sys/dev/sfxge/sfxge.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2010-2016 Solarflare Communications Inc.
   5  * All rights reserved.
   6  *
   7  * This software was developed in part by Philip Paeps under contract for
   8  * Solarflare Communications, Inc.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions are met:
  12  *
  13  * 1. Redistributions of source code must retain the above copyright notice,
  14  *    this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright notice,
  16  *    this list of conditions and the following disclaimer in the documentation
  17  *    and/or other materials provided with the distribution.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  20  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  21  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  22  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
  23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
  26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
  29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  30  *
  31  * The views and conclusions contained in the software and documentation are
  32  * those of the authors and should not be interpreted as representing official
  33  * policies, either expressed or implied, of the FreeBSD Project.
  34  */
  35
  36 #include <sys/cdefs.h>
  37 __FBSDID("$FreeBSD$");
  38
  39 #include "opt_rss.h"
  40
  41 #include <sys/param.h>
  42 #include <sys/kernel.h>
  43 #include <sys/bus.h>
  44 #include <sys/rman.h>
  45 #include <sys/lock.h>
  46 #include <sys/module.h>
  47 #include <sys/mutex.h>
  48 #include <sys/smp.h>
  49 #include <sys/socket.h>
  50 #include <sys/taskqueue.h>
  51 #include <sys/sockio.h>
  52 #include <sys/sysctl.h>
  53 #include <sys/priv.h>
  54 #include <sys/syslog.h>
  55
  56 #include <dev/pci/pcireg.h>
  57 #include <dev/pci/pcivar.h>
  58
  59 #include <net/ethernet.h>
  60 #include <net/if.h>
  61 #include <net/if_var.h>
  62 #include <net/if_media.h>
  63 #include <net/if_types.h>
  64
  65 #ifdef RSS
  66 #include <net/rss_config.h>
  67 #endif
  68
  69 #include "common/efx.h"
  70
  71 #include "sfxge.h"
  72 #include "sfxge_rx.h"
  73 #include "sfxge_ioc.h"
  74 #include "sfxge_version.h"
  75
  76 #define SFXGE_CAP (IFCAP_VLAN_MTU | IFCAP_VLAN_HWCSUM |                 \
  77                    IFCAP_RXCSUM | IFCAP_TXCSUM |                        \
  78                    IFCAP_RXCSUM_IPV6 | IFCAP_TXCSUM_IPV6 |              \
  79                    IFCAP_TSO4 | IFCAP_TSO6 |                            \
  80                    IFCAP_JUMBO_MTU |                                    \
  81                    IFCAP_VLAN_HWTSO | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  82 #define SFXGE_CAP_ENABLE SFXGE_CAP
  83 #define SFXGE_CAP_FIXED (IFCAP_VLAN_MTU |                               \
  84                          IFCAP_JUMBO_MTU | IFCAP_LINKSTATE | IFCAP_HWSTATS)
  85
  86 MALLOC_DEFINE(M_SFXGE, "sfxge", "Solarflare 10GigE driver");
  87
  88
  89 SYSCTL_NODE(_hw, OID_AUTO, sfxge, CTLFLAG_RD, 0,
  90             "SFXGE driver parameters");
  91
  92 #define SFXGE_PARAM_RX_RING     SFXGE_PARAM(rx_ring)
  93 static int sfxge_rx_ring_entries = SFXGE_NDESCS;
  94 TUNABLE_INT(SFXGE_PARAM_RX_RING, &sfxge_rx_ring_entries);
  95 SYSCTL_INT(_hw_sfxge, OID_AUTO, rx_ring, CTLFLAG_RDTUN,
  96            &sfxge_rx_ring_entries, 0,
  97            "Maximum number of descriptors in a receive ring");
  98
  99 #define SFXGE_PARAM_TX_RING     SFXGE_PARAM(tx_ring)
 100 static int sfxge_tx_ring_entries = SFXGE_NDESCS;
 101 TUNABLE_INT(SFXGE_PARAM_TX_RING, &sfxge_tx_ring_entries);
 102 SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_ring, CTLFLAG_RDTUN,
 103            &sfxge_tx_ring_entries, 0,
 104            "Maximum number of descriptors in a transmit ring");
 105
 106 #define SFXGE_PARAM_RESTART_ATTEMPTS    SFXGE_PARAM(restart_attempts)
 107 static int sfxge_restart_attempts = 3;
 108 TUNABLE_INT(SFXGE_PARAM_RESTART_ATTEMPTS, &sfxge_restart_attempts);
 109 SYSCTL_INT(_hw_sfxge, OID_AUTO, restart_attempts, CTLFLAG_RDTUN,
 110            &sfxge_restart_attempts, 0,
 111            "Maximum number of attempts to bring interface up after reset");
 112
 113 #if EFSYS_OPT_MCDI_LOGGING
 114 #define SFXGE_PARAM_MCDI_LOGGING        SFXGE_PARAM(mcdi_logging)
 115 static int sfxge_mcdi_logging = 0;
 116 TUNABLE_INT(SFXGE_PARAM_MCDI_LOGGING, &sfxge_mcdi_logging);
 117 #endif
 118
 119 static void
 120 sfxge_reset(void *arg, int npending);
 121
 122 static int
 123 sfxge_estimate_rsrc_limits(struct sfxge_softc *sc)
 124 {
 125         efx_drv_limits_t limits;
 126         int rc;
 127         unsigned int evq_max;
 128         uint32_t evq_allocated;
 129         uint32_t rxq_allocated;
 130         uint32_t txq_allocated;
 131
 132         /*
 133          * Limit the number of event queues to:
 134          *  - number of CPUs
 135          *  - hardwire maximum RSS channels
 136          *  - administratively specified maximum RSS channels
 137          */
 138 #ifdef RSS
 139         /*
 140          * Avoid extra limitations so that the number of queues
 141          * may be configured at administrator's will
 142          */
 143         evq_max = MIN(MAX(rss_getnumbuckets(), 1), EFX_MAXRSS);
 144 #else
 145         evq_max = MIN(mp_ncpus, EFX_MAXRSS);
 146 #endif
 147         if (sc->max_rss_channels > 0)
 148                 evq_max = MIN(evq_max, sc->max_rss_channels);
 149
 150         memset(&limits, 0, sizeof(limits));
 151
 152         limits.edl_min_evq_count = 1;
 153         limits.edl_max_evq_count = evq_max;
 154         limits.edl_min_txq_count = SFXGE_TXQ_NTYPES;
 155         limits.edl_max_txq_count = evq_max + SFXGE_TXQ_NTYPES - 1;
 156         limits.edl_min_rxq_count = 1;
 157         limits.edl_max_rxq_count = evq_max;
 158
 159         efx_nic_set_drv_limits(sc->enp, &limits);
 160
 161         if ((rc = efx_nic_init(sc->enp)) != 0)
 162                 return (rc);
 163
 164         rc = efx_nic_get_vi_pool(sc->enp, &evq_allocated, &rxq_allocated,
 165                                  &txq_allocated);
 166         if (rc != 0) {
 167                 efx_nic_fini(sc->enp);
 168                 return (rc);
 169         }
 170
 171         KASSERT(txq_allocated >= SFXGE_TXQ_NTYPES,
 172                 ("txq_allocated < SFXGE_TXQ_NTYPES"));
 173
 174         sc->evq_max = MIN(evq_allocated, evq_max);
 175         sc->evq_max = MIN(rxq_allocated, sc->evq_max);
 176         sc->evq_max = MIN(txq_allocated - (SFXGE_TXQ_NTYPES - 1),
 177                           sc->evq_max);
 178
 179         KASSERT(sc->evq_max <= evq_max,
 180                 ("allocated more than maximum requested"));
 181
 182 #ifdef RSS
 183         if (sc->evq_max < rss_getnumbuckets())
 184                 device_printf(sc->dev, "The number of allocated queues (%u) "
 185                               "is less than the number of RSS buckets (%u); "
 186                               "performance degradation might be observed",
 187                               sc->evq_max, rss_getnumbuckets());
 188 #endif
 189
 190         /*
 191          * NIC is kept initialized in the case of success to be able to
 192          * initialize port to find out media types.
 193          */
 194         return (0);
 195 }
 196
 197 static int
 198 sfxge_set_drv_limits(struct sfxge_softc *sc)
 199 {
 200         efx_drv_limits_t limits;
 201
 202         memset(&limits, 0, sizeof(limits));
 203
 204         /* Limits are strict since take into account initial estimation */
 205         limits.edl_min_evq_count = limits.edl_max_evq_count =
 206             sc->intr.n_alloc;
 207         limits.edl_min_txq_count = limits.edl_max_txq_count =
 208             sc->intr.n_alloc + SFXGE_TXQ_NTYPES - 1;
 209         limits.edl_min_rxq_count = limits.edl_max_rxq_count =
 210             sc->intr.n_alloc;
 211
 212         return (efx_nic_set_drv_limits(sc->enp, &limits));
 213 }
 214
 215 static int
 216 sfxge_start(struct sfxge_softc *sc)
 217 {
 218         int rc;
 219
 220         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 221
 222         if (sc->init_state == SFXGE_STARTED)
 223                 return (0);
 224
 225         if (sc->init_state != SFXGE_REGISTERED) {
 226                 rc = EINVAL;
 227                 goto fail;
 228         }
 229
 230         /* Set required resource limits */
 231         if ((rc = sfxge_set_drv_limits(sc)) != 0)
 232                 goto fail;
 233
 234         if ((rc = efx_nic_init(sc->enp)) != 0)
 235                 goto fail;
 236
 237         /* Start processing interrupts. */
 238         if ((rc = sfxge_intr_start(sc)) != 0)
 239                 goto fail2;
 240
 241         /* Start processing events. */
 242         if ((rc = sfxge_ev_start(sc)) != 0)
 243                 goto fail3;
 244
 245         /* Fire up the port. */
 246         if ((rc = sfxge_port_start(sc)) != 0)
 247                 goto fail4;
 248
 249         /* Start the receiver side. */
 250         if ((rc = sfxge_rx_start(sc)) != 0)
 251                 goto fail5;
 252
 253         /* Start the transmitter side. */
 254         if ((rc = sfxge_tx_start(sc)) != 0)
 255                 goto fail6;
 256
 257         sc->init_state = SFXGE_STARTED;
 258
 259         /* Tell the stack we're running. */
 260         sc->ifnet->if_drv_flags |= IFF_DRV_RUNNING;
 261         sc->ifnet->if_drv_flags &= ~IFF_DRV_OACTIVE;
 262
 263         return (0);
 264
 265 fail6:
 266         sfxge_rx_stop(sc);
 267
 268 fail5:
 269         sfxge_port_stop(sc);
 270
 271 fail4:
 272         sfxge_ev_stop(sc);
 273
 274 fail3:
 275         sfxge_intr_stop(sc);
 276
 277 fail2:
 278         efx_nic_fini(sc->enp);
 279
 280 fail:
 281         device_printf(sc->dev, "sfxge_start: %d\n", rc);
 282
 283         return (rc);
 284 }
 285
 286 static void
 287 sfxge_if_init(void *arg)
 288 {
 289         struct sfxge_softc *sc;
 290
 291         sc = (struct sfxge_softc *)arg;
 292
 293         SFXGE_ADAPTER_LOCK(sc);
 294         (void)sfxge_start(sc);
 295         SFXGE_ADAPTER_UNLOCK(sc);
 296 }
 297
 298 static void
 299 sfxge_stop(struct sfxge_softc *sc)
 300 {
 301         SFXGE_ADAPTER_LOCK_ASSERT_OWNED(sc);
 302
 303         if (sc->init_state != SFXGE_STARTED)
 304                 return;
 305
 306         sc->init_state = SFXGE_REGISTERED;
 307
 308         /* Stop the transmitter. */
 309         sfxge_tx_stop(sc);
 310
 311         /* Stop the receiver. */
 312         sfxge_rx_stop(sc);
 313
 314         /* Stop the port. */
 315         sfxge_port_stop(sc);
 316
 317         /* Stop processing events. */
 318         sfxge_ev_stop(sc);
 319
 320         /* Stop processing interrupts. */
 321         sfxge_intr_stop(sc);
 322
 323         efx_nic_fini(sc->enp);
 324
 325         sc->ifnet->if_drv_flags &= ~IFF_DRV_RUNNING;
 326 }
 327
 328
 329 static int
 330 sfxge_vpd_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 331 {
 332         efx_vpd_value_t value;
 333         int rc = 0;
 334
 335         switch (ioc->u.vpd.op) {
 336         case SFXGE_VPD_OP_GET_KEYWORD:
 337                 value.evv_tag = ioc->u.vpd.tag;
 338                 value.evv_keyword = ioc->u.vpd.keyword;
 339                 rc = efx_vpd_get(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 340                 if (rc != 0)
 341                         break;
 342                 ioc->u.vpd.len = MIN(ioc->u.vpd.len, value.evv_length);
 343                 if (ioc->u.vpd.payload != 0) {
 344                         rc = copyout(value.evv_value, ioc->u.vpd.payload,
 345                                      ioc->u.vpd.len);
 346                 }
 347                 break;
 348         case SFXGE_VPD_OP_SET_KEYWORD:
 349                 if (ioc->u.vpd.len > sizeof(value.evv_value))
 350                         return (EINVAL);
 351                 value.evv_tag = ioc->u.vpd.tag;
 352                 value.evv_keyword = ioc->u.vpd.keyword;
 353                 value.evv_length = ioc->u.vpd.len;
 354                 rc = copyin(ioc->u.vpd.payload, value.evv_value, value.evv_length);
 355                 if (rc != 0)
 356                         break;
 357                 rc = efx_vpd_set(sc->enp, sc->vpd_data, sc->vpd_size, &value);
 358                 if (rc != 0)
 359                         break;
 360                 rc = efx_vpd_verify(sc->enp, sc->vpd_data, sc->vpd_size);
 361                 if (rc != 0)
 362                         break;
 363                 rc = efx_vpd_write(sc->enp, sc->vpd_data, sc->vpd_size);
 364                 break;
 365         default:
 366                 rc = EOPNOTSUPP;
 367                 break;
 368         }
 369
 370         return (rc);
 371 }
 372
 373 static int
 374 sfxge_private_ioctl(struct sfxge_softc *sc, sfxge_ioc_t *ioc)
 375 {
 376         switch (ioc->op) {
 377         case SFXGE_MCDI_IOC:
 378                 return (sfxge_mcdi_ioctl(sc, ioc));
 379         case SFXGE_NVRAM_IOC:
 380                 return (sfxge_nvram_ioctl(sc, ioc));
 381         case SFXGE_VPD_IOC:
 382                 return (sfxge_vpd_ioctl(sc, ioc));
 383         default:
 384                 return (EOPNOTSUPP);
 385         }
 386 }
 387
 388
 389 static int
 390 sfxge_if_ioctl(struct ifnet *ifp, unsigned long command, caddr_t data)
 391 {
 392         struct sfxge_softc *sc;
 393         struct ifreq *ifr;
 394         sfxge_ioc_t ioc;
 395         int error;
 396
 397         ifr = (struct ifreq *)data;
 398         sc = ifp->if_softc;
 399         error = 0;
 400
 401         switch (command) {
 402         case SIOCSIFFLAGS:
 403                 SFXGE_ADAPTER_LOCK(sc);
 404                 if (ifp->if_flags & IFF_UP) {
 405                         if (ifp->if_drv_flags & IFF_DRV_RUNNING) {
 406                                 if ((ifp->if_flags ^ sc->if_flags) &
 407                                     (IFF_PROMISC | IFF_ALLMULTI)) {
 408                                         sfxge_mac_filter_set(sc);
 409                                 }
 410                         } else
 411                                 sfxge_start(sc);
 412                 } else
 413                         if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 414                                 sfxge_stop(sc);
 415                 sc->if_flags = ifp->if_flags;
 416                 SFXGE_ADAPTER_UNLOCK(sc);
 417                 break;
 418         case SIOCSIFMTU:
 419                 if (ifr->ifr_mtu == ifp->if_mtu) {
 420                         /* Nothing to do */
 421                         error = 0;
 422                 } else if (ifr->ifr_mtu > SFXGE_MAX_MTU) {
 423                         error = EINVAL;
 424                 } else if (!(ifp->if_drv_flags & IFF_DRV_RUNNING)) {
 425                         ifp->if_mtu = ifr->ifr_mtu;
 426                         error = 0;
 427                 } else {
 428                         /* Restart required */
 429                         SFXGE_ADAPTER_LOCK(sc);
 430                         sfxge_stop(sc);
 431                         ifp->if_mtu = ifr->ifr_mtu;
 432                         error = sfxge_start(sc);
 433                         SFXGE_ADAPTER_UNLOCK(sc);
 434                         if (error != 0) {
 435                                 ifp->if_flags &= ~IFF_UP;
 436                                 ifp->if_drv_flags &= ~IFF_DRV_RUNNING;
 437                                 if_down(ifp);
 438                         }
 439                 }
 440                 break;
 441         case SIOCADDMULTI:
 442         case SIOCDELMULTI:
 443                 if (ifp->if_drv_flags & IFF_DRV_RUNNING)
 444                         sfxge_mac_filter_set(sc);
 445                 break;
 446         case SIOCSIFCAP:
 447         {
 448                 int reqcap = ifr->ifr_reqcap;
 449                 int capchg_mask;
 450
 451                 SFXGE_ADAPTER_LOCK(sc);
 452
 453                 /* Capabilities to be changed in accordance with request */
 454                 capchg_mask = ifp->if_capenable ^ reqcap;
 455
 456                 /*
 457                  * The networking core already rejects attempts to
 458                  * enable capabilities we don't have.  We still have
 459                  * to reject attempts to disable capabilities that we
 460                  * can't (yet) disable.
 461                  */
 462                 KASSERT((reqcap & ~ifp->if_capabilities) == 0,
 463                     ("Unsupported capabilities 0x%x requested 0x%x vs "
 464                      "supported 0x%x",
 465                      reqcap & ~ifp->if_capabilities,
 466                      reqcap , ifp->if_capabilities));
 467                 if (capchg_mask & SFXGE_CAP_FIXED) {
 468                         error = EINVAL;
 469                         SFXGE_ADAPTER_UNLOCK(sc);
 470                         break;
 471                 }
 472
 473                 /* Check request before any changes */
 474                 if ((capchg_mask & IFCAP_TSO4) &&
 475                     (reqcap & (IFCAP_TSO4 | IFCAP_TXCSUM)) == IFCAP_TSO4) {
 476                         error = EAGAIN;
 477                         SFXGE_ADAPTER_UNLOCK(sc);
 478                         if_printf(ifp, "enable txcsum before tso4\n");
 479                         break;
 480                 }
 481                 if ((capchg_mask & IFCAP_TSO6) &&
 482                     (reqcap & (IFCAP_TSO6 | IFCAP_TXCSUM_IPV6)) == IFCAP_TSO6) {
 483                         error = EAGAIN;
 484                         SFXGE_ADAPTER_UNLOCK(sc);
 485                         if_printf(ifp, "enable txcsum6 before tso6\n");
 486                         break;
 487                 }
 488
 489                 if (reqcap & IFCAP_TXCSUM) {
 490                         ifp->if_hwassist |= (CSUM_IP | CSUM_TCP | CSUM_UDP);
 491                 } else {
 492                         ifp->if_hwassist &= ~(CSUM_IP | CSUM_TCP | CSUM_UDP);
 493                         if (reqcap & IFCAP_TSO4) {
 494                                 reqcap &= ~IFCAP_TSO4;
 495                                 if_printf(ifp,
 496                                     "tso4 disabled due to -txcsum\n");
 497                         }
 498                 }
 499                 if (reqcap & IFCAP_TXCSUM_IPV6) {
 500                         ifp->if_hwassist |= (CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 501                 } else {
 502                         ifp->if_hwassist &= ~(CSUM_TCP_IPV6 | CSUM_UDP_IPV6);
 503                         if (reqcap & IFCAP_TSO6) {
 504                                 reqcap &= ~IFCAP_TSO6;
 505                                 if_printf(ifp,
 506                                     "tso6 disabled due to -txcsum6\n");
 507                         }
 508                 }
 509
 510                 /*
 511                  * The kernel takes both IFCAP_TSOx and CSUM_TSO into
 512                  * account before using TSO. So, we do not touch
 513                  * checksum flags when IFCAP_TSOx is modified.
 514                  * Note that CSUM_TSO is (CSUM_IP_TSO|CSUM_IP6_TSO),
 515                  * but both bits are set in IPv4 and IPv6 mbufs.
 516                  */
 517
 518                 ifp->if_capenable = reqcap;
 519
 520                 SFXGE_ADAPTER_UNLOCK(sc);
 521                 break;
 522         }
 523         case SIOCSIFMEDIA:
 524         case SIOCGIFMEDIA:
 525                 error = ifmedia_ioctl(ifp, ifr, &sc->media, command);
 526                 break;
 527 #ifdef SIOCGI2C
 528         case SIOCGI2C:
 529         {
 530                 struct ifi2creq i2c;
 531
 532                 error = copyin(ifr->ifr_data, &i2c, sizeof(i2c));
 533                 if (error != 0)
 534                         break;
 535
 536                 if (i2c.len > sizeof(i2c.data)) {
 537                         error = EINVAL;
 538                         break;
 539                 }
 540
 541                 SFXGE_ADAPTER_LOCK(sc);
 542                 error = efx_phy_module_get_info(sc->enp, i2c.dev_addr,
 543                                                 i2c.offset, i2c.len,
 544                                                 &i2c.data[0]);
 545                 SFXGE_ADAPTER_UNLOCK(sc);
 546                 if (error == 0)
 547                         error = copyout(&i2c, ifr->ifr_data, sizeof(i2c));
 548                 break;
 549         }
 550 #endif
 551         case SIOCGPRIVATE_0:
 552                 error = priv_check(curthread, PRIV_DRIVER);
 553                 if (error != 0)
 554                         break;
 555                 error = copyin(ifr->ifr_data, &ioc, sizeof(ioc));
 556                 if (error != 0)
 557                         return (error);
 558                 error = sfxge_private_ioctl(sc, &ioc);
 559                 if (error == 0) {
 560                         error = copyout(&ioc, ifr->ifr_data, sizeof(ioc));
 561                 }
 562                 break;
 563         default:
 564                 error = ether_ioctl(ifp, command, data);
 565         }
 566
 567         return (error);
 568 }
 569
 570 static void
 571 sfxge_ifnet_fini(struct ifnet *ifp)
 572 {
 573         struct sfxge_softc *sc = ifp->if_softc;
 574
 575         SFXGE_ADAPTER_LOCK(sc);
 576         sfxge_stop(sc);
 577         SFXGE_ADAPTER_UNLOCK(sc);
 578
 579         ifmedia_removeall(&sc->media);
 580         ether_ifdetach(ifp);
 581         if_free(ifp);
 582 }
 583
 584 static int
 585 sfxge_ifnet_init(struct ifnet *ifp, struct sfxge_softc *sc)
 586 {
 587         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sc->enp);
 588         device_t dev;
 589         int rc;
 590
 591         dev = sc->dev;
 592         sc->ifnet = ifp;
 593
 594         if_initname(ifp, device_get_name(dev), device_get_unit(dev));
 595         ifp->if_init = sfxge_if_init;
 596         ifp->if_softc = sc;
 597         ifp->if_flags = IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST;
 598         ifp->if_ioctl = sfxge_if_ioctl;
 599
 600         ifp->if_capabilities = SFXGE_CAP;
 601         ifp->if_capenable = SFXGE_CAP_ENABLE;
 602         ifp->if_hw_tsomax = SFXGE_TSO_MAX_SIZE;
 603         ifp->if_hw_tsomaxsegcount = SFXGE_TX_MAPPING_MAX_SEG;
 604         ifp->if_hw_tsomaxsegsize = PAGE_SIZE;
 605
 606 #ifdef SFXGE_LRO
 607         ifp->if_capabilities |= IFCAP_LRO;
 608         ifp->if_capenable |= IFCAP_LRO;
 609 #endif
 610
 611         if (encp->enc_hw_tx_insert_vlan_enabled) {
 612                 ifp->if_capabilities |= IFCAP_VLAN_HWTAGGING;
 613                 ifp->if_capenable |= IFCAP_VLAN_HWTAGGING;
 614         }
 615         ifp->if_hwassist = CSUM_TCP | CSUM_UDP | CSUM_IP | CSUM_TSO |
 616                            CSUM_TCP_IPV6 | CSUM_UDP_IPV6;
 617
 618         ether_ifattach(ifp, encp->enc_mac_addr);
 619
 620         ifp->if_transmit = sfxge_if_transmit;
 621         ifp->if_qflush = sfxge_if_qflush;
 622
 623         ifp->if_get_counter = sfxge_get_counter;
 624
 625         DBGPRINT(sc->dev, "ifmedia_init");
 626         if ((rc = sfxge_port_ifmedia_init(sc)) != 0)
 627                 goto fail;
 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         /* Initialize MCDI to talk to the microcontroller. */
 743         DBGPRINT(sc->dev, "mcdi_init...");
 744         if ((error = sfxge_mcdi_init(sc)) != 0)
 745                 goto fail4;
 746
 747         /* Probe the NIC and build the configuration data area. */
 748         DBGPRINT(sc->dev, "nic_probe...");
 749         if ((error = efx_nic_probe(enp)) != 0)
 750                 goto fail5;
 751
 752         if (!ISP2(sfxge_rx_ring_entries) ||
 753             (sfxge_rx_ring_entries < EFX_RXQ_MINNDESCS) ||
 754             (sfxge_rx_ring_entries > EFX_RXQ_MAXNDESCS)) {
 755                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 756                     SFXGE_PARAM_RX_RING, sfxge_rx_ring_entries,
 757                     EFX_RXQ_MINNDESCS, EFX_RXQ_MAXNDESCS);
 758                 error = EINVAL;
 759                 goto fail_rx_ring_entries;
 760         }
 761         sc->rxq_entries = sfxge_rx_ring_entries;
 762
 763         if (!ISP2(sfxge_tx_ring_entries) ||
 764             (sfxge_tx_ring_entries < EFX_TXQ_MINNDESCS) ||
 765             (sfxge_tx_ring_entries > EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)))) {
 766                 log(LOG_ERR, "%s=%d must be power of 2 from %u to %u",
 767                     SFXGE_PARAM_TX_RING, sfxge_tx_ring_entries,
 768                     EFX_TXQ_MINNDESCS, EFX_TXQ_MAXNDESCS(efx_nic_cfg_get(enp)));
 769                 error = EINVAL;
 770                 goto fail_tx_ring_entries;
 771         }
 772         sc->txq_entries = sfxge_tx_ring_entries;
 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 fail_tx_ring_entries:
 867 fail_rx_ring_entries:
 868         efx_nic_unprobe(enp);
 869
 870 fail5:
 871         sfxge_mcdi_fini(sc);
 872
 873 fail4:
 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         if (family == EFX_FAMILY_MEDFORD) {
1178                 device_set_desc(dev, "Solarflare SFC9200 family");
1179                 return (0);
1180         }
1181
1182         DBGPRINT(dev, "impossible controller family %d", family);
1183         return (ENXIO);
1184 }
1185
1186 static device_method_t sfxge_methods[] = {
1187         DEVMETHOD(device_probe,         sfxge_probe),
1188         DEVMETHOD(device_attach,        sfxge_attach),
1189         DEVMETHOD(device_detach,        sfxge_detach),
1190
1191         DEVMETHOD_END
1192 };
1193
1194 static devclass_t sfxge_devclass;
1195
1196 static driver_t sfxge_driver = {
1197         "sfxge",
1198         sfxge_methods,
1199         sizeof(struct sfxge_softc)
1200 };
1201
1202 DRIVER_MODULE(sfxge, pci, sfxge_driver, sfxge_devclass, 0, 0);