drivers/net/vmxnet3/vmxnet3_ethdev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2015 Intel Corporation
   3  */
   4
   5 #include <sys/queue.h>
   6 #include <stdio.h>
   7 #include <errno.h>
   8 #include <stdint.h>
   9 #include <string.h>
  10 #include <unistd.h>
  11 #include <stdarg.h>
  12 #include <fcntl.h>
  13 #include <inttypes.h>
  14 #include <rte_byteorder.h>
  15 #include <rte_common.h>
  16 #include <rte_cycles.h>
  17
  18 #include <rte_interrupts.h>
  19 #include <rte_log.h>
  20 #include <rte_debug.h>
  21 #include <rte_pci.h>
  22 #include <rte_bus_pci.h>
  23 #include <rte_branch_prediction.h>
  24 #include <rte_memory.h>
  25 #include <rte_memzone.h>
  26 #include <rte_eal.h>
  27 #include <rte_alarm.h>
  28 #include <rte_ether.h>
  29 #include <rte_ethdev_driver.h>
  30 #include <rte_ethdev_pci.h>
  31 #include <rte_string_fns.h>
  32 #include <rte_malloc.h>
  33 #include <rte_dev.h>
  34
  35 #include "base/vmxnet3_defs.h"
  36
  37 #include "vmxnet3_ring.h"
  38 #include "vmxnet3_logs.h"
  39 #include "vmxnet3_ethdev.h"
  40
  41 #define PROCESS_SYS_EVENTS 0
  42
  43 #define VMXNET3_TX_MAX_SEG      UINT8_MAX
  44
  45 #define VMXNET3_TX_OFFLOAD_CAP          \
  46         (DEV_TX_OFFLOAD_VLAN_INSERT |   \
  47          DEV_TX_OFFLOAD_IPV4_CKSUM |    \
  48          DEV_TX_OFFLOAD_TCP_CKSUM |     \
  49          DEV_TX_OFFLOAD_UDP_CKSUM |     \
  50          DEV_TX_OFFLOAD_TCP_TSO |       \
  51          DEV_TX_OFFLOAD_MULTI_SEGS)
  52
  53 #define VMXNET3_RX_OFFLOAD_CAP          \
  54         (DEV_RX_OFFLOAD_VLAN_STRIP |    \
  55          DEV_RX_OFFLOAD_SCATTER |       \
  56          DEV_RX_OFFLOAD_IPV4_CKSUM |    \
  57          DEV_RX_OFFLOAD_UDP_CKSUM |     \
  58          DEV_RX_OFFLOAD_TCP_CKSUM |     \
  59          DEV_RX_OFFLOAD_TCP_LRO |       \
  60          DEV_RX_OFFLOAD_JUMBO_FRAME)
  61
  62 static int eth_vmxnet3_dev_init(struct rte_eth_dev *eth_dev);
  63 static int eth_vmxnet3_dev_uninit(struct rte_eth_dev *eth_dev);
  64 static int vmxnet3_dev_configure(struct rte_eth_dev *dev);
  65 static int vmxnet3_dev_start(struct rte_eth_dev *dev);
  66 static void vmxnet3_dev_stop(struct rte_eth_dev *dev);
  67 static void vmxnet3_dev_close(struct rte_eth_dev *dev);
  68 static void vmxnet3_dev_set_rxmode(struct vmxnet3_hw *hw, uint32_t feature, int set);
  69 static void vmxnet3_dev_promiscuous_enable(struct rte_eth_dev *dev);
  70 static void vmxnet3_dev_promiscuous_disable(struct rte_eth_dev *dev);
  71 static void vmxnet3_dev_allmulticast_enable(struct rte_eth_dev *dev);
  72 static void vmxnet3_dev_allmulticast_disable(struct rte_eth_dev *dev);
  73 static int __vmxnet3_dev_link_update(struct rte_eth_dev *dev,
  74                                      int wait_to_complete);
  75 static int vmxnet3_dev_link_update(struct rte_eth_dev *dev,
  76                                    int wait_to_complete);
  77 static void vmxnet3_hw_stats_save(struct vmxnet3_hw *hw);
  78 static int vmxnet3_dev_stats_get(struct rte_eth_dev *dev,
  79                                   struct rte_eth_stats *stats);
  80 static int vmxnet3_dev_xstats_get_names(struct rte_eth_dev *dev,
  81                                         struct rte_eth_xstat_name *xstats,
  82                                         unsigned int n);
  83 static int vmxnet3_dev_xstats_get(struct rte_eth_dev *dev,
  84                                   struct rte_eth_xstat *xstats, unsigned int n);
  85 static void vmxnet3_dev_info_get(struct rte_eth_dev *dev,
  86                                  struct rte_eth_dev_info *dev_info);
  87 static const uint32_t *
  88 vmxnet3_dev_supported_ptypes_get(struct rte_eth_dev *dev);
  89 static int vmxnet3_dev_vlan_filter_set(struct rte_eth_dev *dev,
  90                                        uint16_t vid, int on);
  91 static int vmxnet3_dev_vlan_offload_set(struct rte_eth_dev *dev, int mask);
  92 static int vmxnet3_mac_addr_set(struct rte_eth_dev *dev,
  93                                  struct ether_addr *mac_addr);
  94 static void vmxnet3_interrupt_handler(void *param);
  95
  96 int vmxnet3_logtype_init;
  97 int vmxnet3_logtype_driver;
  98
  99 /*
 100  * The set of PCI devices this driver supports
 101  */
 102 #define VMWARE_PCI_VENDOR_ID 0x15AD
 103 #define VMWARE_DEV_ID_VMXNET3 0x07B0
 104 static const struct rte_pci_id pci_id_vmxnet3_map[] = {
 105         { RTE_PCI_DEVICE(VMWARE_PCI_VENDOR_ID, VMWARE_DEV_ID_VMXNET3) },
 106         { .vendor_id = 0, /* sentinel */ },
 107 };
 108
 109 static const struct eth_dev_ops vmxnet3_eth_dev_ops = {
 110         .dev_configure        = vmxnet3_dev_configure,
 111         .dev_start            = vmxnet3_dev_start,
 112         .dev_stop             = vmxnet3_dev_stop,
 113         .dev_close            = vmxnet3_dev_close,
 114         .promiscuous_enable   = vmxnet3_dev_promiscuous_enable,
 115         .promiscuous_disable  = vmxnet3_dev_promiscuous_disable,
 116         .allmulticast_enable  = vmxnet3_dev_allmulticast_enable,
 117         .allmulticast_disable = vmxnet3_dev_allmulticast_disable,
 118         .link_update          = vmxnet3_dev_link_update,
 119         .stats_get            = vmxnet3_dev_stats_get,
 120         .xstats_get_names     = vmxnet3_dev_xstats_get_names,
 121         .xstats_get           = vmxnet3_dev_xstats_get,
 122         .mac_addr_set         = vmxnet3_mac_addr_set,
 123         .dev_infos_get        = vmxnet3_dev_info_get,
 124         .dev_supported_ptypes_get = vmxnet3_dev_supported_ptypes_get,
 125         .vlan_filter_set      = vmxnet3_dev_vlan_filter_set,
 126         .vlan_offload_set     = vmxnet3_dev_vlan_offload_set,
 127         .rx_queue_setup       = vmxnet3_dev_rx_queue_setup,
 128         .rx_queue_release     = vmxnet3_dev_rx_queue_release,
 129         .tx_queue_setup       = vmxnet3_dev_tx_queue_setup,
 130         .tx_queue_release     = vmxnet3_dev_tx_queue_release,
 131 };
 132
 133 struct vmxnet3_xstats_name_off {
 134         char name[RTE_ETH_XSTATS_NAME_SIZE];
 135         unsigned int offset;
 136 };
 137
 138 /* tx_qX_ is prepended to the name string here */
 139 static const struct vmxnet3_xstats_name_off vmxnet3_txq_stat_strings[] = {
 140         {"drop_total",         offsetof(struct vmxnet3_txq_stats, drop_total)},
 141         {"drop_too_many_segs", offsetof(struct vmxnet3_txq_stats, drop_too_many_segs)},
 142         {"drop_tso",           offsetof(struct vmxnet3_txq_stats, drop_tso)},
 143         {"tx_ring_full",       offsetof(struct vmxnet3_txq_stats, tx_ring_full)},
 144 };
 145
 146 /* rx_qX_ is prepended to the name string here */
 147 static const struct vmxnet3_xstats_name_off vmxnet3_rxq_stat_strings[] = {
 148         {"drop_total",           offsetof(struct vmxnet3_rxq_stats, drop_total)},
 149         {"drop_err",             offsetof(struct vmxnet3_rxq_stats, drop_err)},
 150         {"drop_fcs",             offsetof(struct vmxnet3_rxq_stats, drop_fcs)},
 151         {"rx_buf_alloc_failure", offsetof(struct vmxnet3_rxq_stats, rx_buf_alloc_failure)},
 152 };
 153
 154 static const struct rte_memzone *
 155 gpa_zone_reserve(struct rte_eth_dev *dev, uint32_t size,
 156                  const char *post_string, int socket_id,
 157                  uint16_t align, bool reuse)
 158 {
 159         char z_name[RTE_MEMZONE_NAMESIZE];
 160         const struct rte_memzone *mz;
 161
 162         snprintf(z_name, sizeof(z_name), "%s_%d_%s",
 163                  dev->device->driver->name, dev->data->port_id, post_string);
 164
 165         mz = rte_memzone_lookup(z_name);
 166         if (!reuse) {
 167                 if (mz)
 168                         rte_memzone_free(mz);
 169                 return rte_memzone_reserve_aligned(z_name, size, socket_id,
 170                                 RTE_MEMZONE_IOVA_CONTIG, align);
 171         }
 172
 173         if (mz)
 174                 return mz;
 175
 176         return rte_memzone_reserve_aligned(z_name, size, socket_id,
 177                         RTE_MEMZONE_IOVA_CONTIG, align);
 178 }
 179
 180 /*
 181  * This function is based on vmxnet3_disable_intr()
 182  */
 183 static void
 184 vmxnet3_disable_intr(struct vmxnet3_hw *hw)
 185 {
 186         int i;
 187
 188         PMD_INIT_FUNC_TRACE();
 189
 190         hw->shared->devRead.intrConf.intrCtrl |= VMXNET3_IC_DISABLE_ALL;
 191         for (i = 0; i < hw->num_intrs; i++)
 192                 VMXNET3_WRITE_BAR0_REG(hw, VMXNET3_REG_IMR + i * 8, 1);
 193 }
 194
 195 static void
 196 vmxnet3_enable_intr(struct vmxnet3_hw *hw)
 197 {
 198         int i;
 199
 200         PMD_INIT_FUNC_TRACE();
 201
 202         hw->shared->devRead.intrConf.intrCtrl &= ~VMXNET3_IC_DISABLE_ALL;
 203         for (i = 0; i < hw->num_intrs; i++)
 204                 VMXNET3_WRITE_BAR0_REG(hw, VMXNET3_REG_IMR + i * 8, 0);
 205 }
 206
 207 /*
 208  * Gets tx data ring descriptor size.
 209  */
 210 static uint16_t
 211 eth_vmxnet3_txdata_get(struct vmxnet3_hw *hw)
 212 {
 213         uint16 txdata_desc_size;
 214
 215         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
 216                                VMXNET3_CMD_GET_TXDATA_DESC_SIZE);
 217         txdata_desc_size = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 218
 219         return (txdata_desc_size < VMXNET3_TXDATA_DESC_MIN_SIZE ||
 220                 txdata_desc_size > VMXNET3_TXDATA_DESC_MAX_SIZE ||
 221                 txdata_desc_size & VMXNET3_TXDATA_DESC_SIZE_MASK) ?
 222                 sizeof(struct Vmxnet3_TxDataDesc) : txdata_desc_size;
 223 }
 224
 225 /*
 226  * It returns 0 on success.
 227  */
 228 static int
 229 eth_vmxnet3_dev_init(struct rte_eth_dev *eth_dev)
 230 {
 231         struct rte_pci_device *pci_dev;
 232         struct vmxnet3_hw *hw = eth_dev->data->dev_private;
 233         uint32_t mac_hi, mac_lo, ver;
 234         struct rte_eth_link link;
 235
 236         PMD_INIT_FUNC_TRACE();
 237
 238         eth_dev->dev_ops = &vmxnet3_eth_dev_ops;
 239         eth_dev->rx_pkt_burst = &vmxnet3_recv_pkts;
 240         eth_dev->tx_pkt_burst = &vmxnet3_xmit_pkts;
 241         eth_dev->tx_pkt_prepare = vmxnet3_prep_pkts;
 242         pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
 243
 244         /*
 245          * for secondary processes, we don't initialize any further as primary
 246          * has already done this work.
 247          */
 248         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 249                 return 0;
 250
 251         rte_eth_copy_pci_info(eth_dev, pci_dev);
 252
 253         /* Vendor and Device ID need to be set before init of shared code */
 254         hw->device_id = pci_dev->id.device_id;
 255         hw->vendor_id = pci_dev->id.vendor_id;
 256         hw->hw_addr0 = (void *)pci_dev->mem_resource[0].addr;
 257         hw->hw_addr1 = (void *)pci_dev->mem_resource[1].addr;
 258
 259         hw->num_rx_queues = 1;
 260         hw->num_tx_queues = 1;
 261         hw->bufs_per_pkt = 1;
 262
 263         /* Check h/w version compatibility with driver. */
 264         ver = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_VRRS);
 265         PMD_INIT_LOG(DEBUG, "Hardware version : %d", ver);
 266
 267         if (ver & (1 << VMXNET3_REV_3)) {
 268                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 269                                        1 << VMXNET3_REV_3);
 270                 hw->version = VMXNET3_REV_3 + 1;
 271         } else if (ver & (1 << VMXNET3_REV_2)) {
 272                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 273                                        1 << VMXNET3_REV_2);
 274                 hw->version = VMXNET3_REV_2 + 1;
 275         } else if (ver & (1 << VMXNET3_REV_1)) {
 276                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 277                                        1 << VMXNET3_REV_1);
 278                 hw->version = VMXNET3_REV_1 + 1;
 279         } else {
 280                 PMD_INIT_LOG(ERR, "Incompatible hardware version: %d", ver);
 281                 return -EIO;
 282         }
 283
 284         PMD_INIT_LOG(DEBUG, "Using device version %d\n", hw->version);
 285
 286         /* Check UPT version compatibility with driver. */
 287         ver = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_UVRS);
 288         PMD_INIT_LOG(DEBUG, "UPT hardware version : %d", ver);
 289         if (ver & 0x1)
 290                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_UVRS, 1);
 291         else {
 292                 PMD_INIT_LOG(ERR, "Incompatible UPT version.");
 293                 return -EIO;
 294         }
 295
 296         /* Getting MAC Address */
 297         mac_lo = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_MACL);
 298         mac_hi = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_MACH);
 299         memcpy(hw->perm_addr, &mac_lo, 4);
 300         memcpy(hw->perm_addr + 4, &mac_hi, 2);
 301
 302         /* Allocate memory for storing MAC addresses */
 303         eth_dev->data->mac_addrs = rte_zmalloc("vmxnet3", ETHER_ADDR_LEN *
 304                                                VMXNET3_MAX_MAC_ADDRS, 0);
 305         if (eth_dev->data->mac_addrs == NULL) {
 306                 PMD_INIT_LOG(ERR,
 307                              "Failed to allocate %d bytes needed to store MAC addresses",
 308                              ETHER_ADDR_LEN * VMXNET3_MAX_MAC_ADDRS);
 309                 return -ENOMEM;
 310         }
 311         /* Copy the permanent MAC address */
 312         ether_addr_copy((struct ether_addr *) hw->perm_addr,
 313                         &eth_dev->data->mac_addrs[0]);
 314
 315         PMD_INIT_LOG(DEBUG, "MAC Address : %02x:%02x:%02x:%02x:%02x:%02x",
 316                      hw->perm_addr[0], hw->perm_addr[1], hw->perm_addr[2],
 317                      hw->perm_addr[3], hw->perm_addr[4], hw->perm_addr[5]);
 318
 319         /* Put device in Quiesce Mode */
 320         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_QUIESCE_DEV);
 321
 322         /* allow untagged pkts */
 323         VMXNET3_SET_VFTABLE_ENTRY(hw->shadow_vfta, 0);
 324
 325         hw->txdata_desc_size = VMXNET3_VERSION_GE_3(hw) ?
 326                 eth_vmxnet3_txdata_get(hw) : sizeof(struct Vmxnet3_TxDataDesc);
 327
 328         hw->rxdata_desc_size = VMXNET3_VERSION_GE_3(hw) ?
 329                 VMXNET3_DEF_RXDATA_DESC_SIZE : 0;
 330         RTE_ASSERT((hw->rxdata_desc_size & ~VMXNET3_RXDATA_DESC_SIZE_MASK) ==
 331                    hw->rxdata_desc_size);
 332
 333         /* clear shadow stats */
 334         memset(hw->saved_tx_stats, 0, sizeof(hw->saved_tx_stats));
 335         memset(hw->saved_rx_stats, 0, sizeof(hw->saved_rx_stats));
 336
 337         /* set the initial link status */
 338         memset(&link, 0, sizeof(link));
 339         link.link_duplex = ETH_LINK_FULL_DUPLEX;
 340         link.link_speed = ETH_SPEED_NUM_10G;
 341         link.link_autoneg = ETH_LINK_FIXED;
 342         rte_eth_linkstatus_set(eth_dev, &link);
 343
 344         return 0;
 345 }
 346
 347 static int
 348 eth_vmxnet3_dev_uninit(struct rte_eth_dev *eth_dev)
 349 {
 350         struct vmxnet3_hw *hw = eth_dev->data->dev_private;
 351
 352         PMD_INIT_FUNC_TRACE();
 353
 354         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 355                 return 0;
 356
 357         if (hw->adapter_stopped == 0)
 358                 vmxnet3_dev_close(eth_dev);
 359
 360         eth_dev->dev_ops = NULL;
 361         eth_dev->rx_pkt_burst = NULL;
 362         eth_dev->tx_pkt_burst = NULL;
 363         eth_dev->tx_pkt_prepare = NULL;
 364
 365         rte_free(eth_dev->data->mac_addrs);
 366         eth_dev->data->mac_addrs = NULL;
 367
 368         return 0;
 369 }
 370
 371 static int eth_vmxnet3_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 372         struct rte_pci_device *pci_dev)
 373 {
 374         return rte_eth_dev_pci_generic_probe(pci_dev,
 375                 sizeof(struct vmxnet3_hw), eth_vmxnet3_dev_init);
 376 }
 377
 378 static int eth_vmxnet3_pci_remove(struct rte_pci_device *pci_dev)
 379 {
 380         return rte_eth_dev_pci_generic_remove(pci_dev, eth_vmxnet3_dev_uninit);
 381 }
 382
 383 static struct rte_pci_driver rte_vmxnet3_pmd = {
 384         .id_table = pci_id_vmxnet3_map,
 385         .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
 386         .probe = eth_vmxnet3_pci_probe,
 387         .remove = eth_vmxnet3_pci_remove,
 388 };
 389
 390 static int
 391 vmxnet3_dev_configure(struct rte_eth_dev *dev)
 392 {
 393         const struct rte_memzone *mz;
 394         struct vmxnet3_hw *hw = dev->data->dev_private;
 395         size_t size;
 396         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
 397         uint64_t tx_offloads = dev->data->dev_conf.txmode.offloads;
 398
 399         PMD_INIT_FUNC_TRACE();
 400
 401         if ((rx_offloads & VMXNET3_RX_OFFLOAD_CAP) != rx_offloads) {
 402                 RTE_LOG(ERR, PMD, "Requested RX offloads 0x%" PRIx64
 403                         " do not match supported 0x%" PRIx64,
 404                         rx_offloads, (uint64_t)VMXNET3_RX_OFFLOAD_CAP);
 405                 return -ENOTSUP;
 406         }
 407
 408         if ((tx_offloads & VMXNET3_TX_OFFLOAD_CAP) != tx_offloads) {
 409                 RTE_LOG(ERR, PMD, "Requested TX offloads 0x%" PRIx64
 410                         " do not match supported 0x%" PRIx64,
 411                         tx_offloads, (uint64_t)VMXNET3_TX_OFFLOAD_CAP);
 412                 return -ENOTSUP;
 413         }
 414
 415         if (dev->data->nb_tx_queues > VMXNET3_MAX_TX_QUEUES ||
 416             dev->data->nb_rx_queues > VMXNET3_MAX_RX_QUEUES) {
 417                 PMD_INIT_LOG(ERR, "ERROR: Number of queues not supported");
 418                 return -EINVAL;
 419         }
 420
 421         if (!rte_is_power_of_2(dev->data->nb_rx_queues)) {
 422                 PMD_INIT_LOG(ERR, "ERROR: Number of rx queues not power of 2");
 423                 return -EINVAL;
 424         }
 425
 426         size = dev->data->nb_rx_queues * sizeof(struct Vmxnet3_TxQueueDesc) +
 427                 dev->data->nb_tx_queues * sizeof(struct Vmxnet3_RxQueueDesc);
 428
 429         if (size > UINT16_MAX)
 430                 return -EINVAL;
 431
 432         hw->num_rx_queues = (uint8_t)dev->data->nb_rx_queues;
 433         hw->num_tx_queues = (uint8_t)dev->data->nb_tx_queues;
 434
 435         /*
 436          * Allocate a memzone for Vmxnet3_DriverShared - Vmxnet3_DSDevRead
 437          * on current socket
 438          */
 439         mz = gpa_zone_reserve(dev, sizeof(struct Vmxnet3_DriverShared),
 440                               "shared", rte_socket_id(), 8, 1);
 441
 442         if (mz == NULL) {
 443                 PMD_INIT_LOG(ERR, "ERROR: Creating shared zone");
 444                 return -ENOMEM;
 445         }
 446         memset(mz->addr, 0, mz->len);
 447
 448         hw->shared = mz->addr;
 449         hw->sharedPA = mz->iova;
 450
 451         /*
 452          * Allocate a memzone for Vmxnet3_RxQueueDesc - Vmxnet3_TxQueueDesc
 453          * on current socket.
 454          *
 455          * We cannot reuse this memzone from previous allocation as its size
 456          * depends on the number of tx and rx queues, which could be different
 457          * from one config to another.
 458          */
 459         mz = gpa_zone_reserve(dev, size, "queuedesc", rte_socket_id(),
 460                               VMXNET3_QUEUE_DESC_ALIGN, 0);
 461         if (mz == NULL) {
 462                 PMD_INIT_LOG(ERR, "ERROR: Creating queue descriptors zone");
 463                 return -ENOMEM;
 464         }
 465         memset(mz->addr, 0, mz->len);
 466
 467         hw->tqd_start = (Vmxnet3_TxQueueDesc *)mz->addr;
 468         hw->rqd_start = (Vmxnet3_RxQueueDesc *)(hw->tqd_start + hw->num_tx_queues);
 469
 470         hw->queueDescPA = mz->iova;
 471         hw->queue_desc_len = (uint16_t)size;
 472
 473         if (dev->data->dev_conf.rxmode.mq_mode == ETH_MQ_RX_RSS) {
 474                 /* Allocate memory structure for UPT1_RSSConf and configure */
 475                 mz = gpa_zone_reserve(dev, sizeof(struct VMXNET3_RSSConf),
 476                                       "rss_conf", rte_socket_id(),
 477                                       RTE_CACHE_LINE_SIZE, 1);
 478                 if (mz == NULL) {
 479                         PMD_INIT_LOG(ERR,
 480                                      "ERROR: Creating rss_conf structure zone");
 481                         return -ENOMEM;
 482                 }
 483                 memset(mz->addr, 0, mz->len);
 484
 485                 hw->rss_conf = mz->addr;
 486                 hw->rss_confPA = mz->iova;
 487         }
 488
 489         return 0;
 490 }
 491
 492 static void
 493 vmxnet3_write_mac(struct vmxnet3_hw *hw, const uint8_t *addr)
 494 {
 495         uint32_t val;
 496
 497         PMD_INIT_LOG(DEBUG,
 498                      "Writing MAC Address : %02x:%02x:%02x:%02x:%02x:%02x",
 499                      addr[0], addr[1], addr[2],
 500                      addr[3], addr[4], addr[5]);
 501
 502         memcpy(&val, addr, 4);
 503         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_MACL, val);
 504
 505         memcpy(&val, addr + 4, 2);
 506         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_MACH, val);
 507 }
 508
 509 static int
 510 vmxnet3_dev_setup_memreg(struct rte_eth_dev *dev)
 511 {
 512         struct vmxnet3_hw *hw = dev->data->dev_private;
 513         Vmxnet3_DriverShared *shared = hw->shared;
 514         Vmxnet3_CmdInfo *cmdInfo;
 515         struct rte_mempool *mp[VMXNET3_MAX_RX_QUEUES];
 516         uint8_t index[VMXNET3_MAX_RX_QUEUES + VMXNET3_MAX_TX_QUEUES];
 517         uint32_t num, i, j, size;
 518
 519         if (hw->memRegsPA == 0) {
 520                 const struct rte_memzone *mz;
 521
 522                 size = sizeof(Vmxnet3_MemRegs) +
 523                         (VMXNET3_MAX_RX_QUEUES + VMXNET3_MAX_TX_QUEUES) *
 524                         sizeof(Vmxnet3_MemoryRegion);
 525
 526                 mz = gpa_zone_reserve(dev, size, "memRegs", rte_socket_id(), 8,
 527                                       1);
 528                 if (mz == NULL) {
 529                         PMD_INIT_LOG(ERR, "ERROR: Creating memRegs zone");
 530                         return -ENOMEM;
 531                 }
 532                 memset(mz->addr, 0, mz->len);
 533                 hw->memRegs = mz->addr;
 534                 hw->memRegsPA = mz->iova;
 535         }
 536
 537         num = hw->num_rx_queues;
 538
 539         for (i = 0; i < num; i++) {
 540                 vmxnet3_rx_queue_t *rxq = dev->data->rx_queues[i];
 541
 542                 mp[i] = rxq->mp;
 543                 index[i] = 1 << i;
 544         }
 545
 546         /*
 547          * The same mempool could be used by multiple queues. In such a case,
 548          * remove duplicate mempool entries. Only one entry is kept with
 549          * bitmask indicating queues that are using this mempool.
 550          */
 551         for (i = 1; i < num; i++) {
 552                 for (j = 0; j < i; j++) {
 553                         if (mp[i] == mp[j]) {
 554                                 mp[i] = NULL;
 555                                 index[j] |= 1 << i;
 556                                 break;
 557                         }
 558                 }
 559         }
 560
 561         j = 0;
 562         for (i = 0; i < num; i++) {
 563                 if (mp[i] == NULL)
 564                         continue;
 565
 566                 Vmxnet3_MemoryRegion *mr = &hw->memRegs->memRegs[j];
 567
 568                 mr->startPA =
 569                         (uintptr_t)STAILQ_FIRST(&mp[i]->mem_list)->iova;
 570                 mr->length = STAILQ_FIRST(&mp[i]->mem_list)->len <= INT32_MAX ?
 571                         STAILQ_FIRST(&mp[i]->mem_list)->len : INT32_MAX;
 572                 mr->txQueueBits = index[i];
 573                 mr->rxQueueBits = index[i];
 574
 575                 PMD_INIT_LOG(INFO,
 576                              "index: %u startPA: %" PRIu64 " length: %u, "
 577                              "rxBits: %x",
 578                              j, mr->startPA, mr->length, mr->rxQueueBits);
 579                 j++;
 580         }
 581         hw->memRegs->numRegs = j;
 582         PMD_INIT_LOG(INFO, "numRegs: %u", j);
 583
 584         size = sizeof(Vmxnet3_MemRegs) +
 585                 (j - 1) * sizeof(Vmxnet3_MemoryRegion);
 586
 587         cmdInfo = &shared->cu.cmdInfo;
 588         cmdInfo->varConf.confVer = 1;
 589         cmdInfo->varConf.confLen = size;
 590         cmdInfo->varConf.confPA = hw->memRegsPA;
 591
 592         return 0;
 593 }
 594
 595 static int
 596 vmxnet3_setup_driver_shared(struct rte_eth_dev *dev)
 597 {
 598         struct rte_eth_conf port_conf = dev->data->dev_conf;
 599         struct vmxnet3_hw *hw = dev->data->dev_private;
 600         uint32_t mtu = dev->data->mtu;
 601         Vmxnet3_DriverShared *shared = hw->shared;
 602         Vmxnet3_DSDevRead *devRead = &shared->devRead;
 603         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
 604         uint32_t i;
 605         int ret;
 606
 607         hw->mtu = mtu;
 608
 609         shared->magic = VMXNET3_REV1_MAGIC;
 610         devRead->misc.driverInfo.version = VMXNET3_DRIVER_VERSION_NUM;
 611
 612         /* Setting up Guest OS information */
 613         devRead->misc.driverInfo.gos.gosBits   = sizeof(void *) == 4 ?
 614                 VMXNET3_GOS_BITS_32 : VMXNET3_GOS_BITS_64;
 615         devRead->misc.driverInfo.gos.gosType   = VMXNET3_GOS_TYPE_LINUX;
 616         devRead->misc.driverInfo.vmxnet3RevSpt = 1;
 617         devRead->misc.driverInfo.uptVerSpt     = 1;
 618
 619         devRead->misc.mtu = rte_le_to_cpu_32(mtu);
 620         devRead->misc.queueDescPA  = hw->queueDescPA;
 621         devRead->misc.queueDescLen = hw->queue_desc_len;
 622         devRead->misc.numTxQueues  = hw->num_tx_queues;
 623         devRead->misc.numRxQueues  = hw->num_rx_queues;
 624
 625         /*
 626          * Set number of interrupts to 1
 627          * PMD by default disables all the interrupts but this is MUST
 628          * to activate device. It needs at least one interrupt for
 629          * link events to handle
 630          */
 631         hw->num_intrs = devRead->intrConf.numIntrs = 1;
 632         devRead->intrConf.intrCtrl |= VMXNET3_IC_DISABLE_ALL;
 633
 634         for (i = 0; i < hw->num_tx_queues; i++) {
 635                 Vmxnet3_TxQueueDesc *tqd = &hw->tqd_start[i];
 636                 vmxnet3_tx_queue_t *txq  = dev->data->tx_queues[i];
 637
 638                 txq->shared = &hw->tqd_start[i];
 639
 640                 tqd->ctrl.txNumDeferred  = 0;
 641                 tqd->ctrl.txThreshold    = 1;
 642                 tqd->conf.txRingBasePA   = txq->cmd_ring.basePA;
 643                 tqd->conf.compRingBasePA = txq->comp_ring.basePA;
 644                 tqd->conf.dataRingBasePA = txq->data_ring.basePA;
 645
 646                 tqd->conf.txRingSize   = txq->cmd_ring.size;
 647                 tqd->conf.compRingSize = txq->comp_ring.size;
 648                 tqd->conf.dataRingSize = txq->data_ring.size;
 649                 tqd->conf.txDataRingDescSize = txq->txdata_desc_size;
 650                 tqd->conf.intrIdx      = txq->comp_ring.intr_idx;
 651                 tqd->status.stopped    = TRUE;
 652                 tqd->status.error      = 0;
 653                 memset(&tqd->stats, 0, sizeof(tqd->stats));
 654         }
 655
 656         for (i = 0; i < hw->num_rx_queues; i++) {
 657                 Vmxnet3_RxQueueDesc *rqd  = &hw->rqd_start[i];
 658                 vmxnet3_rx_queue_t *rxq   = dev->data->rx_queues[i];
 659
 660                 rxq->shared = &hw->rqd_start[i];
 661
 662                 rqd->conf.rxRingBasePA[0] = rxq->cmd_ring[0].basePA;
 663                 rqd->conf.rxRingBasePA[1] = rxq->cmd_ring[1].basePA;
 664                 rqd->conf.compRingBasePA  = rxq->comp_ring.basePA;
 665
 666                 rqd->conf.rxRingSize[0]   = rxq->cmd_ring[0].size;
 667                 rqd->conf.rxRingSize[1]   = rxq->cmd_ring[1].size;
 668                 rqd->conf.compRingSize    = rxq->comp_ring.size;
 669                 rqd->conf.intrIdx         = rxq->comp_ring.intr_idx;
 670                 if (VMXNET3_VERSION_GE_3(hw)) {
 671                         rqd->conf.rxDataRingBasePA = rxq->data_ring.basePA;
 672                         rqd->conf.rxDataRingDescSize = rxq->data_desc_size;
 673                 }
 674                 rqd->status.stopped       = TRUE;
 675                 rqd->status.error         = 0;
 676                 memset(&rqd->stats, 0, sizeof(rqd->stats));
 677         }
 678
 679         /* RxMode set to 0 of VMXNET3_RXM_xxx */
 680         devRead->rxFilterConf.rxMode = 0;
 681
 682         /* Setting up feature flags */
 683         if (rx_offloads & DEV_RX_OFFLOAD_CHECKSUM)
 684                 devRead->misc.uptFeatures |= VMXNET3_F_RXCSUM;
 685
 686         if (rx_offloads & DEV_RX_OFFLOAD_TCP_LRO) {
 687                 devRead->misc.uptFeatures |= VMXNET3_F_LRO;
 688                 devRead->misc.maxNumRxSG = 0;
 689         }
 690
 691         if (port_conf.rxmode.mq_mode == ETH_MQ_RX_RSS) {
 692                 ret = vmxnet3_rss_configure(dev);
 693                 if (ret != VMXNET3_SUCCESS)
 694                         return ret;
 695
 696                 devRead->misc.uptFeatures |= VMXNET3_F_RSS;
 697                 devRead->rssConfDesc.confVer = 1;
 698                 devRead->rssConfDesc.confLen = sizeof(struct VMXNET3_RSSConf);
 699                 devRead->rssConfDesc.confPA  = hw->rss_confPA;
 700         }
 701
 702         ret = vmxnet3_dev_vlan_offload_set(dev,
 703                         ETH_VLAN_STRIP_MASK | ETH_VLAN_FILTER_MASK);
 704         if (ret)
 705                 return ret;
 706
 707         vmxnet3_write_mac(hw, dev->data->mac_addrs->addr_bytes);
 708
 709         return VMXNET3_SUCCESS;
 710 }
 711
 712 /*
 713  * Configure device link speed and setup link.
 714  * Must be called after eth_vmxnet3_dev_init. Other wise it might fail
 715  * It returns 0 on success.
 716  */
 717 static int
 718 vmxnet3_dev_start(struct rte_eth_dev *dev)
 719 {
 720         int ret;
 721         struct vmxnet3_hw *hw = dev->data->dev_private;
 722
 723         PMD_INIT_FUNC_TRACE();
 724
 725         /* Save stats before it is reset by CMD_ACTIVATE */
 726         vmxnet3_hw_stats_save(hw);
 727
 728         ret = vmxnet3_setup_driver_shared(dev);
 729         if (ret != VMXNET3_SUCCESS)
 730                 return ret;
 731
 732         /* check if lsc interrupt feature is enabled */
 733         if (dev->data->dev_conf.intr_conf.lsc) {
 734                 struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
 735
 736                 /* Setup interrupt callback  */
 737                 rte_intr_callback_register(&pci_dev->intr_handle,
 738                                            vmxnet3_interrupt_handler, dev);
 739
 740                 if (rte_intr_enable(&pci_dev->intr_handle) < 0) {
 741                         PMD_INIT_LOG(ERR, "interrupt enable failed");
 742                         return -EIO;
 743                 }
 744         }
 745
 746         /* Exchange shared data with device */
 747         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAL,
 748                                VMXNET3_GET_ADDR_LO(hw->sharedPA));
 749         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAH,
 750                                VMXNET3_GET_ADDR_HI(hw->sharedPA));
 751
 752         /* Activate device by register write */
 753         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_ACTIVATE_DEV);
 754         ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 755
 756         if (ret != 0) {
 757                 PMD_INIT_LOG(ERR, "Device activation: UNSUCCESSFUL");
 758                 return -EINVAL;
 759         }
 760
 761         /* Setup memory region for rx buffers */
 762         ret = vmxnet3_dev_setup_memreg(dev);
 763         if (ret == 0) {
 764                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
 765                                        VMXNET3_CMD_REGISTER_MEMREGS);
 766                 ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 767                 if (ret != 0)
 768                         PMD_INIT_LOG(DEBUG,
 769                                      "Failed in setup memory region cmd\n");
 770                 ret = 0;
 771         } else {
 772                 PMD_INIT_LOG(DEBUG, "Failed to setup memory region\n");
 773         }
 774
 775         /* Disable interrupts */
 776         vmxnet3_disable_intr(hw);
 777
 778         /*
 779          * Load RX queues with blank mbufs and update next2fill index for device
 780          * Update RxMode of the device
 781          */
 782         ret = vmxnet3_dev_rxtx_init(dev);
 783         if (ret != VMXNET3_SUCCESS) {
 784                 PMD_INIT_LOG(ERR, "Device queue init: UNSUCCESSFUL");
 785                 return ret;
 786         }
 787
 788         hw->adapter_stopped = FALSE;
 789
 790         /* Setting proper Rx Mode and issue Rx Mode Update command */
 791         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_UCAST | VMXNET3_RXM_BCAST, 1);
 792
 793         if (dev->data->dev_conf.intr_conf.lsc) {
 794                 vmxnet3_enable_intr(hw);
 795
 796                 /*
 797                  * Update link state from device since this won't be
 798                  * done upon starting with lsc in use. This is done
 799                  * only after enabling interrupts to avoid any race
 800                  * where the link state could change without an
 801                  * interrupt being fired.
 802                  */
 803                 __vmxnet3_dev_link_update(dev, 0);
 804         }
 805
 806         return VMXNET3_SUCCESS;
 807 }
 808
 809 /*
 810  * Stop device: disable rx and tx functions to allow for reconfiguring.
 811  */
 812 static void
 813 vmxnet3_dev_stop(struct rte_eth_dev *dev)
 814 {
 815         struct rte_eth_link link;
 816         struct vmxnet3_hw *hw = dev->data->dev_private;
 817
 818         PMD_INIT_FUNC_TRACE();
 819
 820         if (hw->adapter_stopped == 1) {
 821                 PMD_INIT_LOG(DEBUG, "Device already closed.");
 822                 return;
 823         }
 824
 825         /* disable interrupts */
 826         vmxnet3_disable_intr(hw);
 827
 828         if (dev->data->dev_conf.intr_conf.lsc) {
 829                 struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
 830
 831                 rte_intr_disable(&pci_dev->intr_handle);
 832
 833                 rte_intr_callback_unregister(&pci_dev->intr_handle,
 834                                              vmxnet3_interrupt_handler, dev);
 835         }
 836
 837         /* quiesce the device first */
 838         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_QUIESCE_DEV);
 839         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAL, 0);
 840         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAH, 0);
 841
 842         /* reset the device */
 843         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_RESET_DEV);
 844         PMD_INIT_LOG(DEBUG, "Device reset.");
 845         hw->adapter_stopped = 0;
 846
 847         vmxnet3_dev_clear_queues(dev);
 848
 849         /* Clear recorded link status */
 850         memset(&link, 0, sizeof(link));
 851         link.link_duplex = ETH_LINK_FULL_DUPLEX;
 852         link.link_speed = ETH_SPEED_NUM_10G;
 853         link.link_autoneg = ETH_LINK_FIXED;
 854         rte_eth_linkstatus_set(dev, &link);
 855 }
 856
 857 /*
 858  * Reset and stop device.
 859  */
 860 static void
 861 vmxnet3_dev_close(struct rte_eth_dev *dev)
 862 {
 863         struct vmxnet3_hw *hw = dev->data->dev_private;
 864
 865         PMD_INIT_FUNC_TRACE();
 866
 867         vmxnet3_dev_stop(dev);
 868         hw->adapter_stopped = 1;
 869 }
 870
 871 static void
 872 vmxnet3_hw_tx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 873                         struct UPT1_TxStats *res)
 874 {
 875 #define VMXNET3_UPDATE_TX_STAT(h, i, f, r)              \
 876                 ((r)->f = (h)->tqd_start[(i)].stats.f + \
 877                         (h)->saved_tx_stats[(i)].f)
 878
 879         VMXNET3_UPDATE_TX_STAT(hw, q, ucastPktsTxOK, res);
 880         VMXNET3_UPDATE_TX_STAT(hw, q, mcastPktsTxOK, res);
 881         VMXNET3_UPDATE_TX_STAT(hw, q, bcastPktsTxOK, res);
 882         VMXNET3_UPDATE_TX_STAT(hw, q, ucastBytesTxOK, res);
 883         VMXNET3_UPDATE_TX_STAT(hw, q, mcastBytesTxOK, res);
 884         VMXNET3_UPDATE_TX_STAT(hw, q, bcastBytesTxOK, res);
 885         VMXNET3_UPDATE_TX_STAT(hw, q, pktsTxError, res);
 886         VMXNET3_UPDATE_TX_STAT(hw, q, pktsTxDiscard, res);
 887
 888 #undef VMXNET3_UPDATE_TX_STAT
 889 }
 890
 891 static void
 892 vmxnet3_hw_rx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 893                         struct UPT1_RxStats *res)
 894 {
 895 #define VMXNET3_UPDATE_RX_STAT(h, i, f, r)              \
 896                 ((r)->f = (h)->rqd_start[(i)].stats.f + \
 897                         (h)->saved_rx_stats[(i)].f)
 898
 899         VMXNET3_UPDATE_RX_STAT(hw, q, ucastPktsRxOK, res);
 900         VMXNET3_UPDATE_RX_STAT(hw, q, mcastPktsRxOK, res);
 901         VMXNET3_UPDATE_RX_STAT(hw, q, bcastPktsRxOK, res);
 902         VMXNET3_UPDATE_RX_STAT(hw, q, ucastBytesRxOK, res);
 903         VMXNET3_UPDATE_RX_STAT(hw, q, mcastBytesRxOK, res);
 904         VMXNET3_UPDATE_RX_STAT(hw, q, bcastBytesRxOK, res);
 905         VMXNET3_UPDATE_RX_STAT(hw, q, pktsRxError, res);
 906         VMXNET3_UPDATE_RX_STAT(hw, q, pktsRxOutOfBuf, res);
 907
 908 #undef VMXNET3_UPDATE_RX_STATS
 909 }
 910
 911 static void
 912 vmxnet3_hw_stats_save(struct vmxnet3_hw *hw)
 913 {
 914         unsigned int i;
 915
 916         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_STATS);
 917
 918         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_TX_QUEUES);
 919
 920         for (i = 0; i < hw->num_tx_queues; i++)
 921                 vmxnet3_hw_tx_stats_get(hw, i, &hw->saved_tx_stats[i]);
 922         for (i = 0; i < hw->num_rx_queues; i++)
 923                 vmxnet3_hw_rx_stats_get(hw, i, &hw->saved_rx_stats[i]);
 924 }
 925
 926 static int
 927 vmxnet3_dev_xstats_get_names(struct rte_eth_dev *dev,
 928                              struct rte_eth_xstat_name *xstats_names,
 929                              unsigned int n)
 930 {
 931         unsigned int i, t, count = 0;
 932         unsigned int nstats =
 933                 dev->data->nb_tx_queues * RTE_DIM(vmxnet3_txq_stat_strings) +
 934                 dev->data->nb_rx_queues * RTE_DIM(vmxnet3_rxq_stat_strings);
 935
 936         if (!xstats_names || n < nstats)
 937                 return nstats;
 938
 939         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 940                 if (!dev->data->rx_queues[i])
 941                         continue;
 942
 943                 for (t = 0; t < RTE_DIM(vmxnet3_rxq_stat_strings); t++) {
 944                         snprintf(xstats_names[count].name,
 945                                  sizeof(xstats_names[count].name),
 946                                  "rx_q%u_%s", i,
 947                                  vmxnet3_rxq_stat_strings[t].name);
 948                         count++;
 949                 }
 950         }
 951
 952         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 953                 if (!dev->data->tx_queues[i])
 954                         continue;
 955
 956                 for (t = 0; t < RTE_DIM(vmxnet3_txq_stat_strings); t++) {
 957                         snprintf(xstats_names[count].name,
 958                                  sizeof(xstats_names[count].name),
 959                                  "tx_q%u_%s", i,
 960                                  vmxnet3_txq_stat_strings[t].name);
 961                         count++;
 962                 }
 963         }
 964
 965         return count;
 966 }
 967
 968 static int
 969 vmxnet3_dev_xstats_get(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 970                        unsigned int n)
 971 {
 972         unsigned int i, t, count = 0;
 973         unsigned int nstats =
 974                 dev->data->nb_tx_queues * RTE_DIM(vmxnet3_txq_stat_strings) +
 975                 dev->data->nb_rx_queues * RTE_DIM(vmxnet3_rxq_stat_strings);
 976
 977         if (n < nstats)
 978                 return nstats;
 979
 980         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 981                 struct vmxnet3_rx_queue *rxq = dev->data->rx_queues[i];
 982
 983                 if (rxq == NULL)
 984                         continue;
 985
 986                 for (t = 0; t < RTE_DIM(vmxnet3_rxq_stat_strings); t++) {
 987                         xstats[count].value = *(uint64_t *)(((char *)&rxq->stats) +
 988                                 vmxnet3_rxq_stat_strings[t].offset);
 989                         xstats[count].id = count;
 990                         count++;
 991                 }
 992         }
 993
 994         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 995                 struct vmxnet3_tx_queue *txq = dev->data->tx_queues[i];
 996
 997                 if (txq == NULL)
 998                         continue;
 999
1000                 for (t = 0; t < RTE_DIM(vmxnet3_txq_stat_strings); t++) {
1001                         xstats[count].value = *(uint64_t *)(((char *)&txq->stats) +
1002                                 vmxnet3_txq_stat_strings[t].offset);
1003                         xstats[count].id = count;
1004                         count++;
1005                 }
1006         }
1007
1008         return count;
1009 }
1010
1011 static int
1012 vmxnet3_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1013 {
1014         unsigned int i;
1015         struct vmxnet3_hw *hw = dev->data->dev_private;
1016         struct UPT1_TxStats txStats;
1017         struct UPT1_RxStats rxStats;
1018
1019         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_STATS);
1020
1021         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_TX_QUEUES);
1022         for (i = 0; i < hw->num_tx_queues; i++) {
1023                 vmxnet3_hw_tx_stats_get(hw, i, &txStats);
1024
1025                 stats->q_opackets[i] = txStats.ucastPktsTxOK +
1026                         txStats.mcastPktsTxOK +
1027                         txStats.bcastPktsTxOK;
1028
1029                 stats->q_obytes[i] = txStats.ucastBytesTxOK +
1030                         txStats.mcastBytesTxOK +
1031                         txStats.bcastBytesTxOK;
1032
1033                 stats->opackets += stats->q_opackets[i];
1034                 stats->obytes += stats->q_obytes[i];
1035                 stats->oerrors += txStats.pktsTxError + txStats.pktsTxDiscard;
1036         }
1037
1038         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_RX_QUEUES);
1039         for (i = 0; i < hw->num_rx_queues; i++) {
1040                 vmxnet3_hw_rx_stats_get(hw, i, &rxStats);
1041
1042                 stats->q_ipackets[i] = rxStats.ucastPktsRxOK +
1043                         rxStats.mcastPktsRxOK +
1044                         rxStats.bcastPktsRxOK;
1045
1046                 stats->q_ibytes[i] = rxStats.ucastBytesRxOK +
1047                         rxStats.mcastBytesRxOK +
1048                         rxStats.bcastBytesRxOK;
1049
1050                 stats->ipackets += stats->q_ipackets[i];
1051                 stats->ibytes += stats->q_ibytes[i];
1052
1053                 stats->q_errors[i] = rxStats.pktsRxError;
1054                 stats->ierrors += rxStats.pktsRxError;
1055                 stats->imissed += rxStats.pktsRxOutOfBuf;
1056         }
1057
1058         return 0;
1059 }
1060
1061 static void
1062 vmxnet3_dev_info_get(struct rte_eth_dev *dev __rte_unused,
1063                      struct rte_eth_dev_info *dev_info)
1064 {
1065         dev_info->max_rx_queues = VMXNET3_MAX_RX_QUEUES;
1066         dev_info->max_tx_queues = VMXNET3_MAX_TX_QUEUES;
1067         dev_info->min_rx_bufsize = 1518 + RTE_PKTMBUF_HEADROOM;
1068         dev_info->max_rx_pktlen = 16384; /* includes CRC, cf MAXFRS register */
1069         dev_info->speed_capa = ETH_LINK_SPEED_10G;
1070         dev_info->max_mac_addrs = VMXNET3_MAX_MAC_ADDRS;
1071
1072         dev_info->default_txconf.txq_flags = ETH_TXQ_FLAGS_NOXSUMSCTP;
1073         dev_info->flow_type_rss_offloads = VMXNET3_RSS_OFFLOAD_ALL;
1074
1075         dev_info->rx_desc_lim = (struct rte_eth_desc_lim) {
1076                 .nb_max = VMXNET3_RX_RING_MAX_SIZE,
1077                 .nb_min = VMXNET3_DEF_RX_RING_SIZE,
1078                 .nb_align = 1,
1079         };
1080
1081         dev_info->tx_desc_lim = (struct rte_eth_desc_lim) {
1082                 .nb_max = VMXNET3_TX_RING_MAX_SIZE,
1083                 .nb_min = VMXNET3_DEF_TX_RING_SIZE,
1084                 .nb_align = 1,
1085                 .nb_seg_max = VMXNET3_TX_MAX_SEG,
1086                 .nb_mtu_seg_max = VMXNET3_MAX_TXD_PER_PKT,
1087         };
1088
1089         dev_info->rx_offload_capa = VMXNET3_RX_OFFLOAD_CAP;
1090         dev_info->rx_queue_offload_capa = 0;
1091         dev_info->tx_offload_capa = VMXNET3_TX_OFFLOAD_CAP;
1092         dev_info->tx_queue_offload_capa = 0;
1093 }
1094
1095 static const uint32_t *
1096 vmxnet3_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1097 {
1098         static const uint32_t ptypes[] = {
1099                 RTE_PTYPE_L3_IPV4_EXT,
1100                 RTE_PTYPE_L3_IPV4,
1101                 RTE_PTYPE_UNKNOWN
1102         };
1103
1104         if (dev->rx_pkt_burst == vmxnet3_recv_pkts)
1105                 return ptypes;
1106         return NULL;
1107 }
1108
1109 static int
1110 vmxnet3_mac_addr_set(struct rte_eth_dev *dev, struct ether_addr *mac_addr)
1111 {
1112         struct vmxnet3_hw *hw = dev->data->dev_private;
1113
1114         ether_addr_copy(mac_addr, (struct ether_addr *)(hw->perm_addr));
1115         vmxnet3_write_mac(hw, mac_addr->addr_bytes);
1116         return 0;
1117 }
1118
1119 /* return 0 means link status changed, -1 means not changed */
1120 static int
1121 __vmxnet3_dev_link_update(struct rte_eth_dev *dev,
1122                           __rte_unused int wait_to_complete)
1123 {
1124         struct vmxnet3_hw *hw = dev->data->dev_private;
1125         struct rte_eth_link link;
1126         uint32_t ret;
1127
1128         memset(&link, 0, sizeof(link));
1129
1130         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_LINK);
1131         ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
1132
1133         if (ret & 0x1)
1134                 link.link_status = ETH_LINK_UP;
1135         link.link_duplex = ETH_LINK_FULL_DUPLEX;
1136         link.link_speed = ETH_SPEED_NUM_10G;
1137         link.link_autoneg = ETH_LINK_FIXED;
1138
1139         return rte_eth_linkstatus_set(dev, &link);
1140 }
1141
1142 static int
1143 vmxnet3_dev_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1144 {
1145         /* Link status doesn't change for stopped dev */
1146         if (dev->data->dev_started == 0)
1147                 return -1;
1148
1149         return __vmxnet3_dev_link_update(dev, wait_to_complete);
1150 }
1151
1152 /* Updating rxmode through Vmxnet3_DriverShared structure in adapter */
1153 static void
1154 vmxnet3_dev_set_rxmode(struct vmxnet3_hw *hw, uint32_t feature, int set)
1155 {
1156         struct Vmxnet3_RxFilterConf *rxConf = &hw->shared->devRead.rxFilterConf;
1157
1158         if (set)
1159                 rxConf->rxMode = rxConf->rxMode | feature;
1160         else
1161                 rxConf->rxMode = rxConf->rxMode & (~feature);
1162
1163         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_UPDATE_RX_MODE);
1164 }
1165
1166 /* Promiscuous supported only if Vmxnet3_DriverShared is initialized in adapter */
1167 static void
1168 vmxnet3_dev_promiscuous_enable(struct rte_eth_dev *dev)
1169 {
1170         struct vmxnet3_hw *hw = dev->data->dev_private;
1171         uint32_t *vf_table = hw->shared->devRead.rxFilterConf.vfTable;
1172
1173         memset(vf_table, 0, VMXNET3_VFT_TABLE_SIZE);
1174         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_PROMISC, 1);
1175
1176         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1177                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1178 }
1179
1180 /* Promiscuous supported only if Vmxnet3_DriverShared is initialized in adapter */
1181 static void
1182 vmxnet3_dev_promiscuous_disable(struct rte_eth_dev *dev)
1183 {
1184         struct vmxnet3_hw *hw = dev->data->dev_private;
1185         uint32_t *vf_table = hw->shared->devRead.rxFilterConf.vfTable;
1186         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
1187
1188         if (rx_offloads & DEV_RX_OFFLOAD_VLAN_FILTER)
1189                 memcpy(vf_table, hw->shadow_vfta, VMXNET3_VFT_TABLE_SIZE);
1190         else
1191                 memset(vf_table, 0xff, VMXNET3_VFT_TABLE_SIZE);
1192         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_PROMISC, 0);
1193         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1194                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1195 }
1196
1197 /* Allmulticast supported only if Vmxnet3_DriverShared is initialized in adapter */
1198 static void
1199 vmxnet3_dev_allmulticast_enable(struct rte_eth_dev *dev)
1200 {
1201         struct vmxnet3_hw *hw = dev->data->dev_private;
1202
1203         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_ALL_MULTI, 1);
1204 }
1205
1206 /* Allmulticast supported only if Vmxnet3_DriverShared is initialized in adapter */
1207 static void
1208 vmxnet3_dev_allmulticast_disable(struct rte_eth_dev *dev)
1209 {
1210         struct vmxnet3_hw *hw = dev->data->dev_private;
1211
1212         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_ALL_MULTI, 0);
1213 }
1214
1215 /* Enable/disable filter on vlan */
1216 static int
1217 vmxnet3_dev_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vid, int on)
1218 {
1219         struct vmxnet3_hw *hw = dev->data->dev_private;
1220         struct Vmxnet3_RxFilterConf *rxConf = &hw->shared->devRead.rxFilterConf;
1221         uint32_t *vf_table = rxConf->vfTable;
1222
1223         /* save state for restore */
1224         if (on)
1225                 VMXNET3_SET_VFTABLE_ENTRY(hw->shadow_vfta, vid);
1226         else
1227                 VMXNET3_CLEAR_VFTABLE_ENTRY(hw->shadow_vfta, vid);
1228
1229         /* don't change active filter if in promiscuous mode */
1230         if (rxConf->rxMode & VMXNET3_RXM_PROMISC)
1231                 return 0;
1232
1233         /* set in hardware */
1234         if (on)
1235                 VMXNET3_SET_VFTABLE_ENTRY(vf_table, vid);
1236         else
1237                 VMXNET3_CLEAR_VFTABLE_ENTRY(vf_table, vid);
1238
1239         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1240                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1241         return 0;
1242 }
1243
1244 static int
1245 vmxnet3_dev_vlan_offload_set(struct rte_eth_dev *dev, int mask)
1246 {
1247         struct vmxnet3_hw *hw = dev->data->dev_private;
1248         Vmxnet3_DSDevRead *devRead = &hw->shared->devRead;
1249         uint32_t *vf_table = devRead->rxFilterConf.vfTable;
1250         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
1251
1252         if (mask & ETH_VLAN_STRIP_MASK) {
1253                 if (rx_offloads & DEV_RX_OFFLOAD_VLAN_STRIP)
1254                         devRead->misc.uptFeatures |= UPT1_F_RXVLAN;
1255                 else
1256                         devRead->misc.uptFeatures &= ~UPT1_F_RXVLAN;
1257
1258                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1259                                        VMXNET3_CMD_UPDATE_FEATURE);
1260         }
1261
1262         if (mask & ETH_VLAN_FILTER_MASK) {
1263                 if (rx_offloads & DEV_RX_OFFLOAD_VLAN_FILTER)
1264                         memcpy(vf_table, hw->shadow_vfta, VMXNET3_VFT_TABLE_SIZE);
1265                 else
1266                         memset(vf_table, 0xff, VMXNET3_VFT_TABLE_SIZE);
1267
1268                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1269                                        VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1270         }
1271
1272         return 0;
1273 }
1274
1275 static void
1276 vmxnet3_process_events(struct rte_eth_dev *dev)
1277 {
1278         struct vmxnet3_hw *hw = dev->data->dev_private;
1279         uint32_t events = hw->shared->ecr;
1280
1281         if (!events)
1282                 return;
1283
1284         /*
1285          * ECR bits when written with 1b are cleared. Hence write
1286          * events back to ECR so that the bits which were set will be reset.
1287          */
1288         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_ECR, events);
1289
1290         /* Check if link state has changed */
1291         if (events & VMXNET3_ECR_LINK) {
1292                 PMD_DRV_LOG(DEBUG, "Process events: VMXNET3_ECR_LINK event");
1293                 if (vmxnet3_dev_link_update(dev, 0) == 0)
1294                         _rte_eth_dev_callback_process(dev,
1295                                                       RTE_ETH_EVENT_INTR_LSC,
1296                                                       NULL);
1297         }
1298
1299         /* Check if there is an error on xmit/recv queues */
1300         if (events & (VMXNET3_ECR_TQERR | VMXNET3_ECR_RQERR)) {
1301                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1302                                        VMXNET3_CMD_GET_QUEUE_STATUS);
1303
1304                 if (hw->tqd_start->status.stopped)
1305                         PMD_DRV_LOG(ERR, "tq error 0x%x",
1306                                     hw->tqd_start->status.error);
1307
1308                 if (hw->rqd_start->status.stopped)
1309                         PMD_DRV_LOG(ERR, "rq error 0x%x",
1310                                      hw->rqd_start->status.error);
1311
1312                 /* Reset the device */
1313                 /* Have to reset the device */
1314         }
1315
1316         if (events & VMXNET3_ECR_DIC)
1317                 PMD_DRV_LOG(DEBUG, "Device implementation change event.");
1318
1319         if (events & VMXNET3_ECR_DEBUG)
1320                 PMD_DRV_LOG(DEBUG, "Debug event generated by device.");
1321 }
1322
1323 static void
1324 vmxnet3_interrupt_handler(void *param)
1325 {
1326         struct rte_eth_dev *dev = param;
1327         struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
1328
1329         vmxnet3_process_events(dev);
1330
1331         if (rte_intr_enable(&pci_dev->intr_handle) < 0)
1332                 PMD_DRV_LOG(ERR, "interrupt enable failed");
1333 }
1334
1335 RTE_PMD_REGISTER_PCI(net_vmxnet3, rte_vmxnet3_pmd);
1336 RTE_PMD_REGISTER_PCI_TABLE(net_vmxnet3, pci_id_vmxnet3_map);
1337 RTE_PMD_REGISTER_KMOD_DEP(net_vmxnet3, "* igb_uio | uio_pci_generic | vfio-pci");
1338
1339 RTE_INIT(vmxnet3_init_log);
1340 static void
1341 vmxnet3_init_log(void)
1342 {
1343         vmxnet3_logtype_init = rte_log_register("pmd.net.vmxnet3.init");
1344         if (vmxnet3_logtype_init >= 0)
1345                 rte_log_set_level(vmxnet3_logtype_init, RTE_LOG_NOTICE);
1346         vmxnet3_logtype_driver = rte_log_register("pmd.net.vmxnet3.driver");
1347         if (vmxnet3_logtype_driver >= 0)
1348                 rte_log_set_level(vmxnet3_logtype_driver, RTE_LOG_NOTICE);
1349 }