats(void *arg)
 647 {
 648         struct mlx5e_priv *priv = arg;
 649
 650         queue_work(priv->wq, &priv->update_stats_work);
 651
 652         callout_reset(&priv->watchdog, hz, &mlx5e_update_stats, priv);
 653 }
 654
 655 static void
 656 mlx5e_async_event_sub(struct mlx5e_priv *priv,
 657     enum mlx5_dev_event event)
 658 {
 659         switch (event) {
 660         case MLX5_DEV_EVENT_PORT_UP:
 661         case MLX5_DEV_EVENT_PORT_DOWN:
 662                 queue_work(priv->wq, &priv->update_carrier_work);
 663                 break;
 664
 665         default:
 666                 break;
 667         }
 668 }
 669
 670 static void
 671 mlx5e_async_event(struct mlx5_core_dev *mdev, void *vpriv,
 672     enum mlx5_dev_event event, unsigned long param)
 673 {
 674         struct mlx5e_priv *priv = vpriv;
 675
 676         mtx_lock(&priv->async_events_mtx);
 677         if (test_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state))
 678                 mlx5e_async_event_sub(priv, event);
 679         mtx_unlock(&priv->async_events_mtx);
 680 }
 681
 682 static void
 683 mlx5e_enable_async_events(struct mlx5e_priv *priv)
 684 {
 685         set_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
 686 }
 687
 688 static void
 689 mlx5e_disable_async_events(struct mlx5e_priv *priv)
 690 {
 691         mtx_lock(&priv->async_events_mtx);
 692         clear_bit(MLX5E_STATE_ASYNC_EVENTS_ENABLE, &priv->state);
 693         mtx_unlock(&priv->async_events_mtx);
 694 }
 695
 696 static void mlx5e_calibration_callout(void *arg);
 697 static int mlx5e_calibration_duration = 20;
 698 static int mlx5e_fast_calibration = 1;
 699 static int mlx5e_normal_calibration = 30;
 700
 701 static SYSCTL_NODE(_hw_mlx5, OID_AUTO, calibr, CTLFLAG_RW, 0,
 702     "MLX5 timestamp calibration parameteres");
 703
 704 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, duration, CTLFLAG_RWTUN,
 705     &mlx5e_calibration_duration, 0,
 706     "Duration of initial calibration");
 707 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, fast, CTLFLAG_RWTUN,
 708     &mlx5e_fast_calibration, 0,
 709     "Recalibration interval during initial calibration");
 710 SYSCTL_INT(_hw_mlx5_calibr, OID_AUTO, normal, CTLFLAG_RWTUN,
 711     &mlx5e_normal_calibration, 0,
 712     "Recalibration interval during normal operations");
 713
 714 /*
 715  * Ignites the calibration process.
 716  */
 717 static void
 718 mlx5e_reset_calibration_callout(struct mlx5e_priv *priv)
 719 {
 720
 721         if (priv->clbr_done == 0)
 722                 mlx5e_calibration_callout(priv);
 723         else
 724                 callout_reset_curcpu(&priv->tstmp_clbr, (priv->clbr_done <
 725                     mlx5e_calibration_duration ? mlx5e_fast_calibration :
 726                     mlx5e_normal_calibration) * hz, mlx5e_calibration_callout,
 727                     priv);
 728 }
 729
 730 static uint64_t
 731 mlx5e_timespec2usec(const struct timespec *ts)
 732 {
 733
 734         return ((uint64_t)ts->tv_sec * 1000000000 + ts->tv_nsec);
 735 }
 736
 737 static uint64_t
 738 mlx5e_hw_clock(struct mlx5e_priv *priv)
 739 {
 740         struct mlx5_init_seg *iseg;
 741         uint32_t hw_h, hw_h1, hw_l;
 742
 743         iseg = priv->mdev->iseg;
 744         do {
 745                 hw_h = ioread32be(&iseg->internal_timer_h);
 746                 hw_l = ioread32be(&iseg->internal_timer_l);
 747                 hw_h1 = ioread32be(&iseg->internal_timer_h);
 748         } while (hw_h1 != hw_h);
 749         return (((uint64_t)hw_h << 32) | hw_l);
 750 }
 751
 752 /*
 753  * The calibration callout, it runs either in the context of the
 754  * thread which enables calibration, or in callout.  It takes the
 755  * snapshot of system and adapter clocks, then advances the pointers to
 756  * the calibration point to allow rx path to read the consistent data
 757  * lockless.
 758  */
 759 static void
 760 mlx5e_calibration_callout(void *arg)
 761 {
 762         struct mlx5e_priv *priv;
 763         struct mlx5e_clbr_point *next, *curr;
 764         struct timespec ts;
 765         int clbr_curr_next;
 766
 767         priv = arg;
 768         curr = &priv->clbr_points[priv->clbr_curr];
 769         clbr_curr_next = priv->clbr_curr + 1;
 770         if (clbr_curr_next >= nitems(priv->clbr_points))
 771                 clbr_curr_next = 0;
 772         next = &priv->clbr_points[clbr_curr_next];
 773
 774         next->base_prev = curr->base_curr;
 775         next->clbr_hw_prev = curr->clbr_hw_curr;
 776
 777         next->clbr_hw_curr = mlx5e_hw_clock(priv);
 778         if (((next->clbr_hw_curr - curr->clbr_hw_prev) >> MLX5E_TSTMP_PREC) ==
 779             0) {
 780                 if_printf(priv->ifp, "HW failed tstmp frozen %#jx %#jx,"
 781                     "disabling\n", next->clbr_hw_curr, curr->clbr_hw_prev);
 782                 priv->clbr_done = 0;
 783                 return;
 784         }
 785
 786         nanouptime(&ts);
 787         next->base_curr = mlx5e_timespec2usec(&ts);
 788
 789         curr->clbr_gen = 0;
 790         atomic_thread_fence_rel();
 791         priv->clbr_curr = clbr_curr_next;
 792         atomic_store_rel_int(&next->clbr_gen, ++(priv->clbr_gen));
 793
 794         if (priv->clbr_done < mlx5e_calibration_duration)
 795                 priv->clbr_done++;
 796         mlx5e_reset_calibration_callout(priv);
 797 }
 798
 799 static const char *mlx5e_rq_stats_desc[] = {
 800         MLX5E_RQ_STATS(MLX5E_STATS_DESC)
 801 };
 802
 803 static int
 804 mlx5e_create_rq(struct mlx5e_channel *c,
 805     struct mlx5e_rq_param *param,
 806     struct mlx5e_rq *rq)
 807 {
 808         struct mlx5e_priv *priv = c->priv;
 809         struct mlx5_core_dev *mdev = priv->mdev;
 810         char buffer[16];
 811         void *rqc = param->rqc;
 812         void *rqc_wq = MLX5_ADDR_OF(rqc, rqc, wq);
 813         int wq_sz;
 814         int err;
 815         int i;
 816         u32 nsegs, wqe_sz;
 817
 818         err = mlx5e_get_wqe_sz(priv, &wqe_sz, &nsegs);
 819         if (err != 0)
 820                 goto done;
 821
 822         /* Create DMA descriptor TAG */
 823         if ((err = -bus_dma_tag_create(
 824             bus_get_dma_tag(mdev->pdev->dev.bsddev),
 825             1,                          /* any alignment */
 826             0,                          /* no boundary */
 827             BUS_SPACE_MAXADDR,          /* lowaddr */
 828             BUS_SPACE_MAXADDR,          /* highaddr */
 829             NULL, NULL,                 /* filter, filterarg */
 830             nsegs * MLX5E_MAX_RX_BYTES, /* maxsize */
 831             nsegs,                      /* nsegments */
 832             nsegs * MLX5E_MAX_RX_BYTES, /* maxsegsize */
 833             0,                          /* flags */
 834             NULL, NULL,                 /* lockfunc, lockfuncarg */
 835             &rq->dma_tag)))
 836                 goto done;
 837
 838         err = mlx5_wq_ll_create(mdev, &param->wq, rqc_wq, &rq->wq,
 839             &rq->wq_ctrl);
 840         if (err)
 841                 goto err_free_dma_tag;
 842
 843         rq->wq.db = &rq->wq.db[MLX5_RCV_DBR];
 844
 845         err = mlx5e_get_wqe_sz(priv, &rq->wqe_sz, &rq->nsegs);
 846         if (err != 0)
 847                 goto err_rq_wq_destroy;
 848
 849         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 850
 851         err = -tcp_lro_init_args(&rq->lro, c->ifp, TCP_LRO_ENTRIES, wq_sz);
 852         if (err)
 853                 goto err_rq_wq_destroy;
 854
 855         rq->mbuf = malloc(wq_sz * sizeof(rq->mbuf[0]), M_MLX5EN, M_WAITOK | M_ZERO);
 856         for (i = 0; i != wq_sz; i++) {
 857                 struct mlx5e_rx_wqe *wqe = mlx5_wq_ll_get_wqe(&rq->wq, i);
 858 #if (MLX5E_MAX_RX_SEGS == 1)
 859                 uint32_t byte_count = rq->wqe_sz - MLX5E_NET_IP_ALIGN;
 860 #else
 861                 int j;
 862 #endif
 863
 864                 err = -bus_dmamap_create(rq->dma_tag, 0, &rq->mbuf[i].dma_map);
 865                 if (err != 0) {
 866                         while (i--)
 867                                 bus_dmamap_destroy(rq->dma_tag, rq->mbuf[i].dma_map);
 868                         goto err_rq_mbuf_free;
 869                 }
 870
 871                 /* set value for constant fields */
 872 #if (MLX5E_MAX_RX_SEGS == 1)
 873                 wqe->data[0].lkey = c->mkey_be;
 874                 wqe->data[0].byte_count = cpu_to_be32(byte_count | MLX5_HW_START_PADDING);
 875 #else
 876                 for (j = 0; j < rq->nsegs; j++)
 877                         wqe->data[j].lkey = c->mkey_be;
 878 #endif
 879         }
 880
 881         rq->ifp = c->ifp;
 882         rq->channel = c;
 883         rq->ix = c->ix;
 884
 885         snprintf(buffer, sizeof(buffer), "rxstat%d", c->ix);
 886         mlx5e_create_stats(&rq->stats.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
 887             buffer, mlx5e_rq_stats_desc, MLX5E_RQ_STATS_NUM,
 888             rq->stats.arg);
 889         return (0);
 890
 891 err_rq_mbuf_free:
 892         free(rq->mbuf, M_MLX5EN);
 893         tcp_lro_free(&rq->lro);
 894 err_rq_wq_destroy:
 895         mlx5_wq_destroy(&rq->wq_ctrl);
 896 err_free_dma_tag:
 897         bus_dma_tag_destroy(rq->dma_tag);
 898 done:
 899         return (err);
 900 }
 901
 902 static void
 903 mlx5e_destroy_rq(struct mlx5e_rq *rq)
 904 {
 905         int wq_sz;
 906         int i;
 907
 908         /* destroy all sysctl nodes */
 909         sysctl_ctx_free(&rq->stats.ctx);
 910
 911         /* free leftover LRO packets, if any */
 912         tcp_lro_free(&rq->lro);
 913
 914         wq_sz = mlx5_wq_ll_get_size(&rq->wq);
 915         for (i = 0; i != wq_sz; i++) {
 916                 if (rq->mbuf[i].mbuf != NULL) {
 917                         bus_dmamap_unload(rq->dma_tag, rq->mbuf[i].dma_map);
 918                         m_freem(rq->mbuf[i].mbuf);
 919                 }
 920                 bus_dmamap_destroy(rq->dma_tag, rq->mbuf[i].dma_map);
 921         }
 922         free(rq->mbuf, M_MLX5EN);
 923         mlx5_wq_destroy(&rq->wq_ctrl);
 924 }
 925
 926 static int
 927 mlx5e_enable_rq(struct mlx5e_rq *rq, struct mlx5e_rq_param *param)
 928 {
 929         struct mlx5e_channel *c = rq->channel;
 930         struct mlx5e_priv *priv = c->priv;
 931         struct mlx5_core_dev *mdev = priv->mdev;
 932
 933         void *in;
 934         void *rqc;
 935         void *wq;
 936         int inlen;
 937         int err;
 938
 939         inlen = MLX5_ST_SZ_BYTES(create_rq_in) +
 940             sizeof(u64) * rq->wq_ctrl.buf.npages;
 941         in = mlx5_vzalloc(inlen);
 942         if (in == NULL)
 943                 return (-ENOMEM);
 944
 945         rqc = MLX5_ADDR_OF(create_rq_in, in, ctx);
 946         wq = MLX5_ADDR_OF(rqc, rqc, wq);
 947
 948         memcpy(rqc, param->rqc, sizeof(param->rqc));
 949
 950         MLX5_SET(rqc, rqc, cqn, c->rq.cq.mcq.cqn);
 951         MLX5_SET(rqc, rqc, state, MLX5_RQC_STATE_RST);
 952         MLX5_SET(rqc, rqc, flush_in_error_en, 1);
 953         if (priv->counter_set_id >= 0)
 954                 MLX5_SET(rqc, rqc, counter_set_id, priv->counter_set_id);
 955         MLX5_SET(wq, wq, log_wq_pg_sz, rq->wq_ctrl.buf.page_shift -
 956             PAGE_SHIFT);
 957         MLX5_SET64(wq, wq, dbr_addr, rq->wq_ctrl.db.dma);
 958
 959         mlx5_fill_page_array(&rq->wq_ctrl.buf,
 960             (__be64 *) MLX5_ADDR_OF(wq, wq, pas));
 961
 962         err = mlx5_core_create_rq(mdev, in, inlen, &rq->rqn);
 963
 964         kvfree(in);
 965
 966         return (err);
 967 }
 968
 969 static int
 970 mlx5e_modify_rq(struct mlx5e_rq *rq, int curr_state, int next_state)
 971 {
 972         struct mlx5e_channel *c = rq->channel;
 973         struct mlx5e_priv *priv = c->priv;
 974         struct mlx5_core_dev *mdev = priv->mdev;
 975
 976         void *in;
 977         void *rqc;
 978         int inlen;
 979         int err;
 980
 981         inlen = MLX5_ST_SZ_BYTES(modify_rq_in);
 982         in = mlx5_vzalloc(inlen);
 983         if (in == NULL)
 984                 return (-ENOMEM);
 985
 986         rqc = MLX5_ADDR_OF(modify_rq_in, in, ctx);
 987
 988         MLX5_SET(modify_rq_in, in, rqn, rq->rqn);
 989         MLX5_SET(modify_rq_in, in, rq_state, curr_state);
 990         MLX5_SET(rqc, rqc, state, next_state);
 991
 992         err = mlx5_core_modify_rq(mdev, in, inlen);
 993
 994         kvfree(in);
 995
 996         return (err);
 997 }
 998
 999 static void
1000 mlx5e_disable_rq(struct mlx5e_rq *rq)
1001 {
1002         struct mlx5e_channel *c = rq->channel;
1003         struct mlx5e_priv *priv = c->priv;
1004         struct mlx5_core_dev *mdev = priv->mdev;
1005
1006         mlx5_core_destroy_rq(mdev, rq->rqn);
1007 }
1008
1009 static int
1010 mlx5e_wait_for_min_rx_wqes(struct mlx5e_rq *rq)
1011 {
1012         struct mlx5e_channel *c = rq->channel;
1013         struct mlx5e_priv *priv = c->priv;
1014         struct mlx5_wq_ll *wq = &rq->wq;
1015         int i;
1016
1017         for (i = 0; i < 1000; i++) {
1018                 if (wq->cur_sz >= priv->params.min_rx_wqes)
1019                         return (0);
1020
1021                 msleep(4);
1022         }
1023         return (-ETIMEDOUT);
1024 }
1025
1026 static int
1027 mlx5e_open_rq(struct mlx5e_channel *c,
1028     struct mlx5e_rq_param *param,
1029     struct mlx5e_rq *rq)
1030 {
1031         int err;
1032
1033         err = mlx5e_create_rq(c, param, rq);
1034         if (err)
1035                 return (err);
1036
1037         err = mlx5e_enable_rq(rq, param);
1038         if (err)
1039                 goto err_destroy_rq;
1040
1041         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
1042         if (err)
1043                 goto err_disable_rq;
1044
1045         c->rq.enabled = 1;
1046
1047         return (0);
1048
1049 err_disable_rq:
1050         mlx5e_disable_rq(rq);
1051 err_destroy_rq:
1052         mlx5e_destroy_rq(rq);
1053
1054         return (err);
1055 }
1056
1057 static void
1058 mlx5e_close_rq(struct mlx5e_rq *rq)
1059 {
1060         mtx_lock(&rq->mtx);
1061         rq->enabled = 0;
1062         callout_stop(&rq->watchdog);
1063         mtx_unlock(&rq->mtx);
1064
1065         callout_drain(&rq->watchdog);
1066
1067         mlx5e_modify_rq(rq, MLX5_RQC_STATE_RDY, MLX5_RQC_STATE_ERR);
1068 }
1069
1070 static void
1071 mlx5e_close_rq_wait(struct mlx5e_rq *rq)
1072 {
1073         struct mlx5_core_dev *mdev = rq->channel->priv->mdev;
1074
1075         /* wait till RQ is empty */
1076         while (!mlx5_wq_ll_is_empty(&rq->wq) &&
1077                (mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR)) {
1078                 msleep(4);
1079                 rq->cq.mcq.comp(&rq->cq.mcq);
1080         }
1081
1082         mlx5e_disable_rq(rq);
1083         mlx5e_destroy_rq(rq);
1084 }
1085
1086 void
1087 mlx5e_free_sq_db(struct mlx5e_sq *sq)
1088 {
1089         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1090         int x;
1091
1092         for (x = 0; x != wq_sz; x++)
1093                 bus_dmamap_destroy(sq->dma_tag, sq->mbuf[x].dma_map);
1094         free(sq->mbuf, M_MLX5EN);
1095 }
1096
1097 int
1098 mlx5e_alloc_sq_db(struct mlx5e_sq *sq)
1099 {
1100         int wq_sz = mlx5_wq_cyc_get_size(&sq->wq);
1101         int err;
1102         int x;
1103
1104         sq->mbuf = malloc(wq_sz * sizeof(sq->mbuf[0]), M_MLX5EN, M_WAITOK | M_ZERO);
1105
1106         /* Create DMA descriptor MAPs */
1107         for (x = 0; x != wq_sz; x++) {
1108                 err = -bus_dmamap_create(sq->dma_tag, 0, &sq->mbuf[x].dma_map);
1109                 if (err != 0) {
1110                         while (x--)
1111                                 bus_dmamap_destroy(sq->dma_tag, sq->mbuf[x].dma_map);
1112                         free(sq->mbuf, M_MLX5EN);
1113                         return (err);
1114                 }
1115         }
1116         return (0);
1117 }
1118
1119 static const char *mlx5e_sq_stats_desc[] = {
1120         MLX5E_SQ_STATS(MLX5E_STATS_DESC)
1121 };
1122
1123 static int
1124 mlx5e_create_sq(struct mlx5e_channel *c,
1125     int tc,
1126     struct mlx5e_sq_param *param,
1127     struct mlx5e_sq *sq)
1128 {
1129         struct mlx5e_priv *priv = c->priv;
1130         struct mlx5_core_dev *mdev = priv->mdev;
1131         char buffer[16];
1132
1133         void *sqc = param->sqc;
1134         void *sqc_wq = MLX5_ADDR_OF(sqc, sqc, wq);
1135 #ifdef RSS
1136         cpuset_t cpu_mask;
1137         int cpu_id;
1138 #endif
1139         int err;
1140
1141         /* Create DMA descriptor TAG */
1142         if ((err = -bus_dma_tag_create(
1143             bus_get_dma_tag(mdev->pdev->dev.bsddev),
1144             1,                          /* any alignment */
1145             0,                          /* no boundary */
1146             BUS_SPACE_MAXADDR,          /* lowaddr */
1147             BUS_SPACE_MAXADDR,          /* highaddr */
1148             NULL, NULL,                 /* filter, filterarg */
1149             MLX5E_MAX_TX_PAYLOAD_SIZE,  /* maxsize */
1150             MLX5E_MAX_TX_MBUF_FRAGS,    /* nsegments */
1151             MLX5E_MAX_TX_MBUF_SIZE,     /* maxsegsize */
1152             0,                          /* flags */
1153             NULL, NULL,                 /* lockfunc, lockfuncarg */
1154             &sq->dma_tag)))
1155                 goto done;
1156
1157         err = mlx5_alloc_map_uar(mdev, &sq->uar);
1158         if (err)
1159                 goto err_free_dma_tag;
1160
1161         err = mlx5_wq_cyc_create(mdev, &param->wq, sqc_wq, &sq->wq,
1162             &sq->wq_ctrl);
1163         if (err)
1164                 goto err_unmap_free_uar;
1165
1166         sq->wq.db = &sq->wq.db[MLX5_SND_DBR];
1167         sq->bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
1168
1169         err = mlx5e_alloc_sq_db(sq);
1170         if (err)
1171                 goto err_sq_wq_destroy;
1172
1173         sq->mkey_be = c->mkey_be;
1174         sq->ifp = priv->ifp;
1175         sq->priv = priv;
1176         sq->tc = tc;
1177         sq->max_inline = priv->params.tx_max_inline;
1178         sq->min_inline_mode = priv->params.tx_min_inline_mode;
1179         sq->vlan_inline_cap = MLX5_CAP_ETH(mdev, wqe_vlan_insert);
1180
1181         /* check if we should allocate a second packet buffer */
1182         if (priv->params_ethtool.tx_bufring_disable == 0) {
1183                 sq->br = buf_ring_alloc(MLX5E_SQ_TX_QUEUE_SIZE, M_MLX5EN,
1184                     M_WAITOK, &sq->lock);
1185                 if (sq->br == NULL) {
1186                         if_printf(c->ifp, "%s: Failed allocating sq drbr buffer\n",
1187                             __func__);
1188                         err = -ENOMEM;
1189                         goto err_free_sq_db;
1190                 }
1191
1192                 sq->sq_tq = taskqueue_create_fast("mlx5e_que", M_WAITOK,
1193                     taskqueue_thread_enqueue, &sq->sq_tq);
1194                 if (sq->sq_tq == NULL) {
1195                         if_printf(c->ifp, "%s: Failed allocating taskqueue\n",
1196                             __func__);
1197                         err = -ENOMEM;
1198                         goto err_free_drbr;
1199                 }
1200
1201                 TASK_INIT(&sq->sq_task, 0, mlx5e_tx_que, sq);
1202 #ifdef RSS
1203                 cpu_id = rss_getcpu(c->ix % rss_getnumbuckets());
1204                 CPU_SETOF(cpu_id, &cpu_mask);
1205                 taskqueue_start_threads_cpuset(&sq->sq_tq, 1, PI_NET, &cpu_mask,
1206                     "%s TX SQ%d.%d CPU%d", c->ifp->if_xname, c->ix, tc, cpu_id);
1207 #else
1208                 taskqueue_start_threads(&sq->sq_tq, 1, PI_NET,
1209                     "%s TX SQ%d.%d", c->ifp->if_xname, c->ix, tc);
1210 #endif
1211         }
1212         snprintf(buffer, sizeof(buffer), "txstat%dtc%d", c->ix, tc);
1213         mlx5e_create_stats(&sq->stats.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
1214             buffer, mlx5e_sq_stats_desc, MLX5E_SQ_STATS_NUM,
1215             sq->stats.arg);
1216
1217         return (0);
1218
1219 err_free_drbr:
1220         buf_ring_free(sq->br, M_MLX5EN);
1221 err_free_sq_db:
1222         mlx5e_free_sq_db(sq);
1223 err_sq_wq_destroy:
1224         mlx5_wq_destroy(&sq->wq_ctrl);
1225
1226 err_unmap_free_uar:
1227         mlx5_unmap_free_uar(mdev, &sq->uar);
1228
1229 err_free_dma_tag:
1230         bus_dma_tag_destroy(sq->dma_tag);
1231 done:
1232         return (err);
1233 }
1234
1235 static void
1236 mlx5e_destroy_sq(struct mlx5e_sq *sq)
1237 {
1238         /* destroy all sysctl nodes */
1239         sysctl_ctx_free(&sq->stats.ctx);
1240
1241         mlx5e_free_sq_db(sq);
1242         mlx5_wq_destroy(&sq->wq_ctrl);
1243         mlx5_unmap_free_uar(sq->priv->mdev, &sq->uar);
1244         if (sq->sq_tq != NULL) {
1245                 taskqueue_drain(sq->sq_tq, &sq->sq_task);
1246                 taskqueue_free(sq->sq_tq);
1247         }
1248         if (sq->br != NULL)
1249                 buf_ring_free(sq->br, M_MLX5EN);
1250 }
1251
1252 int
1253 mlx5e_enable_sq(struct mlx5e_sq *sq, struct mlx5e_sq_param *param,
1254     int tis_num)
1255 {
1256         void *in;
1257         void *sqc;
1258         void *wq;
1259         int inlen;
1260         int err;
1261
1262         inlen = MLX5_ST_SZ_BYTES(create_sq_in) +
1263             sizeof(u64) * sq->wq_ctrl.buf.npages;
1264         in = mlx5_vzalloc(inlen);
1265         if (in == NULL)
1266                 return (-ENOMEM);
1267
1268         sqc = MLX5_ADDR_OF(create_sq_in, in, ctx);
1269         wq = MLX5_ADDR_OF(sqc, sqc, wq);
1270
1271         memcpy(sqc, param->sqc, sizeof(param->sqc));
1272
1273         MLX5_SET(sqc, sqc, tis_num_0, tis_num);
1274         MLX5_SET(sqc, sqc, cqn, sq->cq.mcq.cqn);
1275         MLX5_SET(sqc, sqc, state, MLX5_SQC_STATE_RST);
1276         MLX5_SET(sqc, sqc, tis_lst_sz, 1);
1277         MLX5_SET(sqc, sqc, flush_in_error_en, 1);
1278
1279         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_CYCLIC);
1280         MLX5_SET(wq, wq, uar_page, sq->uar.index);
1281         MLX5_SET(wq, wq, log_wq_pg_sz, sq->wq_ctrl.buf.page_shift -
1282             PAGE_SHIFT);
1283         MLX5_SET64(wq, wq, dbr_addr, sq->wq_ctrl.db.dma);
1284
1285         mlx5_fill_page_array(&sq->wq_ctrl.buf,
1286             (__be64 *) MLX5_ADDR_OF(wq, wq, pas));
1287
1288         err = mlx5_core_create_sq(sq->priv->mdev, in, inlen, &sq->sqn);
1289
1290         kvfree(in);
1291
1292         return (err);
1293 }
1294
1295 int
1296 mlx5e_modify_sq(struct mlx5e_sq *sq, int curr_state, int next_state)
1297 {
1298         void *in;
1299         void *sqc;
1300         int inlen;
1301         int err;
1302
1303         inlen = MLX5_ST_SZ_BYTES(modify_sq_in);
1304         in = mlx5_vzalloc(inlen);
1305         if (in == NULL)
1306                 return (-ENOMEM);
1307
1308         sqc = MLX5_ADDR_OF(modify_sq_in, in, ctx);
1309
1310         MLX5_SET(modify_sq_in, in, sqn, sq->sqn);
1311         MLX5_SET(modify_sq_in, in, sq_state, curr_state);
1312         MLX5_SET(sqc, sqc, state, next_state);
1313
1314         err = mlx5_core_modify_sq(sq->priv->mdev, in, inlen);
1315
1316         kvfree(in);
1317
1318         return (err);
1319 }
1320
1321 void
1322 mlx5e_disable_sq(struct mlx5e_sq *sq)
1323 {
1324
1325         mlx5_core_destroy_sq(sq->priv->mdev, sq->sqn);
1326 }
1327
1328 static int
1329 mlx5e_open_sq(struct mlx5e_channel *c,
1330     int tc,
1331     struct mlx5e_sq_param *param,
1332     struct mlx5e_sq *sq)
1333 {
1334         int err;
1335
1336         err = mlx5e_create_sq(c, tc, param, sq);
1337         if (err)
1338                 return (err);
1339
1340         err = mlx5e_enable_sq(sq, param, c->priv->tisn[tc]);
1341         if (err)
1342                 goto err_destroy_sq;
1343
1344         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST, MLX5_SQC_STATE_RDY);
1345         if (err)
1346                 goto err_disable_sq;
1347
1348         WRITE_ONCE(sq->queue_state, MLX5E_SQ_READY);
1349
1350         return (0);
1351
1352 err_disable_sq:
1353         mlx5e_disable_sq(sq);
1354 err_destroy_sq:
1355         mlx5e_destroy_sq(sq);
1356
1357         return (err);
1358 }
1359
1360 static void
1361 mlx5e_sq_send_nops_locked(struct mlx5e_sq *sq, int can_sleep)
1362 {
1363         /* fill up remainder with NOPs */
1364         while (sq->cev_counter != 0) {
1365                 while (!mlx5e_sq_has_room_for(sq, 1)) {
1366                         if (can_sleep != 0) {
1367                                 mtx_unlock(&sq->lock);
1368                                 msleep(4);
1369                                 mtx_lock(&sq->lock);
1370                         } else {
1371                                 goto done;
1372                         }
1373                 }
1374                 /* send a single NOP */
1375                 mlx5e_send_nop(sq, 1);
1376                 atomic_thread_fence_rel();
1377         }
1378 done:
1379         /* Check if we need to write the doorbell */
1380         if (likely(sq->doorbell.d64 != 0)) {
1381                 mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
1382                 sq->doorbell.d64 = 0;
1383         }
1384 }
1385
1386 void
1387 mlx5e_sq_cev_timeout(void *arg)
1388 {
1389         struct mlx5e_sq *sq = arg;
1390
1391         mtx_assert(&sq->lock, MA_OWNED);
1392
1393         /* check next state */
1394         switch (sq->cev_next_state) {
1395         case MLX5E_CEV_STATE_SEND_NOPS:
1396                 /* fill TX ring with NOPs, if any */
1397                 mlx5e_sq_send_nops_locked(sq, 0);
1398
1399                 /* check if completed */
1400                 if (sq->cev_counter == 0) {
1401                         sq->cev_next_state = MLX5E_CEV_STATE_INITIAL;
1402                         return;
1403                 }
1404                 break;
1405         default:
1406                 /* send NOPs on next timeout */
1407                 sq->cev_next_state = MLX5E_CEV_STATE_SEND_NOPS;
1408                 break;
1409         }
1410
1411         /* restart timer */
1412         callout_reset_curcpu(&sq->cev_callout, hz, mlx5e_sq_cev_timeout, sq);
1413 }
1414
1415 void
1416 mlx5e_drain_sq(struct mlx5e_sq *sq)
1417 {
1418         int error;
1419         struct mlx5_core_dev *mdev= sq->priv->mdev;
1420
1421         /*
1422          * Check if already stopped.
1423          *
1424          * NOTE: The "stopped" variable is only written when both the
1425          * priv's configuration lock and the SQ's lock is locked. It
1426          * can therefore safely be read when only one of the two locks
1427          * is locked. This function is always called when the priv's
1428          * configuration lock is locked.
1429          */
1430         if (sq->stopped != 0)
1431                 return;
1432
1433         mtx_lock(&sq->lock);
1434
1435         /* don't put more packets into the SQ */
1436         sq->stopped = 1;
1437
1438         /* teardown event factor timer, if any */
1439         sq->cev_next_state = MLX5E_CEV_STATE_HOLD_NOPS;
1440         callout_stop(&sq->cev_callout);
1441
1442         /* send dummy NOPs in order to flush the transmit ring */
1443         mlx5e_sq_send_nops_locked(sq, 1);
1444         mtx_unlock(&sq->lock);
1445
1446         /* make sure it is safe to free the callout */
1447         callout_drain(&sq->cev_callout);
1448
1449         /* wait till SQ is empty or link is down */
1450         mtx_lock(&sq->lock);
1451         while (sq->cc != sq->pc &&
1452             (sq->priv->media_status_last & IFM_ACTIVE) != 0 &&
1453             mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
1454                 mtx_unlock(&sq->lock);
1455                 msleep(1);
1456                 sq->cq.mcq.comp(&sq->cq.mcq);
1457                 mtx_lock(&sq->lock);
1458         }
1459         mtx_unlock(&sq->lock);
1460
1461         /* error out remaining requests */
1462         error = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RDY, MLX5_SQC_STATE_ERR);
1463         if (error != 0) {
1464                 if_printf(sq->ifp,
1465                     "mlx5e_modify_sq() from RDY to ERR failed: %d\n", error);
1466         }
1467
1468         /* wait till SQ is empty */
1469         mtx_lock(&sq->lock);
1470         while (sq->cc != sq->pc &&
1471                mdev->state != MLX5_DEVICE_STATE_INTERNAL_ERROR) {
1472                 mtx_unlock(&sq->lock);
1473                 msleep(1);
1474                 sq->cq.mcq.comp(&sq->cq.mcq);
1475                 mtx_lock(&sq->lock);
1476         }
1477         mtx_unlock(&sq->lock);
1478 }
1479
1480 static void
1481 mlx5e_close_sq_wait(struct mlx5e_sq *sq)
1482 {
1483
1484         mlx5e_drain_sq(sq);
1485         mlx5e_disable_sq(sq);
1486         mlx5e_destroy_sq(sq);
1487 }
1488
1489 static int
1490 mlx5e_create_cq(struct mlx5e_priv *priv,
1491     struct mlx5e_cq_param *param,
1492     struct mlx5e_cq *cq,
1493     mlx5e_cq_comp_t *comp,
1494     int eq_ix)
1495 {
1496         struct mlx5_core_dev *mdev = priv->mdev;
1497         struct mlx5_core_cq *mcq = &cq->mcq;
1498         int eqn_not_used;
1499         int irqn;
1500         int err;
1501         u32 i;
1502
1503         param->wq.buf_numa_node = 0;
1504         param->wq.db_numa_node = 0;
1505
1506         err = mlx5_cqwq_create(mdev, &param->wq, param->cqc, &cq->wq,
1507             &cq->wq_ctrl);
1508         if (err)
1509                 return (err);
1510
1511         mlx5_vector2eqn(mdev, eq_ix, &eqn_not_used, &irqn);
1512
1513         mcq->cqe_sz = 64;
1514         mcq->set_ci_db = cq->wq_ctrl.db.db;
1515         mcq->arm_db = cq->wq_ctrl.db.db + 1;
1516         *mcq->set_ci_db = 0;
1517         *mcq->arm_db = 0;
1518         mcq->vector = eq_ix;
1519         mcq->comp = comp;
1520         mcq->event = mlx5e_cq_error_event;
1521         mcq->irqn = irqn;
1522         mcq->uar = &priv->cq_uar;
1523
1524         for (i = 0; i < mlx5_cqwq_get_size(&cq->wq); i++) {
1525                 struct mlx5_cqe64 *cqe = mlx5_cqwq_get_wqe(&cq->wq, i);
1526
1527                 cqe->op_own = 0xf1;
1528         }
1529
1530         cq->priv = priv;
1531
1532         return (0);
1533 }
1534
1535 static void
1536 mlx5e_destroy_cq(struct mlx5e_cq *cq)
1537 {
1538         mlx5_wq_destroy(&cq->wq_ctrl);
1539 }
1540
1541 static int
1542 mlx5e_enable_cq(struct mlx5e_cq *cq, struct mlx5e_cq_param *param, int eq_ix)
1543 {
1544         struct mlx5_core_cq *mcq = &cq->mcq;
1545         void *in;
1546         void *cqc;
1547         int inlen;
1548         int irqn_not_used;
1549         int eqn;
1550         int err;
1551
1552         inlen = MLX5_ST_SZ_BYTES(create_cq_in) +
1553             sizeof(u64) * cq->wq_ctrl.buf.npages;
1554         in = mlx5_vzalloc(inlen);
1555         if (in == NULL)
1556                 return (-ENOMEM);
1557
1558         cqc = MLX5_ADDR_OF(create_cq_in, in, cq_context);
1559
1560         memcpy(cqc, param->cqc, sizeof(param->cqc));
1561
1562         mlx5_fill_page_array(&cq->wq_ctrl.buf,
1563             (__be64 *) MLX5_ADDR_OF(create_cq_in, in, pas));
1564
1565         mlx5_vector2eqn(cq->priv->mdev, eq_ix, &eqn, &irqn_not_used);
1566
1567         MLX5_SET(cqc, cqc, c_eqn, eqn);
1568         MLX5_SET(cqc, cqc, uar_page, mcq->uar->index);
1569         MLX5_SET(cqc, cqc, log_page_size, cq->wq_ctrl.buf.page_shift -
1570             PAGE_SHIFT);
1571         MLX5_SET64(cqc, cqc, dbr_addr, cq->wq_ctrl.db.dma);
1572
1573         err = mlx5_core_create_cq(cq->priv->mdev, mcq, in, inlen);
1574
1575         kvfree(in);
1576
1577         if (err)
1578                 return (err);
1579
1580         mlx5e_cq_arm(cq, MLX5_GET_DOORBELL_LOCK(&cq->priv->doorbell_lock));
1581
1582         return (0);
1583 }
1584
1585 static void
1586 mlx5e_disable_cq(struct mlx5e_cq *cq)
1587 {
1588
1589         mlx5_core_destroy_cq(cq->priv->mdev, &cq->mcq);
1590 }
1591
1592 int
1593 mlx5e_open_cq(struct mlx5e_priv *priv,
1594     struct mlx5e_cq_param *param,
1595     struct mlx5e_cq *cq,
1596     mlx5e_cq_comp_t *comp,
1597     int eq_ix)
1598 {
1599         int err;
1600
1601         err = mlx5e_create_cq(priv, param, cq, comp, eq_ix);
1602         if (err)
1603                 return (err);
1604
1605         err = mlx5e_enable_cq(cq, param, eq_ix);
1606         if (err)
1607                 goto err_destroy_cq;
1608
1609         return (0);
1610
1611 err_destroy_cq:
1612         mlx5e_destroy_cq(cq);
1613
1614         return (err);
1615 }
1616
1617 void
1618 mlx5e_close_cq(struct mlx5e_cq *cq)
1619 {
1620         mlx5e_disable_cq(cq);
1621         mlx5e_destroy_cq(cq);
1622 }
1623
1624 static int
1625 mlx5e_open_tx_cqs(struct mlx5e_channel *c,
1626     struct mlx5e_channel_param *cparam)
1627 {
1628         int err;
1629         int tc;
1630
1631         for (tc = 0; tc < c->num_tc; tc++) {
1632                 /* open completion queue */
1633                 err = mlx5e_open_cq(c->priv, &cparam->tx_cq, &c->sq[tc].cq,
1634                     &mlx5e_tx_cq_comp, c->ix);
1635                 if (err)
1636                         goto err_close_tx_cqs;
1637         }
1638         return (0);
1639
1640 err_close_tx_cqs:
1641         for (tc--; tc >= 0; tc--)
1642                 mlx5e_close_cq(&c->sq[tc].cq);
1643
1644         return (err);
1645 }
1646
1647 static void
1648 mlx5e_close_tx_cqs(struct mlx5e_channel *c)
1649 {
1650         int tc;
1651
1652         for (tc = 0; tc < c->num_tc; tc++)
1653                 mlx5e_close_cq(&c->sq[tc].cq);
1654 }
1655
1656 static int
1657 mlx5e_open_sqs(struct mlx5e_channel *c,
1658     struct mlx5e_channel_param *cparam)
1659 {
1660         int err;
1661         int tc;
1662
1663         for (tc = 0; tc < c->num_tc; tc++) {
1664                 err = mlx5e_open_sq(c, tc, &cparam->sq, &c->sq[tc]);
1665                 if (err)
1666                         goto err_close_sqs;
1667         }
1668
1669         return (0);
1670
1671 err_close_sqs:
1672         for (tc--; tc >= 0; tc--)
1673                 mlx5e_close_sq_wait(&c->sq[tc]);
1674
1675         return (err);
1676 }
1677
1678 static void
1679 mlx5e_close_sqs_wait(struct mlx5e_channel *c)
1680 {
1681         int tc;
1682
1683         for (tc = 0; tc < c->num_tc; tc++)
1684                 mlx5e_close_sq_wait(&c->sq[tc]);
1685 }
1686
1687 static void
1688 mlx5e_chan_mtx_init(struct mlx5e_channel *c)
1689 {
1690         int tc;
1691
1692         mtx_init(&c->rq.mtx, "mlx5rx", MTX_NETWORK_LOCK, MTX_DEF);
1693
1694         callout_init_mtx(&c->rq.watchdog, &c->rq.mtx, 0);
1695
1696         for (tc = 0; tc < c->num_tc; tc++) {
1697                 struct mlx5e_sq *sq = c->sq + tc;
1698
1699                 mtx_init(&sq->lock, "mlx5tx",
1700                     MTX_NETWORK_LOCK " TX", MTX_DEF);
1701                 mtx_init(&sq->comp_lock, "mlx5comp",
1702                     MTX_NETWORK_LOCK " TX", MTX_DEF);
1703
1704                 callout_init_mtx(&sq->cev_callout, &sq->lock, 0);
1705
1706                 sq->cev_factor = c->priv->params_ethtool.tx_completion_fact;
1707
1708                 /* ensure the TX completion event factor is not zero */
1709                 if (sq->cev_factor == 0)
1710                         sq->cev_factor = 1;
1711         }
1712 }
1713
1714 static void
1715 mlx5e_chan_mtx_destroy(struct mlx5e_channel *c)
1716 {
1717         int tc;
1718
1719         mtx_destroy(&c->rq.mtx);
1720
1721         for (tc = 0; tc < c->num_tc; tc++) {
1722                 mtx_destroy(&c->sq[tc].lock);
1723                 mtx_destroy(&c->sq[tc].comp_lock);
1724         }
1725 }
1726
1727 static int
1728 mlx5e_open_channel(struct mlx5e_priv *priv, int ix,
1729     struct mlx5e_channel_param *cparam,
1730     struct mlx5e_channel *volatile *cp)
1731 {
1732         struct mlx5e_channel *c;
1733         int err;
1734
1735         c = malloc(sizeof(*c), M_MLX5EN, M_WAITOK | M_ZERO);
1736         c->priv = priv;
1737         c->ix = ix;
1738         c->cpu = 0;
1739         c->ifp = priv->ifp;
1740         c->mkey_be = cpu_to_be32(priv->mr.key);
1741         c->num_tc = priv->num_tc;
1742
1743         /* init mutexes */
1744         mlx5e_chan_mtx_init(c);
1745
1746         /* open transmit completion queue */
1747         err = mlx5e_open_tx_cqs(c, cparam);
1748         if (err)
1749                 goto err_free;
1750
1751         /* open receive completion queue */
1752         err = mlx5e_open_cq(c->priv, &cparam->rx_cq, &c->rq.cq,
1753             &mlx5e_rx_cq_comp, c->ix);
1754         if (err)
1755                 goto err_close_tx_cqs;
1756
1757         err = mlx5e_open_sqs(c, cparam);
1758         if (err)
1759                 goto err_close_rx_cq;
1760
1761         err = mlx5e_open_rq(c, &cparam->rq, &c->rq);
1762         if (err)
1763                 goto err_close_sqs;
1764
1765         /* store channel pointer */
1766         *cp = c;
1767
1768         /* poll receive queue initially */
1769         c->rq.cq.mcq.comp(&c->rq.cq.mcq);
1770
1771         return (0);
1772
1773 err_close_sqs:
1774         mlx5e_close_sqs_wait(c);
1775
1776 err_close_rx_cq:
1777         mlx5e_close_cq(&c->rq.cq);
1778
1779 err_close_tx_cqs:
1780         mlx5e_close_tx_cqs(c);
1781
1782 err_free:
1783         /* destroy mutexes */
1784         mlx5e_chan_mtx_destroy(c);
1785         free(c, M_MLX5EN);
1786         return (err);
1787 }
1788
1789 static void
1790 mlx5e_close_channel(struct mlx5e_channel *volatile *pp)
1791 {
1792         struct mlx5e_channel *c = *pp;
1793
1794         /* check if channel is already closed */
1795         if (c == NULL)
1796                 return;
1797         mlx5e_close_rq(&c->rq);
1798 }
1799
1800 static void
1801 mlx5e_close_channel_wait(struct mlx5e_channel *volatile *pp)
1802 {
1803         struct mlx5e_channel *c = *pp;
1804
1805         /* check if channel is already closed */
1806         if (c == NULL)
1807                 return;
1808         /* ensure channel pointer is no longer used */
1809         *pp = NULL;
1810
1811         mlx5e_close_rq_wait(&c->rq);
1812         mlx5e_close_sqs_wait(c);
1813         mlx5e_close_cq(&c->rq.cq);
1814         mlx5e_close_tx_cqs(c);
1815         /* destroy mutexes */
1816         mlx5e_chan_mtx_destroy(c);
1817         free(c, M_MLX5EN);
1818 }
1819
1820 static int
1821 mlx5e_get_wqe_sz(struct mlx5e_priv *priv, u32 *wqe_sz, u32 *nsegs)
1822 {
1823         u32 r, n;
1824
1825         r = priv->params.hw_lro_en ? priv->params.lro_wqe_sz :
1826             MLX5E_SW2MB_MTU(priv->ifp->if_mtu);
1827         if (r > MJUM16BYTES)
1828                 return (-ENOMEM);
1829
1830         if (r > MJUM9BYTES)
1831                 r = MJUM16BYTES;
1832         else if (r > MJUMPAGESIZE)
1833                 r = MJUM9BYTES;
1834         else if (r > MCLBYTES)
1835                 r = MJUMPAGESIZE;
1836         else
1837                 r = MCLBYTES;
1838
1839         /*
1840          * n + 1 must be a power of two, because stride size must be.
1841          * Stride size is 16 * (n + 1), as the first segment is
1842          * control.
1843          */
1844         for (n = howmany(r, MLX5E_MAX_RX_BYTES); !powerof2(n + 1); n++)
1845                 ;
1846
1847         *wqe_sz = r;
1848         *nsegs = n;
1849         return (0);
1850 }
1851
1852 static void
1853 mlx5e_build_rq_param(struct mlx5e_priv *priv,
1854     struct mlx5e_rq_param *param)
1855 {
1856         void *rqc = param->rqc;
1857         void *wq = MLX5_ADDR_OF(rqc, rqc, wq);
1858         u32 wqe_sz, nsegs;
1859
1860         mlx5e_get_wqe_sz(priv, &wqe_sz, &nsegs);
1861         MLX5_SET(wq, wq, wq_type, MLX5_WQ_TYPE_LINKED_LIST);
1862         MLX5_SET(wq, wq, end_padding_mode, MLX5_WQ_END_PAD_MODE_ALIGN);
1863         MLX5_SET(wq, wq, log_wq_stride, ilog2(sizeof(struct mlx5e_rx_wqe) +
1864             nsegs * sizeof(struct mlx5_wqe_data_seg)));
1865         MLX5_SET(wq, wq, log_wq_sz, priv->params.log_rq_size);
1866         MLX5_SET(wq, wq, pd, priv->pdn);
1867
1868         param->wq.buf_numa_node = 0;
1869         param->wq.db_numa_node = 0;
1870         param->wq.linear = 1;
1871 }
1872
1873 static void
1874 mlx5e_build_sq_param(struct mlx5e_priv *priv,
1875     struct mlx5e_sq_param *param)
1876 {
1877         void *sqc = param->sqc;
1878         void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
1879
1880         MLX5_SET(wq, wq, log_wq_sz, priv->params.log_sq_size);
1881         MLX5_SET(wq, wq, log_wq_stride, ilog2(MLX5_SEND_WQE_BB));
1882         MLX5_SET(wq, wq, pd, priv->pdn);
1883
1884         param->wq.buf_numa_node = 0;
1885         param->wq.db_numa_node = 0;
1886         param->wq.linear = 1;
1887 }
1888
1889 static void
1890 mlx5e_build_common_cq_param(struct mlx5e_priv *priv,
1891     struct mlx5e_cq_param *param)
1892 {
1893         void *cqc = param->cqc;
1894
1895         MLX5_SET(cqc, cqc, uar_page, priv->cq_uar.index);
1896 }
1897
1898 static void
1899 mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
1900     struct mlx5e_cq_param *param)
1901 {
1902         void *cqc = param->cqc;
1903
1904
1905         /*
1906          * TODO The sysctl to control on/off is a bool value for now, which means
1907          * we only support CSUM, once HASH is implemnted we'll need to address that.
1908          */
1909         if (priv->params.cqe_zipping_en) {
1910                 MLX5_SET(cqc, cqc, mini_cqe_res_format, MLX5_CQE_FORMAT_CSUM);
1911                 MLX5_SET(cqc, cqc, cqe_compression_en, 1);
1912         }
1913
1914         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_rq_size);
1915         MLX5_SET(cqc, cqc, cq_period, priv->params.rx_cq_moderation_usec);
1916         MLX5_SET(cqc, cqc, cq_max_count, priv->params.rx_cq_moderation_pkts);
1917
1918         switch (priv->params.rx_cq_moderation_mode) {
1919         case 0:
1920                 MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
1921                 break;
1922         default:
1923                 if (MLX5_CAP_GEN(priv->mdev, cq_period_start_from_cqe))
1924                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
1925                 else
1926                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
1927                 break;
1928         }
1929
1930         mlx5e_build_common_cq_param(priv, param);
1931 }
1932
1933 static void
1934 mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
1935     struct mlx5e_cq_param *param)
1936 {
1937         void *cqc = param->cqc;
1938
1939         MLX5_SET(cqc, cqc, log_cq_size, priv->params.log_sq_size);
1940         MLX5_SET(cqc, cqc, cq_period, priv->params.tx_cq_moderation_usec);
1941         MLX5_SET(cqc, cqc, cq_max_count, priv->params.tx_cq_moderation_pkts);
1942
1943         switch (priv->params.tx_cq_moderation_mode) {
1944         case 0:
1945                 MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
1946                 break;
1947         default:
1948                 if (MLX5_CAP_GEN(priv->mdev, cq_period_start_from_cqe))
1949                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_CQE);
1950                 else
1951                         MLX5_SET(cqc, cqc, cq_period_mode, MLX5_CQ_PERIOD_MODE_START_FROM_EQE);
1952                 break;
1953         }
1954
1955         mlx5e_build_common_cq_param(priv, param);
1956 }
1957
1958 static void
1959 mlx5e_build_channel_param(struct mlx5e_priv *priv,
1960     struct mlx5e_channel_param *cparam)
1961 {
1962         memset(cparam, 0, sizeof(*cparam));
1963
1964         mlx5e_build_rq_param(priv, &cparam->rq);
1965         mlx5e_build_sq_param(priv, &cparam->sq);
1966         mlx5e_build_rx_cq_param(priv, &cparam->rx_cq);
1967         mlx5e_build_tx_cq_param(priv, &cparam->tx_cq);
1968 }
1969
1970 static int
1971 mlx5e_open_channels(struct mlx5e_priv *priv)
1972 {
1973         struct mlx5e_channel_param cparam;
1974         void *ptr;
1975         int err;
1976         int i;
1977         int j;
1978
1979         priv->channel = malloc(priv->params.num_channels *
1980             sizeof(struct mlx5e_channel *), M_MLX5EN, M_WAITOK | M_ZERO);
1981
1982         mlx5e_build_channel_param(priv, &cparam);
1983         for (i = 0; i < priv->params.num_channels; i++) {
1984                 err = mlx5e_open_channel(priv, i, &cparam, &priv->channel[i]);
1985                 if (err)
1986                         goto err_close_channels;
1987         }
1988
1989         for (j = 0; j < priv->params.num_channels; j++) {
1990                 err = mlx5e_wait_for_min_rx_wqes(&priv->channel[j]->rq);
1991                 if (err)
1992                         goto err_close_channels;
1993         }
1994
1995         return (0);
1996
1997 err_close_channels:
1998         for (i--; i >= 0; i--) {
1999                 mlx5e_close_channel(&priv->channel[i]);
2000                 mlx5e_close_channel_wait(&priv->channel[i]);
2001         }
2002
2003         /* remove "volatile" attribute from "channel" pointer */
2004         ptr = __DECONST(void *, priv->channel);
2005         priv->channel = NULL;
2006
2007         free(ptr, M_MLX5EN);
2008
2009         return (err);
2010 }
2011
2012 static void
2013 mlx5e_close_channels(struct mlx5e_priv *priv)
2014 {
2015         void *ptr;
2016         int i;
2017
2018         if (priv->channel == NULL)
2019                 return;
2020
2021         for (i = 0; i < priv->params.num_channels; i++)
2022                 mlx5e_close_channel(&priv->channel[i]);
2023         for (i = 0; i < priv->params.num_channels; i++)
2024                 mlx5e_close_channel_wait(&priv->channel[i]);
2025
2026         /* remove "volatile" attribute from "channel" pointer */
2027         ptr = __DECONST(void *, priv->channel);
2028         priv->channel = NULL;
2029
2030         free(ptr, M_MLX5EN);
2031 }
2032
2033 static int
2034 mlx5e_refresh_sq_params(struct mlx5e_priv *priv, struct mlx5e_sq *sq)
2035 {
2036
2037         if (MLX5_CAP_GEN(priv->mdev, cq_period_mode_modify)) {
2038                 uint8_t cq_mode;
2039
2040                 switch (priv->params.tx_cq_moderation_mode) {
2041                 case 0:
2042                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
2043                         break;
2044                 default:
2045                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_CQE;
2046                         break;
2047                 }
2048
2049                 return (mlx5_core_modify_cq_moderation_mode(priv->mdev, &sq->cq.mcq,
2050                     priv->params.tx_cq_moderation_usec,
2051                     priv->params.tx_cq_moderation_pkts,
2052                     cq_mode));
2053         }
2054
2055         return (mlx5_core_modify_cq_moderation(priv->mdev, &sq->cq.mcq,
2056             priv->params.tx_cq_moderation_usec,
2057             priv->params.tx_cq_moderation_pkts));
2058 }
2059
2060 static int
2061 mlx5e_refresh_rq_params(struct mlx5e_priv *priv, struct mlx5e_rq *rq)
2062 {
2063
2064         if (MLX5_CAP_GEN(priv->mdev, cq_period_mode_modify)) {
2065                 uint8_t cq_mode;
2066                 int retval;
2067
2068                 switch (priv->params.rx_cq_moderation_mode) {
2069                 case 0:
2070                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_EQE;
2071                         break;
2072                 default:
2073                         cq_mode = MLX5_CQ_PERIOD_MODE_START_FROM_CQE;
2074                         break;
2075                 }
2076
2077                 retval = mlx5_core_modify_cq_moderation_mode(priv->mdev, &rq->cq.mcq,
2078                     priv->params.rx_cq_moderation_usec,
2079                     priv->params.rx_cq_moderation_pkts,
2080                     cq_mode);
2081
2082                 return (retval);
2083         }
2084
2085         return (mlx5_core_modify_cq_moderation(priv->mdev, &rq->cq.mcq,
2086             priv->params.rx_cq_moderation_usec,
2087             priv->params.rx_cq_moderation_pkts));
2088 }
2089
2090 static int
2091 mlx5e_refresh_channel_params_sub(struct mlx5e_priv *priv, struct mlx5e_channel *c)
2092 {
2093         int err;
2094         int i;
2095
2096         if (c == NULL)
2097                 return (EINVAL);
2098
2099         err = mlx5e_refresh_rq_params(priv, &c->rq);
2100         if (err)
2101                 goto done;
2102
2103         for (i = 0; i != c->num_tc; i++) {
2104                 err = mlx5e_refresh_sq_params(priv, &c->sq[i]);
2105                 if (err)
2106                         goto done;
2107         }
2108 done:
2109         return (err);
2110 }
2111
2112 int
2113 mlx5e_refresh_channel_params(struct mlx5e_priv *priv)
2114 {
2115         int i;
2116
2117         if (priv->channel == NULL)
2118                 return (EINVAL);
2119
2120         for (i = 0; i < priv->params.num_channels; i++) {
2121                 int err;
2122
2123                 err = mlx5e_refresh_channel_params_sub(priv, priv->channel[i]);
2124                 if (err)
2125                         return (err);
2126         }
2127         return (0);
2128 }
2129
2130 static int
2131 mlx5e_open_tis(struct mlx5e_priv *priv, int tc)
2132 {
2133         struct mlx5_core_dev *mdev = priv->mdev;
2134         u32 in[MLX5_ST_SZ_DW(create_tis_in)];
2135         void *tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
2136
2137         memset(in, 0, sizeof(in));
2138
2139         MLX5_SET(tisc, tisc, prio, tc);
2140         MLX5_SET(tisc, tisc, transport_domain, priv->tdn);
2141
2142         return (mlx5_core_create_tis(mdev, in, sizeof(in), &priv->tisn[tc]));
2143 }
2144
2145 static void
2146 mlx5e_close_tis(struct mlx5e_priv *priv, int tc)
2147 {
2148         mlx5_core_destroy_tis(priv->mdev, priv->tisn[tc]);
2149 }
2150
2151 static int
2152 mlx5e_open_tises(struct mlx5e_priv *priv)
2153 {
2154         int num_tc = priv->num_tc;
2155         int err;
2156         int tc;
2157
2158         for (tc = 0; tc < num_tc; tc++) {
2159                 err = mlx5e_open_tis(priv, tc);
2160                 if (err)
2161                         goto err_close_tises;
2162         }
2163
2164         return (0);
2165
2166 err_close_tises:
2167         for (tc--; tc >= 0; tc--)
2168                 mlx5e_close_tis(priv, tc);
2169
2170         return (err);
2171 }
2172
2173 static void
2174 mlx5e_close_tises(struct mlx5e_priv *priv)
2175 {
2176         int num_tc = priv->num_tc;
2177         int tc;
2178
2179         for (tc = 0; tc < num_tc; tc++)
2180                 mlx5e_close_tis(priv, tc);
2181 }
2182
2183 static int
2184 mlx5e_open_rqt(struct mlx5e_priv *priv)
2185 {
2186         struct mlx5_core_dev *mdev = priv->mdev;
2187         u32 *in;
2188         u32 out[MLX5_ST_SZ_DW(create_rqt_out)] = {0};
2189         void *rqtc;
2190         int inlen;
2191         int err;
2192         int sz;
2193         int i;
2194
2195         sz = 1 << priv->params.rx_hash_log_tbl_sz;
2196
2197         inlen = MLX5_ST_SZ_BYTES(create_rqt_in) + sizeof(u32) * sz;
2198         in = mlx5_vzalloc(inlen);
2199         if (in == NULL)
2200                 return (-ENOMEM);
2201         rqtc = MLX5_ADDR_OF(create_rqt_in, in, rqt_context);
2202
2203         MLX5_SET(rqtc, rqtc, rqt_actual_size, sz);
2204         MLX5_SET(rqtc, rqtc, rqt_max_size, sz);
2205
2206         for (i = 0; i < sz; i++) {
2207                 int ix;
2208 #ifdef RSS
2209                 ix = rss_get_indirection_to_bucket(i);
2210 #else
2211                 ix = i;
2212 #endif
2213                 /* ensure we don't overflow */
2214                 ix %= priv->params.num_channels;
2215                 MLX5_SET(rqtc, rqtc, rq_num[i], priv->channel[ix]->rq.rqn);
2216         }
2217
2218         MLX5_SET(create_rqt_in, in, opcode, MLX5_CMD_OP_CREATE_RQT);
2219
2220         err = mlx5_cmd_exec(mdev, in, inlen, out, sizeof(out));
2221         if (!err)
2222                 priv->rqtn = MLX5_GET(create_rqt_out, out, rqtn);
2223
2224         kvfree(in);
2225
2226         return (err);
2227 }
2228
2229 static void
2230 mlx5e_close_rqt(struct mlx5e_priv *priv)
2231 {
2232         u32 in[MLX5_ST_SZ_DW(destroy_rqt_in)] = {0};
2233         u32 out[MLX5_ST_SZ_DW(destroy_rqt_out)] = {0};
2234
2235         MLX5_SET(destroy_rqt_in, in, opcode, MLX5_CMD_OP_DESTROY_RQT);
2236         MLX5_SET(destroy_rqt_in, in, rqtn, priv->rqtn);
2237
2238         mlx5_cmd_exec(priv->mdev, in, sizeof(in), out, sizeof(out));
2239 }
2240
2241 static void
2242 mlx5e_build_tir_ctx(struct mlx5e_priv *priv, u32 * tirc, int tt)
2243 {
2244         void *hfso = MLX5_ADDR_OF(tirc, tirc, rx_hash_field_selector_outer);
2245         __be32 *hkey;
2246
2247         MLX5_SET(tirc, tirc, transport_domain, priv->tdn);
2248
2249 #define ROUGH_MAX_L2_L3_HDR_SZ 256
2250
2251 #define MLX5_HASH_IP     (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2252                           MLX5_HASH_FIELD_SEL_DST_IP)
2253
2254 #define MLX5_HASH_ALL    (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2255                           MLX5_HASH_FIELD_SEL_DST_IP   |\
2256                           MLX5_HASH_FIELD_SEL_L4_SPORT |\
2257                           MLX5_HASH_FIELD_SEL_L4_DPORT)
2258
2259 #define MLX5_HASH_IP_IPSEC_SPI  (MLX5_HASH_FIELD_SEL_SRC_IP   |\
2260                                  MLX5_HASH_FIELD_SEL_DST_IP   |\
2261                                  MLX5_HASH_FIELD_SEL_IPSEC_SPI)
2262
2263         if (priv->params.hw_lro_en) {
2264                 MLX5_SET(tirc, tirc, lro_enable_mask,
2265                     MLX5_TIRC_LRO_ENABLE_MASK_IPV4_LRO |
2266                     MLX5_TIRC_LRO_ENABLE_MASK_IPV6_LRO);
2267                 MLX5_SET(tirc, tirc, lro_max_msg_sz,
2268                     (priv->params.lro_wqe_sz -
2269                     ROUGH_MAX_L2_L3_HDR_SZ) >> 8);
2270                 /* TODO: add the option to choose timer value dynamically */
2271                 MLX5_SET(tirc, tirc, lro_timeout_period_usecs,
2272                     MLX5_CAP_ETH(priv->mdev,
2273                     lro_timer_supported_periods[2]));
2274         }
2275
2276         /* setup parameters for hashing TIR type, if any */
2277         switch (tt) {
2278         case MLX5E_TT_ANY:
2279                 MLX5_SET(tirc, tirc, disp_type,
2280                     MLX5_TIRC_DISP_TYPE_DIRECT);
2281                 MLX5_SET(tirc, tirc, inline_rqn,
2282                     priv->channel[0]->rq.rqn);
2283                 break;
2284         default:
2285                 MLX5_SET(tirc, tirc, disp_type,
2286                     MLX5_TIRC_DISP_TYPE_INDIRECT);
2287                 MLX5_SET(tirc, tirc, indirect_table,
2288                     priv->rqtn);
2289                 MLX5_SET(tirc, tirc, rx_hash_fn,
2290                     MLX5_TIRC_RX_HASH_FN_HASH_TOEPLITZ);
2291                 hkey = (__be32 *) MLX5_ADDR_OF(tirc, tirc, rx_hash_toeplitz_key);
2292 #ifdef RSS
2293                 /*
2294                  * The FreeBSD RSS implementation does currently not
2295                  * support symmetric Toeplitz hashes:
2296                  */
2297                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 0);
2298                 rss_getkey((uint8_t *)hkey);
2299 #else
2300                 MLX5_SET(tirc, tirc, rx_hash_symmetric, 1);
2301                 hkey[0] = cpu_to_be32(0xD181C62C);
2302                 hkey[1] = cpu_to_be32(0xF7F4DB5B);
2303                 hkey[2] = cpu_to_be32(0x1983A2FC);
2304                 hkey[3] = cpu_to_be32(0x943E1ADB);
2305                 hkey[4] = cpu_to_be32(0xD9389E6B);
2306                 hkey[5] = cpu_to_be32(0xD1039C2C);
2307                 hkey[6] = cpu_to_be32(0xA74499AD);
2308                 hkey[7] = cpu_to_be32(0x593D56D9);
2309                 hkey[8] = cpu_to_be32(0xF3253C06);
2310                 hkey[9] = cpu_to_be32(0x2ADC1FFC);
2311 #endif
2312                 break;
2313         }
2314
2315         switch (tt) {
2316         case MLX5E_TT_IPV4_TCP:
2317                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2318                     MLX5_L3_PROT_TYPE_IPV4);
2319                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2320                     MLX5_L4_PROT_TYPE_TCP);
2321 #ifdef RSS
2322                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV4)) {
2323                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2324                             MLX5_HASH_IP);
2325                 } else
2326 #endif
2327                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2328                     MLX5_HASH_ALL);
2329                 break;
2330
2331         case MLX5E_TT_IPV6_TCP:
2332                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2333                     MLX5_L3_PROT_TYPE_IPV6);
2334                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2335                     MLX5_L4_PROT_TYPE_TCP);
2336 #ifdef RSS
2337                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_TCP_IPV6)) {
2338                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2339                             MLX5_HASH_IP);
2340                 } else
2341 #endif
2342                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2343                     MLX5_HASH_ALL);
2344                 break;
2345
2346         case MLX5E_TT_IPV4_UDP:
2347                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2348                     MLX5_L3_PROT_TYPE_IPV4);
2349                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2350                     MLX5_L4_PROT_TYPE_UDP);
2351 #ifdef RSS
2352                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV4)) {
2353                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2354                             MLX5_HASH_IP);
2355                 } else
2356 #endif
2357                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2358                     MLX5_HASH_ALL);
2359                 break;
2360
2361         case MLX5E_TT_IPV6_UDP:
2362                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2363                     MLX5_L3_PROT_TYPE_IPV6);
2364                 MLX5_SET(rx_hash_field_select, hfso, l4_prot_type,
2365                     MLX5_L4_PROT_TYPE_UDP);
2366 #ifdef RSS
2367                 if (!(rss_gethashconfig() & RSS_HASHTYPE_RSS_UDP_IPV6)) {
2368                         MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2369                             MLX5_HASH_IP);
2370                 } else
2371 #endif
2372                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2373                     MLX5_HASH_ALL);
2374                 break;
2375
2376         case MLX5E_TT_IPV4_IPSEC_AH:
2377                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2378                     MLX5_L3_PROT_TYPE_IPV4);
2379                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2380                     MLX5_HASH_IP_IPSEC_SPI);
2381                 break;
2382
2383         case MLX5E_TT_IPV6_IPSEC_AH:
2384                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2385                     MLX5_L3_PROT_TYPE_IPV6);
2386                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2387                     MLX5_HASH_IP_IPSEC_SPI);
2388                 break;
2389
2390         case MLX5E_TT_IPV4_IPSEC_ESP:
2391                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2392                     MLX5_L3_PROT_TYPE_IPV4);
2393                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2394                     MLX5_HASH_IP_IPSEC_SPI);
2395                 break;
2396
2397         case MLX5E_TT_IPV6_IPSEC_ESP:
2398                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2399                     MLX5_L3_PROT_TYPE_IPV6);
2400                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2401                     MLX5_HASH_IP_IPSEC_SPI);
2402                 break;
2403
2404         case MLX5E_TT_IPV4:
2405                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2406                     MLX5_L3_PROT_TYPE_IPV4);
2407                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2408                     MLX5_HASH_IP);
2409                 break;
2410
2411         case MLX5E_TT_IPV6:
2412                 MLX5_SET(rx_hash_field_select, hfso, l3_prot_type,
2413                     MLX5_L3_PROT_TYPE_IPV6);
2414                 MLX5_SET(rx_hash_field_select, hfso, selected_fields,
2415                     MLX5_HASH_IP);
2416                 break;
2417
2418         default:
2419                 break;
2420         }
2421 }
2422
2423 static int
2424 mlx5e_open_tir(struct mlx5e_priv *priv, int tt)
2425 {
2426         struct mlx5_core_dev *mdev = priv->mdev;
2427         u32 *in;
2428         void *tirc;
2429         int inlen;
2430         int err;
2431
2432         inlen = MLX5_ST_SZ_BYTES(create_tir_in);
2433         in = mlx5_vzalloc(inlen);
2434         if (in == NULL)
2435                 return (-ENOMEM);
2436         tirc = MLX5_ADDR_OF(create_tir_in, in, tir_context);
2437
2438         mlx5e_build_tir_ctx(priv, tirc, tt);
2439
2440         err = mlx5_core_create_tir(mdev, in, inlen, &priv->tirn[tt]);
2441
2442         kvfree(in);
2443
2444         return (err);
2445 }
2446
2447 static void
2448 mlx5e_close_tir(struct mlx5e_priv *priv, int tt)
2449 {
2450         mlx5_core_destroy_tir(priv->mdev, priv->tirn[tt]);
2451 }
2452
2453 static int
2454 mlx5e_open_tirs(struct mlx5e_priv *priv)
2455 {
2456         int err;
2457         int i;
2458
2459         for (i = 0; i < MLX5E_NUM_TT; i++) {
2460                 err = mlx5e_open_tir(priv, i);
2461                 if (err)
2462                         goto err_close_tirs;
2463         }
2464
2465         return (0);
2466
2467 err_close_tirs:
2468         for (i--; i >= 0; i--)
2469                 mlx5e_close_tir(priv, i);
2470
2471         return (err);
2472 }
2473
2474 static void
2475 mlx5e_close_tirs(struct mlx5e_priv *priv)
2476 {
2477         int i;
2478
2479         for (i = 0; i < MLX5E_NUM_TT; i++)
2480                 mlx5e_close_tir(priv, i);
2481 }
2482
2483 /*
2484  * SW MTU does not include headers,
2485  * HW MTU includes all headers and checksums.
2486  */
2487 static int
2488 mlx5e_set_dev_port_mtu(struct ifnet *ifp, int sw_mtu)
2489 {
2490         struct mlx5e_priv *priv = ifp->if_softc;
2491         struct mlx5_core_dev *mdev = priv->mdev;
2492         int hw_mtu;
2493         int err;
2494
2495         hw_mtu = MLX5E_SW2HW_MTU(sw_mtu);
2496
2497         err = mlx5_set_port_mtu(mdev, hw_mtu);
2498         if (err) {
2499                 if_printf(ifp, "%s: mlx5_set_port_mtu failed setting %d, err=%d\n",
2500                     __func__, sw_mtu, err);
2501                 return (err);
2502         }
2503
2504         /* Update vport context MTU */
2505         err = mlx5_set_vport_mtu(mdev, hw_mtu);
2506         if (err) {
2507                 if_printf(ifp, "%s: Failed updating vport context with MTU size, err=%d\n",
2508                     __func__, err);
2509         }
2510
2511         ifp->if_mtu = sw_mtu;
2512
2513         err = mlx5_query_vport_mtu(mdev, &hw_mtu);
2514         if (err || !hw_mtu) {
2515                 /* fallback to port oper mtu */
2516                 err = mlx5_query_port_oper_mtu(mdev, &hw_mtu);
2517         }
2518         if (err) {
2519                 if_printf(ifp, "Query port MTU, after setting new "
2520                     "MTU value, failed\n");
2521                 return (err);
2522         } else if (MLX5E_HW2SW_MTU(hw_mtu) < sw_mtu) {
2523                 err = -E2BIG,
2524                 if_printf(ifp, "Port MTU %d is smaller than "
2525                     "ifp mtu %d\n", hw_mtu, sw_mtu);
2526         } else if (MLX5E_HW2SW_MTU(hw_mtu) > sw_mtu) {
2527                 err = -EINVAL;
2528                 if_printf(ifp, "Port MTU %d is bigger than "
2529                     "ifp mtu %d\n", hw_mtu, sw_mtu);
2530         }
2531         priv->params_ethtool.hw_mtu = hw_mtu;
2532
2533         return (err);
2534 }
2535
2536 int
2537 mlx5e_open_locked(struct ifnet *ifp)
2538 {
2539         struct mlx5e_priv *priv = ifp->if_softc;
2540         int err;
2541         u16 set_id;
2542
2543         /* check if already opened */
2544         if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0)
2545                 return (0);
2546
2547 #ifdef RSS
2548         if (rss_getnumbuckets() > priv->params.num_channels) {
2549                 if_printf(ifp, "NOTE: There are more RSS buckets(%u) than "
2550                     "channels(%u) available\n", rss_getnumbuckets(),
2551                     priv->params.num_channels);
2552         }
2553 #endif
2554         err = mlx5e_open_tises(priv);
2555         if (err) {
2556                 if_printf(ifp, "%s: mlx5e_open_tises failed, %d\n",
2557                     __func__, err);
2558                 return (err);
2559         }
2560         err = mlx5_vport_alloc_q_counter(priv->mdev,
2561             MLX5_INTERFACE_PROTOCOL_ETH, &set_id);
2562         if (err) {
2563                 if_printf(priv->ifp,
2564                     "%s: mlx5_vport_alloc_q_counter failed: %d\n",
2565                     __func__, err);
2566                 goto err_close_tises;
2567         }
2568         /* store counter set ID */
2569         priv->counter_set_id = set_id;
2570
2571         err = mlx5e_open_channels(priv);
2572         if (err) {
2573                 if_printf(ifp, "%s: mlx5e_open_channels failed, %d\n",
2574                     __func__, err);
2575                 goto err_dalloc_q_counter;
2576         }
2577         err = mlx5e_open_rqt(priv);
2578         if (err) {
2579                 if_printf(ifp, "%s: mlx5e_open_rqt failed, %d\n",
2580                     __func__, err);
2581                 goto err_close_channels;
2582         }
2583         err = mlx5e_open_tirs(priv);
2584         if (err) {
2585                 if_printf(ifp, "%s: mlx5e_open_tir failed, %d\n",
2586                     __func__, err);
2587                 goto err_close_rqls;
2588         }
2589         err = mlx5e_open_flow_table(priv);
2590         if (err) {
2591                 if_printf(ifp, "%s: mlx5e_open_flow_table failed, %d\n",
2592                     __func__, err);
2593                 goto err_close_tirs;
2594         }
2595         err = mlx5e_add_all_vlan_rules(priv);
2596         if (err) {
2597                 if_printf(ifp, "%s: mlx5e_add_all_vlan_rules failed, %d\n",
2598                     __func__, err);
2599                 goto err_close_flow_table;
2600         }
2601         set_bit(MLX5E_STATE_OPENED, &priv->state);
2602
2603         mlx5e_update_carrier(priv);
2604         mlx5e_set_rx_mode_core(priv);
2605
2606         return (0);
2607
2608 err_close_flow_table:
2609         mlx5e_close_flow_table(priv);
2610
2611 err_close_tirs:
2612         mlx5e_close_tirs(priv);
2613
2614 err_close_rqls:
2615         mlx5e_close_rqt(priv);
2616
2617 err_close_channels:
2618         mlx5e_close_channels(priv);
2619
2620 err_dalloc_q_counter:
2621         mlx5_vport_dealloc_q_counter(priv->mdev,
2622             MLX5_INTERFACE_PROTOCOL_ETH, priv->counter_set_id);
2623
2624 err_close_tises:
2625         mlx5e_close_tises(priv);
2626
2627         return (err);
2628 }
2629
2630 static void
2631 mlx5e_open(void *arg)
2632 {
2633         struct mlx5e_priv *priv = arg;
2634
2635         PRIV_LOCK(priv);
2636         if (mlx5_set_port_status(priv->mdev, MLX5_PORT_UP))
2637                 if_printf(priv->ifp,
2638                     "%s: Setting port status to up failed\n",
2639                     __func__);
2640
2641         mlx5e_open_locked(priv->ifp);
2642         priv->ifp->if_drv_flags |= IFF_DRV_RUNNING;
2643         PRIV_UNLOCK(priv);
2644 }
2645
2646 int
2647 mlx5e_close_locked(struct ifnet *ifp)
2648 {
2649         struct mlx5e_priv *priv = ifp->if_softc;
2650
2651         /* check if already closed */
2652         if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
2653                 return (0);
2654
2655         clear_bit(MLX5E_STATE_OPENED, &priv->state);
2656
2657         mlx5e_set_rx_mode_core(priv);
2658         mlx5e_del_all_vlan_rules(priv);
2659         if_link_state_change(priv->ifp, LINK_STATE_DOWN);
2660         mlx5e_close_flow_table(priv);
2661         mlx5e_close_tirs(priv);
2662         mlx5e_close_rqt(priv);
2663         mlx5e_close_channels(priv);
2664         mlx5_vport_dealloc_q_counter(priv->mdev,
2665             MLX5_INTERFACE_PROTOCOL_ETH, priv->counter_set_id);
2666         mlx5e_close_tises(priv);
2667
2668         return (0);
2669 }
2670
2671 #if (__FreeBSD_version >= 1100000)
2672 static uint64_t
2673 mlx5e_get_counter(struct ifnet *ifp, ift_counter cnt)
2674 {
2675         struct mlx5e_priv *priv = ifp->if_softc;
2676         u64 retval;
2677
2678         /* PRIV_LOCK(priv); XXX not allowed */
2679         switch (cnt) {
2680         case IFCOUNTER_IPACKETS:
2681                 retval = priv->stats.vport.rx_packets;
2682                 break;
2683         case IFCOUNTER_IERRORS:
2684                 retval = priv->stats.vport.rx_error_packets +
2685                     priv->stats.pport.alignment_err +
2686                     priv->stats.pport.check_seq_err +
2687                     priv->stats.pport.crc_align_errors +
2688                     priv->stats.pport.in_range_len_errors +
2689                     priv->stats.pport.jabbers +
2690                     priv->stats.pport.out_of_range_len +
2691                     priv->stats.pport.oversize_pkts +
2692                     priv->stats.pport.symbol_err +
2693                     priv->stats.pport.too_long_errors +
2694                     priv->stats.pport.undersize_pkts +
2695                     priv->stats.pport.unsupported_op_rx;
2696                 break;
2697         case IFCOUNTER_IQDROPS:
2698                 retval = priv->stats.vport.rx_out_of_buffer +
2699                     priv->stats.pport.drop_events;
2700                 break;
2701         case IFCOUNTER_OPACKETS:
2702                 retval = priv->stats.vport.tx_packets;
2703                 break;
2704         case IFCOUNTER_OERRORS:
2705                 retval = priv->stats.vport.tx_error_packets;
2706                 break;
2707         case IFCOUNTER_IBYTES:
2708                 retval = priv->stats.vport.rx_bytes;
2709                 break;
2710         case IFCOUNTER_OBYTES:
2711                 retval = priv->stats.vport.tx_bytes;
2712                 break;
2713         case IFCOUNTER_IMCASTS:
2714                 retval = priv->stats.vport.rx_multicast_packets;
2715                 break;
2716         case IFCOUNTER_OMCASTS:
2717                 retval = priv->stats.vport.tx_multicast_packets;
2718                 break;
2719         case IFCOUNTER_OQDROPS:
2720                 retval = priv->stats.vport.tx_queue_dropped;
2721                 break;
2722         case IFCOUNTER_COLLISIONS:
2723                 retval = priv->stats.pport.collisions;
2724                 break;
2725         default:
2726                 retval = if_get_counter_default(ifp, cnt);
2727                 break;
2728         }
2729         /* PRIV_UNLOCK(priv); XXX not allowed */
2730         return (retval);
2731 }
2732 #endif
2733
2734 static void
2735 mlx5e_set_rx_mode(struct ifnet *ifp)
2736 {
2737         struct mlx5e_priv *priv = ifp->if_softc;
2738
2739         queue_work(priv->wq, &priv->set_rx_mode_work);
2740 }
2741
2742 static int
2743 mlx5e_ioctl(struct ifnet *ifp, u_long command, caddr_t data)
2744 {
2745         struct mlx5e_priv *priv;
2746         struct ifreq *ifr;
2747         struct ifi2creq i2c;
2748         int error = 0;
2749         int mask = 0;
2750         int size_read = 0;
2751         int module_status;
2752         int module_num;
2753         int max_mtu;
2754         uint8_t read_addr;
2755
2756         priv = ifp->if_softc;
2757
2758         /* check if detaching */
2759         if (priv == NULL || priv->gone != 0)
2760                 return (ENXIO);
2761
2762         switch (command) {
2763         case SIOCSIFMTU:
2764                 ifr = (struct ifreq *)data;
2765
2766                 PRIV_LOCK(priv);
2767                 mlx5_query_port_max_mtu(priv->mdev, &max_mtu);
2768
2769                 if (ifr->ifr_mtu >= MLX5E_MTU_MIN &&
2770                     ifr->ifr_mtu <= MIN(MLX5E_MTU_MAX, max_mtu)) {
2771                         int was_opened;
2772
2773                         was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2774                         if (was_opened)
2775                                 mlx5e_close_locked(ifp);
2776
2777                         /* set new MTU */
2778                         mlx5e_set_dev_port_mtu(ifp, ifr->ifr_mtu);
2779
2780                         if (was_opened)
2781                                 mlx5e_open_locked(ifp);
2782                 } else {
2783                         error = EINVAL;
2784                         if_printf(ifp, "Invalid MTU value. Min val: %d, Max val: %d\n",
2785                             MLX5E_MTU_MIN, MIN(MLX5E_MTU_MAX, max_mtu));
2786                 }
2787                 PRIV_UNLOCK(priv);
2788                 break;
2789         case SIOCSIFFLAGS:
2790                 if ((ifp->if_flags & IFF_UP) &&
2791                     (ifp->if_drv_flags & IFF_DRV_RUNNING)) {
2792                         mlx5e_set_rx_mode(ifp);
2793                         break;
2794                 }
2795                 PRIV_LOCK(priv);
2796                 if (ifp->if_flags & IFF_UP) {
2797                         if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) {
2798                                 if (test_bit(MLX5E_STATE_OPENED, &priv->state) == 0)
2799                                         mlx5e_open_locked(ifp);
2800                                 ifp->if_drv_flags |= IFF_DRV_RUNNING;
2801                                 mlx5_set_port_status(priv->mdev, MLX5_PORT_UP);
2802                         }
2803                 } else {
2804                         if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
2805                                 mlx5_set_port_status(priv->mdev,
2806                                     MLX5_PORT_DOWN);
2807                                 if (test_bit(MLX5E_STATE_OPENED, &priv->state) != 0)
2808                                         mlx5e_close_locked(ifp);
2809                                 mlx5e_update_carrier(priv);
2810                                 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
2811                         }
2812                 }
2813                 PRIV_UNLOCK(priv);
2814                 break;
2815         case SIOCADDMULTI:
2816         case SIOCDELMULTI:
2817                 mlx5e_set_rx_mode(ifp);
2818                 break;
2819         case SIOCSIFMEDIA:
2820         case SIOCGIFMEDIA:
2821         case SIOCGIFXMEDIA:
2822                 ifr = (struct ifreq *)data;
2823                 error = ifmedia_ioctl(ifp, ifr, &priv->media, command);
2824                 break;
2825         case SIOCSIFCAP:
2826                 ifr = (struct ifreq *)data;
2827                 PRIV_LOCK(priv);
2828                 mask = ifr->ifr_reqcap ^ ifp->if_capenable;
2829
2830                 if (mask & IFCAP_TXCSUM) {
2831                         ifp->if_capenable ^= IFCAP_TXCSUM;
2832                         ifp->if_hwassist ^= (CSUM_TCP | CSUM_UDP | CSUM_IP);
2833
2834                         if (IFCAP_TSO4 & ifp->if_capenable &&
2835                             !(IFCAP_TXCSUM & ifp->if_capenable)) {
2836                                 ifp->if_capenable &= ~IFCAP_TSO4;
2837                                 ifp->if_hwassist &= ~CSUM_IP_TSO;
2838                                 if_printf(ifp,
2839                                     "tso4 disabled due to -txcsum.\n");
2840                         }
2841                 }
2842                 if (mask & IFCAP_TXCSUM_IPV6) {
2843                         ifp->if_capenable ^= IFCAP_TXCSUM_IPV6;
2844                         ifp->if_hwassist ^= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6);
2845
2846                         if (IFCAP_TSO6 & ifp->if_capenable &&
2847                             !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) {
2848                                 ifp->if_capenable &= ~IFCAP_TSO6;
2849                                 ifp->if_hwassist &= ~CSUM_IP6_TSO;
2850                                 if_printf(ifp,
2851                                     "tso6 disabled due to -txcsum6.\n");
2852                         }
2853                 }
2854                 if (mask & IFCAP_RXCSUM)
2855                         ifp->if_capenable ^= IFCAP_RXCSUM;
2856                 if (mask & IFCAP_RXCSUM_IPV6)
2857                         ifp->if_capenable ^= IFCAP_RXCSUM_IPV6;
2858                 if (mask & IFCAP_TSO4) {
2859                         if (!(IFCAP_TSO4 & ifp->if_capenable) &&
2860                             !(IFCAP_TXCSUM & ifp->if_capenable)) {
2861                                 if_printf(ifp, "enable txcsum first.\n");
2862                                 error = EAGAIN;
2863                                 goto out;
2864                         }
2865                         ifp->if_capenable ^= IFCAP_TSO4;
2866                         ifp->if_hwassist ^= CSUM_IP_TSO;
2867                 }
2868                 if (mask & IFCAP_TSO6) {
2869                         if (!(IFCAP_TSO6 & ifp->if_capenable) &&
2870                             !(IFCAP_TXCSUM_IPV6 & ifp->if_capenable)) {
2871                                 if_printf(ifp, "enable txcsum6 first.\n");
2872                                 error = EAGAIN;
2873                                 goto out;
2874                         }
2875                         ifp->if_capenable ^= IFCAP_TSO6;
2876                         ifp->if_hwassist ^= CSUM_IP6_TSO;
2877                 }
2878                 if (mask & IFCAP_VLAN_HWFILTER) {
2879                         if (ifp->if_capenable & IFCAP_VLAN_HWFILTER)
2880                                 mlx5e_disable_vlan_filter(priv);
2881                         else
2882                                 mlx5e_enable_vlan_filter(priv);
2883
2884                         ifp->if_capenable ^= IFCAP_VLAN_HWFILTER;
2885                 }
2886                 if (mask & IFCAP_VLAN_HWTAGGING)
2887                         ifp->if_capenable ^= IFCAP_VLAN_HWTAGGING;
2888                 if (mask & IFCAP_WOL_MAGIC)
2889                         ifp->if_capenable ^= IFCAP_WOL_MAGIC;
2890
2891                 VLAN_CAPABILITIES(ifp);
2892                 /* turn off LRO means also turn of HW LRO - if it's on */
2893                 if (mask & IFCAP_LRO) {
2894                         int was_opened = test_bit(MLX5E_STATE_OPENED, &priv->state);
2895                         bool need_restart = false;
2896
2897                         ifp->if_capenable ^= IFCAP_LRO;
2898                         if (!(ifp->if_capenable & IFCAP_LRO)) {
2899                                 if (priv->params.hw_lro_en) {
2900                                         priv->params.hw_lro_en = false;
2901                                         need_restart = true;
2902                                         /* Not sure this is the correct way */
2903                                         priv->params_ethtool.hw_lro = priv->params.hw_lro_en;
2904                                 }
2905                         }
2906                         if (was_opened && need_restart) {
2907                                 mlx5e_close_locked(ifp);
2908                                 mlx5e_open_locked(ifp);
2909                         }
2910                 }
2911                 if (mask & IFCAP_HWRXTSTMP) {
2912                         ifp->if_capenable ^= IFCAP_HWRXTSTMP;
2913                         if (ifp->if_capenable & IFCAP_HWRXTSTMP) {
2914                                 if (priv->clbr_done == 0)
2915                                         mlx5e_reset_calibration_callout(priv);
2916                         } else {
2917                                 callout_drain(&priv->tstmp_clbr);
2918                                 priv->clbr_done = 0;
2919                         }
2920                 }
2921 out:
2922                 PRIV_UNLOCK(priv);
2923                 break;
2924
2925         case SIOCGI2C:
2926                 ifr = (struct ifreq *)data;
2927
2928                 /*
2929                  * Copy from the user-space address ifr_data to the
2930                  * kernel-space address i2c
2931                  */
2932                 error = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c));
2933                 if (error)
2934                         break;
2935
2936                 if (i2c.len > sizeof(i2c.data)) {
2937                         error = EINVAL;
2938                         break;
2939                 }
2940
2941                 PRIV_LOCK(priv);
2942                 /* Get module_num which is required for the query_eeprom */
2943                 error = mlx5_query_module_num(priv->mdev, &module_num);
2944                 if (error) {
2945                         if_printf(ifp, "Query module num failed, eeprom "
2946                             "reading is not supported\n");
2947                         error = EINVAL;
2948                         goto err_i2c;
2949                 }
2950                 /* Check if module is present before doing an access */
2951                 module_status = mlx5_query_module_status(priv->mdev, module_num);
2952                 if (module_status != MLX5_MODULE_STATUS_PLUGGED_ENABLED &&
2953                     module_status != MLX5_MODULE_STATUS_PLUGGED_DISABLED) {
2954                         error = EINVAL;
2955                         goto err_i2c;
2956                 }
2957                 /*
2958                  * Currently 0XA0 and 0xA2 are the only addresses permitted.
2959                  * The internal conversion is as follows:
2960                  */
2961                 if (i2c.dev_addr == 0xA0)
2962                         read_addr = MLX5E_I2C_ADDR_LOW;
2963                 else if (i2c.dev_addr == 0xA2)
2964                         read_addr = MLX5E_I2C_ADDR_HIGH;
2965                 else {
2966                         if_printf(ifp, "Query eeprom failed, "
2967                             "Invalid Address: %X\n", i2c.dev_addr);
2968                         error = EINVAL;
2969                         goto err_i2c;
2970                 }
2971                 error = mlx5_query_eeprom(priv->mdev,
2972                     read_addr, MLX5E_EEPROM_LOW_PAGE,
2973                     (uint32_t)i2c.offset, (uint32_t)i2c.len, module_num,
2974                     (uint32_t *)i2c.data, &size_read);
2975                 if (error) {
2976                         if_printf(ifp, "Query eeprom failed, eeprom "
2977                             "reading is not supported\n");
2978                         error = EINVAL;
2979                         goto err_i2c;
2980                 }
2981
2982                 if (i2c.len > MLX5_EEPROM_MAX_BYTES) {
2983                         error = mlx5_query_eeprom(priv->mdev,
2984                             read_addr, MLX5E_EEPROM_LOW_PAGE,
2985                             (uint32_t)(i2c.offset + size_read),
2986                             (uint32_t)(i2c.len - size_read), module_num,
2987                             (uint32_t *)(i2c.data + size_read), &size_read);
2988                 }
2989                 if (error) {
2990                         if_printf(ifp, "Query eeprom failed, eeprom "
2991                             "reading is not supported\n");
2992                         error = EINVAL;
2993                         goto err_i2c;
2994                 }
2995
2996                 error = copyout(&i2c, ifr_data_get_ptr(ifr), sizeof(i2c));
2997 err_i2c:
2998                 PRIV_UNLOCK(priv);
2999                 break;
3000
3001         default:
3002                 error = ether_ioctl(ifp, command, data);
3003                 break;
3004         }
3005         return (error);
3006 }
3007
3008 static int
3009 mlx5e_check_required_hca_cap(struct mlx5_core_dev *mdev)
3010 {
3011         /*
3012          * TODO: uncoment once FW really sets all these bits if
3013          * (!mdev->caps.eth.rss_ind_tbl_cap || !mdev->caps.eth.csum_cap ||
3014          * !mdev->caps.eth.max_lso_cap || !mdev->caps.eth.vlan_cap ||
3015          * !(mdev->caps.gen.flags & MLX5_DEV_CAP_FLAG_SCQE_BRK_MOD)) return
3016          * -ENOTSUPP;
3017          */
3018
3019         /* TODO: add more must-to-have features */
3020
3021         if (MLX5_CAP_GEN(mdev, port_type) != MLX5_CAP_PORT_TYPE_ETH)
3022                 return (-ENODEV);
3023
3024         return (0);
3025 }
3026
3027 static u16
3028 mlx5e_get_max_inline_cap(struct mlx5_core_dev *mdev)
3029 {
3030         int bf_buf_size = (1 << MLX5_CAP_GEN(mdev, log_bf_reg_size)) / 2;
3031
3032         return bf_buf_size -
3033                sizeof(struct mlx5e_tx_wqe) +
3034                2 /*sizeof(mlx5e_tx_wqe.inline_hdr_start)*/;
3035 }
3036
3037 static void
3038 mlx5e_build_ifp_priv(struct mlx5_core_dev *mdev,
3039     struct mlx5e_priv *priv,
3040     int num_comp_vectors)
3041 {
3042         /*
3043          * TODO: Consider link speed for setting "log_sq_size",
3044          * "log_rq_size" and "cq_moderation_xxx":
3045          */
3046         priv->params.log_sq_size =
3047             MLX5E_PARAMS_DEFAULT_LOG_SQ_SIZE;
3048         priv->params.log_rq_size =
3049             MLX5E_PARAMS_DEFAULT_LOG_RQ_SIZE;
3050         priv->params.rx_cq_moderation_usec =
3051             MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ?
3052             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC_FROM_CQE :
3053             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_USEC;
3054         priv->params.rx_cq_moderation_mode =
3055             MLX5_CAP_GEN(mdev, cq_period_start_from_cqe) ? 1 : 0;
3056         priv->params.rx_cq_moderation_pkts =
3057             MLX5E_PARAMS_DEFAULT_RX_CQ_MODERATION_PKTS;
3058         priv->params.tx_cq_moderation_usec =
3059             MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_USEC;
3060         priv->params.tx_cq_moderation_pkts =
3061             MLX5E_PARAMS_DEFAULT_TX_CQ_MODERATION_PKTS;
3062         priv->params.min_rx_wqes =
3063             MLX5E_PARAMS_DEFAULT_MIN_RX_WQES;
3064         priv->params.rx_hash_log_tbl_sz =
3065             (order_base_2(num_comp_vectors) >
3066             MLX5E_PARAMS_DEFAULT_RX_HASH_LOG_TBL_SZ) ?
3067             order_base_2(num_comp_vectors) :
3068             MLX5E_PARAMS_DEFAULT_RX_HASH_LOG_TBL_SZ;
3069         priv->params.num_tc = 1;
3070         priv->params.default_vlan_prio = 0;
3071         priv->counter_set_id = -1;
3072         priv->params.tx_max_inline = mlx5e_get_max_inline_cap(mdev);
3073         mlx5_query_min_inline(mdev, &priv->params.tx_min_inline_mode);
3074
3075         /*
3076          * hw lro is currently defaulted to off. when it won't anymore we
3077          * will consider the HW capability: "!!MLX5_CAP_ETH(mdev, lro_cap)"
3078          */
3079         priv->params.hw_lro_en = false;
3080         priv->params.lro_wqe_sz = MLX5E_PARAMS_DEFAULT_LRO_WQE_SZ;
3081
3082         priv->params.cqe_zipping_en = !!MLX5_CAP_GEN(mdev, cqe_compression);
3083
3084         priv->mdev = mdev;
3085         priv->params.num_channels = num_comp_vectors;
3086         priv->order_base_2_num_channels = order_base_2(num_comp_vectors);
3087         priv->queue_mapping_channel_mask =
3088             roundup_pow_of_two(num_comp_vectors) - 1;
3089         priv->num_tc = priv->params.num_tc;
3090         priv->default_vlan_prio = priv->params.default_vlan_prio;
3091
3092         INIT_WORK(&priv->update_stats_work, mlx5e_update_stats_work);
3093         INIT_WORK(&priv->update_carrier_work, mlx5e_update_carrier_work);
3094         INIT_WORK(&priv->set_rx_mode_work, mlx5e_set_rx_mode_work);
3095 }
3096
3097 static int
3098 mlx5e_create_mkey(struct mlx5e_priv *priv, u32 pdn,
3099                   struct mlx5_core_mr *mkey)
3100 {
3101         struct ifnet *ifp = priv->ifp;
3102         struct mlx5_core_dev *mdev = priv->mdev;
3103         int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
3104         void *mkc;
3105         u32 *in;
3106         int err;
3107
3108         in = mlx5_vzalloc(inlen);
3109         if (in == NULL) {
3110                 if_printf(ifp, "%s: failed to allocate inbox\n", __func__);
3111                 return (-ENOMEM);
3112         }
3113
3114         mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
3115         MLX5_SET(mkc, mkc, access_mode, MLX5_ACCESS_MODE_PA);
3116         MLX5_SET(mkc, mkc, lw, 1);
3117         MLX5_SET(mkc, mkc, lr, 1);
3118
3119         MLX5_SET(mkc, mkc, pd, pdn);
3120         MLX5_SET(mkc, mkc, length64, 1);
3121         MLX5_SET(mkc, mkc, qpn, 0xffffff);
3122
3123         err = mlx5_core_create_mkey(mdev, mkey, in, inlen);
3124         if (err)
3125                 if_printf(ifp, "%s: mlx5_core_create_mkey failed, %d\n",
3126                     __func__, err);
3127
3128         kvfree(in);
3129         return (err);
3130 }
3131
3132 static const char *mlx5e_vport_stats_desc[] = {
3133         MLX5E_VPORT_STATS(MLX5E_STATS_DESC)
3134 };
3135
3136 static const char *mlx5e_pport_stats_desc[] = {
3137         MLX5E_PPORT_STATS(MLX5E_STATS_DESC)
3138 };
3139
3140 static void
3141 mlx5e_priv_mtx_init(struct mlx5e_priv *priv)
3142 {
3143         mtx_init(&priv->async_events_mtx, "mlx5async", MTX_NETWORK_LOCK, MTX_DEF);
3144         sx_init(&priv->state_lock, "mlx5state");
3145         callout_init_mtx(&priv->watchdog, &priv->async_events_mtx, 0);
3146         MLX5_INIT_DOORBELL_LOCK(&priv->doorbell_lock);
3147 }
3148
3149 static void
3150 mlx5e_priv_mtx_destroy(struct mlx5e_priv *priv)
3151 {
3152         mtx_destroy(&priv->async_events_mtx);
3153         sx_destroy(&priv->state_lock);
3154 }
3155
3156 static int
3157 sysctl_firmware(SYSCTL_HANDLER_ARGS)
3158 {
3159         /*
3160          * %d.%d%.d the string format.
3161          * fw_rev_{maj,min,sub} return u16, 2^16 = 65536.
3162          * We need at most 5 chars to store that.
3163          * It also has: two "." and NULL at the end, which means we need 18
3164          * (5*3 + 3) chars at most.
3165          */
3166         char fw[18];
3167         struct mlx5e_priv *priv = arg1;
3168         int error;
3169
3170         snprintf(fw, sizeof(fw), "%d.%d.%d", fw_rev_maj(priv->mdev), fw_rev_min(priv->mdev),
3171             fw_rev_sub(priv->mdev));
3172         error = sysctl_handle_string(oidp, fw, sizeof(fw), req);
3173         return (error);
3174 }
3175
3176 static void
3177 mlx5e_disable_tx_dma(struct mlx5e_channel *ch)
3178 {
3179         int i;
3180
3181         for (i = 0; i < ch->num_tc; i++)
3182                 mlx5e_drain_sq(&ch->sq[i]);
3183 }
3184
3185 static void
3186 mlx5e_reset_sq_doorbell_record(struct mlx5e_sq *sq)
3187 {
3188
3189         sq->doorbell.d32[0] = cpu_to_be32(MLX5_OPCODE_NOP);
3190         sq->doorbell.d32[1] = cpu_to_be32(sq->sqn << 8);
3191         mlx5e_tx_notify_hw(sq, sq->doorbell.d32, 0);
3192         sq->doorbell.d64 = 0;
3193 }
3194
3195 void
3196 mlx5e_resume_sq(struct mlx5e_sq *sq)
3197 {
3198         int err;
3199
3200         /* check if already enabled */
3201         if (sq->stopped == 0)
3202                 return;
3203
3204         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_ERR,
3205             MLX5_SQC_STATE_RST);
3206         if (err != 0) {
3207                 if_printf(sq->ifp,
3208                     "mlx5e_modify_sq() from ERR to RST failed: %d\n", err);
3209         }
3210
3211         sq->cc = 0;
3212         sq->pc = 0;
3213
3214         /* reset doorbell prior to moving from RST to RDY */
3215         mlx5e_reset_sq_doorbell_record(sq);
3216
3217         err = mlx5e_modify_sq(sq, MLX5_SQC_STATE_RST,
3218             MLX5_SQC_STATE_RDY);
3219         if (err != 0) {
3220                 if_printf(sq->ifp,
3221                     "mlx5e_modify_sq() from RST to RDY failed: %d\n", err);
3222         }
3223
3224         mtx_lock(&sq->lock);
3225         sq->cev_next_state = MLX5E_CEV_STATE_INITIAL;
3226         sq->stopped = 0;
3227         mtx_unlock(&sq->lock);
3228
3229 }
3230
3231 static void
3232 mlx5e_enable_tx_dma(struct mlx5e_channel *ch)
3233 {
3234         int i;
3235
3236         for (i = 0; i < ch->num_tc; i++)
3237                 mlx5e_resume_sq(&ch->sq[i]);
3238 }
3239
3240 static void
3241 mlx5e_disable_rx_dma(struct mlx5e_channel *ch)
3242 {
3243         struct mlx5e_rq *rq = &ch->rq;
3244         int err;
3245
3246         mtx_lock(&rq->mtx);
3247         rq->enabled = 0;
3248         callout_stop(&rq->watchdog);
3249         mtx_unlock(&rq->mtx);
3250
3251         callout_drain(&rq->watchdog);
3252
3253         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RDY, MLX5_RQC_STATE_ERR);
3254         if (err != 0) {
3255                 if_printf(rq->ifp,
3256                     "mlx5e_modify_rq() from RDY to RST failed: %d\n", err);
3257         }
3258
3259         while (!mlx5_wq_ll_is_empty(&rq->wq)) {
3260                 msleep(1);
3261                 rq->cq.mcq.comp(&rq->cq.mcq);
3262         }
3263
3264         /*
3265          * Transitioning into RST state will allow the FW to track less ERR state queues,
3266          * thus reducing the recv queue flushing time
3267          */
3268         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_ERR, MLX5_RQC_STATE_RST);
3269         if (err != 0) {
3270                 if_printf(rq->ifp,
3271                     "mlx5e_modify_rq() from ERR to RST failed: %d\n", err);
3272         }
3273 }
3274
3275 static void
3276 mlx5e_enable_rx_dma(struct mlx5e_channel *ch)
3277 {
3278         struct mlx5e_rq *rq = &ch->rq;
3279         int err;
3280
3281         rq->wq.wqe_ctr = 0;
3282         mlx5_wq_ll_update_db_record(&rq->wq);
3283         err = mlx5e_modify_rq(rq, MLX5_RQC_STATE_RST, MLX5_RQC_STATE_RDY);
3284         if (err != 0) {
3285                 if_printf(rq->ifp,
3286                     "mlx5e_modify_rq() from RST to RDY failed: %d\n", err);
3287         }
3288
3289         rq->enabled = 1;
3290
3291         rq->cq.mcq.comp(&rq->cq.mcq);
3292 }
3293
3294 void
3295 mlx5e_modify_tx_dma(struct mlx5e_priv *priv, uint8_t value)
3296 {
3297         int i;
3298
3299         if (priv->channel == NULL)
3300                 return;
3301
3302         for (i = 0; i < priv->params.num_channels; i++) {
3303
3304                 if (!priv->channel[i])
3305                         continue;
3306
3307                 if (value)
3308                         mlx5e_disable_tx_dma(priv->channel[i]);
3309                 else
3310                         mlx5e_enable_tx_dma(priv->channel[i]);
3311         }
3312 }
3313
3314 void
3315 mlx5e_modify_rx_dma(struct mlx5e_priv *priv, uint8_t value)
3316 {
3317         int i;
3318
3319         if (priv->channel == NULL)
3320                 return;
3321
3322         for (i = 0; i < priv->params.num_channels; i++) {
3323
3324                 if (!priv->channel[i])
3325                         continue;
3326
3327                 if (value)
3328                         mlx5e_disable_rx_dma(priv->channel[i]);
3329                 else
3330                         mlx5e_enable_rx_dma(priv->channel[i]);
3331         }
3332 }
3333
3334 u8
3335 mlx5e_params_calculate_tx_min_inline(struct mlx5_core_dev *mdev)
3336 {
3337         u8 min_inline_mode;
3338
3339         min_inline_mode = MLX5_INLINE_MODE_L2;
3340         mlx5_query_min_inline(mdev, &min_inline_mode);
3341         if (min_inline_mode == MLX5_INLINE_MODE_NONE &&
3342             !MLX5_CAP_ETH(mdev, wqe_vlan_insert))
3343                 min_inline_mode = MLX5_INLINE_MODE_L2;
3344
3345         return (min_inline_mode);
3346 }
3347
3348 static void
3349 mlx5e_add_hw_stats(struct mlx5e_priv *priv)
3350 {
3351         SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_hw),
3352             OID_AUTO, "fw_version", CTLTYPE_STRING | CTLFLAG_RD, priv, 0,
3353             sysctl_firmware, "A", "HCA firmware version");
3354
3355         SYSCTL_ADD_STRING(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_hw),
3356             OID_AUTO, "board_id", CTLFLAG_RD, priv->mdev->board_id, 0,
3357             "Board ID");
3358 }
3359
3360 static int
3361 mlx5e_sysctl_tx_priority_flow_control(SYSCTL_HANDLER_ARGS)
3362 {
3363         struct mlx5e_priv *priv = arg1;
3364         uint32_t tx_pfc;
3365         uint32_t value;
3366         int error;
3367
3368         PRIV_LOCK(priv);
3369
3370         tx_pfc = priv->params.tx_priority_flow_control;
3371
3372         /* get current value */
3373         value = (tx_pfc >> arg2) & 1;
3374
3375         error = sysctl_handle_32(oidp, &value, 0, req);
3376
3377         /* range check value */
3378         if (value != 0)
3379                 priv->params.tx_priority_flow_control |= (1 << arg2);
3380         else
3381                 priv->params.tx_priority_flow_control &= ~(1 << arg2);
3382
3383         /* check if update is required */
3384         if (error == 0 && priv->gone == 0 &&
3385             tx_pfc != priv->params.tx_priority_flow_control) {
3386                 error = -mlx5e_set_port_pfc(priv);
3387                 /* restore previous value */
3388                 if (error != 0)
3389                         priv->params.tx_priority_flow_control= tx_pfc;
3390         }
3391         PRIV_UNLOCK(priv);
3392
3393         return (error);
3394 }
3395
3396 static int
3397 mlx5e_sysctl_rx_priority_flow_control(SYSCTL_HANDLER_ARGS)
3398 {
3399         struct mlx5e_priv *priv = arg1;
3400         uint32_t rx_pfc;
3401         uint32_t value;
3402         int error;
3403
3404         PRIV_LOCK(priv);
3405
3406         rx_pfc = priv->params.rx_priority_flow_control;
3407
3408         /* get current value */
3409         value = (rx_pfc >> arg2) & 1;
3410
3411         error = sysctl_handle_32(oidp, &value, 0, req);
3412
3413         /* range check value */
3414         if (value != 0)
3415                 priv->params.rx_priority_flow_control |= (1 << arg2);
3416         else
3417                 priv->params.rx_priority_flow_control &= ~(1 << arg2);
3418
3419         /* check if update is required */
3420         if (error == 0 && priv->gone == 0 &&
3421             rx_pfc != priv->params.rx_priority_flow_control) {
3422                 error = -mlx5e_set_port_pfc(priv);
3423                 /* restore previous value */
3424                 if (error != 0)
3425                         priv->params.rx_priority_flow_control= rx_pfc;
3426         }
3427         PRIV_UNLOCK(priv);
3428
3429         return (error);
3430 }
3431
3432 static void
3433 mlx5e_setup_pauseframes(struct mlx5e_priv *priv)
3434 {
3435         unsigned int x;
3436         char path[96];
3437         int error;
3438
3439         /* enable pauseframes by default */
3440         priv->params.tx_pauseframe_control = 1;
3441         priv->params.rx_pauseframe_control = 1;
3442
3443         /* disable ports flow control, PFC, by default */
3444         priv->params.tx_priority_flow_control = 0;
3445         priv->params.rx_priority_flow_control = 0;
3446
3447 #if (__FreeBSD_version < 1100000)
3448         /* compute path for sysctl */
3449         snprintf(path, sizeof(path), "dev.mce.%d.tx_pauseframe_control",
3450             device_get_unit(priv->mdev->pdev->dev.bsddev));
3451
3452         /* try to fetch tunable, if any */
3453         TUNABLE_INT_FETCH(path, &priv->params.tx_pauseframe_control);
3454
3455         /* compute path for sysctl */
3456         snprintf(path, sizeof(path), "dev.mce.%d.rx_pauseframe_control",
3457             device_get_unit(priv->mdev->pdev->dev.bsddev));
3458
3459         /* try to fetch tunable, if any */
3460         TUNABLE_INT_FETCH(path, &priv->params.rx_pauseframe_control);
3461
3462         for (x = 0; x != 8; x++) {
3463
3464                 /* compute path for sysctl */
3465                 snprintf(path, sizeof(path), "dev.mce.%d.tx_priority_flow_control_%u",
3466                     device_get_unit(priv->mdev->pdev->dev.bsddev), x);
3467
3468                 /* try to fetch tunable, if any */
3469                 if (TUNABLE_INT_FETCH(path, &value) == 0 && value != 0)
3470                         priv->params.tx_priority_flow_control |= 1 << x;
3471
3472                 /* compute path for sysctl */
3473                 snprintf(path, sizeof(path), "dev.mce.%d.rx_priority_flow_control_%u",
3474                     device_get_unit(priv->mdev->pdev->dev.bsddev), x);
3475
3476                 /* try to fetch tunable, if any */
3477                 if (TUNABLE_INT_FETCH(path, &value) == 0 && value != 0)
3478                         priv->params.rx_priority_flow_control |= 1 << x;
3479         }
3480 #endif
3481
3482         /* register pauseframe SYSCTLs */
3483         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3484             OID_AUTO, "tx_pauseframe_control", CTLFLAG_RDTUN,
3485             &priv->params.tx_pauseframe_control, 0,
3486             "Set to enable TX pause frames. Clear to disable.");
3487
3488         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3489             OID_AUTO, "rx_pauseframe_control", CTLFLAG_RDTUN,
3490             &priv->params.rx_pauseframe_control, 0,
3491             "Set to enable RX pause frames. Clear to disable.");
3492
3493         /* register priority_flow control, PFC, SYSCTLs */
3494         for (x = 0; x != 8; x++) {
3495                 snprintf(path, sizeof(path), "tx_priority_flow_control_%u", x);
3496
3497                 SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3498                     OID_AUTO, path, CTLTYPE_UINT | CTLFLAG_RWTUN |
3499                     CTLFLAG_MPSAFE, priv, x, &mlx5e_sysctl_tx_priority_flow_control, "IU",
3500                     "Set to enable TX ports flow control frames for given priority. Clear to disable.");
3501
3502                 snprintf(path, sizeof(path), "rx_priority_flow_control_%u", x);
3503
3504                 SYSCTL_ADD_PROC(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3505                     OID_AUTO, path, CTLTYPE_UINT | CTLFLAG_RWTUN |
3506                     CTLFLAG_MPSAFE, priv, x, &mlx5e_sysctl_rx_priority_flow_control, "IU",
3507                     "Set to enable RX ports flow control frames for given priority. Clear to disable.");
3508         }
3509
3510         PRIV_LOCK(priv);
3511
3512         /* range check */
3513         priv->params.tx_pauseframe_control =
3514             priv->params.tx_pauseframe_control ? 1 : 0;
3515         priv->params.rx_pauseframe_control =
3516             priv->params.rx_pauseframe_control ? 1 : 0;
3517
3518         /* update firmware */
3519         error = mlx5e_set_port_pause_and_pfc(priv);
3520         if (error == -EINVAL) {
3521                 if_printf(priv->ifp,
3522                     "Global pauseframes must be disabled before enabling PFC.\n");
3523                 priv->params.rx_priority_flow_control = 0;
3524                 priv->params.tx_priority_flow_control = 0;
3525
3526                 /* update firmware */
3527                 (void) mlx5e_set_port_pause_and_pfc(priv);
3528         }
3529         PRIV_UNLOCK(priv);
3530 }
3531
3532 static void *
3533 mlx5e_create_ifp(struct mlx5_core_dev *mdev)
3534 {
3535         struct ifnet *ifp;
3536         struct mlx5e_priv *priv;
3537         u8 dev_addr[ETHER_ADDR_LEN] __aligned(4);
3538         struct sysctl_oid_list *child;
3539         int ncv = mdev->priv.eq_table.num_comp_vectors;
3540         char unit[16];
3541         int err;
3542         int i;
3543         u32 eth_proto_cap;
3544
3545         if (mlx5e_check_required_hca_cap(mdev)) {
3546                 mlx5_core_dbg(mdev, "mlx5e_check_required_hca_cap() failed\n");
3547                 return (NULL);
3548         }
3549         priv = malloc(sizeof(*priv), M_MLX5EN, M_WAITOK | M_ZERO);
3550         mlx5e_priv_mtx_init(priv);
3551
3552         ifp = priv->ifp = if_alloc(IFT_ETHER);
3553         if (ifp == NULL) {
3554                 mlx5_core_err(mdev, "if_alloc() failed\n");
3555                 goto err_free_priv;
3556         }
3557         ifp->if_softc = priv;
3558         if_initname(ifp, "mce", device_get_unit(mdev->pdev->dev.bsddev));
3559         ifp->if_mtu = ETHERMTU;
3560         ifp->if_init = mlx5e_open;
3561         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
3562         ifp->if_ioctl = mlx5e_ioctl;
3563         ifp->if_transmit = mlx5e_xmit;
3564         ifp->if_qflush = if_qflush;
3565 #if (__FreeBSD_version >= 1100000)
3566         ifp->if_get_counter = mlx5e_get_counter;
3567 #endif
3568         ifp->if_snd.ifq_maxlen = ifqmaxlen;
3569         /*
3570          * Set driver features
3571          */
3572         ifp->if_capabilities |= IFCAP_HWCSUM | IFCAP_HWCSUM_IPV6;
3573         ifp->if_capabilities |= IFCAP_VLAN_MTU | IFCAP_VLAN_HWTAGGING;
3574         ifp->if_capabilities |= IFCAP_VLAN_HWCSUM | IFCAP_VLAN_HWFILTER;
3575         ifp->if_capabilities |= IFCAP_LINKSTATE | IFCAP_JUMBO_MTU;
3576         ifp->if_capabilities |= IFCAP_LRO;
3577         ifp->if_capabilities |= IFCAP_TSO | IFCAP_VLAN_HWTSO;
3578         ifp->if_capabilities |= IFCAP_HWSTATS | IFCAP_HWRXTSTMP;
3579 #ifdef RATELIMIT
3580         ifp->if_capabilities |= IFCAP_TXRTLMT;
3581         ifp->if_snd_tag_alloc = mlx5e_rl_snd_tag_alloc;
3582         ifp->if_snd_tag_free = mlx5e_rl_snd_tag_free;
3583         ifp->if_snd_tag_modify = mlx5e_rl_snd_tag_modify;
3584         ifp->if_snd_tag_query = mlx5e_rl_snd_tag_query;
3585 #endif
3586
3587         /* set TSO limits so that we don't have to drop TX packets */
3588         ifp->if_hw_tsomax = MLX5E_MAX_TX_PAYLOAD_SIZE - (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN);
3589         ifp->if_hw_tsomaxsegcount = MLX5E_MAX_TX_MBUF_FRAGS - 1 /* hdr */;
3590         ifp->if_hw_tsomaxsegsize = MLX5E_MAX_TX_MBUF_SIZE;
3591
3592         ifp->if_capenable = ifp->if_capabilities;
3593         ifp->if_hwassist = 0;
3594         if (ifp->if_capenable & IFCAP_TSO)
3595                 ifp->if_hwassist |= CSUM_TSO;
3596         if (ifp->if_capenable & IFCAP_TXCSUM)
3597                 ifp->if_hwassist |= (CSUM_TCP | CSUM_UDP | CSUM_IP);
3598         if (ifp->if_capenable & IFCAP_TXCSUM_IPV6)
3599                 ifp->if_hwassist |= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6);
3600
3601         sysctl_ctx_init(&priv->sysctl_ctx_channel_debug);
3602
3603         /* ifnet sysctl tree */
3604         sysctl_ctx_init(&priv->sysctl_ctx);
3605         priv->sysctl_ifnet = SYSCTL_ADD_NODE(&priv->sysctl_ctx, SYSCTL_STATIC_CHILDREN(_dev),
3606             OID_AUTO, ifp->if_dname, CTLFLAG_RD, 0, "MLX5 ethernet - interface name");
3607         if (priv->sysctl_ifnet == NULL) {
3608                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
3609                 goto err_free_sysctl;
3610         }
3611         snprintf(unit, sizeof(unit), "%d", ifp->if_dunit);
3612         priv->sysctl_ifnet = SYSCTL_ADD_NODE(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3613             OID_AUTO, unit, CTLFLAG_RD, 0, "MLX5 ethernet - interface unit");
3614         if (priv->sysctl_ifnet == NULL) {
3615                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
3616                 goto err_free_sysctl;
3617         }
3618
3619         /* HW sysctl tree */
3620         child = SYSCTL_CHILDREN(device_get_sysctl_tree(mdev->pdev->dev.bsddev));
3621         priv->sysctl_hw = SYSCTL_ADD_NODE(&priv->sysctl_ctx, child,
3622             OID_AUTO, "hw", CTLFLAG_RD, 0, "MLX5 ethernet dev hw");
3623         if (priv->sysctl_hw == NULL) {
3624                 mlx5_core_err(mdev, "SYSCTL_ADD_NODE() failed\n");
3625                 goto err_free_sysctl;
3626         }
3627         mlx5e_build_ifp_priv(mdev, priv, ncv);
3628
3629         snprintf(unit, sizeof(unit), "mce%u_wq",
3630             device_get_unit(mdev->pdev->dev.bsddev));
3631         priv->wq = alloc_workqueue(unit, 0, 1);
3632         if (priv->wq == NULL) {
3633                 if_printf(ifp, "%s: alloc_workqueue failed\n", __func__);
3634                 goto err_free_sysctl;
3635         }
3636
3637         err = mlx5_alloc_map_uar(mdev, &priv->cq_uar);
3638         if (err) {
3639                 if_printf(ifp, "%s: mlx5_alloc_map_uar failed, %d\n",
3640                     __func__, err);
3641                 goto err_free_wq;
3642         }
3643         err = mlx5_core_alloc_pd(mdev, &priv->pdn);
3644         if (err) {
3645                 if_printf(ifp, "%s: mlx5_core_alloc_pd failed, %d\n",
3646                     __func__, err);
3647                 goto err_unmap_free_uar;
3648         }
3649         err = mlx5_alloc_transport_domain(mdev, &priv->tdn);
3650         if (err) {
3651                 if_printf(ifp, "%s: mlx5_alloc_transport_domain failed, %d\n",
3652                     __func__, err);
3653                 goto err_dealloc_pd;
3654         }
3655         err = mlx5e_create_mkey(priv, priv->pdn, &priv->mr);
3656         if (err) {
3657                 if_printf(ifp, "%s: mlx5e_create_mkey failed, %d\n",
3658                     __func__, err);
3659                 goto err_dealloc_transport_domain;
3660         }
3661         mlx5_query_nic_vport_mac_address(priv->mdev, 0, dev_addr);
3662
3663         /* check if we should generate a random MAC address */
3664         if (MLX5_CAP_GEN(priv->mdev, vport_group_manager) == 0 &&
3665             is_zero_ether_addr(dev_addr)) {
3666                 random_ether_addr(dev_addr);
3667                 if_printf(ifp, "Assigned random MAC address\n");
3668         }
3669 #ifdef RATELIMIT
3670         err = mlx5e_rl_init(priv);
3671         if (err) {
3672                 if_printf(ifp, "%s: mlx5e_rl_init failed, %d\n",
3673                     __func__, err);
3674                 goto err_create_mkey;
3675         }
3676 #endif
3677
3678         /* set default MTU */
3679         mlx5e_set_dev_port_mtu(ifp, ifp->if_mtu);
3680
3681         /* Set desc */
3682         device_set_desc(mdev->pdev->dev.bsddev, mlx5e_version);
3683
3684         /* Set default media status */
3685         priv->media_status_last = IFM_AVALID;
3686         priv->media_active_last = IFM_ETHER | IFM_AUTO |
3687             IFM_ETH_RXPAUSE | IFM_FDX;
3688
3689         /* setup default pauseframes configuration */
3690         mlx5e_setup_pauseframes(priv);
3691
3692         err = mlx5_query_port_proto_cap(mdev, &eth_proto_cap, MLX5_PTYS_EN);
3693         if (err) {
3694                 eth_proto_cap = 0;
3695                 if_printf(ifp, "%s: Query port media capability failed, %d\n",
3696                     __func__, err);
3697         }
3698
3699         /* Setup supported medias */
3700         ifmedia_init(&priv->media, IFM_IMASK | IFM_ETH_FMASK,
3701             mlx5e_media_change, mlx5e_media_status);
3702
3703         for (i = 0; i < MLX5E_LINK_MODES_NUMBER; ++i) {
3704                 if (mlx5e_mode_table[i].baudrate == 0)
3705                         continue;
3706                 if (MLX5E_PROT_MASK(i) & eth_proto_cap) {
3707                         ifmedia_add(&priv->media,
3708                             mlx5e_mode_table[i].subtype |
3709                             IFM_ETHER, 0, NULL);
3710                         ifmedia_add(&priv->media,
3711                             mlx5e_mode_table[i].subtype |
3712                             IFM_ETHER | IFM_FDX |
3713                             IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE, 0, NULL);
3714                 }
3715         }
3716
3717         ifmedia_add(&priv->media, IFM_ETHER | IFM_AUTO, 0, NULL);
3718         ifmedia_add(&priv->media, IFM_ETHER | IFM_AUTO | IFM_FDX |
3719             IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE, 0, NULL);
3720
3721         /* Set autoselect by default */
3722         ifmedia_set(&priv->media, IFM_ETHER | IFM_AUTO | IFM_FDX |
3723             IFM_ETH_RXPAUSE | IFM_ETH_TXPAUSE);
3724         ether_ifattach(ifp, dev_addr);
3725
3726         /* Register for VLAN events */
3727         priv->vlan_attach = EVENTHANDLER_REGISTER(vlan_config,
3728             mlx5e_vlan_rx_add_vid, priv, EVENTHANDLER_PRI_FIRST);
3729         priv->vlan_detach = EVENTHANDLER_REGISTER(vlan_unconfig,
3730             mlx5e_vlan_rx_kill_vid, priv, EVENTHANDLER_PRI_FIRST);
3731
3732         /* Link is down by default */
3733         if_link_state_change(ifp, LINK_STATE_DOWN);
3734
3735         mlx5e_enable_async_events(priv);
3736
3737         mlx5e_add_hw_stats(priv);
3738
3739         mlx5e_create_stats(&priv->stats.vport.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3740             "vstats", mlx5e_vport_stats_desc, MLX5E_VPORT_STATS_NUM,
3741             priv->stats.vport.arg);
3742
3743         mlx5e_create_stats(&priv->stats.pport.ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3744             "pstats", mlx5e_pport_stats_desc, MLX5E_PPORT_STATS_NUM,
3745             priv->stats.pport.arg);
3746
3747         mlx5e_create_ethtool(priv);
3748
3749         mtx_lock(&priv->async_events_mtx);
3750         mlx5e_update_stats(priv);
3751         mtx_unlock(&priv->async_events_mtx);
3752
3753         SYSCTL_ADD_INT(&priv->sysctl_ctx, SYSCTL_CHILDREN(priv->sysctl_ifnet),
3754             OID_AUTO, "rx_clbr_done", CTLFLAG_RD,
3755             &priv->clbr_done, 0,
3756             "RX timestamps calibration state");
3757         callout_init(&priv->tstmp_clbr, CALLOUT_DIRECT);
3758         mlx5e_reset_calibration_callout(priv);
3759
3760         return (priv);
3761
3762 #ifdef RATELIMIT
3763 err_create_mkey:
3764         mlx5_core_destroy_mkey(priv->mdev, &priv->mr);
3765 #endif
3766 err_dealloc_transport_domain:
3767         mlx5_dealloc_transport_domain(mdev, priv->tdn);
3768
3769 err_dealloc_pd:
3770         mlx5_core_dealloc_pd(mdev, priv->pdn);
3771
3772 err_unmap_free_uar:
3773         mlx5_unmap_free_uar(mdev, &priv->cq_uar);
3774
3775 err_free_wq:
3776         destroy_workqueue(priv->wq);
3777
3778 err_free_sysctl:
3779         sysctl_ctx_free(&priv->sysctl_ctx);
3780         sysctl_ctx_free(&priv->sysctl_ctx_channel_debug);
3781
3782         if_free(ifp);
3783
3784 err_free_priv:
3785         mlx5e_priv_mtx_destroy(priv);
3786         free(priv, M_MLX5EN);
3787         return (NULL);
3788 }
3789
3790 static void
3791 mlx5e_destroy_ifp(struct mlx5_core_dev *mdev, void *vpriv)
3792 {
3793         struct mlx5e_priv *priv = vpriv;
3794         struct ifnet *ifp = priv->ifp;
3795
3796         /* don't allow more IOCTLs */
3797         priv->gone = 1;
3798
3799         /*
3800          * Clear the device description to avoid use after free,
3801          * because the bsddev is not destroyed when this module is
3802          * unloaded:
3803          */
3804         device_set_desc(mdev->pdev->dev.bsddev, NULL);
3805
3806         /* XXX wait a bit to allow IOCTL handlers to complete */
3807         pause("W", hz);
3808
3809 #ifdef RATELIMIT
3810         /*
3811          * The kernel can have reference(s) via the m_snd_tag's into
3812          * the ratelimit channels, and these must go away before
3813          * detaching:
3814          */
3815         while (READ_ONCE(priv->rl.stats.tx_active_connections) != 0) {
3816                 if_printf(priv->ifp, "Waiting for all ratelimit connections "
3817                     "to terminate\n");
3818                 pause("W", hz);
3819         }
3820 #endif
3821         /* stop watchdog timer */
3822         callout_drain(&priv->watchdog);
3823
3824         callout_drain(&priv->tstmp_clbr);
3825
3826         if (priv->vlan_attach != NULL)
3827                 EVENTHANDLER_DEREGISTER(vlan_config, priv->vlan_attach);
3828         if (priv->vlan_detach != NULL)
3829                 EVENTHANDLER_DEREGISTER(vlan_unconfig, priv->vlan_detach);
3830
3831         /* make sure device gets closed */
3832         PRIV_LOCK(priv);
3833         mlx5e_close_locked(ifp);
3834         PRIV_UNLOCK(priv);
3835
3836         /* unregister device */
3837         ifmedia_removeall(&priv->media);
3838         ether_ifdetach(ifp);
3839         if_free(ifp);
3840
3841 #ifdef RATELIMIT
3842         mlx5e_rl_cleanup(priv);
3843 #endif
3844         /* destroy all remaining sysctl nodes */
3845         if (priv->sysctl_debug) {
3846                 sysctl_ctx_free(&priv->sysctl_ctx_channel_debug);
3847                 sysctl_ctx_free(&priv->stats.port_stats_debug.ctx);
3848         }
3849         sysctl_ctx_free(&priv->stats.vport.ctx);
3850         sysctl_ctx_free(&priv->stats.pport.ctx);
3851         sysctl_ctx_free(&priv->sysctl_ctx);
3852
3853         mlx5_core_destroy_mkey(priv->mdev, &priv->mr);
3854         mlx5_dealloc_transport_domain(priv->mdev, priv->tdn);
3855         mlx5_core_dealloc_pd(priv->mdev, priv->pdn);
3856         mlx5_unmap_free_uar(priv->mdev, &priv->cq_uar);
3857         mlx5e_disable_async_events(priv);
3858         destroy_workqueue(priv->wq);
3859         mlx5e_priv_mtx_destroy(priv);
3860         free(priv, M_MLX5EN);
3861 }
3862
3863 static void *
3864 mlx5e_get_ifp(void *vpriv)
3865 {
3866         struct mlx5e_priv *priv = vpriv;
3867
3868         return (priv->ifp);
3869 }
3870
3871 static struct mlx5_interface mlx5e_interface = {
3872         .add = mlx5e_create_ifp,
3873         .remove = mlx5e_destroy_ifp,
3874         .event = mlx5e_async_event,
3875         .protocol = MLX5_INTERFACE_PROTOCOL_ETH,
3876         .get_dev = mlx5e_get_ifp,
3877 };
3878
3879 void
3880 mlx5e_init(void)
3881 {
3882         mlx5_register_interface(&mlx5e_interface);
3883 }
3884
3885 void
3886 mlx5e_cleanup(void)
3887 {
3888         mlx5_unregister_interface(&mlx5e_interface);
3889 }
3890
3891 module_init_order(mlx5e_init, SI_ORDER_THIRD);
3892 module_exit_order(mlx5e_cleanup, SI_ORDER_THIRD);
3893
3894 #if (__FreeBSD_version >= 1100000)
3895 MODULE_DEPEND(mlx5en, linuxkpi, 1, 1, 1);
3896 #endif
3897 MODULE_DEPEND(mlx5en, mlx5, 1, 1, 1);
3898 MODULE_VERSION(mlx5en, 1);