drivers/net/hyperv/netvsc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/wait.h>
  14 #include <linux/mm.h>
  15 #include <linux/delay.h>
  16 #include <linux/io.h>
  17 #include <linux/slab.h>
  18 #include <linux/netdevice.h>
  19 #include <linux/if_ether.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/prefetch.h>
  23
  24 #include <asm/sync_bitops.h>
  25
  26 #include "hyperv_net.h"
  27 #include "netvsc_trace.h"
  28
  29 /*
  30  * Switch the data path from the synthetic interface to the VF
  31  * interface.
  32  */
  33 void netvsc_switch_datapath(struct net_device *ndev, bool vf)
  34 {
  35         struct net_device_context *net_device_ctx = netdev_priv(ndev);
  36         struct hv_device *dev = net_device_ctx->device_ctx;
  37         struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
  38         struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
  39
  40         memset(init_pkt, 0, sizeof(struct nvsp_message));
  41         init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
  42         if (vf)
  43                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  44                         NVSP_DATAPATH_VF;
  45         else
  46                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  47                         NVSP_DATAPATH_SYNTHETIC;
  48
  49         trace_nvsp_send(ndev, init_pkt);
  50
  51         vmbus_sendpacket(dev->channel, init_pkt,
  52                                sizeof(struct nvsp_message),
  53                                (unsigned long)init_pkt,
  54                                VM_PKT_DATA_INBAND, 0);
  55 }
  56
  57 /* Worker to setup sub channels on initial setup
  58  * Initial hotplug event occurs in softirq context
  59  * and can't wait for channels.
  60  */
  61 static void netvsc_subchan_work(struct work_struct *w)
  62 {
  63         struct netvsc_device *nvdev =
  64                 container_of(w, struct netvsc_device, subchan_work);
  65         struct rndis_device *rdev;
  66         int i, ret;
  67
  68         /* Avoid deadlock with device removal already under RTNL */
  69         if (!rtnl_trylock()) {
  70                 schedule_work(w);
  71                 return;
  72         }
  73
  74         rdev = nvdev->extension;
  75         if (rdev) {
  76                 ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL);
  77                 if (ret == 0) {
  78                         netif_device_attach(rdev->ndev);
  79                 } else {
  80                         /* fallback to only primary channel */
  81                         for (i = 1; i < nvdev->num_chn; i++)
  82                                 netif_napi_del(&nvdev->chan_table[i].napi);
  83
  84                         nvdev->max_chn = 1;
  85                         nvdev->num_chn = 1;
  86                 }
  87         }
  88
  89         rtnl_unlock();
  90 }
  91
  92 static struct netvsc_device *alloc_net_device(void)
  93 {
  94         struct netvsc_device *net_device;
  95
  96         net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
  97         if (!net_device)
  98                 return NULL;
  99
 100         init_waitqueue_head(&net_device->wait_drain);
 101         net_device->destroy = false;
 102         net_device->tx_disable = false;
 103
 104         net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
 105         net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
 106
 107         init_completion(&net_device->channel_init_wait);
 108         init_waitqueue_head(&net_device->subchan_open);
 109         INIT_WORK(&net_device->subchan_work, netvsc_subchan_work);
 110
 111         return net_device;
 112 }
 113
 114 static void free_netvsc_device(struct rcu_head *head)
 115 {
 116         struct netvsc_device *nvdev
 117                 = container_of(head, struct netvsc_device, rcu);
 118         int i;
 119
 120         kfree(nvdev->extension);
 121         vfree(nvdev->recv_buf);
 122         vfree(nvdev->send_buf);
 123         kfree(nvdev->send_section_map);
 124
 125         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 126                 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
 127                 vfree(nvdev->chan_table[i].mrc.slots);
 128         }
 129
 130         kfree(nvdev);
 131 }
 132
 133 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 134 {
 135         call_rcu(&nvdev->rcu, free_netvsc_device);
 136 }
 137
 138 static void netvsc_revoke_recv_buf(struct hv_device *device,
 139                                    struct netvsc_device *net_device,
 140                                    struct net_device *ndev)
 141 {
 142         struct nvsp_message *revoke_packet;
 143         int ret;
 144
 145         /*
 146          * If we got a section count, it means we received a
 147          * SendReceiveBufferComplete msg (ie sent
 148          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 149          * to send a revoke msg here
 150          */
 151         if (net_device->recv_section_cnt) {
 152                 /* Send the revoke receive buffer */
 153                 revoke_packet = &net_device->revoke_packet;
 154                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 155
 156                 revoke_packet->hdr.msg_type =
 157                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 158                 revoke_packet->msg.v1_msg.
 159                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 160
 161                 trace_nvsp_send(ndev, revoke_packet);
 162
 163                 ret = vmbus_sendpacket(device->channel,
 164                                        revoke_packet,
 165                                        sizeof(struct nvsp_message),
 166                                        (unsigned long)revoke_packet,
 167                                        VM_PKT_DATA_INBAND, 0);
 168                 /* If the failure is because the channel is rescinded;
 169                  * ignore the failure since we cannot send on a rescinded
 170                  * channel. This would allow us to properly cleanup
 171                  * even when the channel is rescinded.
 172                  */
 173                 if (device->channel->rescind)
 174                         ret = 0;
 175                 /*
 176                  * If we failed here, we might as well return and
 177                  * have a leak rather than continue and a bugchk
 178                  */
 179                 if (ret != 0) {
 180                         netdev_err(ndev, "unable to send "
 181                                 "revoke receive buffer to netvsp\n");
 182                         return;
 183                 }
 184                 net_device->recv_section_cnt = 0;
 185         }
 186 }
 187
 188 static void netvsc_revoke_send_buf(struct hv_device *device,
 189                                    struct netvsc_device *net_device,
 190                                    struct net_device *ndev)
 191 {
 192         struct nvsp_message *revoke_packet;
 193         int ret;
 194
 195         /* Deal with the send buffer we may have setup.
 196          * If we got a  send section size, it means we received a
 197          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 198          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 199          * to send a revoke msg here
 200          */
 201         if (net_device->send_section_cnt) {
 202                 /* Send the revoke receive buffer */
 203                 revoke_packet = &net_device->revoke_packet;
 204                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 205
 206                 revoke_packet->hdr.msg_type =
 207                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 208                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 209                         NETVSC_SEND_BUFFER_ID;
 210
 211                 trace_nvsp_send(ndev, revoke_packet);
 212
 213                 ret = vmbus_sendpacket(device->channel,
 214                                        revoke_packet,
 215                                        sizeof(struct nvsp_message),
 216                                        (unsigned long)revoke_packet,
 217                                        VM_PKT_DATA_INBAND, 0);
 218
 219                 /* If the failure is because the channel is rescinded;
 220                  * ignore the failure since we cannot send on a rescinded
 221                  * channel. This would allow us to properly cleanup
 222                  * even when the channel is rescinded.
 223                  */
 224                 if (device->channel->rescind)
 225                         ret = 0;
 226
 227                 /* If we failed here, we might as well return and
 228                  * have a leak rather than continue and a bugchk
 229                  */
 230                 if (ret != 0) {
 231                         netdev_err(ndev, "unable to send "
 232                                    "revoke send buffer to netvsp\n");
 233                         return;
 234                 }
 235                 net_device->send_section_cnt = 0;
 236         }
 237 }
 238
 239 static void netvsc_teardown_recv_gpadl(struct hv_device *device,
 240                                        struct netvsc_device *net_device,
 241                                        struct net_device *ndev)
 242 {
 243         int ret;
 244
 245         if (net_device->recv_buf_gpadl_handle) {
 246                 ret = vmbus_teardown_gpadl(device->channel,
 247                                            net_device->recv_buf_gpadl_handle);
 248
 249                 /* If we failed here, we might as well return and have a leak
 250                  * rather than continue and a bugchk
 251                  */
 252                 if (ret != 0) {
 253                         netdev_err(ndev,
 254                                    "unable to teardown receive buffer's gpadl\n");
 255                         return;
 256                 }
 257                 net_device->recv_buf_gpadl_handle = 0;
 258         }
 259 }
 260
 261 static void netvsc_teardown_send_gpadl(struct hv_device *device,
 262                                        struct netvsc_device *net_device,
 263                                        struct net_device *ndev)
 264 {
 265         int ret;
 266
 267         if (net_device->send_buf_gpadl_handle) {
 268                 ret = vmbus_teardown_gpadl(device->channel,
 269                                            net_device->send_buf_gpadl_handle);
 270
 271                 /* If we failed here, we might as well return and have a leak
 272                  * rather than continue and a bugchk
 273                  */
 274                 if (ret != 0) {
 275                         netdev_err(ndev,
 276                                    "unable to teardown send buffer's gpadl\n");
 277                         return;
 278                 }
 279                 net_device->send_buf_gpadl_handle = 0;
 280         }
 281 }
 282
 283 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 284 {
 285         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 286         int node = cpu_to_node(nvchan->channel->target_cpu);
 287         size_t size;
 288
 289         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 290         nvchan->mrc.slots = vzalloc_node(size, node);
 291         if (!nvchan->mrc.slots)
 292                 nvchan->mrc.slots = vzalloc(size);
 293
 294         return nvchan->mrc.slots ? 0 : -ENOMEM;
 295 }
 296
 297 static int netvsc_init_buf(struct hv_device *device,
 298                            struct netvsc_device *net_device,
 299                            const struct netvsc_device_info *device_info)
 300 {
 301         struct nvsp_1_message_send_receive_buffer_complete *resp;
 302         struct net_device *ndev = hv_get_drvdata(device);
 303         struct nvsp_message *init_packet;
 304         unsigned int buf_size;
 305         size_t map_words;
 306         int ret = 0;
 307
 308         /* Get receive buffer area. */
 309         buf_size = device_info->recv_sections * device_info->recv_section_size;
 310         buf_size = roundup(buf_size, PAGE_SIZE);
 311
 312         /* Legacy hosts only allow smaller receive buffer */
 313         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 314                 buf_size = min_t(unsigned int, buf_size,
 315                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 316
 317         net_device->recv_buf = vzalloc(buf_size);
 318         if (!net_device->recv_buf) {
 319                 netdev_err(ndev,
 320                            "unable to allocate receive buffer of size %u\n",
 321                            buf_size);
 322                 ret = -ENOMEM;
 323                 goto cleanup;
 324         }
 325
 326         net_device->recv_buf_size = buf_size;
 327
 328         /*
 329          * Establish the gpadl handle for this buffer on this
 330          * channel.  Note: This call uses the vmbus connection rather
 331          * than the channel to establish the gpadl handle.
 332          */
 333         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 334                                     buf_size,
 335                                     &net_device->recv_buf_gpadl_handle);
 336         if (ret != 0) {
 337                 netdev_err(ndev,
 338                         "unable to establish receive buffer's gpadl\n");
 339                 goto cleanup;
 340         }
 341
 342         /* Notify the NetVsp of the gpadl handle */
 343         init_packet = &net_device->channel_init_pkt;
 344         memset(init_packet, 0, sizeof(struct nvsp_message));
 345         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 346         init_packet->msg.v1_msg.send_recv_buf.
 347                 gpadl_handle = net_device->recv_buf_gpadl_handle;
 348         init_packet->msg.v1_msg.
 349                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 350
 351         trace_nvsp_send(ndev, init_packet);
 352
 353         /* Send the gpadl notification request */
 354         ret = vmbus_sendpacket(device->channel, init_packet,
 355                                sizeof(struct nvsp_message),
 356                                (unsigned long)init_packet,
 357                                VM_PKT_DATA_INBAND,
 358                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 359         if (ret != 0) {
 360                 netdev_err(ndev,
 361                         "unable to send receive buffer's gpadl to netvsp\n");
 362                 goto cleanup;
 363         }
 364
 365         wait_for_completion(&net_device->channel_init_wait);
 366
 367         /* Check the response */
 368         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 369         if (resp->status != NVSP_STAT_SUCCESS) {
 370                 netdev_err(ndev,
 371                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 372                            resp->status);
 373                 ret = -EINVAL;
 374                 goto cleanup;
 375         }
 376
 377         /* Parse the response */
 378         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 379                    resp->num_sections, resp->sections[0].sub_alloc_size,
 380                    resp->sections[0].num_sub_allocs);
 381
 382         /* There should only be one section for the entire receive buffer */
 383         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 384                 ret = -EINVAL;
 385                 goto cleanup;
 386         }
 387
 388         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 389         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 390
 391         /* Setup receive completion ring */
 392         net_device->recv_completion_cnt
 393                 = round_up(net_device->recv_section_cnt + 1,
 394                            PAGE_SIZE / sizeof(u64));
 395         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 396         if (ret)
 397                 goto cleanup;
 398
 399         /* Now setup the send buffer. */
 400         buf_size = device_info->send_sections * device_info->send_section_size;
 401         buf_size = round_up(buf_size, PAGE_SIZE);
 402
 403         net_device->send_buf = vzalloc(buf_size);
 404         if (!net_device->send_buf) {
 405                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 406                            buf_size);
 407                 ret = -ENOMEM;
 408                 goto cleanup;
 409         }
 410
 411         /* Establish the gpadl handle for this buffer on this
 412          * channel.  Note: This call uses the vmbus connection rather
 413          * than the channel to establish the gpadl handle.
 414          */
 415         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 416                                     buf_size,
 417                                     &net_device->send_buf_gpadl_handle);
 418         if (ret != 0) {
 419                 netdev_err(ndev,
 420                            "unable to establish send buffer's gpadl\n");
 421                 goto cleanup;
 422         }
 423
 424         /* Notify the NetVsp of the gpadl handle */
 425         init_packet = &net_device->channel_init_pkt;
 426         memset(init_packet, 0, sizeof(struct nvsp_message));
 427         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 428         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 429                 net_device->send_buf_gpadl_handle;
 430         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 431
 432         trace_nvsp_send(ndev, init_packet);
 433
 434         /* Send the gpadl notification request */
 435         ret = vmbus_sendpacket(device->channel, init_packet,
 436                                sizeof(struct nvsp_message),
 437                                (unsigned long)init_packet,
 438                                VM_PKT_DATA_INBAND,
 439                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 440         if (ret != 0) {
 441                 netdev_err(ndev,
 442                            "unable to send send buffer's gpadl to netvsp\n");
 443                 goto cleanup;
 444         }
 445
 446         wait_for_completion(&net_device->channel_init_wait);
 447
 448         /* Check the response */
 449         if (init_packet->msg.v1_msg.
 450             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 451                 netdev_err(ndev, "Unable to complete send buffer "
 452                            "initialization with NetVsp - status %d\n",
 453                            init_packet->msg.v1_msg.
 454                            send_send_buf_complete.status);
 455                 ret = -EINVAL;
 456                 goto cleanup;
 457         }
 458
 459         /* Parse the response */
 460         net_device->send_section_size = init_packet->msg.
 461                                 v1_msg.send_send_buf_complete.section_size;
 462
 463         /* Section count is simply the size divided by the section size. */
 464         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 465
 466         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 467                    net_device->send_section_size, net_device->send_section_cnt);
 468
 469         /* Setup state for managing the send buffer. */
 470         map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 471
 472         net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 473         if (net_device->send_section_map == NULL) {
 474                 ret = -ENOMEM;
 475                 goto cleanup;
 476         }
 477
 478         goto exit;
 479
 480 cleanup:
 481         netvsc_revoke_recv_buf(device, net_device, ndev);
 482         netvsc_revoke_send_buf(device, net_device, ndev);
 483         netvsc_teardown_recv_gpadl(device, net_device, ndev);
 484         netvsc_teardown_send_gpadl(device, net_device, ndev);
 485
 486 exit:
 487         return ret;
 488 }
 489
 490 /* Negotiate NVSP protocol version */
 491 static int negotiate_nvsp_ver(struct hv_device *device,
 492                               struct netvsc_device *net_device,
 493                               struct nvsp_message *init_packet,
 494                               u32 nvsp_ver)
 495 {
 496         struct net_device *ndev = hv_get_drvdata(device);
 497         int ret;
 498
 499         memset(init_packet, 0, sizeof(struct nvsp_message));
 500         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 501         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 502         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 503         trace_nvsp_send(ndev, init_packet);
 504
 505         /* Send the init request */
 506         ret = vmbus_sendpacket(device->channel, init_packet,
 507                                sizeof(struct nvsp_message),
 508                                (unsigned long)init_packet,
 509                                VM_PKT_DATA_INBAND,
 510                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 511
 512         if (ret != 0)
 513                 return ret;
 514
 515         wait_for_completion(&net_device->channel_init_wait);
 516
 517         if (init_packet->msg.init_msg.init_complete.status !=
 518             NVSP_STAT_SUCCESS)
 519                 return -EINVAL;
 520
 521         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 522                 return 0;
 523
 524         /* NVSPv2 or later: Send NDIS config */
 525         memset(init_packet, 0, sizeof(struct nvsp_message));
 526         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 527         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 528         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 529
 530         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 531                 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 532
 533                 /* Teaming bit is needed to receive link speed updates */
 534                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 535         }
 536
 537         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 538                 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 539
 540         trace_nvsp_send(ndev, init_packet);
 541
 542         ret = vmbus_sendpacket(device->channel, init_packet,
 543                                 sizeof(struct nvsp_message),
 544                                 (unsigned long)init_packet,
 545                                 VM_PKT_DATA_INBAND, 0);
 546
 547         return ret;
 548 }
 549
 550 static int netvsc_connect_vsp(struct hv_device *device,
 551                               struct netvsc_device *net_device,
 552                               const struct netvsc_device_info *device_info)
 553 {
 554         struct net_device *ndev = hv_get_drvdata(device);
 555         static const u32 ver_list[] = {
 556                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 557                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 558                 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 559         };
 560         struct nvsp_message *init_packet;
 561         int ndis_version, i, ret;
 562
 563         init_packet = &net_device->channel_init_pkt;
 564
 565         /* Negotiate the latest NVSP protocol supported */
 566         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 567                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 568                                        ver_list[i])  == 0) {
 569                         net_device->nvsp_version = ver_list[i];
 570                         break;
 571                 }
 572
 573         if (i < 0) {
 574                 ret = -EPROTO;
 575                 goto cleanup;
 576         }
 577
 578         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 579
 580         /* Send the ndis version */
 581         memset(init_packet, 0, sizeof(struct nvsp_message));
 582
 583         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 584                 ndis_version = 0x00060001;
 585         else
 586                 ndis_version = 0x0006001e;
 587
 588         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 589         init_packet->msg.v1_msg.
 590                 send_ndis_ver.ndis_major_ver =
 591                                 (ndis_version & 0xFFFF0000) >> 16;
 592         init_packet->msg.v1_msg.
 593                 send_ndis_ver.ndis_minor_ver =
 594                                 ndis_version & 0xFFFF;
 595
 596         trace_nvsp_send(ndev, init_packet);
 597
 598         /* Send the init request */
 599         ret = vmbus_sendpacket(device->channel, init_packet,
 600                                 sizeof(struct nvsp_message),
 601                                 (unsigned long)init_packet,
 602                                 VM_PKT_DATA_INBAND, 0);
 603         if (ret != 0)
 604                 goto cleanup;
 605
 606
 607         ret = netvsc_init_buf(device, net_device, device_info);
 608
 609 cleanup:
 610         return ret;
 611 }
 612
 613 /*
 614  * netvsc_device_remove - Callback when the root bus device is removed
 615  */
 616 void netvsc_device_remove(struct hv_device *device)
 617 {
 618         struct net_device *ndev = hv_get_drvdata(device);
 619         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 620         struct netvsc_device *net_device
 621                 = rtnl_dereference(net_device_ctx->nvdev);
 622         int i;
 623
 624         /*
 625          * Revoke receive buffer. If host is pre-Win2016 then tear down
 626          * receive buffer GPADL. Do the same for send buffer.
 627          */
 628         netvsc_revoke_recv_buf(device, net_device, ndev);
 629         if (vmbus_proto_version < VERSION_WIN10)
 630                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 631
 632         netvsc_revoke_send_buf(device, net_device, ndev);
 633         if (vmbus_proto_version < VERSION_WIN10)
 634                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 635
 636         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 637
 638         /* And disassociate NAPI context from device */
 639         for (i = 0; i < net_device->num_chn; i++)
 640                 netif_napi_del(&net_device->chan_table[i].napi);
 641
 642         /*
 643          * At this point, no one should be accessing net_device
 644          * except in here
 645          */
 646         netdev_dbg(ndev, "net device safe to remove\n");
 647
 648         /* Now, we can close the channel safely */
 649         vmbus_close(device->channel);
 650
 651         /*
 652          * If host is Win2016 or higher then we do the GPADL tear down
 653          * here after VMBus is closed.
 654         */
 655         if (vmbus_proto_version >= VERSION_WIN10) {
 656                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 657                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 658         }
 659
 660         /* Release all resources */
 661         free_netvsc_device_rcu(net_device);
 662 }
 663
 664 #define RING_AVAIL_PERCENT_HIWATER 20
 665 #define RING_AVAIL_PERCENT_LOWATER 10
 666
 667 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 668                                          u32 index)
 669 {
 670         sync_change_bit(index, net_device->send_section_map);
 671 }
 672
 673 static void netvsc_send_tx_complete(struct net_device *ndev,
 674                                     struct netvsc_device *net_device,
 675                                     struct vmbus_channel *channel,
 676                                     const struct vmpacket_descriptor *desc,
 677                                     int budget)
 678 {
 679         struct sk_buff *skb = (struct sk_buff *)(unsigned long)desc->trans_id;
 680         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 681         u16 q_idx = 0;
 682         int queue_sends;
 683
 684         /* Notify the layer above us */
 685         if (likely(skb)) {
 686                 const struct hv_netvsc_packet *packet
 687                         = (struct hv_netvsc_packet *)skb->cb;
 688                 u32 send_index = packet->send_buf_index;
 689                 struct netvsc_stats *tx_stats;
 690
 691                 if (send_index != NETVSC_INVALID_INDEX)
 692                         netvsc_free_send_slot(net_device, send_index);
 693                 q_idx = packet->q_idx;
 694
 695                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 696
 697                 u64_stats_update_begin(&tx_stats->syncp);
 698                 tx_stats->packets += packet->total_packets;
 699                 tx_stats->bytes += packet->total_bytes;
 700                 u64_stats_update_end(&tx_stats->syncp);
 701
 702                 napi_consume_skb(skb, budget);
 703         }
 704
 705         queue_sends =
 706                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 707
 708         if (unlikely(net_device->destroy)) {
 709                 if (queue_sends == 0)
 710                         wake_up(&net_device->wait_drain);
 711         } else {
 712                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 713
 714                 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 715                     (hv_get_avail_to_write_percent(&channel->outbound) >
 716                      RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 717                         netif_tx_wake_queue(txq);
 718                         ndev_ctx->eth_stats.wake_queue++;
 719                 }
 720         }
 721 }
 722
 723 static void netvsc_send_completion(struct net_device *ndev,
 724                                    struct netvsc_device *net_device,
 725                                    struct vmbus_channel *incoming_channel,
 726                                    const struct vmpacket_descriptor *desc,
 727                                    int budget)
 728 {
 729         const struct nvsp_message *nvsp_packet = hv_pkt_data(desc);
 730
 731         switch (nvsp_packet->hdr.msg_type) {
 732         case NVSP_MSG_TYPE_INIT_COMPLETE:
 733         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 734         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 735         case NVSP_MSG5_TYPE_SUBCHANNEL:
 736                 /* Copy the response back */
 737                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 738                        sizeof(struct nvsp_message));
 739                 complete(&net_device->channel_init_wait);
 740                 break;
 741
 742         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 743                 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 744                                         desc, budget);
 745                 break;
 746
 747         default:
 748                 netdev_err(ndev,
 749                            "Unknown send completion type %d received!!\n",
 750                            nvsp_packet->hdr.msg_type);
 751         }
 752 }
 753
 754 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 755 {
 756         unsigned long *map_addr = net_device->send_section_map;
 757         unsigned int i;
 758
 759         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 760                 if (sync_test_and_set_bit(i, map_addr) == 0)
 761                         return i;
 762         }
 763
 764         return NETVSC_INVALID_INDEX;
 765 }
 766
 767 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 768                                     unsigned int section_index,
 769                                     u32 pend_size,
 770                                     struct hv_netvsc_packet *packet,
 771                                     struct rndis_message *rndis_msg,
 772                                     struct hv_page_buffer *pb,
 773                                     bool xmit_more)
 774 {
 775         char *start = net_device->send_buf;
 776         char *dest = start + (section_index * net_device->send_section_size)
 777                      + pend_size;
 778         int i;
 779         u32 padding = 0;
 780         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 781                 packet->page_buf_cnt;
 782         u32 remain;
 783
 784         /* Add padding */
 785         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 786         if (xmit_more && remain) {
 787                 padding = net_device->pkt_align - remain;
 788                 rndis_msg->msg_len += padding;
 789                 packet->total_data_buflen += padding;
 790         }
 791
 792         for (i = 0; i < page_count; i++) {
 793                 char *src = phys_to_virt(pb[i].pfn << PAGE_SHIFT);
 794                 u32 offset = pb[i].offset;
 795                 u32 len = pb[i].len;
 796
 797                 memcpy(dest, (src + offset), len);
 798                 dest += len;
 799         }
 800
 801         if (padding)
 802                 memset(dest, 0, padding);
 803 }
 804
 805 static inline int netvsc_send_pkt(
 806         struct hv_device *device,
 807         struct hv_netvsc_packet *packet,
 808         struct netvsc_device *net_device,
 809         struct hv_page_buffer *pb,
 810         struct sk_buff *skb)
 811 {
 812         struct nvsp_message nvmsg;
 813         struct nvsp_1_message_send_rndis_packet *rpkt =
 814                 &nvmsg.msg.v1_msg.send_rndis_pkt;
 815         struct netvsc_channel * const nvchan =
 816                 &net_device->chan_table[packet->q_idx];
 817         struct vmbus_channel *out_channel = nvchan->channel;
 818         struct net_device *ndev = hv_get_drvdata(device);
 819         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 820         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
 821         u64 req_id;
 822         int ret;
 823         u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
 824
 825         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 826         if (skb)
 827                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
 828         else
 829                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
 830
 831         rpkt->send_buf_section_index = packet->send_buf_index;
 832         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
 833                 rpkt->send_buf_section_size = 0;
 834         else
 835                 rpkt->send_buf_section_size = packet->total_data_buflen;
 836
 837         req_id = (ulong)skb;
 838
 839         if (out_channel->rescind)
 840                 return -ENODEV;
 841
 842         trace_nvsp_send_pkt(ndev, out_channel, rpkt);
 843
 844         if (packet->page_buf_cnt) {
 845                 if (packet->cp_partial)
 846                         pb += packet->rmsg_pgcnt;
 847
 848                 ret = vmbus_sendpacket_pagebuffer(out_channel,
 849                                                   pb, packet->page_buf_cnt,
 850                                                   &nvmsg, sizeof(nvmsg),
 851                                                   req_id);
 852         } else {
 853                 ret = vmbus_sendpacket(out_channel,
 854                                        &nvmsg, sizeof(nvmsg),
 855                                        req_id, VM_PKT_DATA_INBAND,
 856                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 857         }
 858
 859         if (ret == 0) {
 860                 atomic_inc_return(&nvchan->queue_sends);
 861
 862                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
 863                         netif_tx_stop_queue(txq);
 864                         ndev_ctx->eth_stats.stop_queue++;
 865                 }
 866         } else if (ret == -EAGAIN) {
 867                 netif_tx_stop_queue(txq);
 868                 ndev_ctx->eth_stats.stop_queue++;
 869         } else {
 870                 netdev_err(ndev,
 871                            "Unable to send packet pages %u len %u, ret %d\n",
 872                            packet->page_buf_cnt, packet->total_data_buflen,
 873                            ret);
 874         }
 875
 876         if (netif_tx_queue_stopped(txq) &&
 877             atomic_read(&nvchan->queue_sends) < 1 &&
 878             !net_device->tx_disable) {
 879                 netif_tx_wake_queue(txq);
 880                 ndev_ctx->eth_stats.wake_queue++;
 881                 if (ret == -EAGAIN)
 882                         ret = -ENOSPC;
 883         }
 884
 885         return ret;
 886 }
 887
 888 /* Move packet out of multi send data (msd), and clear msd */
 889 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
 890                                 struct sk_buff **msd_skb,
 891                                 struct multi_send_data *msdp)
 892 {
 893         *msd_skb = msdp->skb;
 894         *msd_send = msdp->pkt;
 895         msdp->skb = NULL;
 896         msdp->pkt = NULL;
 897         msdp->count = 0;
 898 }
 899
 900 /* RCU already held by caller */
 901 int netvsc_send(struct net_device *ndev,
 902                 struct hv_netvsc_packet *packet,
 903                 struct rndis_message *rndis_msg,
 904                 struct hv_page_buffer *pb,
 905                 struct sk_buff *skb,
 906                 bool xdp_tx)
 907 {
 908         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 909         struct netvsc_device *net_device
 910                 = rcu_dereference_bh(ndev_ctx->nvdev);
 911         struct hv_device *device = ndev_ctx->device_ctx;
 912         int ret = 0;
 913         struct netvsc_channel *nvchan;
 914         u32 pktlen = packet->total_data_buflen, msd_len = 0;
 915         unsigned int section_index = NETVSC_INVALID_INDEX;
 916         struct multi_send_data *msdp;
 917         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
 918         struct sk_buff *msd_skb = NULL;
 919         bool try_batch, xmit_more;
 920
 921         /* If device is rescinded, return error and packet will get dropped. */
 922         if (unlikely(!net_device || net_device->destroy))
 923                 return -ENODEV;
 924
 925         nvchan = &net_device->chan_table[packet->q_idx];
 926         packet->send_buf_index = NETVSC_INVALID_INDEX;
 927         packet->cp_partial = false;
 928
 929         /* Send a control message or XDP packet directly without accessing
 930          * msd (Multi-Send Data) field which may be changed during data packet
 931          * processing.
 932          */
 933         if (!skb || xdp_tx)
 934                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
 935
 936         /* batch packets in send buffer if possible */
 937         msdp = &nvchan->msd;
 938         if (msdp->pkt)
 939                 msd_len = msdp->pkt->total_data_buflen;
 940
 941         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
 942         if (try_batch && msd_len + pktlen + net_device->pkt_align <
 943             net_device->send_section_size) {
 944                 section_index = msdp->pkt->send_buf_index;
 945
 946         } else if (try_batch && msd_len + packet->rmsg_size <
 947                    net_device->send_section_size) {
 948                 section_index = msdp->pkt->send_buf_index;
 949                 packet->cp_partial = true;
 950
 951         } else if (pktlen + net_device->pkt_align <
 952                    net_device->send_section_size) {
 953                 section_index = netvsc_get_next_send_section(net_device);
 954                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
 955                         ++ndev_ctx->eth_stats.tx_send_full;
 956                 } else {
 957                         move_pkt_msd(&msd_send, &msd_skb, msdp);
 958                         msd_len = 0;
 959                 }
 960         }
 961
 962         /* Keep aggregating only if stack says more data is coming
 963          * and not doing mixed modes send and not flow blocked
 964          */
 965         xmit_more = netdev_xmit_more() &&
 966                 !packet->cp_partial &&
 967                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
 968
 969         if (section_index != NETVSC_INVALID_INDEX) {
 970                 netvsc_copy_to_send_buf(net_device,
 971                                         section_index, msd_len,
 972                                         packet, rndis_msg, pb, xmit_more);
 973
 974                 packet->send_buf_index = section_index;
 975
 976                 if (packet->cp_partial) {
 977                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
 978                         packet->total_data_buflen = msd_len + packet->rmsg_size;
 979                 } else {
 980                         packet->page_buf_cnt = 0;
 981                         packet->total_data_buflen += msd_len;
 982                 }
 983
 984                 if (msdp->pkt) {
 985                         packet->total_packets += msdp->pkt->total_packets;
 986                         packet->total_bytes += msdp->pkt->total_bytes;
 987                 }
 988
 989                 if (msdp->skb)
 990                         dev_consume_skb_any(msdp->skb);
 991
 992                 if (xmit_more) {
 993                         msdp->skb = skb;
 994                         msdp->pkt = packet;
 995                         msdp->count++;
 996                 } else {
 997                         cur_send = packet;
 998                         msdp->skb = NULL;
 999                         msdp->pkt = NULL;
1000                         msdp->count = 0;
1001                 }
1002         } else {
1003                 move_pkt_msd(&msd_send, &msd_skb, msdp);
1004                 cur_send = packet;
1005         }
1006
1007         if (msd_send) {
1008                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1009                                             NULL, msd_skb);
1010
1011                 if (m_ret != 0) {
1012                         netvsc_free_send_slot(net_device,
1013                                               msd_send->send_buf_index);
1014                         dev_kfree_skb_any(msd_skb);
1015                 }
1016         }
1017
1018         if (cur_send)
1019                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1020
1021         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1022                 netvsc_free_send_slot(net_device, section_index);
1023
1024         return ret;
1025 }
1026
1027 /* Send pending recv completions */
1028 static int send_recv_completions(struct net_device *ndev,
1029                                  struct netvsc_device *nvdev,
1030                                  struct netvsc_channel *nvchan)
1031 {
1032         struct multi_recv_comp *mrc = &nvchan->mrc;
1033         struct recv_comp_msg {
1034                 struct nvsp_message_header hdr;
1035                 u32 status;
1036         }  __packed;
1037         struct recv_comp_msg msg = {
1038                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1039         };
1040         int ret;
1041
1042         while (mrc->first != mrc->next) {
1043                 const struct recv_comp_data *rcd
1044                         = mrc->slots + mrc->first;
1045
1046                 msg.status = rcd->status;
1047                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1048                                        rcd->tid, VM_PKT_COMP, 0);
1049                 if (unlikely(ret)) {
1050                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1051
1052                         ++ndev_ctx->eth_stats.rx_comp_busy;
1053                         return ret;
1054                 }
1055
1056                 if (++mrc->first == nvdev->recv_completion_cnt)
1057                         mrc->first = 0;
1058         }
1059
1060         /* receive completion ring has been emptied */
1061         if (unlikely(nvdev->destroy))
1062                 wake_up(&nvdev->wait_drain);
1063
1064         return 0;
1065 }
1066
1067 /* Count how many receive completions are outstanding */
1068 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1069                                  const struct multi_recv_comp *mrc,
1070                                  u32 *filled, u32 *avail)
1071 {
1072         u32 count = nvdev->recv_completion_cnt;
1073
1074         if (mrc->next >= mrc->first)
1075                 *filled = mrc->next - mrc->first;
1076         else
1077                 *filled = (count - mrc->first) + mrc->next;
1078
1079         *avail = count - *filled - 1;
1080 }
1081
1082 /* Add receive complete to ring to send to host. */
1083 static void enq_receive_complete(struct net_device *ndev,
1084                                  struct netvsc_device *nvdev, u16 q_idx,
1085                                  u64 tid, u32 status)
1086 {
1087         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1088         struct multi_recv_comp *mrc = &nvchan->mrc;
1089         struct recv_comp_data *rcd;
1090         u32 filled, avail;
1091
1092         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1093
1094         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1095                 send_recv_completions(ndev, nvdev, nvchan);
1096                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1097         }
1098
1099         if (unlikely(!avail)) {
1100                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1101                            q_idx, tid);
1102                 return;
1103         }
1104
1105         rcd = mrc->slots + mrc->next;
1106         rcd->tid = tid;
1107         rcd->status = status;
1108
1109         if (++mrc->next == nvdev->recv_completion_cnt)
1110                 mrc->next = 0;
1111 }
1112
1113 static int netvsc_receive(struct net_device *ndev,
1114                           struct netvsc_device *net_device,
1115                           struct netvsc_channel *nvchan,
1116                           const struct vmpacket_descriptor *desc,
1117                           const struct nvsp_message *nvsp)
1118 {
1119         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1120         struct vmbus_channel *channel = nvchan->channel;
1121         const struct vmtransfer_page_packet_header *vmxferpage_packet
1122                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1123         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1124         char *recv_buf = net_device->recv_buf;
1125         u32 status = NVSP_STAT_SUCCESS;
1126         int i;
1127         int count = 0;
1128
1129         /* Make sure this is a valid nvsp packet */
1130         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1131                 netif_err(net_device_ctx, rx_err, ndev,
1132                           "Unknown nvsp packet type received %u\n",
1133                           nvsp->hdr.msg_type);
1134                 return 0;
1135         }
1136
1137         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1138                 netif_err(net_device_ctx, rx_err, ndev,
1139                           "Invalid xfer page set id - expecting %x got %x\n",
1140                           NETVSC_RECEIVE_BUFFER_ID,
1141                           vmxferpage_packet->xfer_pageset_id);
1142                 return 0;
1143         }
1144
1145         count = vmxferpage_packet->range_cnt;
1146
1147         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1148         for (i = 0; i < count; i++) {
1149                 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1150                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1151                 void *data;
1152                 int ret;
1153
1154                 if (unlikely(offset + buflen > net_device->recv_buf_size)) {
1155                         nvchan->rsc.cnt = 0;
1156                         status = NVSP_STAT_FAIL;
1157                         netif_err(net_device_ctx, rx_err, ndev,
1158                                   "Packet offset:%u + len:%u too big\n",
1159                                   offset, buflen);
1160
1161                         continue;
1162                 }
1163
1164                 data = recv_buf + offset;
1165
1166                 nvchan->rsc.is_last = (i == count - 1);
1167
1168                 trace_rndis_recv(ndev, q_idx, data);
1169
1170                 /* Pass it to the upper layer */
1171                 ret = rndis_filter_receive(ndev, net_device,
1172                                            nvchan, data, buflen);
1173
1174                 if (unlikely(ret != NVSP_STAT_SUCCESS))
1175                         status = NVSP_STAT_FAIL;
1176         }
1177
1178         enq_receive_complete(ndev, net_device, q_idx,
1179                              vmxferpage_packet->d.trans_id, status);
1180
1181         return count;
1182 }
1183
1184 static void netvsc_send_table(struct net_device *ndev,
1185                               struct netvsc_device *nvscdev,
1186                               const struct nvsp_message *nvmsg,
1187                               u32 msglen)
1188 {
1189         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1190         u32 count, offset, *tab;
1191         int i;
1192
1193         count = nvmsg->msg.v5_msg.send_table.count;
1194         offset = nvmsg->msg.v5_msg.send_table.offset;
1195
1196         if (count != VRSS_SEND_TAB_SIZE) {
1197                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1198                 return;
1199         }
1200
1201         /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1202          * wrong due to a host bug. So fix the offset here.
1203          */
1204         if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1205             msglen >= sizeof(struct nvsp_message_header) +
1206             sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1207                 offset = sizeof(struct nvsp_message_header) +
1208                          sizeof(union nvsp_6_message_uber);
1209
1210         /* Boundary check for all versions */
1211         if (offset > msglen - count * sizeof(u32)) {
1212                 netdev_err(ndev, "Received send-table offset too big:%u\n",
1213                            offset);
1214                 return;
1215         }
1216
1217         tab = (void *)nvmsg + offset;
1218
1219         for (i = 0; i < count; i++)
1220                 net_device_ctx->tx_table[i] = tab[i];
1221 }
1222
1223 static void netvsc_send_vf(struct net_device *ndev,
1224                            const struct nvsp_message *nvmsg)
1225 {
1226         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1227
1228         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1229         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1230         netdev_info(ndev, "VF slot %u %s\n",
1231                     net_device_ctx->vf_serial,
1232                     net_device_ctx->vf_alloc ? "added" : "removed");
1233 }
1234
1235 static void netvsc_receive_inband(struct net_device *ndev,
1236                                   struct netvsc_device *nvscdev,
1237                                   const struct nvsp_message *nvmsg,
1238                                   u32 msglen)
1239 {
1240         switch (nvmsg->hdr.msg_type) {
1241         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1242                 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1243                 break;
1244
1245         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1246                 netvsc_send_vf(ndev, nvmsg);
1247                 break;
1248         }
1249 }
1250
1251 static int netvsc_process_raw_pkt(struct hv_device *device,
1252                                   struct netvsc_channel *nvchan,
1253                                   struct netvsc_device *net_device,
1254                                   struct net_device *ndev,
1255                                   const struct vmpacket_descriptor *desc,
1256                                   int budget)
1257 {
1258         struct vmbus_channel *channel = nvchan->channel;
1259         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1260         u32 msglen = hv_pkt_datalen(desc);
1261
1262         trace_nvsp_recv(ndev, channel, nvmsg);
1263
1264         switch (desc->type) {
1265         case VM_PKT_COMP:
1266                 netvsc_send_completion(ndev, net_device, channel,
1267                                        desc, budget);
1268                 break;
1269
1270         case VM_PKT_DATA_USING_XFER_PAGES:
1271                 return netvsc_receive(ndev, net_device, nvchan,
1272                                       desc, nvmsg);
1273                 break;
1274
1275         case VM_PKT_DATA_INBAND:
1276                 netvsc_receive_inband(ndev, net_device, nvmsg, msglen);
1277                 break;
1278
1279         default:
1280                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1281                            desc->type, desc->trans_id);
1282                 break;
1283         }
1284
1285         return 0;
1286 }
1287
1288 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1289 {
1290         struct vmbus_channel *primary = channel->primary_channel;
1291
1292         return primary ? primary->device_obj : channel->device_obj;
1293 }
1294
1295 /* Network processing softirq
1296  * Process data in incoming ring buffer from host
1297  * Stops when ring is empty or budget is met or exceeded.
1298  */
1299 int netvsc_poll(struct napi_struct *napi, int budget)
1300 {
1301         struct netvsc_channel *nvchan
1302                 = container_of(napi, struct netvsc_channel, napi);
1303         struct netvsc_device *net_device = nvchan->net_device;
1304         struct vmbus_channel *channel = nvchan->channel;
1305         struct hv_device *device = netvsc_channel_to_device(channel);
1306         struct net_device *ndev = hv_get_drvdata(device);
1307         int work_done = 0;
1308         int ret;
1309
1310         /* If starting a new interval */
1311         if (!nvchan->desc)
1312                 nvchan->desc = hv_pkt_iter_first(channel);
1313
1314         while (nvchan->desc && work_done < budget) {
1315                 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1316                                                     ndev, nvchan->desc, budget);
1317                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1318         }
1319
1320         /* Send any pending receive completions */
1321         ret = send_recv_completions(ndev, net_device, nvchan);
1322
1323         /* If it did not exhaust NAPI budget this time
1324          *  and not doing busy poll
1325          * then re-enable host interrupts
1326          *  and reschedule if ring is not empty
1327          *   or sending receive completion failed.
1328          */
1329         if (work_done < budget &&
1330             napi_complete_done(napi, work_done) &&
1331             (ret || hv_end_read(&channel->inbound)) &&
1332             napi_schedule_prep(napi)) {
1333                 hv_begin_read(&channel->inbound);
1334                 __napi_schedule(napi);
1335         }
1336
1337         /* Driver may overshoot since multiple packets per descriptor */
1338         return min(work_done, budget);
1339 }
1340
1341 /* Call back when data is available in host ring buffer.
1342  * Processing is deferred until network softirq (NAPI)
1343  */
1344 void netvsc_channel_cb(void *context)
1345 {
1346         struct netvsc_channel *nvchan = context;
1347         struct vmbus_channel *channel = nvchan->channel;
1348         struct hv_ring_buffer_info *rbi = &channel->inbound;
1349
1350         /* preload first vmpacket descriptor */
1351         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1352
1353         if (napi_schedule_prep(&nvchan->napi)) {
1354                 /* disable interrupts from host */
1355                 hv_begin_read(rbi);
1356
1357                 __napi_schedule_irqoff(&nvchan->napi);
1358         }
1359 }
1360
1361 /*
1362  * netvsc_device_add - Callback when the device belonging to this
1363  * driver is added
1364  */
1365 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1366                                 const struct netvsc_device_info *device_info)
1367 {
1368         int i, ret = 0;
1369         struct netvsc_device *net_device;
1370         struct net_device *ndev = hv_get_drvdata(device);
1371         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1372
1373         net_device = alloc_net_device();
1374         if (!net_device)
1375                 return ERR_PTR(-ENOMEM);
1376
1377         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1378                 net_device_ctx->tx_table[i] = 0;
1379
1380         /* Because the device uses NAPI, all the interrupt batching and
1381          * control is done via Net softirq, not the channel handling
1382          */
1383         set_channel_read_mode(device->channel, HV_CALL_ISR);
1384
1385         /* If we're reopening the device we may have multiple queues, fill the
1386          * chn_table with the default channel to use it before subchannels are
1387          * opened.
1388          * Initialize the channel state before we open;
1389          * we can be interrupted as soon as we open the channel.
1390          */
1391
1392         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1393                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1394
1395                 nvchan->channel = device->channel;
1396                 nvchan->net_device = net_device;
1397                 u64_stats_init(&nvchan->tx_stats.syncp);
1398                 u64_stats_init(&nvchan->rx_stats.syncp);
1399
1400                 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i);
1401
1402                 if (ret) {
1403                         netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1404                         goto cleanup2;
1405                 }
1406
1407                 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1408                                                  MEM_TYPE_PAGE_SHARED, NULL);
1409
1410                 if (ret) {
1411                         netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1412                         goto cleanup2;
1413                 }
1414         }
1415
1416         /* Enable NAPI handler before init callbacks */
1417         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1418                        netvsc_poll, NAPI_POLL_WEIGHT);
1419
1420         /* Open the channel */
1421         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1422                          netvsc_ring_bytes,  NULL, 0,
1423                          netvsc_channel_cb, net_device->chan_table);
1424
1425         if (ret != 0) {
1426                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1427                 goto cleanup;
1428         }
1429
1430         /* Channel is opened */
1431         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1432
1433         napi_enable(&net_device->chan_table[0].napi);
1434
1435         /* Connect with the NetVsp */
1436         ret = netvsc_connect_vsp(device, net_device, device_info);
1437         if (ret != 0) {
1438                 netdev_err(ndev,
1439                         "unable to connect to NetVSP - %d\n", ret);
1440                 goto close;
1441         }
1442
1443         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1444          * populated.
1445          */
1446         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1447
1448         return net_device;
1449
1450 close:
1451         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1452         napi_disable(&net_device->chan_table[0].napi);
1453
1454         /* Now, we can close the channel safely */
1455         vmbus_close(device->channel);
1456
1457 cleanup:
1458         netif_napi_del(&net_device->chan_table[0].napi);
1459
1460 cleanup2:
1461         free_netvsc_device(&net_device->rcu);
1462
1463         return ERR_PTR(ret);
1464 }