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