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