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
 158         if (nvdev->recv_original_buf)
 159                 vfree(nvdev->recv_original_buf);
 160         else
 161                 vfree(nvdev->recv_buf);
 162
 163         if (nvdev->send_original_buf)
 164                 vfree(nvdev->send_original_buf);
 165         else
 166                 vfree(nvdev->send_buf);
 167
 168         bitmap_free(nvdev->send_section_map);
 169
 170         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 171                 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
 172                 kfree(nvdev->chan_table[i].recv_buf);
 173                 vfree(nvdev->chan_table[i].mrc.slots);
 174         }
 175
 176         kfree(nvdev);
 177 }
 178
 179 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 180 {
 181         call_rcu(&nvdev->rcu, free_netvsc_device);
 182 }
 183
 184 static void netvsc_revoke_recv_buf(struct hv_device *device,
 185                                    struct netvsc_device *net_device,
 186                                    struct net_device *ndev)
 187 {
 188         struct nvsp_message *revoke_packet;
 189         int ret;
 190
 191         /*
 192          * If we got a section count, it means we received a
 193          * SendReceiveBufferComplete msg (ie sent
 194          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 195          * to send a revoke msg here
 196          */
 197         if (net_device->recv_section_cnt) {
 198                 /* Send the revoke receive buffer */
 199                 revoke_packet = &net_device->revoke_packet;
 200                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 201
 202                 revoke_packet->hdr.msg_type =
 203                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 204                 revoke_packet->msg.v1_msg.
 205                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 206
 207                 trace_nvsp_send(ndev, revoke_packet);
 208
 209                 ret = vmbus_sendpacket(device->channel,
 210                                        revoke_packet,
 211                                        sizeof(struct nvsp_message),
 212                                        VMBUS_RQST_ID_NO_RESPONSE,
 213                                        VM_PKT_DATA_INBAND, 0);
 214                 /* If the failure is because the channel is rescinded;
 215                  * ignore the failure since we cannot send on a rescinded
 216                  * channel. This would allow us to properly cleanup
 217                  * even when the channel is rescinded.
 218                  */
 219                 if (device->channel->rescind)
 220                         ret = 0;
 221                 /*
 222                  * If we failed here, we might as well return and
 223                  * have a leak rather than continue and a bugchk
 224                  */
 225                 if (ret != 0) {
 226                         netdev_err(ndev, "unable to send "
 227                                 "revoke receive buffer to netvsp\n");
 228                         return;
 229                 }
 230                 net_device->recv_section_cnt = 0;
 231         }
 232 }
 233
 234 static void netvsc_revoke_send_buf(struct hv_device *device,
 235                                    struct netvsc_device *net_device,
 236                                    struct net_device *ndev)
 237 {
 238         struct nvsp_message *revoke_packet;
 239         int ret;
 240
 241         /* Deal with the send buffer we may have setup.
 242          * If we got a  send section size, it means we received a
 243          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 244          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 245          * to send a revoke msg here
 246          */
 247         if (net_device->send_section_cnt) {
 248                 /* Send the revoke receive buffer */
 249                 revoke_packet = &net_device->revoke_packet;
 250                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 251
 252                 revoke_packet->hdr.msg_type =
 253                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 254                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 255                         NETVSC_SEND_BUFFER_ID;
 256
 257                 trace_nvsp_send(ndev, revoke_packet);
 258
 259                 ret = vmbus_sendpacket(device->channel,
 260                                        revoke_packet,
 261                                        sizeof(struct nvsp_message),
 262                                        VMBUS_RQST_ID_NO_RESPONSE,
 263                                        VM_PKT_DATA_INBAND, 0);
 264
 265                 /* If the failure is because the channel is rescinded;
 266                  * ignore the failure since we cannot send on a rescinded
 267                  * channel. This would allow us to properly cleanup
 268                  * even when the channel is rescinded.
 269                  */
 270                 if (device->channel->rescind)
 271                         ret = 0;
 272
 273                 /* If we failed here, we might as well return and
 274                  * have a leak rather than continue and a bugchk
 275                  */
 276                 if (ret != 0) {
 277                         netdev_err(ndev, "unable to send "
 278                                    "revoke send buffer to netvsp\n");
 279                         return;
 280                 }
 281                 net_device->send_section_cnt = 0;
 282         }
 283 }
 284
 285 static void netvsc_teardown_recv_gpadl(struct hv_device *device,
 286                                        struct netvsc_device *net_device,
 287                                        struct net_device *ndev)
 288 {
 289         int ret;
 290
 291         if (net_device->recv_buf_gpadl_handle.gpadl_handle) {
 292                 ret = vmbus_teardown_gpadl(device->channel,
 293                                            &net_device->recv_buf_gpadl_handle);
 294
 295                 /* If we failed here, we might as well return and have a leak
 296                  * rather than continue and a bugchk
 297                  */
 298                 if (ret != 0) {
 299                         netdev_err(ndev,
 300                                    "unable to teardown receive buffer's gpadl\n");
 301                         return;
 302                 }
 303         }
 304 }
 305
 306 static void netvsc_teardown_send_gpadl(struct hv_device *device,
 307                                        struct netvsc_device *net_device,
 308                                        struct net_device *ndev)
 309 {
 310         int ret;
 311
 312         if (net_device->send_buf_gpadl_handle.gpadl_handle) {
 313                 ret = vmbus_teardown_gpadl(device->channel,
 314                                            &net_device->send_buf_gpadl_handle);
 315
 316                 /* If we failed here, we might as well return and have a leak
 317                  * rather than continue and a bugchk
 318                  */
 319                 if (ret != 0) {
 320                         netdev_err(ndev,
 321                                    "unable to teardown send buffer's gpadl\n");
 322                         return;
 323                 }
 324         }
 325 }
 326
 327 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 328 {
 329         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 330         int node = cpu_to_node(nvchan->channel->target_cpu);
 331         size_t size;
 332
 333         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 334         nvchan->mrc.slots = vzalloc_node(size, node);
 335         if (!nvchan->mrc.slots)
 336                 nvchan->mrc.slots = vzalloc(size);
 337
 338         return nvchan->mrc.slots ? 0 : -ENOMEM;
 339 }
 340
 341 static int netvsc_init_buf(struct hv_device *device,
 342                            struct netvsc_device *net_device,
 343                            const struct netvsc_device_info *device_info)
 344 {
 345         struct nvsp_1_message_send_receive_buffer_complete *resp;
 346         struct net_device *ndev = hv_get_drvdata(device);
 347         struct nvsp_message *init_packet;
 348         unsigned int buf_size;
 349         int i, ret = 0;
 350         void *vaddr;
 351
 352         /* Get receive buffer area. */
 353         buf_size = device_info->recv_sections * device_info->recv_section_size;
 354         buf_size = roundup(buf_size, PAGE_SIZE);
 355
 356         /* Legacy hosts only allow smaller receive buffer */
 357         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 358                 buf_size = min_t(unsigned int, buf_size,
 359                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 360
 361         net_device->recv_buf = vzalloc(buf_size);
 362         if (!net_device->recv_buf) {
 363                 netdev_err(ndev,
 364                            "unable to allocate receive buffer of size %u\n",
 365                            buf_size);
 366                 ret = -ENOMEM;
 367                 goto cleanup;
 368         }
 369
 370         net_device->recv_buf_size = buf_size;
 371
 372         /*
 373          * Establish the gpadl handle for this buffer on this
 374          * channel.  Note: This call uses the vmbus connection rather
 375          * than the channel to establish the gpadl handle.
 376          */
 377         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 378                                     buf_size,
 379                                     &net_device->recv_buf_gpadl_handle);
 380         if (ret != 0) {
 381                 netdev_err(ndev,
 382                         "unable to establish receive buffer's gpadl\n");
 383                 goto cleanup;
 384         }
 385
 386         if (hv_isolation_type_snp()) {
 387                 vaddr = hv_map_memory(net_device->recv_buf, buf_size);
 388                 if (!vaddr) {
 389                         ret = -ENOMEM;
 390                         goto cleanup;
 391                 }
 392
 393                 net_device->recv_original_buf = net_device->recv_buf;
 394                 net_device->recv_buf = vaddr;
 395         }
 396
 397         /* Notify the NetVsp of the gpadl handle */
 398         init_packet = &net_device->channel_init_pkt;
 399         memset(init_packet, 0, sizeof(struct nvsp_message));
 400         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 401         init_packet->msg.v1_msg.send_recv_buf.
 402                 gpadl_handle = net_device->recv_buf_gpadl_handle.gpadl_handle;
 403         init_packet->msg.v1_msg.
 404                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 405
 406         trace_nvsp_send(ndev, init_packet);
 407
 408         /* Send the gpadl notification request */
 409         ret = vmbus_sendpacket(device->channel, init_packet,
 410                                sizeof(struct nvsp_message),
 411                                (unsigned long)init_packet,
 412                                VM_PKT_DATA_INBAND,
 413                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 414         if (ret != 0) {
 415                 netdev_err(ndev,
 416                         "unable to send receive buffer's gpadl to netvsp\n");
 417                 goto cleanup;
 418         }
 419
 420         wait_for_completion(&net_device->channel_init_wait);
 421
 422         /* Check the response */
 423         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 424         if (resp->status != NVSP_STAT_SUCCESS) {
 425                 netdev_err(ndev,
 426                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 427                            resp->status);
 428                 ret = -EINVAL;
 429                 goto cleanup;
 430         }
 431
 432         /* Parse the response */
 433         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 434                    resp->num_sections, resp->sections[0].sub_alloc_size,
 435                    resp->sections[0].num_sub_allocs);
 436
 437         /* There should only be one section for the entire receive buffer */
 438         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 439                 ret = -EINVAL;
 440                 goto cleanup;
 441         }
 442
 443         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 444         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 445
 446         /* Ensure buffer will not overflow */
 447         if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
 448             (u64)net_device->recv_section_cnt > (u64)buf_size) {
 449                 netdev_err(ndev, "invalid recv_section_size %u\n",
 450                            net_device->recv_section_size);
 451                 ret = -EINVAL;
 452                 goto cleanup;
 453         }
 454
 455         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 456                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
 457
 458                 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
 459                 if (nvchan->recv_buf == NULL) {
 460                         ret = -ENOMEM;
 461                         goto cleanup;
 462                 }
 463         }
 464
 465         /* Setup receive completion ring.
 466          * Add 1 to the recv_section_cnt because at least one entry in a
 467          * ring buffer has to be empty.
 468          */
 469         net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
 470         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 471         if (ret)
 472                 goto cleanup;
 473
 474         /* Now setup the send buffer. */
 475         buf_size = device_info->send_sections * device_info->send_section_size;
 476         buf_size = round_up(buf_size, PAGE_SIZE);
 477
 478         net_device->send_buf = vzalloc(buf_size);
 479         if (!net_device->send_buf) {
 480                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 481                            buf_size);
 482                 ret = -ENOMEM;
 483                 goto cleanup;
 484         }
 485         net_device->send_buf_size = buf_size;
 486
 487         /* Establish the gpadl handle for this buffer on this
 488          * channel.  Note: This call uses the vmbus connection rather
 489          * than the channel to establish the gpadl handle.
 490          */
 491         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 492                                     buf_size,
 493                                     &net_device->send_buf_gpadl_handle);
 494         if (ret != 0) {
 495                 netdev_err(ndev,
 496                            "unable to establish send buffer's gpadl\n");
 497                 goto cleanup;
 498         }
 499
 500         if (hv_isolation_type_snp()) {
 501                 vaddr = hv_map_memory(net_device->send_buf, buf_size);
 502                 if (!vaddr) {
 503                         ret = -ENOMEM;
 504                         goto cleanup;
 505                 }
 506
 507                 net_device->send_original_buf = net_device->send_buf;
 508                 net_device->send_buf = vaddr;
 509         }
 510
 511         /* Notify the NetVsp of the gpadl handle */
 512         init_packet = &net_device->channel_init_pkt;
 513         memset(init_packet, 0, sizeof(struct nvsp_message));
 514         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 515         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 516                 net_device->send_buf_gpadl_handle.gpadl_handle;
 517         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 518
 519         trace_nvsp_send(ndev, init_packet);
 520
 521         /* Send the gpadl notification request */
 522         ret = vmbus_sendpacket(device->channel, init_packet,
 523                                sizeof(struct nvsp_message),
 524                                (unsigned long)init_packet,
 525                                VM_PKT_DATA_INBAND,
 526                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 527         if (ret != 0) {
 528                 netdev_err(ndev,
 529                            "unable to send send buffer's gpadl to netvsp\n");
 530                 goto cleanup;
 531         }
 532
 533         wait_for_completion(&net_device->channel_init_wait);
 534
 535         /* Check the response */
 536         if (init_packet->msg.v1_msg.
 537             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 538                 netdev_err(ndev, "Unable to complete send buffer "
 539                            "initialization with NetVsp - status %d\n",
 540                            init_packet->msg.v1_msg.
 541                            send_send_buf_complete.status);
 542                 ret = -EINVAL;
 543                 goto cleanup;
 544         }
 545
 546         /* Parse the response */
 547         net_device->send_section_size = init_packet->msg.
 548                                 v1_msg.send_send_buf_complete.section_size;
 549         if (net_device->send_section_size < NETVSC_MTU_MIN) {
 550                 netdev_err(ndev, "invalid send_section_size %u\n",
 551                            net_device->send_section_size);
 552                 ret = -EINVAL;
 553                 goto cleanup;
 554         }
 555
 556         /* Section count is simply the size divided by the section size. */
 557         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 558
 559         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 560                    net_device->send_section_size, net_device->send_section_cnt);
 561
 562         /* Setup state for managing the send buffer. */
 563         net_device->send_section_map = bitmap_zalloc(net_device->send_section_cnt,
 564                                                      GFP_KERNEL);
 565         if (!net_device->send_section_map) {
 566                 ret = -ENOMEM;
 567                 goto cleanup;
 568         }
 569
 570         goto exit;
 571
 572 cleanup:
 573         netvsc_revoke_recv_buf(device, net_device, ndev);
 574         netvsc_revoke_send_buf(device, net_device, ndev);
 575         netvsc_teardown_recv_gpadl(device, net_device, ndev);
 576         netvsc_teardown_send_gpadl(device, net_device, ndev);
 577
 578 exit:
 579         return ret;
 580 }
 581
 582 /* Negotiate NVSP protocol version */
 583 static int negotiate_nvsp_ver(struct hv_device *device,
 584                               struct netvsc_device *net_device,
 585                               struct nvsp_message *init_packet,
 586                               u32 nvsp_ver)
 587 {
 588         struct net_device *ndev = hv_get_drvdata(device);
 589         int ret;
 590
 591         memset(init_packet, 0, sizeof(struct nvsp_message));
 592         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 593         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 594         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 595         trace_nvsp_send(ndev, init_packet);
 596
 597         /* Send the init request */
 598         ret = vmbus_sendpacket(device->channel, init_packet,
 599                                sizeof(struct nvsp_message),
 600                                (unsigned long)init_packet,
 601                                VM_PKT_DATA_INBAND,
 602                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 603
 604         if (ret != 0)
 605                 return ret;
 606
 607         wait_for_completion(&net_device->channel_init_wait);
 608
 609         if (init_packet->msg.init_msg.init_complete.status !=
 610             NVSP_STAT_SUCCESS)
 611                 return -EINVAL;
 612
 613         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 614                 return 0;
 615
 616         /* NVSPv2 or later: Send NDIS config */
 617         memset(init_packet, 0, sizeof(struct nvsp_message));
 618         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 619         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 620         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 621
 622         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 623                 if (hv_is_isolation_supported())
 624                         netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n");
 625                 else
 626                         init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 627
 628                 /* Teaming bit is needed to receive link speed updates */
 629                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 630         }
 631
 632         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 633                 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 634
 635         trace_nvsp_send(ndev, init_packet);
 636
 637         ret = vmbus_sendpacket(device->channel, init_packet,
 638                                 sizeof(struct nvsp_message),
 639                                 VMBUS_RQST_ID_NO_RESPONSE,
 640                                 VM_PKT_DATA_INBAND, 0);
 641
 642         return ret;
 643 }
 644
 645 static int netvsc_connect_vsp(struct hv_device *device,
 646                               struct netvsc_device *net_device,
 647                               const struct netvsc_device_info *device_info)
 648 {
 649         struct net_device *ndev = hv_get_drvdata(device);
 650         static const u32 ver_list[] = {
 651                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 652                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 653                 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 654         };
 655         struct nvsp_message *init_packet;
 656         int ndis_version, i, ret;
 657
 658         init_packet = &net_device->channel_init_pkt;
 659
 660         /* Negotiate the latest NVSP protocol supported */
 661         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 662                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 663                                        ver_list[i])  == 0) {
 664                         net_device->nvsp_version = ver_list[i];
 665                         break;
 666                 }
 667
 668         if (i < 0) {
 669                 ret = -EPROTO;
 670                 goto cleanup;
 671         }
 672
 673         if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) {
 674                 netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n",
 675                            net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61);
 676                 ret = -EPROTO;
 677                 goto cleanup;
 678         }
 679
 680         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 681
 682         /* Send the ndis version */
 683         memset(init_packet, 0, sizeof(struct nvsp_message));
 684
 685         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 686                 ndis_version = 0x00060001;
 687         else
 688                 ndis_version = 0x0006001e;
 689
 690         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 691         init_packet->msg.v1_msg.
 692                 send_ndis_ver.ndis_major_ver =
 693                                 (ndis_version & 0xFFFF0000) >> 16;
 694         init_packet->msg.v1_msg.
 695                 send_ndis_ver.ndis_minor_ver =
 696                                 ndis_version & 0xFFFF;
 697
 698         trace_nvsp_send(ndev, init_packet);
 699
 700         /* Send the init request */
 701         ret = vmbus_sendpacket(device->channel, init_packet,
 702                                 sizeof(struct nvsp_message),
 703                                 VMBUS_RQST_ID_NO_RESPONSE,
 704                                 VM_PKT_DATA_INBAND, 0);
 705         if (ret != 0)
 706                 goto cleanup;
 707
 708
 709         ret = netvsc_init_buf(device, net_device, device_info);
 710
 711 cleanup:
 712         return ret;
 713 }
 714
 715 /*
 716  * netvsc_device_remove - Callback when the root bus device is removed
 717  */
 718 void netvsc_device_remove(struct hv_device *device)
 719 {
 720         struct net_device *ndev = hv_get_drvdata(device);
 721         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 722         struct netvsc_device *net_device
 723                 = rtnl_dereference(net_device_ctx->nvdev);
 724         int i;
 725
 726         /*
 727          * Revoke receive buffer. If host is pre-Win2016 then tear down
 728          * receive buffer GPADL. Do the same for send buffer.
 729          */
 730         netvsc_revoke_recv_buf(device, net_device, ndev);
 731         if (vmbus_proto_version < VERSION_WIN10)
 732                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 733
 734         netvsc_revoke_send_buf(device, net_device, ndev);
 735         if (vmbus_proto_version < VERSION_WIN10)
 736                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 737
 738         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 739
 740         /* Disable NAPI and disassociate its context from the device. */
 741         for (i = 0; i < net_device->num_chn; i++) {
 742                 /* See also vmbus_reset_channel_cb(). */
 743                 napi_disable(&net_device->chan_table[i].napi);
 744                 netif_napi_del(&net_device->chan_table[i].napi);
 745         }
 746
 747         /*
 748          * At this point, no one should be accessing net_device
 749          * except in here
 750          */
 751         netdev_dbg(ndev, "net device safe to remove\n");
 752
 753         /* Now, we can close the channel safely */
 754         vmbus_close(device->channel);
 755
 756         /*
 757          * If host is Win2016 or higher then we do the GPADL tear down
 758          * here after VMBus is closed.
 759         */
 760         if (vmbus_proto_version >= VERSION_WIN10) {
 761                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 762                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 763         }
 764
 765         if (net_device->recv_original_buf)
 766                 hv_unmap_memory(net_device->recv_buf);
 767
 768         if (net_device->send_original_buf)
 769                 hv_unmap_memory(net_device->send_buf);
 770
 771         /* Release all resources */
 772         free_netvsc_device_rcu(net_device);
 773 }
 774
 775 #define RING_AVAIL_PERCENT_HIWATER 20
 776 #define RING_AVAIL_PERCENT_LOWATER 10
 777
 778 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 779                                          u32 index)
 780 {
 781         sync_change_bit(index, net_device->send_section_map);
 782 }
 783
 784 static void netvsc_send_tx_complete(struct net_device *ndev,
 785                                     struct netvsc_device *net_device,
 786                                     struct vmbus_channel *channel,
 787                                     const struct vmpacket_descriptor *desc,
 788                                     int budget)
 789 {
 790         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 791         struct sk_buff *skb;
 792         u16 q_idx = 0;
 793         int queue_sends;
 794         u64 cmd_rqst;
 795
 796         cmd_rqst = channel->request_addr_callback(channel, desc->trans_id);
 797         if (cmd_rqst == VMBUS_RQST_ERROR) {
 798                 netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 799                 return;
 800         }
 801
 802         skb = (struct sk_buff *)(unsigned long)cmd_rqst;
 803
 804         /* Notify the layer above us */
 805         if (likely(skb)) {
 806                 struct hv_netvsc_packet *packet
 807                         = (struct hv_netvsc_packet *)skb->cb;
 808                 u32 send_index = packet->send_buf_index;
 809                 struct netvsc_stats_tx *tx_stats;
 810
 811                 if (send_index != NETVSC_INVALID_INDEX)
 812                         netvsc_free_send_slot(net_device, send_index);
 813                 q_idx = packet->q_idx;
 814
 815                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 816
 817                 u64_stats_update_begin(&tx_stats->syncp);
 818                 tx_stats->packets += packet->total_packets;
 819                 tx_stats->bytes += packet->total_bytes;
 820                 u64_stats_update_end(&tx_stats->syncp);
 821
 822                 netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
 823                 napi_consume_skb(skb, budget);
 824         }
 825
 826         queue_sends =
 827                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 828
 829         if (unlikely(net_device->destroy)) {
 830                 if (queue_sends == 0)
 831                         wake_up(&net_device->wait_drain);
 832         } else {
 833                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 834
 835                 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 836                     (hv_get_avail_to_write_percent(&channel->outbound) >
 837                      RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 838                         netif_tx_wake_queue(txq);
 839                         ndev_ctx->eth_stats.wake_queue++;
 840                 }
 841         }
 842 }
 843
 844 static void netvsc_send_completion(struct net_device *ndev,
 845                                    struct netvsc_device *net_device,
 846                                    struct vmbus_channel *incoming_channel,
 847                                    const struct vmpacket_descriptor *desc,
 848                                    int budget)
 849 {
 850         const struct nvsp_message *nvsp_packet;
 851         u32 msglen = hv_pkt_datalen(desc);
 852         struct nvsp_message *pkt_rqst;
 853         u64 cmd_rqst;
 854
 855         /* First check if this is a VMBUS completion without data payload */
 856         if (!msglen) {
 857                 cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
 858                                                                    desc->trans_id);
 859                 if (cmd_rqst == VMBUS_RQST_ERROR) {
 860                         netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
 861                         return;
 862                 }
 863
 864                 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
 865                 switch (pkt_rqst->hdr.msg_type) {
 866                 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
 867                         complete(&net_device->channel_init_wait);
 868                         break;
 869
 870                 default:
 871                         netdev_err(ndev, "Unexpected VMBUS completion!!\n");
 872                 }
 873                 return;
 874         }
 875
 876         /* Ensure packet is big enough to read header fields */
 877         if (msglen < sizeof(struct nvsp_message_header)) {
 878                 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
 879                 return;
 880         }
 881
 882         nvsp_packet = hv_pkt_data(desc);
 883         switch (nvsp_packet->hdr.msg_type) {
 884         case NVSP_MSG_TYPE_INIT_COMPLETE:
 885                 if (msglen < sizeof(struct nvsp_message_header) +
 886                                 sizeof(struct nvsp_message_init_complete)) {
 887                         netdev_err(ndev, "nvsp_msg length too small: %u\n",
 888                                    msglen);
 889                         return;
 890                 }
 891                 fallthrough;
 892
 893         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 894                 if (msglen < sizeof(struct nvsp_message_header) +
 895                                 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
 896                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 897                                    msglen);
 898                         return;
 899                 }
 900                 fallthrough;
 901
 902         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 903                 if (msglen < sizeof(struct nvsp_message_header) +
 904                                 sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
 905                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 906                                    msglen);
 907                         return;
 908                 }
 909                 fallthrough;
 910
 911         case NVSP_MSG5_TYPE_SUBCHANNEL:
 912                 if (msglen < sizeof(struct nvsp_message_header) +
 913                                 sizeof(struct nvsp_5_subchannel_complete)) {
 914                         netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
 915                                    msglen);
 916                         return;
 917                 }
 918                 /* Copy the response back */
 919                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 920                        sizeof(struct nvsp_message));
 921                 complete(&net_device->channel_init_wait);
 922                 break;
 923
 924         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 925                 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 926                                         desc, budget);
 927                 break;
 928
 929         default:
 930                 netdev_err(ndev,
 931                            "Unknown send completion type %d received!!\n",
 932                            nvsp_packet->hdr.msg_type);
 933         }
 934 }
 935
 936 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 937 {
 938         unsigned long *map_addr = net_device->send_section_map;
 939         unsigned int i;
 940
 941         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 942                 if (sync_test_and_set_bit(i, map_addr) == 0)
 943                         return i;
 944         }
 945
 946         return NETVSC_INVALID_INDEX;
 947 }
 948
 949 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 950                                     unsigned int section_index,
 951                                     u32 pend_size,
 952                                     struct hv_netvsc_packet *packet,
 953                                     struct rndis_message *rndis_msg,
 954                                     struct hv_page_buffer *pb,
 955                                     bool xmit_more)
 956 {
 957         char *start = net_device->send_buf;
 958         char *dest = start + (section_index * net_device->send_section_size)
 959                      + pend_size;
 960         int i;
 961         u32 padding = 0;
 962         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 963                 packet->page_buf_cnt;
 964         u32 remain;
 965
 966         /* Add padding */
 967         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 968         if (xmit_more && remain) {
 969                 padding = net_device->pkt_align - remain;
 970                 rndis_msg->msg_len += padding;
 971                 packet->total_data_buflen += padding;
 972         }
 973
 974         for (i = 0; i < page_count; i++) {
 975                 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
 976                 u32 offset = pb[i].offset;
 977                 u32 len = pb[i].len;
 978
 979                 memcpy(dest, (src + offset), len);
 980                 dest += len;
 981         }
 982
 983         if (padding)
 984                 memset(dest, 0, padding);
 985 }
 986
 987 void netvsc_dma_unmap(struct hv_device *hv_dev,
 988                       struct hv_netvsc_packet *packet)
 989 {
 990         u32 page_count = packet->cp_partial ?
 991                 packet->page_buf_cnt - packet->rmsg_pgcnt :
 992                 packet->page_buf_cnt;
 993         int i;
 994
 995         if (!hv_is_isolation_supported())
 996                 return;
 997
 998         if (!packet->dma_range)
 999                 return;
1000
1001         for (i = 0; i < page_count; i++)
1002                 dma_unmap_single(&hv_dev->device, packet->dma_range[i].dma,
1003                                  packet->dma_range[i].mapping_size,
1004                                  DMA_TO_DEVICE);
1005
1006         kfree(packet->dma_range);
1007 }
1008
1009 /* netvsc_dma_map - Map swiotlb bounce buffer with data page of
1010  * packet sent by vmbus_sendpacket_pagebuffer() in the Isolation
1011  * VM.
1012  *
1013  * In isolation VM, netvsc send buffer has been marked visible to
1014  * host and so the data copied to send buffer doesn't need to use
1015  * bounce buffer. The data pages handled by vmbus_sendpacket_pagebuffer()
1016  * may not be copied to send buffer and so these pages need to be
1017  * mapped with swiotlb bounce buffer. netvsc_dma_map() is to do
1018  * that. The pfns in the struct hv_page_buffer need to be converted
1019  * to bounce buffer's pfn. The loop here is necessary because the
1020  * entries in the page buffer array are not necessarily full
1021  * pages of data.  Each entry in the array has a separate offset and
1022  * len that may be non-zero, even for entries in the middle of the
1023  * array.  And the entries are not physically contiguous.  So each
1024  * entry must be individually mapped rather than as a contiguous unit.
1025  * So not use dma_map_sg() here.
1026  */
1027 static int netvsc_dma_map(struct hv_device *hv_dev,
1028                           struct hv_netvsc_packet *packet,
1029                           struct hv_page_buffer *pb)
1030 {
1031         u32 page_count =  packet->cp_partial ?
1032                 packet->page_buf_cnt - packet->rmsg_pgcnt :
1033                 packet->page_buf_cnt;
1034         dma_addr_t dma;
1035         int i;
1036
1037         if (!hv_is_isolation_supported())
1038                 return 0;
1039
1040         packet->dma_range = kcalloc(page_count,
1041                                     sizeof(*packet->dma_range),
1042                                     GFP_KERNEL);
1043         if (!packet->dma_range)
1044                 return -ENOMEM;
1045
1046         for (i = 0; i < page_count; i++) {
1047                 char *src = phys_to_virt((pb[i].pfn << HV_HYP_PAGE_SHIFT)
1048                                          + pb[i].offset);
1049                 u32 len = pb[i].len;
1050
1051                 dma = dma_map_single(&hv_dev->device, src, len,
1052                                      DMA_TO_DEVICE);
1053                 if (dma_mapping_error(&hv_dev->device, dma)) {
1054                         kfree(packet->dma_range);
1055                         return -ENOMEM;
1056                 }
1057
1058                 /* pb[].offset and pb[].len are not changed during dma mapping
1059                  * and so not reassign.
1060                  */
1061                 packet->dma_range[i].dma = dma;
1062                 packet->dma_range[i].mapping_size = len;
1063                 pb[i].pfn = dma >> HV_HYP_PAGE_SHIFT;
1064         }
1065
1066         return 0;
1067 }
1068
1069 static inline int netvsc_send_pkt(
1070         struct hv_device *device,
1071         struct hv_netvsc_packet *packet,
1072         struct netvsc_device *net_device,
1073         struct hv_page_buffer *pb,
1074         struct sk_buff *skb)
1075 {
1076         struct nvsp_message nvmsg;
1077         struct nvsp_1_message_send_rndis_packet *rpkt =
1078                 &nvmsg.msg.v1_msg.send_rndis_pkt;
1079         struct netvsc_channel * const nvchan =
1080                 &net_device->chan_table[packet->q_idx];
1081         struct vmbus_channel *out_channel = nvchan->channel;
1082         struct net_device *ndev = hv_get_drvdata(device);
1083         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1084         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
1085         u64 req_id;
1086         int ret;
1087         u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
1088
1089         memset(&nvmsg, 0, sizeof(struct nvsp_message));
1090         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
1091         if (skb)
1092                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
1093         else
1094                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
1095
1096         rpkt->send_buf_section_index = packet->send_buf_index;
1097         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
1098                 rpkt->send_buf_section_size = 0;
1099         else
1100                 rpkt->send_buf_section_size = packet->total_data_buflen;
1101
1102         req_id = (ulong)skb;
1103
1104         if (out_channel->rescind)
1105                 return -ENODEV;
1106
1107         trace_nvsp_send_pkt(ndev, out_channel, rpkt);
1108
1109         packet->dma_range = NULL;
1110         if (packet->page_buf_cnt) {
1111                 if (packet->cp_partial)
1112                         pb += packet->rmsg_pgcnt;
1113
1114                 ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb);
1115                 if (ret) {
1116                         ret = -EAGAIN;
1117                         goto exit;
1118                 }
1119
1120                 ret = vmbus_sendpacket_pagebuffer(out_channel,
1121                                                   pb, packet->page_buf_cnt,
1122                                                   &nvmsg, sizeof(nvmsg),
1123                                                   req_id);
1124
1125                 if (ret)
1126                         netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
1127         } else {
1128                 ret = vmbus_sendpacket(out_channel,
1129                                        &nvmsg, sizeof(nvmsg),
1130                                        req_id, VM_PKT_DATA_INBAND,
1131                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1132         }
1133
1134 exit:
1135         if (ret == 0) {
1136                 atomic_inc_return(&nvchan->queue_sends);
1137
1138                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
1139                         netif_tx_stop_queue(txq);
1140                         ndev_ctx->eth_stats.stop_queue++;
1141                 }
1142         } else if (ret == -EAGAIN) {
1143                 netif_tx_stop_queue(txq);
1144                 ndev_ctx->eth_stats.stop_queue++;
1145         } else {
1146                 netdev_err(ndev,
1147                            "Unable to send packet pages %u len %u, ret %d\n",
1148                            packet->page_buf_cnt, packet->total_data_buflen,
1149                            ret);
1150         }
1151
1152         if (netif_tx_queue_stopped(txq) &&
1153             atomic_read(&nvchan->queue_sends) < 1 &&
1154             !net_device->tx_disable) {
1155                 netif_tx_wake_queue(txq);
1156                 ndev_ctx->eth_stats.wake_queue++;
1157                 if (ret == -EAGAIN)
1158                         ret = -ENOSPC;
1159         }
1160
1161         return ret;
1162 }
1163
1164 /* Move packet out of multi send data (msd), and clear msd */
1165 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
1166                                 struct sk_buff **msd_skb,
1167                                 struct multi_send_data *msdp)
1168 {
1169         *msd_skb = msdp->skb;
1170         *msd_send = msdp->pkt;
1171         msdp->skb = NULL;
1172         msdp->pkt = NULL;
1173         msdp->count = 0;
1174 }
1175
1176 /* RCU already held by caller */
1177 /* Batching/bouncing logic is designed to attempt to optimize
1178  * performance.
1179  *
1180  * For small, non-LSO packets we copy the packet to a send buffer
1181  * which is pre-registered with the Hyper-V side. This enables the
1182  * hypervisor to avoid remapping the aperture to access the packet
1183  * descriptor and data.
1184  *
1185  * If we already started using a buffer and the netdev is transmitting
1186  * a burst of packets, keep on copying into the buffer until it is
1187  * full or we are done collecting a burst. If there is an existing
1188  * buffer with space for the RNDIS descriptor but not the packet, copy
1189  * the RNDIS descriptor to the buffer, keeping the packet in place.
1190  *
1191  * If we do batching and send more than one packet using a single
1192  * NetVSC message, free the SKBs of the packets copied, except for the
1193  * last packet. This is done to streamline the handling of the case
1194  * where the last packet only had the RNDIS descriptor copied to the
1195  * send buffer, with the data pointers included in the NetVSC message.
1196  */
1197 int netvsc_send(struct net_device *ndev,
1198                 struct hv_netvsc_packet *packet,
1199                 struct rndis_message *rndis_msg,
1200                 struct hv_page_buffer *pb,
1201                 struct sk_buff *skb,
1202                 bool xdp_tx)
1203 {
1204         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1205         struct netvsc_device *net_device
1206                 = rcu_dereference_bh(ndev_ctx->nvdev);
1207         struct hv_device *device = ndev_ctx->device_ctx;
1208         int ret = 0;
1209         struct netvsc_channel *nvchan;
1210         u32 pktlen = packet->total_data_buflen, msd_len = 0;
1211         unsigned int section_index = NETVSC_INVALID_INDEX;
1212         struct multi_send_data *msdp;
1213         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1214         struct sk_buff *msd_skb = NULL;
1215         bool try_batch, xmit_more;
1216
1217         /* If device is rescinded, return error and packet will get dropped. */
1218         if (unlikely(!net_device || net_device->destroy))
1219                 return -ENODEV;
1220
1221         nvchan = &net_device->chan_table[packet->q_idx];
1222         packet->send_buf_index = NETVSC_INVALID_INDEX;
1223         packet->cp_partial = false;
1224
1225         /* Send a control message or XDP packet directly without accessing
1226          * msd (Multi-Send Data) field which may be changed during data packet
1227          * processing.
1228          */
1229         if (!skb || xdp_tx)
1230                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
1231
1232         /* batch packets in send buffer if possible */
1233         msdp = &nvchan->msd;
1234         if (msdp->pkt)
1235                 msd_len = msdp->pkt->total_data_buflen;
1236
1237         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
1238         if (try_batch && msd_len + pktlen + net_device->pkt_align <
1239             net_device->send_section_size) {
1240                 section_index = msdp->pkt->send_buf_index;
1241
1242         } else if (try_batch && msd_len + packet->rmsg_size <
1243                    net_device->send_section_size) {
1244                 section_index = msdp->pkt->send_buf_index;
1245                 packet->cp_partial = true;
1246
1247         } else if (pktlen + net_device->pkt_align <
1248                    net_device->send_section_size) {
1249                 section_index = netvsc_get_next_send_section(net_device);
1250                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1251                         ++ndev_ctx->eth_stats.tx_send_full;
1252                 } else {
1253                         move_pkt_msd(&msd_send, &msd_skb, msdp);
1254                         msd_len = 0;
1255                 }
1256         }
1257
1258         /* Keep aggregating only if stack says more data is coming
1259          * and not doing mixed modes send and not flow blocked
1260          */
1261         xmit_more = netdev_xmit_more() &&
1262                 !packet->cp_partial &&
1263                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1264
1265         if (section_index != NETVSC_INVALID_INDEX) {
1266                 netvsc_copy_to_send_buf(net_device,
1267                                         section_index, msd_len,
1268                                         packet, rndis_msg, pb, xmit_more);
1269
1270                 packet->send_buf_index = section_index;
1271
1272                 if (packet->cp_partial) {
1273                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
1274                         packet->total_data_buflen = msd_len + packet->rmsg_size;
1275                 } else {
1276                         packet->page_buf_cnt = 0;
1277                         packet->total_data_buflen += msd_len;
1278                 }
1279
1280                 if (msdp->pkt) {
1281                         packet->total_packets += msdp->pkt->total_packets;
1282                         packet->total_bytes += msdp->pkt->total_bytes;
1283                 }
1284
1285                 if (msdp->skb)
1286                         dev_consume_skb_any(msdp->skb);
1287
1288                 if (xmit_more) {
1289                         msdp->skb = skb;
1290                         msdp->pkt = packet;
1291                         msdp->count++;
1292                 } else {
1293                         cur_send = packet;
1294                         msdp->skb = NULL;
1295                         msdp->pkt = NULL;
1296                         msdp->count = 0;
1297                 }
1298         } else {
1299                 move_pkt_msd(&msd_send, &msd_skb, msdp);
1300                 cur_send = packet;
1301         }
1302
1303         if (msd_send) {
1304                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1305                                             NULL, msd_skb);
1306
1307                 if (m_ret != 0) {
1308                         netvsc_free_send_slot(net_device,
1309                                               msd_send->send_buf_index);
1310                         dev_kfree_skb_any(msd_skb);
1311                 }
1312         }
1313
1314         if (cur_send)
1315                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1316
1317         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1318                 netvsc_free_send_slot(net_device, section_index);
1319
1320         return ret;
1321 }
1322
1323 /* Send pending recv completions */
1324 static int send_recv_completions(struct net_device *ndev,
1325                                  struct netvsc_device *nvdev,
1326                                  struct netvsc_channel *nvchan)
1327 {
1328         struct multi_recv_comp *mrc = &nvchan->mrc;
1329         struct recv_comp_msg {
1330                 struct nvsp_message_header hdr;
1331                 u32 status;
1332         }  __packed;
1333         struct recv_comp_msg msg = {
1334                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1335         };
1336         int ret;
1337
1338         while (mrc->first != mrc->next) {
1339                 const struct recv_comp_data *rcd
1340                         = mrc->slots + mrc->first;
1341
1342                 msg.status = rcd->status;
1343                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1344                                        rcd->tid, VM_PKT_COMP, 0);
1345                 if (unlikely(ret)) {
1346                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1347
1348                         ++ndev_ctx->eth_stats.rx_comp_busy;
1349                         return ret;
1350                 }
1351
1352                 if (++mrc->first == nvdev->recv_completion_cnt)
1353                         mrc->first = 0;
1354         }
1355
1356         /* receive completion ring has been emptied */
1357         if (unlikely(nvdev->destroy))
1358                 wake_up(&nvdev->wait_drain);
1359
1360         return 0;
1361 }
1362
1363 /* Count how many receive completions are outstanding */
1364 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1365                                  const struct multi_recv_comp *mrc,
1366                                  u32 *filled, u32 *avail)
1367 {
1368         u32 count = nvdev->recv_completion_cnt;
1369
1370         if (mrc->next >= mrc->first)
1371                 *filled = mrc->next - mrc->first;
1372         else
1373                 *filled = (count - mrc->first) + mrc->next;
1374
1375         *avail = count - *filled - 1;
1376 }
1377
1378 /* Add receive complete to ring to send to host. */
1379 static void enq_receive_complete(struct net_device *ndev,
1380                                  struct netvsc_device *nvdev, u16 q_idx,
1381                                  u64 tid, u32 status)
1382 {
1383         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1384         struct multi_recv_comp *mrc = &nvchan->mrc;
1385         struct recv_comp_data *rcd;
1386         u32 filled, avail;
1387
1388         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1389
1390         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1391                 send_recv_completions(ndev, nvdev, nvchan);
1392                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1393         }
1394
1395         if (unlikely(!avail)) {
1396                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1397                            q_idx, tid);
1398                 return;
1399         }
1400
1401         rcd = mrc->slots + mrc->next;
1402         rcd->tid = tid;
1403         rcd->status = status;
1404
1405         if (++mrc->next == nvdev->recv_completion_cnt)
1406                 mrc->next = 0;
1407 }
1408
1409 static int netvsc_receive(struct net_device *ndev,
1410                           struct netvsc_device *net_device,
1411                           struct netvsc_channel *nvchan,
1412                           const struct vmpacket_descriptor *desc)
1413 {
1414         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1415         struct vmbus_channel *channel = nvchan->channel;
1416         const struct vmtransfer_page_packet_header *vmxferpage_packet
1417                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1418         const struct nvsp_message *nvsp = hv_pkt_data(desc);
1419         u32 msglen = hv_pkt_datalen(desc);
1420         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1421         char *recv_buf = net_device->recv_buf;
1422         u32 status = NVSP_STAT_SUCCESS;
1423         int i;
1424         int count = 0;
1425
1426         /* Ensure packet is big enough to read header fields */
1427         if (msglen < sizeof(struct nvsp_message_header)) {
1428                 netif_err(net_device_ctx, rx_err, ndev,
1429                           "invalid nvsp header, length too small: %u\n",
1430                           msglen);
1431                 return 0;
1432         }
1433
1434         /* Make sure this is a valid nvsp packet */
1435         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1436                 netif_err(net_device_ctx, rx_err, ndev,
1437                           "Unknown nvsp packet type received %u\n",
1438                           nvsp->hdr.msg_type);
1439                 return 0;
1440         }
1441
1442         /* Validate xfer page pkt header */
1443         if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1444                 netif_err(net_device_ctx, rx_err, ndev,
1445                           "Invalid xfer page pkt, offset too small: %u\n",
1446                           desc->offset8 << 3);
1447                 return 0;
1448         }
1449
1450         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1451                 netif_err(net_device_ctx, rx_err, ndev,
1452                           "Invalid xfer page set id - expecting %x got %x\n",
1453                           NETVSC_RECEIVE_BUFFER_ID,
1454                           vmxferpage_packet->xfer_pageset_id);
1455                 return 0;
1456         }
1457
1458         count = vmxferpage_packet->range_cnt;
1459
1460         /* Check count for a valid value */
1461         if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1462                 netif_err(net_device_ctx, rx_err, ndev,
1463                           "Range count is not valid: %d\n",
1464                           count);
1465                 return 0;
1466         }
1467
1468         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1469         for (i = 0; i < count; i++) {
1470                 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1471                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1472                 void *data;
1473                 int ret;
1474
1475                 if (unlikely(offset > net_device->recv_buf_size ||
1476                              buflen > net_device->recv_buf_size - offset)) {
1477                         nvchan->rsc.cnt = 0;
1478                         status = NVSP_STAT_FAIL;
1479                         netif_err(net_device_ctx, rx_err, ndev,
1480                                   "Packet offset:%u + len:%u too big\n",
1481                                   offset, buflen);
1482
1483                         continue;
1484                 }
1485
1486                 /* We're going to copy (sections of) the packet into nvchan->recv_buf;
1487                  * make sure that nvchan->recv_buf is large enough to hold the packet.
1488                  */
1489                 if (unlikely(buflen > net_device->recv_section_size)) {
1490                         nvchan->rsc.cnt = 0;
1491                         status = NVSP_STAT_FAIL;
1492                         netif_err(net_device_ctx, rx_err, ndev,
1493                                   "Packet too big: buflen=%u recv_section_size=%u\n",
1494                                   buflen, net_device->recv_section_size);
1495
1496                         continue;
1497                 }
1498
1499                 data = recv_buf + offset;
1500
1501                 nvchan->rsc.is_last = (i == count - 1);
1502
1503                 trace_rndis_recv(ndev, q_idx, data);
1504
1505                 /* Pass it to the upper layer */
1506                 ret = rndis_filter_receive(ndev, net_device,
1507                                            nvchan, data, buflen);
1508
1509                 if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1510                         /* Drop incomplete packet */
1511                         nvchan->rsc.cnt = 0;
1512                         status = NVSP_STAT_FAIL;
1513                 }
1514         }
1515
1516         enq_receive_complete(ndev, net_device, q_idx,
1517                              vmxferpage_packet->d.trans_id, status);
1518
1519         return count;
1520 }
1521
1522 static void netvsc_send_table(struct net_device *ndev,
1523                               struct netvsc_device *nvscdev,
1524                               const struct nvsp_message *nvmsg,
1525                               u32 msglen)
1526 {
1527         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1528         u32 count, offset, *tab;
1529         int i;
1530
1531         /* Ensure packet is big enough to read send_table fields */
1532         if (msglen < sizeof(struct nvsp_message_header) +
1533                      sizeof(struct nvsp_5_send_indirect_table)) {
1534                 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1535                 return;
1536         }
1537
1538         count = nvmsg->msg.v5_msg.send_table.count;
1539         offset = nvmsg->msg.v5_msg.send_table.offset;
1540
1541         if (count != VRSS_SEND_TAB_SIZE) {
1542                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1543                 return;
1544         }
1545
1546         /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1547          * wrong due to a host bug. So fix the offset here.
1548          */
1549         if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1550             msglen >= sizeof(struct nvsp_message_header) +
1551             sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1552                 offset = sizeof(struct nvsp_message_header) +
1553                          sizeof(union nvsp_6_message_uber);
1554
1555         /* Boundary check for all versions */
1556         if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1557                 netdev_err(ndev, "Received send-table offset too big:%u\n",
1558                            offset);
1559                 return;
1560         }
1561
1562         tab = (void *)nvmsg + offset;
1563
1564         for (i = 0; i < count; i++)
1565                 net_device_ctx->tx_table[i] = tab[i];
1566 }
1567
1568 static void netvsc_send_vf(struct net_device *ndev,
1569                            const struct nvsp_message *nvmsg,
1570                            u32 msglen)
1571 {
1572         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1573
1574         /* Ensure packet is big enough to read its fields */
1575         if (msglen < sizeof(struct nvsp_message_header) +
1576                      sizeof(struct nvsp_4_send_vf_association)) {
1577                 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1578                 return;
1579         }
1580
1581         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1582         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1583         netdev_info(ndev, "VF slot %u %s\n",
1584                     net_device_ctx->vf_serial,
1585                     net_device_ctx->vf_alloc ? "added" : "removed");
1586 }
1587
1588 static void netvsc_receive_inband(struct net_device *ndev,
1589                                   struct netvsc_device *nvscdev,
1590                                   const struct vmpacket_descriptor *desc)
1591 {
1592         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1593         u32 msglen = hv_pkt_datalen(desc);
1594
1595         /* Ensure packet is big enough to read header fields */
1596         if (msglen < sizeof(struct nvsp_message_header)) {
1597                 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1598                 return;
1599         }
1600
1601         switch (nvmsg->hdr.msg_type) {
1602         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1603                 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1604                 break;
1605
1606         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1607                 if (hv_is_isolation_supported())
1608                         netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n");
1609                 else
1610                         netvsc_send_vf(ndev, nvmsg, msglen);
1611                 break;
1612         }
1613 }
1614
1615 static int netvsc_process_raw_pkt(struct hv_device *device,
1616                                   struct netvsc_channel *nvchan,
1617                                   struct netvsc_device *net_device,
1618                                   struct net_device *ndev,
1619                                   const struct vmpacket_descriptor *desc,
1620                                   int budget)
1621 {
1622         struct vmbus_channel *channel = nvchan->channel;
1623         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1624
1625         trace_nvsp_recv(ndev, channel, nvmsg);
1626
1627         switch (desc->type) {
1628         case VM_PKT_COMP:
1629                 netvsc_send_completion(ndev, net_device, channel, desc, budget);
1630                 break;
1631
1632         case VM_PKT_DATA_USING_XFER_PAGES:
1633                 return netvsc_receive(ndev, net_device, nvchan, desc);
1634
1635         case VM_PKT_DATA_INBAND:
1636                 netvsc_receive_inband(ndev, net_device, desc);
1637                 break;
1638
1639         default:
1640                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1641                            desc->type, desc->trans_id);
1642                 break;
1643         }
1644
1645         return 0;
1646 }
1647
1648 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1649 {
1650         struct vmbus_channel *primary = channel->primary_channel;
1651
1652         return primary ? primary->device_obj : channel->device_obj;
1653 }
1654
1655 /* Network processing softirq
1656  * Process data in incoming ring buffer from host
1657  * Stops when ring is empty or budget is met or exceeded.
1658  */
1659 int netvsc_poll(struct napi_struct *napi, int budget)
1660 {
1661         struct netvsc_channel *nvchan
1662                 = container_of(napi, struct netvsc_channel, napi);
1663         struct netvsc_device *net_device = nvchan->net_device;
1664         struct vmbus_channel *channel = nvchan->channel;
1665         struct hv_device *device = netvsc_channel_to_device(channel);
1666         struct net_device *ndev = hv_get_drvdata(device);
1667         int work_done = 0;
1668         int ret;
1669
1670         /* If starting a new interval */
1671         if (!nvchan->desc)
1672                 nvchan->desc = hv_pkt_iter_first(channel);
1673
1674         nvchan->xdp_flush = false;
1675
1676         while (nvchan->desc && work_done < budget) {
1677                 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1678                                                     ndev, nvchan->desc, budget);
1679                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1680         }
1681
1682         if (nvchan->xdp_flush)
1683                 xdp_do_flush();
1684
1685         /* Send any pending receive completions */
1686         ret = send_recv_completions(ndev, net_device, nvchan);
1687
1688         /* If it did not exhaust NAPI budget this time
1689          *  and not doing busy poll
1690          * then re-enable host interrupts
1691          *  and reschedule if ring is not empty
1692          *   or sending receive completion failed.
1693          */
1694         if (work_done < budget &&
1695             napi_complete_done(napi, work_done) &&
1696             (ret || hv_end_read(&channel->inbound)) &&
1697             napi_schedule_prep(napi)) {
1698                 hv_begin_read(&channel->inbound);
1699                 __napi_schedule(napi);
1700         }
1701
1702         /* Driver may overshoot since multiple packets per descriptor */
1703         return min(work_done, budget);
1704 }
1705
1706 /* Call back when data is available in host ring buffer.
1707  * Processing is deferred until network softirq (NAPI)
1708  */
1709 void netvsc_channel_cb(void *context)
1710 {
1711         struct netvsc_channel *nvchan = context;
1712         struct vmbus_channel *channel = nvchan->channel;
1713         struct hv_ring_buffer_info *rbi = &channel->inbound;
1714
1715         /* preload first vmpacket descriptor */
1716         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1717
1718         if (napi_schedule_prep(&nvchan->napi)) {
1719                 /* disable interrupts from host */
1720                 hv_begin_read(rbi);
1721
1722                 __napi_schedule_irqoff(&nvchan->napi);
1723         }
1724 }
1725
1726 /*
1727  * netvsc_device_add - Callback when the device belonging to this
1728  * driver is added
1729  */
1730 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1731                                 const struct netvsc_device_info *device_info)
1732 {
1733         int i, ret = 0;
1734         struct netvsc_device *net_device;
1735         struct net_device *ndev = hv_get_drvdata(device);
1736         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1737
1738         net_device = alloc_net_device();
1739         if (!net_device)
1740                 return ERR_PTR(-ENOMEM);
1741
1742         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1743                 net_device_ctx->tx_table[i] = 0;
1744
1745         /* Because the device uses NAPI, all the interrupt batching and
1746          * control is done via Net softirq, not the channel handling
1747          */
1748         set_channel_read_mode(device->channel, HV_CALL_ISR);
1749
1750         /* If we're reopening the device we may have multiple queues, fill the
1751          * chn_table with the default channel to use it before subchannels are
1752          * opened.
1753          * Initialize the channel state before we open;
1754          * we can be interrupted as soon as we open the channel.
1755          */
1756
1757         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1758                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1759
1760                 nvchan->channel = device->channel;
1761                 nvchan->net_device = net_device;
1762                 u64_stats_init(&nvchan->tx_stats.syncp);
1763                 u64_stats_init(&nvchan->rx_stats.syncp);
1764
1765                 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1766
1767                 if (ret) {
1768                         netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1769                         goto cleanup2;
1770                 }
1771
1772                 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1773                                                  MEM_TYPE_PAGE_SHARED, NULL);
1774
1775                 if (ret) {
1776                         netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1777                         goto cleanup2;
1778                 }
1779         }
1780
1781         /* Enable NAPI handler before init callbacks */
1782         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1783                        netvsc_poll, NAPI_POLL_WEIGHT);
1784
1785         /* Open the channel */
1786         device->channel->next_request_id_callback = vmbus_next_request_id;
1787         device->channel->request_addr_callback = vmbus_request_addr;
1788         device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1789         device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE;
1790
1791         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1792                          netvsc_ring_bytes,  NULL, 0,
1793                          netvsc_channel_cb, net_device->chan_table);
1794
1795         if (ret != 0) {
1796                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1797                 goto cleanup;
1798         }
1799
1800         /* Channel is opened */
1801         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1802
1803         napi_enable(&net_device->chan_table[0].napi);
1804
1805         /* Connect with the NetVsp */
1806         ret = netvsc_connect_vsp(device, net_device, device_info);
1807         if (ret != 0) {
1808                 netdev_err(ndev,
1809                         "unable to connect to NetVSP - %d\n", ret);
1810                 goto close;
1811         }
1812
1813         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1814          * populated.
1815          */
1816         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1817
1818         return net_device;
1819
1820 close:
1821         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1822         napi_disable(&net_device->chan_table[0].napi);
1823
1824         /* Now, we can close the channel safely */
1825         vmbus_close(device->channel);
1826
1827 cleanup:
1828         netif_napi_del(&net_device->chan_table[0].napi);
1829
1830 cleanup2:
1831         if (net_device->recv_original_buf)
1832                 hv_unmap_memory(net_device->recv_buf);
1833
1834         if (net_device->send_original_buf)
1835                 hv_unmap_memory(net_device->send_buf);
1836
1837         free_netvsc_device(&net_device->rcu);
1838
1839         return ERR_PTR(ret);
1840 }