sys/dev/hyperv/netvsc/hv_net_vsc.c

   1 /*-
   2  * Copyright (c) 2009-2012 Microsoft Corp.
   3  * Copyright (c) 2010-2012 Citrix Inc.
   4  * Copyright (c) 2012 NetApp Inc.
   5  * All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice unmodified, this list of conditions, and the following
  12  *    disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
  18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
  19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
  20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
  21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
  22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
  26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  27  *
  28  * $FreeBSD$
  29  */
  30
  31 /**
  32  * HyperV vmbus network VSC (virtual services client) module
  33  *
  34  */
  35
  36
  37 #include <sys/param.h>
  38 #include <sys/kernel.h>
  39 #include <sys/socket.h>
  40 #include <sys/lock.h>
  41 #include <net/if.h>
  42 #include <net/if_arp.h>
  43 #include <machine/bus.h>
  44 #include <machine/atomic.h>
  45
  46 #include <dev/hyperv/include/hyperv.h>
  47 #include "hv_net_vsc.h"
  48 #include "hv_rndis.h"
  49 #include "hv_rndis_filter.h"
  50
  51 MALLOC_DEFINE(M_NETVSC, "netvsc", "Hyper-V netvsc driver");
  52
  53 /*
  54  * Forward declarations
  55  */
  56 static void hv_nv_on_channel_callback(void *context);
  57 static int  hv_nv_init_send_buffer_with_net_vsp(struct hv_device *device);
  58 static int  hv_nv_init_rx_buffer_with_net_vsp(struct hv_device *device);
  59 static int  hv_nv_destroy_send_buffer(netvsc_dev *net_dev);
  60 static int  hv_nv_destroy_rx_buffer(netvsc_dev *net_dev);
  61 static int  hv_nv_connect_to_vsp(struct hv_device *device);
  62 static void hv_nv_on_send_completion(netvsc_dev *net_dev,
  63     struct hv_device *device, hv_vm_packet_descriptor *pkt);
  64 static void hv_nv_on_receive(netvsc_dev *net_dev,
  65     struct hv_device *device, struct hv_vmbus_channel *chan,
  66     hv_vm_packet_descriptor *pkt);
  67
  68 /*
  69  *
  70  */
  71 static inline netvsc_dev *
  72 hv_nv_alloc_net_device(struct hv_device *device)
  73 {
  74         netvsc_dev *net_dev;
  75         hn_softc_t *sc = device_get_softc(device->device);
  76
  77         net_dev = malloc(sizeof(netvsc_dev), M_NETVSC, M_WAITOK | M_ZERO);
  78
  79         net_dev->dev = device;
  80         net_dev->destroy = FALSE;
  81         sc->net_dev = net_dev;
  82
  83         return (net_dev);
  84 }
  85
  86 /*
  87  *
  88  */
  89 static inline netvsc_dev *
  90 hv_nv_get_outbound_net_device(struct hv_device *device)
  91 {
  92         hn_softc_t *sc = device_get_softc(device->device);
  93         netvsc_dev *net_dev = sc->net_dev;;
  94
  95         if ((net_dev != NULL) && net_dev->destroy) {
  96                 return (NULL);
  97         }
  98
  99         return (net_dev);
 100 }
 101
 102 /*
 103  *
 104  */
 105 static inline netvsc_dev *
 106 hv_nv_get_inbound_net_device(struct hv_device *device)
 107 {
 108         hn_softc_t *sc = device_get_softc(device->device);
 109         netvsc_dev *net_dev = sc->net_dev;;
 110
 111         if (net_dev == NULL) {
 112                 return (net_dev);
 113         }
 114         /*
 115          * When the device is being destroyed; we only
 116          * permit incoming packets if and only if there
 117          * are outstanding sends.
 118          */
 119         if (net_dev->destroy) {
 120                 return (NULL);
 121         }
 122
 123         return (net_dev);
 124 }
 125
 126 int
 127 hv_nv_get_next_send_section(netvsc_dev *net_dev)
 128 {
 129         unsigned long bitsmap_words = net_dev->bitsmap_words;
 130         unsigned long *bitsmap = net_dev->send_section_bitsmap;
 131         unsigned long idx;
 132         int ret = NVSP_1_CHIMNEY_SEND_INVALID_SECTION_INDEX;
 133         int i;
 134
 135         for (i = 0; i < bitsmap_words; i++) {
 136                 idx = ffsl(~bitsmap[i]);
 137                 if (0 == idx)
 138                         continue;
 139
 140                 idx--;
 141                 KASSERT(i * BITS_PER_LONG + idx < net_dev->send_section_count,
 142                     ("invalid i %d and idx %lu", i, idx));
 143
 144                 if (atomic_testandset_long(&bitsmap[i], idx))
 145                         continue;
 146
 147                 ret = i * BITS_PER_LONG + idx;
 148                 break;
 149         }
 150
 151         return (ret);
 152 }
 153
 154 /*
 155  * Net VSC initialize receive buffer with net VSP
 156  *
 157  * Net VSP:  Network virtual services client, also known as the
 158  *     Hyper-V extensible switch and the synthetic data path.
 159  */
 160 static int
 161 hv_nv_init_rx_buffer_with_net_vsp(struct hv_device *device)
 162 {
 163         netvsc_dev *net_dev;
 164         nvsp_msg *init_pkt;
 165         int ret = 0;
 166
 167         net_dev = hv_nv_get_outbound_net_device(device);
 168         if (!net_dev) {
 169                 return (ENODEV);
 170         }
 171
 172         net_dev->rx_buf = contigmalloc(net_dev->rx_buf_size, M_NETVSC,
 173             M_ZERO, 0UL, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
 174
 175         /*
 176          * Establish the GPADL handle for this buffer on this channel.
 177          * Note:  This call uses the vmbus connection rather than the
 178          * channel to establish the gpadl handle.
 179          * GPADL:  Guest physical address descriptor list.
 180          */
 181         ret = hv_vmbus_channel_establish_gpadl(
 182                 device->channel, net_dev->rx_buf,
 183                 net_dev->rx_buf_size, &net_dev->rx_buf_gpadl_handle);
 184         if (ret != 0) {
 185                 goto cleanup;
 186         }
 187
 188         /* sema_wait(&ext->channel_init_sema); KYS CHECK */
 189
 190         /* Notify the NetVsp of the gpadl handle */
 191         init_pkt = &net_dev->channel_init_packet;
 192
 193         memset(init_pkt, 0, sizeof(nvsp_msg));
 194
 195         init_pkt->hdr.msg_type = nvsp_msg_1_type_send_rx_buf;
 196         init_pkt->msgs.vers_1_msgs.send_rx_buf.gpadl_handle =
 197             net_dev->rx_buf_gpadl_handle;
 198         init_pkt->msgs.vers_1_msgs.send_rx_buf.id =
 199             NETVSC_RECEIVE_BUFFER_ID;
 200
 201         /* Send the gpadl notification request */
 202
 203         ret = hv_vmbus_channel_send_packet(device->channel, init_pkt,
 204             sizeof(nvsp_msg), (uint64_t)(uintptr_t)init_pkt,
 205             HV_VMBUS_PACKET_TYPE_DATA_IN_BAND,
 206             HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 207         if (ret != 0) {
 208                 goto cleanup;
 209         }
 210
 211         sema_wait(&net_dev->channel_init_sema);
 212
 213         /* Check the response */
 214         if (init_pkt->msgs.vers_1_msgs.send_rx_buf_complete.status
 215             != nvsp_status_success) {
 216                 ret = EINVAL;
 217                 goto cleanup;
 218         }
 219
 220         net_dev->rx_section_count =
 221             init_pkt->msgs.vers_1_msgs.send_rx_buf_complete.num_sections;
 222
 223         net_dev->rx_sections = malloc(net_dev->rx_section_count *
 224             sizeof(nvsp_1_rx_buf_section), M_NETVSC, M_WAITOK);
 225         memcpy(net_dev->rx_sections,
 226             init_pkt->msgs.vers_1_msgs.send_rx_buf_complete.sections,
 227             net_dev->rx_section_count * sizeof(nvsp_1_rx_buf_section));
 228
 229
 230         /*
 231          * For first release, there should only be 1 section that represents
 232          * the entire receive buffer
 233          */
 234         if (net_dev->rx_section_count != 1
 235             || net_dev->rx_sections->offset != 0) {
 236                 ret = EINVAL;
 237                 goto cleanup;
 238         }
 239
 240         goto exit;
 241
 242 cleanup:
 243         hv_nv_destroy_rx_buffer(net_dev);
 244
 245 exit:
 246         return (ret);
 247 }
 248
 249 /*
 250  * Net VSC initialize send buffer with net VSP
 251  */
 252 static int
 253 hv_nv_init_send_buffer_with_net_vsp(struct hv_device *device)
 254 {
 255         netvsc_dev *net_dev;
 256         nvsp_msg *init_pkt;
 257         int ret = 0;
 258
 259         net_dev = hv_nv_get_outbound_net_device(device);
 260         if (!net_dev) {
 261                 return (ENODEV);
 262         }
 263
 264         net_dev->send_buf  = contigmalloc(net_dev->send_buf_size, M_NETVSC,
 265             M_ZERO, 0UL, BUS_SPACE_MAXADDR, PAGE_SIZE, 0);
 266         if (net_dev->send_buf == NULL) {
 267                 ret = ENOMEM;
 268                 goto cleanup;
 269         }
 270
 271         /*
 272          * Establish the gpadl handle for this buffer on this channel.
 273          * Note:  This call uses the vmbus connection rather than the
 274          * channel to establish the gpadl handle.
 275          */
 276         ret = hv_vmbus_channel_establish_gpadl(device->channel,
 277             net_dev->send_buf, net_dev->send_buf_size,
 278             &net_dev->send_buf_gpadl_handle);
 279         if (ret != 0) {
 280                 goto cleanup;
 281         }
 282
 283         /* Notify the NetVsp of the gpadl handle */
 284
 285         init_pkt = &net_dev->channel_init_packet;
 286
 287         memset(init_pkt, 0, sizeof(nvsp_msg));
 288
 289         init_pkt->hdr.msg_type = nvsp_msg_1_type_send_send_buf;
 290         init_pkt->msgs.vers_1_msgs.send_rx_buf.gpadl_handle =
 291             net_dev->send_buf_gpadl_handle;
 292         init_pkt->msgs.vers_1_msgs.send_rx_buf.id =
 293             NETVSC_SEND_BUFFER_ID;
 294
 295         /* Send the gpadl notification request */
 296
 297         ret = hv_vmbus_channel_send_packet(device->channel, init_pkt,
 298             sizeof(nvsp_msg), (uint64_t)init_pkt,
 299             HV_VMBUS_PACKET_TYPE_DATA_IN_BAND,
 300             HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 301         if (ret != 0) {
 302                 goto cleanup;
 303         }
 304
 305         sema_wait(&net_dev->channel_init_sema);
 306
 307         /* Check the response */
 308         if (init_pkt->msgs.vers_1_msgs.send_send_buf_complete.status
 309             != nvsp_status_success) {
 310                 ret = EINVAL;
 311                 goto cleanup;
 312         }
 313
 314         net_dev->send_section_size =
 315             init_pkt->msgs.vers_1_msgs.send_send_buf_complete.section_size;
 316         net_dev->send_section_count =
 317             net_dev->send_buf_size / net_dev->send_section_size;
 318         net_dev->bitsmap_words = howmany(net_dev->send_section_count,
 319             BITS_PER_LONG);
 320         net_dev->send_section_bitsmap =
 321             malloc(net_dev->bitsmap_words * sizeof(long), M_NETVSC,
 322             M_WAITOK | M_ZERO);
 323
 324         goto exit;
 325
 326 cleanup:
 327         hv_nv_destroy_send_buffer(net_dev);
 328
 329 exit:
 330         return (ret);
 331 }
 332
 333 /*
 334  * Net VSC destroy receive buffer
 335  */
 336 static int
 337 hv_nv_destroy_rx_buffer(netvsc_dev *net_dev)
 338 {
 339         nvsp_msg *revoke_pkt;
 340         int ret = 0;
 341
 342         /*
 343          * If we got a section count, it means we received a
 344          * send_rx_buf_complete msg
 345          * (ie sent nvsp_msg_1_type_send_rx_buf msg) therefore,
 346          * we need to send a revoke msg here
 347          */
 348         if (net_dev->rx_section_count) {
 349                 /* Send the revoke receive buffer */
 350                 revoke_pkt = &net_dev->revoke_packet;
 351                 memset(revoke_pkt, 0, sizeof(nvsp_msg));
 352
 353                 revoke_pkt->hdr.msg_type = nvsp_msg_1_type_revoke_rx_buf;
 354                 revoke_pkt->msgs.vers_1_msgs.revoke_rx_buf.id =
 355                     NETVSC_RECEIVE_BUFFER_ID;
 356
 357                 ret = hv_vmbus_channel_send_packet(net_dev->dev->channel,
 358                     revoke_pkt, sizeof(nvsp_msg),
 359                     (uint64_t)(uintptr_t)revoke_pkt,
 360                     HV_VMBUS_PACKET_TYPE_DATA_IN_BAND, 0);
 361
 362                 /*
 363                  * If we failed here, we might as well return and have a leak
 364                  * rather than continue and a bugchk
 365                  */
 366                 if (ret != 0) {
 367                         return (ret);
 368                 }
 369         }
 370
 371         /* Tear down the gpadl on the vsp end */
 372         if (net_dev->rx_buf_gpadl_handle) {
 373                 ret = hv_vmbus_channel_teardown_gpdal(net_dev->dev->channel,
 374                     net_dev->rx_buf_gpadl_handle);
 375                 /*
 376                  * If we failed here, we might as well return and have a leak
 377                  * rather than continue and a bugchk
 378                  */
 379                 if (ret != 0) {
 380                         return (ret);
 381                 }
 382                 net_dev->rx_buf_gpadl_handle = 0;
 383         }
 384
 385         if (net_dev->rx_buf) {
 386                 /* Free up the receive buffer */
 387                 contigfree(net_dev->rx_buf, net_dev->rx_buf_size, M_NETVSC);
 388                 net_dev->rx_buf = NULL;
 389         }
 390
 391         if (net_dev->rx_sections) {
 392                 free(net_dev->rx_sections, M_NETVSC);
 393                 net_dev->rx_sections = NULL;
 394                 net_dev->rx_section_count = 0;
 395         }
 396
 397         return (ret);
 398 }
 399
 400 /*
 401  * Net VSC destroy send buffer
 402  */
 403 static int
 404 hv_nv_destroy_send_buffer(netvsc_dev *net_dev)
 405 {
 406         nvsp_msg *revoke_pkt;
 407         int ret = 0;
 408
 409         /*
 410          * If we got a section count, it means we received a
 411          * send_rx_buf_complete msg
 412          * (ie sent nvsp_msg_1_type_send_rx_buf msg) therefore,
 413          * we need to send a revoke msg here
 414          */
 415         if (net_dev->send_section_size) {
 416                 /* Send the revoke send buffer */
 417                 revoke_pkt = &net_dev->revoke_packet;
 418                 memset(revoke_pkt, 0, sizeof(nvsp_msg));
 419
 420                 revoke_pkt->hdr.msg_type =
 421                     nvsp_msg_1_type_revoke_send_buf;
 422                 revoke_pkt->msgs.vers_1_msgs.revoke_send_buf.id =
 423                     NETVSC_SEND_BUFFER_ID;
 424
 425                 ret = hv_vmbus_channel_send_packet(net_dev->dev->channel,
 426                     revoke_pkt, sizeof(nvsp_msg),
 427                     (uint64_t)(uintptr_t)revoke_pkt,
 428                     HV_VMBUS_PACKET_TYPE_DATA_IN_BAND, 0);
 429                 /*
 430                  * If we failed here, we might as well return and have a leak
 431                  * rather than continue and a bugchk
 432                  */
 433                 if (ret != 0) {
 434                         return (ret);
 435                 }
 436         }
 437
 438         /* Tear down the gpadl on the vsp end */
 439         if (net_dev->send_buf_gpadl_handle) {
 440                 ret = hv_vmbus_channel_teardown_gpdal(net_dev->dev->channel,
 441                     net_dev->send_buf_gpadl_handle);
 442
 443                 /*
 444                  * If we failed here, we might as well return and have a leak
 445                  * rather than continue and a bugchk
 446                  */
 447                 if (ret != 0) {
 448                         return (ret);
 449                 }
 450                 net_dev->send_buf_gpadl_handle = 0;
 451         }
 452
 453         if (net_dev->send_buf) {
 454                 /* Free up the receive buffer */
 455                 contigfree(net_dev->send_buf, net_dev->send_buf_size, M_NETVSC);
 456                 net_dev->send_buf = NULL;
 457         }
 458
 459         if (net_dev->send_section_bitsmap) {
 460                 free(net_dev->send_section_bitsmap, M_NETVSC);
 461         }
 462
 463         return (ret);
 464 }
 465
 466
 467 /*
 468  * Attempt to negotiate the caller-specified NVSP version
 469  *
 470  * For NVSP v2, Server 2008 R2 does not set
 471  * init_pkt->msgs.init_msgs.init_compl.negotiated_prot_vers
 472  * to the negotiated version, so we cannot rely on that.
 473  */
 474 static int
 475 hv_nv_negotiate_nvsp_protocol(struct hv_device *device, netvsc_dev *net_dev,
 476     uint32_t nvsp_ver)
 477 {
 478         nvsp_msg *init_pkt;
 479         int ret;
 480
 481         init_pkt = &net_dev->channel_init_packet;
 482         memset(init_pkt, 0, sizeof(nvsp_msg));
 483         init_pkt->hdr.msg_type = nvsp_msg_type_init;
 484
 485         /*
 486          * Specify parameter as the only acceptable protocol version
 487          */
 488         init_pkt->msgs.init_msgs.init.p1.protocol_version = nvsp_ver;
 489         init_pkt->msgs.init_msgs.init.protocol_version_2 = nvsp_ver;
 490
 491         /* Send the init request */
 492         ret = hv_vmbus_channel_send_packet(device->channel, init_pkt,
 493             sizeof(nvsp_msg), (uint64_t)(uintptr_t)init_pkt,
 494             HV_VMBUS_PACKET_TYPE_DATA_IN_BAND,
 495             HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 496         if (ret != 0)
 497                 return (-1);
 498
 499         sema_wait(&net_dev->channel_init_sema);
 500
 501         if (init_pkt->msgs.init_msgs.init_compl.status != nvsp_status_success)
 502                 return (EINVAL);
 503
 504         return (0);
 505 }
 506
 507 /*
 508  * Send NDIS version 2 config packet containing MTU.
 509  *
 510  * Not valid for NDIS version 1.
 511  */
 512 static int
 513 hv_nv_send_ndis_config(struct hv_device *device, uint32_t mtu)
 514 {
 515         netvsc_dev *net_dev;
 516         nvsp_msg *init_pkt;
 517         int ret;
 518
 519         net_dev = hv_nv_get_outbound_net_device(device);
 520         if (!net_dev)
 521                 return (-ENODEV);
 522
 523         /*
 524          * Set up configuration packet, write MTU
 525          * Indicate we are capable of handling VLAN tags
 526          */
 527         init_pkt = &net_dev->channel_init_packet;
 528         memset(init_pkt, 0, sizeof(nvsp_msg));
 529         init_pkt->hdr.msg_type = nvsp_msg_2_type_send_ndis_config;
 530         init_pkt->msgs.vers_2_msgs.send_ndis_config.mtu = mtu;
 531         init_pkt->
 532                 msgs.vers_2_msgs.send_ndis_config.capabilities.u1.u2.ieee8021q
 533                 = 1;
 534
 535         /* Send the configuration packet */
 536         ret = hv_vmbus_channel_send_packet(device->channel, init_pkt,
 537             sizeof(nvsp_msg), (uint64_t)(uintptr_t)init_pkt,
 538             HV_VMBUS_PACKET_TYPE_DATA_IN_BAND, 0);
 539         if (ret != 0)
 540                 return (-EINVAL);
 541
 542         return (0);
 543 }
 544
 545 /*
 546  * Net VSC connect to VSP
 547  */
 548 static int
 549 hv_nv_connect_to_vsp(struct hv_device *device)
 550 {
 551         netvsc_dev *net_dev;
 552         nvsp_msg *init_pkt;
 553         uint32_t ndis_version;
 554         uint32_t protocol_list[] = { NVSP_PROTOCOL_VERSION_1,
 555             NVSP_PROTOCOL_VERSION_2,
 556             NVSP_PROTOCOL_VERSION_4,
 557             NVSP_PROTOCOL_VERSION_5 };
 558         int i;
 559         int protocol_number = nitems(protocol_list);
 560         int ret = 0;
 561         device_t dev = device->device;
 562         hn_softc_t *sc = device_get_softc(dev);
 563         struct ifnet *ifp = sc->arpcom.ac_ifp;
 564
 565         net_dev = hv_nv_get_outbound_net_device(device);
 566         if (!net_dev) {
 567                 return (ENODEV);
 568         }
 569
 570         /*
 571          * Negotiate the NVSP version.  Try the latest NVSP first.
 572          */
 573         for (i = protocol_number - 1; i >= 0; i--) {
 574                 if (hv_nv_negotiate_nvsp_protocol(device, net_dev,
 575                     protocol_list[i]) == 0) {
 576                         net_dev->nvsp_version = protocol_list[i];
 577                         if (bootverbose)
 578                                 device_printf(dev, "Netvsc: got version 0x%x\n",
 579                                     net_dev->nvsp_version);
 580                         break;
 581                 }
 582         }
 583
 584         if (i < 0) {
 585                 if (bootverbose)
 586                         device_printf(dev, "failed to negotiate a valid "
 587                             "protocol.\n");
 588                 return (EPROTO);
 589         }
 590
 591         /*
 592          * Set the MTU if supported by this NVSP protocol version
 593          * This needs to be right after the NVSP init message per Haiyang
 594          */
 595         if (net_dev->nvsp_version >= NVSP_PROTOCOL_VERSION_2)
 596                 ret = hv_nv_send_ndis_config(device, ifp->if_mtu);
 597
 598         /*
 599          * Send the NDIS version
 600          */
 601         init_pkt = &net_dev->channel_init_packet;
 602
 603         memset(init_pkt, 0, sizeof(nvsp_msg));
 604
 605         if (net_dev->nvsp_version <= NVSP_PROTOCOL_VERSION_4) {
 606                 ndis_version = NDIS_VERSION_6_1;
 607         } else {
 608                 ndis_version = NDIS_VERSION_6_30;
 609         }
 610
 611         init_pkt->hdr.msg_type = nvsp_msg_1_type_send_ndis_vers;
 612         init_pkt->msgs.vers_1_msgs.send_ndis_vers.ndis_major_vers =
 613             (ndis_version & 0xFFFF0000) >> 16;
 614         init_pkt->msgs.vers_1_msgs.send_ndis_vers.ndis_minor_vers =
 615             ndis_version & 0xFFFF;
 616
 617         /* Send the init request */
 618
 619         ret = hv_vmbus_channel_send_packet(device->channel, init_pkt,
 620             sizeof(nvsp_msg), (uint64_t)(uintptr_t)init_pkt,
 621             HV_VMBUS_PACKET_TYPE_DATA_IN_BAND, 0);
 622         if (ret != 0) {
 623                 goto cleanup;
 624         }
 625         /*
 626          * TODO:  BUGBUG - We have to wait for the above msg since the netvsp
 627          * uses KMCL which acknowledges packet (completion packet)
 628          * since our Vmbus always set the
 629          * HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED flag
 630          */
 631         /* sema_wait(&NetVscChannel->channel_init_sema); */
 632
 633         /* Post the big receive buffer to NetVSP */
 634         if (net_dev->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
 635                 net_dev->rx_buf_size = NETVSC_RECEIVE_BUFFER_SIZE_LEGACY;
 636         else
 637                 net_dev->rx_buf_size = NETVSC_RECEIVE_BUFFER_SIZE;
 638         net_dev->send_buf_size = NETVSC_SEND_BUFFER_SIZE;
 639
 640         ret = hv_nv_init_rx_buffer_with_net_vsp(device);
 641         if (ret == 0)
 642                 ret = hv_nv_init_send_buffer_with_net_vsp(device);
 643
 644 cleanup:
 645         return (ret);
 646 }
 647
 648 /*
 649  * Net VSC disconnect from VSP
 650  */
 651 static void
 652 hv_nv_disconnect_from_vsp(netvsc_dev *net_dev)
 653 {
 654         hv_nv_destroy_rx_buffer(net_dev);
 655         hv_nv_destroy_send_buffer(net_dev);
 656 }
 657
 658 /*
 659  * Net VSC on device add
 660  *
 661  * Callback when the device belonging to this driver is added
 662  */
 663 netvsc_dev *
 664 hv_nv_on_device_add(struct hv_device *device, void *additional_info)
 665 {
 666         netvsc_dev *net_dev;
 667         int ret = 0;
 668
 669         net_dev = hv_nv_alloc_net_device(device);
 670         if (!net_dev)
 671                 goto cleanup;
 672
 673         /* Initialize the NetVSC channel extension */
 674
 675         sema_init(&net_dev->channel_init_sema, 0, "netdev_sema");
 676
 677         /*
 678          * Open the channel
 679          */
 680         ret = hv_vmbus_channel_open(device->channel,
 681             NETVSC_DEVICE_RING_BUFFER_SIZE, NETVSC_DEVICE_RING_BUFFER_SIZE,
 682             NULL, 0, hv_nv_on_channel_callback, device->channel);
 683         if (ret != 0)
 684                 goto cleanup;
 685
 686         /*
 687          * Connect with the NetVsp
 688          */
 689         ret = hv_nv_connect_to_vsp(device);
 690         if (ret != 0)
 691                 goto close;
 692
 693         return (net_dev);
 694
 695 close:
 696         /* Now, we can close the channel safely */
 697
 698         hv_vmbus_channel_close(device->channel);
 699
 700 cleanup:
 701         /*
 702          * Free the packet buffers on the netvsc device packet queue.
 703          * Release other resources.
 704          */
 705         if (net_dev) {
 706                 sema_destroy(&net_dev->channel_init_sema);
 707                 free(net_dev, M_NETVSC);
 708         }
 709
 710         return (NULL);
 711 }
 712
 713 /*
 714  * Net VSC on device remove
 715  */
 716 int
 717 hv_nv_on_device_remove(struct hv_device *device, boolean_t destroy_channel)
 718 {
 719         hn_softc_t *sc = device_get_softc(device->device);
 720         netvsc_dev *net_dev = sc->net_dev;;
 721
 722         /* Stop outbound traffic ie sends and receives completions */
 723         net_dev->destroy = TRUE;
 724
 725         hv_nv_disconnect_from_vsp(net_dev);
 726
 727         /* At this point, no one should be accessing net_dev except in here */
 728
 729         /* Now, we can close the channel safely */
 730
 731         if (!destroy_channel) {
 732                 device->channel->state =
 733                     HV_CHANNEL_CLOSING_NONDESTRUCTIVE_STATE;
 734         }
 735
 736         hv_vmbus_channel_close(device->channel);
 737
 738         sema_destroy(&net_dev->channel_init_sema);
 739         free(net_dev, M_NETVSC);
 740
 741         return (0);
 742 }
 743
 744 /*
 745  * Net VSC on send completion
 746  */
 747 static void
 748 hv_nv_on_send_completion(netvsc_dev *net_dev,
 749     struct hv_device *device, hv_vm_packet_descriptor *pkt)
 750 {
 751         nvsp_msg *nvsp_msg_pkt;
 752         netvsc_packet *net_vsc_pkt;
 753
 754         nvsp_msg_pkt =
 755             (nvsp_msg *)((unsigned long)pkt + (pkt->data_offset8 << 3));
 756
 757         if (nvsp_msg_pkt->hdr.msg_type == nvsp_msg_type_init_complete
 758                 || nvsp_msg_pkt->hdr.msg_type
 759                         == nvsp_msg_1_type_send_rx_buf_complete
 760                 || nvsp_msg_pkt->hdr.msg_type
 761                         == nvsp_msg_1_type_send_send_buf_complete) {
 762                 /* Copy the response back */
 763                 memcpy(&net_dev->channel_init_packet, nvsp_msg_pkt,
 764                     sizeof(nvsp_msg));
 765                 sema_post(&net_dev->channel_init_sema);
 766         } else if (nvsp_msg_pkt->hdr.msg_type ==
 767                     nvsp_msg_1_type_send_rndis_pkt_complete) {
 768                 /* Get the send context */
 769                 net_vsc_pkt =
 770                     (netvsc_packet *)(unsigned long)pkt->transaction_id;
 771                 if (NULL != net_vsc_pkt) {
 772                         if (net_vsc_pkt->send_buf_section_idx !=
 773                             NVSP_1_CHIMNEY_SEND_INVALID_SECTION_INDEX) {
 774                                 u_long mask;
 775                                 int idx;
 776
 777                                 idx = net_vsc_pkt->send_buf_section_idx /
 778                                     BITS_PER_LONG;
 779                                 KASSERT(idx < net_dev->bitsmap_words,
 780                                     ("invalid section index %u",
 781                                      net_vsc_pkt->send_buf_section_idx));
 782                                 mask = 1UL <<
 783                                     (net_vsc_pkt->send_buf_section_idx %
 784                                      BITS_PER_LONG);
 785
 786                                 KASSERT(net_dev->send_section_bitsmap[idx] &
 787                                     mask,
 788                                     ("index bitmap 0x%lx, section index %u, "
 789                                      "bitmap idx %d, bitmask 0x%lx",
 790                                      net_dev->send_section_bitsmap[idx],
 791                                      net_vsc_pkt->send_buf_section_idx,
 792                                      idx, mask));
 793                                 atomic_clear_long(
 794                                     &net_dev->send_section_bitsmap[idx], mask);
 795                         }
 796
 797                         /* Notify the layer above us */
 798                         net_vsc_pkt->compl.send.on_send_completion(
 799                             net_vsc_pkt->compl.send.send_completion_context);
 800
 801                 }
 802         }
 803 }
 804
 805 /*
 806  * Net VSC on send
 807  * Sends a packet on the specified Hyper-V device.
 808  * Returns 0 on success, non-zero on failure.
 809  */
 810 int
 811 hv_nv_on_send(struct hv_vmbus_channel *chan, netvsc_packet *pkt)
 812 {
 813         nvsp_msg send_msg;
 814         int ret;
 815
 816         send_msg.hdr.msg_type = nvsp_msg_1_type_send_rndis_pkt;
 817         if (pkt->is_data_pkt) {
 818                 /* 0 is RMC_DATA */
 819                 send_msg.msgs.vers_1_msgs.send_rndis_pkt.chan_type = 0;
 820         } else {
 821                 /* 1 is RMC_CONTROL */
 822                 send_msg.msgs.vers_1_msgs.send_rndis_pkt.chan_type = 1;
 823         }
 824
 825         send_msg.msgs.vers_1_msgs.send_rndis_pkt.send_buf_section_idx =
 826             pkt->send_buf_section_idx;
 827         send_msg.msgs.vers_1_msgs.send_rndis_pkt.send_buf_section_size =
 828             pkt->send_buf_section_size;
 829
 830         if (pkt->page_buf_count) {
 831                 ret = hv_vmbus_channel_send_packet_pagebuffer(chan,
 832                     pkt->page_buffers, pkt->page_buf_count,
 833                     &send_msg, sizeof(nvsp_msg), (uint64_t)(uintptr_t)pkt);
 834         } else {
 835                 ret = hv_vmbus_channel_send_packet(chan,
 836                     &send_msg, sizeof(nvsp_msg), (uint64_t)(uintptr_t)pkt,
 837                     HV_VMBUS_PACKET_TYPE_DATA_IN_BAND,
 838                     HV_VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
 839         }
 840
 841         return (ret);
 842 }
 843
 844 /*
 845  * Net VSC on receive
 846  *
 847  * In the FreeBSD Hyper-V virtual world, this function deals exclusively
 848  * with virtual addresses.
 849  */
 850 static void
 851 hv_nv_on_receive(netvsc_dev *net_dev, struct hv_device *device,
 852     struct hv_vmbus_channel *chan, hv_vm_packet_descriptor *pkt)
 853 {
 854         hv_vm_transfer_page_packet_header *vm_xfer_page_pkt;
 855         nvsp_msg *nvsp_msg_pkt;
 856         netvsc_packet vsc_pkt;
 857         netvsc_packet *net_vsc_pkt = &vsc_pkt;
 858         device_t dev = device->device;
 859         int count = 0;
 860         int i = 0;
 861         int status = nvsp_status_success;
 862
 863         /*
 864          * All inbound packets other than send completion should be
 865          * xfer page packet.
 866          */
 867         if (pkt->type != HV_VMBUS_PACKET_TYPE_DATA_USING_TRANSFER_PAGES) {
 868                 device_printf(dev, "packet type %d is invalid!\n", pkt->type);
 869                 return;
 870         }
 871
 872         nvsp_msg_pkt = (nvsp_msg *)((unsigned long)pkt
 873                 + (pkt->data_offset8 << 3));
 874
 875         /* Make sure this is a valid nvsp packet */
 876         if (nvsp_msg_pkt->hdr.msg_type != nvsp_msg_1_type_send_rndis_pkt) {
 877                 device_printf(dev, "packet hdr type %d is invalid!\n",
 878                     pkt->type);
 879                 return;
 880         }
 881
 882         vm_xfer_page_pkt = (hv_vm_transfer_page_packet_header *)pkt;
 883
 884         if (vm_xfer_page_pkt->transfer_page_set_id !=
 885             NETVSC_RECEIVE_BUFFER_ID) {
 886                 device_printf(dev, "transfer_page_set_id %d is invalid!\n",
 887                     vm_xfer_page_pkt->transfer_page_set_id);
 888                 return;
 889         }
 890
 891         count = vm_xfer_page_pkt->range_count;
 892         net_vsc_pkt->device = device;
 893
 894         /* Each range represents 1 RNDIS pkt that contains 1 Ethernet frame */
 895         for (i = 0; i < count; i++) {
 896                 net_vsc_pkt->status = nvsp_status_success;
 897                 net_vsc_pkt->data = (void *)((unsigned long)net_dev->rx_buf +
 898                     vm_xfer_page_pkt->ranges[i].byte_offset);
 899                 net_vsc_pkt->tot_data_buf_len =
 900                     vm_xfer_page_pkt->ranges[i].byte_count;
 901
 902                 hv_rf_on_receive(net_dev, device, chan, net_vsc_pkt);
 903                 if (net_vsc_pkt->status != nvsp_status_success) {
 904                         status = nvsp_status_failure;
 905                 }
 906         }
 907
 908         /*
 909          * Moved completion call back here so that all received
 910          * messages (not just data messages) will trigger a response
 911          * message back to the host.
 912          */
 913         hv_nv_on_receive_completion(device, vm_xfer_page_pkt->d.transaction_id,
 914             status);
 915 }
 916
 917 /*
 918  * Net VSC on receive completion
 919  *
 920  * Send a receive completion packet to RNDIS device (ie NetVsp)
 921  */
 922 void
 923 hv_nv_on_receive_completion(struct hv_device *device, uint64_t tid,
 924     uint32_t status)
 925 {
 926         nvsp_msg rx_comp_msg;
 927         int retries = 0;
 928         int ret = 0;
 929
 930         rx_comp_msg.hdr.msg_type = nvsp_msg_1_type_send_rndis_pkt_complete;
 931
 932         /* Pass in the status */
 933         rx_comp_msg.msgs.vers_1_msgs.send_rndis_pkt_complete.status =
 934             status;
 935
 936 retry_send_cmplt:
 937         /* Send the completion */
 938         ret = hv_vmbus_channel_send_packet(device->channel, &rx_comp_msg,
 939             sizeof(nvsp_msg), tid, HV_VMBUS_PACKET_TYPE_COMPLETION, 0);
 940         if (ret == 0) {
 941                 /* success */
 942                 /* no-op */
 943         } else if (ret == EAGAIN) {
 944                 /* no more room... wait a bit and attempt to retry 3 times */
 945                 retries++;
 946
 947                 if (retries < 4) {
 948                         DELAY(100);
 949                         goto retry_send_cmplt;
 950                 }
 951         }
 952 }
 953
 954 /*
 955  * Net VSC on channel callback
 956  */
 957 static void
 958 hv_nv_on_channel_callback(void *xchan)
 959 {
 960         struct hv_vmbus_channel *chan = xchan;
 961         struct hv_device *device = chan->device;
 962         netvsc_dev *net_dev;
 963         device_t dev = device->device;
 964         uint32_t bytes_rxed;
 965         uint64_t request_id;
 966         hv_vm_packet_descriptor *desc;
 967         uint8_t *buffer;
 968         int bufferlen = NETVSC_PACKET_SIZE;
 969         int ret = 0;
 970
 971         net_dev = hv_nv_get_inbound_net_device(device);
 972         if (net_dev == NULL)
 973                 return;
 974
 975         buffer = net_dev->callback_buf;
 976
 977         do {
 978                 ret = hv_vmbus_channel_recv_packet_raw(chan,
 979                     buffer, bufferlen, &bytes_rxed, &request_id);
 980                 if (ret == 0) {
 981                         if (bytes_rxed > 0) {
 982                                 desc = (hv_vm_packet_descriptor *)buffer;
 983                                 switch (desc->type) {
 984                                 case HV_VMBUS_PACKET_TYPE_COMPLETION:
 985                                         hv_nv_on_send_completion(net_dev, device, desc);
 986                                         break;
 987                                 case HV_VMBUS_PACKET_TYPE_DATA_USING_TRANSFER_PAGES:
 988                                         hv_nv_on_receive(net_dev, device, chan, desc);
 989                                         break;
 990                                 default:
 991                                         device_printf(dev,
 992                                             "hv_cb recv unknow type %d "
 993                                             " packet\n", desc->type);
 994                                         break;
 995                                 }
 996                         } else {
 997                                 break;
 998                         }
 999                 } else if (ret == ENOBUFS) {
1000                         /* Handle large packet */
1001                         if (bufferlen > NETVSC_PACKET_SIZE) {
1002                                 free(buffer, M_NETVSC);
1003                                 buffer = NULL;
1004                         }
1005
1006                         /* alloc new buffer */
1007                         buffer = malloc(bytes_rxed, M_NETVSC, M_NOWAIT);
1008                         if (buffer == NULL) {
1009                                 device_printf(dev,
1010                                     "hv_cb malloc buffer failed, len=%u\n",
1011                                     bytes_rxed);
1012                                 bufferlen = 0;
1013                                 break;
1014                         }
1015                         bufferlen = bytes_rxed;
1016                 }
1017         } while (1);
1018
1019         if (bufferlen > NETVSC_PACKET_SIZE)
1020                 free(buffer, M_NETVSC);
1021
1022         hv_rf_channel_rollup(chan);
1023 }