drivers/net/hyperv/netvsc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2009, Microsoft Corporation.
   4  *
   5  * Authors:
   6  *   Haiyang Zhang <haiyangz@microsoft.com>
   7  *   Hank Janssen  <hjanssen@microsoft.com>
   8  */
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/wait.h>
  14 #include <linux/mm.h>
  15 #include <linux/delay.h>
  16 #include <linux/io.h>
  17 #include <linux/slab.h>
  18 #include <linux/netdevice.h>
  19 #include <linux/if_ether.h>
  20 #include <linux/vmalloc.h>
  21 #include <linux/rtnetlink.h>
  22 #include <linux/prefetch.h>
  23
  24 #include <asm/sync_bitops.h>
  25 #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) {
 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                 net_device->recv_buf_gpadl_handle = 0;
 294         }
 295 }
 296
 297 static void netvsc_teardown_send_gpadl(struct hv_device *device,
 298                                        struct netvsc_device *net_device,
 299                                        struct net_device *ndev)
 300 {
 301         int ret;
 302
 303         if (net_device->send_buf_gpadl_handle) {
 304                 ret = vmbus_teardown_gpadl(device->channel,
 305                                            net_device->send_buf_gpadl_handle);
 306
 307                 /* If we failed here, we might as well return and have a leak
 308                  * rather than continue and a bugchk
 309                  */
 310                 if (ret != 0) {
 311                         netdev_err(ndev,
 312                                    "unable to teardown send buffer's gpadl\n");
 313                         return;
 314                 }
 315                 net_device->send_buf_gpadl_handle = 0;
 316         }
 317 }
 318
 319 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 320 {
 321         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 322         int node = cpu_to_node(nvchan->channel->target_cpu);
 323         size_t size;
 324
 325         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 326         nvchan->mrc.slots = vzalloc_node(size, node);
 327         if (!nvchan->mrc.slots)
 328                 nvchan->mrc.slots = vzalloc(size);
 329
 330         return nvchan->mrc.slots ? 0 : -ENOMEM;
 331 }
 332
 333 static int netvsc_init_buf(struct hv_device *device,
 334                            struct netvsc_device *net_device,
 335                            const struct netvsc_device_info *device_info)
 336 {
 337         struct nvsp_1_message_send_receive_buffer_complete *resp;
 338         struct net_device *ndev = hv_get_drvdata(device);
 339         struct nvsp_message *init_packet;
 340         unsigned int buf_size;
 341         size_t map_words;
 342         int i, ret = 0;
 343
 344         /* Get receive buffer area. */
 345         buf_size = device_info->recv_sections * device_info->recv_section_size;
 346         buf_size = roundup(buf_size, PAGE_SIZE);
 347
 348         /* Legacy hosts only allow smaller receive buffer */
 349         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 350                 buf_size = min_t(unsigned int, buf_size,
 351                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 352
 353         net_device->recv_buf = vzalloc(buf_size);
 354         if (!net_device->recv_buf) {
 355                 netdev_err(ndev,
 356                            "unable to allocate receive buffer of size %u\n",
 357                            buf_size);
 358                 ret = -ENOMEM;
 359                 goto cleanup;
 360         }
 361
 362         net_device->recv_buf_size = buf_size;
 363
 364         /*
 365          * Establish the gpadl handle for this buffer on this
 366          * channel.  Note: This call uses the vmbus connection rather
 367          * than the channel to establish the gpadl handle.
 368          */
 369         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 370                                     buf_size,
 371                                     &net_device->recv_buf_gpadl_handle);
 372         if (ret != 0) {
 373                 netdev_err(ndev,
 374                         "unable to establish receive buffer's gpadl\n");
 375                 goto cleanup;
 376         }
 377
 378         /* Notify the NetVsp of the gpadl handle */
 379         init_packet = &net_device->channel_init_pkt;
 380         memset(init_packet, 0, sizeof(struct nvsp_message));
 381         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 382         init_packet->msg.v1_msg.send_recv_buf.
 383                 gpadl_handle = net_device->recv_buf_gpadl_handle;
 384         init_packet->msg.v1_msg.
 385                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 386
 387         trace_nvsp_send(ndev, init_packet);
 388
 389         /* Send the gpadl notification request */
 390         ret = vmbus_sendpacket(device->channel, init_packet,
 391                                sizeof(struct nvsp_message),
 392                                (unsigned long)init_packet,
 393                                VM_PKT_DATA_INBAND,
 394                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 395         if (ret != 0) {
 396                 netdev_err(ndev,
 397                         "unable to send receive buffer's gpadl to netvsp\n");
 398                 goto cleanup;
 399         }
 400
 401         wait_for_completion(&net_device->channel_init_wait);
 402
 403         /* Check the response */
 404         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 405         if (resp->status != NVSP_STAT_SUCCESS) {
 406                 netdev_err(ndev,
 407                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 408                            resp->status);
 409                 ret = -EINVAL;
 410                 goto cleanup;
 411         }
 412
 413         /* Parse the response */
 414         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 415                    resp->num_sections, resp->sections[0].sub_alloc_size,
 416                    resp->sections[0].num_sub_allocs);
 417
 418         /* There should only be one section for the entire receive buffer */
 419         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 420                 ret = -EINVAL;
 421                 goto cleanup;
 422         }
 423
 424         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 425         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 426
 427         /* Ensure buffer will not overflow */
 428         if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
 429             (u64)net_device->recv_section_cnt > (u64)buf_size) {
 430                 netdev_err(ndev, "invalid recv_section_size %u\n",
 431                            net_device->recv_section_size);
 432                 ret = -EINVAL;
 433                 goto cleanup;
 434         }
 435
 436         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
 437                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
 438
 439                 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
 440                 if (nvchan->recv_buf == NULL) {
 441                         ret = -ENOMEM;
 442                         goto cleanup;
 443                 }
 444         }
 445
 446         /* Setup receive completion ring.
 447          * Add 1 to the recv_section_cnt because at least one entry in a
 448          * ring buffer has to be empty.
 449          */
 450         net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
 451         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 452         if (ret)
 453                 goto cleanup;
 454
 455         /* Now setup the send buffer. */
 456         buf_size = device_info->send_sections * device_info->send_section_size;
 457         buf_size = round_up(buf_size, PAGE_SIZE);
 458
 459         net_device->send_buf = vzalloc(buf_size);
 460         if (!net_device->send_buf) {
 461                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 462                            buf_size);
 463                 ret = -ENOMEM;
 464                 goto cleanup;
 465         }
 466
 467         /* Establish the gpadl handle for this buffer on this
 468          * channel.  Note: This call uses the vmbus connection rather
 469          * than the channel to establish the gpadl handle.
 470          */
 471         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 472                                     buf_size,
 473                                     &net_device->send_buf_gpadl_handle);
 474         if (ret != 0) {
 475                 netdev_err(ndev,
 476                            "unable to establish send buffer's gpadl\n");
 477                 goto cleanup;
 478         }
 479
 480         /* Notify the NetVsp of the gpadl handle */
 481         init_packet = &net_device->channel_init_pkt;
 482         memset(init_packet, 0, sizeof(struct nvsp_message));
 483         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 484         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 485                 net_device->send_buf_gpadl_handle;
 486         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 487
 488         trace_nvsp_send(ndev, init_packet);
 489
 490         /* Send the gpadl notification request */
 491         ret = vmbus_sendpacket(device->channel, init_packet,
 492                                sizeof(struct nvsp_message),
 493                                (unsigned long)init_packet,
 494                                VM_PKT_DATA_INBAND,
 495                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 496         if (ret != 0) {
 497                 netdev_err(ndev,
 498                            "unable to send send buffer's gpadl to netvsp\n");
 499                 goto cleanup;
 500         }
 501
 502         wait_for_completion(&net_device->channel_init_wait);
 503
 504         /* Check the response */
 505         if (init_packet->msg.v1_msg.
 506             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 507                 netdev_err(ndev, "Unable to complete send buffer "
 508                            "initialization with NetVsp - status %d\n",
 509                            init_packet->msg.v1_msg.
 510                            send_send_buf_complete.status);
 511                 ret = -EINVAL;
 512                 goto cleanup;
 513         }
 514
 515         /* Parse the response */
 516         net_device->send_section_size = init_packet->msg.
 517                                 v1_msg.send_send_buf_complete.section_size;
 518         if (net_device->send_section_size < NETVSC_MTU_MIN) {
 519                 netdev_err(ndev, "invalid send_section_size %u\n",
 520                            net_device->send_section_size);
 521                 ret = -EINVAL;
 522                 goto cleanup;
 523         }
 524
 525         /* Section count is simply the size divided by the section size. */
 526         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 527
 528         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 529                    net_device->send_section_size, net_device->send_section_cnt);
 530
 531         /* Setup state for managing the send buffer. */
 532         map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 533
 534         net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 535         if (net_device->send_section_map == NULL) {
 536                 ret = -ENOMEM;
 537                 goto cleanup;
 538         }
 539
 540         goto exit;
 541
 542 cleanup:
 543         netvsc_revoke_recv_buf(device, net_device, ndev);
 544         netvsc_revoke_send_buf(device, net_device, ndev);
 545         netvsc_teardown_recv_gpadl(device, net_device, ndev);
 546         netvsc_teardown_send_gpadl(device, net_device, ndev);
 547
 548 exit:
 549         return ret;
 550 }
 551
 552 /* Negotiate NVSP protocol version */
 553 static int negotiate_nvsp_ver(struct hv_device *device,
 554                               struct netvsc_device *net_device,
 555                               struct nvsp_message *init_packet,
 556                               u32 nvsp_ver)
 557 {
 558         struct net_device *ndev = hv_get_drvdata(device);
 559         int ret;
 560
 561         memset(init_packet, 0, sizeof(struct nvsp_message));
 562         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 563         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 564         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 565         trace_nvsp_send(ndev, init_packet);
 566
 567         /* Send the init request */
 568         ret = vmbus_sendpacket(device->channel, init_packet,
 569                                sizeof(struct nvsp_message),
 570                                (unsigned long)init_packet,
 571                                VM_PKT_DATA_INBAND,
 572                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 573
 574         if (ret != 0)
 575                 return ret;
 576
 577         wait_for_completion(&net_device->channel_init_wait);
 578
 579         if (init_packet->msg.init_msg.init_complete.status !=
 580             NVSP_STAT_SUCCESS)
 581                 return -EINVAL;
 582
 583         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 584                 return 0;
 585
 586         /* NVSPv2 or later: Send NDIS config */
 587         memset(init_packet, 0, sizeof(struct nvsp_message));
 588         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 589         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 590         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 591
 592         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 593                 if (hv_is_isolation_supported())
 594                         netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n");
 595                 else
 596                         init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 597
 598                 /* Teaming bit is needed to receive link speed updates */
 599                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 600         }
 601
 602         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
 603                 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
 604
 605         trace_nvsp_send(ndev, init_packet);
 606
 607         ret = vmbus_sendpacket(device->channel, init_packet,
 608                                 sizeof(struct nvsp_message),
 609                                 VMBUS_RQST_ID_NO_RESPONSE,
 610                                 VM_PKT_DATA_INBAND, 0);
 611
 612         return ret;
 613 }
 614
 615 static int netvsc_connect_vsp(struct hv_device *device,
 616                               struct netvsc_device *net_device,
 617                               const struct netvsc_device_info *device_info)
 618 {
 619         struct net_device *ndev = hv_get_drvdata(device);
 620         static const u32 ver_list[] = {
 621                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 622                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
 623                 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
 624         };
 625         struct nvsp_message *init_packet;
 626         int ndis_version, i, ret;
 627
 628         init_packet = &net_device->channel_init_pkt;
 629
 630         /* Negotiate the latest NVSP protocol supported */
 631         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 632                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 633                                        ver_list[i])  == 0) {
 634                         net_device->nvsp_version = ver_list[i];
 635                         break;
 636                 }
 637
 638         if (i < 0) {
 639                 ret = -EPROTO;
 640                 goto cleanup;
 641         }
 642
 643         if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) {
 644                 netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n",
 645                            net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61);
 646                 ret = -EPROTO;
 647                 goto cleanup;
 648         }
 649
 650         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 651
 652         /* Send the ndis version */
 653         memset(init_packet, 0, sizeof(struct nvsp_message));
 654
 655         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 656                 ndis_version = 0x00060001;
 657         else
 658                 ndis_version = 0x0006001e;
 659
 660         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 661         init_packet->msg.v1_msg.
 662                 send_ndis_ver.ndis_major_ver =
 663                                 (ndis_version & 0xFFFF0000) >> 16;
 664         init_packet->msg.v1_msg.
 665                 send_ndis_ver.ndis_minor_ver =
 666                                 ndis_version & 0xFFFF;
 667
 668         trace_nvsp_send(ndev, init_packet);
 669
 670         /* Send the init request */
 671         ret = vmbus_sendpacket(device->channel, init_packet,
 672                                 sizeof(struct nvsp_message),
 673                                 VMBUS_RQST_ID_NO_RESPONSE,
 674                                 VM_PKT_DATA_INBAND, 0);
 675         if (ret != 0)
 676                 goto cleanup;
 677
 678
 679         ret = netvsc_init_buf(device, net_device, device_info);
 680
 681 cleanup:
 682         return ret;
 683 }
 684
 685 /*
 686  * netvsc_device_remove - Callback when the root bus device is removed
 687  */
 688 void netvsc_device_remove(struct hv_device *device)
 689 {
 690         struct net_device *ndev = hv_get_drvdata(device);
 691         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 692         struct netvsc_device *net_device
 693                 = rtnl_dereference(net_device_ctx->nvdev);
 694         int i;
 695
 696         /*
 697          * Revoke receive buffer. If host is pre-Win2016 then tear down
 698          * receive buffer GPADL. Do the same for send buffer.
 699          */
 700         netvsc_revoke_recv_buf(device, net_device, ndev);
 701         if (vmbus_proto_version < VERSION_WIN10)
 702                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 703
 704         netvsc_revoke_send_buf(device, net_device, ndev);
 705         if (vmbus_proto_version < VERSION_WIN10)
 706                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 707
 708         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 709
 710         /* Disable NAPI and disassociate its context from the device. */
 711         for (i = 0; i < net_device->num_chn; i++) {
 712                 /* See also vmbus_reset_channel_cb(). */
 713                 napi_disable(&net_device->chan_table[i].napi);
 714                 netif_napi_del(&net_device->chan_table[i].napi);
 715         }
 716
 717         /*
 718          * At this point, no one should be accessing net_device
 719          * except in here
 720          */
 721         netdev_dbg(ndev, "net device safe to remove\n");
 722
 723         /* Now, we can close the channel safely */
 724         vmbus_close(device->channel);
 725
 726         /*
 727          * If host is Win2016 or higher then we do the GPADL tear down
 728          * here after VMBus is closed.
 729         */
 730         if (vmbus_proto_version >= VERSION_WIN10) {
 731                 netvsc_teardown_recv_gpadl(device, net_device, ndev);
 732                 netvsc_teardown_send_gpadl(device, net_device, ndev);
 733         }
 734
 735         /* Release all resources */
 736         free_netvsc_device_rcu(net_device);
 737 }
 738
 739 #define RING_AVAIL_PERCENT_HIWATER 20
 740 #define RING_AVAIL_PERCENT_LOWATER 10
 741
 742 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 743                                          u32 index)
 744 {
 745         sync_change_bit(index, net_device->send_section_map);
 746 }
 747
 748 static void netvsc_send_tx_complete(struct net_device *ndev,
 749                                     struct netvsc_device *net_device,
 750                                     struct vmbus_channel *channel,
 751                                     const struct vmpacket_descriptor *desc,
 752                                     int budget)
 753 {
 754         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 755         struct sk_buff *skb;
 756         u16 q_idx = 0;
 757         int queue_sends;
 758         u64 cmd_rqst;
 759
 760         cmd_rqst = channel->request_addr_callback(channel, (u64)desc->trans_id);
 761         if (cmd_rqst == VMBUS_RQST_ERROR) {
 762                 netdev_err(ndev, "Incorrect transaction id\n");
 763                 return;
 764         }
 765
 766         skb = (struct sk_buff *)(unsigned long)cmd_rqst;
 767
 768         /* Notify the layer above us */
 769         if (likely(skb)) {
 770                 const struct hv_netvsc_packet *packet
 771                         = (struct hv_netvsc_packet *)skb->cb;
 772                 u32 send_index = packet->send_buf_index;
 773                 struct netvsc_stats *tx_stats;
 774
 775                 if (send_index != NETVSC_INVALID_INDEX)
 776                         netvsc_free_send_slot(net_device, send_index);
 777                 q_idx = packet->q_idx;
 778
 779                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 780
 781                 u64_stats_update_begin(&tx_stats->syncp);
 782                 tx_stats->packets += packet->total_packets;
 783                 tx_stats->bytes += packet->total_bytes;
 784                 u64_stats_update_end(&tx_stats->syncp);
 785
 786                 napi_consume_skb(skb, budget);
 787         }
 788
 789         queue_sends =
 790                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 791
 792         if (unlikely(net_device->destroy)) {
 793                 if (queue_sends == 0)
 794                         wake_up(&net_device->wait_drain);
 795         } else {
 796                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 797
 798                 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
 799                     (hv_get_avail_to_write_percent(&channel->outbound) >
 800                      RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
 801                         netif_tx_wake_queue(txq);
 802                         ndev_ctx->eth_stats.wake_queue++;
 803                 }
 804         }
 805 }
 806
 807 static void netvsc_send_completion(struct net_device *ndev,
 808                                    struct netvsc_device *net_device,
 809                                    struct vmbus_channel *incoming_channel,
 810                                    const struct vmpacket_descriptor *desc,
 811                                    int budget)
 812 {
 813         const struct nvsp_message *nvsp_packet;
 814         u32 msglen = hv_pkt_datalen(desc);
 815         struct nvsp_message *pkt_rqst;
 816         u64 cmd_rqst;
 817
 818         /* First check if this is a VMBUS completion without data payload */
 819         if (!msglen) {
 820                 cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
 821                                                                    (u64)desc->trans_id);
 822                 if (cmd_rqst == VMBUS_RQST_ERROR) {
 823                         netdev_err(ndev, "Invalid transaction id\n");
 824                         return;
 825                 }
 826
 827                 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
 828                 switch (pkt_rqst->hdr.msg_type) {
 829                 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
 830                         complete(&net_device->channel_init_wait);
 831                         break;
 832
 833                 default:
 834                         netdev_err(ndev, "Unexpected VMBUS completion!!\n");
 835                 }
 836                 return;
 837         }
 838
 839         /* Ensure packet is big enough to read header fields */
 840         if (msglen < sizeof(struct nvsp_message_header)) {
 841                 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
 842                 return;
 843         }
 844
 845         nvsp_packet = hv_pkt_data(desc);
 846         switch (nvsp_packet->hdr.msg_type) {
 847         case NVSP_MSG_TYPE_INIT_COMPLETE:
 848                 if (msglen < sizeof(struct nvsp_message_header) +
 849                                 sizeof(struct nvsp_message_init_complete)) {
 850                         netdev_err(ndev, "nvsp_msg length too small: %u\n",
 851                                    msglen);
 852                         return;
 853                 }
 854                 fallthrough;
 855
 856         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 857                 if (msglen < sizeof(struct nvsp_message_header) +
 858                                 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
 859                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 860                                    msglen);
 861                         return;
 862                 }
 863                 fallthrough;
 864
 865         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 866                 if (msglen < sizeof(struct nvsp_message_header) +
 867                                 sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
 868                         netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
 869                                    msglen);
 870                         return;
 871                 }
 872                 fallthrough;
 873
 874         case NVSP_MSG5_TYPE_SUBCHANNEL:
 875                 if (msglen < sizeof(struct nvsp_message_header) +
 876                                 sizeof(struct nvsp_5_subchannel_complete)) {
 877                         netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
 878                                    msglen);
 879                         return;
 880                 }
 881                 /* Copy the response back */
 882                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 883                        sizeof(struct nvsp_message));
 884                 complete(&net_device->channel_init_wait);
 885                 break;
 886
 887         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 888                 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
 889                                         desc, budget);
 890                 break;
 891
 892         default:
 893                 netdev_err(ndev,
 894                            "Unknown send completion type %d received!!\n",
 895                            nvsp_packet->hdr.msg_type);
 896         }
 897 }
 898
 899 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 900 {
 901         unsigned long *map_addr = net_device->send_section_map;
 902         unsigned int i;
 903
 904         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 905                 if (sync_test_and_set_bit(i, map_addr) == 0)
 906                         return i;
 907         }
 908
 909         return NETVSC_INVALID_INDEX;
 910 }
 911
 912 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 913                                     unsigned int section_index,
 914                                     u32 pend_size,
 915                                     struct hv_netvsc_packet *packet,
 916                                     struct rndis_message *rndis_msg,
 917                                     struct hv_page_buffer *pb,
 918                                     bool xmit_more)
 919 {
 920         char *start = net_device->send_buf;
 921         char *dest = start + (section_index * net_device->send_section_size)
 922                      + pend_size;
 923         int i;
 924         u32 padding = 0;
 925         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 926                 packet->page_buf_cnt;
 927         u32 remain;
 928
 929         /* Add padding */
 930         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 931         if (xmit_more && remain) {
 932                 padding = net_device->pkt_align - remain;
 933                 rndis_msg->msg_len += padding;
 934                 packet->total_data_buflen += padding;
 935         }
 936
 937         for (i = 0; i < page_count; i++) {
 938                 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
 939                 u32 offset = pb[i].offset;
 940                 u32 len = pb[i].len;
 941
 942                 memcpy(dest, (src + offset), len);
 943                 dest += len;
 944         }
 945
 946         if (padding)
 947                 memset(dest, 0, padding);
 948 }
 949
 950 static inline int netvsc_send_pkt(
 951         struct hv_device *device,
 952         struct hv_netvsc_packet *packet,
 953         struct netvsc_device *net_device,
 954         struct hv_page_buffer *pb,
 955         struct sk_buff *skb)
 956 {
 957         struct nvsp_message nvmsg;
 958         struct nvsp_1_message_send_rndis_packet *rpkt =
 959                 &nvmsg.msg.v1_msg.send_rndis_pkt;
 960         struct netvsc_channel * const nvchan =
 961                 &net_device->chan_table[packet->q_idx];
 962         struct vmbus_channel *out_channel = nvchan->channel;
 963         struct net_device *ndev = hv_get_drvdata(device);
 964         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 965         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
 966         u64 req_id;
 967         int ret;
 968         u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
 969
 970         memset(&nvmsg, 0, sizeof(struct nvsp_message));
 971         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 972         if (skb)
 973                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
 974         else
 975                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
 976
 977         rpkt->send_buf_section_index = packet->send_buf_index;
 978         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
 979                 rpkt->send_buf_section_size = 0;
 980         else
 981                 rpkt->send_buf_section_size = packet->total_data_buflen;
 982
 983         req_id = (ulong)skb;
 984
 985         if (out_channel->rescind)
 986                 return -ENODEV;
 987
 988         trace_nvsp_send_pkt(ndev, out_channel, rpkt);
 989
 990         if (packet->page_buf_cnt) {
 991                 if (packet->cp_partial)
 992                         pb += packet->rmsg_pgcnt;
 993
 994                 ret = vmbus_sendpacket_pagebuffer(out_channel,
 995                                                   pb, packet->page_buf_cnt,
 996                                                   &nvmsg, sizeof(nvmsg),
 997                                                   req_id);
 998         } else {
 999                 ret = vmbus_sendpacket(out_channel,
1000                                        &nvmsg, sizeof(nvmsg),
1001                                        req_id, VM_PKT_DATA_INBAND,
1002                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1003         }
1004
1005         if (ret == 0) {
1006                 atomic_inc_return(&nvchan->queue_sends);
1007
1008                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
1009                         netif_tx_stop_queue(txq);
1010                         ndev_ctx->eth_stats.stop_queue++;
1011                 }
1012         } else if (ret == -EAGAIN) {
1013                 netif_tx_stop_queue(txq);
1014                 ndev_ctx->eth_stats.stop_queue++;
1015         } else {
1016                 netdev_err(ndev,
1017                            "Unable to send packet pages %u len %u, ret %d\n",
1018                            packet->page_buf_cnt, packet->total_data_buflen,
1019                            ret);
1020         }
1021
1022         if (netif_tx_queue_stopped(txq) &&
1023             atomic_read(&nvchan->queue_sends) < 1 &&
1024             !net_device->tx_disable) {
1025                 netif_tx_wake_queue(txq);
1026                 ndev_ctx->eth_stats.wake_queue++;
1027                 if (ret == -EAGAIN)
1028                         ret = -ENOSPC;
1029         }
1030
1031         return ret;
1032 }
1033
1034 /* Move packet out of multi send data (msd), and clear msd */
1035 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
1036                                 struct sk_buff **msd_skb,
1037                                 struct multi_send_data *msdp)
1038 {
1039         *msd_skb = msdp->skb;
1040         *msd_send = msdp->pkt;
1041         msdp->skb = NULL;
1042         msdp->pkt = NULL;
1043         msdp->count = 0;
1044 }
1045
1046 /* RCU already held by caller */
1047 /* Batching/bouncing logic is designed to attempt to optimize
1048  * performance.
1049  *
1050  * For small, non-LSO packets we copy the packet to a send buffer
1051  * which is pre-registered with the Hyper-V side. This enables the
1052  * hypervisor to avoid remapping the aperture to access the packet
1053  * descriptor and data.
1054  *
1055  * If we already started using a buffer and the netdev is transmitting
1056  * a burst of packets, keep on copying into the buffer until it is
1057  * full or we are done collecting a burst. If there is an existing
1058  * buffer with space for the RNDIS descriptor but not the packet, copy
1059  * the RNDIS descriptor to the buffer, keeping the packet in place.
1060  *
1061  * If we do batching and send more than one packet using a single
1062  * NetVSC message, free the SKBs of the packets copied, except for the
1063  * last packet. This is done to streamline the handling of the case
1064  * where the last packet only had the RNDIS descriptor copied to the
1065  * send buffer, with the data pointers included in the NetVSC message.
1066  */
1067 int netvsc_send(struct net_device *ndev,
1068                 struct hv_netvsc_packet *packet,
1069                 struct rndis_message *rndis_msg,
1070                 struct hv_page_buffer *pb,
1071                 struct sk_buff *skb,
1072                 bool xdp_tx)
1073 {
1074         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1075         struct netvsc_device *net_device
1076                 = rcu_dereference_bh(ndev_ctx->nvdev);
1077         struct hv_device *device = ndev_ctx->device_ctx;
1078         int ret = 0;
1079         struct netvsc_channel *nvchan;
1080         u32 pktlen = packet->total_data_buflen, msd_len = 0;
1081         unsigned int section_index = NETVSC_INVALID_INDEX;
1082         struct multi_send_data *msdp;
1083         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1084         struct sk_buff *msd_skb = NULL;
1085         bool try_batch, xmit_more;
1086
1087         /* If device is rescinded, return error and packet will get dropped. */
1088         if (unlikely(!net_device || net_device->destroy))
1089                 return -ENODEV;
1090
1091         nvchan = &net_device->chan_table[packet->q_idx];
1092         packet->send_buf_index = NETVSC_INVALID_INDEX;
1093         packet->cp_partial = false;
1094
1095         /* Send a control message or XDP packet directly without accessing
1096          * msd (Multi-Send Data) field which may be changed during data packet
1097          * processing.
1098          */
1099         if (!skb || xdp_tx)
1100                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
1101
1102         /* batch packets in send buffer if possible */
1103         msdp = &nvchan->msd;
1104         if (msdp->pkt)
1105                 msd_len = msdp->pkt->total_data_buflen;
1106
1107         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
1108         if (try_batch && msd_len + pktlen + net_device->pkt_align <
1109             net_device->send_section_size) {
1110                 section_index = msdp->pkt->send_buf_index;
1111
1112         } else if (try_batch && msd_len + packet->rmsg_size <
1113                    net_device->send_section_size) {
1114                 section_index = msdp->pkt->send_buf_index;
1115                 packet->cp_partial = true;
1116
1117         } else if (pktlen + net_device->pkt_align <
1118                    net_device->send_section_size) {
1119                 section_index = netvsc_get_next_send_section(net_device);
1120                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1121                         ++ndev_ctx->eth_stats.tx_send_full;
1122                 } else {
1123                         move_pkt_msd(&msd_send, &msd_skb, msdp);
1124                         msd_len = 0;
1125                 }
1126         }
1127
1128         /* Keep aggregating only if stack says more data is coming
1129          * and not doing mixed modes send and not flow blocked
1130          */
1131         xmit_more = netdev_xmit_more() &&
1132                 !packet->cp_partial &&
1133                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1134
1135         if (section_index != NETVSC_INVALID_INDEX) {
1136                 netvsc_copy_to_send_buf(net_device,
1137                                         section_index, msd_len,
1138                                         packet, rndis_msg, pb, xmit_more);
1139
1140                 packet->send_buf_index = section_index;
1141
1142                 if (packet->cp_partial) {
1143                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
1144                         packet->total_data_buflen = msd_len + packet->rmsg_size;
1145                 } else {
1146                         packet->page_buf_cnt = 0;
1147                         packet->total_data_buflen += msd_len;
1148                 }
1149
1150                 if (msdp->pkt) {
1151                         packet->total_packets += msdp->pkt->total_packets;
1152                         packet->total_bytes += msdp->pkt->total_bytes;
1153                 }
1154
1155                 if (msdp->skb)
1156                         dev_consume_skb_any(msdp->skb);
1157
1158                 if (xmit_more) {
1159                         msdp->skb = skb;
1160                         msdp->pkt = packet;
1161                         msdp->count++;
1162                 } else {
1163                         cur_send = packet;
1164                         msdp->skb = NULL;
1165                         msdp->pkt = NULL;
1166                         msdp->count = 0;
1167                 }
1168         } else {
1169                 move_pkt_msd(&msd_send, &msd_skb, msdp);
1170                 cur_send = packet;
1171         }
1172
1173         if (msd_send) {
1174                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1175                                             NULL, msd_skb);
1176
1177                 if (m_ret != 0) {
1178                         netvsc_free_send_slot(net_device,
1179                                               msd_send->send_buf_index);
1180                         dev_kfree_skb_any(msd_skb);
1181                 }
1182         }
1183
1184         if (cur_send)
1185                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1186
1187         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1188                 netvsc_free_send_slot(net_device, section_index);
1189
1190         return ret;
1191 }
1192
1193 /* Send pending recv completions */
1194 static int send_recv_completions(struct net_device *ndev,
1195                                  struct netvsc_device *nvdev,
1196                                  struct netvsc_channel *nvchan)
1197 {
1198         struct multi_recv_comp *mrc = &nvchan->mrc;
1199         struct recv_comp_msg {
1200                 struct nvsp_message_header hdr;
1201                 u32 status;
1202         }  __packed;
1203         struct recv_comp_msg msg = {
1204                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1205         };
1206         int ret;
1207
1208         while (mrc->first != mrc->next) {
1209                 const struct recv_comp_data *rcd
1210                         = mrc->slots + mrc->first;
1211
1212                 msg.status = rcd->status;
1213                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1214                                        rcd->tid, VM_PKT_COMP, 0);
1215                 if (unlikely(ret)) {
1216                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
1217
1218                         ++ndev_ctx->eth_stats.rx_comp_busy;
1219                         return ret;
1220                 }
1221
1222                 if (++mrc->first == nvdev->recv_completion_cnt)
1223                         mrc->first = 0;
1224         }
1225
1226         /* receive completion ring has been emptied */
1227         if (unlikely(nvdev->destroy))
1228                 wake_up(&nvdev->wait_drain);
1229
1230         return 0;
1231 }
1232
1233 /* Count how many receive completions are outstanding */
1234 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1235                                  const struct multi_recv_comp *mrc,
1236                                  u32 *filled, u32 *avail)
1237 {
1238         u32 count = nvdev->recv_completion_cnt;
1239
1240         if (mrc->next >= mrc->first)
1241                 *filled = mrc->next - mrc->first;
1242         else
1243                 *filled = (count - mrc->first) + mrc->next;
1244
1245         *avail = count - *filled - 1;
1246 }
1247
1248 /* Add receive complete to ring to send to host. */
1249 static void enq_receive_complete(struct net_device *ndev,
1250                                  struct netvsc_device *nvdev, u16 q_idx,
1251                                  u64 tid, u32 status)
1252 {
1253         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1254         struct multi_recv_comp *mrc = &nvchan->mrc;
1255         struct recv_comp_data *rcd;
1256         u32 filled, avail;
1257
1258         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1259
1260         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1261                 send_recv_completions(ndev, nvdev, nvchan);
1262                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1263         }
1264
1265         if (unlikely(!avail)) {
1266                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1267                            q_idx, tid);
1268                 return;
1269         }
1270
1271         rcd = mrc->slots + mrc->next;
1272         rcd->tid = tid;
1273         rcd->status = status;
1274
1275         if (++mrc->next == nvdev->recv_completion_cnt)
1276                 mrc->next = 0;
1277 }
1278
1279 static int netvsc_receive(struct net_device *ndev,
1280                           struct netvsc_device *net_device,
1281                           struct netvsc_channel *nvchan,
1282                           const struct vmpacket_descriptor *desc)
1283 {
1284         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1285         struct vmbus_channel *channel = nvchan->channel;
1286         const struct vmtransfer_page_packet_header *vmxferpage_packet
1287                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1288         const struct nvsp_message *nvsp = hv_pkt_data(desc);
1289         u32 msglen = hv_pkt_datalen(desc);
1290         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1291         char *recv_buf = net_device->recv_buf;
1292         u32 status = NVSP_STAT_SUCCESS;
1293         int i;
1294         int count = 0;
1295
1296         /* Ensure packet is big enough to read header fields */
1297         if (msglen < sizeof(struct nvsp_message_header)) {
1298                 netif_err(net_device_ctx, rx_err, ndev,
1299                           "invalid nvsp header, length too small: %u\n",
1300                           msglen);
1301                 return 0;
1302         }
1303
1304         /* Make sure this is a valid nvsp packet */
1305         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1306                 netif_err(net_device_ctx, rx_err, ndev,
1307                           "Unknown nvsp packet type received %u\n",
1308                           nvsp->hdr.msg_type);
1309                 return 0;
1310         }
1311
1312         /* Validate xfer page pkt header */
1313         if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1314                 netif_err(net_device_ctx, rx_err, ndev,
1315                           "Invalid xfer page pkt, offset too small: %u\n",
1316                           desc->offset8 << 3);
1317                 return 0;
1318         }
1319
1320         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1321                 netif_err(net_device_ctx, rx_err, ndev,
1322                           "Invalid xfer page set id - expecting %x got %x\n",
1323                           NETVSC_RECEIVE_BUFFER_ID,
1324                           vmxferpage_packet->xfer_pageset_id);
1325                 return 0;
1326         }
1327
1328         count = vmxferpage_packet->range_cnt;
1329
1330         /* Check count for a valid value */
1331         if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1332                 netif_err(net_device_ctx, rx_err, ndev,
1333                           "Range count is not valid: %d\n",
1334                           count);
1335                 return 0;
1336         }
1337
1338         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1339         for (i = 0; i < count; i++) {
1340                 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1341                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1342                 void *data;
1343                 int ret;
1344
1345                 if (unlikely(offset > net_device->recv_buf_size ||
1346                              buflen > net_device->recv_buf_size - offset)) {
1347                         nvchan->rsc.cnt = 0;
1348                         status = NVSP_STAT_FAIL;
1349                         netif_err(net_device_ctx, rx_err, ndev,
1350                                   "Packet offset:%u + len:%u too big\n",
1351                                   offset, buflen);
1352
1353                         continue;
1354                 }
1355
1356                 /* We're going to copy (sections of) the packet into nvchan->recv_buf;
1357                  * make sure that nvchan->recv_buf is large enough to hold the packet.
1358                  */
1359                 if (unlikely(buflen > net_device->recv_section_size)) {
1360                         nvchan->rsc.cnt = 0;
1361                         status = NVSP_STAT_FAIL;
1362                         netif_err(net_device_ctx, rx_err, ndev,
1363                                   "Packet too big: buflen=%u recv_section_size=%u\n",
1364                                   buflen, net_device->recv_section_size);
1365
1366                         continue;
1367                 }
1368
1369                 data = recv_buf + offset;
1370
1371                 nvchan->rsc.is_last = (i == count - 1);
1372
1373                 trace_rndis_recv(ndev, q_idx, data);
1374
1375                 /* Pass it to the upper layer */
1376                 ret = rndis_filter_receive(ndev, net_device,
1377                                            nvchan, data, buflen);
1378
1379                 if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1380                         /* Drop incomplete packet */
1381                         nvchan->rsc.cnt = 0;
1382                         status = NVSP_STAT_FAIL;
1383                 }
1384         }
1385
1386         enq_receive_complete(ndev, net_device, q_idx,
1387                              vmxferpage_packet->d.trans_id, status);
1388
1389         return count;
1390 }
1391
1392 static void netvsc_send_table(struct net_device *ndev,
1393                               struct netvsc_device *nvscdev,
1394                               const struct nvsp_message *nvmsg,
1395                               u32 msglen)
1396 {
1397         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1398         u32 count, offset, *tab;
1399         int i;
1400
1401         /* Ensure packet is big enough to read send_table fields */
1402         if (msglen < sizeof(struct nvsp_message_header) +
1403                      sizeof(struct nvsp_5_send_indirect_table)) {
1404                 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1405                 return;
1406         }
1407
1408         count = nvmsg->msg.v5_msg.send_table.count;
1409         offset = nvmsg->msg.v5_msg.send_table.offset;
1410
1411         if (count != VRSS_SEND_TAB_SIZE) {
1412                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1413                 return;
1414         }
1415
1416         /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1417          * wrong due to a host bug. So fix the offset here.
1418          */
1419         if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1420             msglen >= sizeof(struct nvsp_message_header) +
1421             sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1422                 offset = sizeof(struct nvsp_message_header) +
1423                          sizeof(union nvsp_6_message_uber);
1424
1425         /* Boundary check for all versions */
1426         if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1427                 netdev_err(ndev, "Received send-table offset too big:%u\n",
1428                            offset);
1429                 return;
1430         }
1431
1432         tab = (void *)nvmsg + offset;
1433
1434         for (i = 0; i < count; i++)
1435                 net_device_ctx->tx_table[i] = tab[i];
1436 }
1437
1438 static void netvsc_send_vf(struct net_device *ndev,
1439                            const struct nvsp_message *nvmsg,
1440                            u32 msglen)
1441 {
1442         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1443
1444         /* Ensure packet is big enough to read its fields */
1445         if (msglen < sizeof(struct nvsp_message_header) +
1446                      sizeof(struct nvsp_4_send_vf_association)) {
1447                 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1448                 return;
1449         }
1450
1451         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1452         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1453         netdev_info(ndev, "VF slot %u %s\n",
1454                     net_device_ctx->vf_serial,
1455                     net_device_ctx->vf_alloc ? "added" : "removed");
1456 }
1457
1458 static void netvsc_receive_inband(struct net_device *ndev,
1459                                   struct netvsc_device *nvscdev,
1460                                   const struct vmpacket_descriptor *desc)
1461 {
1462         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1463         u32 msglen = hv_pkt_datalen(desc);
1464
1465         /* Ensure packet is big enough to read header fields */
1466         if (msglen < sizeof(struct nvsp_message_header)) {
1467                 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1468                 return;
1469         }
1470
1471         switch (nvmsg->hdr.msg_type) {
1472         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1473                 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1474                 break;
1475
1476         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1477                 if (hv_is_isolation_supported())
1478                         netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n");
1479                 else
1480                         netvsc_send_vf(ndev, nvmsg, msglen);
1481                 break;
1482         }
1483 }
1484
1485 static int netvsc_process_raw_pkt(struct hv_device *device,
1486                                   struct netvsc_channel *nvchan,
1487                                   struct netvsc_device *net_device,
1488                                   struct net_device *ndev,
1489                                   const struct vmpacket_descriptor *desc,
1490                                   int budget)
1491 {
1492         struct vmbus_channel *channel = nvchan->channel;
1493         const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1494
1495         trace_nvsp_recv(ndev, channel, nvmsg);
1496
1497         switch (desc->type) {
1498         case VM_PKT_COMP:
1499                 netvsc_send_completion(ndev, net_device, channel, desc, budget);
1500                 break;
1501
1502         case VM_PKT_DATA_USING_XFER_PAGES:
1503                 return netvsc_receive(ndev, net_device, nvchan, desc);
1504                 break;
1505
1506         case VM_PKT_DATA_INBAND:
1507                 netvsc_receive_inband(ndev, net_device, desc);
1508                 break;
1509
1510         default:
1511                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1512                            desc->type, desc->trans_id);
1513                 break;
1514         }
1515
1516         return 0;
1517 }
1518
1519 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1520 {
1521         struct vmbus_channel *primary = channel->primary_channel;
1522
1523         return primary ? primary->device_obj : channel->device_obj;
1524 }
1525
1526 /* Network processing softirq
1527  * Process data in incoming ring buffer from host
1528  * Stops when ring is empty or budget is met or exceeded.
1529  */
1530 int netvsc_poll(struct napi_struct *napi, int budget)
1531 {
1532         struct netvsc_channel *nvchan
1533                 = container_of(napi, struct netvsc_channel, napi);
1534         struct netvsc_device *net_device = nvchan->net_device;
1535         struct vmbus_channel *channel = nvchan->channel;
1536         struct hv_device *device = netvsc_channel_to_device(channel);
1537         struct net_device *ndev = hv_get_drvdata(device);
1538         int work_done = 0;
1539         int ret;
1540
1541         /* If starting a new interval */
1542         if (!nvchan->desc)
1543                 nvchan->desc = hv_pkt_iter_first(channel);
1544
1545         while (nvchan->desc && work_done < budget) {
1546                 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1547                                                     ndev, nvchan->desc, budget);
1548                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1549         }
1550
1551         /* Send any pending receive completions */
1552         ret = send_recv_completions(ndev, net_device, nvchan);
1553
1554         /* If it did not exhaust NAPI budget this time
1555          *  and not doing busy poll
1556          * then re-enable host interrupts
1557          *  and reschedule if ring is not empty
1558          *   or sending receive completion failed.
1559          */
1560         if (work_done < budget &&
1561             napi_complete_done(napi, work_done) &&
1562             (ret || hv_end_read(&channel->inbound)) &&
1563             napi_schedule_prep(napi)) {
1564                 hv_begin_read(&channel->inbound);
1565                 __napi_schedule(napi);
1566         }
1567
1568         /* Driver may overshoot since multiple packets per descriptor */
1569         return min(work_done, budget);
1570 }
1571
1572 /* Call back when data is available in host ring buffer.
1573  * Processing is deferred until network softirq (NAPI)
1574  */
1575 void netvsc_channel_cb(void *context)
1576 {
1577         struct netvsc_channel *nvchan = context;
1578         struct vmbus_channel *channel = nvchan->channel;
1579         struct hv_ring_buffer_info *rbi = &channel->inbound;
1580
1581         /* preload first vmpacket descriptor */
1582         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1583
1584         if (napi_schedule_prep(&nvchan->napi)) {
1585                 /* disable interrupts from host */
1586                 hv_begin_read(rbi);
1587
1588                 __napi_schedule_irqoff(&nvchan->napi);
1589         }
1590 }
1591
1592 /*
1593  * netvsc_device_add - Callback when the device belonging to this
1594  * driver is added
1595  */
1596 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1597                                 const struct netvsc_device_info *device_info)
1598 {
1599         int i, ret = 0;
1600         struct netvsc_device *net_device;
1601         struct net_device *ndev = hv_get_drvdata(device);
1602         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1603
1604         net_device = alloc_net_device();
1605         if (!net_device)
1606                 return ERR_PTR(-ENOMEM);
1607
1608         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1609                 net_device_ctx->tx_table[i] = 0;
1610
1611         /* Because the device uses NAPI, all the interrupt batching and
1612          * control is done via Net softirq, not the channel handling
1613          */
1614         set_channel_read_mode(device->channel, HV_CALL_ISR);
1615
1616         /* If we're reopening the device we may have multiple queues, fill the
1617          * chn_table with the default channel to use it before subchannels are
1618          * opened.
1619          * Initialize the channel state before we open;
1620          * we can be interrupted as soon as we open the channel.
1621          */
1622
1623         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1624                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1625
1626                 nvchan->channel = device->channel;
1627                 nvchan->net_device = net_device;
1628                 u64_stats_init(&nvchan->tx_stats.syncp);
1629                 u64_stats_init(&nvchan->rx_stats.syncp);
1630
1631                 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1632
1633                 if (ret) {
1634                         netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1635                         goto cleanup2;
1636                 }
1637
1638                 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1639                                                  MEM_TYPE_PAGE_SHARED, NULL);
1640
1641                 if (ret) {
1642                         netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1643                         goto cleanup2;
1644                 }
1645         }
1646
1647         /* Enable NAPI handler before init callbacks */
1648         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1649                        netvsc_poll, NAPI_POLL_WEIGHT);
1650
1651         /* Open the channel */
1652         device->channel->next_request_id_callback = vmbus_next_request_id;
1653         device->channel->request_addr_callback = vmbus_request_addr;
1654         device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1655         device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE;
1656
1657         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1658                          netvsc_ring_bytes,  NULL, 0,
1659                          netvsc_channel_cb, net_device->chan_table);
1660
1661         if (ret != 0) {
1662                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1663                 goto cleanup;
1664         }
1665
1666         /* Channel is opened */
1667         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1668
1669         napi_enable(&net_device->chan_table[0].napi);
1670
1671         /* Connect with the NetVsp */
1672         ret = netvsc_connect_vsp(device, net_device, device_info);
1673         if (ret != 0) {
1674                 netdev_err(ndev,
1675                         "unable to connect to NetVSP - %d\n", ret);
1676                 goto close;
1677         }
1678
1679         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1680          * populated.
1681          */
1682         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1683
1684         return net_device;
1685
1686 close:
1687         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1688         napi_disable(&net_device->chan_table[0].napi);
1689
1690         /* Now, we can close the channel safely */
1691         vmbus_close(device->channel);
1692
1693 cleanup:
1694         netif_napi_del(&net_device->chan_table[0].napi);
1695
1696 cleanup2:
1697         free_netvsc_device(&net_device->rcu);
1698
1699         return ERR_PTR(ret);
1700 }