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 <[email protected]>
  18  *   Hank Janssen  <[email protected]>
  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 static struct netvsc_device *alloc_net_device(void)
  66 {
  67         struct netvsc_device *net_device;
  68
  69         net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
  70         if (!net_device)
  71                 return NULL;
  72
  73         init_waitqueue_head(&net_device->wait_drain);
  74         net_device->destroy = false;
  75         atomic_set(&net_device->open_cnt, 0);
  76         net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
  77         net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
  78
  79         init_completion(&net_device->channel_init_wait);
  80         init_waitqueue_head(&net_device->subchan_open);
  81         INIT_WORK(&net_device->subchan_work, rndis_set_subchannel);
  82
  83         return net_device;
  84 }
  85
  86 static void free_netvsc_device(struct rcu_head *head)
  87 {
  88         struct netvsc_device *nvdev
  89                 = container_of(head, struct netvsc_device, rcu);
  90         int i;
  91
  92         for (i = 0; i < VRSS_CHANNEL_MAX; i++)
  93                 vfree(nvdev->chan_table[i].mrc.slots);
  94
  95         kfree(nvdev);
  96 }
  97
  98 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
  99 {
 100         call_rcu(&nvdev->rcu, free_netvsc_device);
 101 }
 102
 103 static void netvsc_destroy_buf(struct hv_device *device)
 104 {
 105         struct nvsp_message *revoke_packet;
 106         struct net_device *ndev = hv_get_drvdata(device);
 107         struct net_device_context *ndc = netdev_priv(ndev);
 108         struct netvsc_device *net_device = rtnl_dereference(ndc->nvdev);
 109         int ret;
 110
 111         /*
 112          * If we got a section count, it means we received a
 113          * SendReceiveBufferComplete msg (ie sent
 114          * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
 115          * to send a revoke msg here
 116          */
 117         if (net_device->recv_section_cnt) {
 118                 /* Send the revoke receive buffer */
 119                 revoke_packet = &net_device->revoke_packet;
 120                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 121
 122                 revoke_packet->hdr.msg_type =
 123                         NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
 124                 revoke_packet->msg.v1_msg.
 125                 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 126
 127                 ret = vmbus_sendpacket(device->channel,
 128                                        revoke_packet,
 129                                        sizeof(struct nvsp_message),
 130                                        (unsigned long)revoke_packet,
 131                                        VM_PKT_DATA_INBAND, 0);
 132                 /* If the failure is because the channel is rescinded;
 133                  * ignore the failure since we cannot send on a rescinded
 134                  * channel. This would allow us to properly cleanup
 135                  * even when the channel is rescinded.
 136                  */
 137                 if (device->channel->rescind)
 138                         ret = 0;
 139                 /*
 140                  * If we failed here, we might as well return and
 141                  * have a leak rather than continue and a bugchk
 142                  */
 143                 if (ret != 0) {
 144                         netdev_err(ndev, "unable to send "
 145                                 "revoke receive buffer to netvsp\n");
 146                         return;
 147                 }
 148                 net_device->recv_section_cnt = 0;
 149         }
 150
 151         /* Teardown the gpadl on the vsp end */
 152         if (net_device->recv_buf_gpadl_handle) {
 153                 ret = vmbus_teardown_gpadl(device->channel,
 154                                            net_device->recv_buf_gpadl_handle);
 155
 156                 /* If we failed here, we might as well return and have a leak
 157                  * rather than continue and a bugchk
 158                  */
 159                 if (ret != 0) {
 160                         netdev_err(ndev,
 161                                    "unable to teardown receive buffer's gpadl\n");
 162                         return;
 163                 }
 164                 net_device->recv_buf_gpadl_handle = 0;
 165         }
 166
 167         if (net_device->recv_buf) {
 168                 /* Free up the receive buffer */
 169                 vfree(net_device->recv_buf);
 170                 net_device->recv_buf = NULL;
 171         }
 172
 173         /* Deal with the send buffer we may have setup.
 174          * If we got a  send section size, it means we received a
 175          * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
 176          * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
 177          * to send a revoke msg here
 178          */
 179         if (net_device->send_section_cnt) {
 180                 /* Send the revoke receive buffer */
 181                 revoke_packet = &net_device->revoke_packet;
 182                 memset(revoke_packet, 0, sizeof(struct nvsp_message));
 183
 184                 revoke_packet->hdr.msg_type =
 185                         NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
 186                 revoke_packet->msg.v1_msg.revoke_send_buf.id =
 187                         NETVSC_SEND_BUFFER_ID;
 188
 189                 ret = vmbus_sendpacket(device->channel,
 190                                        revoke_packet,
 191                                        sizeof(struct nvsp_message),
 192                                        (unsigned long)revoke_packet,
 193                                        VM_PKT_DATA_INBAND, 0);
 194
 195                 /* If the failure is because the channel is rescinded;
 196                  * ignore the failure since we cannot send on a rescinded
 197                  * channel. This would allow us to properly cleanup
 198                  * even when the channel is rescinded.
 199                  */
 200                 if (device->channel->rescind)
 201                         ret = 0;
 202
 203                 /* If we failed here, we might as well return and
 204                  * have a leak rather than continue and a bugchk
 205                  */
 206                 if (ret != 0) {
 207                         netdev_err(ndev, "unable to send "
 208                                    "revoke send buffer to netvsp\n");
 209                         return;
 210                 }
 211                 net_device->send_section_cnt = 0;
 212         }
 213         /* Teardown the gpadl on the vsp end */
 214         if (net_device->send_buf_gpadl_handle) {
 215                 ret = vmbus_teardown_gpadl(device->channel,
 216                                            net_device->send_buf_gpadl_handle);
 217
 218                 /* If we failed here, we might as well return and have a leak
 219                  * rather than continue and a bugchk
 220                  */
 221                 if (ret != 0) {
 222                         netdev_err(ndev,
 223                                    "unable to teardown send buffer's gpadl\n");
 224                         return;
 225                 }
 226                 net_device->send_buf_gpadl_handle = 0;
 227         }
 228         if (net_device->send_buf) {
 229                 /* Free up the send buffer */
 230                 vfree(net_device->send_buf);
 231                 net_device->send_buf = NULL;
 232         }
 233         kfree(net_device->send_section_map);
 234 }
 235
 236 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
 237 {
 238         struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
 239         int node = cpu_to_node(nvchan->channel->target_cpu);
 240         size_t size;
 241
 242         size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
 243         nvchan->mrc.slots = vzalloc_node(size, node);
 244         if (!nvchan->mrc.slots)
 245                 nvchan->mrc.slots = vzalloc(size);
 246
 247         return nvchan->mrc.slots ? 0 : -ENOMEM;
 248 }
 249
 250 static int netvsc_init_buf(struct hv_device *device,
 251                            struct netvsc_device *net_device,
 252                            const struct netvsc_device_info *device_info)
 253 {
 254         struct nvsp_1_message_send_receive_buffer_complete *resp;
 255         struct net_device *ndev = hv_get_drvdata(device);
 256         struct nvsp_message *init_packet;
 257         unsigned int buf_size;
 258         size_t map_words;
 259         int ret = 0;
 260
 261         /* Get receive buffer area. */
 262         buf_size = device_info->recv_sections * device_info->recv_section_size;
 263         buf_size = roundup(buf_size, PAGE_SIZE);
 264
 265         net_device->recv_buf = vzalloc(buf_size);
 266         if (!net_device->recv_buf) {
 267                 netdev_err(ndev,
 268                            "unable to allocate receive buffer of size %u\n",
 269                            buf_size);
 270                 ret = -ENOMEM;
 271                 goto cleanup;
 272         }
 273
 274         /*
 275          * Establish the gpadl handle for this buffer on this
 276          * channel.  Note: This call uses the vmbus connection rather
 277          * than the channel to establish the gpadl handle.
 278          */
 279         ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
 280                                     buf_size,
 281                                     &net_device->recv_buf_gpadl_handle);
 282         if (ret != 0) {
 283                 netdev_err(ndev,
 284                         "unable to establish receive buffer's gpadl\n");
 285                 goto cleanup;
 286         }
 287
 288         /* Notify the NetVsp of the gpadl handle */
 289         init_packet = &net_device->channel_init_pkt;
 290         memset(init_packet, 0, sizeof(struct nvsp_message));
 291         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
 292         init_packet->msg.v1_msg.send_recv_buf.
 293                 gpadl_handle = net_device->recv_buf_gpadl_handle;
 294         init_packet->msg.v1_msg.
 295                 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
 296
 297         /* Send the gpadl notification request */
 298         ret = vmbus_sendpacket(device->channel, init_packet,
 299                                sizeof(struct nvsp_message),
 300                                (unsigned long)init_packet,
 301                                VM_PKT_DATA_INBAND,
 302                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 303         if (ret != 0) {
 304                 netdev_err(ndev,
 305                         "unable to send receive buffer's gpadl to netvsp\n");
 306                 goto cleanup;
 307         }
 308
 309         wait_for_completion(&net_device->channel_init_wait);
 310
 311         /* Check the response */
 312         resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
 313         if (resp->status != NVSP_STAT_SUCCESS) {
 314                 netdev_err(ndev,
 315                            "Unable to complete receive buffer initialization with NetVsp - status %d\n",
 316                            resp->status);
 317                 ret = -EINVAL;
 318                 goto cleanup;
 319         }
 320
 321         /* Parse the response */
 322         netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
 323                    resp->num_sections, resp->sections[0].sub_alloc_size,
 324                    resp->sections[0].num_sub_allocs);
 325
 326         /* There should only be one section for the entire receive buffer */
 327         if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
 328                 ret = -EINVAL;
 329                 goto cleanup;
 330         }
 331
 332         net_device->recv_section_size = resp->sections[0].sub_alloc_size;
 333         net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
 334
 335         /* Setup receive completion ring */
 336         net_device->recv_completion_cnt
 337                 = round_up(net_device->recv_section_cnt + 1,
 338                            PAGE_SIZE / sizeof(u64));
 339         ret = netvsc_alloc_recv_comp_ring(net_device, 0);
 340         if (ret)
 341                 goto cleanup;
 342
 343         /* Now setup the send buffer. */
 344         buf_size = device_info->send_sections * device_info->send_section_size;
 345         buf_size = round_up(buf_size, PAGE_SIZE);
 346
 347         net_device->send_buf = vzalloc(buf_size);
 348         if (!net_device->send_buf) {
 349                 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
 350                            buf_size);
 351                 ret = -ENOMEM;
 352                 goto cleanup;
 353         }
 354
 355         /* Establish the gpadl handle for this buffer on this
 356          * channel.  Note: This call uses the vmbus connection rather
 357          * than the channel to establish the gpadl handle.
 358          */
 359         ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
 360                                     buf_size,
 361                                     &net_device->send_buf_gpadl_handle);
 362         if (ret != 0) {
 363                 netdev_err(ndev,
 364                            "unable to establish send buffer's gpadl\n");
 365                 goto cleanup;
 366         }
 367
 368         /* Notify the NetVsp of the gpadl handle */
 369         init_packet = &net_device->channel_init_pkt;
 370         memset(init_packet, 0, sizeof(struct nvsp_message));
 371         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
 372         init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
 373                 net_device->send_buf_gpadl_handle;
 374         init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
 375
 376         /* Send the gpadl notification request */
 377         ret = vmbus_sendpacket(device->channel, init_packet,
 378                                sizeof(struct nvsp_message),
 379                                (unsigned long)init_packet,
 380                                VM_PKT_DATA_INBAND,
 381                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 382         if (ret != 0) {
 383                 netdev_err(ndev,
 384                            "unable to send send buffer's gpadl to netvsp\n");
 385                 goto cleanup;
 386         }
 387
 388         wait_for_completion(&net_device->channel_init_wait);
 389
 390         /* Check the response */
 391         if (init_packet->msg.v1_msg.
 392             send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
 393                 netdev_err(ndev, "Unable to complete send buffer "
 394                            "initialization with NetVsp - status %d\n",
 395                            init_packet->msg.v1_msg.
 396                            send_send_buf_complete.status);
 397                 ret = -EINVAL;
 398                 goto cleanup;
 399         }
 400
 401         /* Parse the response */
 402         net_device->send_section_size = init_packet->msg.
 403                                 v1_msg.send_send_buf_complete.section_size;
 404
 405         /* Section count is simply the size divided by the section size. */
 406         net_device->send_section_cnt = buf_size / net_device->send_section_size;
 407
 408         netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
 409                    net_device->send_section_size, net_device->send_section_cnt);
 410
 411         /* Setup state for managing the send buffer. */
 412         map_words = DIV_ROUND_UP(net_device->send_section_cnt, BITS_PER_LONG);
 413
 414         net_device->send_section_map = kcalloc(map_words, sizeof(ulong), GFP_KERNEL);
 415         if (net_device->send_section_map == NULL) {
 416                 ret = -ENOMEM;
 417                 goto cleanup;
 418         }
 419
 420         goto exit;
 421
 422 cleanup:
 423         netvsc_destroy_buf(device);
 424
 425 exit:
 426         return ret;
 427 }
 428
 429 /* Negotiate NVSP protocol version */
 430 static int negotiate_nvsp_ver(struct hv_device *device,
 431                               struct netvsc_device *net_device,
 432                               struct nvsp_message *init_packet,
 433                               u32 nvsp_ver)
 434 {
 435         struct net_device *ndev = hv_get_drvdata(device);
 436         int ret;
 437
 438         memset(init_packet, 0, sizeof(struct nvsp_message));
 439         init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
 440         init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
 441         init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
 442
 443         /* Send the init request */
 444         ret = vmbus_sendpacket(device->channel, init_packet,
 445                                sizeof(struct nvsp_message),
 446                                (unsigned long)init_packet,
 447                                VM_PKT_DATA_INBAND,
 448                                VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 449
 450         if (ret != 0)
 451                 return ret;
 452
 453         wait_for_completion(&net_device->channel_init_wait);
 454
 455         if (init_packet->msg.init_msg.init_complete.status !=
 456             NVSP_STAT_SUCCESS)
 457                 return -EINVAL;
 458
 459         if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
 460                 return 0;
 461
 462         /* NVSPv2 or later: Send NDIS config */
 463         memset(init_packet, 0, sizeof(struct nvsp_message));
 464         init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
 465         init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
 466         init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
 467
 468         if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
 469                 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
 470
 471                 /* Teaming bit is needed to receive link speed updates */
 472                 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
 473         }
 474
 475         ret = vmbus_sendpacket(device->channel, init_packet,
 476                                 sizeof(struct nvsp_message),
 477                                 (unsigned long)init_packet,
 478                                 VM_PKT_DATA_INBAND, 0);
 479
 480         return ret;
 481 }
 482
 483 static int netvsc_connect_vsp(struct hv_device *device,
 484                               struct netvsc_device *net_device,
 485                               const struct netvsc_device_info *device_info)
 486 {
 487         static const u32 ver_list[] = {
 488                 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
 489                 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5
 490         };
 491         struct nvsp_message *init_packet;
 492         int ndis_version, i, ret;
 493
 494         init_packet = &net_device->channel_init_pkt;
 495
 496         /* Negotiate the latest NVSP protocol supported */
 497         for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
 498                 if (negotiate_nvsp_ver(device, net_device, init_packet,
 499                                        ver_list[i])  == 0) {
 500                         net_device->nvsp_version = ver_list[i];
 501                         break;
 502                 }
 503
 504         if (i < 0) {
 505                 ret = -EPROTO;
 506                 goto cleanup;
 507         }
 508
 509         pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
 510
 511         /* Send the ndis version */
 512         memset(init_packet, 0, sizeof(struct nvsp_message));
 513
 514         if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
 515                 ndis_version = 0x00060001;
 516         else
 517                 ndis_version = 0x0006001e;
 518
 519         init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
 520         init_packet->msg.v1_msg.
 521                 send_ndis_ver.ndis_major_ver =
 522                                 (ndis_version & 0xFFFF0000) >> 16;
 523         init_packet->msg.v1_msg.
 524                 send_ndis_ver.ndis_minor_ver =
 525                                 ndis_version & 0xFFFF;
 526
 527         /* Send the init request */
 528         ret = vmbus_sendpacket(device->channel, init_packet,
 529                                 sizeof(struct nvsp_message),
 530                                 (unsigned long)init_packet,
 531                                 VM_PKT_DATA_INBAND, 0);
 532         if (ret != 0)
 533                 goto cleanup;
 534
 535
 536         ret = netvsc_init_buf(device, net_device, device_info);
 537
 538 cleanup:
 539         return ret;
 540 }
 541
 542 static void netvsc_disconnect_vsp(struct hv_device *device)
 543 {
 544         netvsc_destroy_buf(device);
 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         cancel_work_sync(&net_device->subchan_work);
 559
 560         netvsc_disconnect_vsp(device);
 561
 562         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
 563
 564         /*
 565          * At this point, no one should be accessing net_device
 566          * except in here
 567          */
 568         netdev_dbg(ndev, "net device safe to remove\n");
 569
 570         /* Now, we can close the channel safely */
 571         vmbus_close(device->channel);
 572
 573         /* And dissassociate NAPI context from device */
 574         for (i = 0; i < net_device->num_chn; i++)
 575                 netif_napi_del(&net_device->chan_table[i].napi);
 576
 577         /* Release all resources */
 578         free_netvsc_device_rcu(net_device);
 579 }
 580
 581 #define RING_AVAIL_PERCENT_HIWATER 20
 582 #define RING_AVAIL_PERCENT_LOWATER 10
 583
 584 /*
 585  * Get the percentage of available bytes to write in the ring.
 586  * The return value is in range from 0 to 100.
 587  */
 588 static inline u32 hv_ringbuf_avail_percent(
 589                 struct hv_ring_buffer_info *ring_info)
 590 {
 591         u32 avail_read, avail_write;
 592
 593         hv_get_ringbuffer_availbytes(ring_info, &avail_read, &avail_write);
 594
 595         return avail_write * 100 / ring_info->ring_datasize;
 596 }
 597
 598 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
 599                                          u32 index)
 600 {
 601         sync_change_bit(index, net_device->send_section_map);
 602 }
 603
 604 static void netvsc_send_tx_complete(struct netvsc_device *net_device,
 605                                     struct vmbus_channel *incoming_channel,
 606                                     struct hv_device *device,
 607                                     const struct vmpacket_descriptor *desc,
 608                                     int budget)
 609 {
 610         struct sk_buff *skb = (struct sk_buff *)(unsigned long)desc->trans_id;
 611         struct net_device *ndev = hv_get_drvdata(device);
 612         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 613         struct vmbus_channel *channel = device->channel;
 614         u16 q_idx = 0;
 615         int queue_sends;
 616
 617         /* Notify the layer above us */
 618         if (likely(skb)) {
 619                 const struct hv_netvsc_packet *packet
 620                         = (struct hv_netvsc_packet *)skb->cb;
 621                 u32 send_index = packet->send_buf_index;
 622                 struct netvsc_stats *tx_stats;
 623
 624                 if (send_index != NETVSC_INVALID_INDEX)
 625                         netvsc_free_send_slot(net_device, send_index);
 626                 q_idx = packet->q_idx;
 627                 channel = incoming_channel;
 628
 629                 tx_stats = &net_device->chan_table[q_idx].tx_stats;
 630
 631                 u64_stats_update_begin(&tx_stats->syncp);
 632                 tx_stats->packets += packet->total_packets;
 633                 tx_stats->bytes += packet->total_bytes;
 634                 u64_stats_update_end(&tx_stats->syncp);
 635
 636                 napi_consume_skb(skb, budget);
 637         }
 638
 639         queue_sends =
 640                 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
 641
 642         if (net_device->destroy && queue_sends == 0)
 643                 wake_up(&net_device->wait_drain);
 644
 645         if (netif_tx_queue_stopped(netdev_get_tx_queue(ndev, q_idx)) &&
 646             (hv_ringbuf_avail_percent(&channel->outbound) > RING_AVAIL_PERCENT_HIWATER ||
 647              queue_sends < 1)) {
 648                 netif_tx_wake_queue(netdev_get_tx_queue(ndev, q_idx));
 649                 ndev_ctx->eth_stats.wake_queue++;
 650         }
 651 }
 652
 653 static void netvsc_send_completion(struct netvsc_device *net_device,
 654                                    struct vmbus_channel *incoming_channel,
 655                                    struct hv_device *device,
 656                                    const struct vmpacket_descriptor *desc,
 657                                    int budget)
 658 {
 659         struct nvsp_message *nvsp_packet = hv_pkt_data(desc);
 660         struct net_device *ndev = hv_get_drvdata(device);
 661
 662         switch (nvsp_packet->hdr.msg_type) {
 663         case NVSP_MSG_TYPE_INIT_COMPLETE:
 664         case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
 665         case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
 666         case NVSP_MSG5_TYPE_SUBCHANNEL:
 667                 /* Copy the response back */
 668                 memcpy(&net_device->channel_init_pkt, nvsp_packet,
 669                        sizeof(struct nvsp_message));
 670                 complete(&net_device->channel_init_wait);
 671                 break;
 672
 673         case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
 674                 netvsc_send_tx_complete(net_device, incoming_channel,
 675                                         device, desc, budget);
 676                 break;
 677
 678         default:
 679                 netdev_err(ndev,
 680                            "Unknown send completion type %d received!!\n",
 681                            nvsp_packet->hdr.msg_type);
 682         }
 683 }
 684
 685 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
 686 {
 687         unsigned long *map_addr = net_device->send_section_map;
 688         unsigned int i;
 689
 690         for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
 691                 if (sync_test_and_set_bit(i, map_addr) == 0)
 692                         return i;
 693         }
 694
 695         return NETVSC_INVALID_INDEX;
 696 }
 697
 698 static u32 netvsc_copy_to_send_buf(struct netvsc_device *net_device,
 699                                    unsigned int section_index,
 700                                    u32 pend_size,
 701                                    struct hv_netvsc_packet *packet,
 702                                    struct rndis_message *rndis_msg,
 703                                    struct hv_page_buffer *pb,
 704                                    struct sk_buff *skb)
 705 {
 706         char *start = net_device->send_buf;
 707         char *dest = start + (section_index * net_device->send_section_size)
 708                      + pend_size;
 709         int i;
 710         u32 msg_size = 0;
 711         u32 padding = 0;
 712         u32 remain = packet->total_data_buflen % net_device->pkt_align;
 713         u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
 714                 packet->page_buf_cnt;
 715
 716         /* Add padding */
 717         if (skb->xmit_more && remain && !packet->cp_partial) {
 718                 padding = net_device->pkt_align - remain;
 719                 rndis_msg->msg_len += padding;
 720                 packet->total_data_buflen += padding;
 721         }
 722
 723         for (i = 0; i < page_count; i++) {
 724                 char *src = phys_to_virt(pb[i].pfn << PAGE_SHIFT);
 725                 u32 offset = pb[i].offset;
 726                 u32 len = pb[i].len;
 727
 728                 memcpy(dest, (src + offset), len);
 729                 msg_size += len;
 730                 dest += len;
 731         }
 732
 733         if (padding) {
 734                 memset(dest, 0, padding);
 735                 msg_size += padding;
 736         }
 737
 738         return msg_size;
 739 }
 740
 741 static inline int netvsc_send_pkt(
 742         struct hv_device *device,
 743         struct hv_netvsc_packet *packet,
 744         struct netvsc_device *net_device,
 745         struct hv_page_buffer *pb,
 746         struct sk_buff *skb)
 747 {
 748         struct nvsp_message nvmsg;
 749         struct nvsp_1_message_send_rndis_packet * const rpkt =
 750                 &nvmsg.msg.v1_msg.send_rndis_pkt;
 751         struct netvsc_channel * const nvchan =
 752                 &net_device->chan_table[packet->q_idx];
 753         struct vmbus_channel *out_channel = nvchan->channel;
 754         struct net_device *ndev = hv_get_drvdata(device);
 755         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 756         struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
 757         u64 req_id;
 758         int ret;
 759         u32 ring_avail = hv_ringbuf_avail_percent(&out_channel->outbound);
 760
 761         nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
 762         if (skb)
 763                 rpkt->channel_type = 0;         /* 0 is RMC_DATA */
 764         else
 765                 rpkt->channel_type = 1;         /* 1 is RMC_CONTROL */
 766
 767         rpkt->send_buf_section_index = packet->send_buf_index;
 768         if (packet->send_buf_index == NETVSC_INVALID_INDEX)
 769                 rpkt->send_buf_section_size = 0;
 770         else
 771                 rpkt->send_buf_section_size = packet->total_data_buflen;
 772
 773         req_id = (ulong)skb;
 774
 775         if (out_channel->rescind)
 776                 return -ENODEV;
 777
 778         if (packet->page_buf_cnt) {
 779                 if (packet->cp_partial)
 780                         pb += packet->rmsg_pgcnt;
 781
 782                 ret = vmbus_sendpacket_pagebuffer(out_channel,
 783                                                   pb, packet->page_buf_cnt,
 784                                                   &nvmsg, sizeof(nvmsg),
 785                                                   req_id);
 786         } else {
 787                 ret = vmbus_sendpacket(out_channel,
 788                                        &nvmsg, sizeof(nvmsg),
 789                                        req_id, VM_PKT_DATA_INBAND,
 790                                        VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 791         }
 792
 793         if (ret == 0) {
 794                 atomic_inc_return(&nvchan->queue_sends);
 795
 796                 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
 797                         netif_tx_stop_queue(txq);
 798                         ndev_ctx->eth_stats.stop_queue++;
 799                 }
 800         } else if (ret == -EAGAIN) {
 801                 netif_tx_stop_queue(txq);
 802                 ndev_ctx->eth_stats.stop_queue++;
 803                 if (atomic_read(&nvchan->queue_sends) < 1) {
 804                         netif_tx_wake_queue(txq);
 805                         ndev_ctx->eth_stats.wake_queue++;
 806                         ret = -ENOSPC;
 807                 }
 808         } else {
 809                 netdev_err(ndev,
 810                            "Unable to send packet pages %u len %u, ret %d\n",
 811                            packet->page_buf_cnt, packet->total_data_buflen,
 812                            ret);
 813         }
 814
 815         return ret;
 816 }
 817
 818 /* Move packet out of multi send data (msd), and clear msd */
 819 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
 820                                 struct sk_buff **msd_skb,
 821                                 struct multi_send_data *msdp)
 822 {
 823         *msd_skb = msdp->skb;
 824         *msd_send = msdp->pkt;
 825         msdp->skb = NULL;
 826         msdp->pkt = NULL;
 827         msdp->count = 0;
 828 }
 829
 830 /* RCU already held by caller */
 831 int netvsc_send(struct net_device_context *ndev_ctx,
 832                 struct hv_netvsc_packet *packet,
 833                 struct rndis_message *rndis_msg,
 834                 struct hv_page_buffer *pb,
 835                 struct sk_buff *skb)
 836 {
 837         struct netvsc_device *net_device
 838                 = rcu_dereference_bh(ndev_ctx->nvdev);
 839         struct hv_device *device = ndev_ctx->device_ctx;
 840         int ret = 0;
 841         struct netvsc_channel *nvchan;
 842         u32 pktlen = packet->total_data_buflen, msd_len = 0;
 843         unsigned int section_index = NETVSC_INVALID_INDEX;
 844         struct multi_send_data *msdp;
 845         struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
 846         struct sk_buff *msd_skb = NULL;
 847         bool try_batch;
 848         bool xmit_more = (skb != NULL) ? skb->xmit_more : false;
 849
 850         /* If device is rescinded, return error and packet will get dropped. */
 851         if (unlikely(!net_device || net_device->destroy))
 852                 return -ENODEV;
 853
 854         /* We may race with netvsc_connect_vsp()/netvsc_init_buf() and get
 855          * here before the negotiation with the host is finished and
 856          * send_section_map may not be allocated yet.
 857          */
 858         if (unlikely(!net_device->send_section_map))
 859                 return -EAGAIN;
 860
 861         nvchan = &net_device->chan_table[packet->q_idx];
 862         packet->send_buf_index = NETVSC_INVALID_INDEX;
 863         packet->cp_partial = false;
 864
 865         /* Send control message directly without accessing msd (Multi-Send
 866          * Data) field which may be changed during data packet processing.
 867          */
 868         if (!skb) {
 869                 cur_send = packet;
 870                 goto send_now;
 871         }
 872
 873         /* batch packets in send buffer if possible */
 874         msdp = &nvchan->msd;
 875         if (msdp->pkt)
 876                 msd_len = msdp->pkt->total_data_buflen;
 877
 878         try_batch =  msd_len > 0 && msdp->count < net_device->max_pkt;
 879         if (try_batch && msd_len + pktlen + net_device->pkt_align <
 880             net_device->send_section_size) {
 881                 section_index = msdp->pkt->send_buf_index;
 882
 883         } else if (try_batch && msd_len + packet->rmsg_size <
 884                    net_device->send_section_size) {
 885                 section_index = msdp->pkt->send_buf_index;
 886                 packet->cp_partial = true;
 887
 888         } else if (pktlen + net_device->pkt_align <
 889                    net_device->send_section_size) {
 890                 section_index = netvsc_get_next_send_section(net_device);
 891                 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
 892                         ++ndev_ctx->eth_stats.tx_send_full;
 893                 } else {
 894                         move_pkt_msd(&msd_send, &msd_skb, msdp);
 895                         msd_len = 0;
 896                 }
 897         }
 898
 899         if (section_index != NETVSC_INVALID_INDEX) {
 900                 netvsc_copy_to_send_buf(net_device,
 901                                         section_index, msd_len,
 902                                         packet, rndis_msg, pb, skb);
 903
 904                 packet->send_buf_index = section_index;
 905
 906                 if (packet->cp_partial) {
 907                         packet->page_buf_cnt -= packet->rmsg_pgcnt;
 908                         packet->total_data_buflen = msd_len + packet->rmsg_size;
 909                 } else {
 910                         packet->page_buf_cnt = 0;
 911                         packet->total_data_buflen += msd_len;
 912                 }
 913
 914                 if (msdp->pkt) {
 915                         packet->total_packets += msdp->pkt->total_packets;
 916                         packet->total_bytes += msdp->pkt->total_bytes;
 917                 }
 918
 919                 if (msdp->skb)
 920                         dev_consume_skb_any(msdp->skb);
 921
 922                 if (xmit_more && !packet->cp_partial) {
 923                         msdp->skb = skb;
 924                         msdp->pkt = packet;
 925                         msdp->count++;
 926                 } else {
 927                         cur_send = packet;
 928                         msdp->skb = NULL;
 929                         msdp->pkt = NULL;
 930                         msdp->count = 0;
 931                 }
 932         } else {
 933                 move_pkt_msd(&msd_send, &msd_skb, msdp);
 934                 cur_send = packet;
 935         }
 936
 937         if (msd_send) {
 938                 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
 939                                             NULL, msd_skb);
 940
 941                 if (m_ret != 0) {
 942                         netvsc_free_send_slot(net_device,
 943                                               msd_send->send_buf_index);
 944                         dev_kfree_skb_any(msd_skb);
 945                 }
 946         }
 947
 948 send_now:
 949         if (cur_send)
 950                 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
 951
 952         if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
 953                 netvsc_free_send_slot(net_device, section_index);
 954
 955         return ret;
 956 }
 957
 958 /* Send pending recv completions */
 959 static int send_recv_completions(struct net_device *ndev,
 960                                  struct netvsc_device *nvdev,
 961                                  struct netvsc_channel *nvchan)
 962 {
 963         struct multi_recv_comp *mrc = &nvchan->mrc;
 964         struct recv_comp_msg {
 965                 struct nvsp_message_header hdr;
 966                 u32 status;
 967         }  __packed;
 968         struct recv_comp_msg msg = {
 969                 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
 970         };
 971         int ret;
 972
 973         while (mrc->first != mrc->next) {
 974                 const struct recv_comp_data *rcd
 975                         = mrc->slots + mrc->first;
 976
 977                 msg.status = rcd->status;
 978                 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
 979                                        rcd->tid, VM_PKT_COMP, 0);
 980                 if (unlikely(ret)) {
 981                         struct net_device_context *ndev_ctx = netdev_priv(ndev);
 982
 983                         ++ndev_ctx->eth_stats.rx_comp_busy;
 984                         return ret;
 985                 }
 986
 987                 if (++mrc->first == nvdev->recv_completion_cnt)
 988                         mrc->first = 0;
 989         }
 990
 991         /* receive completion ring has been emptied */
 992         if (unlikely(nvdev->destroy))
 993                 wake_up(&nvdev->wait_drain);
 994
 995         return 0;
 996 }
 997
 998 /* Count how many receive completions are outstanding */
 999 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1000                                  const struct multi_recv_comp *mrc,
1001                                  u32 *filled, u32 *avail)
1002 {
1003         u32 count = nvdev->recv_completion_cnt;
1004
1005         if (mrc->next >= mrc->first)
1006                 *filled = mrc->next - mrc->first;
1007         else
1008                 *filled = (count - mrc->first) + mrc->next;
1009
1010         *avail = count - *filled - 1;
1011 }
1012
1013 /* Add receive complete to ring to send to host. */
1014 static void enq_receive_complete(struct net_device *ndev,
1015                                  struct netvsc_device *nvdev, u16 q_idx,
1016                                  u64 tid, u32 status)
1017 {
1018         struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1019         struct multi_recv_comp *mrc = &nvchan->mrc;
1020         struct recv_comp_data *rcd;
1021         u32 filled, avail;
1022
1023         recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1024
1025         if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1026                 send_recv_completions(ndev, nvdev, nvchan);
1027                 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1028         }
1029
1030         if (unlikely(!avail)) {
1031                 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1032                            q_idx, tid);
1033                 return;
1034         }
1035
1036         rcd = mrc->slots + mrc->next;
1037         rcd->tid = tid;
1038         rcd->status = status;
1039
1040         if (++mrc->next == nvdev->recv_completion_cnt)
1041                 mrc->next = 0;
1042 }
1043
1044 static int netvsc_receive(struct net_device *ndev,
1045                           struct netvsc_device *net_device,
1046                           struct net_device_context *net_device_ctx,
1047                           struct hv_device *device,
1048                           struct vmbus_channel *channel,
1049                           const struct vmpacket_descriptor *desc,
1050                           struct nvsp_message *nvsp)
1051 {
1052         const struct vmtransfer_page_packet_header *vmxferpage_packet
1053                 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1054         u16 q_idx = channel->offermsg.offer.sub_channel_index;
1055         char *recv_buf = net_device->recv_buf;
1056         u32 status = NVSP_STAT_SUCCESS;
1057         int i;
1058         int count = 0;
1059
1060         /* Make sure this is a valid nvsp packet */
1061         if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1062                 netif_err(net_device_ctx, rx_err, ndev,
1063                           "Unknown nvsp packet type received %u\n",
1064                           nvsp->hdr.msg_type);
1065                 return 0;
1066         }
1067
1068         if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1069                 netif_err(net_device_ctx, rx_err, ndev,
1070                           "Invalid xfer page set id - expecting %x got %x\n",
1071                           NETVSC_RECEIVE_BUFFER_ID,
1072                           vmxferpage_packet->xfer_pageset_id);
1073                 return 0;
1074         }
1075
1076         count = vmxferpage_packet->range_cnt;
1077
1078         /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1079         for (i = 0; i < count; i++) {
1080                 void *data = recv_buf
1081                         + vmxferpage_packet->ranges[i].byte_offset;
1082                 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1083
1084                 /* Pass it to the upper layer */
1085                 status = rndis_filter_receive(ndev, net_device, device,
1086                                               channel, data, buflen);
1087         }
1088
1089         enq_receive_complete(ndev, net_device, q_idx,
1090                              vmxferpage_packet->d.trans_id, status);
1091
1092         return count;
1093 }
1094
1095 static void netvsc_send_table(struct hv_device *hdev,
1096                               struct nvsp_message *nvmsg)
1097 {
1098         struct net_device *ndev = hv_get_drvdata(hdev);
1099         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1100         int i;
1101         u32 count, *tab;
1102
1103         count = nvmsg->msg.v5_msg.send_table.count;
1104         if (count != VRSS_SEND_TAB_SIZE) {
1105                 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1106                 return;
1107         }
1108
1109         tab = (u32 *)((unsigned long)&nvmsg->msg.v5_msg.send_table +
1110                       nvmsg->msg.v5_msg.send_table.offset);
1111
1112         for (i = 0; i < count; i++)
1113                 net_device_ctx->tx_send_table[i] = tab[i];
1114 }
1115
1116 static void netvsc_send_vf(struct net_device_context *net_device_ctx,
1117                            struct nvsp_message *nvmsg)
1118 {
1119         net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1120         net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1121 }
1122
1123 static inline void netvsc_receive_inband(struct hv_device *hdev,
1124                                  struct net_device_context *net_device_ctx,
1125                                  struct nvsp_message *nvmsg)
1126 {
1127         switch (nvmsg->hdr.msg_type) {
1128         case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1129                 netvsc_send_table(hdev, nvmsg);
1130                 break;
1131
1132         case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1133                 netvsc_send_vf(net_device_ctx, nvmsg);
1134                 break;
1135         }
1136 }
1137
1138 static int netvsc_process_raw_pkt(struct hv_device *device,
1139                                   struct vmbus_channel *channel,
1140                                   struct netvsc_device *net_device,
1141                                   struct net_device *ndev,
1142                                   const struct vmpacket_descriptor *desc,
1143                                   int budget)
1144 {
1145         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1146         struct nvsp_message *nvmsg = hv_pkt_data(desc);
1147
1148         switch (desc->type) {
1149         case VM_PKT_COMP:
1150                 netvsc_send_completion(net_device, channel, device,
1151                                        desc, budget);
1152                 break;
1153
1154         case VM_PKT_DATA_USING_XFER_PAGES:
1155                 return netvsc_receive(ndev, net_device, net_device_ctx,
1156                                       device, channel, desc, nvmsg);
1157                 break;
1158
1159         case VM_PKT_DATA_INBAND:
1160                 netvsc_receive_inband(device, net_device_ctx, nvmsg);
1161                 break;
1162
1163         default:
1164                 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1165                            desc->type, desc->trans_id);
1166                 break;
1167         }
1168
1169         return 0;
1170 }
1171
1172 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1173 {
1174         struct vmbus_channel *primary = channel->primary_channel;
1175
1176         return primary ? primary->device_obj : channel->device_obj;
1177 }
1178
1179 /* Network processing softirq
1180  * Process data in incoming ring buffer from host
1181  * Stops when ring is empty or budget is met or exceeded.
1182  */
1183 int netvsc_poll(struct napi_struct *napi, int budget)
1184 {
1185         struct netvsc_channel *nvchan
1186                 = container_of(napi, struct netvsc_channel, napi);
1187         struct netvsc_device *net_device = nvchan->net_device;
1188         struct vmbus_channel *channel = nvchan->channel;
1189         struct hv_device *device = netvsc_channel_to_device(channel);
1190         struct net_device *ndev = hv_get_drvdata(device);
1191         int work_done = 0;
1192
1193         /* If starting a new interval */
1194         if (!nvchan->desc)
1195                 nvchan->desc = hv_pkt_iter_first(channel);
1196
1197         while (nvchan->desc && work_done < budget) {
1198                 work_done += netvsc_process_raw_pkt(device, channel, net_device,
1199                                                     ndev, nvchan->desc, budget);
1200                 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1201         }
1202
1203         /* If send of pending receive completions suceeded
1204          *   and did not exhaust NAPI budget this time
1205          *   and not doing busy poll
1206          * then re-enable host interrupts
1207          *     and reschedule if ring is not empty.
1208          */
1209         if (send_recv_completions(ndev, net_device, nvchan) == 0 &&
1210             work_done < budget &&
1211             napi_complete_done(napi, work_done) &&
1212             hv_end_read(&channel->inbound)) {
1213                 hv_begin_read(&channel->inbound);
1214                 napi_reschedule(napi);
1215         }
1216
1217         /* Driver may overshoot since multiple packets per descriptor */
1218         return min(work_done, budget);
1219 }
1220
1221 /* Call back when data is available in host ring buffer.
1222  * Processing is deferred until network softirq (NAPI)
1223  */
1224 void netvsc_channel_cb(void *context)
1225 {
1226         struct netvsc_channel *nvchan = context;
1227         struct vmbus_channel *channel = nvchan->channel;
1228         struct hv_ring_buffer_info *rbi = &channel->inbound;
1229
1230         /* preload first vmpacket descriptor */
1231         prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1232
1233         if (napi_schedule_prep(&nvchan->napi)) {
1234                 /* disable interupts from host */
1235                 hv_begin_read(rbi);
1236
1237                 __napi_schedule(&nvchan->napi);
1238         }
1239 }
1240
1241 /*
1242  * netvsc_device_add - Callback when the device belonging to this
1243  * driver is added
1244  */
1245 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1246                                 const struct netvsc_device_info *device_info)
1247 {
1248         int i, ret = 0;
1249         int ring_size = device_info->ring_size;
1250         struct netvsc_device *net_device;
1251         struct net_device *ndev = hv_get_drvdata(device);
1252         struct net_device_context *net_device_ctx = netdev_priv(ndev);
1253
1254         net_device = alloc_net_device();
1255         if (!net_device)
1256                 return ERR_PTR(-ENOMEM);
1257
1258         net_device->ring_size = ring_size;
1259
1260         /* Because the device uses NAPI, all the interrupt batching and
1261          * control is done via Net softirq, not the channel handling
1262          */
1263         set_channel_read_mode(device->channel, HV_CALL_ISR);
1264
1265         /* If we're reopening the device we may have multiple queues, fill the
1266          * chn_table with the default channel to use it before subchannels are
1267          * opened.
1268          * Initialize the channel state before we open;
1269          * we can be interrupted as soon as we open the channel.
1270          */
1271
1272         for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1273                 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1274
1275                 nvchan->channel = device->channel;
1276                 nvchan->net_device = net_device;
1277                 u64_stats_init(&nvchan->tx_stats.syncp);
1278                 u64_stats_init(&nvchan->rx_stats.syncp);
1279         }
1280
1281         /* Enable NAPI handler before init callbacks */
1282         netif_napi_add(ndev, &net_device->chan_table[0].napi,
1283                        netvsc_poll, NAPI_POLL_WEIGHT);
1284
1285         /* Open the channel */
1286         ret = vmbus_open(device->channel, ring_size * PAGE_SIZE,
1287                          ring_size * PAGE_SIZE, NULL, 0,
1288                          netvsc_channel_cb,
1289                          net_device->chan_table);
1290
1291         if (ret != 0) {
1292                 netif_napi_del(&net_device->chan_table[0].napi);
1293                 netdev_err(ndev, "unable to open channel: %d\n", ret);
1294                 goto cleanup;
1295         }
1296
1297         /* Channel is opened */
1298         netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1299
1300         napi_enable(&net_device->chan_table[0].napi);
1301
1302         /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1303          * populated.
1304          */
1305         rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1306
1307         /* Connect with the NetVsp */
1308         ret = netvsc_connect_vsp(device, net_device, device_info);
1309         if (ret != 0) {
1310                 netdev_err(ndev,
1311                         "unable to connect to NetVSP - %d\n", ret);
1312                 goto close;
1313         }
1314
1315         return net_device;
1316
1317 close:
1318         RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1319         napi_disable(&net_device->chan_table[0].napi);
1320
1321         /* Now, we can close the channel safely */
1322         vmbus_close(device->channel);
1323
1324 cleanup:
1325         free_netvsc_device(&net_device->rcu);
1326
1327         return ERR_PTR(ret);
1328 }