sys/dev/sfxge/sfxge.c

   1 /*-
   2  * Copyright (c) 2010-2016 Solarflare Communications Inc.
   3  * All rights reserved.
   4  *
   5  * This software was developed in part by Philip Paeps under contract for
   6  * Solarflare Communications, Inc.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright notice,
  12  *    this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright notice,
  14  *    this list of conditions and the following disclaimer in the documentation
  15  *    and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  18  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  19  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  20  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
  21  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  22  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  23  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
  24  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  25  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  26  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
  27  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  28  *
  29  * The views and conclusions contained in the software and documentation are
  30  * those of the authors and should not be interpreted as representing official
  31  * policies, either expressed or implied, of the FreeBSD Project.
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_rss.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/kernel.h>
  41 #include <sys/bus.h>
  42 #include <sys/rman.h>
  43 #include <sys/lock.h>
  44 #include <sys/module.h>
  45 #include <sys/mutex.h>
  46 #include <sys/smp.h>
  47 #include <sys/socket.h>
  48 #include <sys/taskqueue.h>
  49 #include <sys/sockio.h>
  50 #include <sys/sysctl.h>
  51 #include <sys/priv.h>
  52 #include <sys/syslog.h>
  53
  54 #include <dev/pci/pcireg.h>
  55 #include <dev/pci/pcivar.h>
  56
  57 #include <net/ethernet.h>
  58 #include <net/if.h>
  59 #include <net/if_var.h>
  60 #include <net/if_media.h>
  61 #include <net/if_types.h>
  62
  63 #ifdef RSS
  64 #include <net/rss_config.h>
  65 #endif
  66
  67 #include "common/efx.h"
  68
  69 #include "sfxge.h"
  70 #include "sfxge_rx.h"
  71 #include "sfxge_ioc.h"
  72 #include "sfxge_version.h"
  73
  74 #define SFXGE_CAP (IFCAP_VLAN_MTU | IFCAP_VLAN_HWCSUM |                 \
  75                    IFCAP_RXCSUM | IFCAP_TXCSUM |                        \
  76                    IFCAP_RXCSUM_IPV6 | IFCAP_TXCSUM_IPV6 |              \
  77                    IFCAP_TSO4 | IFCAP_TSO6 |                            \
  78                    IFCAP_JUMBO_MTU |                                    \
  79                    IFCAP_VLAN_HWTSO | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  80 #define SFXGE_CAP_ENABLE SFXGE_CAP
  81 #define SFXGE_CAP_FIXED (IFCAP_VLAN_MTU |                               \
  82                          IFCAP_JUMBO_MTU | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  83
  84 MALLOC_DEFINE(M_SFXGE, "sfxge", "Solarflare 10GigE driver");
  85
  86
  87 SYSCTL_NODE(_hw, OID_AUTO, sfxge, CTLFLAG_RD, 0,
  88             "SFXGE driver parameters");
  89
  90 #define SFXGE_PARAM_RX_RING     SFXGE_PARAM(rx_ring)
  91 static int sfxge_rx_ring_entries = SFXGE_NDESCS;
  92 TUNABLE_INT(SFXGE_PARAM_RX_RING, &sfxge_rx_ring_entries);
  93 SYSCTL_INT(_hw_sfxge, OID_AUTO, rx_ring, CTLFLAG_RDTUN,
  94            &sfxge_rx_ring_entries, 0,
  95            "Maximum number of descriptors in a receive ring");
  96
  97 #define SFXGE_PARAM_TX_RING     SFXGE_PARAM(tx_ring)
  98 static int sfxge_tx_ring_entries = SFXGE_NDESCS;
  99 TUNABLE_INT(SFXGE_PARAM_TX_RING, &sfxge_tx_ring_entries);
 100 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_ring, CTLFLAG_RDTUN,
 101            &sfxge_tx_ring_entries, 0,
 102            "Maximum number of descriptors in a transmit ring");
 103
 104 #define SFXGE_PARAM_RESTART_ATTEMPTS    SFXGE_PARAM(restart_attempts)
 105 static int sfxge_restart_attempts = 3;
 106 TUNABLE_INT(SFXGE_PARAM_RESTART_ATTEMPTS, &sfxge_restart_attempts);
 107 SYSCTL_INT(_hw_sfxge, OID_AUTO, restart_attempts, CTLFLAG_RDTUN,
 108            &sfxge_restart_attempts, 0,
 109            "Maximum number of attempts to bring interface up after reset");
 110
 111 #if EFSYS_OPT_MCDI_LOGGING
 112 #define SFXGE_PARAM_MCDI_LOGGING        SFXGE_PARAM(mcdi_logging)
 113 static int sfxge_mcdi_logging = 0;
 114 TUNABLE_INT(SFXGE_PARAM_MCDI_LOGGING, &sfxge_mcdi_logging);
 115 #endif
 116
 117 static void
 118 sfxge_reset(void *arg, int npending);
 119
 120 static int
 121 sfxge_estimate_rsrc_limits(struct sfxge_softc *sc)
 122 {
 123         efx_drv_limits_t limits;
 124         int rc;
 125         unsigned int evq_max;
 126         uint32_t evq_allocated;
 127         uint32_t rxq_allocated;
 128         uint32_t txq_allocated;
 129
 130         /*
 131          * Limit the number of event queues to:
 132          *  - number of CPUs
 133          *  - hardwire maximum RSS channels
 134          *  - administratively specified maximum RSS channels
 135          */
 136 #ifdef RSS
 137         /*
 138          * Avoid extra limitations so that the number of queues
 139          * may be configured at administrator's will
 140          */
 141         evq_max = MIN(MAX(rss_getnumbuckets(), 1), EFX_MAXRSS);
 142 #else
 143         evq_max = MIN(mp_ncpus, EFX_MAXRSS);
 144 #endif
 145         if (sc->max_rss_channels > 0)
 146                 evq_max = MIN(evq_max, sc->max_rss_channels);
 147
 148         memset(&limits, 0, sizeof(limits));
 149
 150         limits.edl_min_evq_count = 1;
 151         limits.edl_max_evq_count = evq_max;
 152         limits.edl_min_txq_count = SFXGE_TXQ_NTYPES;
 153         limits.edl_max_txq_count = evq_max + SFXGE_TXQ_NTYPES - 1;
 154         limits.edl_min_rxq_count = 1;
 155         limits.edl_max_rxq_count = evq_max;
 156
 157         efx_nic_set_drv_limits(sc->enp, &limits);
 158
 159         if ((rc = efx_nic_init(sc->enp)) != 0)
 160                 return (rc);
 161
 162         rc = efx_nic_get_vi_pool(sc->enp, &evq_allocated, &rxq_allocated,
 163                                  &txq_allocated);
 164         if (rc != 0) {
 165                 efx_nic_fini(sc->enp);
 166                 return (rc);
 167         }
 168
 169         KASSERT(txq_allocated >= SFXGE_TXQ_NTYPES,
 170                 ("txq_allocated < SFXGE_TXQ_NTYPES"));
 171
 172         sc->evq_max = MIN(evq_allocated, evq_max);
 173         sc->evq_max = MIN(rxq_allocated, sc->evq_max);
 174         sc->evq_max = MIN(txq_allocated - (SFXGE_TXQ_NTYPES - 1),
 175                           sc->evq_max);
 176
 177         KASSERT(sc->evq_max <= evq_max,
 178                 ("allocated more than maximum requested"));
 179
 180 #ifdef RSS
 181         if (sc->evq_max < rss_getnumbuckets())
 182                 device_printf(sc->dev, "The number of allocated queues (%u) "
 183                               "is less than the number of RSS buckets (%u); "
 184                               "performance degradation might be observed",
 185                               sc->evq_max, rss_getnumbuckets());
 186 #endif
 187
 188         /*
 189          * NIC is kept initialized in the case of success to be able to
 190          * initialize port to find out media types.
 191          */
 192         return (0);
 193 }
 194
 195 static int
 196 sfxge_set_drv_limits(struct sfxge_softc *sc)
 197 {
 198         efx_drv_limits_t limits;
 199
 200         memset(&limits, 0, sizeof(limits));
 201
 202         /* Limits are strict since take into account initial estimation */
 203         limits.edl_min_evq_count = limits.edl_max_evq_count =
 204             sc->intr.n_alloc;
 205         limits.edl_min_txq_count = limits.edl_max_txq_count =
 206             sc->intr.n_alloc + SFXGE_TXQ_NTYPES - 1;
 207         limits.edl_min_rxq_count = limits.edl_max_rxq_count =
 208             sc->intr.n_alloc;
 209
 210         return (efx_nic_set_drv_limits(sc->enp, &limits));
 211 }
 212
 213 static int
 214 sfxge_start(struct sfxge_softc *sc)
 215 {
 216         int rc;
 217
 218         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 219
 220         if (sc->init_state == SFXGE_STARTED)
 221                 return (0);
 222
 223         if (sc->init_state != SFXGE_REGISTERED) {
 224                 rc = EINVAL;
 225                 goto fail;
 226         }
 227
 228         /* Set required resource limits */
 229         if ((rc = sfxge_set_drv_limits(sc)) != 0)
 230                 goto fail;
 231
 232         if ((rc = efx_nic_init(sc->enp)) != 0)
 233                 goto fail;
 234
 235         /* Start processing interrupts. */
 236         if ((rc = sfxge_intr_start(sc)) != 0)
 237                 goto fail2;
 238
 239         /* Start processing events. */
 240         if ((rc = sfxge_ev_start(sc)) != 0)
 241                 goto fail3;
 242
 243         /* Fire up the port. */
 244         if ((rc = sfxge_port_start(sc)) != 0)
 245                 goto fail4;
 246
 247         /* Start the receiver side. */
 248         if ((rc = sfxge_rx_start(sc)) != 0)
 249                 goto fail5;
 250
 251         /* Start the transmitter side. */
 252         if ((rc = sfxge_tx_start(sc)) != 0)
 253                 goto fail6;
 254
 255         sc->init_state = SFXGE_STARTED;
 256
 257         /* Tell the stack we're running. */
 258         sc->ifnet->if_drv_flags |= IFF_DRV_RUNNING;
 259         sc->ifnet->if_drv_flags &= ~IFF_DRV_OACTIVE;
 260
 261         return (0);
 262
 263 fail6:
 264         sfxge_rx_stop(sc);
 265
 266 fail5:
 267         sfxge_port_stop(sc);
 268
 269 fail4:
 270         sfxge_ev_stop(sc);
 271
 272 fail3:
 273         sfxge_intr_stop(sc);
 274
 275 fail2:
 276         efx_nic_fini(sc->enp);
 277
 278 fail:
 279         device_printf(sc->dev, "sfxge_start: %d\n", rc);
 280
 281         return (rc);
 282 }
 283
 284 static void
 285 sfxge_if_init(void *arg)
 286 {
 287         struct sfxge_softc *sc;
 288
 289         sc = (struct sfxge_softc *)arg;
 290
 291         SFXGE_ADAPTER_LOCK(sc);
 292         (void)sfxge_start(sc);
 293         SFXGE_ADAPTER_UNLOCK(sc);
 294 }
 295
 296 static void
 297 sfxge_stop(struct sfxge_softc *sc)
 298 {
 299         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 300
 301         if (sc->init_state != SFXGE_STARTED)
 302                 return;
 303
 304         sc->init_state = SFXGE_REGISTERED;
 305
 306         /* Stop the transmitter. */
 307         sfxge_tx_stop(sc);
 308
 309         /* Stop the receiver. */
 310         sfxge_rx_stop(sc);
 311
 312         /* Stop the port. */
 313         sfxge_port_stop(sc);
 314
 315         /* Stop processing events. */
 316         sfxge_ev_stop(sc);
 317
 318         /* Stop processing interrupts. */
 319         sfxge_intr_stop(sc);
 320
 321         efx_nic_fini(sc->enp);
 322
 323         sc->ifnet->if_drv_flags &= ~IFF_DRV_RUNNING;
 324 }
 325
 326
 327 static int
 328 sfxge_vpd_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 329 {
 330         efx_vpd_value_t value;
 331         int rc = 0;
 332
 333         switch (ioc->u.vpd.op) {
 334         case SFXGE_VPD_OP_GET_KEYWORD:
 335                 value.evv_tag = ioc->u.vpd.tag;
 336                 value.evv_keyword = ioc->u.vpd.keyword;
 337                 rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 338                 if (rc != 0)
 339                         break;
 340                 ioc->u.vpd.len = MIN(ioc->u.vpd.len, value.evv_length);
 341                 if (ioc->u.vpd.payload != 0) {
 342                         rc = copyout(value.evv_value, ioc->u.vpd.payload,
 343                                      ioc->u.vpd.len);
 344                 }
 345                 break;
 346         case SFXGE_VPD_OP_SET_KEYWORD:
 347                 if (ioc->u.vpd.len > sizeof(value.evv_value))
 348                         return (EINVAL);
 349                 value.evv_tag = ioc->u.vpd.tag;
 350                 value.evv_keyword = ioc->u.vpd.keyword;
 351                 value.evv_length = ioc->u.vpd.len;
 352                 rc = copyin(ioc->u.vpd.payload, value.evv_value, value.evv_length);
 353                 if (rc != 0)
 354                         break;
 355                 rc = efx_vpd_set(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 356                 if (rc != 0)
 357                         break;
 358                 rc = efx_vpd_verify(sc->enp, sc->vpd_data, sc->vpd_size);
 359                 if (rc != 0)
 360                         break;
 361                 rc = efx_vpd_write(sc->enp, sc->vpd_data, sc->vpd_size);
 362                 break;
 363         default:
 364                 rc = EOPNOTSUPP;
 365                 break;
 366         }
 367
 368         return (rc);
 369 }
 370
 371 static int
 372 sfxge_private_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 373 {
 374         switch (ioc->op) {
 375         case SFXGE_MCDI_IOC:
 376                 return (sfxge_mcdi_ioctl(sc, ioc));
 377         case SFXGE_NVRAM_IOC:
 378                 return (sfxge_nvram_ioctl(sc, ioc));
 379         case SFXGE_VPD_IOC:
 380                 return (sfxge_vpd_ioctl(sc, ioc));
 381         default:
 382                 return (EOPNOTSUPP);
 383         }
 384 }
 385
 386
 387 static int
 388 sfxge_if_ioctl(struct ifnet *ifp, unsigned long command, caddr_t data)
 389 {
 390         struct sfxge_softc *sc;
 391         struct ifreq *ifr;
 392         sfxge_ioc_t ioc;
 393         int error;
 394
 395         ifr = (struct ifreq *)data;
 396         sc = ifp->if_softc;
 397         error = 0;
 398
 399         switch (command) {
 400         case SIOCSIFFLAGS:
 401                 SFXGE_ADAPTER_LOCK(sc);
 402                 if (ifp->if_flags & IFF_UP) {
 403                         if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
 404                                 if ((ifp->if_flags ^ sc->if_flags) &
 405                                     (IFF_PROMISC | IFF_ALLMULTI)) {
 406                                         sfxge_mac_filter_set(sc);
 407                                 }
 408                         } else
 409                                 sfxge_start(sc);
 410                 } else
 411                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 412                                 sfxge_stop(sc);
 413                 sc->if_flags = ifp->if_flags;
 414                 SFXGE_ADAPTER_UNLOCK(sc);
 415                 break;
 416         case SIOCSIFMTU:
 417                 if (ifr->ifr_mtu == ifp->if_mtu) {
 418                         /* Nothing to do */
 419                         error = 0;
 420                 } else if (ifr->ifr_mtu > SFXGE_MAX_MTU) {
 421                         error = EINVAL;
 422                 } else if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
 423                         ifp->if_mtu = ifr->ifr_mtu;
 424                         error = 0;
 425                 } else {
 426                         /* Restart required */
 427                         SFXGE_ADAPTER_LOCK(sc);
 428                         sfxge_stop(sc);
 429                         ifp->if_mtu = ifr->ifr_mtu;
 430                         error = sfxge_start(sc);
 431                         SFXGE_ADAPTER_UNLOCK(sc);
 432                         if (error != 0) {
 433                                 ifp->if_flags &= ~IFF_UP;
 434                                 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
 435                                 if_down(ifp);
 436                         }
 437                 }
 438                 break;
 439         case SIOCADDMULTI:
 440         case SIOCDELMULTI:
 441                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 442                         sfxge_mac_filter_set(sc);
 443                 break;
 444         case SIOCSIFCAP:
 445         {
 446                 int reqcap = ifr->ifr_reqcap;
 447                 int capchg_mask;
 448
 449                 SFXGE_ADAPTER_LOCK(sc);
 450
 451                 /* Capabilities to be changed in accordance with request */
 452                 capchg_mask = ifp->if_capenable ^ reqcap;
 453
 454                 /*
 455                  * The networking core already rejects attempts to
 456                  * enable capabilities we don't have.  We still have
 457                  * to reject attempts to disable capabilities that we
 458                  * can't (yet) disable.
 459                  */
 460                 KASSERT((reqcap & ~ifp->if_capabilities) == 0,
 461                     ("Unsupported capabilities 0x%x requested 0x%x vs "
 462                      "supported 0x%x",
 463                      reqcap & ~ifp->if_capabilities,
 464                      reqcap , ifp->if_capabilities));
 465                 if (capchg_mask & SFXGE_CAP_FIXED) {
 466                         error = EINVAL;
 467                         SFXGE_ADAPTER_UNLOCK(sc);
 468                         break;
 469                 }
 470
 471                 /* Check request before any changes */
 472                 if ((capchg_mask & IFCAP_TSO4) &&
 473                     (reqcap & (IFCAP_TSO4 | IFCAP_TXCSUM)) == IFCAP_TSO4) {
 474                         error = EAGAIN;
 475                         SFXGE_ADAPTER_UNLOCK(sc);
 476                         if_printf(ifp, "enable txcsum before tso4\n");
 477                         break;
 478                 }
 479                 if ((capchg_mask & IFCAP_TSO6) &&
 480                     (reqcap & (IFCAP_TSO6 | IFCAP_TXCSUM_IPV6)) == IFCAP_TSO6) {
 481                         error = EAGAIN;
 482                         SFXGE_ADAPTER_UNLOCK(sc);
 483                         if_printf(ifp, "enable txcsum6 before tso6\n");
 484                         break;
 485                 }
 486
 487                 if (reqcap & IFCAP_TXCSUM) {
 488                         ifp->if_hwassist |= (CSUM_IP | CSUM_TCP | CSUM_UDP);
 489                 } else {
 490                         ifp->if_hwassist &= ~(CSUM_IP | CSUM_TCP | CSUM_UDP);
 491                         if (reqcap & IFCAP_TSO4) {
 492                                 reqcap &= ~IFCAP_TSO4;
 493                                 if_printf(ifp,
 494                                     "tso4 disabled due to -txcsum\n");
 495                         }
 496                 }
 497                 if (reqcap & IFCAP_TXCSUM_IPV6) {
 498                         ifp->if_hwassist |= (CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 499                 } else {
 500                         ifp->if_hwassist &= ~(CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 501                         if (reqcap & IFCAP_TSO6) {
 502                                 reqcap &= ~IFCAP_TSO6;
 503                                 if_printf(ifp,
 504                                     "tso6 disabled due to -txcsum6\n");
 505                         }
 506                 }
 507
 508                 /*
 509                  * The kernel takes both IFCAP_TSOx and CSUM_TSO into
 510                  * account before using TSO. So, we do not touch
 511                  * checksum flags when IFCAP_TSOx is modified.
 512                  * Note that CSUM_TSO is (CSUM_IP_TSO|CSUM_IP6_TSO),
 513                  * but both bits are set in IPv4 and IPv6 mbufs.
 514                  */
 515
 516                 ifp->if_capenable = reqcap;
 517
 518                 SFXGE_ADAPTER_UNLOCK(sc);
 519                 break;
 520         }
 521         case SIOCSIFMEDIA:
 522         case SIOCGIFMEDIA:
 523                 error = ifmedia_ioctl(ifp, ifr, &sc->media, command);
 524                 break;
 525 #ifdef SIOCGI2C
 526         case SIOCGI2C:
 527         {
 528                 struct ifi2creq i2c;
 529
 530                 error = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
 531                 if (error != 0)
 532                         break;
 533
 534                 if (i2c.len > sizeof(i2c.data)) {
 535                         error = EINVAL;
 536                         break;
 537                 }
 538
 539                 SFXGE_ADAPTER_LOCK(sc);
 540                 error = efx_phy_module_get_info(sc->enp, i2c.dev_addr,
 541                                                 i2c.offset, i2c.len,
 542                                                 &i2c.data[0]);
 543                 SFXGE_ADAPTER_UNLOCK(sc);
 544                 if (error == 0)
 545                         error = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
 546                 break;
 547         }
 548 #endif
 549         case SIOCGPRIVATE_0:
 550                 error = priv_check(curthread, PRIV_DRIVER);
 551                 if (error != 0)
 552                         break;
 553                 error = copyin(ifr->ifr_data, &ioc, sizeof(ioc));
 554                 if (error != 0)
 555                         return (error);
 556                 error = sfxge_private_ioctl(sc, &ioc);
 557                 if (error == 0) {
 558                         error = copyout(&ioc, ifr->ifr_data, sizeof(ioc));
 559                 }
 560                 break;
 561         default:
 562                 error = ether_ioctl(ifp, command, data);
 563         }
 564
 565         return (error);
 566 }
 567
 568 static void
 569 sfxge_ifnet_fini(struct ifnet *ifp)
 570 {
 571         struct sfxge_softc *sc = ifp->if_softc;
 572
 573         SFXGE_ADAPTER_LOCK(sc);
 574         sfxge_stop(sc);
 575         SFXGE_ADAPTER_UNLOCK(sc);
 576
 577         ifmedia_removeall(&sc->media);
 578         ether_ifdetach(ifp);
 579         if_free(ifp);
 580 }
 581
 582 static int
 583 sfxge_ifnet_init(struct ifnet *ifp, struct sfxge_softc *sc)
 584 {
 585         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
 586         device_t dev;
 587         int rc;
 588
 589         dev = sc->dev;
 590         sc->ifnet = ifp;
 591
 592         if_initname(ifp, device_get_name(dev), device_get_unit(dev));
 593         ifp->if_init = sfxge_if_init;
 594         ifp->if_softc = sc;
 595         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 596         ifp->if_ioctl = sfxge_if_ioctl;
 597
 598         ifp->if_capabilities = SFXGE_CAP;
 599         ifp->if_capenable = SFXGE_CAP_ENABLE;
 600         ifp->if_hw_tsomax = SFXGE_TSO_MAX_SIZE;
 601         ifp->if_hw_tsomaxsegcount = SFXGE_TX_MAPPING_MAX_SEG;
 602         ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
 603
 604 #ifdef SFXGE_LRO
 605         ifp->if_capabilities |= IFCAP_LRO;
 606         ifp->if_capenable |= IFCAP_LRO;
 607 #endif
 608
 609         if (encp->enc_hw_tx_insert_vlan_enabled) {
 610                 ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING;
 611                 ifp->if_capenable |= IFCAP_VLAN_HWTAGGING;
 612         }
 613         ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO |
 614                            CSUM_TCP_IPV6 | CSUM_UDP_IPV6;
 615
 616         ether_ifattach(ifp, encp->enc_mac_addr);
 617
 618         ifp->if_transmit = sfxge_if_transmit;
 619         ifp->if_qflush = sfxge_if_qflush;
 620
 621         ifp->if_get_counter = sfxge_get_counter;
 622
 623         DBGPRINT(sc->dev, "ifmedia_init");
 624         if ((rc = sfxge_port_ifmedia_init(sc)) != 0)
 625                 goto fail;
 626
 627         return (0);
 628
 629 fail:
 630         ether_ifdetach(sc->ifnet);
 631         return (rc);
 632 }
 633
 634 void
 635 sfxge_sram_buf_tbl_alloc(struct sfxge_softc *sc, size_t n, uint32_t *idp)
 636 {
 637         KASSERT(sc->buffer_table_next + n <=
 638                 efx_nic_cfg_get(sc->enp)->enc_buftbl_limit,
 639                 ("buffer table full"));
 640
 641         *idp = sc->buffer_table_next;
 642         sc->buffer_table_next += n;
 643 }
 644
 645 static int
 646 sfxge_bar_init(struct sfxge_softc *sc)
 647 {
 648         efsys_bar_t *esbp = &sc->bar;
 649
 650         esbp->esb_rid = PCIR_BAR(EFX_MEM_BAR);
 651         if ((esbp->esb_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY,
 652             &esbp->esb_rid, RF_ACTIVE)) == NULL) {
 653                 device_printf(sc->dev, "Cannot allocate BAR region %d\n",
 654                     EFX_MEM_BAR);
 655                 return (ENXIO);
 656         }
 657         esbp->esb_tag = rman_get_bustag(esbp->esb_res);
 658         esbp->esb_handle = rman_get_bushandle(esbp->esb_res);
 659
 660         SFXGE_BAR_LOCK_INIT(esbp, device_get_nameunit(sc->dev));
 661
 662         return (0);
 663 }
 664
 665 static void
 666 sfxge_bar_fini(struct sfxge_softc *sc)
 667 {
 668         efsys_bar_t *esbp = &sc->bar;
 669
 670         bus_release_resource(sc->dev, SYS_RES_MEMORY, esbp->esb_rid,
 671             esbp->esb_res);
 672         SFXGE_BAR_LOCK_DESTROY(esbp);
 673 }
 674
 675 static int
 676 sfxge_create(struct sfxge_softc *sc)
 677 {
 678         device_t dev;
 679         efx_nic_t *enp;
 680         int error;
 681         char rss_param_name[sizeof(SFXGE_PARAM(%d.max_rss_channels))];
 682 #if EFSYS_OPT_MCDI_LOGGING
 683         char mcdi_log_param_name[sizeof(SFXGE_PARAM(%d.mcdi_logging))];
 684 #endif
 685
 686         dev = sc->dev;
 687
 688         SFXGE_ADAPTER_LOCK_INIT(sc, device_get_nameunit(sc->dev));
 689
 690         sc->max_rss_channels = 0;
 691         snprintf(rss_param_name, sizeof(rss_param_name),
 692                  SFXGE_PARAM(%d.max_rss_channels),
 693                  (int)device_get_unit(dev));
 694         TUNABLE_INT_FETCH(rss_param_name, &sc->max_rss_channels);
 695 #if EFSYS_OPT_MCDI_LOGGING
 696         sc->mcdi_logging = sfxge_mcdi_logging;
 697         snprintf(mcdi_log_param_name, sizeof(mcdi_log_param_name),
 698                  SFXGE_PARAM(%d.mcdi_logging),
 699                  (int)device_get_unit(dev));
 700         TUNABLE_INT_FETCH(mcdi_log_param_name, &sc->mcdi_logging);
 701 #endif
 702
 703         sc->stats_node = SYSCTL_ADD_NODE(
 704                 device_get_sysctl_ctx(dev),
 705                 SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 706                 OID_AUTO, "stats", CTLFLAG_RD, NULL, "Statistics");
 707         if (sc->stats_node == NULL) {
 708                 error = ENOMEM;
 709                 goto fail;
 710         }
 711
 712         TASK_INIT(&sc->task_reset, 0, sfxge_reset, sc);
 713
 714         (void) pci_enable_busmaster(dev);
 715
 716         /* Initialize DMA mappings. */
 717         DBGPRINT(sc->dev, "dma_init...");
 718         if ((error = sfxge_dma_init(sc)) != 0)
 719                 goto fail;
 720
 721         /* Map the device registers. */
 722         DBGPRINT(sc->dev, "bar_init...");
 723         if ((error = sfxge_bar_init(sc)) != 0)
 724                 goto fail;
 725
 726         error = efx_family(pci_get_vendor(dev), pci_get_device(dev),
 727             &sc->family);
 728         KASSERT(error == 0, ("Family should be filtered by sfxge_probe()"));
 729
 730         DBGPRINT(sc->dev, "nic_create...");
 731
 732         /* Create the common code nic object. */
 733         SFXGE_EFSYS_LOCK_INIT(&sc->enp_lock,
 734                               device_get_nameunit(sc->dev), "nic");
 735         if ((error = efx_nic_create(sc->family, (efsys_identifier_t *)sc,
 736             &sc->bar, &sc->enp_lock, &enp)) != 0)
 737                 goto fail3;
 738         sc->enp = enp;
 739
 740         /* Initialize MCDI to talk to the microcontroller. */
 741         DBGPRINT(sc->dev, "mcdi_init...");
 742         if ((error = sfxge_mcdi_init(sc)) != 0)
 743                 goto fail4;
 744
 745         /* Probe the NIC and build the configuration data area. */
 746         DBGPRINT(sc->dev, "nic_probe...");
 747         if ((error = efx_nic_probe(enp)) != 0)
 748                 goto fail5;
 749
 750         if (!ISP2(sfxge_rx_ring_entries) ||
 751             (sfxge_rx_ring_entries < EFX_RXQ_MINNDESCS) ||
 752             (sfxge_rx_ring_entries > EFX_RXQ_MAXNDESCS)) {
 753                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 754                     SFXGE_PARAM_RX_RING, sfxge_rx_ring_entries,
 755                     EFX_RXQ_MINNDESCS, EFX_RXQ_MAXNDESCS);
 756                 error = EINVAL;
 757                 goto fail_rx_ring_entries;
 758         }
 759         sc->rxq_entries = sfxge_rx_ring_entries;
 760
 761         if (!ISP2(sfxge_tx_ring_entries) ||
 762             (sfxge_tx_ring_entries < EFX_TXQ_MINNDESCS) ||
 763             (sfxge_tx_ring_entries > EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)))) {
 764                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 765                     SFXGE_PARAM_TX_RING, sfxge_tx_ring_entries,
 766                     EFX_TXQ_MINNDESCS, EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)));
 767                 error = EINVAL;
 768                 goto fail_tx_ring_entries;
 769         }
 770         sc->txq_entries = sfxge_tx_ring_entries;
 771
 772         SYSCTL_ADD_STRING(device_get_sysctl_ctx(dev),
 773                           SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 774                           OID_AUTO, "version", CTLFLAG_RD,
 775                           SFXGE_VERSION_STRING, 0,
 776                           "Driver version");
 777
 778         SYSCTL_ADD_UINT(device_get_sysctl_ctx(dev),
 779                         SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 780                         OID_AUTO, "phy_type", CTLFLAG_RD,
 781                         NULL, efx_nic_cfg_get(enp)->enc_phy_type,
 782                         "PHY type");
 783
 784         /* Initialize the NVRAM. */
 785         DBGPRINT(sc->dev, "nvram_init...");
 786         if ((error = efx_nvram_init(enp)) != 0)
 787                 goto fail6;
 788
 789         /* Initialize the VPD. */
 790         DBGPRINT(sc->dev, "vpd_init...");
 791         if ((error = efx_vpd_init(enp)) != 0)
 792                 goto fail7;
 793
 794         efx_mcdi_new_epoch(enp);
 795
 796         /* Reset the NIC. */
 797         DBGPRINT(sc->dev, "nic_reset...");
 798         if ((error = efx_nic_reset(enp)) != 0)
 799                 goto fail8;
 800
 801         /* Initialize buffer table allocation. */
 802         sc->buffer_table_next = 0;
 803
 804         /*
 805          * Guarantee minimum and estimate maximum number of event queues
 806          * to take it into account when MSI-X interrupts are allocated.
 807          * It initializes NIC and keeps it initialized on success.
 808          */
 809         if ((error = sfxge_estimate_rsrc_limits(sc)) != 0)
 810                 goto fail8;
 811
 812         /* Set up interrupts. */
 813         DBGPRINT(sc->dev, "intr_init...");
 814         if ((error = sfxge_intr_init(sc)) != 0)
 815                 goto fail9;
 816
 817         /* Initialize event processing state. */
 818         DBGPRINT(sc->dev, "ev_init...");
 819         if ((error = sfxge_ev_init(sc)) != 0)
 820                 goto fail11;
 821
 822         /* Initialize port state. */
 823         DBGPRINT(sc->dev, "port_init...");
 824         if ((error = sfxge_port_init(sc)) != 0)
 825                 goto fail12;
 826
 827         /* Initialize receive state. */
 828         DBGPRINT(sc->dev, "rx_init...");
 829         if ((error = sfxge_rx_init(sc)) != 0)
 830                 goto fail13;
 831
 832         /* Initialize transmit state. */
 833         DBGPRINT(sc->dev, "tx_init...");
 834         if ((error = sfxge_tx_init(sc)) != 0)
 835                 goto fail14;
 836
 837         sc->init_state = SFXGE_INITIALIZED;
 838
 839         DBGPRINT(sc->dev, "success");
 840         return (0);
 841
 842 fail14:
 843         sfxge_rx_fini(sc);
 844
 845 fail13:
 846         sfxge_port_fini(sc);
 847
 848 fail12:
 849         sfxge_ev_fini(sc);
 850
 851 fail11:
 852         sfxge_intr_fini(sc);
 853
 854 fail9:
 855         efx_nic_fini(sc->enp);
 856
 857 fail8:
 858         efx_vpd_fini(enp);
 859
 860 fail7:
 861         efx_nvram_fini(enp);
 862
 863 fail6:
 864 fail_tx_ring_entries:
 865 fail_rx_ring_entries:
 866         efx_nic_unprobe(enp);
 867
 868 fail5:
 869         sfxge_mcdi_fini(sc);
 870
 871 fail4:
 872         sc->enp = NULL;
 873         efx_nic_destroy(enp);
 874         SFXGE_EFSYS_LOCK_DESTROY(&sc->enp_lock);
 875
 876 fail3:
 877         sfxge_bar_fini(sc);
 878         (void) pci_disable_busmaster(sc->dev);
 879
 880 fail:
 881         DBGPRINT(sc->dev, "failed %d", error);
 882         sc->dev = NULL;
 883         SFXGE_ADAPTER_LOCK_DESTROY(sc);
 884         return (error);
 885 }
 886
 887 static void
 888 sfxge_destroy(struct sfxge_softc *sc)
 889 {
 890         efx_nic_t *enp;
 891
 892         /* Clean up transmit state. */
 893         sfxge_tx_fini(sc);
 894
 895         /* Clean up receive state. */
 896         sfxge_rx_fini(sc);
 897
 898         /* Clean up port state. */
 899         sfxge_port_fini(sc);
 900
 901         /* Clean up event processing state. */
 902         sfxge_ev_fini(sc);
 903
 904         /* Clean up interrupts. */
 905         sfxge_intr_fini(sc);
 906
 907         /* Tear down common code subsystems. */
 908         efx_nic_reset(sc->enp);
 909         efx_vpd_fini(sc->enp);
 910         efx_nvram_fini(sc->enp);
 911         efx_nic_unprobe(sc->enp);
 912
 913         /* Tear down MCDI. */
 914         sfxge_mcdi_fini(sc);
 915
 916         /* Destroy common code context. */
 917         enp = sc->enp;
 918         sc->enp = NULL;
 919         efx_nic_destroy(enp);
 920
 921         /* Free DMA memory. */
 922         sfxge_dma_fini(sc);
 923
 924         /* Free mapped BARs. */
 925         sfxge_bar_fini(sc);
 926
 927         (void) pci_disable_busmaster(sc->dev);
 928
 929         taskqueue_drain(taskqueue_thread, &sc->task_reset);
 930
 931         /* Destroy the softc lock. */
 932         SFXGE_ADAPTER_LOCK_DESTROY(sc);
 933 }
 934
 935 static int
 936 sfxge_vpd_handler(SYSCTL_HANDLER_ARGS)
 937 {
 938         struct sfxge_softc *sc = arg1;
 939         efx_vpd_value_t value;
 940         int rc;
 941
 942         value.evv_tag = arg2 >> 16;
 943         value.evv_keyword = arg2 & 0xffff;
 944         if ((rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value))
 945             != 0)
 946                 return (rc);
 947
 948         return (SYSCTL_OUT(req, value.evv_value, value.evv_length));
 949 }
 950
 951 static void
 952 sfxge_vpd_try_add(struct sfxge_softc *sc, struct sysctl_oid_list *list,
 953                   efx_vpd_tag_t tag, const char *keyword)
 954 {
 955         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
 956         efx_vpd_value_t value;
 957
 958         /* Check whether VPD tag/keyword is present */
 959         value.evv_tag = tag;
 960         value.evv_keyword = EFX_VPD_KEYWORD(keyword[0], keyword[1]);
 961         if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) != 0)
 962                 return;
 963
 964         SYSCTL_ADD_PROC(
 965                 ctx, list, OID_AUTO, keyword, CTLTYPE_STRING|CTLFLAG_RD,
 966                 sc, tag << 16 | EFX_VPD_KEYWORD(keyword[0], keyword[1]),
 967                 sfxge_vpd_handler, "A", "");
 968 }
 969
 970 static int
 971 sfxge_vpd_init(struct sfxge_softc *sc)
 972 {
 973         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
 974         struct sysctl_oid *vpd_node;
 975         struct sysctl_oid_list *vpd_list;
 976         char keyword[3];
 977         efx_vpd_value_t value;
 978         int rc;
 979
 980         if ((rc = efx_vpd_size(sc->enp, &sc->vpd_size)) != 0) {
 981                 /*
 982                  * Unpriviledged functions deny VPD access.
 983                  * Simply skip VPD in this case.
 984                  */
 985                 if (rc == EACCES)
 986                         goto done;
 987                 goto fail;
 988         }
 989         sc->vpd_data = malloc(sc->vpd_size, M_SFXGE, M_WAITOK);
 990         if ((rc = efx_vpd_read(sc->enp, sc->vpd_data, sc->vpd_size)) != 0)
 991                 goto fail2;
 992
 993         /* Copy ID (product name) into device description, and log it. */
 994         value.evv_tag = EFX_VPD_ID;
 995         if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) == 0) {
 996                 value.evv_value[value.evv_length] = 0;
 997                 device_set_desc_copy(sc->dev, value.evv_value);
 998                 device_printf(sc->dev, "%s\n", value.evv_value);
 999         }
1000
1001         vpd_node = SYSCTL_ADD_NODE(
1002                 ctx, SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1003                 OID_AUTO, "vpd", CTLFLAG_RD, NULL, "Vital Product Data");
1004         vpd_list = SYSCTL_CHILDREN(vpd_node);
1005
1006         /* Add sysctls for all expected and any vendor-defined keywords. */
1007         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "PN");
1008         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "EC");
1009         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "SN");
1010         keyword[0] = 'V';
1011         keyword[2] = 0;
1012         for (keyword[1] = '0'; keyword[1] <= '9'; keyword[1]++)
1013                 sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword);
1014         for (keyword[1] = 'A'; keyword[1] <= 'Z'; keyword[1]++)
1015                 sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword);
1016
1017 done:
1018         return (0);
1019
1020 fail2:
1021         free(sc->vpd_data, M_SFXGE);
1022 fail:
1023         return (rc);
1024 }
1025
1026 static void
1027 sfxge_vpd_fini(struct sfxge_softc *sc)
1028 {
1029         free(sc->vpd_data, M_SFXGE);
1030 }
1031
1032 static void
1033 sfxge_reset(void *arg, int npending)
1034 {
1035         struct sfxge_softc *sc;
1036         int rc;
1037         unsigned attempt;
1038
1039         (void)npending;
1040
1041         sc = (struct sfxge_softc *)arg;
1042
1043         SFXGE_ADAPTER_LOCK(sc);
1044
1045         if (sc->init_state != SFXGE_STARTED)
1046                 goto done;
1047
1048         sfxge_stop(sc);
1049         efx_nic_reset(sc->enp);
1050         for (attempt = 0; attempt < sfxge_restart_attempts; ++attempt) {
1051                 if ((rc = sfxge_start(sc)) == 0)
1052                         goto done;
1053
1054                 device_printf(sc->dev, "start on reset failed (%d)\n", rc);
1055                 DELAY(100000);
1056         }
1057
1058         device_printf(sc->dev, "reset failed; interface is now stopped\n");
1059
1060 done:
1061         SFXGE_ADAPTER_UNLOCK(sc);
1062 }
1063
1064 void
1065 sfxge_schedule_reset(struct sfxge_softc *sc)
1066 {
1067         taskqueue_enqueue(taskqueue_thread, &sc->task_reset);
1068 }
1069
1070 static int
1071 sfxge_attach(device_t dev)
1072 {
1073         struct sfxge_softc *sc;
1074         struct ifnet *ifp;
1075         int error;
1076
1077         sc = device_get_softc(dev);
1078         sc->dev = dev;
1079
1080         /* Allocate ifnet. */
1081         ifp = if_alloc(IFT_ETHER);
1082         if (ifp == NULL) {
1083                 device_printf(dev, "Couldn't allocate ifnet\n");
1084                 error = ENOMEM;
1085                 goto fail;
1086         }
1087         sc->ifnet = ifp;
1088
1089         /* Initialize hardware. */
1090         DBGPRINT(sc->dev, "create nic");
1091         if ((error = sfxge_create(sc)) != 0)
1092                 goto fail2;
1093
1094         /* Create the ifnet for the port. */
1095         DBGPRINT(sc->dev, "init ifnet");
1096         if ((error = sfxge_ifnet_init(ifp, sc)) != 0)
1097                 goto fail3;
1098
1099         DBGPRINT(sc->dev, "init vpd");
1100         if ((error = sfxge_vpd_init(sc)) != 0)
1101                 goto fail4;
1102
1103         /*
1104          * NIC is initialized inside sfxge_create() and kept inialized
1105          * to be able to initialize port to discover media types in
1106          * sfxge_ifnet_init().
1107          */
1108         efx_nic_fini(sc->enp);
1109
1110         sc->init_state = SFXGE_REGISTERED;
1111
1112         DBGPRINT(sc->dev, "success");
1113         return (0);
1114
1115 fail4:
1116         sfxge_ifnet_fini(ifp);
1117 fail3:
1118         efx_nic_fini(sc->enp);
1119         sfxge_destroy(sc);
1120
1121 fail2:
1122         if_free(sc->ifnet);
1123
1124 fail:
1125         DBGPRINT(sc->dev, "failed %d", error);
1126         return (error);
1127 }
1128
1129 static int
1130 sfxge_detach(device_t dev)
1131 {
1132         struct sfxge_softc *sc;
1133
1134         sc = device_get_softc(dev);
1135
1136         sfxge_vpd_fini(sc);
1137
1138         /* Destroy the ifnet. */
1139         sfxge_ifnet_fini(sc->ifnet);
1140
1141         /* Tear down hardware. */
1142         sfxge_destroy(sc);
1143
1144         return (0);
1145 }
1146
1147 static int
1148 sfxge_probe(device_t dev)
1149 {
1150         uint16_t pci_vendor_id;
1151         uint16_t pci_device_id;
1152         efx_family_t family;
1153         int rc;
1154
1155         pci_vendor_id = pci_get_vendor(dev);
1156         pci_device_id = pci_get_device(dev);
1157
1158         DBGPRINT(dev, "PCI ID %04x:%04x", pci_vendor_id, pci_device_id);
1159         rc = efx_family(pci_vendor_id, pci_device_id, &family);
1160         if (rc != 0) {
1161                 DBGPRINT(dev, "efx_family fail %d", rc);
1162                 return (ENXIO);
1163         }
1164
1165         if (family == EFX_FAMILY_SIENA) {
1166                 device_set_desc(dev, "Solarflare SFC9000 family");
1167                 return (0);
1168         }
1169
1170         if (family == EFX_FAMILY_HUNTINGTON) {
1171                 device_set_desc(dev, "Solarflare SFC9100 family");
1172                 return (0);
1173         }
1174
1175         if (family == EFX_FAMILY_MEDFORD) {
1176                 device_set_desc(dev, "Solarflare SFC9200 family");
1177                 return (0);
1178         }
1179
1180         DBGPRINT(dev, "impossible controller family %d", family);
1181         return (ENXIO);
1182 }
1183
1184 static device_method_t sfxge_methods[] = {
1185         DEVMETHOD(device_probe,         sfxge_probe),
1186         DEVMETHOD(device_attach,        sfxge_attach),
1187         DEVMETHOD(device_detach,        sfxge_detach),
1188
1189         DEVMETHOD_END
1190 };
1191
1192 static devclass_t sfxge_devclass;
1193
1194 static driver_t sfxge_driver = {
1195         "sfxge",
1196         sfxge_methods,
1197         sizeof(struct sfxge_softc)
1198 };
1199
1200 DRIVER_MODULE(sfxge, pci, sfxge_driver, sfxge_devclass, 0, 0);