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