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_TCP_CKSUM |     \
  48          DEV_TX_OFFLOAD_UDP_CKSUM |     \
  49          DEV_TX_OFFLOAD_TCP_TSO |       \
  50          DEV_TX_OFFLOAD_MULTI_SEGS)
  51
  52 #define VMXNET3_RX_OFFLOAD_CAP          \
  53         (DEV_RX_OFFLOAD_VLAN_STRIP |    \
  54          DEV_RX_OFFLOAD_VLAN_FILTER |   \
  55          DEV_RX_OFFLOAD_SCATTER |       \
  56          DEV_RX_OFFLOAD_UDP_CKSUM |     \
  57          DEV_RX_OFFLOAD_TCP_CKSUM |     \
  58          DEV_RX_OFFLOAD_TCP_LRO |       \
  59          DEV_RX_OFFLOAD_JUMBO_FRAME |   \
  60          DEV_RX_OFFLOAD_RSS_HASH)
  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 int vmxnet3_dev_promiscuous_enable(struct rte_eth_dev *dev);
  70 static int vmxnet3_dev_promiscuous_disable(struct rte_eth_dev *dev);
  71 static int vmxnet3_dev_allmulticast_enable(struct rte_eth_dev *dev);
  72 static int 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_stats_reset(struct rte_eth_dev *dev);
  81 static int vmxnet3_dev_xstats_get_names(struct rte_eth_dev *dev,
  82                                         struct rte_eth_xstat_name *xstats,
  83                                         unsigned int n);
  84 static int vmxnet3_dev_xstats_get(struct rte_eth_dev *dev,
  85                                   struct rte_eth_xstat *xstats, unsigned int n);
  86 static int vmxnet3_dev_info_get(struct rte_eth_dev *dev,
  87                                 struct rte_eth_dev_info *dev_info);
  88 static const uint32_t *
  89 vmxnet3_dev_supported_ptypes_get(struct rte_eth_dev *dev);
  90 static int vmxnet3_dev_vlan_filter_set(struct rte_eth_dev *dev,
  91                                        uint16_t vid, int on);
  92 static int vmxnet3_dev_vlan_offload_set(struct rte_eth_dev *dev, int mask);
  93 static int vmxnet3_mac_addr_set(struct rte_eth_dev *dev,
  94                                  struct rte_ether_addr *mac_addr);
  95 static void vmxnet3_interrupt_handler(void *param);
  96
  97 int vmxnet3_logtype_init;
  98 int vmxnet3_logtype_driver;
  99
 100 /*
 101  * The set of PCI devices this driver supports
 102  */
 103 #define VMWARE_PCI_VENDOR_ID 0x15AD
 104 #define VMWARE_DEV_ID_VMXNET3 0x07B0
 105 static const struct rte_pci_id pci_id_vmxnet3_map[] = {
 106         { RTE_PCI_DEVICE(VMWARE_PCI_VENDOR_ID, VMWARE_DEV_ID_VMXNET3) },
 107         { .vendor_id = 0, /* sentinel */ },
 108 };
 109
 110 static const struct eth_dev_ops vmxnet3_eth_dev_ops = {
 111         .dev_configure        = vmxnet3_dev_configure,
 112         .dev_start            = vmxnet3_dev_start,
 113         .dev_stop             = vmxnet3_dev_stop,
 114         .dev_close            = vmxnet3_dev_close,
 115         .promiscuous_enable   = vmxnet3_dev_promiscuous_enable,
 116         .promiscuous_disable  = vmxnet3_dev_promiscuous_disable,
 117         .allmulticast_enable  = vmxnet3_dev_allmulticast_enable,
 118         .allmulticast_disable = vmxnet3_dev_allmulticast_disable,
 119         .link_update          = vmxnet3_dev_link_update,
 120         .stats_get            = vmxnet3_dev_stats_get,
 121         .xstats_get_names     = vmxnet3_dev_xstats_get_names,
 122         .xstats_get           = vmxnet3_dev_xstats_get,
 123         .stats_reset          = vmxnet3_dev_stats_reset,
 124         .mac_addr_set         = vmxnet3_mac_addr_set,
 125         .dev_infos_get        = vmxnet3_dev_info_get,
 126         .dev_supported_ptypes_get = vmxnet3_dev_supported_ptypes_get,
 127         .vlan_filter_set      = vmxnet3_dev_vlan_filter_set,
 128         .vlan_offload_set     = vmxnet3_dev_vlan_offload_set,
 129         .rx_queue_setup       = vmxnet3_dev_rx_queue_setup,
 130         .rx_queue_release     = vmxnet3_dev_rx_queue_release,
 131         .tx_queue_setup       = vmxnet3_dev_tx_queue_setup,
 132         .tx_queue_release     = vmxnet3_dev_tx_queue_release,
 133 };
 134
 135 struct vmxnet3_xstats_name_off {
 136         char name[RTE_ETH_XSTATS_NAME_SIZE];
 137         unsigned int offset;
 138 };
 139
 140 /* tx_qX_ is prepended to the name string here */
 141 static const struct vmxnet3_xstats_name_off vmxnet3_txq_stat_strings[] = {
 142         {"drop_total",         offsetof(struct vmxnet3_txq_stats, drop_total)},
 143         {"drop_too_many_segs", offsetof(struct vmxnet3_txq_stats, drop_too_many_segs)},
 144         {"drop_tso",           offsetof(struct vmxnet3_txq_stats, drop_tso)},
 145         {"tx_ring_full",       offsetof(struct vmxnet3_txq_stats, tx_ring_full)},
 146 };
 147
 148 /* rx_qX_ is prepended to the name string here */
 149 static const struct vmxnet3_xstats_name_off vmxnet3_rxq_stat_strings[] = {
 150         {"drop_total",           offsetof(struct vmxnet3_rxq_stats, drop_total)},
 151         {"drop_err",             offsetof(struct vmxnet3_rxq_stats, drop_err)},
 152         {"drop_fcs",             offsetof(struct vmxnet3_rxq_stats, drop_fcs)},
 153         {"rx_buf_alloc_failure", offsetof(struct vmxnet3_rxq_stats, rx_buf_alloc_failure)},
 154 };
 155
 156 static const struct rte_memzone *
 157 gpa_zone_reserve(struct rte_eth_dev *dev, uint32_t size,
 158                  const char *post_string, int socket_id,
 159                  uint16_t align, bool reuse)
 160 {
 161         char z_name[RTE_MEMZONE_NAMESIZE];
 162         const struct rte_memzone *mz;
 163
 164         snprintf(z_name, sizeof(z_name), "eth_p%d_%s",
 165                         dev->data->port_id, post_string);
 166
 167         mz = rte_memzone_lookup(z_name);
 168         if (!reuse) {
 169                 if (mz)
 170                         rte_memzone_free(mz);
 171                 return rte_memzone_reserve_aligned(z_name, size, socket_id,
 172                                 RTE_MEMZONE_IOVA_CONTIG, align);
 173         }
 174
 175         if (mz)
 176                 return mz;
 177
 178         return rte_memzone_reserve_aligned(z_name, size, socket_id,
 179                         RTE_MEMZONE_IOVA_CONTIG, align);
 180 }
 181
 182 /*
 183  * This function is based on vmxnet3_disable_intr()
 184  */
 185 static void
 186 vmxnet3_disable_intr(struct vmxnet3_hw *hw)
 187 {
 188         int i;
 189
 190         PMD_INIT_FUNC_TRACE();
 191
 192         hw->shared->devRead.intrConf.intrCtrl |= VMXNET3_IC_DISABLE_ALL;
 193         for (i = 0; i < hw->num_intrs; i++)
 194                 VMXNET3_WRITE_BAR0_REG(hw, VMXNET3_REG_IMR + i * 8, 1);
 195 }
 196
 197 static void
 198 vmxnet3_enable_intr(struct vmxnet3_hw *hw)
 199 {
 200         int i;
 201
 202         PMD_INIT_FUNC_TRACE();
 203
 204         hw->shared->devRead.intrConf.intrCtrl &= ~VMXNET3_IC_DISABLE_ALL;
 205         for (i = 0; i < hw->num_intrs; i++)
 206                 VMXNET3_WRITE_BAR0_REG(hw, VMXNET3_REG_IMR + i * 8, 0);
 207 }
 208
 209 /*
 210  * Gets tx data ring descriptor size.
 211  */
 212 static uint16_t
 213 eth_vmxnet3_txdata_get(struct vmxnet3_hw *hw)
 214 {
 215         uint16 txdata_desc_size;
 216
 217         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
 218                                VMXNET3_CMD_GET_TXDATA_DESC_SIZE);
 219         txdata_desc_size = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 220
 221         return (txdata_desc_size < VMXNET3_TXDATA_DESC_MIN_SIZE ||
 222                 txdata_desc_size > VMXNET3_TXDATA_DESC_MAX_SIZE ||
 223                 txdata_desc_size & VMXNET3_TXDATA_DESC_SIZE_MASK) ?
 224                 sizeof(struct Vmxnet3_TxDataDesc) : txdata_desc_size;
 225 }
 226
 227 /*
 228  * It returns 0 on success.
 229  */
 230 static int
 231 eth_vmxnet3_dev_init(struct rte_eth_dev *eth_dev)
 232 {
 233         struct rte_pci_device *pci_dev;
 234         struct vmxnet3_hw *hw = eth_dev->data->dev_private;
 235         uint32_t mac_hi, mac_lo, ver;
 236         struct rte_eth_link link;
 237
 238         PMD_INIT_FUNC_TRACE();
 239
 240         eth_dev->dev_ops = &vmxnet3_eth_dev_ops;
 241         eth_dev->rx_pkt_burst = &vmxnet3_recv_pkts;
 242         eth_dev->tx_pkt_burst = &vmxnet3_xmit_pkts;
 243         eth_dev->tx_pkt_prepare = vmxnet3_prep_pkts;
 244         pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
 245
 246         /*
 247          * for secondary processes, we don't initialize any further as primary
 248          * has already done this work.
 249          */
 250         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 251                 return 0;
 252
 253         rte_eth_copy_pci_info(eth_dev, pci_dev);
 254
 255         /* Vendor and Device ID need to be set before init of shared code */
 256         hw->device_id = pci_dev->id.device_id;
 257         hw->vendor_id = pci_dev->id.vendor_id;
 258         hw->hw_addr0 = (void *)pci_dev->mem_resource[0].addr;
 259         hw->hw_addr1 = (void *)pci_dev->mem_resource[1].addr;
 260
 261         hw->num_rx_queues = 1;
 262         hw->num_tx_queues = 1;
 263         hw->bufs_per_pkt = 1;
 264
 265         /* Check h/w version compatibility with driver. */
 266         ver = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_VRRS);
 267         PMD_INIT_LOG(DEBUG, "Hardware version : %d", ver);
 268
 269         if (ver & (1 << VMXNET3_REV_4)) {
 270                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 271                                        1 << VMXNET3_REV_4);
 272                 hw->version = VMXNET3_REV_4 + 1;
 273         } else if (ver & (1 << VMXNET3_REV_3)) {
 274                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 275                                        1 << VMXNET3_REV_3);
 276                 hw->version = VMXNET3_REV_3 + 1;
 277         } else if (ver & (1 << VMXNET3_REV_2)) {
 278                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 279                                        1 << VMXNET3_REV_2);
 280                 hw->version = VMXNET3_REV_2 + 1;
 281         } else if (ver & (1 << VMXNET3_REV_1)) {
 282                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_VRRS,
 283                                        1 << VMXNET3_REV_1);
 284                 hw->version = VMXNET3_REV_1 + 1;
 285         } else {
 286                 PMD_INIT_LOG(ERR, "Incompatible hardware version: %d", ver);
 287                 return -EIO;
 288         }
 289
 290         PMD_INIT_LOG(DEBUG, "Using device version %d\n", hw->version);
 291
 292         /* Check UPT version compatibility with driver. */
 293         ver = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_UVRS);
 294         PMD_INIT_LOG(DEBUG, "UPT hardware version : %d", ver);
 295         if (ver & 0x1)
 296                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_UVRS, 1);
 297         else {
 298                 PMD_INIT_LOG(ERR, "Incompatible UPT version.");
 299                 return -EIO;
 300         }
 301
 302         /* Getting MAC Address */
 303         mac_lo = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_MACL);
 304         mac_hi = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_MACH);
 305         memcpy(hw->perm_addr, &mac_lo, 4);
 306         memcpy(hw->perm_addr + 4, &mac_hi, 2);
 307
 308         /* Allocate memory for storing MAC addresses */
 309         eth_dev->data->mac_addrs = rte_zmalloc("vmxnet3", RTE_ETHER_ADDR_LEN *
 310                                                VMXNET3_MAX_MAC_ADDRS, 0);
 311         if (eth_dev->data->mac_addrs == NULL) {
 312                 PMD_INIT_LOG(ERR,
 313                              "Failed to allocate %d bytes needed to store MAC addresses",
 314                              RTE_ETHER_ADDR_LEN * VMXNET3_MAX_MAC_ADDRS);
 315                 return -ENOMEM;
 316         }
 317         /* Copy the permanent MAC address */
 318         rte_ether_addr_copy((struct rte_ether_addr *)hw->perm_addr,
 319                         &eth_dev->data->mac_addrs[0]);
 320
 321         PMD_INIT_LOG(DEBUG, "MAC Address : %02x:%02x:%02x:%02x:%02x:%02x",
 322                      hw->perm_addr[0], hw->perm_addr[1], hw->perm_addr[2],
 323                      hw->perm_addr[3], hw->perm_addr[4], hw->perm_addr[5]);
 324
 325         /* Flag to call rte_eth_dev_release_port() in rte_eth_dev_close(). */
 326         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 327
 328         /* Put device in Quiesce Mode */
 329         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_QUIESCE_DEV);
 330
 331         /* allow untagged pkts */
 332         VMXNET3_SET_VFTABLE_ENTRY(hw->shadow_vfta, 0);
 333
 334         hw->txdata_desc_size = VMXNET3_VERSION_GE_3(hw) ?
 335                 eth_vmxnet3_txdata_get(hw) : sizeof(struct Vmxnet3_TxDataDesc);
 336
 337         hw->rxdata_desc_size = VMXNET3_VERSION_GE_3(hw) ?
 338                 VMXNET3_DEF_RXDATA_DESC_SIZE : 0;
 339         RTE_ASSERT((hw->rxdata_desc_size & ~VMXNET3_RXDATA_DESC_SIZE_MASK) ==
 340                    hw->rxdata_desc_size);
 341
 342         /* clear shadow stats */
 343         memset(hw->saved_tx_stats, 0, sizeof(hw->saved_tx_stats));
 344         memset(hw->saved_rx_stats, 0, sizeof(hw->saved_rx_stats));
 345
 346         /* clear snapshot stats */
 347         memset(hw->snapshot_tx_stats, 0, sizeof(hw->snapshot_tx_stats));
 348         memset(hw->snapshot_rx_stats, 0, sizeof(hw->snapshot_rx_stats));
 349
 350         /* set the initial link status */
 351         memset(&link, 0, sizeof(link));
 352         link.link_duplex = ETH_LINK_FULL_DUPLEX;
 353         link.link_speed = ETH_SPEED_NUM_10G;
 354         link.link_autoneg = ETH_LINK_FIXED;
 355         rte_eth_linkstatus_set(eth_dev, &link);
 356
 357         return 0;
 358 }
 359
 360 static int
 361 eth_vmxnet3_dev_uninit(struct rte_eth_dev *eth_dev)
 362 {
 363         struct vmxnet3_hw *hw = eth_dev->data->dev_private;
 364
 365         PMD_INIT_FUNC_TRACE();
 366
 367         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 368                 return 0;
 369
 370         if (hw->adapter_stopped == 0) {
 371                 PMD_INIT_LOG(DEBUG, "Device has not been closed.");
 372                 return -EBUSY;
 373         }
 374
 375         eth_dev->dev_ops = NULL;
 376         eth_dev->rx_pkt_burst = NULL;
 377         eth_dev->tx_pkt_burst = NULL;
 378         eth_dev->tx_pkt_prepare = NULL;
 379
 380         return 0;
 381 }
 382
 383 static int eth_vmxnet3_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 384         struct rte_pci_device *pci_dev)
 385 {
 386         return rte_eth_dev_pci_generic_probe(pci_dev,
 387                 sizeof(struct vmxnet3_hw), eth_vmxnet3_dev_init);
 388 }
 389
 390 static int eth_vmxnet3_pci_remove(struct rte_pci_device *pci_dev)
 391 {
 392         return rte_eth_dev_pci_generic_remove(pci_dev, eth_vmxnet3_dev_uninit);
 393 }
 394
 395 static struct rte_pci_driver rte_vmxnet3_pmd = {
 396         .id_table = pci_id_vmxnet3_map,
 397         .drv_flags = RTE_PCI_DRV_NEED_MAPPING | RTE_PCI_DRV_INTR_LSC,
 398         .probe = eth_vmxnet3_pci_probe,
 399         .remove = eth_vmxnet3_pci_remove,
 400 };
 401
 402 static int
 403 vmxnet3_dev_configure(struct rte_eth_dev *dev)
 404 {
 405         const struct rte_memzone *mz;
 406         struct vmxnet3_hw *hw = dev->data->dev_private;
 407         size_t size;
 408
 409         PMD_INIT_FUNC_TRACE();
 410
 411         if (dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
 412                 dev->data->dev_conf.rxmode.offloads |= DEV_RX_OFFLOAD_RSS_HASH;
 413
 414         if (dev->data->nb_tx_queues > VMXNET3_MAX_TX_QUEUES ||
 415             dev->data->nb_rx_queues > VMXNET3_MAX_RX_QUEUES) {
 416                 PMD_INIT_LOG(ERR, "ERROR: Number of queues not supported");
 417                 return -EINVAL;
 418         }
 419
 420         if (!rte_is_power_of_2(dev->data->nb_rx_queues)) {
 421                 PMD_INIT_LOG(ERR, "ERROR: Number of rx queues not power of 2");
 422                 return -EINVAL;
 423         }
 424
 425         size = dev->data->nb_rx_queues * sizeof(struct Vmxnet3_TxQueueDesc) +
 426                 dev->data->nb_tx_queues * sizeof(struct Vmxnet3_RxQueueDesc);
 427
 428         if (size > UINT16_MAX)
 429                 return -EINVAL;
 430
 431         hw->num_rx_queues = (uint8_t)dev->data->nb_rx_queues;
 432         hw->num_tx_queues = (uint8_t)dev->data->nb_tx_queues;
 433
 434         /*
 435          * Allocate a memzone for Vmxnet3_DriverShared - Vmxnet3_DSDevRead
 436          * on current socket
 437          */
 438         mz = gpa_zone_reserve(dev, sizeof(struct Vmxnet3_DriverShared),
 439                               "shared", rte_socket_id(), 8, 1);
 440
 441         if (mz == NULL) {
 442                 PMD_INIT_LOG(ERR, "ERROR: Creating shared zone");
 443                 return -ENOMEM;
 444         }
 445         memset(mz->addr, 0, mz->len);
 446
 447         hw->shared = mz->addr;
 448         hw->sharedPA = mz->iova;
 449
 450         /*
 451          * Allocate a memzone for Vmxnet3_RxQueueDesc - Vmxnet3_TxQueueDesc
 452          * on current socket.
 453          *
 454          * We cannot reuse this memzone from previous allocation as its size
 455          * depends on the number of tx and rx queues, which could be different
 456          * from one config to another.
 457          */
 458         mz = gpa_zone_reserve(dev, size, "queuedesc", rte_socket_id(),
 459                               VMXNET3_QUEUE_DESC_ALIGN, 0);
 460         if (mz == NULL) {
 461                 PMD_INIT_LOG(ERR, "ERROR: Creating queue descriptors zone");
 462                 return -ENOMEM;
 463         }
 464         memset(mz->addr, 0, mz->len);
 465
 466         hw->tqd_start = (Vmxnet3_TxQueueDesc *)mz->addr;
 467         hw->rqd_start = (Vmxnet3_RxQueueDesc *)(hw->tqd_start + hw->num_tx_queues);
 468
 469         hw->queueDescPA = mz->iova;
 470         hw->queue_desc_len = (uint16_t)size;
 471
 472         if (dev->data->dev_conf.rxmode.mq_mode == ETH_MQ_RX_RSS) {
 473                 /* Allocate memory structure for UPT1_RSSConf and configure */
 474                 mz = gpa_zone_reserve(dev, sizeof(struct VMXNET3_RSSConf),
 475                                       "rss_conf", rte_socket_id(),
 476                                       RTE_CACHE_LINE_SIZE, 1);
 477                 if (mz == NULL) {
 478                         PMD_INIT_LOG(ERR,
 479                                      "ERROR: Creating rss_conf structure zone");
 480                         return -ENOMEM;
 481                 }
 482                 memset(mz->addr, 0, mz->len);
 483
 484                 hw->rss_conf = mz->addr;
 485                 hw->rss_confPA = mz->iova;
 486         }
 487
 488         return 0;
 489 }
 490
 491 static void
 492 vmxnet3_write_mac(struct vmxnet3_hw *hw, const uint8_t *addr)
 493 {
 494         uint32_t val;
 495
 496         PMD_INIT_LOG(DEBUG,
 497                      "Writing MAC Address : %02x:%02x:%02x:%02x:%02x:%02x",
 498                      addr[0], addr[1], addr[2],
 499                      addr[3], addr[4], addr[5]);
 500
 501         memcpy(&val, addr, 4);
 502         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_MACL, val);
 503
 504         memcpy(&val, addr + 4, 2);
 505         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_MACH, val);
 506 }
 507
 508 static int
 509 vmxnet3_dev_setup_memreg(struct rte_eth_dev *dev)
 510 {
 511         struct vmxnet3_hw *hw = dev->data->dev_private;
 512         Vmxnet3_DriverShared *shared = hw->shared;
 513         Vmxnet3_CmdInfo *cmdInfo;
 514         struct rte_mempool *mp[VMXNET3_MAX_RX_QUEUES];
 515         uint8_t index[VMXNET3_MAX_RX_QUEUES + VMXNET3_MAX_TX_QUEUES];
 516         uint32_t num, i, j, size;
 517
 518         if (hw->memRegsPA == 0) {
 519                 const struct rte_memzone *mz;
 520
 521                 size = sizeof(Vmxnet3_MemRegs) +
 522                         (VMXNET3_MAX_RX_QUEUES + VMXNET3_MAX_TX_QUEUES) *
 523                         sizeof(Vmxnet3_MemoryRegion);
 524
 525                 mz = gpa_zone_reserve(dev, size, "memRegs", rte_socket_id(), 8,
 526                                       1);
 527                 if (mz == NULL) {
 528                         PMD_INIT_LOG(ERR, "ERROR: Creating memRegs zone");
 529                         return -ENOMEM;
 530                 }
 531                 memset(mz->addr, 0, mz->len);
 532                 hw->memRegs = mz->addr;
 533                 hw->memRegsPA = mz->iova;
 534         }
 535
 536         num = hw->num_rx_queues;
 537
 538         for (i = 0; i < num; i++) {
 539                 vmxnet3_rx_queue_t *rxq = dev->data->rx_queues[i];
 540
 541                 mp[i] = rxq->mp;
 542                 index[i] = 1 << i;
 543         }
 544
 545         /*
 546          * The same mempool could be used by multiple queues. In such a case,
 547          * remove duplicate mempool entries. Only one entry is kept with
 548          * bitmask indicating queues that are using this mempool.
 549          */
 550         for (i = 1; i < num; i++) {
 551                 for (j = 0; j < i; j++) {
 552                         if (mp[i] == mp[j]) {
 553                                 mp[i] = NULL;
 554                                 index[j] |= 1 << i;
 555                                 break;
 556                         }
 557                 }
 558         }
 559
 560         j = 0;
 561         for (i = 0; i < num; i++) {
 562                 if (mp[i] == NULL)
 563                         continue;
 564
 565                 Vmxnet3_MemoryRegion *mr = &hw->memRegs->memRegs[j];
 566
 567                 mr->startPA =
 568                         (uintptr_t)STAILQ_FIRST(&mp[i]->mem_list)->iova;
 569                 mr->length = STAILQ_FIRST(&mp[i]->mem_list)->len <= INT32_MAX ?
 570                         STAILQ_FIRST(&mp[i]->mem_list)->len : INT32_MAX;
 571                 mr->txQueueBits = index[i];
 572                 mr->rxQueueBits = index[i];
 573
 574                 PMD_INIT_LOG(INFO,
 575                              "index: %u startPA: %" PRIu64 " length: %u, "
 576                              "rxBits: %x",
 577                              j, mr->startPA, mr->length, mr->rxQueueBits);
 578                 j++;
 579         }
 580         hw->memRegs->numRegs = j;
 581         PMD_INIT_LOG(INFO, "numRegs: %u", j);
 582
 583         size = sizeof(Vmxnet3_MemRegs) +
 584                 (j - 1) * sizeof(Vmxnet3_MemoryRegion);
 585
 586         cmdInfo = &shared->cu.cmdInfo;
 587         cmdInfo->varConf.confVer = 1;
 588         cmdInfo->varConf.confLen = size;
 589         cmdInfo->varConf.confPA = hw->memRegsPA;
 590
 591         return 0;
 592 }
 593
 594 static int
 595 vmxnet3_setup_driver_shared(struct rte_eth_dev *dev)
 596 {
 597         struct rte_eth_conf port_conf = dev->data->dev_conf;
 598         struct vmxnet3_hw *hw = dev->data->dev_private;
 599         uint32_t mtu = dev->data->mtu;
 600         Vmxnet3_DriverShared *shared = hw->shared;
 601         Vmxnet3_DSDevRead *devRead = &shared->devRead;
 602         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
 603         uint32_t i;
 604         int ret;
 605
 606         hw->mtu = mtu;
 607
 608         shared->magic = VMXNET3_REV1_MAGIC;
 609         devRead->misc.driverInfo.version = VMXNET3_DRIVER_VERSION_NUM;
 610
 611         /* Setting up Guest OS information */
 612         devRead->misc.driverInfo.gos.gosBits   = sizeof(void *) == 4 ?
 613                 VMXNET3_GOS_BITS_32 : VMXNET3_GOS_BITS_64;
 614         devRead->misc.driverInfo.gos.gosType   = VMXNET3_GOS_TYPE_LINUX;
 615         devRead->misc.driverInfo.vmxnet3RevSpt = 1;
 616         devRead->misc.driverInfo.uptVerSpt     = 1;
 617
 618         devRead->misc.mtu = rte_le_to_cpu_32(mtu);
 619         devRead->misc.queueDescPA  = hw->queueDescPA;
 620         devRead->misc.queueDescLen = hw->queue_desc_len;
 621         devRead->misc.numTxQueues  = hw->num_tx_queues;
 622         devRead->misc.numRxQueues  = hw->num_rx_queues;
 623
 624         /*
 625          * Set number of interrupts to 1
 626          * PMD by default disables all the interrupts but this is MUST
 627          * to activate device. It needs at least one interrupt for
 628          * link events to handle
 629          */
 630         hw->num_intrs = devRead->intrConf.numIntrs = 1;
 631         devRead->intrConf.intrCtrl |= VMXNET3_IC_DISABLE_ALL;
 632
 633         for (i = 0; i < hw->num_tx_queues; i++) {
 634                 Vmxnet3_TxQueueDesc *tqd = &hw->tqd_start[i];
 635                 vmxnet3_tx_queue_t *txq  = dev->data->tx_queues[i];
 636
 637                 txq->shared = &hw->tqd_start[i];
 638
 639                 tqd->ctrl.txNumDeferred  = 0;
 640                 tqd->ctrl.txThreshold    = 1;
 641                 tqd->conf.txRingBasePA   = txq->cmd_ring.basePA;
 642                 tqd->conf.compRingBasePA = txq->comp_ring.basePA;
 643                 tqd->conf.dataRingBasePA = txq->data_ring.basePA;
 644
 645                 tqd->conf.txRingSize   = txq->cmd_ring.size;
 646                 tqd->conf.compRingSize = txq->comp_ring.size;
 647                 tqd->conf.dataRingSize = txq->data_ring.size;
 648                 tqd->conf.txDataRingDescSize = txq->txdata_desc_size;
 649                 tqd->conf.intrIdx      = txq->comp_ring.intr_idx;
 650                 tqd->status.stopped    = TRUE;
 651                 tqd->status.error      = 0;
 652                 memset(&tqd->stats, 0, sizeof(tqd->stats));
 653         }
 654
 655         for (i = 0; i < hw->num_rx_queues; i++) {
 656                 Vmxnet3_RxQueueDesc *rqd  = &hw->rqd_start[i];
 657                 vmxnet3_rx_queue_t *rxq   = dev->data->rx_queues[i];
 658
 659                 rxq->shared = &hw->rqd_start[i];
 660
 661                 rqd->conf.rxRingBasePA[0] = rxq->cmd_ring[0].basePA;
 662                 rqd->conf.rxRingBasePA[1] = rxq->cmd_ring[1].basePA;
 663                 rqd->conf.compRingBasePA  = rxq->comp_ring.basePA;
 664
 665                 rqd->conf.rxRingSize[0]   = rxq->cmd_ring[0].size;
 666                 rqd->conf.rxRingSize[1]   = rxq->cmd_ring[1].size;
 667                 rqd->conf.compRingSize    = rxq->comp_ring.size;
 668                 rqd->conf.intrIdx         = rxq->comp_ring.intr_idx;
 669                 if (VMXNET3_VERSION_GE_3(hw)) {
 670                         rqd->conf.rxDataRingBasePA = rxq->data_ring.basePA;
 671                         rqd->conf.rxDataRingDescSize = rxq->data_desc_size;
 672                 }
 673                 rqd->status.stopped       = TRUE;
 674                 rqd->status.error         = 0;
 675                 memset(&rqd->stats, 0, sizeof(rqd->stats));
 676         }
 677
 678         /* RxMode set to 0 of VMXNET3_RXM_xxx */
 679         devRead->rxFilterConf.rxMode = 0;
 680
 681         /* Setting up feature flags */
 682         if (rx_offloads & DEV_RX_OFFLOAD_CHECKSUM)
 683                 devRead->misc.uptFeatures |= VMXNET3_F_RXCSUM;
 684
 685         if (rx_offloads & DEV_RX_OFFLOAD_TCP_LRO) {
 686                 devRead->misc.uptFeatures |= VMXNET3_F_LRO;
 687                 devRead->misc.maxNumRxSG = 0;
 688         }
 689
 690         if (port_conf.rxmode.mq_mode == ETH_MQ_RX_RSS) {
 691                 ret = vmxnet3_rss_configure(dev);
 692                 if (ret != VMXNET3_SUCCESS)
 693                         return ret;
 694
 695                 devRead->misc.uptFeatures |= VMXNET3_F_RSS;
 696                 devRead->rssConfDesc.confVer = 1;
 697                 devRead->rssConfDesc.confLen = sizeof(struct VMXNET3_RSSConf);
 698                 devRead->rssConfDesc.confPA  = hw->rss_confPA;
 699         }
 700
 701         ret = vmxnet3_dev_vlan_offload_set(dev,
 702                         ETH_VLAN_STRIP_MASK | ETH_VLAN_FILTER_MASK);
 703         if (ret)
 704                 return ret;
 705
 706         vmxnet3_write_mac(hw, dev->data->mac_addrs->addr_bytes);
 707
 708         return VMXNET3_SUCCESS;
 709 }
 710
 711 /*
 712  * Configure device link speed and setup link.
 713  * Must be called after eth_vmxnet3_dev_init. Other wise it might fail
 714  * It returns 0 on success.
 715  */
 716 static int
 717 vmxnet3_dev_start(struct rte_eth_dev *dev)
 718 {
 719         int ret;
 720         struct vmxnet3_hw *hw = dev->data->dev_private;
 721
 722         PMD_INIT_FUNC_TRACE();
 723
 724         /* Save stats before it is reset by CMD_ACTIVATE */
 725         vmxnet3_hw_stats_save(hw);
 726
 727         ret = vmxnet3_setup_driver_shared(dev);
 728         if (ret != VMXNET3_SUCCESS)
 729                 return ret;
 730
 731         /* check if lsc interrupt feature is enabled */
 732         if (dev->data->dev_conf.intr_conf.lsc) {
 733                 struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
 734
 735                 /* Setup interrupt callback  */
 736                 rte_intr_callback_register(&pci_dev->intr_handle,
 737                                            vmxnet3_interrupt_handler, dev);
 738
 739                 if (rte_intr_enable(&pci_dev->intr_handle) < 0) {
 740                         PMD_INIT_LOG(ERR, "interrupt enable failed");
 741                         return -EIO;
 742                 }
 743         }
 744
 745         /* Exchange shared data with device */
 746         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAL,
 747                                VMXNET3_GET_ADDR_LO(hw->sharedPA));
 748         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAH,
 749                                VMXNET3_GET_ADDR_HI(hw->sharedPA));
 750
 751         /* Activate device by register write */
 752         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_ACTIVATE_DEV);
 753         ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 754
 755         if (ret != 0) {
 756                 PMD_INIT_LOG(ERR, "Device activation: UNSUCCESSFUL");
 757                 return -EINVAL;
 758         }
 759
 760         /* Setup memory region for rx buffers */
 761         ret = vmxnet3_dev_setup_memreg(dev);
 762         if (ret == 0) {
 763                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
 764                                        VMXNET3_CMD_REGISTER_MEMREGS);
 765                 ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
 766                 if (ret != 0)
 767                         PMD_INIT_LOG(DEBUG,
 768                                      "Failed in setup memory region cmd\n");
 769                 ret = 0;
 770         } else {
 771                 PMD_INIT_LOG(DEBUG, "Failed to setup memory region\n");
 772         }
 773
 774         if (VMXNET3_VERSION_GE_4(hw)) {
 775                 /* Check for additional RSS  */
 776                 ret = vmxnet3_v4_rss_configure(dev);
 777                 if (ret != VMXNET3_SUCCESS) {
 778                         PMD_INIT_LOG(ERR, "Failed to configure v4 RSS");
 779                         return ret;
 780                 }
 781         }
 782
 783         /* Disable interrupts */
 784         vmxnet3_disable_intr(hw);
 785
 786         /*
 787          * Load RX queues with blank mbufs and update next2fill index for device
 788          * Update RxMode of the device
 789          */
 790         ret = vmxnet3_dev_rxtx_init(dev);
 791         if (ret != VMXNET3_SUCCESS) {
 792                 PMD_INIT_LOG(ERR, "Device queue init: UNSUCCESSFUL");
 793                 return ret;
 794         }
 795
 796         hw->adapter_stopped = FALSE;
 797
 798         /* Setting proper Rx Mode and issue Rx Mode Update command */
 799         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_UCAST | VMXNET3_RXM_BCAST, 1);
 800
 801         if (dev->data->dev_conf.intr_conf.lsc) {
 802                 vmxnet3_enable_intr(hw);
 803
 804                 /*
 805                  * Update link state from device since this won't be
 806                  * done upon starting with lsc in use. This is done
 807                  * only after enabling interrupts to avoid any race
 808                  * where the link state could change without an
 809                  * interrupt being fired.
 810                  */
 811                 __vmxnet3_dev_link_update(dev, 0);
 812         }
 813
 814         return VMXNET3_SUCCESS;
 815 }
 816
 817 /*
 818  * Stop device: disable rx and tx functions to allow for reconfiguring.
 819  */
 820 static void
 821 vmxnet3_dev_stop(struct rte_eth_dev *dev)
 822 {
 823         struct rte_eth_link link;
 824         struct vmxnet3_hw *hw = dev->data->dev_private;
 825
 826         PMD_INIT_FUNC_TRACE();
 827
 828         if (hw->adapter_stopped == 1) {
 829                 PMD_INIT_LOG(DEBUG, "Device already stopped.");
 830                 return;
 831         }
 832
 833         /* disable interrupts */
 834         vmxnet3_disable_intr(hw);
 835
 836         if (dev->data->dev_conf.intr_conf.lsc) {
 837                 struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
 838
 839                 rte_intr_disable(&pci_dev->intr_handle);
 840
 841                 rte_intr_callback_unregister(&pci_dev->intr_handle,
 842                                              vmxnet3_interrupt_handler, dev);
 843         }
 844
 845         /* quiesce the device first */
 846         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_QUIESCE_DEV);
 847         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAL, 0);
 848         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_DSAH, 0);
 849
 850         /* reset the device */
 851         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_RESET_DEV);
 852         PMD_INIT_LOG(DEBUG, "Device reset.");
 853
 854         vmxnet3_dev_clear_queues(dev);
 855
 856         /* Clear recorded link status */
 857         memset(&link, 0, sizeof(link));
 858         link.link_duplex = ETH_LINK_FULL_DUPLEX;
 859         link.link_speed = ETH_SPEED_NUM_10G;
 860         link.link_autoneg = ETH_LINK_FIXED;
 861         rte_eth_linkstatus_set(dev, &link);
 862
 863         hw->adapter_stopped = 1;
 864 }
 865
 866 static void
 867 vmxnet3_free_queues(struct rte_eth_dev *dev)
 868 {
 869         int i;
 870
 871         PMD_INIT_FUNC_TRACE();
 872
 873         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 874                 void *rxq = dev->data->rx_queues[i];
 875
 876                 vmxnet3_dev_rx_queue_release(rxq);
 877         }
 878         dev->data->nb_rx_queues = 0;
 879
 880         for (i = 0; i < dev->data->nb_tx_queues; i++) {
 881                 void *txq = dev->data->tx_queues[i];
 882
 883                 vmxnet3_dev_tx_queue_release(txq);
 884         }
 885         dev->data->nb_tx_queues = 0;
 886 }
 887
 888 /*
 889  * Reset and stop device.
 890  */
 891 static void
 892 vmxnet3_dev_close(struct rte_eth_dev *dev)
 893 {
 894         PMD_INIT_FUNC_TRACE();
 895
 896         vmxnet3_dev_stop(dev);
 897         vmxnet3_free_queues(dev);
 898 }
 899
 900 static void
 901 vmxnet3_hw_tx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 902                         struct UPT1_TxStats *res)
 903 {
 904 #define VMXNET3_UPDATE_TX_STAT(h, i, f, r)              \
 905                 ((r)->f = (h)->tqd_start[(i)].stats.f + \
 906                         (h)->saved_tx_stats[(i)].f)
 907
 908         VMXNET3_UPDATE_TX_STAT(hw, q, ucastPktsTxOK, res);
 909         VMXNET3_UPDATE_TX_STAT(hw, q, mcastPktsTxOK, res);
 910         VMXNET3_UPDATE_TX_STAT(hw, q, bcastPktsTxOK, res);
 911         VMXNET3_UPDATE_TX_STAT(hw, q, ucastBytesTxOK, res);
 912         VMXNET3_UPDATE_TX_STAT(hw, q, mcastBytesTxOK, res);
 913         VMXNET3_UPDATE_TX_STAT(hw, q, bcastBytesTxOK, res);
 914         VMXNET3_UPDATE_TX_STAT(hw, q, pktsTxError, res);
 915         VMXNET3_UPDATE_TX_STAT(hw, q, pktsTxDiscard, res);
 916
 917 #undef VMXNET3_UPDATE_TX_STAT
 918 }
 919
 920 static void
 921 vmxnet3_hw_rx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 922                         struct UPT1_RxStats *res)
 923 {
 924 #define VMXNET3_UPDATE_RX_STAT(h, i, f, r)              \
 925                 ((r)->f = (h)->rqd_start[(i)].stats.f + \
 926                         (h)->saved_rx_stats[(i)].f)
 927
 928         VMXNET3_UPDATE_RX_STAT(hw, q, ucastPktsRxOK, res);
 929         VMXNET3_UPDATE_RX_STAT(hw, q, mcastPktsRxOK, res);
 930         VMXNET3_UPDATE_RX_STAT(hw, q, bcastPktsRxOK, res);
 931         VMXNET3_UPDATE_RX_STAT(hw, q, ucastBytesRxOK, res);
 932         VMXNET3_UPDATE_RX_STAT(hw, q, mcastBytesRxOK, res);
 933         VMXNET3_UPDATE_RX_STAT(hw, q, bcastBytesRxOK, res);
 934         VMXNET3_UPDATE_RX_STAT(hw, q, pktsRxError, res);
 935         VMXNET3_UPDATE_RX_STAT(hw, q, pktsRxOutOfBuf, res);
 936
 937 #undef VMXNET3_UPDATE_RX_STAT
 938 }
 939
 940 static void
 941 vmxnet3_tx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 942                                         struct UPT1_TxStats *res)
 943 {
 944                 vmxnet3_hw_tx_stats_get(hw, q, res);
 945
 946 #define VMXNET3_REDUCE_SNAPSHOT_TX_STAT(h, i, f, r)     \
 947                 ((r)->f -= (h)->snapshot_tx_stats[(i)].f)
 948
 949         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, ucastPktsTxOK, res);
 950         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, mcastPktsTxOK, res);
 951         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, bcastPktsTxOK, res);
 952         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, ucastBytesTxOK, res);
 953         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, mcastBytesTxOK, res);
 954         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, bcastBytesTxOK, res);
 955         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, pktsTxError, res);
 956         VMXNET3_REDUCE_SNAPSHOT_TX_STAT(hw, q, pktsTxDiscard, res);
 957
 958 #undef VMXNET3_REDUCE_SNAPSHOT_TX_STAT
 959 }
 960
 961 static void
 962 vmxnet3_rx_stats_get(struct vmxnet3_hw *hw, unsigned int q,
 963                                         struct UPT1_RxStats *res)
 964 {
 965                 vmxnet3_hw_rx_stats_get(hw, q, res);
 966
 967 #define VMXNET3_REDUCE_SNAPSHOT_RX_STAT(h, i, f, r)     \
 968                 ((r)->f -= (h)->snapshot_rx_stats[(i)].f)
 969
 970         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, ucastPktsRxOK, res);
 971         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, mcastPktsRxOK, res);
 972         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, bcastPktsRxOK, res);
 973         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, ucastBytesRxOK, res);
 974         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, mcastBytesRxOK, res);
 975         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, bcastBytesRxOK, res);
 976         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, pktsRxError, res);
 977         VMXNET3_REDUCE_SNAPSHOT_RX_STAT(hw, q, pktsRxOutOfBuf, res);
 978
 979 #undef VMXNET3_REDUCE_SNAPSHOT_RX_STAT
 980 }
 981
 982 static void
 983 vmxnet3_hw_stats_save(struct vmxnet3_hw *hw)
 984 {
 985         unsigned int i;
 986
 987         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_STATS);
 988
 989         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_TX_QUEUES);
 990
 991         for (i = 0; i < hw->num_tx_queues; i++)
 992                 vmxnet3_hw_tx_stats_get(hw, i, &hw->saved_tx_stats[i]);
 993         for (i = 0; i < hw->num_rx_queues; i++)
 994                 vmxnet3_hw_rx_stats_get(hw, i, &hw->saved_rx_stats[i]);
 995 }
 996
 997 static int
 998 vmxnet3_dev_xstats_get_names(struct rte_eth_dev *dev,
 999                              struct rte_eth_xstat_name *xstats_names,
1000                              unsigned int n)
1001 {
1002         unsigned int i, t, count = 0;
1003         unsigned int nstats =
1004                 dev->data->nb_tx_queues * RTE_DIM(vmxnet3_txq_stat_strings) +
1005                 dev->data->nb_rx_queues * RTE_DIM(vmxnet3_rxq_stat_strings);
1006
1007         if (!xstats_names || n < nstats)
1008                 return nstats;
1009
1010         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1011                 if (!dev->data->rx_queues[i])
1012                         continue;
1013
1014                 for (t = 0; t < RTE_DIM(vmxnet3_rxq_stat_strings); t++) {
1015                         snprintf(xstats_names[count].name,
1016                                  sizeof(xstats_names[count].name),
1017                                  "rx_q%u_%s", i,
1018                                  vmxnet3_rxq_stat_strings[t].name);
1019                         count++;
1020                 }
1021         }
1022
1023         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1024                 if (!dev->data->tx_queues[i])
1025                         continue;
1026
1027                 for (t = 0; t < RTE_DIM(vmxnet3_txq_stat_strings); t++) {
1028                         snprintf(xstats_names[count].name,
1029                                  sizeof(xstats_names[count].name),
1030                                  "tx_q%u_%s", i,
1031                                  vmxnet3_txq_stat_strings[t].name);
1032                         count++;
1033                 }
1034         }
1035
1036         return count;
1037 }
1038
1039 static int
1040 vmxnet3_dev_xstats_get(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
1041                        unsigned int n)
1042 {
1043         unsigned int i, t, count = 0;
1044         unsigned int nstats =
1045                 dev->data->nb_tx_queues * RTE_DIM(vmxnet3_txq_stat_strings) +
1046                 dev->data->nb_rx_queues * RTE_DIM(vmxnet3_rxq_stat_strings);
1047
1048         if (n < nstats)
1049                 return nstats;
1050
1051         for (i = 0; i < dev->data->nb_rx_queues; i++) {
1052                 struct vmxnet3_rx_queue *rxq = dev->data->rx_queues[i];
1053
1054                 if (rxq == NULL)
1055                         continue;
1056
1057                 for (t = 0; t < RTE_DIM(vmxnet3_rxq_stat_strings); t++) {
1058                         xstats[count].value = *(uint64_t *)(((char *)&rxq->stats) +
1059                                 vmxnet3_rxq_stat_strings[t].offset);
1060                         xstats[count].id = count;
1061                         count++;
1062                 }
1063         }
1064
1065         for (i = 0; i < dev->data->nb_tx_queues; i++) {
1066                 struct vmxnet3_tx_queue *txq = dev->data->tx_queues[i];
1067
1068                 if (txq == NULL)
1069                         continue;
1070
1071                 for (t = 0; t < RTE_DIM(vmxnet3_txq_stat_strings); t++) {
1072                         xstats[count].value = *(uint64_t *)(((char *)&txq->stats) +
1073                                 vmxnet3_txq_stat_strings[t].offset);
1074                         xstats[count].id = count;
1075                         count++;
1076                 }
1077         }
1078
1079         return count;
1080 }
1081
1082 static int
1083 vmxnet3_dev_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
1084 {
1085         unsigned int i;
1086         struct vmxnet3_hw *hw = dev->data->dev_private;
1087         struct UPT1_TxStats txStats;
1088         struct UPT1_RxStats rxStats;
1089
1090         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_STATS);
1091
1092         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_TX_QUEUES);
1093         for (i = 0; i < hw->num_tx_queues; i++) {
1094                 vmxnet3_tx_stats_get(hw, i, &txStats);
1095
1096                 stats->q_opackets[i] = txStats.ucastPktsTxOK +
1097                         txStats.mcastPktsTxOK +
1098                         txStats.bcastPktsTxOK;
1099
1100                 stats->q_obytes[i] = txStats.ucastBytesTxOK +
1101                         txStats.mcastBytesTxOK +
1102                         txStats.bcastBytesTxOK;
1103
1104                 stats->opackets += stats->q_opackets[i];
1105                 stats->obytes += stats->q_obytes[i];
1106                 stats->oerrors += txStats.pktsTxError + txStats.pktsTxDiscard;
1107         }
1108
1109         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_RX_QUEUES);
1110         for (i = 0; i < hw->num_rx_queues; i++) {
1111                 vmxnet3_rx_stats_get(hw, i, &rxStats);
1112
1113                 stats->q_ipackets[i] = rxStats.ucastPktsRxOK +
1114                         rxStats.mcastPktsRxOK +
1115                         rxStats.bcastPktsRxOK;
1116
1117                 stats->q_ibytes[i] = rxStats.ucastBytesRxOK +
1118                         rxStats.mcastBytesRxOK +
1119                         rxStats.bcastBytesRxOK;
1120
1121                 stats->ipackets += stats->q_ipackets[i];
1122                 stats->ibytes += stats->q_ibytes[i];
1123
1124                 stats->q_errors[i] = rxStats.pktsRxError;
1125                 stats->ierrors += rxStats.pktsRxError;
1126                 stats->imissed += rxStats.pktsRxOutOfBuf;
1127         }
1128
1129         return 0;
1130 }
1131
1132 static int
1133 vmxnet3_dev_stats_reset(struct rte_eth_dev *dev)
1134 {
1135         unsigned int i;
1136         struct vmxnet3_hw *hw = dev->data->dev_private;
1137         struct UPT1_TxStats txStats = {0};
1138         struct UPT1_RxStats rxStats = {0};
1139
1140         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_STATS);
1141
1142         RTE_BUILD_BUG_ON(RTE_ETHDEV_QUEUE_STAT_CNTRS < VMXNET3_MAX_TX_QUEUES);
1143
1144         for (i = 0; i < hw->num_tx_queues; i++) {
1145                 vmxnet3_hw_tx_stats_get(hw, i, &txStats);
1146                 memcpy(&hw->snapshot_tx_stats[i], &txStats,
1147                         sizeof(hw->snapshot_tx_stats[0]));
1148         }
1149         for (i = 0; i < hw->num_rx_queues; i++) {
1150                 vmxnet3_hw_rx_stats_get(hw, i, &rxStats);
1151                 memcpy(&hw->snapshot_rx_stats[i], &rxStats,
1152                         sizeof(hw->snapshot_rx_stats[0]));
1153         }
1154
1155         return 0;
1156 }
1157
1158 static int
1159 vmxnet3_dev_info_get(struct rte_eth_dev *dev,
1160                      struct rte_eth_dev_info *dev_info)
1161 {
1162         struct vmxnet3_hw *hw = dev->data->dev_private;
1163
1164         dev_info->max_rx_queues = VMXNET3_MAX_RX_QUEUES;
1165         dev_info->max_tx_queues = VMXNET3_MAX_TX_QUEUES;
1166         dev_info->min_rx_bufsize = 1518 + RTE_PKTMBUF_HEADROOM;
1167         dev_info->max_rx_pktlen = 16384; /* includes CRC, cf MAXFRS register */
1168         dev_info->speed_capa = ETH_LINK_SPEED_10G;
1169         dev_info->max_mac_addrs = VMXNET3_MAX_MAC_ADDRS;
1170
1171         dev_info->flow_type_rss_offloads = VMXNET3_RSS_OFFLOAD_ALL;
1172
1173         if (VMXNET3_VERSION_GE_4(hw)) {
1174                 dev_info->flow_type_rss_offloads |= VMXNET3_V4_RSS_MASK;
1175         }
1176
1177         dev_info->rx_desc_lim = (struct rte_eth_desc_lim) {
1178                 .nb_max = VMXNET3_RX_RING_MAX_SIZE,
1179                 .nb_min = VMXNET3_DEF_RX_RING_SIZE,
1180                 .nb_align = 1,
1181         };
1182
1183         dev_info->tx_desc_lim = (struct rte_eth_desc_lim) {
1184                 .nb_max = VMXNET3_TX_RING_MAX_SIZE,
1185                 .nb_min = VMXNET3_DEF_TX_RING_SIZE,
1186                 .nb_align = 1,
1187                 .nb_seg_max = VMXNET3_TX_MAX_SEG,
1188                 .nb_mtu_seg_max = VMXNET3_MAX_TXD_PER_PKT,
1189         };
1190
1191         dev_info->rx_offload_capa = VMXNET3_RX_OFFLOAD_CAP;
1192         dev_info->rx_queue_offload_capa = 0;
1193         dev_info->tx_offload_capa = VMXNET3_TX_OFFLOAD_CAP;
1194         dev_info->tx_queue_offload_capa = 0;
1195
1196         return 0;
1197 }
1198
1199 static const uint32_t *
1200 vmxnet3_dev_supported_ptypes_get(struct rte_eth_dev *dev)
1201 {
1202         static const uint32_t ptypes[] = {
1203                 RTE_PTYPE_L3_IPV4_EXT,
1204                 RTE_PTYPE_L3_IPV4,
1205                 RTE_PTYPE_UNKNOWN
1206         };
1207
1208         if (dev->rx_pkt_burst == vmxnet3_recv_pkts)
1209                 return ptypes;
1210         return NULL;
1211 }
1212
1213 static int
1214 vmxnet3_mac_addr_set(struct rte_eth_dev *dev, struct rte_ether_addr *mac_addr)
1215 {
1216         struct vmxnet3_hw *hw = dev->data->dev_private;
1217
1218         rte_ether_addr_copy(mac_addr, (struct rte_ether_addr *)(hw->perm_addr));
1219         vmxnet3_write_mac(hw, mac_addr->addr_bytes);
1220         return 0;
1221 }
1222
1223 /* return 0 means link status changed, -1 means not changed */
1224 static int
1225 __vmxnet3_dev_link_update(struct rte_eth_dev *dev,
1226                           __rte_unused int wait_to_complete)
1227 {
1228         struct vmxnet3_hw *hw = dev->data->dev_private;
1229         struct rte_eth_link link;
1230         uint32_t ret;
1231
1232         memset(&link, 0, sizeof(link));
1233
1234         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_GET_LINK);
1235         ret = VMXNET3_READ_BAR1_REG(hw, VMXNET3_REG_CMD);
1236
1237         if (ret & 0x1)
1238                 link.link_status = ETH_LINK_UP;
1239         link.link_duplex = ETH_LINK_FULL_DUPLEX;
1240         link.link_speed = ETH_SPEED_NUM_10G;
1241         link.link_autoneg = ETH_LINK_FIXED;
1242
1243         return rte_eth_linkstatus_set(dev, &link);
1244 }
1245
1246 static int
1247 vmxnet3_dev_link_update(struct rte_eth_dev *dev, int wait_to_complete)
1248 {
1249         /* Link status doesn't change for stopped dev */
1250         if (dev->data->dev_started == 0)
1251                 return -1;
1252
1253         return __vmxnet3_dev_link_update(dev, wait_to_complete);
1254 }
1255
1256 /* Updating rxmode through Vmxnet3_DriverShared structure in adapter */
1257 static void
1258 vmxnet3_dev_set_rxmode(struct vmxnet3_hw *hw, uint32_t feature, int set)
1259 {
1260         struct Vmxnet3_RxFilterConf *rxConf = &hw->shared->devRead.rxFilterConf;
1261
1262         if (set)
1263                 rxConf->rxMode = rxConf->rxMode | feature;
1264         else
1265                 rxConf->rxMode = rxConf->rxMode & (~feature);
1266
1267         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD, VMXNET3_CMD_UPDATE_RX_MODE);
1268 }
1269
1270 /* Promiscuous supported only if Vmxnet3_DriverShared is initialized in adapter */
1271 static int
1272 vmxnet3_dev_promiscuous_enable(struct rte_eth_dev *dev)
1273 {
1274         struct vmxnet3_hw *hw = dev->data->dev_private;
1275         uint32_t *vf_table = hw->shared->devRead.rxFilterConf.vfTable;
1276
1277         memset(vf_table, 0, VMXNET3_VFT_TABLE_SIZE);
1278         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_PROMISC, 1);
1279
1280         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1281                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1282
1283         return 0;
1284 }
1285
1286 /* Promiscuous supported only if Vmxnet3_DriverShared is initialized in adapter */
1287 static int
1288 vmxnet3_dev_promiscuous_disable(struct rte_eth_dev *dev)
1289 {
1290         struct vmxnet3_hw *hw = dev->data->dev_private;
1291         uint32_t *vf_table = hw->shared->devRead.rxFilterConf.vfTable;
1292         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
1293
1294         if (rx_offloads & DEV_RX_OFFLOAD_VLAN_FILTER)
1295                 memcpy(vf_table, hw->shadow_vfta, VMXNET3_VFT_TABLE_SIZE);
1296         else
1297                 memset(vf_table, 0xff, VMXNET3_VFT_TABLE_SIZE);
1298         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_PROMISC, 0);
1299         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1300                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1301
1302         return 0;
1303 }
1304
1305 /* Allmulticast supported only if Vmxnet3_DriverShared is initialized in adapter */
1306 static int
1307 vmxnet3_dev_allmulticast_enable(struct rte_eth_dev *dev)
1308 {
1309         struct vmxnet3_hw *hw = dev->data->dev_private;
1310
1311         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_ALL_MULTI, 1);
1312
1313         return 0;
1314 }
1315
1316 /* Allmulticast supported only if Vmxnet3_DriverShared is initialized in adapter */
1317 static int
1318 vmxnet3_dev_allmulticast_disable(struct rte_eth_dev *dev)
1319 {
1320         struct vmxnet3_hw *hw = dev->data->dev_private;
1321
1322         vmxnet3_dev_set_rxmode(hw, VMXNET3_RXM_ALL_MULTI, 0);
1323
1324         return 0;
1325 }
1326
1327 /* Enable/disable filter on vlan */
1328 static int
1329 vmxnet3_dev_vlan_filter_set(struct rte_eth_dev *dev, uint16_t vid, int on)
1330 {
1331         struct vmxnet3_hw *hw = dev->data->dev_private;
1332         struct Vmxnet3_RxFilterConf *rxConf = &hw->shared->devRead.rxFilterConf;
1333         uint32_t *vf_table = rxConf->vfTable;
1334
1335         /* save state for restore */
1336         if (on)
1337                 VMXNET3_SET_VFTABLE_ENTRY(hw->shadow_vfta, vid);
1338         else
1339                 VMXNET3_CLEAR_VFTABLE_ENTRY(hw->shadow_vfta, vid);
1340
1341         /* don't change active filter if in promiscuous mode */
1342         if (rxConf->rxMode & VMXNET3_RXM_PROMISC)
1343                 return 0;
1344
1345         /* set in hardware */
1346         if (on)
1347                 VMXNET3_SET_VFTABLE_ENTRY(vf_table, vid);
1348         else
1349                 VMXNET3_CLEAR_VFTABLE_ENTRY(vf_table, vid);
1350
1351         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1352                                VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1353         return 0;
1354 }
1355
1356 static int
1357 vmxnet3_dev_vlan_offload_set(struct rte_eth_dev *dev, int mask)
1358 {
1359         struct vmxnet3_hw *hw = dev->data->dev_private;
1360         Vmxnet3_DSDevRead *devRead = &hw->shared->devRead;
1361         uint32_t *vf_table = devRead->rxFilterConf.vfTable;
1362         uint64_t rx_offloads = dev->data->dev_conf.rxmode.offloads;
1363
1364         if (mask & ETH_VLAN_STRIP_MASK) {
1365                 if (rx_offloads & DEV_RX_OFFLOAD_VLAN_STRIP)
1366                         devRead->misc.uptFeatures |= UPT1_F_RXVLAN;
1367                 else
1368                         devRead->misc.uptFeatures &= ~UPT1_F_RXVLAN;
1369
1370                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1371                                        VMXNET3_CMD_UPDATE_FEATURE);
1372         }
1373
1374         if (mask & ETH_VLAN_FILTER_MASK) {
1375                 if (rx_offloads & DEV_RX_OFFLOAD_VLAN_FILTER)
1376                         memcpy(vf_table, hw->shadow_vfta, VMXNET3_VFT_TABLE_SIZE);
1377                 else
1378                         memset(vf_table, 0xff, VMXNET3_VFT_TABLE_SIZE);
1379
1380                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1381                                        VMXNET3_CMD_UPDATE_VLAN_FILTERS);
1382         }
1383
1384         return 0;
1385 }
1386
1387 static void
1388 vmxnet3_process_events(struct rte_eth_dev *dev)
1389 {
1390         struct vmxnet3_hw *hw = dev->data->dev_private;
1391         uint32_t events = hw->shared->ecr;
1392
1393         if (!events)
1394                 return;
1395
1396         /*
1397          * ECR bits when written with 1b are cleared. Hence write
1398          * events back to ECR so that the bits which were set will be reset.
1399          */
1400         VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_ECR, events);
1401
1402         /* Check if link state has changed */
1403         if (events & VMXNET3_ECR_LINK) {
1404                 PMD_DRV_LOG(DEBUG, "Process events: VMXNET3_ECR_LINK event");
1405                 if (vmxnet3_dev_link_update(dev, 0) == 0)
1406                         _rte_eth_dev_callback_process(dev,
1407                                                       RTE_ETH_EVENT_INTR_LSC,
1408                                                       NULL);
1409         }
1410
1411         /* Check if there is an error on xmit/recv queues */
1412         if (events & (VMXNET3_ECR_TQERR | VMXNET3_ECR_RQERR)) {
1413                 VMXNET3_WRITE_BAR1_REG(hw, VMXNET3_REG_CMD,
1414                                        VMXNET3_CMD_GET_QUEUE_STATUS);
1415
1416                 if (hw->tqd_start->status.stopped)
1417                         PMD_DRV_LOG(ERR, "tq error 0x%x",
1418                                     hw->tqd_start->status.error);
1419
1420                 if (hw->rqd_start->status.stopped)
1421                         PMD_DRV_LOG(ERR, "rq error 0x%x",
1422                                      hw->rqd_start->status.error);
1423
1424                 /* Reset the device */
1425                 /* Have to reset the device */
1426         }
1427
1428         if (events & VMXNET3_ECR_DIC)
1429                 PMD_DRV_LOG(DEBUG, "Device implementation change event.");
1430
1431         if (events & VMXNET3_ECR_DEBUG)
1432                 PMD_DRV_LOG(DEBUG, "Debug event generated by device.");
1433 }
1434
1435 static void
1436 vmxnet3_interrupt_handler(void *param)
1437 {
1438         struct rte_eth_dev *dev = param;
1439         struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
1440
1441         vmxnet3_process_events(dev);
1442
1443         if (rte_intr_ack(&pci_dev->intr_handle) < 0)
1444                 PMD_DRV_LOG(ERR, "interrupt enable failed");
1445 }
1446
1447 RTE_PMD_REGISTER_PCI(net_vmxnet3, rte_vmxnet3_pmd);
1448 RTE_PMD_REGISTER_PCI_TABLE(net_vmxnet3, pci_id_vmxnet3_map);
1449 RTE_PMD_REGISTER_KMOD_DEP(net_vmxnet3, "* igb_uio | uio_pci_generic | vfio-pci");
1450
1451 RTE_INIT(vmxnet3_init_log)
1452 {
1453         vmxnet3_logtype_init = rte_log_register("pmd.net.vmxnet3.init");
1454         if (vmxnet3_logtype_init >= 0)
1455                 rte_log_set_level(vmxnet3_logtype_init, RTE_LOG_NOTICE);
1456         vmxnet3_logtype_driver = rte_log_register("pmd.net.vmxnet3.driver");
1457         if (vmxnet3_logtype_driver >= 0)
1458                 rte_log_set_level(vmxnet3_logtype_driver, RTE_LOG_NOTICE);
1459 }