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 <[email protected]>
   7  *   Hank Janssen  <[email protected]>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/wait.h>
  14 #include <linux/mm.h>
  15 #include <linux/delay.h>
  16 #include <linux/io.h>
  17 #include <linux/slab.h>
  18 #include <linux/netdevice.h>
  19 #include <linux/if_ether.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/prefetch.h>
  23
  24 #include <asm/sync_bitops.h>
  25
  26 #include "hyperv_net.h"
  27 #include "netvsc_trace.h"
  28
  29 /*
  30  * Switch the data path from the synthetic interface to the VF
  31  * interface.
  32  */
  33 void netvsc_switch_datapath(struct net_device *ndev, bool vf)
  34 {
  35         struct net_device_context *net_device_ctx = netdev_priv(ndev);
  36         struct hv_device *dev = net_device_ctx->device_ctx;
  37         struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
  38         struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
  39
  40         /* Block sending traffic to VF if it's about to be gone */
  41         if (!vf)
  42                 net_device_ctx->data_path_is_vf = vf;
  43
  44         memset(init_pkt, 0, sizeof(struct nvsp_message));
  45         init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
  46         if (vf)
  47                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  48                         NVSP_DATAPATH_VF;
  49         else
  50                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  51                         NVSP_DATAPATH_SYNTHETIC;
  52
  53         trace_nvsp_send(ndev, init_pkt);
  54
  55         vmbus_sendpacket(dev->channel, init_pkt,
  56                                sizeof(struct nvsp_message),
  57                                (unsigned long)init_pkt,
  58                                VM_PKT_DATA_INBAND,
  59                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
  60         wait_for_completion(&nv_dev->channel_init_wait);
  61         net_device_ctx->data_path_is_vf = vf;
  62 }
  63
  64 /* Worker to setup sub channels on initial setup
  65  * Initial hotplug event occurs in softirq context
  66  * and can't wait for channels.
  67  */
  68 static void netvsc_subchan_work(struct work_struct *w)
  69 {
  70         struct netvsc_device *nvdev =
  71                 container_of(w, struct netvsc_device, subchan_work);
  72         struct rndis_device *rdev;
  73         int i, ret;
  74
  75         /* Avoid deadlock with device removal already under RTNL */
  76         if (!rtnl_trylock()) {
  77                 schedule_work(w);
  78                 return;
  79         }
  80
  81         rdev = nvdev->extension;
  82         if (rdev) {
  83                 ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL);
  84                 if (ret == 0) {
  85                         netif_device_attach(rdev->ndev);
  86                 } else {
  87                         /* fallback to only primary channel */
  88                         for (i = 1; i < nvdev->num_chn; i++)
  89                                 netif_napi_del(&nvdev->chan_table[i].napi);
  90
  91                         nvdev->max_chn = 1;
  92                         nvdev->num_chn = 1;
  93                 }
  94         }
  95
  96         rtnl_unlock();
  97 }
  98
  99 static struct netvsc_device *alloc_net_device(void)
 100 {
 101         struct netvsc_device *net_device;
 102
 103         net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
 104         if (!net_device)
 105                 return NULL;
 106
 107         init_waitqueue_head(&net_device->wait_drain);
 108         net_device->destroy = false;
 109         net_device->tx_disable = true;
 110
 111         net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
 112         net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
 113
 114         init_completion(&net_device->channel_init_wait);
 115         init_waitqueue_head(&net_device->subchan_open);
 116         INIT_WORK(&net_device->subchan_work, netvsc_subchan_work);
 117
 118         return net_device;
 119 }
 120
 121 static void free_netvsc_device(struct rcu_head *head)
 122 {
 123         struct netvsc_device *nvdev
 124                 = container_of(head, struct netvsc_device, rcu);
 125         int i;
 126
 127         kfree(nvdev->extension);
 128         vfree(nvdev->recv_buf);
 129         vfree(nvdev->send_buf);
 130         kfree(nvdev->send_section_map);
 131
 132         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 133                 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
 134                 kfree(nvdev->chan_table[i].recv_buf);
 135                 vfree(nvdev->chan_table[i].mrc.slots);
 136         }
 137
 138         kfree(nvdev);
 139 }
 140
 141 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 142 {
 143         call_rcu(&nvdev->rcu, free_netvsc_device);
 144 }
 145
 146 static void netvsc_revoke_recv_buf(struct hv_device *device,
 147                                    struct netvsc_device *net_device,
 148                                    struct net_device *ndev)
 149 {
 150         struct nvsp_message *revoke_packet;
 151         int ret;
 152
 153         /*
 154          * If we got a section count, it means we received a
 155          * SendReceiveBufferComplete msg (ie sent
 156          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 157          * to send a revoke msg here
 158          */
 159         if (net_device->recv_section_cnt) {
 160                 /* Send the revoke receive buffer */
 161                 revoke_packet = &net_device->revoke_packet;
 162                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 163
 164                 revoke_packet->hdr.msg_type =
 165                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 166                 revoke_packet->msg.v1_msg.
 167                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 168
 169                 trace_nvsp_send(ndev, revoke_packet);
 170
 171                 ret = vmbus_sendpacket(device->channel,
 172                                        revoke_packet,
 173                                        sizeof(struct nvsp_message),
 174                                        VMBUS_RQST_ID_NO_RESPONSE,
 175                                        VM_PKT_DATA_INBAND, 0);
 176                 /* If the failure is because the channel is rescinded;
 177                  * ignore the failure since we cannot send on a rescinded
 178                  * channel. This would allow us to properly cleanup
 179                  * even when the channel is rescinded.
 180                  */
 181                 if (device->channel->rescind)
 182                         ret = 0;
 183                 /*
 184                  * If we failed here, we might as well return and
 185                  * have a leak rather than continue and a bugchk
 186                  */
 187                 if (ret != 0) {
 188                         netdev_err(ndev, "unable to send "
 189                                 "revoke receive buffer to netvsp\n");
 190                         return;
 191                 }
 192                 net_device->recv_section_cnt = 0;
 193         }
 194 }
 195
 196 static void netvsc_revoke_send_buf(struct hv_device *device,
 197                                    struct netvsc_device *net_device,
 198                                    struct net_device *ndev)
 199 {
 200         struct nvsp_message *revoke_packet;
 201         int ret;
 202
 203         /* Deal with the send buffer we may have setup.
 204          * If we got a  send section size, it means we received a
 205          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 206          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 207          * to send a revoke msg here
 208          */
 209         if (net_device->send_section_cnt) {
 210                 /* Send the revoke receive buffer */
 211                 revoke_packet = &net_device->revoke_packet;
 212                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 213
 214                 revoke_packet->hdr.msg_type =
 215                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 216                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 217                         NETVSC_SEND_BUFFER_ID;
 218
 219                 trace_nvsp_send(ndev, revoke_packet);
 220
 221                 ret = vmbus_sendpacket(device->channel,
 222                                        revoke_packet,
 223                                        sizeof(struct nvsp_message),
 224                                        VMBUS_RQST_ID_NO_RESPONSE,
 225                                        VM_PKT_DATA_INBAND, 0);
 226
 227                 /* If the failure is because the channel is rescinded;
 228                  * ignore the failure since we cannot send on a rescinded
 229                  * channel. This would allow us to properly cleanup
 230                  * even when the channel is rescinded.
 231                  */
 232                 if (device->channel->rescind)
 233                         ret = 0;
 234
 235                 /* If we failed here, we might as well return and
 236                  * have a leak rather than continue and a bugchk
 237                  */
 238                 if (ret != 0) {
 239                         netdev_err(ndev, "unable to send "
 240                                    "revoke send buffer to netvsp\n");
 241                         return;
 242                 }
 243                 net_device->send_section_cnt = 0;
 244         }
 245 }
 246
 247 static void netvsc_teardown_recv_gpadl(struct hv_device *device,
 248                                        struct netvsc_device *net_device,
 249                                        struct net_device *ndev)
 250 {
 251         int ret;
 252
 253         if (net_device->recv_buf_gpadl_handle) {
 254                 ret = vmbus_teardown_gpadl(device->channel,
 255                                            net_device->recv_buf_gpadl_handle);
 256
 257                 /* If we failed here, we might as well return and have a leak
 258                  * rather than continue and a bugchk
 259                  */
 260                 if (ret != 0) {
 261                         netdev_err(ndev,
 262                                    "unable to teardown receive buffer's gpadl\n");
 263                         return;
 264                 }
 265                 net_device->recv_buf_gpadl_handle = 0;
 266         }
 267 }
 268
 269 static void netvsc_teardown_send_gpadl(struct hv_device *device,
 270                                        struct netvsc_device *net_device,
 271                                        struct net_device *ndev)
 272 {
 273         int ret;
 274
 275         if (net_device->send_buf_gpadl_handle) {
 276                 ret = vmbus_teardown_gpadl(device->channel,
 277                                            net_device->send_buf_gpadl_handle);
 278
 279                 /* If we failed here, we might as well return and have a leak
 280                  * rather than continue and a bugchk
 281                  */
 282                 if (ret != 0) {
 283                         netdev_err(ndev,
 284                                    "unable to teardown send buffer's gpadl\n");
 285                         return;
 286                 }
 287                 net_device->send_buf_gpadl_handle = 0;
 288         }
 289 }
 290
 291 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 292 {
 293         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 294         int node = cpu_to_node(nvchan->channel->target_cpu);
 295         size_t size;
 296
 297         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 298         nvchan->mrc.slots = vzalloc_node(size, node);
 299         if (!nvchan->mrc.slots)
 300                 nvchan->mrc.slots = vzalloc(size);
 301
 302         return nvchan->mrc.slots ? 0 : -ENOMEM;
 303 }
 304
 305 static int netvsc_init_buf(struct hv_device *device,
 306                            struct netvsc_device *net_device,
 307                            const struct netvsc_device_info *device_info)
 308 {
 309         struct nvsp_1_message_send_receive_buffer_complete *resp;
 310         struct net_device *ndev = hv_get_drvdata(device);
 311         struct nvsp_message *init_packet;
 312         unsigned int buf_size;
 313         size_t map_words;
 314         int i, ret = 0;
 315
 316         /* Get receive buffer area. */
 317         buf_size = device_info->recv_sections * device_info->recv_section_size;
 318         buf_size = roundup(buf_size, PAGE_SIZE);
 319
 320         /* Legacy hosts only allow smaller receive buffer */
 321         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 322                 buf_size = min_t(unsigned int, buf_size,
 323                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 324
 325         net_device->recv_buf = vzalloc(buf_size);
 326         if (!net_device->recv_buf) {
 327                 netdev_err(ndev,
 328                            "unable to allocate receive buffer of size %u\n",
 329                            buf_size);
 330                 ret = -ENOMEM;
 331                 goto cleanup;
 332         }
 333
 334         net_device->recv_buf_size = buf_size;
 335
 336         /*
 337          * Establish the gpadl handle for this buffer on this
 338          * channel.  Note: This call uses the vmbus connection rather
 339          * than the channel to establish the gpadl handle.
 340          */
 341         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 342                                     buf_size,
 343                                     &net_device->recv_buf_gpadl_handle);
 344         if (ret != 0) {
 345                 netdev_err(ndev,
 346                         "unable to establish receive buffer's gpadl\n");
 347                 goto cleanup;
 348         }
 349
 350         /* Notify the NetVsp of the gpadl handle */
 351         init_packet = &net_device->channel_init_pkt;
 352         memset(init_packet, 0, sizeof(struct nvsp_message));
 353         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 354         init_packet->msg.v1_msg.send_recv_buf.
 355                 gpadl_handle = net_device->recv_buf_gpadl_handle;
 356         init_packet->msg.v1_msg.
 357                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 358
 359         trace_nvsp_send(ndev, init_packet);
 360
 361         /* Send the gpadl notification request */
 362         ret = vmbus_sendpacket(device->channel, init_packet,
 363                                sizeof(struct nvsp_message),
 364                                (unsigned long)init_packet,
 365                                VM_PKT_DATA_INBAND,
 366                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 367         if (ret != 0) {
 368                 netdev_err(ndev,
 369                         "unable to send receive buffer's gpadl to netvsp\n");
 370                 goto cleanup;
 371         }
 372
 373         wait_for_completion(&net_device->channel_init_wait);
 374
 375         /* Check the response */
 376         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 377         if (resp->status != NVSP_STAT_SUCCESS) {
 378                 netdev_err(ndev,
 379                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 380                            resp->status);
 381                 ret = -EINVAL;
 382                 goto cleanup;
 383         }
 384
 385         /* Parse the response */
 386         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 387                    resp->num_sections, resp->sections[0].sub_alloc_size,
 388                    resp->sections[0].num_sub_allocs);
 389
 390         /* There should only be one section for the entire receive buffer */
 391         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 392                 ret = -EINVAL;
 393                 goto cleanup;
 394         }
 395
 396         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 397         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 398
 399         /* Ensure buffer will not overflow */
 400         if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
 401             (u64)net_device->recv_section_cnt > (u64)buf_size) {
 402                 netdev_err(ndev, "invalid recv_section_size %u\n",
 403                            net_device->recv_section_size);
 404                 ret = -EINVAL;
 405                 goto cleanup;
 406         }
 407
 408         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 409                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
 410
 411                 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
 412                 if (nvchan->recv_buf == NULL) {
 413                         ret = -ENOMEM;
 414                         goto cleanup;
 415                 }
 416         }
 417
 418         /* Setup receive completion ring.
 419          * Add 1 to the recv_section_cnt because at least one entry in a
 420          * ring buffer has to be empty.
 421          */
 422         net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
 423         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 424         if (ret)
 425                 goto cleanup;
 426
 427         /* Now setup the send buffer. */
 428         buf_size = device_info->send_sections * device_info->send_section_size;
 429         buf_size = round_up(buf_size, PAGE_SIZE);
 430
 431         net_device->send_buf = vzalloc(buf_size);
 432         if (!net_device->send_buf) {
 433                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 434                            buf_size);
 435                 ret = -ENOMEM;
 436                 goto cleanup;
 437         }
 438
 439         /* Establish the gpadl handle for this buffer on this
 440          * channel.  Note: This call uses the vmbus connection rather
 441          * than the channel to establish the gpadl handle.
 442          */
 443         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 444                                     buf_size,
 445                                     &net_device->send_buf_gpadl_handle);
 446         if (ret != 0) {
 447                 netdev_err(ndev,
 448                            "unable to establish send buffer's gpadl\n");
 449                 goto cleanup;
 450         }
 451
 452         /* Notify the NetVsp of the gpadl handle */
 453         init_packet = &net_device->channel_init_pkt;
 454         memset(init_packet, 0, sizeof(struct nvsp_message));
 455         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 456         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 457                 net_device->send_buf_gpadl_handle;
 458         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 459
 460         trace_nvsp_send(ndev, init_packet);
 461
 462         /* Send the gpadl notification request */
 463         ret = vmbus_sendpacket(device->channel, init_packet,
 464                                sizeof(struct nvsp_message),
 465                                (unsigned long)init_packet,
 466                                VM_PKT_DATA_INBAND,
 467                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 468         if (ret != 0) {
 469                 netdev_err(ndev,
 470                            "unable to send send buffer's gpadl to netvsp\n");
 471                 goto cleanup;
 472         }
 473
 474         wait_for_completion(&net_device->channel_init_wait);
 475
 476         /* Check the response */
 477         if (init_packet->msg.v1_msg.
 478             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 479                 netdev_err(ndev, "Unable to complete send buffer "
 480                            "initialization with NetVsp - status %d\n",
 481                            init_packet->msg.v1_msg.
 482                            send_send_buf_complete.status);
 483                 ret = -EINVAL;
 484                 goto cleanup;
 485         }
 486
 487         /* Parse the response */
 488         net_device->send_section_size = init_packet->msg.
 489                                 v1_msg.send_send_buf_complete.section_size;
 490         if (net_device->send_section_size < NETVSC_MTU_MIN) {
 491                 netdev_err(ndev, "invalid send_section_size %u\n",
 492                            net_device->send_section_size);
 493                 ret = -EINVAL;
 494                 goto cleanup;
 495         }
 496
 497         /* Section count is simply the size divided by the section size. */
 498         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 499
 500         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 501                    net_device->send_section_size, net_device->send_section_cnt);
 502
 503         /* Setup state for managing the send buffer. */
 504         map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 505
 506         net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 507         if (net_device->send_section_map == NULL) {
 508                 ret = -ENOMEM;
 509                 goto cleanup;
 510         }
 511
 512         goto exit;
 513
 514 cleanup:
 515         netvsc_revoke_recv_buf(device, net_device, ndev);
 516         netvsc_revoke_send_buf(device, net_device, ndev);
 517         netvsc_teardown_recv_gpadl(device, net_device, ndev);
 518         netvsc_teardown_send_gpadl(device, net_device, ndev);
 519
 520 exit:
 521         return ret;
 522 }
 523
 524 /* Negotiate NVSP protocol version */
 525 static int negotiate_nvsp_ver(struct hv_device *device,
 526                               struct netvsc_device *net_device,
 527                               struct nvsp_message *init_packet,
 528                               u32 nvsp_ver)
 529 {
 530         struct net_device *ndev = hv_get_drvdata(device);
 531         int ret;
 532
 533         memset(init_packet, 0, sizeof(struct nvsp_message));
 534         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 535         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 536         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 537         trace_nvsp_send(ndev, init_packet);
 538
 539         /* Send the init request */
 540         ret = vmbus_sendpacket(device->channel, init_packet,
 541                                sizeof(struct nvsp_message),
 542                                (unsigned long)init_packet,
 543                                VM_PKT_DATA_INBAND,
 544                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 545
 546         if (ret != 0)
 547                 return ret;
 548
 549         wait_for_completion(&net_device->channel_init_wait);
 550
 551         if (init_packet->msg.init_msg.init_complete.status !=
 552             NVSP_STAT_SUCCESS)
 553                 return -EINVAL;
 554
 555         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 556                 return 0;
 557
 558         /* NVSPv2 or later: Send NDIS config */
 559         memset(init_packet, 0, sizeof(struct nvsp_message));
 560         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 561         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 562         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 563
 564         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 565                 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 566
 567                 /* Teaming bit is needed to receive link speed updates */
 568                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 569         }
 570
 571         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 572                 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 573
 574         trace_nvsp_send(ndev, init_packet);
 575
 576         ret = vmbus_sendpacket(device->channel, init_packet,
 577                                 sizeof(struct nvsp_message),
 578                                 VMBUS_RQST_ID_NO_RESPONSE,
 579                                 VM_PKT_DATA_INBAND, 0);
 580
 581         return ret;
 582 }
 583
 584 static int netvsc_connect_vsp(struct hv_device *device,
 585                               struct netvsc_device *net_device,
 586                               const struct netvsc_device_info *device_info)
 587 {
 588         struct net_device *ndev = hv_get_drvdata(device);
 589         static const u32 ver_list[] = {
 590                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 591                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 592                 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 593         };
 594         struct nvsp_message *init_packet;
 595         int ndis_version, i, ret;
 596
 597         init_packet = &net_device->channel_init_pkt;
 598
 599         /* Negotiate the latest NVSP protocol supported */
 600         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 601                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 602                                        ver_list[i])  == 0) {
 603                         net_device->nvsp_version = ver_list[i];
 604                         break;
 605                 }
 606
 607         if (i < 0) {
 608                 ret = -EPROTO;
 609                 goto cleanup;
 610         }
 611
 612         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 613
 614         /* Send the ndis version */
 615         memset(init_packet, 0, sizeof(struct nvsp_message));
 616
 617         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 618                 ndis_version = 0x00060001;
 619         else
 620                 ndis_version = 0x0006001e;
 621
 622         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 623         init_packet->msg.v1_msg.
 624                 send_ndis_ver.ndis_major_ver =
 625                                 (ndis_version & 0xFFFF0000) >> 16;
 626         init_packet->msg.v1_msg.
 627                 send_ndis_ver.ndis_minor_ver =
 628                                 ndis_version & 0xFFFF;
 629
 630         trace_nvsp_send(ndev, init_packet);
 631
 632         /* Send the init request */
 633         ret = vmbus_sendpacket(device->channel, init_packet,
 634                                 sizeof(struct nvsp_message),
 635                                 VMBUS_RQST_ID_NO_RESPONSE,
 636                                 VM_PKT_DATA_INBAND, 0);
 637         if (ret != 0)
 638                 goto cleanup;
 639
 640
 641         ret = netvsc_init_buf(device, net_device, device_info);
 642
 643 cleanup:
 644         return ret;
 645 }
 646
 647 /*
 648  * netvsc_device_remove - Callback when the root bus device is removed
 649  */
 650 void netvsc_device_remove(struct hv_device *device)
 651 {
 652         struct net_device *ndev = hv_get_drvdata(device);
 653         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 654         struct netvsc_device *net_device
 655                 = rtnl_dereference(net_device_ctx->nvdev);
 656         int i;
 657
 658         /*
 659          * Revoke receive buffer. If host is pre-Win2016 then tear down
 660          * receive buffer GPADL. Do the same for send buffer.
 661          */
 662         netvsc_revoke_recv_buf(device, net_device, ndev);
 663         if (vmbus_proto_version < VERSION_WIN10)
 664                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 665
 666         netvsc_revoke_send_buf(device, net_device, ndev);
 667         if (vmbus_proto_version < VERSION_WIN10)
 668                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 669
 670         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 671
 672         /* Disable NAPI and disassociate its context from the device. */
 673         for (i = 0; i < net_device->num_chn; i++) {
 674                 /* See also vmbus_reset_channel_cb(). */
 675                 napi_disable(&net_device->chan_table[i].napi);
 676                 netif_napi_del(&net_device->chan_table[i].napi);
 677         }
 678
 679         /*
 680          * At this point, no one should be accessing net_device
 681          * except in here
 682          */
 683         netdev_dbg(ndev, "net device safe to remove\n");
 684
 685         /* Now, we can close the channel safely */
 686         vmbus_close(device->channel);
 687
 688         /*
 689          * If host is Win2016 or higher then we do the GPADL tear down
 690          * here after VMBus is closed.
 691         */
 692         if (vmbus_proto_version >= VERSION_WIN10) {
 693                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 694                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 695         }
 696
 697         /* Release all resources */
 698         free_netvsc_device_rcu(net_device);
 699 }
 700
 701 #define RING_AVAIL_PERCENT_HIWATER 20
 702 #define RING_AVAIL_PERCENT_LOWATER 10
 703
 704 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 705                                          u32 index)
 706 {
 707         sync_change_bit(index, net_device->send_section_map);
 708 }
 709
 710 static void netvsc_send_tx_complete(struct net_device *ndev,
 711                                     struct netvsc_device *net_device,
 712                                     struct vmbus_channel *channel,
 713                                     const struct vmpacket_descriptor *desc,
 714                                     int budget)
 715 {
 716         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 717         struct sk_buff *skb;
 718         u16 q_idx = 0;
 719         int queue_sends;
 720         u64 cmd_rqst;
 721
 722         cmd_rqst = vmbus_request_addr(&channel->requestor, (u64)desc->trans_id);
 723         if (cmd_rqst == VMBUS_RQST_ERROR) {
 724                 netdev_err(ndev, "Incorrect transaction id\n");
 725                 return;
 726         }
 727
 728         skb = (struct sk_buff *)(unsigned long)cmd_rqst;
 729
 730         /* Notify the layer above us */
 731         if (likely(skb)) {
 732                 const struct hv_netvsc_packet *packet
 733                         = (struct hv_netvsc_packet *)skb->cb;
 734                 u32 send_index = packet->send_buf_index;
 735                 struct netvsc_stats *tx_stats;
 736
 737                 if (send_index != NETVSC_INVALID_INDEX)
 738                         netvsc_free_send_slot(net_device, send_index);
 739                 q_idx = packet->q_idx;
 740
 741                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 742
 743                 u64_stats_update_begin(&tx_stats->syncp);
 744                 tx_stats->packets += packet->total_packets;
 745                 tx_stats->bytes += packet->total_bytes;
 746                 u64_stats_update_end(&tx_stats->syncp);
 747
 748                 napi_consume_skb(skb, budget);
 749         }
 750
 751         queue_sends =
 752                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 753
 754         if (unlikely(net_device->destroy)) {
 755                 if (queue_sends == 0)
 756                         wake_up(&net_device->wait_drain);
 757         } else {
 758                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 759
 760                 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 761                     (hv_get_avail_to_write_percent(&channel->outbound) >
 762                      RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 763                         netif_tx_wake_queue(txq);
 764                         ndev_ctx->eth_stats.wake_queue++;
 765                 }
 766         }
 767 }
 768
 769 static void netvsc_send_completion(struct net_device *ndev,
 770                                    struct netvsc_device *net_device,
 771                                    struct vmbus_channel *incoming_channel,
 772                                    const struct vmpacket_descriptor *desc,
 773                                    int budget)
 774 {
 775         const struct nvsp_message *nvsp_packet;
 776         u32 msglen = hv_pkt_datalen(desc);
 777         struct nvsp_message *pkt_rqst;
 778         u64 cmd_rqst;
 779
 780         /* First check if this is a VMBUS completion without data payload */
 781         if (!msglen) {
 782                 cmd_rqst = vmbus_request_addr(&incoming_channel->requestor,
 783                                               (u64)desc->trans_id);
 784                 if (cmd_rqst == VMBUS_RQST_ERROR) {
 785                         netdev_err(ndev, "Invalid transaction id\n");
 786                         return;
 787                 }
 788
 789                 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
 790                 switch (pkt_rqst->hdr.msg_type) {
 791                 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
 792                         complete(&net_device->channel_init_wait);
 793                         break;
 794
 795                 default:
 796                         netdev_err(ndev, "Unexpected VMBUS completion!!\n");
 797                 }
 798                 return;
 799         }
 800
 801         /* Ensure packet is big enough to read header fields */
 802         if (msglen < sizeof(struct nvsp_message_header)) {
 803                 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
 804                 return;
 805         }
 806
 807         nvsp_packet = hv_pkt_data(desc);
 808         switch (nvsp_packet->hdr.msg_type) {
 809         case NVSP_MSG_TYPE_INIT_COMPLETE:
 810                 if (msglen < sizeof(struct nvsp_message_header) +
 811                                 sizeof(struct nvsp_message_init_complete)) {
 812                         netdev_err(ndev, "nvsp_msg length too small: %u\n",
 813                                    msglen);
 814                         return;
 815                 }
 816                 fallthrough;
 817
 818         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 819                 if (msglen < sizeof(struct nvsp_message_header) +
 820                                 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
 821                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 822                                    msglen);
 823                         return;
 824                 }
 825                 fallthrough;
 826
 827         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 828                 if (msglen < sizeof(struct nvsp_message_header) +
 829                                 sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
 830                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 831                                    msglen);
 832                         return;
 833                 }
 834                 fallthrough;
 835
 836         case NVSP_MSG5_TYPE_SUBCHANNEL:
 837                 if (msglen < sizeof(struct nvsp_message_header) +
 838                                 sizeof(struct nvsp_5_subchannel_complete)) {
 839                         netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
 840                                    msglen);
 841                         return;
 842                 }
 843                 /* Copy the response back */
 844                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 845                        sizeof(struct nvsp_message));
 846                 complete(&net_device->channel_init_wait);
 847                 break;
 848
 849         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 850                 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 851                                         desc, budget);
 852                 break;
 853
 854         default:
 855                 netdev_err(ndev,
 856                            "Unknown send completion type %d received!!\n",
 857                            nvsp_packet->hdr.msg_type);
 858         }
 859 }
 860
 861 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 862 {
 863         unsigned long *map_addr = net_device->send_section_map;
 864         unsigned int i;
 865
 866         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 867                 if (sync_test_and_set_bit(i, map_addr) == 0)
 868                         return i;
 869         }
 870
 871         return NETVSC_INVALID_INDEX;
 872 }
 873
 874 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 875                                     unsigned int section_index,
 876                                     u32 pend_size,
 877                                     struct hv_netvsc_packet *packet,
 878                                     struct rndis_message *rndis_msg,
 879                                     struct hv_page_buffer *pb,
 880                                     bool xmit_more)
 881 {
 882         char *start = net_device->send_buf;
 883         char *dest = start + (section_index * net_device->send_section_size)
 884                      + pend_size;
 885         int i;
 886         u32 padding = 0;
 887         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 888                 packet->page_buf_cnt;
 889         u32 remain;
 890
 891         /* Add padding */
 892         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 893         if (xmit_more && remain) {
 894                 padding = net_device->pkt_align - remain;
 895                 rndis_msg->msg_len += padding;
 896                 packet->total_data_buflen += padding;
 897         }
 898
 899         for (i = 0; i < page_count; i++) {
 900                 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
 901                 u32 offset = pb[i].offset;
 902                 u32 len = pb[i].len;
 903
 904                 memcpy(dest, (src + offset), len);
 905                 dest += len;
 906         }
 907
 908         if (padding)
 909                 memset(dest, 0, padding);
 910 }
 911
 912 static inline int netvsc_send_pkt(
 913         struct hv_device *device,
 914         struct hv_netvsc_packet *packet,
 915         struct netvsc_device *net_device,
 916         struct hv_page_buffer *pb,
 917         struct sk_buff *skb)
 918 {
 919         struct nvsp_message nvmsg;
 920         struct nvsp_1_message_send_rndis_packet *rpkt =
 921                 &nvmsg.msg.v1_msg.send_rndis_pkt;
 922         struct netvsc_channel * const nvchan =
 923                 &net_device->chan_table[packet->q_idx];
 924         struct vmbus_channel *out_channel = nvchan->channel;
 925         struct net_device *ndev = hv_get_drvdata(device);
 926         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 927         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
 928         u64 req_id;
 929         int ret;
 930         u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
 931
 932         memset(&nvmsg, 0, sizeof(struct nvsp_message));
 933         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 934         if (skb)
 935                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
 936         else
 937                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
 938
 939         rpkt->send_buf_section_index = packet->send_buf_index;
 940         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
 941                 rpkt->send_buf_section_size = 0;
 942         else
 943                 rpkt->send_buf_section_size = packet->total_data_buflen;
 944
 945         req_id = (ulong)skb;
 946
 947         if (out_channel->rescind)
 948                 return -ENODEV;
 949
 950         trace_nvsp_send_pkt(ndev, out_channel, rpkt);
 951
 952         if (packet->page_buf_cnt) {
 953                 if (packet->cp_partial)
 954                         pb += packet->rmsg_pgcnt;
 955
 956                 ret = vmbus_sendpacket_pagebuffer(out_channel,
 957                                                   pb, packet->page_buf_cnt,
 958                                                   &nvmsg, sizeof(nvmsg),
 959                                                   req_id);
 960         } else {
 961                 ret = vmbus_sendpacket(out_channel,
 962                                        &nvmsg, sizeof(nvmsg),
 963                                        req_id, VM_PKT_DATA_INBAND,
 964                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 965         }
 966
 967         if (ret == 0) {
 968                 atomic_inc_return(&nvchan->queue_sends);
 969
 970                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
 971                         netif_tx_stop_queue(txq);
 972                         ndev_ctx->eth_stats.stop_queue++;
 973                 }
 974         } else if (ret == -EAGAIN) {
 975                 netif_tx_stop_queue(txq);
 976                 ndev_ctx->eth_stats.stop_queue++;
 977         } else {
 978                 netdev_err(ndev,
 979                            "Unable to send packet pages %u len %u, ret %d\n",
 980                            packet->page_buf_cnt, packet->total_data_buflen,
 981                            ret);
 982         }
 983
 984         if (netif_tx_queue_stopped(txq) &&
 985             atomic_read(&nvchan->queue_sends) < 1 &&
 986             !net_device->tx_disable) {
 987                 netif_tx_wake_queue(txq);
 988                 ndev_ctx->eth_stats.wake_queue++;
 989                 if (ret == -EAGAIN)
 990                         ret = -ENOSPC;
 991         }
 992
 993         return ret;
 994 }
 995
 996 /* Move packet out of multi send data (msd), and clear msd */
 997 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
 998                                 struct sk_buff **msd_skb,
 999                                 struct multi_send_data *msdp)
1000 {
1001         *msd_skb = msdp->skb;
1002         *msd_send = msdp->pkt;
1003         msdp->skb = NULL;
1004         msdp->pkt = NULL;
1005         msdp->count = 0;
1006 }
1007
1008 /* RCU already held by caller */
1009 int netvsc_send(struct net_device *ndev,
1010                 struct hv_netvsc_packet *packet,
1011                 struct rndis_message *rndis_msg,
1012                 struct hv_page_buffer *pb,
1013                 struct sk_buff *skb,
1014                 bool xdp_tx)
1015 {
1016         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1017         struct netvsc_device *net_device
1018                 = rcu_dereference_bh(ndev_ctx->nvdev);
1019         struct hv_device *device = ndev_ctx->device_ctx;
1020         int ret = 0;
1021         struct netvsc_channel *nvchan;
1022         u32 pktlen = packet->total_data_buflen, msd_len = 0;
1023         unsigned int section_index = NETVSC_INVALID_INDEX;
1024         struct multi_send_data *msdp;
1025         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1026         struct sk_buff *msd_skb = NULL;
1027         bool try_batch, xmit_more;
1028
1029         /* If device is rescinded, return error and packet will get dropped. */
1030         if (unlikely(!net_device || net_device->destroy))
1031                 return -ENODEV;
1032
1033         nvchan = &net_device->chan_table[packet->q_idx];
1034         packet->send_buf_index = NETVSC_INVALID_INDEX;
1035         packet->cp_partial = false;
1036
1037         /* Send a control message or XDP packet directly without accessing
1038          * msd (Multi-Send Data) field which may be changed during data packet
1039          * processing.
1040          */
1041         if (!skb || xdp_tx)
1042                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
1043
1044         /* batch packets in send buffer if possible */
1045         msdp = &nvchan->msd;
1046         if (msdp->pkt)
1047                 msd_len = msdp->pkt->total_data_buflen;
1048
1049         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
1050         if (try_batch && msd_len + pktlen + net_device->pkt_align <
1051             net_device->send_section_size) {
1052                 section_index = msdp->pkt->send_buf_index;
1053
1054         } else if (try_batch && msd_len + packet->rmsg_size <
1055                    net_device->send_section_size) {
1056                 section_index = msdp->pkt->send_buf_index;
1057                 packet->cp_partial = true;
1058
1059         } else if (pktlen + net_device->pkt_align <
1060                    net_device->send_section_size) {
1061                 section_index = netvsc_get_next_send_section(net_device);
1062                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1063                         ++ndev_ctx->eth_stats.tx_send_full;
1064                 } else {
1065                         move_pkt_msd(&msd_send, &msd_skb, msdp);
1066                         msd_len = 0;
1067                 }
1068         }
1069
1070         /* Keep aggregating only if stack says more data is coming
1071          * and not doing mixed modes send and not flow blocked
1072          */
1073         xmit_more = netdev_xmit_more() &&
1074                 !packet->cp_partial &&
1075                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1076
1077         if (section_index != NETVSC_INVALID_INDEX) {
1078                 netvsc_copy_to_send_buf(net_device,
1079                                         section_index, msd_len,
1080                                         packet, rndis_msg, pb, xmit_more);
1081
1082                 packet->send_buf_index = section_index;
1083
1084                 if (packet->cp_partial) {
1085                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
1086                         packet->total_data_buflen = msd_len + packet->rmsg_size;
1087                 } else {
1088                         packet->page_buf_cnt = 0;
1089                         packet->total_data_buflen += msd_len;
1090                 }
1091
1092                 if (msdp->pkt) {
1093                         packet->total_packets += msdp->pkt->total_packets;
1094                         packet->total_bytes += msdp->pkt->total_bytes;
1095                 }
1096
1097                 if (msdp->skb)
1098                         dev_consume_skb_any(msdp->skb);
1099
1100                 if (xmit_more) {
1101                         msdp->skb = skb;
1102                         msdp->pkt = packet;
1103                         msdp->count++;
1104                 } else {
1105                         cur_send = packet;
1106                         msdp->skb = NULL;
1107                         msdp->pkt = NULL;
1108                         msdp->count = 0;
1109                 }
1110         } else {
1111                 move_pkt_msd(&msd_send, &msd_skb, msdp);
1112                 cur_send = packet;
1113         }
1114
1115         if (msd_send) {
1116                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1117                                             NULL, msd_skb);
1118
1119                 if (m_ret != 0) {
1120                         netvsc_free_send_slot(net_device,
1121                                               msd_send->send_buf_index);
1122                         dev_kfree_skb_any(msd_skb);
1123                 }
1124         }
1125
1126         if (cur_send)
1127                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1128
1129         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1130                 netvsc_free_send_slot(net_device, section_index);
1131
1132         return ret;
1133 }
1134
1135 /* Send pending recv completions */
1136 static int send_recv_completions(struct net_device *ndev,
1137                                  struct netvsc_device *nvdev,
1138                                  struct netvsc_channel *nvchan)
1139 {
1140         struct multi_recv_comp *mrc = &nvchan->mrc;
1141         struct recv_comp_msg {
1142                 struct nvsp_message_header hdr;
1143                 u32 status;
1144         }  __packed;
1145         struct recv_comp_msg msg = {
1146                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1147         };
1148         int ret;
1149
1150         while (mrc->first != mrc->next) {
1151                 const struct recv_comp_data *rcd
1152                         = mrc->slots + mrc->first;
1153
1154                 msg.status = rcd->status;
1155                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1156                                        rcd->tid, VM_PKT_COMP, 0);
1157                 if (unlikely(ret)) {
1158                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1159
1160                         ++ndev_ctx->eth_stats.rx_comp_busy;
1161                         return ret;
1162                 }
1163
1164                 if (++mrc->first == nvdev->recv_completion_cnt)
1165                         mrc->first = 0;
1166         }
1167
1168         /* receive completion ring has been emptied */
1169         if (unlikely(nvdev->destroy))
1170                 wake_up(&nvdev->wait_drain);
1171
1172         return 0;
1173 }
1174
1175 /* Count how many receive completions are outstanding */
1176 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1177                                  const struct multi_recv_comp *mrc,
1178                                  u32 *filled, u32 *avail)
1179 {
1180         u32 count = nvdev->recv_completion_cnt;
1181
1182         if (mrc->next >= mrc->first)
1183                 *filled = mrc->next - mrc->first;
1184         else
1185                 *filled = (count - mrc->first) + mrc->next;
1186
1187         *avail = count - *filled - 1;
1188 }
1189
1190 /* Add receive complete to ring to send to host. */
1191 static void enq_receive_complete(struct net_device *ndev,
1192                                  struct netvsc_device *nvdev, u16 q_idx,
1193                                  u64 tid, u32 status)
1194 {
1195         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1196         struct multi_recv_comp *mrc = &nvchan->mrc;
1197         struct recv_comp_data *rcd;
1198         u32 filled, avail;
1199
1200         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1201
1202         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1203                 send_recv_completions(ndev, nvdev, nvchan);
1204                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1205         }
1206
1207         if (unlikely(!avail)) {
1208                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1209                            q_idx, tid);
1210                 return;
1211         }
1212
1213         rcd = mrc->slots + mrc->next;
1214         rcd->tid = tid;
1215         rcd->status = status;
1216
1217         if (++mrc->next == nvdev->recv_completion_cnt)
1218                 mrc->next = 0;
1219 }
1220
1221 static int netvsc_receive(struct net_device *ndev,
1222                           struct netvsc_device *net_device,
1223                           struct netvsc_channel *nvchan,
1224                           const struct vmpacket_descriptor *desc)
1225 {
1226         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1227         struct vmbus_channel *channel = nvchan->channel;
1228         const struct vmtransfer_page_packet_header *vmxferpage_packet
1229                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1230         const struct nvsp_message *nvsp = hv_pkt_data(desc);
1231         u32 msglen = hv_pkt_datalen(desc);
1232         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1233         char *recv_buf = net_device->recv_buf;
1234         u32 status = NVSP_STAT_SUCCESS;
1235         int i;
1236         int count = 0;
1237
1238         /* Ensure packet is big enough to read header fields */
1239         if (msglen < sizeof(struct nvsp_message_header)) {
1240                 netif_err(net_device_ctx, rx_err, ndev,
1241                           "invalid nvsp header, length too small: %u\n",
1242                           msglen);
1243                 return 0;
1244         }
1245
1246         /* Make sure this is a valid nvsp packet */
1247         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1248                 netif_err(net_device_ctx, rx_err, ndev,
1249                           "Unknown nvsp packet type received %u\n",
1250                           nvsp->hdr.msg_type);
1251                 return 0;
1252         }
1253
1254         /* Validate xfer page pkt header */
1255         if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1256                 netif_err(net_device_ctx, rx_err, ndev,
1257                           "Invalid xfer page pkt, offset too small: %u\n",
1258                           desc->offset8 << 3);
1259                 return 0;
1260         }
1261
1262         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1263                 netif_err(net_device_ctx, rx_err, ndev,
1264                           "Invalid xfer page set id - expecting %x got %x\n",
1265                           NETVSC_RECEIVE_BUFFER_ID,
1266                           vmxferpage_packet->xfer_pageset_id);
1267                 return 0;
1268         }
1269
1270         count = vmxferpage_packet->range_cnt;
1271
1272         /* Check count for a valid value */
1273         if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1274                 netif_err(net_device_ctx, rx_err, ndev,
1275                           "Range count is not valid: %d\n",
1276                           count);
1277                 return 0;
1278         }
1279
1280         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1281         for (i = 0; i < count; i++) {
1282                 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1283                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1284                 void *data;
1285                 int ret;
1286
1287                 if (unlikely(offset > net_device->recv_buf_size ||
1288                              buflen > net_device->recv_buf_size - offset)) {
1289                         nvchan->rsc.cnt = 0;
1290                         status = NVSP_STAT_FAIL;
1291                         netif_err(net_device_ctx, rx_err, ndev,
1292                                   "Packet offset:%u + len:%u too big\n",
1293                                   offset, buflen);
1294
1295                         continue;
1296                 }
1297
1298                 /* We're going to copy (sections of) the packet into nvchan->recv_buf;
1299                  * make sure that nvchan->recv_buf is large enough to hold the packet.
1300                  */
1301                 if (unlikely(buflen > net_device->recv_section_size)) {
1302                         nvchan->rsc.cnt = 0;
1303                         status = NVSP_STAT_FAIL;
1304                         netif_err(net_device_ctx, rx_err, ndev,
1305                                   "Packet too big: buflen=%u recv_section_size=%u\n",
1306                                   buflen, net_device->recv_section_size);
1307
1308                         continue;
1309                 }
1310
1311                 data = recv_buf + offset;
1312
1313                 nvchan->rsc.is_last = (i == count - 1);
1314
1315                 trace_rndis_recv(ndev, q_idx, data);
1316
1317                 /* Pass it to the upper layer */
1318                 ret = rndis_filter_receive(ndev, net_device,
1319                                            nvchan, data, buflen);
1320
1321                 if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1322                         /* Drop incomplete packet */
1323                         nvchan->rsc.cnt = 0;
1324                         status = NVSP_STAT_FAIL;
1325                 }
1326         }
1327
1328         enq_receive_complete(ndev, net_device, q_idx,
1329                              vmxferpage_packet->d.trans_id, status);
1330
1331         return count;
1332 }
1333
1334 static void netvsc_send_table(struct net_device *ndev,
1335                               struct netvsc_device *nvscdev,
1336                               const struct nvsp_message *nvmsg,
1337                               u32 msglen)
1338 {
1339         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1340         u32 count, offset, *tab;
1341         int i;
1342
1343         /* Ensure packet is big enough to read send_table fields */
1344         if (msglen < sizeof(struct nvsp_message_header) +
1345                      sizeof(struct nvsp_5_send_indirect_table)) {
1346                 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1347                 return;
1348         }
1349
1350         count = nvmsg->msg.v5_msg.send_table.count;
1351         offset = nvmsg->msg.v5_msg.send_table.offset;
1352
1353         if (count != VRSS_SEND_TAB_SIZE) {
1354                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1355                 return;
1356         }
1357
1358         /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1359          * wrong due to a host bug. So fix the offset here.
1360          */
1361         if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1362             msglen >= sizeof(struct nvsp_message_header) +
1363             sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1364                 offset = sizeof(struct nvsp_message_header) +
1365                          sizeof(union nvsp_6_message_uber);
1366
1367         /* Boundary check for all versions */
1368         if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1369                 netdev_err(ndev, "Received send-table offset too big:%u\n",
1370                            offset);
1371                 return;
1372         }
1373
1374         tab = (void *)nvmsg + offset;
1375
1376         for (i = 0; i < count; i++)
1377                 net_device_ctx->tx_table[i] = tab[i];
1378 }
1379
1380 static void netvsc_send_vf(struct net_device *ndev,
1381                            const struct nvsp_message *nvmsg,
1382                            u32 msglen)
1383 {
1384         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1385
1386         /* Ensure packet is big enough to read its fields */
1387         if (msglen < sizeof(struct nvsp_message_header) +
1388                      sizeof(struct nvsp_4_send_vf_association)) {
1389                 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1390                 return;
1391         }
1392
1393         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1394         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1395         netdev_info(ndev, "VF slot %u %s\n",
1396                     net_device_ctx->vf_serial,
1397                     net_device_ctx->vf_alloc ? "added" : "removed");
1398 }
1399
1400 static void netvsc_receive_inband(struct net_device *ndev,
1401                                   struct netvsc_device *nvscdev,
1402                                   const struct vmpacket_descriptor *desc)
1403 {
1404         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1405         u32 msglen = hv_pkt_datalen(desc);
1406
1407         /* Ensure packet is big enough to read header fields */
1408         if (msglen < sizeof(struct nvsp_message_header)) {
1409                 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1410                 return;
1411         }
1412
1413         switch (nvmsg->hdr.msg_type) {
1414         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1415                 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1416                 break;
1417
1418         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1419                 netvsc_send_vf(ndev, nvmsg, msglen);
1420                 break;
1421         }
1422 }
1423
1424 static int netvsc_process_raw_pkt(struct hv_device *device,
1425                                   struct netvsc_channel *nvchan,
1426                                   struct netvsc_device *net_device,
1427                                   struct net_device *ndev,
1428                                   const struct vmpacket_descriptor *desc,
1429                                   int budget)
1430 {
1431         struct vmbus_channel *channel = nvchan->channel;
1432         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1433
1434         trace_nvsp_recv(ndev, channel, nvmsg);
1435
1436         switch (desc->type) {
1437         case VM_PKT_COMP:
1438                 netvsc_send_completion(ndev, net_device, channel, desc, budget);
1439                 break;
1440
1441         case VM_PKT_DATA_USING_XFER_PAGES:
1442                 return netvsc_receive(ndev, net_device, nvchan, desc);
1443                 break;
1444
1445         case VM_PKT_DATA_INBAND:
1446                 netvsc_receive_inband(ndev, net_device, desc);
1447                 break;
1448
1449         default:
1450                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1451                            desc->type, desc->trans_id);
1452                 break;
1453         }
1454
1455         return 0;
1456 }
1457
1458 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1459 {
1460         struct vmbus_channel *primary = channel->primary_channel;
1461
1462         return primary ? primary->device_obj : channel->device_obj;
1463 }
1464
1465 /* Network processing softirq
1466  * Process data in incoming ring buffer from host
1467  * Stops when ring is empty or budget is met or exceeded.
1468  */
1469 int netvsc_poll(struct napi_struct *napi, int budget)
1470 {
1471         struct netvsc_channel *nvchan
1472                 = container_of(napi, struct netvsc_channel, napi);
1473         struct netvsc_device *net_device = nvchan->net_device;
1474         struct vmbus_channel *channel = nvchan->channel;
1475         struct hv_device *device = netvsc_channel_to_device(channel);
1476         struct net_device *ndev = hv_get_drvdata(device);
1477         int work_done = 0;
1478         int ret;
1479
1480         /* If starting a new interval */
1481         if (!nvchan->desc)
1482                 nvchan->desc = hv_pkt_iter_first(channel);
1483
1484         while (nvchan->desc && work_done < budget) {
1485                 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1486                                                     ndev, nvchan->desc, budget);
1487                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1488         }
1489
1490         /* Send any pending receive completions */
1491         ret = send_recv_completions(ndev, net_device, nvchan);
1492
1493         /* If it did not exhaust NAPI budget this time
1494          *  and not doing busy poll
1495          * then re-enable host interrupts
1496          *  and reschedule if ring is not empty
1497          *   or sending receive completion failed.
1498          */
1499         if (work_done < budget &&
1500             napi_complete_done(napi, work_done) &&
1501             (ret || hv_end_read(&channel->inbound)) &&
1502             napi_schedule_prep(napi)) {
1503                 hv_begin_read(&channel->inbound);
1504                 __napi_schedule(napi);
1505         }
1506
1507         /* Driver may overshoot since multiple packets per descriptor */
1508         return min(work_done, budget);
1509 }
1510
1511 /* Call back when data is available in host ring buffer.
1512  * Processing is deferred until network softirq (NAPI)
1513  */
1514 void netvsc_channel_cb(void *context)
1515 {
1516         struct netvsc_channel *nvchan = context;
1517         struct vmbus_channel *channel = nvchan->channel;
1518         struct hv_ring_buffer_info *rbi = &channel->inbound;
1519
1520         /* preload first vmpacket descriptor */
1521         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1522
1523         if (napi_schedule_prep(&nvchan->napi)) {
1524                 /* disable interrupts from host */
1525                 hv_begin_read(rbi);
1526
1527                 __napi_schedule_irqoff(&nvchan->napi);
1528         }
1529 }
1530
1531 /*
1532  * netvsc_device_add - Callback when the device belonging to this
1533  * driver is added
1534  */
1535 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1536                                 const struct netvsc_device_info *device_info)
1537 {
1538         int i, ret = 0;
1539         struct netvsc_device *net_device;
1540         struct net_device *ndev = hv_get_drvdata(device);
1541         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1542
1543         net_device = alloc_net_device();
1544         if (!net_device)
1545                 return ERR_PTR(-ENOMEM);
1546
1547         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1548                 net_device_ctx->tx_table[i] = 0;
1549
1550         /* Because the device uses NAPI, all the interrupt batching and
1551          * control is done via Net softirq, not the channel handling
1552          */
1553         set_channel_read_mode(device->channel, HV_CALL_ISR);
1554
1555         /* If we're reopening the device we may have multiple queues, fill the
1556          * chn_table with the default channel to use it before subchannels are
1557          * opened.
1558          * Initialize the channel state before we open;
1559          * we can be interrupted as soon as we open the channel.
1560          */
1561
1562         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1563                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1564
1565                 nvchan->channel = device->channel;
1566                 nvchan->net_device = net_device;
1567                 u64_stats_init(&nvchan->tx_stats.syncp);
1568                 u64_stats_init(&nvchan->rx_stats.syncp);
1569
1570                 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1571
1572                 if (ret) {
1573                         netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1574                         goto cleanup2;
1575                 }
1576
1577                 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1578                                                  MEM_TYPE_PAGE_SHARED, NULL);
1579
1580                 if (ret) {
1581                         netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1582                         goto cleanup2;
1583                 }
1584         }
1585
1586         /* Enable NAPI handler before init callbacks */
1587         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1588                        netvsc_poll, NAPI_POLL_WEIGHT);
1589
1590         /* Open the channel */
1591         device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1592         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1593                          netvsc_ring_bytes,  NULL, 0,
1594                          netvsc_channel_cb, net_device->chan_table);
1595
1596         if (ret != 0) {
1597                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1598                 goto cleanup;
1599         }
1600
1601         /* Channel is opened */
1602         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1603
1604         napi_enable(&net_device->chan_table[0].napi);
1605
1606         /* Connect with the NetVsp */
1607         ret = netvsc_connect_vsp(device, net_device, device_info);
1608         if (ret != 0) {
1609                 netdev_err(ndev,
1610                         "unable to connect to NetVSP - %d\n", ret);
1611                 goto close;
1612         }
1613
1614         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1615          * populated.
1616          */
1617         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1618
1619         return net_device;
1620
1621 close:
1622         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1623         napi_disable(&net_device->chan_table[0].napi);
1624
1625         /* Now, we can close the channel safely */
1626         vmbus_close(device->channel);
1627
1628 cleanup:
1629         netif_napi_del(&net_device->chan_table[0].napi);
1630
1631 cleanup2:
1632         free_netvsc_device(&net_device->rcu);
1633
1634         return ERR_PTR(ret);
1635 }