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