drivers/net/hyperv/netvsc.c

   1 /*
   2  * Copyright (c) 2009, Microsoft Corporation.
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  * You should have received a copy of the GNU General Public License along with
  14  * this program; if not, see <http://www.gnu.org/licenses/>.
  15  *
  16  * Authors:
  17  *   Haiyang Zhang <haiyangz@microsoft.com>
  18  *   Hank Janssen  <hjanssen@microsoft.com>
  19  */
  20 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  21
  22 #include <linux/kernel.h>
  23 #include <linux/sched.h>
  24 #include <linux/wait.h>
  25 #include <linux/mm.h>
  26 #include <linux/delay.h>
  27 #include <linux/io.h>
  28 #include <linux/slab.h>
  29 #include <linux/netdevice.h>
  30 #include <linux/if_ether.h>
  31 #include <linux/vmalloc.h>
  32 #include <linux/rtnetlink.h>
  33 #include <linux/prefetch.h>
  34 #include <linux/reciprocal_div.h>
  35
  36 #include <asm/sync_bitops.h>
  37
  38 #include "hyperv_net.h"
  39
  40 /*
  41  * Switch the data path from the synthetic interface to the VF
  42  * interface.
  43  */
  44 void netvsc_switch_datapath(struct net_device *ndev, bool vf)
  45 {
  46         struct net_device_context *net_device_ctx = netdev_priv(ndev);
  47         struct hv_device *dev = net_device_ctx->device_ctx;
  48         struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
  49         struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
  50
  51         memset(init_pkt, 0, sizeof(struct nvsp_message));
  52         init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
  53         if (vf)
  54                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  55                         NVSP_DATAPATH_VF;
  56         else
  57                 init_pkt->msg.v4_msg.active_dp.active_datapath =
  58                         NVSP_DATAPATH_SYNTHETIC;
  59
  60         vmbus_sendpacket(dev->channel, init_pkt,
  61                                sizeof(struct nvsp_message),
  62                                (unsigned long)init_pkt,
  63                                VM_PKT_DATA_INBAND, 0);
  64 }
  65
  66 static struct netvsc_device *alloc_net_device(void)
  67 {
  68         struct netvsc_device *net_device;
  69
  70         net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
  71         if (!net_device)
  72                 return NULL;
  73
  74         init_waitqueue_head(&net_device->wait_drain);
  75         net_device->destroy = false;
  76
  77         net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
  78         net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
  79
  80         init_completion(&net_device->channel_init_wait);
  81         init_waitqueue_head(&net_device->subchan_open);
  82         INIT_WORK(&net_device->subchan_work, rndis_set_subchannel);
  83
  84         return net_device;
  85 }
  86
  87 static void free_netvsc_device(struct rcu_head *head)
  88 {
  89         struct netvsc_device *nvdev
  90                 = container_of(head, struct netvsc_device, rcu);
  91         int i;
  92
  93         kfree(nvdev->extension);
  94         vfree(nvdev->recv_buf);
  95         vfree(nvdev->send_buf);
  96         kfree(nvdev->send_section_map);
  97
  98         for (i = 0; i < VRSS_CHANNEL_MAX; i++)
  99                 vfree(nvdev->chan_table[i].mrc.slots);
 100
 101         kfree(nvdev);
 102 }
 103
 104 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
 105 {
 106         call_rcu(&nvdev->rcu, free_netvsc_device);
 107 }
 108
 109 static void netvsc_revoke_buf(struct hv_device *device,
 110                               struct netvsc_device *net_device)
 111 {
 112         struct nvsp_message *revoke_packet;
 113         struct net_device *ndev = hv_get_drvdata(device);
 114         int ret;
 115
 116         /*
 117          * If we got a section count, it means we received a
 118          * SendReceiveBufferComplete msg (ie sent
 119          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 120          * to send a revoke msg here
 121          */
 122         if (net_device->recv_section_cnt) {
 123                 /* Send the revoke receive buffer */
 124                 revoke_packet = &net_device->revoke_packet;
 125                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 126
 127                 revoke_packet->hdr.msg_type =
 128                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 129                 revoke_packet->msg.v1_msg.
 130                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 131
 132                 ret = vmbus_sendpacket(device->channel,
 133                                        revoke_packet,
 134                                        sizeof(struct nvsp_message),
 135                                        (unsigned long)revoke_packet,
 136                                        VM_PKT_DATA_INBAND, 0);
 137                 /* If the failure is because the channel is rescinded;
 138                  * ignore the failure since we cannot send on a rescinded
 139                  * channel. This would allow us to properly cleanup
 140                  * even when the channel is rescinded.
 141                  */
 142                 if (device->channel->rescind)
 143                         ret = 0;
 144                 /*
 145                  * If we failed here, we might as well return and
 146                  * have a leak rather than continue and a bugchk
 147                  */
 148                 if (ret != 0) {
 149                         netdev_err(ndev, "unable to send "
 150                                 "revoke receive buffer to netvsp\n");
 151                         return;
 152                 }
 153                 net_device->recv_section_cnt = 0;
 154         }
 155
 156         /* Deal with the send buffer we may have setup.
 157          * If we got a  send section size, it means we received a
 158          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 159          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 160          * to send a revoke msg here
 161          */
 162         if (net_device->send_section_cnt) {
 163                 /* Send the revoke receive buffer */
 164                 revoke_packet = &net_device->revoke_packet;
 165                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 166
 167                 revoke_packet->hdr.msg_type =
 168                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 169                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 170                         NETVSC_SEND_BUFFER_ID;
 171
 172                 ret = vmbus_sendpacket(device->channel,
 173                                        revoke_packet,
 174                                        sizeof(struct nvsp_message),
 175                                        (unsigned long)revoke_packet,
 176                                        VM_PKT_DATA_INBAND, 0);
 177
 178                 /* If the failure is because the channel is rescinded;
 179                  * ignore the failure since we cannot send on a rescinded
 180                  * channel. This would allow us to properly cleanup
 181                  * even when the channel is rescinded.
 182                  */
 183                 if (device->channel->rescind)
 184                         ret = 0;
 185
 186                 /* If we failed here, we might as well return and
 187                  * have a leak rather than continue and a bugchk
 188                  */
 189                 if (ret != 0) {
 190                         netdev_err(ndev, "unable to send "
 191                                    "revoke send buffer to netvsp\n");
 192                         return;
 193                 }
 194                 net_device->send_section_cnt = 0;
 195         }
 196 }
 197
 198 static void netvsc_teardown_gpadl(struct hv_device *device,
 199                                   struct netvsc_device *net_device)
 200 {
 201         struct net_device *ndev = hv_get_drvdata(device);
 202         int ret;
 203
 204         if (net_device->recv_buf_gpadl_handle) {
 205                 ret = vmbus_teardown_gpadl(device->channel,
 206                                            net_device->recv_buf_gpadl_handle);
 207
 208                 /* If we failed here, we might as well return and have a leak
 209                  * rather than continue and a bugchk
 210                  */
 211                 if (ret != 0) {
 212                         netdev_err(ndev,
 213                                    "unable to teardown receive buffer's gpadl\n");
 214                         return;
 215                 }
 216                 net_device->recv_buf_gpadl_handle = 0;
 217         }
 218
 219         if (net_device->send_buf_gpadl_handle) {
 220                 ret = vmbus_teardown_gpadl(device->channel,
 221                                            net_device->send_buf_gpadl_handle);
 222
 223                 /* If we failed here, we might as well return and have a leak
 224                  * rather than continue and a bugchk
 225                  */
 226                 if (ret != 0) {
 227                         netdev_err(ndev,
 228                                    "unable to teardown send buffer's gpadl\n");
 229                         return;
 230                 }
 231                 net_device->send_buf_gpadl_handle = 0;
 232         }
 233 }
 234
 235 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 236 {
 237         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 238         int node = cpu_to_node(nvchan->channel->target_cpu);
 239         size_t size;
 240
 241         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 242         nvchan->mrc.slots = vzalloc_node(size, node);
 243         if (!nvchan->mrc.slots)
 244                 nvchan->mrc.slots = vzalloc(size);
 245
 246         return nvchan->mrc.slots ? 0 : -ENOMEM;
 247 }
 248
 249 static int netvsc_init_buf(struct hv_device *device,
 250                            struct netvsc_device *net_device,
 251                            const struct netvsc_device_info *device_info)
 252 {
 253         struct nvsp_1_message_send_receive_buffer_complete *resp;
 254         struct net_device *ndev = hv_get_drvdata(device);
 255         struct nvsp_message *init_packet;
 256         unsigned int buf_size;
 257         size_t map_words;
 258         int ret = 0;
 259
 260         /* Get receive buffer area. */
 261         buf_size = device_info->recv_sections * device_info->recv_section_size;
 262         buf_size = roundup(buf_size, PAGE_SIZE);
 263
 264         /* Legacy hosts only allow smaller receive buffer */
 265         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 266                 buf_size = min_t(unsigned int, buf_size,
 267                                  NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
 268
 269         net_device->recv_buf = vzalloc(buf_size);
 270         if (!net_device->recv_buf) {
 271                 netdev_err(ndev,
 272                            "unable to allocate receive buffer of size %u\n",
 273                            buf_size);
 274                 ret = -ENOMEM;
 275                 goto cleanup;
 276         }
 277
 278         /*
 279          * Establish the gpadl handle for this buffer on this
 280          * channel.  Note: This call uses the vmbus connection rather
 281          * than the channel to establish the gpadl handle.
 282          */
 283         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 284                                     buf_size,
 285                                     &net_device->recv_buf_gpadl_handle);
 286         if (ret != 0) {
 287                 netdev_err(ndev,
 288                         "unable to establish receive buffer's gpadl\n");
 289                 goto cleanup;
 290         }
 291
 292         /* Notify the NetVsp of the gpadl handle */
 293         init_packet = &net_device->channel_init_pkt;
 294         memset(init_packet, 0, sizeof(struct nvsp_message));
 295         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 296         init_packet->msg.v1_msg.send_recv_buf.
 297                 gpadl_handle = net_device->recv_buf_gpadl_handle;
 298         init_packet->msg.v1_msg.
 299                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 300
 301         /* Send the gpadl notification request */
 302         ret = vmbus_sendpacket(device->channel, init_packet,
 303                                sizeof(struct nvsp_message),
 304                                (unsigned long)init_packet,
 305                                VM_PKT_DATA_INBAND,
 306                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 307         if (ret != 0) {
 308                 netdev_err(ndev,
 309                         "unable to send receive buffer's gpadl to netvsp\n");
 310                 goto cleanup;
 311         }
 312
 313         wait_for_completion(&net_device->channel_init_wait);
 314
 315         /* Check the response */
 316         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 317         if (resp->status != NVSP_STAT_SUCCESS) {
 318                 netdev_err(ndev,
 319                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 320                            resp->status);
 321                 ret = -EINVAL;
 322                 goto cleanup;
 323         }
 324
 325         /* Parse the response */
 326         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 327                    resp->num_sections, resp->sections[0].sub_alloc_size,
 328                    resp->sections[0].num_sub_allocs);
 329
 330         /* There should only be one section for the entire receive buffer */
 331         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 332                 ret = -EINVAL;
 333                 goto cleanup;
 334         }
 335
 336         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 337         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 338
 339         /* Setup receive completion ring */
 340         net_device->recv_completion_cnt
 341                 = round_up(net_device->recv_section_cnt + 1,
 342                            PAGE_SIZE / sizeof(u64));
 343         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 344         if (ret)
 345                 goto cleanup;
 346
 347         /* Now setup the send buffer. */
 348         buf_size = device_info->send_sections * device_info->send_section_size;
 349         buf_size = round_up(buf_size, PAGE_SIZE);
 350
 351         net_device->send_buf = vzalloc(buf_size);
 352         if (!net_device->send_buf) {
 353                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 354                            buf_size);
 355                 ret = -ENOMEM;
 356                 goto cleanup;
 357         }
 358
 359         /* Establish the gpadl handle for this buffer on this
 360          * channel.  Note: This call uses the vmbus connection rather
 361          * than the channel to establish the gpadl handle.
 362          */
 363         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 364                                     buf_size,
 365                                     &net_device->send_buf_gpadl_handle);
 366         if (ret != 0) {
 367                 netdev_err(ndev,
 368                            "unable to establish send buffer's gpadl\n");
 369                 goto cleanup;
 370         }
 371
 372         /* Notify the NetVsp of the gpadl handle */
 373         init_packet = &net_device->channel_init_pkt;
 374         memset(init_packet, 0, sizeof(struct nvsp_message));
 375         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 376         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 377                 net_device->send_buf_gpadl_handle;
 378         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 379
 380         /* Send the gpadl notification request */
 381         ret = vmbus_sendpacket(device->channel, init_packet,
 382                                sizeof(struct nvsp_message),
 383                                (unsigned long)init_packet,
 384                                VM_PKT_DATA_INBAND,
 385                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 386         if (ret != 0) {
 387                 netdev_err(ndev,
 388                            "unable to send send buffer's gpadl to netvsp\n");
 389                 goto cleanup;
 390         }
 391
 392         wait_for_completion(&net_device->channel_init_wait);
 393
 394         /* Check the response */
 395         if (init_packet->msg.v1_msg.
 396             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 397                 netdev_err(ndev, "Unable to complete send buffer "
 398                            "initialization with NetVsp - status %d\n",
 399                            init_packet->msg.v1_msg.
 400                            send_send_buf_complete.status);
 401                 ret = -EINVAL;
 402                 goto cleanup;
 403         }
 404
 405         /* Parse the response */
 406         net_device->send_section_size = init_packet->msg.
 407                                 v1_msg.send_send_buf_complete.section_size;
 408
 409         /* Section count is simply the size divided by the section size. */
 410         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 411
 412         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 413                    net_device->send_section_size, net_device->send_section_cnt);
 414
 415         /* Setup state for managing the send buffer. */
 416         map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 417
 418         net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 419         if (net_device->send_section_map == NULL) {
 420                 ret = -ENOMEM;
 421                 goto cleanup;
 422         }
 423
 424         goto exit;
 425
 426 cleanup:
 427         netvsc_revoke_buf(device, net_device);
 428         netvsc_teardown_gpadl(device, net_device);
 429
 430 exit:
 431         return ret;
 432 }
 433
 434 /* Negotiate NVSP protocol version */
 435 static int negotiate_nvsp_ver(struct hv_device *device,
 436                               struct netvsc_device *net_device,
 437                               struct nvsp_message *init_packet,
 438                               u32 nvsp_ver)
 439 {
 440         struct net_device *ndev = hv_get_drvdata(device);
 441         int ret;
 442
 443         memset(init_packet, 0, sizeof(struct nvsp_message));
 444         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 445         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 446         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 447
 448         /* Send the init request */
 449         ret = vmbus_sendpacket(device->channel, init_packet,
 450                                sizeof(struct nvsp_message),
 451                                (unsigned long)init_packet,
 452                                VM_PKT_DATA_INBAND,
 453                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 454
 455         if (ret != 0)
 456                 return ret;
 457
 458         wait_for_completion(&net_device->channel_init_wait);
 459
 460         if (init_packet->msg.init_msg.init_complete.status !=
 461             NVSP_STAT_SUCCESS)
 462                 return -EINVAL;
 463
 464         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 465                 return 0;
 466
 467         /* NVSPv2 or later: Send NDIS config */
 468         memset(init_packet, 0, sizeof(struct nvsp_message));
 469         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 470         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 471         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 472
 473         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 474                 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 475
 476                 /* Teaming bit is needed to receive link speed updates */
 477                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 478         }
 479
 480         ret = vmbus_sendpacket(device->channel, init_packet,
 481                                 sizeof(struct nvsp_message),
 482                                 (unsigned long)init_packet,
 483                                 VM_PKT_DATA_INBAND, 0);
 484
 485         return ret;
 486 }
 487
 488 static int netvsc_connect_vsp(struct hv_device *device,
 489                               struct netvsc_device *net_device,
 490                               const struct netvsc_device_info *device_info)
 491 {
 492         static const u32 ver_list[] = {
 493                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 494                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5
 495         };
 496         struct nvsp_message *init_packet;
 497         int ndis_version, i, ret;
 498
 499         init_packet = &net_device->channel_init_pkt;
 500
 501         /* Negotiate the latest NVSP protocol supported */
 502         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 503                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 504                                        ver_list[i])  == 0) {
 505                         net_device->nvsp_version = ver_list[i];
 506                         break;
 507                 }
 508
 509         if (i < 0) {
 510                 ret = -EPROTO;
 511                 goto cleanup;
 512         }
 513
 514         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 515
 516         /* Send the ndis version */
 517         memset(init_packet, 0, sizeof(struct nvsp_message));
 518
 519         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 520                 ndis_version = 0x00060001;
 521         else
 522                 ndis_version = 0x0006001e;
 523
 524         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 525         init_packet->msg.v1_msg.
 526                 send_ndis_ver.ndis_major_ver =
 527                                 (ndis_version & 0xFFFF0000) >> 16;
 528         init_packet->msg.v1_msg.
 529                 send_ndis_ver.ndis_minor_ver =
 530                                 ndis_version & 0xFFFF;
 531
 532         /* Send the init request */
 533         ret = vmbus_sendpacket(device->channel, init_packet,
 534                                 sizeof(struct nvsp_message),
 535                                 (unsigned long)init_packet,
 536                                 VM_PKT_DATA_INBAND, 0);
 537         if (ret != 0)
 538                 goto cleanup;
 539
 540
 541         ret = netvsc_init_buf(device, net_device, device_info);
 542
 543 cleanup:
 544         return ret;
 545 }
 546
 547 /*
 548  * netvsc_device_remove - Callback when the root bus device is removed
 549  */
 550 void netvsc_device_remove(struct hv_device *device)
 551 {
 552         struct net_device *ndev = hv_get_drvdata(device);
 553         struct net_device_context *net_device_ctx = netdev_priv(ndev);
 554         struct netvsc_device *net_device
 555                 = rtnl_dereference(net_device_ctx->nvdev);
 556         int i;
 557
 558         netvsc_revoke_buf(device, net_device);
 559
 560         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 561
 562         /* And disassociate NAPI context from device */
 563         for (i = 0; i < net_device->num_chn; i++)
 564                 netif_napi_del(&net_device->chan_table[i].napi);
 565
 566         /*
 567          * At this point, no one should be accessing net_device
 568          * except in here
 569          */
 570         netdev_dbg(ndev, "net device safe to remove\n");
 571
 572         /* older versions require that buffer be revoked before close */
 573         if (net_device->nvsp_version < NVSP_PROTOCOL_VERSION_4)
 574                 netvsc_teardown_gpadl(device, net_device);
 575
 576         /* Now, we can close the channel safely */
 577         vmbus_close(device->channel);
 578
 579         if (net_device->nvsp_version >= NVSP_PROTOCOL_VERSION_4)
 580                 netvsc_teardown_gpadl(device, net_device);
 581
 582         /* Release all resources */
 583         free_netvsc_device_rcu(net_device);
 584 }
 585
 586 #define RING_AVAIL_PERCENT_HIWATER 20
 587 #define RING_AVAIL_PERCENT_LOWATER 10
 588
 589 /*
 590  * Get the percentage of available bytes to write in the ring.
 591  * The return value is in range from 0 to 100.
 592  */
 593 static u32 hv_ringbuf_avail_percent(const struct hv_ring_buffer_info *ring_info)
 594 {
 595         u32 avail_write = hv_get_bytes_to_write(ring_info);
 596
 597         return reciprocal_divide(avail_write  * 100, netvsc_ring_reciprocal);
 598 }
 599
 600 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 601                                          u32 index)
 602 {
 603         sync_change_bit(index, net_device->send_section_map);
 604 }
 605
 606 static void netvsc_send_tx_complete(struct netvsc_device *net_device,
 607                                     struct vmbus_channel *incoming_channel,
 608                                     struct hv_device *device,
 609                                     const struct vmpacket_descriptor *desc,
 610                                     int budget)
 611 {
 612         struct sk_buff *skb = (struct sk_buff *)(unsigned long)desc->trans_id;
 613         struct net_device *ndev = hv_get_drvdata(device);
 614         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 615         struct vmbus_channel *channel = device->channel;
 616         u16 q_idx = 0;
 617         int queue_sends;
 618
 619         /* Notify the layer above us */
 620         if (likely(skb)) {
 621                 const struct hv_netvsc_packet *packet
 622                         = (struct hv_netvsc_packet *)skb->cb;
 623                 u32 send_index = packet->send_buf_index;
 624                 struct netvsc_stats *tx_stats;
 625
 626                 if (send_index != NETVSC_INVALID_INDEX)
 627                         netvsc_free_send_slot(net_device, send_index);
 628                 q_idx = packet->q_idx;
 629                 channel = incoming_channel;
 630
 631                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 632
 633                 u64_stats_update_begin(&tx_stats->syncp);
 634                 tx_stats->packets += packet->total_packets;
 635                 tx_stats->bytes += packet->total_bytes;
 636                 u64_stats_update_end(&tx_stats->syncp);
 637
 638                 napi_consume_skb(skb, budget);
 639         }
 640
 641         queue_sends =
 642                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 643
 644         if (unlikely(net_device->destroy)) {
 645                 if (queue_sends == 0)
 646                         wake_up(&net_device->wait_drain);
 647         } else {
 648                 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
 649
 650                 if (netif_tx_queue_stopped(txq) &&
 651                     (hv_ringbuf_avail_percent(&channel->outbound) > RING_AVAIL_PERCENT_HIWATER ||
 652                      queue_sends < 1)) {
 653                         netif_tx_wake_queue(txq);
 654                         ndev_ctx->eth_stats.wake_queue++;
 655                 }
 656         }
 657 }
 658
 659 static void netvsc_send_completion(struct netvsc_device *net_device,
 660                                    struct vmbus_channel *incoming_channel,
 661                                    struct hv_device *device,
 662                                    const struct vmpacket_descriptor *desc,
 663                                    int budget)
 664 {
 665         struct nvsp_message *nvsp_packet = hv_pkt_data(desc);
 666         struct net_device *ndev = hv_get_drvdata(device);
 667
 668         switch (nvsp_packet->hdr.msg_type) {
 669         case NVSP_MSG_TYPE_INIT_COMPLETE:
 670         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 671         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 672         case NVSP_MSG5_TYPE_SUBCHANNEL:
 673                 /* Copy the response back */
 674                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 675                        sizeof(struct nvsp_message));
 676                 complete(&net_device->channel_init_wait);
 677                 break;
 678
 679         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 680                 netvsc_send_tx_complete(net_device, incoming_channel,
 681                                         device, desc, budget);
 682                 break;
 683
 684         default:
 685                 netdev_err(ndev,
 686                            "Unknown send completion type %d received!!\n",
 687                            nvsp_packet->hdr.msg_type);
 688         }
 689 }
 690
 691 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 692 {
 693         unsigned long *map_addr = net_device->send_section_map;
 694         unsigned int i;
 695
 696         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 697                 if (sync_test_and_set_bit(i, map_addr) == 0)
 698                         return i;
 699         }
 700
 701         return NETVSC_INVALID_INDEX;
 702 }
 703
 704 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 705                                     unsigned int section_index,
 706                                     u32 pend_size,
 707                                     struct hv_netvsc_packet *packet,
 708                                     struct rndis_message *rndis_msg,
 709                                     struct hv_page_buffer *pb,
 710                                     bool xmit_more)
 711 {
 712         char *start = net_device->send_buf;
 713         char *dest = start + (section_index * net_device->send_section_size)
 714                      + pend_size;
 715         int i;
 716         u32 padding = 0;
 717         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 718                 packet->page_buf_cnt;
 719         u32 remain;
 720
 721         /* Add padding */
 722         remain = packet->total_data_buflen & (net_device->pkt_align - 1);
 723         if (xmit_more && remain) {
 724                 padding = net_device->pkt_align - remain;
 725                 rndis_msg->msg_len += padding;
 726                 packet->total_data_buflen += padding;
 727         }
 728
 729         for (i = 0; i < page_count; i++) {
 730                 char *src = phys_to_virt(pb[i].pfn << PAGE_SHIFT);
 731                 u32 offset = pb[i].offset;
 732                 u32 len = pb[i].len;
 733
 734                 memcpy(dest, (src + offset), len);
 735                 dest += len;
 736         }
 737
 738         if (padding)
 739                 memset(dest, 0, padding);
 740 }
 741
 742 static inline int netvsc_send_pkt(
 743         struct hv_device *device,
 744         struct hv_netvsc_packet *packet,
 745         struct netvsc_device *net_device,
 746         struct hv_page_buffer *pb,
 747         struct sk_buff *skb)
 748 {
 749         struct nvsp_message nvmsg;
 750         struct nvsp_1_message_send_rndis_packet * const rpkt =
 751                 &nvmsg.msg.v1_msg.send_rndis_pkt;
 752         struct netvsc_channel * const nvchan =
 753                 &net_device->chan_table[packet->q_idx];
 754         struct vmbus_channel *out_channel = nvchan->channel;
 755         struct net_device *ndev = hv_get_drvdata(device);
 756         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 757         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
 758         u64 req_id;
 759         int ret;
 760         u32 ring_avail = hv_ringbuf_avail_percent(&out_channel->outbound);
 761
 762         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 763         if (skb)
 764                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
 765         else
 766                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
 767
 768         rpkt->send_buf_section_index = packet->send_buf_index;
 769         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
 770                 rpkt->send_buf_section_size = 0;
 771         else
 772                 rpkt->send_buf_section_size = packet->total_data_buflen;
 773
 774         req_id = (ulong)skb;
 775
 776         if (out_channel->rescind)
 777                 return -ENODEV;
 778
 779         if (packet->page_buf_cnt) {
 780                 if (packet->cp_partial)
 781                         pb += packet->rmsg_pgcnt;
 782
 783                 ret = vmbus_sendpacket_pagebuffer(out_channel,
 784                                                   pb, packet->page_buf_cnt,
 785                                                   &nvmsg, sizeof(nvmsg),
 786                                                   req_id);
 787         } else {
 788                 ret = vmbus_sendpacket(out_channel,
 789                                        &nvmsg, sizeof(nvmsg),
 790                                        req_id, VM_PKT_DATA_INBAND,
 791                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 792         }
 793
 794         if (ret == 0) {
 795                 atomic_inc_return(&nvchan->queue_sends);
 796
 797                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
 798                         netif_tx_stop_queue(txq);
 799                         ndev_ctx->eth_stats.stop_queue++;
 800                 }
 801         } else if (ret == -EAGAIN) {
 802                 netif_tx_stop_queue(txq);
 803                 ndev_ctx->eth_stats.stop_queue++;
 804                 if (atomic_read(&nvchan->queue_sends) < 1) {
 805                         netif_tx_wake_queue(txq);
 806                         ndev_ctx->eth_stats.wake_queue++;
 807                         ret = -ENOSPC;
 808                 }
 809         } else {
 810                 netdev_err(ndev,
 811                            "Unable to send packet pages %u len %u, ret %d\n",
 812                            packet->page_buf_cnt, packet->total_data_buflen,
 813                            ret);
 814         }
 815
 816         return ret;
 817 }
 818
 819 /* Move packet out of multi send data (msd), and clear msd */
 820 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
 821                                 struct sk_buff **msd_skb,
 822                                 struct multi_send_data *msdp)
 823 {
 824         *msd_skb = msdp->skb;
 825         *msd_send = msdp->pkt;
 826         msdp->skb = NULL;
 827         msdp->pkt = NULL;
 828         msdp->count = 0;
 829 }
 830
 831 /* RCU already held by caller */
 832 int netvsc_send(struct net_device *ndev,
 833                 struct hv_netvsc_packet *packet,
 834                 struct rndis_message *rndis_msg,
 835                 struct hv_page_buffer *pb,
 836                 struct sk_buff *skb)
 837 {
 838         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 839         struct netvsc_device *net_device
 840                 = rcu_dereference_bh(ndev_ctx->nvdev);
 841         struct hv_device *device = ndev_ctx->device_ctx;
 842         int ret = 0;
 843         struct netvsc_channel *nvchan;
 844         u32 pktlen = packet->total_data_buflen, msd_len = 0;
 845         unsigned int section_index = NETVSC_INVALID_INDEX;
 846         struct multi_send_data *msdp;
 847         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
 848         struct sk_buff *msd_skb = NULL;
 849         bool try_batch, xmit_more;
 850
 851         /* If device is rescinded, return error and packet will get dropped. */
 852         if (unlikely(!net_device || net_device->destroy))
 853                 return -ENODEV;
 854
 855         nvchan = &net_device->chan_table[packet->q_idx];
 856         packet->send_buf_index = NETVSC_INVALID_INDEX;
 857         packet->cp_partial = false;
 858
 859         /* Send control message directly without accessing msd (Multi-Send
 860          * Data) field which may be changed during data packet processing.
 861          */
 862         if (!skb)
 863                 return netvsc_send_pkt(device, packet, net_device, pb, skb);
 864
 865         /* batch packets in send buffer if possible */
 866         msdp = &nvchan->msd;
 867         if (msdp->pkt)
 868                 msd_len = msdp->pkt->total_data_buflen;
 869
 870         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
 871         if (try_batch && msd_len + pktlen + net_device->pkt_align <
 872             net_device->send_section_size) {
 873                 section_index = msdp->pkt->send_buf_index;
 874
 875         } else if (try_batch && msd_len + packet->rmsg_size <
 876                    net_device->send_section_size) {
 877                 section_index = msdp->pkt->send_buf_index;
 878                 packet->cp_partial = true;
 879
 880         } else if (pktlen + net_device->pkt_align <
 881                    net_device->send_section_size) {
 882                 section_index = netvsc_get_next_send_section(net_device);
 883                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
 884                         ++ndev_ctx->eth_stats.tx_send_full;
 885                 } else {
 886                         move_pkt_msd(&msd_send, &msd_skb, msdp);
 887                         msd_len = 0;
 888                 }
 889         }
 890
 891         /* Keep aggregating only if stack says more data is coming
 892          * and not doing mixed modes send and not flow blocked
 893          */
 894         xmit_more = skb->xmit_more &&
 895                 !packet->cp_partial &&
 896                 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
 897
 898         if (section_index != NETVSC_INVALID_INDEX) {
 899                 netvsc_copy_to_send_buf(net_device,
 900                                         section_index, msd_len,
 901                                         packet, rndis_msg, pb, xmit_more);
 902
 903                 packet->send_buf_index = section_index;
 904
 905                 if (packet->cp_partial) {
 906                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
 907                         packet->total_data_buflen = msd_len + packet->rmsg_size;
 908                 } else {
 909                         packet->page_buf_cnt = 0;
 910                         packet->total_data_buflen += msd_len;
 911                 }
 912
 913                 if (msdp->pkt) {
 914                         packet->total_packets += msdp->pkt->total_packets;
 915                         packet->total_bytes += msdp->pkt->total_bytes;
 916                 }
 917
 918                 if (msdp->skb)
 919                         dev_consume_skb_any(msdp->skb);
 920
 921                 if (xmit_more) {
 922                         msdp->skb = skb;
 923                         msdp->pkt = packet;
 924                         msdp->count++;
 925                 } else {
 926                         cur_send = packet;
 927                         msdp->skb = NULL;
 928                         msdp->pkt = NULL;
 929                         msdp->count = 0;
 930                 }
 931         } else {
 932                 move_pkt_msd(&msd_send, &msd_skb, msdp);
 933                 cur_send = packet;
 934         }
 935
 936         if (msd_send) {
 937                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
 938                                             NULL, msd_skb);
 939
 940                 if (m_ret != 0) {
 941                         netvsc_free_send_slot(net_device,
 942                                               msd_send->send_buf_index);
 943                         dev_kfree_skb_any(msd_skb);
 944                 }
 945         }
 946
 947         if (cur_send)
 948                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
 949
 950         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
 951                 netvsc_free_send_slot(net_device, section_index);
 952
 953         return ret;
 954 }
 955
 956 /* Send pending recv completions */
 957 static int send_recv_completions(struct net_device *ndev,
 958                                  struct netvsc_device *nvdev,
 959                                  struct netvsc_channel *nvchan)
 960 {
 961         struct multi_recv_comp *mrc = &nvchan->mrc;
 962         struct recv_comp_msg {
 963                 struct nvsp_message_header hdr;
 964                 u32 status;
 965         }  __packed;
 966         struct recv_comp_msg msg = {
 967                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
 968         };
 969         int ret;
 970
 971         while (mrc->first != mrc->next) {
 972                 const struct recv_comp_data *rcd
 973                         = mrc->slots + mrc->first;
 974
 975                 msg.status = rcd->status;
 976                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
 977                                        rcd->tid, VM_PKT_COMP, 0);
 978                 if (unlikely(ret)) {
 979                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 980
 981                         ++ndev_ctx->eth_stats.rx_comp_busy;
 982                         return ret;
 983                 }
 984
 985                 if (++mrc->first == nvdev->recv_completion_cnt)
 986                         mrc->first = 0;
 987         }
 988
 989         /* receive completion ring has been emptied */
 990         if (unlikely(nvdev->destroy))
 991                 wake_up(&nvdev->wait_drain);
 992
 993         return 0;
 994 }
 995
 996 /* Count how many receive completions are outstanding */
 997 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
 998                                  const struct multi_recv_comp *mrc,
 999                                  u32 *filled, u32 *avail)
1000 {
1001         u32 count = nvdev->recv_completion_cnt;
1002
1003         if (mrc->next >= mrc->first)
1004                 *filled = mrc->next - mrc->first;
1005         else
1006                 *filled = (count - mrc->first) + mrc->next;
1007
1008         *avail = count - *filled - 1;
1009 }
1010
1011 /* Add receive complete to ring to send to host. */
1012 static void enq_receive_complete(struct net_device *ndev,
1013                                  struct netvsc_device *nvdev, u16 q_idx,
1014                                  u64 tid, u32 status)
1015 {
1016         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1017         struct multi_recv_comp *mrc = &nvchan->mrc;
1018         struct recv_comp_data *rcd;
1019         u32 filled, avail;
1020
1021         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1022
1023         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1024                 send_recv_completions(ndev, nvdev, nvchan);
1025                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1026         }
1027
1028         if (unlikely(!avail)) {
1029                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1030                            q_idx, tid);
1031                 return;
1032         }
1033
1034         rcd = mrc->slots + mrc->next;
1035         rcd->tid = tid;
1036         rcd->status = status;
1037
1038         if (++mrc->next == nvdev->recv_completion_cnt)
1039                 mrc->next = 0;
1040 }
1041
1042 static int netvsc_receive(struct net_device *ndev,
1043                           struct netvsc_device *net_device,
1044                           struct net_device_context *net_device_ctx,
1045                           struct hv_device *device,
1046                           struct vmbus_channel *channel,
1047                           const struct vmpacket_descriptor *desc,
1048                           struct nvsp_message *nvsp)
1049 {
1050         const struct vmtransfer_page_packet_header *vmxferpage_packet
1051                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1052         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1053         char *recv_buf = net_device->recv_buf;
1054         u32 status = NVSP_STAT_SUCCESS;
1055         int i;
1056         int count = 0;
1057
1058         /* Make sure this is a valid nvsp packet */
1059         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1060                 netif_err(net_device_ctx, rx_err, ndev,
1061                           "Unknown nvsp packet type received %u\n",
1062                           nvsp->hdr.msg_type);
1063                 return 0;
1064         }
1065
1066         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1067                 netif_err(net_device_ctx, rx_err, ndev,
1068                           "Invalid xfer page set id - expecting %x got %x\n",
1069                           NETVSC_RECEIVE_BUFFER_ID,
1070                           vmxferpage_packet->xfer_pageset_id);
1071                 return 0;
1072         }
1073
1074         count = vmxferpage_packet->range_cnt;
1075
1076         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1077         for (i = 0; i < count; i++) {
1078                 void *data = recv_buf
1079                         + vmxferpage_packet->ranges[i].byte_offset;
1080                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1081
1082                 /* Pass it to the upper layer */
1083                 status = rndis_filter_receive(ndev, net_device,
1084                                               channel, data, buflen);
1085         }
1086
1087         enq_receive_complete(ndev, net_device, q_idx,
1088                              vmxferpage_packet->d.trans_id, status);
1089
1090         return count;
1091 }
1092
1093 static void netvsc_send_table(struct hv_device *hdev,
1094                               struct nvsp_message *nvmsg)
1095 {
1096         struct net_device *ndev = hv_get_drvdata(hdev);
1097         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1098         int i;
1099         u32 count, *tab;
1100
1101         count = nvmsg->msg.v5_msg.send_table.count;
1102         if (count != VRSS_SEND_TAB_SIZE) {
1103                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1104                 return;
1105         }
1106
1107         tab = (u32 *)((unsigned long)&nvmsg->msg.v5_msg.send_table +
1108                       nvmsg->msg.v5_msg.send_table.offset);
1109
1110         for (i = 0; i < count; i++)
1111                 net_device_ctx->tx_table[i] = tab[i];
1112 }
1113
1114 static void netvsc_send_vf(struct net_device_context *net_device_ctx,
1115                            struct nvsp_message *nvmsg)
1116 {
1117         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1118         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1119 }
1120
1121 static inline void netvsc_receive_inband(struct hv_device *hdev,
1122                                  struct net_device_context *net_device_ctx,
1123                                  struct nvsp_message *nvmsg)
1124 {
1125         switch (nvmsg->hdr.msg_type) {
1126         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1127                 netvsc_send_table(hdev, nvmsg);
1128                 break;
1129
1130         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1131                 netvsc_send_vf(net_device_ctx, nvmsg);
1132                 break;
1133         }
1134 }
1135
1136 static int netvsc_process_raw_pkt(struct hv_device *device,
1137                                   struct vmbus_channel *channel,
1138                                   struct netvsc_device *net_device,
1139                                   struct net_device *ndev,
1140                                   const struct vmpacket_descriptor *desc,
1141                                   int budget)
1142 {
1143         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1144         struct nvsp_message *nvmsg = hv_pkt_data(desc);
1145
1146         switch (desc->type) {
1147         case VM_PKT_COMP:
1148                 netvsc_send_completion(net_device, channel, device,
1149                                        desc, budget);
1150                 break;
1151
1152         case VM_PKT_DATA_USING_XFER_PAGES:
1153                 return netvsc_receive(ndev, net_device, net_device_ctx,
1154                                       device, channel, desc, nvmsg);
1155                 break;
1156
1157         case VM_PKT_DATA_INBAND:
1158                 netvsc_receive_inband(device, net_device_ctx, nvmsg);
1159                 break;
1160
1161         default:
1162                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1163                            desc->type, desc->trans_id);
1164                 break;
1165         }
1166
1167         return 0;
1168 }
1169
1170 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1171 {
1172         struct vmbus_channel *primary = channel->primary_channel;
1173
1174         return primary ? primary->device_obj : channel->device_obj;
1175 }
1176
1177 /* Network processing softirq
1178  * Process data in incoming ring buffer from host
1179  * Stops when ring is empty or budget is met or exceeded.
1180  */
1181 int netvsc_poll(struct napi_struct *napi, int budget)
1182 {
1183         struct netvsc_channel *nvchan
1184                 = container_of(napi, struct netvsc_channel, napi);
1185         struct netvsc_device *net_device = nvchan->net_device;
1186         struct vmbus_channel *channel = nvchan->channel;
1187         struct hv_device *device = netvsc_channel_to_device(channel);
1188         struct net_device *ndev = hv_get_drvdata(device);
1189         int work_done = 0;
1190
1191         /* If starting a new interval */
1192         if (!nvchan->desc)
1193                 nvchan->desc = hv_pkt_iter_first(channel);
1194
1195         while (nvchan->desc && work_done < budget) {
1196                 work_done += netvsc_process_raw_pkt(device, channel, net_device,
1197                                                     ndev, nvchan->desc, budget);
1198                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1199         }
1200
1201         /* If send of pending receive completions suceeded
1202          *   and did not exhaust NAPI budget this time
1203          *   and not doing busy poll
1204          * then re-enable host interrupts
1205          *     and reschedule if ring is not empty.
1206          */
1207         if (send_recv_completions(ndev, net_device, nvchan) == 0 &&
1208             work_done < budget &&
1209             napi_complete_done(napi, work_done) &&
1210             hv_end_read(&channel->inbound) &&
1211             napi_schedule_prep(napi)) {
1212                 hv_begin_read(&channel->inbound);
1213                 __napi_schedule(napi);
1214         }
1215
1216         /* Driver may overshoot since multiple packets per descriptor */
1217         return min(work_done, budget);
1218 }
1219
1220 /* Call back when data is available in host ring buffer.
1221  * Processing is deferred until network softirq (NAPI)
1222  */
1223 void netvsc_channel_cb(void *context)
1224 {
1225         struct netvsc_channel *nvchan = context;
1226         struct vmbus_channel *channel = nvchan->channel;
1227         struct hv_ring_buffer_info *rbi = &channel->inbound;
1228
1229         /* preload first vmpacket descriptor */
1230         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1231
1232         if (napi_schedule_prep(&nvchan->napi)) {
1233                 /* disable interupts from host */
1234                 hv_begin_read(rbi);
1235
1236                 __napi_schedule_irqoff(&nvchan->napi);
1237         }
1238 }
1239
1240 /*
1241  * netvsc_device_add - Callback when the device belonging to this
1242  * driver is added
1243  */
1244 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1245                                 const struct netvsc_device_info *device_info)
1246 {
1247         int i, ret = 0;
1248         struct netvsc_device *net_device;
1249         struct net_device *ndev = hv_get_drvdata(device);
1250         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1251
1252         net_device = alloc_net_device();
1253         if (!net_device)
1254                 return ERR_PTR(-ENOMEM);
1255
1256         for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1257                 net_device_ctx->tx_table[i] = 0;
1258
1259         /* Because the device uses NAPI, all the interrupt batching and
1260          * control is done via Net softirq, not the channel handling
1261          */
1262         set_channel_read_mode(device->channel, HV_CALL_ISR);
1263
1264         /* If we're reopening the device we may have multiple queues, fill the
1265          * chn_table with the default channel to use it before subchannels are
1266          * opened.
1267          * Initialize the channel state before we open;
1268          * we can be interrupted as soon as we open the channel.
1269          */
1270
1271         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1272                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1273
1274                 nvchan->channel = device->channel;
1275                 nvchan->net_device = net_device;
1276                 u64_stats_init(&nvchan->tx_stats.syncp);
1277                 u64_stats_init(&nvchan->rx_stats.syncp);
1278         }
1279
1280         /* Enable NAPI handler before init callbacks */
1281         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1282                        netvsc_poll, NAPI_POLL_WEIGHT);
1283
1284         /* Open the channel */
1285         ret = vmbus_open(device->channel, netvsc_ring_bytes,
1286                          netvsc_ring_bytes,  NULL, 0,
1287                          netvsc_channel_cb, net_device->chan_table);
1288
1289         if (ret != 0) {
1290                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1291                 goto cleanup;
1292         }
1293
1294         /* Channel is opened */
1295         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1296
1297         napi_enable(&net_device->chan_table[0].napi);
1298
1299         /* Connect with the NetVsp */
1300         ret = netvsc_connect_vsp(device, net_device, device_info);
1301         if (ret != 0) {
1302                 netdev_err(ndev,
1303                         "unable to connect to NetVSP - %d\n", ret);
1304                 goto close;
1305         }
1306
1307         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1308          * populated.
1309          */
1310         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1311
1312         return net_device;
1313
1314 close:
1315         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1316         napi_disable(&net_device->chan_table[0].napi);
1317
1318         /* Now, we can close the channel safely */
1319         vmbus_close(device->channel);
1320
1321 cleanup:
1322         netif_napi_del(&net_device->chan_table[0].napi);
1323         free_netvsc_device(&net_device->rcu);
1324
1325         return ERR_PTR(ret);
1326 }