sys/dev/sfxge/sfxge.c

   1 /*-
   2  * Copyright (c) 2010-2015 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 <sys/param.h>
  38 #include <sys/kernel.h>
  39 #include <sys/bus.h>
  40 #include <sys/rman.h>
  41 #include <sys/lock.h>
  42 #include <sys/module.h>
  43 #include <sys/mutex.h>
  44 #include <sys/smp.h>
  45 #include <sys/socket.h>
  46 #include <sys/taskqueue.h>
  47 #include <sys/sockio.h>
  48 #include <sys/sysctl.h>
  49 #include <sys/priv.h>
  50 #include <sys/syslog.h>
  51
  52 #include <dev/pci/pcireg.h>
  53 #include <dev/pci/pcivar.h>
  54
  55 #include <net/ethernet.h>
  56 #include <net/if.h>
  57 #include <net/if_media.h>
  58 #include <net/if_types.h>
  59
  60 #include "common/efx.h"
  61
  62 #include "sfxge.h"
  63 #include "sfxge_rx.h"
  64 #include "sfxge_ioc.h"
  65 #include "sfxge_version.h"
  66
  67 #define SFXGE_CAP (IFCAP_VLAN_MTU | IFCAP_VLAN_HWCSUM |                 \
  68                    IFCAP_RXCSUM | IFCAP_TXCSUM |                        \
  69                    IFCAP_RXCSUM_IPV6 | IFCAP_TXCSUM_IPV6 |              \
  70                    IFCAP_TSO4 | IFCAP_TSO6 |                            \
  71                    IFCAP_JUMBO_MTU |                                    \
  72                    IFCAP_VLAN_HWTSO | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  73 #define SFXGE_CAP_ENABLE SFXGE_CAP
  74 #define SFXGE_CAP_FIXED (IFCAP_VLAN_MTU |                               \
  75                          IFCAP_JUMBO_MTU | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  76
  77 MALLOC_DEFINE(M_SFXGE, "sfxge", "Solarflare 10GigE driver");
  78
  79
  80 SYSCTL_NODE(_hw, OID_AUTO, sfxge, CTLFLAG_RD, 0,
  81             "SFXGE driver parameters");
  82
  83 #define SFXGE_PARAM_RX_RING     SFXGE_PARAM(rx_ring)
  84 static int sfxge_rx_ring_entries = SFXGE_NDESCS;
  85 TUNABLE_INT(SFXGE_PARAM_RX_RING, &sfxge_rx_ring_entries);
  86 SYSCTL_INT(_hw_sfxge, OID_AUTO, rx_ring, CTLFLAG_RDTUN,
  87            &sfxge_rx_ring_entries, 0,
  88            "Maximum number of descriptors in a receive ring");
  89
  90 #define SFXGE_PARAM_TX_RING     SFXGE_PARAM(tx_ring)
  91 static int sfxge_tx_ring_entries = SFXGE_NDESCS;
  92 TUNABLE_INT(SFXGE_PARAM_TX_RING, &sfxge_tx_ring_entries);
  93 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_ring, CTLFLAG_RDTUN,
  94            &sfxge_tx_ring_entries, 0,
  95            "Maximum number of descriptors in a transmit ring");
  96
  97 #define SFXGE_PARAM_STATS_UPDATE_PERIOD SFXGE_PARAM(stats_update_period)
  98 static int sfxge_stats_update_period = SFXGE_CALLOUT_TICKS;
  99 TUNABLE_INT(SFXGE_PARAM_STATS_UPDATE_PERIOD,
 100             &sfxge_stats_update_period);
 101 SYSCTL_INT(_hw_sfxge, OID_AUTO, stats_update_period, CTLFLAG_RDTUN,
 102            &sfxge_stats_update_period, 0,
 103            "netstat interface statistics update period in ticks");
 104
 105 static void
 106 sfxge_reset(void *arg, int npending);
 107
 108 static int
 109 sfxge_estimate_rsrc_limits(struct sfxge_softc *sc)
 110 {
 111         efx_drv_limits_t limits;
 112         int rc;
 113         unsigned int evq_max;
 114         uint32_t evq_allocated;
 115         uint32_t rxq_allocated;
 116         uint32_t txq_allocated;
 117
 118         /*
 119          * Limit the number of event queues to:
 120          *  - number of CPUs
 121          *  - hardwire maximum RSS channels
 122          *  - administratively specified maximum RSS channels
 123          */
 124         evq_max = MIN(mp_ncpus, EFX_MAXRSS);
 125         if (sc->max_rss_channels > 0)
 126                 evq_max = MIN(evq_max, sc->max_rss_channels);
 127
 128         memset(&limits, 0, sizeof(limits));
 129
 130         limits.edl_min_evq_count = 1;
 131         limits.edl_max_evq_count = evq_max;
 132         limits.edl_min_txq_count = SFXGE_TXQ_NTYPES;
 133         limits.edl_max_txq_count = evq_max + SFXGE_TXQ_NTYPES - 1;
 134         limits.edl_min_rxq_count = 1;
 135         limits.edl_max_rxq_count = evq_max;
 136
 137         efx_nic_set_drv_limits(sc->enp, &limits);
 138
 139         if ((rc = efx_nic_init(sc->enp)) != 0)
 140                 return (rc);
 141
 142         rc = efx_nic_get_vi_pool(sc->enp, &evq_allocated, &rxq_allocated,
 143                                  &txq_allocated);
 144         if (rc != 0) {
 145                 efx_nic_fini(sc->enp);
 146                 return (rc);
 147         }
 148
 149         KASSERT(txq_allocated >= SFXGE_TXQ_NTYPES,
 150                 ("txq_allocated < SFXGE_TXQ_NTYPES"));
 151
 152         sc->evq_max = MIN(evq_allocated, evq_max);
 153         sc->evq_max = MIN(rxq_allocated, sc->evq_max);
 154         sc->evq_max = MIN(txq_allocated - (SFXGE_TXQ_NTYPES - 1),
 155                           sc->evq_max);
 156
 157         KASSERT(sc->evq_max <= evq_max,
 158                 ("allocated more than maximum requested"));
 159
 160         /*
 161          * NIC is kept initialized in the case of success to be able to
 162          * initialize port to find out media types.
 163          */
 164         return (0);
 165 }
 166
 167 static int
 168 sfxge_set_drv_limits(struct sfxge_softc *sc)
 169 {
 170         efx_drv_limits_t limits;
 171
 172         memset(&limits, 0, sizeof(limits));
 173
 174         /* Limits are strict since take into account initial estimation */
 175         limits.edl_min_evq_count = limits.edl_max_evq_count =
 176             sc->intr.n_alloc;
 177         limits.edl_min_txq_count = limits.edl_max_txq_count =
 178             sc->intr.n_alloc + SFXGE_TXQ_NTYPES - 1;
 179         limits.edl_min_rxq_count = limits.edl_max_rxq_count =
 180             sc->intr.n_alloc;
 181
 182         return (efx_nic_set_drv_limits(sc->enp, &limits));
 183 }
 184
 185 static int
 186 sfxge_start(struct sfxge_softc *sc)
 187 {
 188         int rc;
 189
 190         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 191
 192         if (sc->init_state == SFXGE_STARTED)
 193                 return (0);
 194
 195         if (sc->init_state != SFXGE_REGISTERED) {
 196                 rc = EINVAL;
 197                 goto fail;
 198         }
 199
 200         /* Set required resource limits */
 201         if ((rc = sfxge_set_drv_limits(sc)) != 0)
 202                 goto fail;
 203
 204         if ((rc = efx_nic_init(sc->enp)) != 0)
 205                 goto fail;
 206
 207         /* Start processing interrupts. */
 208         if ((rc = sfxge_intr_start(sc)) != 0)
 209                 goto fail2;
 210
 211         /* Start processing events. */
 212         if ((rc = sfxge_ev_start(sc)) != 0)
 213                 goto fail3;
 214
 215         /* Fire up the port. */
 216         if ((rc = sfxge_port_start(sc)) != 0)
 217                 goto fail4;
 218
 219         /* Start the receiver side. */
 220         if ((rc = sfxge_rx_start(sc)) != 0)
 221                 goto fail5;
 222
 223         /* Start the transmitter side. */
 224         if ((rc = sfxge_tx_start(sc)) != 0)
 225                 goto fail6;
 226
 227         sc->init_state = SFXGE_STARTED;
 228
 229         /* Tell the stack we're running. */
 230         sc->ifnet->if_drv_flags |= IFF_DRV_RUNNING;
 231         sc->ifnet->if_drv_flags &= ~IFF_DRV_OACTIVE;
 232
 233         return (0);
 234
 235 fail6:
 236         sfxge_rx_stop(sc);
 237
 238 fail5:
 239         sfxge_port_stop(sc);
 240
 241 fail4:
 242         sfxge_ev_stop(sc);
 243
 244 fail3:
 245         sfxge_intr_stop(sc);
 246
 247 fail2:
 248         efx_nic_fini(sc->enp);
 249
 250 fail:
 251         device_printf(sc->dev, "sfxge_start: %d\n", rc);
 252
 253         return (rc);
 254 }
 255
 256 static void
 257 sfxge_if_init(void *arg)
 258 {
 259         struct sfxge_softc *sc;
 260
 261         sc = (struct sfxge_softc *)arg;
 262
 263         SFXGE_ADAPTER_LOCK(sc);
 264         (void)sfxge_start(sc);
 265         SFXGE_ADAPTER_UNLOCK(sc);
 266 }
 267
 268 static void
 269 sfxge_stop(struct sfxge_softc *sc)
 270 {
 271         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 272
 273         if (sc->init_state != SFXGE_STARTED)
 274                 return;
 275
 276         sc->init_state = SFXGE_REGISTERED;
 277
 278         /* Stop the transmitter. */
 279         sfxge_tx_stop(sc);
 280
 281         /* Stop the receiver. */
 282         sfxge_rx_stop(sc);
 283
 284         /* Stop the port. */
 285         sfxge_port_stop(sc);
 286
 287         /* Stop processing events. */
 288         sfxge_ev_stop(sc);
 289
 290         /* Stop processing interrupts. */
 291         sfxge_intr_stop(sc);
 292
 293         efx_nic_fini(sc->enp);
 294
 295         sc->ifnet->if_drv_flags &= ~IFF_DRV_RUNNING;
 296 }
 297
 298
 299 static int
 300 sfxge_vpd_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 301 {
 302         efx_vpd_value_t value;
 303         int rc = 0;
 304
 305         switch (ioc->u.vpd.op) {
 306         case SFXGE_VPD_OP_GET_KEYWORD:
 307                 value.evv_tag = ioc->u.vpd.tag;
 308                 value.evv_keyword = ioc->u.vpd.keyword;
 309                 rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 310                 if (rc != 0)
 311                         break;
 312                 ioc->u.vpd.len = MIN(ioc->u.vpd.len, value.evv_length);
 313                 if (ioc->u.vpd.payload != 0) {
 314                         rc = copyout(value.evv_value, ioc->u.vpd.payload,
 315                                      ioc->u.vpd.len);
 316                 }
 317                 break;
 318         case SFXGE_VPD_OP_SET_KEYWORD:
 319                 if (ioc->u.vpd.len > sizeof(value.evv_value))
 320                         return (EINVAL);
 321                 value.evv_tag = ioc->u.vpd.tag;
 322                 value.evv_keyword = ioc->u.vpd.keyword;
 323                 value.evv_length = ioc->u.vpd.len;
 324                 rc = copyin(ioc->u.vpd.payload, value.evv_value, value.evv_length);
 325                 if (rc != 0)
 326                         break;
 327                 rc = efx_vpd_set(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 328                 if (rc != 0)
 329                         break;
 330                 rc = efx_vpd_verify(sc->enp, sc->vpd_data, sc->vpd_size);
 331                 if (rc != 0)
 332                         break;
 333                 rc = efx_vpd_write(sc->enp, sc->vpd_data, sc->vpd_size);
 334                 break;
 335         default:
 336                 rc = EOPNOTSUPP;
 337                 break;
 338         }
 339
 340         return (rc);
 341 }
 342
 343 static int
 344 sfxge_private_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 345 {
 346         switch (ioc->op) {
 347         case SFXGE_MCDI_IOC:
 348                 return (sfxge_mcdi_ioctl(sc, ioc));
 349         case SFXGE_NVRAM_IOC:
 350                 return (sfxge_nvram_ioctl(sc, ioc));
 351         case SFXGE_VPD_IOC:
 352                 return (sfxge_vpd_ioctl(sc, ioc));
 353         default:
 354                 return (EOPNOTSUPP);
 355         }
 356 }
 357
 358
 359 static int
 360 sfxge_if_ioctl(struct ifnet *ifp, unsigned long command, caddr_t data)
 361 {
 362         struct sfxge_softc *sc;
 363         struct ifreq *ifr;
 364         sfxge_ioc_t ioc;
 365         int error;
 366
 367         ifr = (struct ifreq *)data;
 368         sc = ifp->if_softc;
 369         error = 0;
 370
 371         switch (command) {
 372         case SIOCSIFFLAGS:
 373                 SFXGE_ADAPTER_LOCK(sc);
 374                 if (ifp->if_flags & IFF_UP) {
 375                         if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
 376                                 if ((ifp->if_flags ^ sc->if_flags) &
 377                                     (IFF_PROMISC | IFF_ALLMULTI)) {
 378                                         sfxge_mac_filter_set(sc);
 379                                 }
 380                         } else
 381                                 sfxge_start(sc);
 382                 } else
 383                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 384                                 sfxge_stop(sc);
 385                 sc->if_flags = ifp->if_flags;
 386                 SFXGE_ADAPTER_UNLOCK(sc);
 387                 break;
 388         case SIOCSIFMTU:
 389                 if (ifr->ifr_mtu == ifp->if_mtu) {
 390                         /* Nothing to do */
 391                         error = 0;
 392                 } else if (ifr->ifr_mtu > SFXGE_MAX_MTU) {
 393                         error = EINVAL;
 394                 } else if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
 395                         ifp->if_mtu = ifr->ifr_mtu;
 396                         error = 0;
 397                 } else {
 398                         /* Restart required */
 399                         SFXGE_ADAPTER_LOCK(sc);
 400                         sfxge_stop(sc);
 401                         ifp->if_mtu = ifr->ifr_mtu;
 402                         error = sfxge_start(sc);
 403                         SFXGE_ADAPTER_UNLOCK(sc);
 404                         if (error != 0) {
 405                                 ifp->if_flags &= ~IFF_UP;
 406                                 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
 407                                 if_down(ifp);
 408                         }
 409                 }
 410                 break;
 411         case SIOCADDMULTI:
 412         case SIOCDELMULTI:
 413                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 414                         sfxge_mac_filter_set(sc);
 415                 break;
 416         case SIOCSIFCAP:
 417         {
 418                 int reqcap = ifr->ifr_reqcap;
 419                 int capchg_mask;
 420
 421                 SFXGE_ADAPTER_LOCK(sc);
 422
 423                 /* Capabilities to be changed in accordance with request */
 424                 capchg_mask = ifp->if_capenable ^ reqcap;
 425
 426                 /*
 427                  * The networking core already rejects attempts to
 428                  * enable capabilities we don't have.  We still have
 429                  * to reject attempts to disable capabilities that we
 430                  * can't (yet) disable.
 431                  */
 432                 KASSERT((reqcap & ~ifp->if_capabilities) == 0,
 433                     ("Unsupported capabilities 0x%x requested 0x%x vs "
 434                      "supported 0x%x",
 435                      reqcap & ~ifp->if_capabilities,
 436                      reqcap , ifp->if_capabilities));
 437                 if (capchg_mask & SFXGE_CAP_FIXED) {
 438                         error = EINVAL;
 439                         SFXGE_ADAPTER_UNLOCK(sc);
 440                         break;
 441                 }
 442
 443                 /* Check request before any changes */
 444                 if ((capchg_mask & IFCAP_TSO4) &&
 445                     (reqcap & (IFCAP_TSO4 | IFCAP_TXCSUM)) == IFCAP_TSO4) {
 446                         error = EAGAIN;
 447                         SFXGE_ADAPTER_UNLOCK(sc);
 448                         if_printf(ifp, "enable txcsum before tso4\n");
 449                         break;
 450                 }
 451                 if ((capchg_mask & IFCAP_TSO6) &&
 452                     (reqcap & (IFCAP_TSO6 | IFCAP_TXCSUM_IPV6)) == IFCAP_TSO6) {
 453                         error = EAGAIN;
 454                         SFXGE_ADAPTER_UNLOCK(sc);
 455                         if_printf(ifp, "enable txcsum6 before tso6\n");
 456                         break;
 457                 }
 458
 459                 if (reqcap & IFCAP_TXCSUM) {
 460                         ifp->if_hwassist |= (CSUM_IP | CSUM_TCP | CSUM_UDP);
 461                 } else {
 462                         ifp->if_hwassist &= ~(CSUM_IP | CSUM_TCP | CSUM_UDP);
 463                         if (reqcap & IFCAP_TSO4) {
 464                                 reqcap &= ~IFCAP_TSO4;
 465                                 if_printf(ifp,
 466                                     "tso4 disabled due to -txcsum\n");
 467                         }
 468                 }
 469                 if (reqcap & IFCAP_TXCSUM_IPV6) {
 470                         ifp->if_hwassist |= (CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 471                 } else {
 472                         ifp->if_hwassist &= ~(CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 473                         if (reqcap & IFCAP_TSO6) {
 474                                 reqcap &= ~IFCAP_TSO6;
 475                                 if_printf(ifp,
 476                                     "tso6 disabled due to -txcsum6\n");
 477                         }
 478                 }
 479
 480                 /*
 481                  * The kernel takes both IFCAP_TSOx and CSUM_TSO into
 482                  * account before using TSO. So, we do not touch
 483                  * checksum flags when IFCAP_TSOx is modified.
 484                  * Note that CSUM_TSO is (CSUM_IP_TSO|CSUM_IP6_TSO),
 485                  * but both bits are set in IPv4 and IPv6 mbufs.
 486                  */
 487
 488                 ifp->if_capenable = reqcap;
 489
 490                 SFXGE_ADAPTER_UNLOCK(sc);
 491                 break;
 492         }
 493         case SIOCSIFMEDIA:
 494         case SIOCGIFMEDIA:
 495                 error = ifmedia_ioctl(ifp, ifr, &sc->media, command);
 496                 break;
 497         case SIOCGPRIVATE_0:
 498                 error = priv_check(curthread, PRIV_DRIVER);
 499                 if (error != 0)
 500                         break;
 501                 error = copyin(ifr->ifr_data, &ioc, sizeof(ioc));
 502                 if (error != 0)
 503                         return (error);
 504                 error = sfxge_private_ioctl(sc, &ioc);
 505                 if (error == 0) {
 506                         error = copyout(&ioc, ifr->ifr_data, sizeof(ioc));
 507                 }
 508                 break;
 509         default:
 510                 error = ether_ioctl(ifp, command, data);
 511         }
 512
 513         return (error);
 514 }
 515
 516 static void
 517 sfxge_tick(void *arg)
 518 {
 519         struct sfxge_softc *sc = arg;
 520
 521         sfxge_port_update_stats(sc);
 522         sfxge_tx_update_stats(sc);
 523
 524         callout_reset(&sc->tick_callout, sfxge_stats_update_period,
 525                       sfxge_tick, sc);
 526 }
 527
 528 static void
 529 sfxge_ifnet_fini(struct ifnet *ifp)
 530 {
 531         struct sfxge_softc *sc = ifp->if_softc;
 532
 533         callout_drain(&sc->tick_callout);
 534
 535         SFXGE_ADAPTER_LOCK(sc);
 536         sfxge_stop(sc);
 537         SFXGE_ADAPTER_UNLOCK(sc);
 538
 539         ifmedia_removeall(&sc->media);
 540         ether_ifdetach(ifp);
 541         if_free(ifp);
 542 }
 543
 544 static int
 545 sfxge_ifnet_init(struct ifnet *ifp, struct sfxge_softc *sc)
 546 {
 547         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
 548         device_t dev;
 549         int rc;
 550
 551         dev = sc->dev;
 552         sc->ifnet = ifp;
 553
 554         if_initname(ifp, device_get_name(dev), device_get_unit(dev));
 555         ifp->if_init = sfxge_if_init;
 556         ifp->if_softc = sc;
 557         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 558         ifp->if_ioctl = sfxge_if_ioctl;
 559
 560         ifp->if_capabilities = SFXGE_CAP;
 561         ifp->if_capenable = SFXGE_CAP_ENABLE;
 562
 563 #ifdef SFXGE_LRO
 564         ifp->if_capabilities |= IFCAP_LRO;
 565         ifp->if_capenable |= IFCAP_LRO;
 566 #endif
 567
 568         if (encp->enc_hw_tx_insert_vlan_enabled) {
 569                 ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING;
 570                 ifp->if_capenable |= IFCAP_VLAN_HWTAGGING;
 571         }
 572         ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO |
 573                            CSUM_TCP_IPV6 | CSUM_UDP_IPV6;
 574
 575         ether_ifattach(ifp, encp->enc_mac_addr);
 576
 577         ifp->if_transmit = sfxge_if_transmit;
 578         ifp->if_qflush = sfxge_if_qflush;
 579
 580         callout_init(&sc->tick_callout, B_TRUE);
 581
 582         DBGPRINT(sc->dev, "ifmedia_init");
 583         if ((rc = sfxge_port_ifmedia_init(sc)) != 0)
 584                 goto fail;
 585
 586         callout_reset(&sc->tick_callout, sfxge_stats_update_period,
 587                       sfxge_tick, sc);
 588
 589         return (0);
 590
 591 fail:
 592         ether_ifdetach(sc->ifnet);
 593         return (rc);
 594 }
 595
 596 void
 597 sfxge_sram_buf_tbl_alloc(struct sfxge_softc *sc, size_t n, uint32_t *idp)
 598 {
 599         KASSERT(sc->buffer_table_next + n <=
 600                 efx_nic_cfg_get(sc->enp)->enc_buftbl_limit,
 601                 ("buffer table full"));
 602
 603         *idp = sc->buffer_table_next;
 604         sc->buffer_table_next += n;
 605 }
 606
 607 static int
 608 sfxge_bar_init(struct sfxge_softc *sc)
 609 {
 610         efsys_bar_t *esbp = &sc->bar;
 611
 612         esbp->esb_rid = PCIR_BAR(EFX_MEM_BAR);
 613         if ((esbp->esb_res = bus_alloc_resource_any(sc->dev, SYS_RES_MEMORY,
 614             &esbp->esb_rid, RF_ACTIVE)) == NULL) {
 615                 device_printf(sc->dev, "Cannot allocate BAR region %d\n",
 616                     EFX_MEM_BAR);
 617                 return (ENXIO);
 618         }
 619         esbp->esb_tag = rman_get_bustag(esbp->esb_res);
 620         esbp->esb_handle = rman_get_bushandle(esbp->esb_res);
 621
 622         SFXGE_BAR_LOCK_INIT(esbp, device_get_nameunit(sc->dev));
 623
 624         return (0);
 625 }
 626
 627 static void
 628 sfxge_bar_fini(struct sfxge_softc *sc)
 629 {
 630         efsys_bar_t *esbp = &sc->bar;
 631
 632         bus_release_resource(sc->dev, SYS_RES_MEMORY, esbp->esb_rid,
 633             esbp->esb_res);
 634         SFXGE_BAR_LOCK_DESTROY(esbp);
 635 }
 636
 637 static int
 638 sfxge_create(struct sfxge_softc *sc)
 639 {
 640         device_t dev;
 641         efx_nic_t *enp;
 642         int error;
 643         char rss_param_name[sizeof(SFXGE_PARAM(%d.max_rss_channels))];
 644
 645         dev = sc->dev;
 646
 647         SFXGE_ADAPTER_LOCK_INIT(sc, device_get_nameunit(sc->dev));
 648
 649         sc->max_rss_channels = 0;
 650         snprintf(rss_param_name, sizeof(rss_param_name),
 651                  SFXGE_PARAM(%d.max_rss_channels),
 652                  (int)device_get_unit(dev));
 653         TUNABLE_INT_FETCH(rss_param_name, &sc->max_rss_channels);
 654
 655         sc->stats_node = SYSCTL_ADD_NODE(
 656                 device_get_sysctl_ctx(dev),
 657                 SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 658                 OID_AUTO, "stats", CTLFLAG_RD, NULL, "Statistics");
 659         if (sc->stats_node == NULL) {
 660                 error = ENOMEM;
 661                 goto fail;
 662         }
 663
 664         TASK_INIT(&sc->task_reset, 0, sfxge_reset, sc);
 665
 666         (void) pci_enable_busmaster(dev);
 667
 668         /* Initialize DMA mappings. */
 669         DBGPRINT(sc->dev, "dma_init...");
 670         if ((error = sfxge_dma_init(sc)) != 0)
 671                 goto fail;
 672
 673         /* Map the device registers. */
 674         DBGPRINT(sc->dev, "bar_init...");
 675         if ((error = sfxge_bar_init(sc)) != 0)
 676                 goto fail;
 677
 678         error = efx_family(pci_get_vendor(dev), pci_get_device(dev),
 679             &sc->family);
 680         KASSERT(error == 0, ("Family should be filtered by sfxge_probe()"));
 681
 682         DBGPRINT(sc->dev, "nic_create...");
 683
 684         /* Create the common code nic object. */
 685         SFXGE_EFSYS_LOCK_INIT(&sc->enp_lock,
 686                               device_get_nameunit(sc->dev), "nic");
 687         if ((error = efx_nic_create(sc->family, (efsys_identifier_t *)sc,
 688             &sc->bar, &sc->enp_lock, &enp)) != 0)
 689                 goto fail3;
 690         sc->enp = enp;
 691
 692         if (!ISP2(sfxge_rx_ring_entries) ||
 693             (sfxge_rx_ring_entries < EFX_RXQ_MINNDESCS) ||
 694             (sfxge_rx_ring_entries > EFX_RXQ_MAXNDESCS)) {
 695                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 696                     SFXGE_PARAM_RX_RING, sfxge_rx_ring_entries,
 697                     EFX_RXQ_MINNDESCS, EFX_RXQ_MAXNDESCS);
 698                 error = EINVAL;
 699                 goto fail_rx_ring_entries;
 700         }
 701         sc->rxq_entries = sfxge_rx_ring_entries;
 702
 703         if (!ISP2(sfxge_tx_ring_entries) ||
 704             (sfxge_tx_ring_entries < EFX_TXQ_MINNDESCS) ||
 705             (sfxge_tx_ring_entries > EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)))) {
 706                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 707                     SFXGE_PARAM_TX_RING, sfxge_tx_ring_entries,
 708                     EFX_TXQ_MINNDESCS, EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)));
 709                 error = EINVAL;
 710                 goto fail_tx_ring_entries;
 711         }
 712         sc->txq_entries = sfxge_tx_ring_entries;
 713
 714         /* Initialize MCDI to talk to the microcontroller. */
 715         DBGPRINT(sc->dev, "mcdi_init...");
 716         if ((error = sfxge_mcdi_init(sc)) != 0)
 717                 goto fail4;
 718
 719         /* Probe the NIC and build the configuration data area. */
 720         DBGPRINT(sc->dev, "nic_probe...");
 721         if ((error = efx_nic_probe(enp)) != 0)
 722                 goto fail5;
 723
 724         SYSCTL_ADD_STRING(device_get_sysctl_ctx(dev),
 725                           SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 726                           OID_AUTO, "version", CTLFLAG_RD,
 727                           SFXGE_VERSION_STRING, 0,
 728                           "Driver version");
 729
 730         SYSCTL_ADD_UINT(device_get_sysctl_ctx(dev),
 731                         SYSCTL_CHILDREN(device_get_sysctl_tree(dev)),
 732                         OID_AUTO, "phy_type", CTLFLAG_RD,
 733                         NULL, efx_nic_cfg_get(enp)->enc_phy_type,
 734                         "PHY type");
 735
 736         /* Initialize the NVRAM. */
 737         DBGPRINT(sc->dev, "nvram_init...");
 738         if ((error = efx_nvram_init(enp)) != 0)
 739                 goto fail6;
 740
 741         /* Initialize the VPD. */
 742         DBGPRINT(sc->dev, "vpd_init...");
 743         if ((error = efx_vpd_init(enp)) != 0)
 744                 goto fail7;
 745
 746         efx_mcdi_new_epoch(enp);
 747
 748         /* Reset the NIC. */
 749         DBGPRINT(sc->dev, "nic_reset...");
 750         if ((error = efx_nic_reset(enp)) != 0)
 751                 goto fail8;
 752
 753         /* Initialize buffer table allocation. */
 754         sc->buffer_table_next = 0;
 755
 756         /*
 757          * Guarantee minimum and estimate maximum number of event queues
 758          * to take it into account when MSI-X interrupts are allocated.
 759          * It initializes NIC and keeps it initialized on success.
 760          */
 761         if ((error = sfxge_estimate_rsrc_limits(sc)) != 0)
 762                 goto fail8;
 763
 764         /* Set up interrupts. */
 765         DBGPRINT(sc->dev, "intr_init...");
 766         if ((error = sfxge_intr_init(sc)) != 0)
 767                 goto fail9;
 768
 769         /* Initialize event processing state. */
 770         DBGPRINT(sc->dev, "ev_init...");
 771         if ((error = sfxge_ev_init(sc)) != 0)
 772                 goto fail11;
 773
 774         /* Initialize port state. */
 775         DBGPRINT(sc->dev, "port_init...");
 776         if ((error = sfxge_port_init(sc)) != 0)
 777                 goto fail12;
 778
 779         /* Initialize receive state. */
 780         DBGPRINT(sc->dev, "rx_init...");
 781         if ((error = sfxge_rx_init(sc)) != 0)
 782                 goto fail13;
 783
 784         /* Initialize transmit state. */
 785         DBGPRINT(sc->dev, "tx_init...");
 786         if ((error = sfxge_tx_init(sc)) != 0)
 787                 goto fail14;
 788
 789         sc->init_state = SFXGE_INITIALIZED;
 790
 791         DBGPRINT(sc->dev, "success");
 792         return (0);
 793
 794 fail14:
 795         sfxge_rx_fini(sc);
 796
 797 fail13:
 798         sfxge_port_fini(sc);
 799
 800 fail12:
 801         sfxge_ev_fini(sc);
 802
 803 fail11:
 804         sfxge_intr_fini(sc);
 805
 806 fail9:
 807         efx_nic_fini(sc->enp);
 808
 809 fail8:
 810         efx_vpd_fini(enp);
 811
 812 fail7:
 813         efx_nvram_fini(enp);
 814
 815 fail6:
 816         efx_nic_unprobe(enp);
 817
 818 fail5:
 819         sfxge_mcdi_fini(sc);
 820
 821 fail4:
 822 fail_tx_ring_entries:
 823 fail_rx_ring_entries:
 824         sc->enp = NULL;
 825         efx_nic_destroy(enp);
 826         SFXGE_EFSYS_LOCK_DESTROY(&sc->enp_lock);
 827
 828 fail3:
 829         sfxge_bar_fini(sc);
 830         (void) pci_disable_busmaster(sc->dev);
 831
 832 fail:
 833         DBGPRINT(sc->dev, "failed %d", error);
 834         sc->dev = NULL;
 835         SFXGE_ADAPTER_LOCK_DESTROY(sc);
 836         return (error);
 837 }
 838
 839 static void
 840 sfxge_destroy(struct sfxge_softc *sc)
 841 {
 842         efx_nic_t *enp;
 843
 844         /* Clean up transmit state. */
 845         sfxge_tx_fini(sc);
 846
 847         /* Clean up receive state. */
 848         sfxge_rx_fini(sc);
 849
 850         /* Clean up port state. */
 851         sfxge_port_fini(sc);
 852
 853         /* Clean up event processing state. */
 854         sfxge_ev_fini(sc);
 855
 856         /* Clean up interrupts. */
 857         sfxge_intr_fini(sc);
 858
 859         /* Tear down common code subsystems. */
 860         efx_nic_reset(sc->enp);
 861         efx_vpd_fini(sc->enp);
 862         efx_nvram_fini(sc->enp);
 863         efx_nic_unprobe(sc->enp);
 864
 865         /* Tear down MCDI. */
 866         sfxge_mcdi_fini(sc);
 867
 868         /* Destroy common code context. */
 869         enp = sc->enp;
 870         sc->enp = NULL;
 871         efx_nic_destroy(enp);
 872
 873         /* Free DMA memory. */
 874         sfxge_dma_fini(sc);
 875
 876         /* Free mapped BARs. */
 877         sfxge_bar_fini(sc);
 878
 879         (void) pci_disable_busmaster(sc->dev);
 880
 881         taskqueue_drain(taskqueue_thread, &sc->task_reset);
 882
 883         /* Destroy the softc lock. */
 884         SFXGE_ADAPTER_LOCK_DESTROY(sc);
 885 }
 886
 887 static int
 888 sfxge_vpd_handler(SYSCTL_HANDLER_ARGS)
 889 {
 890         struct sfxge_softc *sc = arg1;
 891         efx_vpd_value_t value;
 892         int rc;
 893
 894         value.evv_tag = arg2 >> 16;
 895         value.evv_keyword = arg2 & 0xffff;
 896         if ((rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value))
 897             != 0)
 898                 return (rc);
 899
 900         return (SYSCTL_OUT(req, value.evv_value, value.evv_length));
 901 }
 902
 903 static void
 904 sfxge_vpd_try_add(struct sfxge_softc *sc, struct sysctl_oid_list *list,
 905                   efx_vpd_tag_t tag, const char *keyword)
 906 {
 907         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
 908         efx_vpd_value_t value;
 909
 910         /* Check whether VPD tag/keyword is present */
 911         value.evv_tag = tag;
 912         value.evv_keyword = EFX_VPD_KEYWORD(keyword[0], keyword[1]);
 913         if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) != 0)
 914                 return;
 915
 916         SYSCTL_ADD_PROC(
 917                 ctx, list, OID_AUTO, keyword, CTLTYPE_STRING|CTLFLAG_RD,
 918                 sc, tag << 16 | EFX_VPD_KEYWORD(keyword[0], keyword[1]),
 919                 sfxge_vpd_handler, "A", "");
 920 }
 921
 922 static int
 923 sfxge_vpd_init(struct sfxge_softc *sc)
 924 {
 925         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
 926         struct sysctl_oid *vpd_node;
 927         struct sysctl_oid_list *vpd_list;
 928         char keyword[3];
 929         efx_vpd_value_t value;
 930         int rc;
 931
 932         if ((rc = efx_vpd_size(sc->enp, &sc->vpd_size)) != 0) {
 933                 /*
 934                  * Unpriviledged functions deny VPD access.
 935                  * Simply skip VPD in this case.
 936                  */
 937                 if (rc == EACCES)
 938                         goto done;
 939                 goto fail;
 940         }
 941         sc->vpd_data = malloc(sc->vpd_size, M_SFXGE, M_WAITOK);
 942         if ((rc = efx_vpd_read(sc->enp, sc->vpd_data, sc->vpd_size)) != 0)
 943                 goto fail2;
 944
 945         /* Copy ID (product name) into device description, and log it. */
 946         value.evv_tag = EFX_VPD_ID;
 947         if (efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value) == 0) {
 948                 value.evv_value[value.evv_length] = 0;
 949                 device_set_desc_copy(sc->dev, value.evv_value);
 950                 device_printf(sc->dev, "%s\n", value.evv_value);
 951         }
 952
 953         vpd_node = SYSCTL_ADD_NODE(
 954                 ctx, SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
 955                 OID_AUTO, "vpd", CTLFLAG_RD, NULL, "Vital Product Data");
 956         vpd_list = SYSCTL_CHILDREN(vpd_node);
 957
 958         /* Add sysctls for all expected and any vendor-defined keywords. */
 959         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "PN");
 960         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "EC");
 961         sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, "SN");
 962         keyword[0] = 'V';
 963         keyword[2] = 0;
 964         for (keyword[1] = '0'; keyword[1] <= '9'; keyword[1]++)
 965                 sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword);
 966         for (keyword[1] = 'A'; keyword[1] <= 'Z'; keyword[1]++)
 967                 sfxge_vpd_try_add(sc, vpd_list, EFX_VPD_RO, keyword);
 968
 969 done:
 970         return (0);
 971
 972 fail2:
 973         free(sc->vpd_data, M_SFXGE);
 974 fail:
 975         return (rc);
 976 }
 977
 978 static void
 979 sfxge_vpd_fini(struct sfxge_softc *sc)
 980 {
 981         free(sc->vpd_data, M_SFXGE);
 982 }
 983
 984 static void
 985 sfxge_reset(void *arg, int npending)
 986 {
 987         struct sfxge_softc *sc;
 988         int rc;
 989         unsigned attempt;
 990
 991         (void)npending;
 992
 993         sc = (struct sfxge_softc *)arg;
 994
 995         SFXGE_ADAPTER_LOCK(sc);
 996
 997         if (sc->init_state != SFXGE_STARTED)
 998                 goto done;
 999
1000         sfxge_stop(sc);
1001         efx_nic_reset(sc->enp);
1002         for (attempt = 0; attempt < 3; ++attempt) {
1003                 if ((rc = sfxge_start(sc)) == 0)
1004                         goto done;
1005
1006                 device_printf(sc->dev, "start on reset failed (%d)\n", rc);
1007                 DELAY(100000);
1008         }
1009
1010         device_printf(sc->dev, "reset failed; interface is now stopped\n");
1011
1012 done:
1013         SFXGE_ADAPTER_UNLOCK(sc);
1014 }
1015
1016 void
1017 sfxge_schedule_reset(struct sfxge_softc *sc)
1018 {
1019         taskqueue_enqueue(taskqueue_thread, &sc->task_reset);
1020 }
1021
1022 static int
1023 sfxge_attach(device_t dev)
1024 {
1025         struct sfxge_softc *sc;
1026         struct ifnet *ifp;
1027         int error;
1028
1029         sc = device_get_softc(dev);
1030         sc->dev = dev;
1031
1032         /* Allocate ifnet. */
1033         ifp = if_alloc(IFT_ETHER);
1034         if (ifp == NULL) {
1035                 device_printf(dev, "Couldn't allocate ifnet\n");
1036                 error = ENOMEM;
1037                 goto fail;
1038         }
1039         sc->ifnet = ifp;
1040
1041         /* Initialize hardware. */
1042         DBGPRINT(sc->dev, "create nic");
1043         if ((error = sfxge_create(sc)) != 0)
1044                 goto fail2;
1045
1046         /* Create the ifnet for the port. */
1047         DBGPRINT(sc->dev, "init ifnet");
1048         if ((error = sfxge_ifnet_init(ifp, sc)) != 0)
1049                 goto fail3;
1050
1051         DBGPRINT(sc->dev, "init vpd");
1052         if ((error = sfxge_vpd_init(sc)) != 0)
1053                 goto fail4;
1054
1055         /*
1056          * NIC is initialized inside sfxge_create() and kept inialized
1057          * to be able to initialize port to discover media types in
1058          * sfxge_ifnet_init().
1059          */
1060         efx_nic_fini(sc->enp);
1061
1062         sc->init_state = SFXGE_REGISTERED;
1063
1064         DBGPRINT(sc->dev, "success");
1065         return (0);
1066
1067 fail4:
1068         sfxge_ifnet_fini(ifp);
1069 fail3:
1070         efx_nic_fini(sc->enp);
1071         sfxge_destroy(sc);
1072
1073 fail2:
1074         if_free(sc->ifnet);
1075
1076 fail:
1077         DBGPRINT(sc->dev, "failed %d", error);
1078         return (error);
1079 }
1080
1081 static int
1082 sfxge_detach(device_t dev)
1083 {
1084         struct sfxge_softc *sc;
1085
1086         sc = device_get_softc(dev);
1087
1088         sfxge_vpd_fini(sc);
1089
1090         /* Destroy the ifnet. */
1091         sfxge_ifnet_fini(sc->ifnet);
1092
1093         /* Tear down hardware. */
1094         sfxge_destroy(sc);
1095
1096         return (0);
1097 }
1098
1099 static int
1100 sfxge_probe(device_t dev)
1101 {
1102         uint16_t pci_vendor_id;
1103         uint16_t pci_device_id;
1104         efx_family_t family;
1105         int rc;
1106
1107         pci_vendor_id = pci_get_vendor(dev);
1108         pci_device_id = pci_get_device(dev);
1109
1110         DBGPRINT(dev, "PCI ID %04x:%04x", pci_vendor_id, pci_device_id);
1111         rc = efx_family(pci_vendor_id, pci_device_id, &family);
1112         if (rc != 0) {
1113                 DBGPRINT(dev, "efx_family fail %d", rc);
1114                 return (ENXIO);
1115         }
1116
1117         if (family == EFX_FAMILY_SIENA) {
1118                 device_set_desc(dev, "Solarflare SFC9000 family");
1119                 return (0);
1120         }
1121
1122         if (family == EFX_FAMILY_HUNTINGTON) {
1123                 device_set_desc(dev, "Solarflare SFC9100 family");
1124                 return (0);
1125         }
1126
1127         DBGPRINT(dev, "impossible controller family %d", family);
1128         return (ENXIO);
1129 }
1130
1131 static device_method_t sfxge_methods[] = {
1132         DEVMETHOD(device_probe,         sfxge_probe),
1133         DEVMETHOD(device_attach,        sfxge_attach),
1134         DEVMETHOD(device_detach,        sfxge_detach),
1135
1136         DEVMETHOD_END
1137 };
1138
1139 static devclass_t sfxge_devclass;
1140
1141 static driver_t sfxge_driver = {
1142         "sfxge",
1143         sfxge_methods,
1144         sizeof(struct sfxge_softc)
1145 };
1146
1147 DRIVER_MODULE(sfxge, pci, sfxge_driver, sfxge_devclass, 0, 0);