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