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 #include <linux/filter.h>
  24
  25 #include <asm/sync_bitops.h>
  26 #include <asm/mshyperv.h>
  27
  28 #include "hyperv_net.h"
  29 #include "netvsc_trace.h"
  30
  31 /*
  32  * Switch the data path from the synthetic interface to the VF
  33  * interface.
  34  */
  35 int netvsc_switch_datapath(struct net_device *ndev, bool vf)
  36 {
  37         struct net_device_context *net_device_ctx = netdev_priv(ndev);
  38         struct hv_device *dev = net_device_ctx->device_ctx;
  39         struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
  40         struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
  41         int ret, retry = 0;
  42
  43         /* Block sending traffic to VF if it's about to be gone */
  44         if (!vf)
  45                 net_device_ctx->data_path_is_vf = vf;
  46
  47         memset(init_pkt, 0, sizeof(struct nvsp_message));
  48         init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
  49         if (vf)
  50                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  51                         NVSP_DATAPATH_VF;
  52         else
  53                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  54                         NVSP_DATAPATH_SYNTHETIC;
  55
  56 again:
  57         trace_nvsp_send(ndev, init_pkt);
  58
  59         ret = vmbus_sendpacket(dev->channel, init_pkt,
  60                                sizeof(struct nvsp_message),
  61                                (unsigned long)init_pkt, VM_PKT_DATA_INBAND,
  62                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
  63
  64         /* If failed to switch to/from VF, let data_path_is_vf stay false,
  65          * so we use synthetic path to send data.
  66          */
  67         if (ret) {
  68                 if (ret != -EAGAIN) {
  69                         netdev_err(ndev,
  70                                    "Unable to send sw datapath msg, err: %d\n",
  71                                    ret);
  72                         return ret;
  73                 }
  74
  75                 if (retry++ < RETRY_MAX) {
  76                         usleep_range(RETRY_US_LO, RETRY_US_HI);
  77                         goto again;
  78                 } else {
  79                         netdev_err(
  80                                 ndev,
  81                                 "Retry failed to send sw datapath msg, err: %d\n",
  82                                 ret);
  83                         return ret;
  84                 }
  85         }
  86
  87         wait_for_completion(&nv_dev->channel_init_wait);
  88         net_device_ctx->data_path_is_vf = vf;
  89
  90         return 0;
  91 }
  92
  93 /* Worker to setup sub channels on initial setup
  94  * Initial hotplug event occurs in softirq context
  95  * and can't wait for channels.
  96  */
  97 static void netvsc_subchan_work(struct work_struct *w)
  98 {
  99         struct netvsc_device *nvdev =
 100                 container_of(w, struct netvsc_device, subchan_work);
 101         struct rndis_device *rdev;
 102         int i, ret;
 103
 104         /* Avoid deadlock with device removal already under RTNL */
 105         if (!rtnl_trylock()) {
 106                 schedule_work(w);
 107                 return;
 108         }
 109
 110         rdev = nvdev->extension;
 111         if (rdev) {
 112                 ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL);
 113                 if (ret == 0) {
 114                         netif_device_attach(rdev->ndev);
 115                 } else {
 116                         /* fallback to only primary channel */
 117                         for (i = 1; i < nvdev->num_chn; i++)
 118                                 netif_napi_del(&nvdev->chan_table[i].napi);
 119
 120                         nvdev->max_chn = 1;
 121                         nvdev->num_chn = 1;
 122                 }
 123         }
 124
 125         rtnl_unlock();
 126 }
 127
 128 static struct netvsc_device *alloc_net_device(void)
 129 {
 130         struct netvsc_device *net_device;
 131
 132         net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
 133         if (!net_device)
 134                 return NULL;
 135
 136         init_waitqueue_head(&net_device->wait_drain);
 137         net_device->destroy = false;
 138         net_device->tx_disable = true;
 139
 140         net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
 141         net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
 142
 143         init_completion(&net_device->channel_init_wait);
 144         init_waitqueue_head(&net_device->subchan_open);
 145         INIT_WORK(&net_device->subchan_work, netvsc_subchan_work);
 146
 147         return net_device;
 148 }
 149
 150 static void free_netvsc_device(struct rcu_head *head)
 151 {
 152         struct netvsc_device *nvdev
 153                 = container_of(head, struct netvsc_device, rcu);
 154         int i;
 155
 156         kfree(nvdev->extension);
 157         vfree(nvdev->recv_buf);
 158         vfree(nvdev->send_buf);
 159         bitmap_free(nvdev->send_section_map);
 160
 161         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 162                 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
 163                 kfree(nvdev->chan_table[i].recv_buf);
 164                 vfree(nvdev->chan_table[i].mrc.slots);
 165         }
 166
 167         kfree(nvdev);
 168 }
 169
 170 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 171 {
 172         call_rcu(&nvdev->rcu, free_netvsc_device);
 173 }
 174
 175 static void netvsc_revoke_recv_buf(struct hv_device *device,
 176                                    struct netvsc_device *net_device,
 177                                    struct net_device *ndev)
 178 {
 179         struct nvsp_message *revoke_packet;
 180         int ret;
 181
 182         /*
 183          * If we got a section count, it means we received a
 184          * SendReceiveBufferComplete msg (ie sent
 185          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 186          * to send a revoke msg here
 187          */
 188         if (net_device->recv_section_cnt) {
 189                 /* Send the revoke receive buffer */
 190                 revoke_packet = &net_device->revoke_packet;
 191                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 192
 193                 revoke_packet->hdr.msg_type =
 194                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 195                 revoke_packet->msg.v1_msg.
 196                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 197
 198                 trace_nvsp_send(ndev, revoke_packet);
 199
 200                 ret = vmbus_sendpacket(device->channel,
 201                                        revoke_packet,
 202                                        sizeof(struct nvsp_message),
 203                                        VMBUS_RQST_ID_NO_RESPONSE,
 204                                        VM_PKT_DATA_INBAND, 0);
 205                 /* If the failure is because the channel is rescinded;
 206                  * ignore the failure since we cannot send on a rescinded
 207                  * channel. This would allow us to properly cleanup
 208                  * even when the channel is rescinded.
 209                  */
 210                 if (device->channel->rescind)
 211                         ret = 0;
 212                 /*
 213                  * If we failed here, we might as well return and
 214                  * have a leak rather than continue and a bugchk
 215                  */
 216                 if (ret != 0) {
 217                         netdev_err(ndev, "unable to send "
 218                                 "revoke receive buffer to netvsp\n");
 219                         return;
 220                 }
 221                 net_device->recv_section_cnt = 0;
 222         }
 223 }
 224
 225 static void netvsc_revoke_send_buf(struct hv_device *device,
 226                                    struct netvsc_device *net_device,
 227                                    struct net_device *ndev)
 228 {
 229         struct nvsp_message *revoke_packet;
 230         int ret;
 231
 232         /* Deal with the send buffer we may have setup.
 233          * If we got a  send section size, it means we received a
 234          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 235          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 236          * to send a revoke msg here
 237          */
 238         if (net_device->send_section_cnt) {
 239                 /* Send the revoke receive buffer */
 240                 revoke_packet = &net_device->revoke_packet;
 241                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 242
 243                 revoke_packet->hdr.msg_type =
 244                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 245                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 246                         NETVSC_SEND_BUFFER_ID;
 247
 248                 trace_nvsp_send(ndev, revoke_packet);
 249
 250                 ret = vmbus_sendpacket(device->channel,
 251                                        revoke_packet,
 252                                        sizeof(struct nvsp_message),
 253                                        VMBUS_RQST_ID_NO_RESPONSE,
 254                                        VM_PKT_DATA_INBAND, 0);
 255
 256                 /* If the failure is because the channel is rescinded;
 257                  * ignore the failure since we cannot send on a rescinded
 258                  * channel. This would allow us to properly cleanup
 259                  * even when the channel is rescinded.
 260                  */
 261                 if (device->channel->rescind)
 262                         ret = 0;
 263
 264                 /* If we failed here, we might as well return and
 265                  * have a leak rather than continue and a bugchk
 266                  */
 267                 if (ret != 0) {
 268                         netdev_err(ndev, "unable to send "
 269                                    "revoke send buffer to netvsp\n");
 270                         return;
 271                 }
 272                 net_device->send_section_cnt = 0;
 273         }
 274 }
 275
 276 static void netvsc_teardown_recv_gpadl(struct hv_device *device,
 277                                        struct netvsc_device *net_device,
 278                                        struct net_device *ndev)
 279 {
 280         int ret;
 281
 282         if (net_device->recv_buf_gpadl_handle.gpadl_handle) {
 283                 ret = vmbus_teardown_gpadl(device->channel,
 284                                            &net_device->recv_buf_gpadl_handle);
 285
 286                 /* If we failed here, we might as well return and have a leak
 287                  * rather than continue and a bugchk
 288                  */
 289                 if (ret != 0) {
 290                         netdev_err(ndev,
 291                                    "unable to teardown receive buffer's gpadl\n");
 292                         return;
 293                 }
 294         }
 295 }
 296
 297 static void netvsc_teardown_send_gpadl(struct hv_device *device,
 298                                        struct netvsc_device *net_device,
 299                                        struct net_device *ndev)
 300 {
 301         int ret;
 302
 303         if (net_device->send_buf_gpadl_handle.gpadl_handle) {
 304                 ret = vmbus_teardown_gpadl(device->channel,
 305                                            &net_device->send_buf_gpadl_handle);
 306
 307                 /* If we failed here, we might as well return and have a leak
 308                  * rather than continue and a bugchk
 309                  */
 310                 if (ret != 0) {
 311                         netdev_err(ndev,
 312                                    "unable to teardown send buffer's gpadl\n");
 313                         return;
 314                 }
 315         }
 316 }
 317
 318 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 319 {
 320         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 321         int node = cpu_to_node(nvchan->channel->target_cpu);
 322         size_t size;
 323
 324         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 325         nvchan->mrc.slots = vzalloc_node(size, node);
 326         if (!nvchan->mrc.slots)
 327                 nvchan->mrc.slots = vzalloc(size);
 328
 329         return nvchan->mrc.slots ? 0 : -ENOMEM;
 330 }
 331
 332 static int netvsc_init_buf(struct hv_device *device,
 333                            struct netvsc_device *net_device,
 334                            const struct netvsc_device_info *device_info)
 335 {
 336         struct nvsp_1_message_send_receive_buffer_complete *resp;
 337         struct net_device *ndev = hv_get_drvdata(device);
 338         struct nvsp_message *init_packet;
 339         unsigned int buf_size;
 340         int i, ret = 0;
 341
 342         /* Get receive buffer area. */
 343         buf_size = device_info->recv_sections * device_info->recv_section_size;
 344         buf_size = roundup(buf_size, PAGE_SIZE);
 345
 346         /* Legacy hosts only allow smaller receive buffer */
 347         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 348                 buf_size = min_t(unsigned int, buf_size,
 349                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 350
 351         net_device->recv_buf = vzalloc(buf_size);
 352         if (!net_device->recv_buf) {
 353                 netdev_err(ndev,
 354                            "unable to allocate receive buffer of size %u\n",
 355                            buf_size);
 356                 ret = -ENOMEM;
 357                 goto cleanup;
 358         }
 359
 360         net_device->recv_buf_size = buf_size;
 361
 362         /*
 363          * Establish the gpadl handle for this buffer on this
 364          * channel.  Note: This call uses the vmbus connection rather
 365          * than the channel to establish the gpadl handle.
 366          */
 367         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 368                                     buf_size,
 369                                     &net_device->recv_buf_gpadl_handle);
 370         if (ret != 0) {
 371                 netdev_err(ndev,
 372                         "unable to establish receive buffer's gpadl\n");
 373                 goto cleanup;
 374         }
 375
 376         /* Notify the NetVsp of the gpadl handle */
 377         init_packet = &net_device->channel_init_pkt;
 378         memset(init_packet, 0, sizeof(struct nvsp_message));
 379         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 380         init_packet->msg.v1_msg.send_recv_buf.
 381                 gpadl_handle = net_device->recv_buf_gpadl_handle.gpadl_handle;
 382         init_packet->msg.v1_msg.
 383                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 384
 385         trace_nvsp_send(ndev, init_packet);
 386
 387         /* Send the gpadl notification request */
 388         ret = vmbus_sendpacket(device->channel, init_packet,
 389                                sizeof(struct nvsp_message),
 390                                (unsigned long)init_packet,
 391                                VM_PKT_DATA_INBAND,
 392                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 393         if (ret != 0) {
 394                 netdev_err(ndev,
 395                         "unable to send receive buffer's gpadl to netvsp\n");
 396                 goto cleanup;
 397         }
 398
 399         wait_for_completion(&net_device->channel_init_wait);
 400
 401         /* Check the response */
 402         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 403         if (resp->status != NVSP_STAT_SUCCESS) {
 404                 netdev_err(ndev,
 405                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 406                            resp->status);
 407                 ret = -EINVAL;
 408                 goto cleanup;
 409         }
 410
 411         /* Parse the response */
 412         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 413                    resp->num_sections, resp->sections[0].sub_alloc_size,
 414                    resp->sections[0].num_sub_allocs);
 415
 416         /* There should only be one section for the entire receive buffer */
 417         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 418                 ret = -EINVAL;
 419                 goto cleanup;
 420         }
 421
 422         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 423         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 424
 425         /* Ensure buffer will not overflow */
 426         if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
 427             (u64)net_device->recv_section_cnt > (u64)buf_size) {
 428                 netdev_err(ndev, "invalid recv_section_size %u\n",
 429                            net_device->recv_section_size);
 430                 ret = -EINVAL;
 431                 goto cleanup;
 432         }
 433
 434         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 435                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
 436
 437                 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
 438                 if (nvchan->recv_buf == NULL) {
 439                         ret = -ENOMEM;
 440                         goto cleanup;
 441                 }
 442         }
 443
 444         /* Setup receive completion ring.
 445          * Add 1 to the recv_section_cnt because at least one entry in a
 446          * ring buffer has to be empty.
 447          */
 448         net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
 449         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 450         if (ret)
 451                 goto cleanup;
 452
 453         /* Now setup the send buffer. */
 454         buf_size = device_info->send_sections * device_info->send_section_size;
 455         buf_size = round_up(buf_size, PAGE_SIZE);
 456
 457         net_device->send_buf = vzalloc(buf_size);
 458         if (!net_device->send_buf) {
 459                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 460                            buf_size);
 461                 ret = -ENOMEM;
 462                 goto cleanup;
 463         }
 464         net_device->send_buf_size = buf_size;
 465
 466         /* Establish the gpadl handle for this buffer on this
 467          * channel.  Note: This call uses the vmbus connection rather
 468          * than the channel to establish the gpadl handle.
 469          */
 470         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 471                                     buf_size,
 472                                     &net_device->send_buf_gpadl_handle);
 473         if (ret != 0) {
 474                 netdev_err(ndev,
 475                            "unable to establish send buffer's gpadl\n");
 476                 goto cleanup;
 477         }
 478
 479         /* Notify the NetVsp of the gpadl handle */
 480         init_packet = &net_device->channel_init_pkt;
 481         memset(init_packet, 0, sizeof(struct nvsp_message));
 482         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 483         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 484                 net_device->send_buf_gpadl_handle.gpadl_handle;
 485         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 486
 487         trace_nvsp_send(ndev, init_packet);
 488
 489         /* Send the gpadl notification request */
 490         ret = vmbus_sendpacket(device->channel, init_packet,
 491                                sizeof(struct nvsp_message),
 492                                (unsigned long)init_packet,
 493                                VM_PKT_DATA_INBAND,
 494                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 495         if (ret != 0) {
 496                 netdev_err(ndev,
 497                            "unable to send send buffer's gpadl to netvsp\n");
 498                 goto cleanup;
 499         }
 500
 501         wait_for_completion(&net_device->channel_init_wait);
 502
 503         /* Check the response */
 504         if (init_packet->msg.v1_msg.
 505             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 506                 netdev_err(ndev, "Unable to complete send buffer "
 507                            "initialization with NetVsp - status %d\n",
 508                            init_packet->msg.v1_msg.
 509                            send_send_buf_complete.status);
 510                 ret = -EINVAL;
 511                 goto cleanup;
 512         }
 513
 514         /* Parse the response */
 515         net_device->send_section_size = init_packet->msg.
 516                                 v1_msg.send_send_buf_complete.section_size;
 517         if (net_device->send_section_size < NETVSC_MTU_MIN) {
 518                 netdev_err(ndev, "invalid send_section_size %u\n",
 519                            net_device->send_section_size);
 520                 ret = -EINVAL;
 521                 goto cleanup;
 522         }
 523
 524         /* Section count is simply the size divided by the section size. */
 525         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 526
 527         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 528                    net_device->send_section_size, net_device->send_section_cnt);
 529
 530         /* Setup state for managing the send buffer. */
 531         net_device->send_section_map = bitmap_zalloc(net_device->send_section_cnt,
 532                                                      GFP_KERNEL);
 533         if (!net_device->send_section_map) {
 534                 ret = -ENOMEM;
 535                 goto cleanup;
 536         }
 537
 538         goto exit;
 539
 540 cleanup:
 541         netvsc_revoke_recv_buf(device, net_device, ndev);
 542         netvsc_revoke_send_buf(device, net_device, ndev);
 543         netvsc_teardown_recv_gpadl(device, net_device, ndev);
 544         netvsc_teardown_send_gpadl(device, net_device, ndev);
 545
 546 exit:
 547         return ret;
 548 }
 549
 550 /* Negotiate NVSP protocol version */
 551 static int negotiate_nvsp_ver(struct hv_device *device,
 552                               struct netvsc_device *net_device,
 553                               struct nvsp_message *init_packet,
 554                               u32 nvsp_ver)
 555 {
 556         struct net_device *ndev = hv_get_drvdata(device);
 557         int ret;
 558
 559         memset(init_packet, 0, sizeof(struct nvsp_message));
 560         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 561         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 562         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 563         trace_nvsp_send(ndev, init_packet);
 564
 565         /* Send the init request */
 566         ret = vmbus_sendpacket(device->channel, init_packet,
 567                                sizeof(struct nvsp_message),
 568                                (unsigned long)init_packet,
 569                                VM_PKT_DATA_INBAND,
 570                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 571
 572         if (ret != 0)
 573                 return ret;
 574
 575         wait_for_completion(&net_device->channel_init_wait);
 576
 577         if (init_packet->msg.init_msg.init_complete.status !=
 578             NVSP_STAT_SUCCESS)
 579                 return -EINVAL;
 580
 581         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 582                 return 0;
 583
 584         /* NVSPv2 or later: Send NDIS config */
 585         memset(init_packet, 0, sizeof(struct nvsp_message));
 586         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 587         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 588         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 589
 590         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 591                 if (hv_is_isolation_supported())
 592                         netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n");
 593                 else
 594                         init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 595
 596                 /* Teaming bit is needed to receive link speed updates */
 597                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 598         }
 599
 600         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 601                 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 602
 603         trace_nvsp_send(ndev, init_packet);
 604
 605         ret = vmbus_sendpacket(device->channel, init_packet,
 606                                 sizeof(struct nvsp_message),
 607                                 VMBUS_RQST_ID_NO_RESPONSE,
 608                                 VM_PKT_DATA_INBAND, 0);
 609
 610         return ret;
 611 }
 612
 613 static int netvsc_connect_vsp(struct hv_device *device,
 614                               struct netvsc_device *net_device,
 615                               const struct netvsc_device_info *device_info)
 616 {
 617         struct net_device *ndev = hv_get_drvdata(device);
 618         static const u32 ver_list[] = {
 619                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 620                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 621                 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 622         };
 623         struct nvsp_message *init_packet;
 624         int ndis_version, i, ret;
 625
 626         init_packet = &net_device->channel_init_pkt;
 627
 628         /* Negotiate the latest NVSP protocol supported */
 629         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 630                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 631                                        ver_list[i])  == 0) {
 632                         net_device->nvsp_version = ver_list[i];
 633                         break;
 634                 }
 635
 636         if (i < 0) {
 637                 ret = -EPROTO;
 638                 goto cleanup;
 639         }
 640
 641         if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) {
 642                 netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n",
 643                            net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61);
 644                 ret = -EPROTO;
 645                 goto cleanup;
 646         }
 647
 648         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 649
 650         /* Send the ndis version */
 651         memset(init_packet, 0, sizeof(struct nvsp_message));
 652
 653         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 654                 ndis_version = 0x00060001;
 655         else
 656                 ndis_version = 0x0006001e;
 657
 658         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 659         init_packet->msg.v1_msg.
 660                 send_ndis_ver.ndis_major_ver =
 661                                 (ndis_version & 0xFFFF0000) >> 16;
 662         init_packet->msg.v1_msg.
 663                 send_ndis_ver.ndis_minor_ver =
 664                                 ndis_version & 0xFFFF;
 665
 666         trace_nvsp_send(ndev, init_packet);
 667
 668         /* Send the init request */
 669         ret = vmbus_sendpacket(device->channel, init_packet,
 670                                 sizeof(struct nvsp_message),
 671                                 VMBUS_RQST_ID_NO_RESPONSE,
 672                                 VM_PKT_DATA_INBAND, 0);
 673         if (ret != 0)
 674                 goto cleanup;
 675
 676
 677         ret = netvsc_init_buf(device, net_device, device_info);
 678
 679 cleanup:
 680         return ret;
 681 }
 682
 683 /*
 684  * netvsc_device_remove - Callback when the root bus device is removed
 685  */
 686 void netvsc_device_remove(struct hv_device *device)
 687 {
 688         struct net_device *ndev = hv_get_drvdata(device);
 689         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 690         struct netvsc_device *net_device
 691                 = rtnl_dereference(net_device_ctx->nvdev);
 692         int i;
 693
 694         /*
 695          * Revoke receive buffer. If host is pre-Win2016 then tear down
 696          * receive buffer GPADL. Do the same for send buffer.
 697          */
 698         netvsc_revoke_recv_buf(device, net_device, ndev);
 699         if (vmbus_proto_version < VERSION_WIN10)
 700                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 701
 702         netvsc_revoke_send_buf(device, net_device, ndev);
 703         if (vmbus_proto_version < VERSION_WIN10)
 704                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 705
 706         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 707
 708         /* Disable NAPI and disassociate its context from the device. */
 709         for (i = 0; i < net_device->num_chn; i++) {
 710                 /* See also vmbus_reset_channel_cb(). */
 711                 napi_disable(&net_device->chan_table[i].napi);
 712                 netif_napi_del(&net_device->chan_table[i].napi);
 713         }
 714
 715         /*
 716          * At this point, no one should be accessing net_device
 717          * except in here
 718          */
 719         netdev_dbg(ndev, "net device safe to remove\n");
 720
 721         /* Now, we can close the channel safely */
 722         vmbus_close(device->channel);
 723
 724         /*
 725          * If host is Win2016 or higher then we do the GPADL tear down
 726          * here after VMBus is closed.
 727         */
 728         if (vmbus_proto_version >= VERSION_WIN10) {
 729                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 730                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 731         }
 732
 733         /* Release all resources */
 734         free_netvsc_device_rcu(net_device);
 735 }
 736
 737 #define RING_AVAIL_PERCENT_HIWATER 20
 738 #define RING_AVAIL_PERCENT_LOWATER 10
 739
 740 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 741                                          u32 index)
 742 {
 743         sync_change_bit(index, net_device->send_section_map);
 744 }
 745
 746 static void netvsc_send_tx_complete(struct net_device *ndev,
 747                                     struct netvsc_device *net_device,
 748                                     struct vmbus_channel *channel,
 749                                     const struct vmpacket_descriptor *desc,
 750                                     int budget)
 751 {
 752         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 753         struct sk_buff *skb;
 754         u16 q_idx = 0;
 755         int queue_sends;
 756         u64 cmd_rqst;
 757
 758         cmd_rqst = channel->request_addr_callback(channel, desc->trans_id);
 759         if (cmd_rqst == VMBUS_RQST_ERROR) {
 760                 netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 761                 return;
 762         }
 763
 764         skb = (struct sk_buff *)(unsigned long)cmd_rqst;
 765
 766         /* Notify the layer above us */
 767         if (likely(skb)) {
 768                 struct hv_netvsc_packet *packet
 769                         = (struct hv_netvsc_packet *)skb->cb;
 770                 u32 send_index = packet->send_buf_index;
 771                 struct netvsc_stats_tx *tx_stats;
 772
 773                 if (send_index != NETVSC_INVALID_INDEX)
 774                         netvsc_free_send_slot(net_device, send_index);
 775                 q_idx = packet->q_idx;
 776
 777                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 778
 779                 u64_stats_update_begin(&tx_stats->syncp);
 780                 tx_stats->packets += packet->total_packets;
 781                 tx_stats->bytes += packet->total_bytes;
 782                 u64_stats_update_end(&tx_stats->syncp);
 783
 784                 netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
 785                 napi_consume_skb(skb, budget);
 786         }
 787
 788         queue_sends =
 789                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 790
 791         if (unlikely(net_device->destroy)) {
 792                 if (queue_sends == 0)
 793                         wake_up(&net_device->wait_drain);
 794         } else {
 795                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 796
 797                 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 798                     (hv_get_avail_to_write_percent(&channel->outbound) >
 799                      RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 800                         netif_tx_wake_queue(txq);
 801                         ndev_ctx->eth_stats.wake_queue++;
 802                 }
 803         }
 804 }
 805
 806 static void netvsc_send_completion(struct net_device *ndev,
 807                                    struct netvsc_device *net_device,
 808                                    struct vmbus_channel *incoming_channel,
 809                                    const struct vmpacket_descriptor *desc,
 810                                    int budget)
 811 {
 812         const struct nvsp_message *nvsp_packet;
 813         u32 msglen = hv_pkt_datalen(desc);
 814         struct nvsp_message *pkt_rqst;
 815         u64 cmd_rqst;
 816         u32 status;
 817
 818         /* First check if this is a VMBUS completion without data payload */
 819         if (!msglen) {
 820                 cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
 821                                                                    desc->trans_id);
 822                 if (cmd_rqst == VMBUS_RQST_ERROR) {
 823                         netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 824                         return;
 825                 }
 826
 827                 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
 828                 switch (pkt_rqst->hdr.msg_type) {
 829                 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
 830                         complete(&net_device->channel_init_wait);
 831                         break;
 832
 833                 default:
 834                         netdev_err(ndev, "Unexpected VMBUS completion!!\n");
 835                 }
 836                 return;
 837         }
 838
 839         /* Ensure packet is big enough to read header fields */
 840         if (msglen < sizeof(struct nvsp_message_header)) {
 841                 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
 842                 return;
 843         }
 844
 845         nvsp_packet = hv_pkt_data(desc);
 846         switch (nvsp_packet->hdr.msg_type) {
 847         case NVSP_MSG_TYPE_INIT_COMPLETE:
 848                 if (msglen < sizeof(struct nvsp_message_header) +
 849                                 sizeof(struct nvsp_message_init_complete)) {
 850                         netdev_err(ndev, "nvsp_msg length too small: %u\n",
 851                                    msglen);
 852                         return;
 853                 }
 854                 fallthrough;
 855
 856         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 857                 if (msglen < sizeof(struct nvsp_message_header) +
 858                                 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
 859                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 860                                    msglen);
 861                         return;
 862                 }
 863                 fallthrough;
 864
 865         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 866                 if (msglen < sizeof(struct nvsp_message_header) +
 867                                 sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
 868                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 869                                    msglen);
 870                         return;
 871                 }
 872                 fallthrough;
 873
 874         case NVSP_MSG5_TYPE_SUBCHANNEL:
 875                 if (msglen < sizeof(struct nvsp_message_header) +
 876                                 sizeof(struct nvsp_5_subchannel_complete)) {
 877                         netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
 878                                    msglen);
 879                         return;
 880                 }
 881                 /* Copy the response back */
 882                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 883                        sizeof(struct nvsp_message));
 884                 complete(&net_device->channel_init_wait);
 885                 break;
 886
 887         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 888                 if (msglen < sizeof(struct nvsp_message_header) +
 889                     sizeof(struct nvsp_1_message_send_rndis_packet_complete)) {
 890                         if (net_ratelimit())
 891                                 netdev_err(ndev, "nvsp_rndis_pkt_complete length too small: %u\n",
 892                                            msglen);
 893                         return;
 894                 }
 895
 896                 /* If status indicates an error, output a message so we know
 897                  * there's a problem. But process the completion anyway so the
 898                  * resources are released.
 899                  */
 900                 status = nvsp_packet->msg.v1_msg.send_rndis_pkt_complete.status;
 901                 if (status != NVSP_STAT_SUCCESS && net_ratelimit())
 902                         netdev_err(ndev, "nvsp_rndis_pkt_complete error status: %x\n",
 903                                    status);
 904
 905                 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 906                                         desc, budget);
 907                 break;
 908
 909         default:
 910                 netdev_err(ndev,
 911                            "Unknown send completion type %d received!!\n",
 912                            nvsp_packet->hdr.msg_type);
 913         }
 914 }
 915
 916 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 917 {
 918         unsigned long *map_addr = net_device->send_section_map;
 919         unsigned int i;
 920
 921         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 922                 if (sync_test_and_set_bit(i, map_addr) == 0)
 923                         return i;
 924         }
 925
 926         return NETVSC_INVALID_INDEX;
 927 }
 928
 929 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 930                                     unsigned int section_index,
 931                                     u32 pend_size,
 932                                     struct hv_netvsc_packet *packet,
 933                                     struct rndis_message *rndis_msg,
 934                                     struct hv_page_buffer *pb,
 935                                     bool xmit_more)
 936 {
 937         char *start = net_device->send_buf;
 938         char *dest = start + (section_index * net_device->send_section_size)
 939                      + pend_size;
 940         int i;
 941         u32 padding = 0;
 942         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 943                 packet->page_buf_cnt;
 944         u32 remain;
 945
 946         /* Add padding */
 947         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 948         if (xmit_more && remain) {
 949                 padding = net_device->pkt_align - remain;
 950                 rndis_msg->msg_len += padding;
 951                 packet->total_data_buflen += padding;
 952         }
 953
 954         for (i = 0; i < page_count; i++) {
 955                 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
 956                 u32 offset = pb[i].offset;
 957                 u32 len = pb[i].len;
 958
 959                 memcpy(dest, (src + offset), len);
 960                 dest += len;
 961         }
 962
 963         if (padding)
 964                 memset(dest, 0, padding);
 965 }
 966
 967 void netvsc_dma_unmap(struct hv_device *hv_dev,
 968                       struct hv_netvsc_packet *packet)
 969 {
 970         int i;
 971
 972         if (!hv_is_isolation_supported())
 973                 return;
 974
 975         if (!packet->dma_range)
 976                 return;
 977
 978         for (i = 0; i < packet->page_buf_cnt; i++)
 979                 dma_unmap_single(&hv_dev->device, packet->dma_range[i].dma,
 980                                  packet->dma_range[i].mapping_size,
 981                                  DMA_TO_DEVICE);
 982
 983         kfree(packet->dma_range);
 984 }
 985
 986 /* netvsc_dma_map - Map swiotlb bounce buffer with data page of
 987  * packet sent by vmbus_sendpacket_pagebuffer() in the Isolation
 988  * VM.
 989  *
 990  * In isolation VM, netvsc send buffer has been marked visible to
 991  * host and so the data copied to send buffer doesn't need to use
 992  * bounce buffer. The data pages handled by vmbus_sendpacket_pagebuffer()
 993  * may not be copied to send buffer and so these pages need to be
 994  * mapped with swiotlb bounce buffer. netvsc_dma_map() is to do
 995  * that. The pfns in the struct hv_page_buffer need to be converted
 996  * to bounce buffer's pfn. The loop here is necessary because the
 997  * entries in the page buffer array are not necessarily full
 998  * pages of data.  Each entry in the array has a separate offset and
 999  * len that may be non-zero, even for entries in the middle of the
1000  * array.  And the entries are not physically contiguous.  So each
1001  * entry must be individually mapped rather than as a contiguous unit.
1002  * So not use dma_map_sg() here.
1003  */
1004 static int netvsc_dma_map(struct hv_device *hv_dev,
1005                           struct hv_netvsc_packet *packet,
1006                           struct hv_page_buffer *pb)
1007 {
1008         u32 page_count = packet->page_buf_cnt;
1009         dma_addr_t dma;
1010         int i;
1011
1012         if (!hv_is_isolation_supported())
1013                 return 0;
1014
1015         packet->dma_range = kcalloc(page_count,
1016                                     sizeof(*packet->dma_range),
1017                                     GFP_ATOMIC);
1018         if (!packet->dma_range)
1019                 return -ENOMEM;
1020
1021         for (i = 0; i < page_count; i++) {
1022                 char *src = phys_to_virt((pb[i].pfn << HV_HYP_PAGE_SHIFT)
1023                                          + pb[i].offset);
1024                 u32 len = pb[i].len;
1025
1026                 dma = dma_map_single(&hv_dev->device, src, len,
1027                                      DMA_TO_DEVICE);
1028                 if (dma_mapping_error(&hv_dev->device, dma)) {
1029                         kfree(packet->dma_range);
1030                         return -ENOMEM;
1031                 }
1032
1033                 /* pb[].offset and pb[].len are not changed during dma mapping
1034                  * and so not reassign.
1035                  */
1036                 packet->dma_range[i].dma = dma;
1037                 packet->dma_range[i].mapping_size = len;
1038                 pb[i].pfn = dma >> HV_HYP_PAGE_SHIFT;
1039         }
1040
1041         return 0;
1042 }
1043
1044 static inline int netvsc_send_pkt(
1045         struct hv_device *device,
1046         struct hv_netvsc_packet *packet,
1047         struct netvsc_device *net_device,
1048         struct hv_page_buffer *pb,
1049         struct sk_buff *skb)
1050 {
1051         struct nvsp_message nvmsg;
1052         struct nvsp_1_message_send_rndis_packet *rpkt =
1053                 &nvmsg.msg.v1_msg.send_rndis_pkt;
1054         struct netvsc_channel * const nvchan =
1055                 &net_device->chan_table[packet->q_idx];
1056         struct vmbus_channel *out_channel = nvchan->channel;
1057         struct net_device *ndev = hv_get_drvdata(device);
1058         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1059         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
1060         u64 req_id;
1061         int ret;
1062         u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
1063
1064         memset(&nvmsg, 0, sizeof(struct nvsp_message));
1065         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
1066         if (skb)
1067                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
1068         else
1069                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
1070
1071         rpkt->send_buf_section_index = packet->send_buf_index;
1072         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
1073                 rpkt->send_buf_section_size = 0;
1074         else
1075                 rpkt->send_buf_section_size = packet->total_data_buflen;
1076
1077         req_id = (ulong)skb;
1078
1079         if (out_channel->rescind)
1080                 return -ENODEV;
1081
1082         trace_nvsp_send_pkt(ndev, out_channel, rpkt);
1083
1084         packet->dma_range = NULL;
1085         if (packet->page_buf_cnt) {
1086                 if (packet->cp_partial)
1087                         pb += packet->rmsg_pgcnt;
1088
1089                 ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb);
1090                 if (ret) {
1091                         ret = -EAGAIN;
1092                         goto exit;
1093                 }
1094
1095                 ret = vmbus_sendpacket_pagebuffer(out_channel,
1096                                                   pb, packet->page_buf_cnt,
1097                                                   &nvmsg, sizeof(nvmsg),
1098                                                   req_id);
1099
1100                 if (ret)
1101                         netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
1102         } else {
1103                 ret = vmbus_sendpacket(out_channel,
1104                                        &nvmsg, sizeof(nvmsg),
1105                                        req_id, VM_PKT_DATA_INBAND,
1106                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1107         }
1108
1109 exit:
1110         if (ret == 0) {
1111                 atomic_inc_return(&nvchan->queue_sends);
1112
1113                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
1114                         netif_tx_stop_queue(txq);
1115                         ndev_ctx->eth_stats.stop_queue++;
1116                 }
1117         } else if (ret == -EAGAIN) {
1118                 netif_tx_stop_queue(txq);
1119                 ndev_ctx->eth_stats.stop_queue++;
1120         } else {
1121                 netdev_err(ndev,
1122                            "Unable to send packet pages %u len %u, ret %d\n",
1123                            packet->page_buf_cnt, packet->total_data_buflen,
1124                            ret);
1125         }
1126
1127         if (netif_tx_queue_stopped(txq) &&
1128             atomic_read(&nvchan->queue_sends) < 1 &&
1129             !net_device->tx_disable) {
1130                 netif_tx_wake_queue(txq);
1131                 ndev_ctx->eth_stats.wake_queue++;
1132                 if (ret == -EAGAIN)
1133                         ret = -ENOSPC;
1134         }
1135
1136         return ret;
1137 }
1138
1139 /* Move packet out of multi send data (msd), and clear msd */
1140 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
1141                                 struct sk_buff **msd_skb,
1142                                 struct multi_send_data *msdp)
1143 {
1144         *msd_skb = msdp->skb;
1145         *msd_send = msdp->pkt;
1146         msdp->skb = NULL;
1147         msdp->pkt = NULL;
1148         msdp->count = 0;
1149 }
1150
1151 /* RCU already held by caller */
1152 /* Batching/bouncing logic is designed to attempt to optimize
1153  * performance.
1154  *
1155  * For small, non-LSO packets we copy the packet to a send buffer
1156  * which is pre-registered with the Hyper-V side. This enables the
1157  * hypervisor to avoid remapping the aperture to access the packet
1158  * descriptor and data.
1159  *
1160  * If we already started using a buffer and the netdev is transmitting
1161  * a burst of packets, keep on copying into the buffer until it is
1162  * full or we are done collecting a burst. If there is an existing
1163  * buffer with space for the RNDIS descriptor but not the packet, copy
1164  * the RNDIS descriptor to the buffer, keeping the packet in place.
1165  *
1166  * If we do batching and send more than one packet using a single
1167  * NetVSC message, free the SKBs of the packets copied, except for the
1168  * last packet. This is done to streamline the handling of the case
1169  * where the last packet only had the RNDIS descriptor copied to the
1170  * send buffer, with the data pointers included in the NetVSC message.
1171  */
1172 int netvsc_send(struct net_device *ndev,
1173                 struct hv_netvsc_packet *packet,
1174                 struct rndis_message *rndis_msg,
1175                 struct hv_page_buffer *pb,
1176                 struct sk_buff *skb,
1177                 bool xdp_tx)
1178 {
1179         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1180         struct netvsc_device *net_device
1181                 = rcu_dereference_bh(ndev_ctx->nvdev);
1182         struct hv_device *device = ndev_ctx->device_ctx;
1183         int ret = 0;
1184         struct netvsc_channel *nvchan;
1185         u32 pktlen = packet->total_data_buflen, msd_len = 0;
1186         unsigned int section_index = NETVSC_INVALID_INDEX;
1187         struct multi_send_data *msdp;
1188         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1189         struct sk_buff *msd_skb = NULL;
1190         bool try_batch, xmit_more;
1191
1192         /* If device is rescinded, return error and packet will get dropped. */
1193         if (unlikely(!net_device || net_device->destroy))
1194                 return -ENODEV;
1195
1196         nvchan = &net_device->chan_table[packet->q_idx];
1197         packet->send_buf_index = NETVSC_INVALID_INDEX;
1198         packet->cp_partial = false;
1199
1200         /* Send a control message or XDP packet directly without accessing
1201          * msd (Multi-Send Data) field which may be changed during data packet
1202          * processing.
1203          */
1204         if (!skb || xdp_tx)
1205                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
1206
1207         /* batch packets in send buffer if possible */
1208         msdp = &nvchan->msd;
1209         if (msdp->pkt)
1210                 msd_len = msdp->pkt->total_data_buflen;
1211
1212         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
1213         if (try_batch && msd_len + pktlen + net_device->pkt_align <
1214             net_device->send_section_size) {
1215                 section_index = msdp->pkt->send_buf_index;
1216
1217         } else if (try_batch && msd_len + packet->rmsg_size <
1218                    net_device->send_section_size) {
1219                 section_index = msdp->pkt->send_buf_index;
1220                 packet->cp_partial = true;
1221
1222         } else if (pktlen + net_device->pkt_align <
1223                    net_device->send_section_size) {
1224                 section_index = netvsc_get_next_send_section(net_device);
1225                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1226                         ++ndev_ctx->eth_stats.tx_send_full;
1227                 } else {
1228                         move_pkt_msd(&msd_send, &msd_skb, msdp);
1229                         msd_len = 0;
1230                 }
1231         }
1232
1233         /* Keep aggregating only if stack says more data is coming
1234          * and not doing mixed modes send and not flow blocked
1235          */
1236         xmit_more = netdev_xmit_more() &&
1237                 !packet->cp_partial &&
1238                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1239
1240         if (section_index != NETVSC_INVALID_INDEX) {
1241                 netvsc_copy_to_send_buf(net_device,
1242                                         section_index, msd_len,
1243                                         packet, rndis_msg, pb, xmit_more);
1244
1245                 packet->send_buf_index = section_index;
1246
1247                 if (packet->cp_partial) {
1248                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
1249                         packet->total_data_buflen = msd_len + packet->rmsg_size;
1250                 } else {
1251                         packet->page_buf_cnt = 0;
1252                         packet->total_data_buflen += msd_len;
1253                 }
1254
1255                 if (msdp->pkt) {
1256                         packet->total_packets += msdp->pkt->total_packets;
1257                         packet->total_bytes += msdp->pkt->total_bytes;
1258                 }
1259
1260                 if (msdp->skb)
1261                         dev_consume_skb_any(msdp->skb);
1262
1263                 if (xmit_more) {
1264                         msdp->skb = skb;
1265                         msdp->pkt = packet;
1266                         msdp->count++;
1267                 } else {
1268                         cur_send = packet;
1269                         msdp->skb = NULL;
1270                         msdp->pkt = NULL;
1271                         msdp->count = 0;
1272                 }
1273         } else {
1274                 move_pkt_msd(&msd_send, &msd_skb, msdp);
1275                 cur_send = packet;
1276         }
1277
1278         if (msd_send) {
1279                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1280                                             NULL, msd_skb);
1281
1282                 if (m_ret != 0) {
1283                         netvsc_free_send_slot(net_device,
1284                                               msd_send->send_buf_index);
1285                         dev_kfree_skb_any(msd_skb);
1286                 }
1287         }
1288
1289         if (cur_send)
1290                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1291
1292         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1293                 netvsc_free_send_slot(net_device, section_index);
1294
1295         return ret;
1296 }
1297
1298 /* Send pending recv completions */
1299 static int send_recv_completions(struct net_device *ndev,
1300                                  struct netvsc_device *nvdev,
1301                                  struct netvsc_channel *nvchan)
1302 {
1303         struct multi_recv_comp *mrc = &nvchan->mrc;
1304         struct recv_comp_msg {
1305                 struct nvsp_message_header hdr;
1306                 u32 status;
1307         }  __packed;
1308         struct recv_comp_msg msg = {
1309                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1310         };
1311         int ret;
1312
1313         while (mrc->first != mrc->next) {
1314                 const struct recv_comp_data *rcd
1315                         = mrc->slots + mrc->first;
1316
1317                 msg.status = rcd->status;
1318                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1319                                        rcd->tid, VM_PKT_COMP, 0);
1320                 if (unlikely(ret)) {
1321                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1322
1323                         ++ndev_ctx->eth_stats.rx_comp_busy;
1324                         return ret;
1325                 }
1326
1327                 if (++mrc->first == nvdev->recv_completion_cnt)
1328                         mrc->first = 0;
1329         }
1330
1331         /* receive completion ring has been emptied */
1332         if (unlikely(nvdev->destroy))
1333                 wake_up(&nvdev->wait_drain);
1334
1335         return 0;
1336 }
1337
1338 /* Count how many receive completions are outstanding */
1339 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1340                                  const struct multi_recv_comp *mrc,
1341                                  u32 *filled, u32 *avail)
1342 {
1343         u32 count = nvdev->recv_completion_cnt;
1344
1345         if (mrc->next >= mrc->first)
1346                 *filled = mrc->next - mrc->first;
1347         else
1348                 *filled = (count - mrc->first) + mrc->next;
1349
1350         *avail = count - *filled - 1;
1351 }
1352
1353 /* Add receive complete to ring to send to host. */
1354 static void enq_receive_complete(struct net_device *ndev,
1355                                  struct netvsc_device *nvdev, u16 q_idx,
1356                                  u64 tid, u32 status)
1357 {
1358         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1359         struct multi_recv_comp *mrc = &nvchan->mrc;
1360         struct recv_comp_data *rcd;
1361         u32 filled, avail;
1362
1363         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1364
1365         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1366                 send_recv_completions(ndev, nvdev, nvchan);
1367                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1368         }
1369
1370         if (unlikely(!avail)) {
1371                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1372                            q_idx, tid);
1373                 return;
1374         }
1375
1376         rcd = mrc->slots + mrc->next;
1377         rcd->tid = tid;
1378         rcd->status = status;
1379
1380         if (++mrc->next == nvdev->recv_completion_cnt)
1381                 mrc->next = 0;
1382 }
1383
1384 static int netvsc_receive(struct net_device *ndev,
1385                           struct netvsc_device *net_device,
1386                           struct netvsc_channel *nvchan,
1387                           const struct vmpacket_descriptor *desc)
1388 {
1389         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1390         struct vmbus_channel *channel = nvchan->channel;
1391         const struct vmtransfer_page_packet_header *vmxferpage_packet
1392                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1393         const struct nvsp_message *nvsp = hv_pkt_data(desc);
1394         u32 msglen = hv_pkt_datalen(desc);
1395         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1396         char *recv_buf = net_device->recv_buf;
1397         u32 status = NVSP_STAT_SUCCESS;
1398         int i;
1399         int count = 0;
1400
1401         /* Ensure packet is big enough to read header fields */
1402         if (msglen < sizeof(struct nvsp_message_header)) {
1403                 netif_err(net_device_ctx, rx_err, ndev,
1404                           "invalid nvsp header, length too small: %u\n",
1405                           msglen);
1406                 return 0;
1407         }
1408
1409         /* Make sure this is a valid nvsp packet */
1410         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1411                 netif_err(net_device_ctx, rx_err, ndev,
1412                           "Unknown nvsp packet type received %u\n",
1413                           nvsp->hdr.msg_type);
1414                 return 0;
1415         }
1416
1417         /* Validate xfer page pkt header */
1418         if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1419                 netif_err(net_device_ctx, rx_err, ndev,
1420                           "Invalid xfer page pkt, offset too small: %u\n",
1421                           desc->offset8 << 3);
1422                 return 0;
1423         }
1424
1425         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1426                 netif_err(net_device_ctx, rx_err, ndev,
1427                           "Invalid xfer page set id - expecting %x got %x\n",
1428                           NETVSC_RECEIVE_BUFFER_ID,
1429                           vmxferpage_packet->xfer_pageset_id);
1430                 return 0;
1431         }
1432
1433         count = vmxferpage_packet->range_cnt;
1434
1435         /* Check count for a valid value */
1436         if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1437                 netif_err(net_device_ctx, rx_err, ndev,
1438                           "Range count is not valid: %d\n",
1439                           count);
1440                 return 0;
1441         }
1442
1443         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1444         for (i = 0; i < count; i++) {
1445                 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1446                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1447                 void *data;
1448                 int ret;
1449
1450                 if (unlikely(offset > net_device->recv_buf_size ||
1451                              buflen > net_device->recv_buf_size - offset)) {
1452                         nvchan->rsc.cnt = 0;
1453                         status = NVSP_STAT_FAIL;
1454                         netif_err(net_device_ctx, rx_err, ndev,
1455                                   "Packet offset:%u + len:%u too big\n",
1456                                   offset, buflen);
1457
1458                         continue;
1459                 }
1460
1461                 /* We're going to copy (sections of) the packet into nvchan->recv_buf;
1462                  * make sure that nvchan->recv_buf is large enough to hold the packet.
1463                  */
1464                 if (unlikely(buflen > net_device->recv_section_size)) {
1465                         nvchan->rsc.cnt = 0;
1466                         status = NVSP_STAT_FAIL;
1467                         netif_err(net_device_ctx, rx_err, ndev,
1468                                   "Packet too big: buflen=%u recv_section_size=%u\n",
1469                                   buflen, net_device->recv_section_size);
1470
1471                         continue;
1472                 }
1473
1474                 data = recv_buf + offset;
1475
1476                 nvchan->rsc.is_last = (i == count - 1);
1477
1478                 trace_rndis_recv(ndev, q_idx, data);
1479
1480                 /* Pass it to the upper layer */
1481                 ret = rndis_filter_receive(ndev, net_device,
1482                                            nvchan, data, buflen);
1483
1484                 if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1485                         /* Drop incomplete packet */
1486                         nvchan->rsc.cnt = 0;
1487                         status = NVSP_STAT_FAIL;
1488                 }
1489         }
1490
1491         enq_receive_complete(ndev, net_device, q_idx,
1492                              vmxferpage_packet->d.trans_id, status);
1493
1494         return count;
1495 }
1496
1497 static void netvsc_send_table(struct net_device *ndev,
1498                               struct netvsc_device *nvscdev,
1499                               const struct nvsp_message *nvmsg,
1500                               u32 msglen)
1501 {
1502         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1503         u32 count, offset, *tab;
1504         int i;
1505
1506         /* Ensure packet is big enough to read send_table fields */
1507         if (msglen < sizeof(struct nvsp_message_header) +
1508                      sizeof(struct nvsp_5_send_indirect_table)) {
1509                 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1510                 return;
1511         }
1512
1513         count = nvmsg->msg.v5_msg.send_table.count;
1514         offset = nvmsg->msg.v5_msg.send_table.offset;
1515
1516         if (count != VRSS_SEND_TAB_SIZE) {
1517                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1518                 return;
1519         }
1520
1521         /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1522          * wrong due to a host bug. So fix the offset here.
1523          */
1524         if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1525             msglen >= sizeof(struct nvsp_message_header) +
1526             sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1527                 offset = sizeof(struct nvsp_message_header) +
1528                          sizeof(union nvsp_6_message_uber);
1529
1530         /* Boundary check for all versions */
1531         if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1532                 netdev_err(ndev, "Received send-table offset too big:%u\n",
1533                            offset);
1534                 return;
1535         }
1536
1537         tab = (void *)nvmsg + offset;
1538
1539         for (i = 0; i < count; i++)
1540                 net_device_ctx->tx_table[i] = tab[i];
1541 }
1542
1543 static void netvsc_send_vf(struct net_device *ndev,
1544                            const struct nvsp_message *nvmsg,
1545                            u32 msglen)
1546 {
1547         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1548
1549         /* Ensure packet is big enough to read its fields */
1550         if (msglen < sizeof(struct nvsp_message_header) +
1551                      sizeof(struct nvsp_4_send_vf_association)) {
1552                 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1553                 return;
1554         }
1555
1556         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1557         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1558
1559         if (net_device_ctx->vf_alloc)
1560                 complete(&net_device_ctx->vf_add);
1561
1562         netdev_info(ndev, "VF slot %u %s\n",
1563                     net_device_ctx->vf_serial,
1564                     net_device_ctx->vf_alloc ? "added" : "removed");
1565 }
1566
1567 static void netvsc_receive_inband(struct net_device *ndev,
1568                                   struct netvsc_device *nvscdev,
1569                                   const struct vmpacket_descriptor *desc)
1570 {
1571         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1572         u32 msglen = hv_pkt_datalen(desc);
1573
1574         /* Ensure packet is big enough to read header fields */
1575         if (msglen < sizeof(struct nvsp_message_header)) {
1576                 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1577                 return;
1578         }
1579
1580         switch (nvmsg->hdr.msg_type) {
1581         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1582                 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1583                 break;
1584
1585         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1586                 if (hv_is_isolation_supported())
1587                         netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n");
1588                 else
1589                         netvsc_send_vf(ndev, nvmsg, msglen);
1590                 break;
1591         }
1592 }
1593
1594 static int netvsc_process_raw_pkt(struct hv_device *device,
1595                                   struct netvsc_channel *nvchan,
1596                                   struct netvsc_device *net_device,
1597                                   struct net_device *ndev,
1598                                   const struct vmpacket_descriptor *desc,
1599                                   int budget)
1600 {
1601         struct vmbus_channel *channel = nvchan->channel;
1602         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1603
1604         trace_nvsp_recv(ndev, channel, nvmsg);
1605
1606         switch (desc->type) {
1607         case VM_PKT_COMP:
1608                 netvsc_send_completion(ndev, net_device, channel, desc, budget);
1609                 break;
1610
1611         case VM_PKT_DATA_USING_XFER_PAGES:
1612                 return netvsc_receive(ndev, net_device, nvchan, desc);
1613
1614         case VM_PKT_DATA_INBAND:
1615                 netvsc_receive_inband(ndev, net_device, desc);
1616                 break;
1617
1618         default:
1619                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1620                            desc->type, desc->trans_id);
1621                 break;
1622         }
1623
1624         return 0;
1625 }
1626
1627 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1628 {
1629         struct vmbus_channel *primary = channel->primary_channel;
1630
1631         return primary ? primary->device_obj : channel->device_obj;
1632 }
1633
1634 /* Network processing softirq
1635  * Process data in incoming ring buffer from host
1636  * Stops when ring is empty or budget is met or exceeded.
1637  */
1638 int netvsc_poll(struct napi_struct *napi, int budget)
1639 {
1640         struct netvsc_channel *nvchan
1641                 = container_of(napi, struct netvsc_channel, napi);
1642         struct netvsc_device *net_device = nvchan->net_device;
1643         struct vmbus_channel *channel = nvchan->channel;
1644         struct hv_device *device = netvsc_channel_to_device(channel);
1645         struct net_device *ndev = hv_get_drvdata(device);
1646         int work_done = 0;
1647         int ret;
1648
1649         /* If starting a new interval */
1650         if (!nvchan->desc)
1651                 nvchan->desc = hv_pkt_iter_first(channel);
1652
1653         nvchan->xdp_flush = false;
1654
1655         while (nvchan->desc && work_done < budget) {
1656                 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1657                                                     ndev, nvchan->desc, budget);
1658                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1659         }
1660
1661         if (nvchan->xdp_flush)
1662                 xdp_do_flush();
1663
1664         /* Send any pending receive completions */
1665         ret = send_recv_completions(ndev, net_device, nvchan);
1666
1667         /* If it did not exhaust NAPI budget this time
1668          *  and not doing busy poll
1669          * then re-enable host interrupts
1670          *  and reschedule if ring is not empty
1671          *   or sending receive completion failed.
1672          */
1673         if (work_done < budget &&
1674             napi_complete_done(napi, work_done) &&
1675             (ret || hv_end_read(&channel->inbound)) &&
1676             napi_schedule_prep(napi)) {
1677                 hv_begin_read(&channel->inbound);
1678                 __napi_schedule(napi);
1679         }
1680
1681         /* Driver may overshoot since multiple packets per descriptor */
1682         return min(work_done, budget);
1683 }
1684
1685 /* Call back when data is available in host ring buffer.
1686  * Processing is deferred until network softirq (NAPI)
1687  */
1688 void netvsc_channel_cb(void *context)
1689 {
1690         struct netvsc_channel *nvchan = context;
1691         struct vmbus_channel *channel = nvchan->channel;
1692         struct hv_ring_buffer_info *rbi = &channel->inbound;
1693
1694         /* preload first vmpacket descriptor */
1695         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1696
1697         if (napi_schedule_prep(&nvchan->napi)) {
1698                 /* disable interrupts from host */
1699                 hv_begin_read(rbi);
1700
1701                 __napi_schedule_irqoff(&nvchan->napi);
1702         }
1703 }
1704
1705 /*
1706  * netvsc_device_add - Callback when the device belonging to this
1707  * driver is added
1708  */
1709 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1710                                 const struct netvsc_device_info *device_info)
1711 {
1712         int i, ret = 0;
1713         struct netvsc_device *net_device;
1714         struct net_device *ndev = hv_get_drvdata(device);
1715         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1716
1717         net_device = alloc_net_device();
1718         if (!net_device)
1719                 return ERR_PTR(-ENOMEM);
1720
1721         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1722                 net_device_ctx->tx_table[i] = 0;
1723
1724         /* Because the device uses NAPI, all the interrupt batching and
1725          * control is done via Net softirq, not the channel handling
1726          */
1727         set_channel_read_mode(device->channel, HV_CALL_ISR);
1728
1729         /* If we're reopening the device we may have multiple queues, fill the
1730          * chn_table with the default channel to use it before subchannels are
1731          * opened.
1732          * Initialize the channel state before we open;
1733          * we can be interrupted as soon as we open the channel.
1734          */
1735
1736         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1737                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1738
1739                 nvchan->channel = device->channel;
1740                 nvchan->net_device = net_device;
1741                 u64_stats_init(&nvchan->tx_stats.syncp);
1742                 u64_stats_init(&nvchan->rx_stats.syncp);
1743
1744                 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1745
1746                 if (ret) {
1747                         netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1748                         goto cleanup2;
1749                 }
1750
1751                 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1752                                                  MEM_TYPE_PAGE_SHARED, NULL);
1753
1754                 if (ret) {
1755                         netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1756                         goto cleanup2;
1757                 }
1758         }
1759
1760         /* Enable NAPI handler before init callbacks */
1761         netif_napi_add(ndev, &net_device->chan_table[0].napi, netvsc_poll);
1762
1763         /* Open the channel */
1764         device->channel->next_request_id_callback = vmbus_next_request_id;
1765         device->channel->request_addr_callback = vmbus_request_addr;
1766         device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1767         device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE;
1768
1769         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1770                          netvsc_ring_bytes,  NULL, 0,
1771                          netvsc_channel_cb, net_device->chan_table);
1772
1773         if (ret != 0) {
1774                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1775                 goto cleanup;
1776         }
1777
1778         /* Channel is opened */
1779         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1780
1781         napi_enable(&net_device->chan_table[0].napi);
1782
1783         /* Connect with the NetVsp */
1784         ret = netvsc_connect_vsp(device, net_device, device_info);
1785         if (ret != 0) {
1786                 netdev_err(ndev,
1787                         "unable to connect to NetVSP - %d\n", ret);
1788                 goto close;
1789         }
1790
1791         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1792          * populated.
1793          */
1794         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1795
1796         return net_device;
1797
1798 close:
1799         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1800         napi_disable(&net_device->chan_table[0].napi);
1801
1802         /* Now, we can close the channel safely */
1803         vmbus_close(device->channel);
1804
1805 cleanup:
1806         netif_napi_del(&net_device->chan_table[0].napi);
1807
1808 cleanup2:
1809         free_netvsc_device(&net_device->rcu);
1810
1811         return ERR_PTR(ret);
1812 }