sys/dev/hyperv/vmbus/hv_hv.c

   1 /*-
   2  * Copyright (c) 2009-2012 Microsoft Corp.
   3  * Copyright (c) 2012 NetApp Inc.
   4  * Copyright (c) 2012 Citrix 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
  29 /**
  30  * Implements low-level interactions with Hypver-V/Azure
  31  */
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/kernel.h>
  37 #include <sys/malloc.h>
  38 #include <sys/pcpu.h>
  39 #include <sys/timetc.h>
  40 #include <machine/bus.h>
  41 #include <machine/md_var.h>
  42 #include <vm/vm.h>
  43 #include <vm/vm_param.h>
  44 #include <vm/pmap.h>
  45
  46
  47 #include "hv_vmbus_priv.h"
  48
  49 #define HV_NANOSECONDS_PER_SEC          1000000000L
  50
  51
  52 static u_int hv_get_timecount(struct timecounter *tc);
  53
  54 u_int   hyperv_features;
  55 u_int   hyperv_recommends;
  56
  57 /**
  58  * Globals
  59  */
  60 hv_vmbus_context hv_vmbus_g_context = {
  61         .syn_ic_initialized = FALSE,
  62         .hypercall_page = NULL,
  63 };
  64
  65 static struct timecounter hv_timecounter = {
  66         hv_get_timecount, 0, ~0u, HV_NANOSECONDS_PER_SEC/100, "Hyper-V", HV_NANOSECONDS_PER_SEC/100
  67 };
  68
  69 static u_int
  70 hv_get_timecount(struct timecounter *tc)
  71 {
  72         u_int now = rdmsr(HV_X64_MSR_TIME_REF_COUNT);
  73         return (now);
  74 }
  75
  76 /**
  77  * @brief Query the cpuid for presence of windows hypervisor
  78  */
  79 int
  80 hv_vmbus_query_hypervisor_presence(void)
  81 {
  82         if (vm_guest != VM_GUEST_HV)
  83                 return (0);
  84
  85         return (hv_high >= HV_X64_CPUID_MIN && hv_high <= HV_X64_CPUID_MAX);
  86 }
  87
  88 /**
  89  * @brief Get version of the windows hypervisor
  90  */
  91 static int
  92 hv_vmbus_get_hypervisor_version(void)
  93 {
  94         u_int regs[4];
  95         unsigned int maxLeaf;
  96         unsigned int op;
  97
  98         /*
  99          * Its assumed that this is called after confirming that
 100          * Viridian is present
 101          * Query id and revision.
 102          */
 103         op = HV_CPU_ID_FUNCTION_HV_VENDOR_AND_MAX_FUNCTION;
 104         do_cpuid(op, regs);
 105
 106         maxLeaf = regs[0];
 107         op = HV_CPU_ID_FUNCTION_HV_INTERFACE;
 108         do_cpuid(op, regs);
 109
 110         if (maxLeaf >= HV_CPU_ID_FUNCTION_MS_HV_VERSION) {
 111             op = HV_CPU_ID_FUNCTION_MS_HV_VERSION;
 112             do_cpuid(op, regs);
 113         }
 114         return (maxLeaf);
 115 }
 116
 117 /**
 118  * @brief Invoke the specified hypercall
 119  */
 120 static uint64_t
 121 hv_vmbus_do_hypercall(uint64_t control, void* input, void* output)
 122 {
 123 #ifdef __x86_64__
 124         uint64_t hv_status = 0;
 125         uint64_t input_address = (input) ? hv_get_phys_addr(input) : 0;
 126         uint64_t output_address = (output) ? hv_get_phys_addr(output) : 0;
 127         volatile void* hypercall_page = hv_vmbus_g_context.hypercall_page;
 128
 129         __asm__ __volatile__ ("mov %0, %%r8" : : "r" (output_address): "r8");
 130         __asm__ __volatile__ ("call *%3" : "=a"(hv_status):
 131                                 "c" (control), "d" (input_address),
 132                                 "m" (hypercall_page));
 133         return (hv_status);
 134 #else
 135         uint32_t control_high = control >> 32;
 136         uint32_t control_low = control & 0xFFFFFFFF;
 137         uint32_t hv_status_high = 1;
 138         uint32_t hv_status_low = 1;
 139         uint64_t input_address = (input) ? hv_get_phys_addr(input) : 0;
 140         uint32_t input_address_high = input_address >> 32;
 141         uint32_t input_address_low = input_address & 0xFFFFFFFF;
 142         uint64_t output_address = (output) ? hv_get_phys_addr(output) : 0;
 143         uint32_t output_address_high = output_address >> 32;
 144         uint32_t output_address_low = output_address & 0xFFFFFFFF;
 145         volatile void* hypercall_page = hv_vmbus_g_context.hypercall_page;
 146
 147         __asm__ __volatile__ ("call *%8" : "=d"(hv_status_high),
 148                                 "=a"(hv_status_low) : "d" (control_high),
 149                                 "a" (control_low), "b" (input_address_high),
 150                                 "c" (input_address_low),
 151                                 "D"(output_address_high),
 152                                 "S"(output_address_low), "m" (hypercall_page));
 153         return (hv_status_low | ((uint64_t)hv_status_high << 32));
 154 #endif /* __x86_64__ */
 155 }
 156
 157 /**
 158  *  @brief Main initialization routine.
 159  *
 160  *  This routine must be called
 161  *  before any other routines in here are called
 162  */
 163 int
 164 hv_vmbus_init(void)
 165 {
 166         int                                     max_leaf;
 167         hv_vmbus_x64_msr_hypercall_contents     hypercall_msr;
 168         void*                                   virt_addr = 0;
 169
 170         memset(
 171             hv_vmbus_g_context.syn_ic_event_page,
 172             0,
 173             sizeof(hv_vmbus_handle) * MAXCPU);
 174
 175         memset(
 176             hv_vmbus_g_context.syn_ic_msg_page,
 177             0,
 178             sizeof(hv_vmbus_handle) * MAXCPU);
 179
 180         if (vm_guest != VM_GUEST_HV)
 181             goto cleanup;
 182
 183         max_leaf = hv_vmbus_get_hypervisor_version();
 184
 185         /*
 186          * Write our OS info
 187          */
 188         uint64_t os_guest_info = HV_FREEBSD_GUEST_ID;
 189         wrmsr(HV_X64_MSR_GUEST_OS_ID, os_guest_info);
 190         hv_vmbus_g_context.guest_id = os_guest_info;
 191
 192         /*
 193          * See if the hypercall page is already set
 194          */
 195         hypercall_msr.as_uint64_t = rdmsr(HV_X64_MSR_HYPERCALL);
 196         virt_addr = malloc(PAGE_SIZE, M_DEVBUF, M_NOWAIT | M_ZERO);
 197         KASSERT(virt_addr != NULL,
 198             ("Error VMBUS: malloc failed to allocate page during init!"));
 199         if (virt_addr == NULL)
 200             goto cleanup;
 201
 202         hypercall_msr.u.enable = 1;
 203         hypercall_msr.u.guest_physical_address =
 204             (hv_get_phys_addr(virt_addr) >> PAGE_SHIFT);
 205         wrmsr(HV_X64_MSR_HYPERCALL, hypercall_msr.as_uint64_t);
 206
 207         /*
 208          * Confirm that hypercall page did get set up
 209          */
 210         hypercall_msr.as_uint64_t = 0;
 211         hypercall_msr.as_uint64_t = rdmsr(HV_X64_MSR_HYPERCALL);
 212
 213         if (!hypercall_msr.u.enable)
 214             goto cleanup;
 215
 216         hv_vmbus_g_context.hypercall_page = virt_addr;
 217
 218         hv_et_init();
 219
 220         return (0);
 221
 222         cleanup:
 223         if (virt_addr != NULL) {
 224             if (hypercall_msr.u.enable) {
 225                 hypercall_msr.as_uint64_t = 0;
 226                 wrmsr(HV_X64_MSR_HYPERCALL,
 227                                         hypercall_msr.as_uint64_t);
 228             }
 229
 230             free(virt_addr, M_DEVBUF);
 231         }
 232         return (ENOTSUP);
 233 }
 234
 235 /**
 236  * @brief Cleanup routine, called normally during driver unloading or exiting
 237  */
 238 void
 239 hv_vmbus_cleanup(void)
 240 {
 241         hv_vmbus_x64_msr_hypercall_contents hypercall_msr;
 242
 243         if (hv_vmbus_g_context.guest_id == HV_FREEBSD_GUEST_ID) {
 244             if (hv_vmbus_g_context.hypercall_page != NULL) {
 245                 hypercall_msr.as_uint64_t = 0;
 246                 wrmsr(HV_X64_MSR_HYPERCALL,
 247                                         hypercall_msr.as_uint64_t);
 248                 free(hv_vmbus_g_context.hypercall_page, M_DEVBUF);
 249                 hv_vmbus_g_context.hypercall_page = NULL;
 250             }
 251         }
 252 }
 253
 254 /**
 255  * @brief Post a message using the hypervisor message IPC.
 256  * (This involves a hypercall.)
 257  */
 258 hv_vmbus_status
 259 hv_vmbus_post_msg_via_msg_ipc(
 260         hv_vmbus_connection_id  connection_id,
 261         hv_vmbus_msg_type       message_type,
 262         void*                   payload,
 263         size_t                  payload_size)
 264 {
 265         struct alignedinput {
 266             uint64_t alignment8;
 267             hv_vmbus_input_post_message msg;
 268         };
 269
 270         hv_vmbus_input_post_message*    aligned_msg;
 271         hv_vmbus_status                 status;
 272         size_t                          addr;
 273
 274         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
 275             return (EMSGSIZE);
 276
 277         addr = (size_t) malloc(sizeof(struct alignedinput), M_DEVBUF,
 278                             M_ZERO | M_NOWAIT);
 279         KASSERT(addr != 0,
 280             ("Error VMBUS: malloc failed to allocate message buffer!"));
 281         if (addr == 0)
 282             return (ENOMEM);
 283
 284         aligned_msg = (hv_vmbus_input_post_message*)
 285             (HV_ALIGN_UP(addr, HV_HYPERCALL_PARAM_ALIGN));
 286
 287         aligned_msg->connection_id = connection_id;
 288         aligned_msg->message_type = message_type;
 289         aligned_msg->payload_size = payload_size;
 290         memcpy((void*) aligned_msg->payload, payload, payload_size);
 291
 292         status = hv_vmbus_do_hypercall(
 293                     HV_CALL_POST_MESSAGE, aligned_msg, 0) & 0xFFFF;
 294
 295         free((void *) addr, M_DEVBUF);
 296         return (status);
 297 }
 298
 299 /**
 300  * @brief Signal an event on the specified connection using the hypervisor
 301  * event IPC. (This involves a hypercall.)
 302  */
 303 hv_vmbus_status
 304 hv_vmbus_signal_event(void *con_id)
 305 {
 306         hv_vmbus_status status;
 307
 308         status = hv_vmbus_do_hypercall(
 309                     HV_CALL_SIGNAL_EVENT,
 310                     con_id,
 311                     0) & 0xFFFF;
 312
 313         return (status);
 314 }
 315
 316 /**
 317  * @brief hv_vmbus_synic_init
 318  */
 319 void
 320 hv_vmbus_synic_init(void *arg)
 321
 322 {
 323         int                     cpu;
 324         uint64_t                hv_vcpu_index;
 325         hv_vmbus_synic_simp     simp;
 326         hv_vmbus_synic_siefp    siefp;
 327         hv_vmbus_synic_scontrol sctrl;
 328         hv_vmbus_synic_sint     shared_sint;
 329         uint64_t                version;
 330         hv_setup_args*          setup_args = (hv_setup_args *)arg;
 331
 332         cpu = PCPU_GET(cpuid);
 333
 334         if (hv_vmbus_g_context.hypercall_page == NULL)
 335             return;
 336
 337         /*
 338          * TODO: Check the version
 339          */
 340         version = rdmsr(HV_X64_MSR_SVERSION);
 341
 342         hv_vmbus_g_context.syn_ic_msg_page[cpu] =
 343             setup_args->page_buffers[2 * cpu];
 344         hv_vmbus_g_context.syn_ic_event_page[cpu] =
 345             setup_args->page_buffers[2 * cpu + 1];
 346
 347         /*
 348          * Setup the Synic's message page
 349          */
 350
 351         simp.as_uint64_t = rdmsr(HV_X64_MSR_SIMP);
 352         simp.u.simp_enabled = 1;
 353         simp.u.base_simp_gpa = ((hv_get_phys_addr(
 354             hv_vmbus_g_context.syn_ic_msg_page[cpu])) >> PAGE_SHIFT);
 355
 356         wrmsr(HV_X64_MSR_SIMP, simp.as_uint64_t);
 357
 358         /*
 359          * Setup the Synic's event page
 360          */
 361         siefp.as_uint64_t = rdmsr(HV_X64_MSR_SIEFP);
 362         siefp.u.siefp_enabled = 1;
 363         siefp.u.base_siefp_gpa = ((hv_get_phys_addr(
 364             hv_vmbus_g_context.syn_ic_event_page[cpu])) >> PAGE_SHIFT);
 365
 366         wrmsr(HV_X64_MSR_SIEFP, siefp.as_uint64_t);
 367
 368         /*HV_SHARED_SINT_IDT_VECTOR + 0x20; */
 369         shared_sint.as_uint64_t = 0;
 370         shared_sint.u.vector = setup_args->vector;
 371         shared_sint.u.masked = FALSE;
 372         shared_sint.u.auto_eoi = TRUE;
 373
 374         wrmsr(HV_X64_MSR_SINT0 + HV_VMBUS_MESSAGE_SINT,
 375             shared_sint.as_uint64_t);
 376
 377         /* Enable the global synic bit */
 378         sctrl.as_uint64_t = rdmsr(HV_X64_MSR_SCONTROL);
 379         sctrl.u.enable = 1;
 380
 381         wrmsr(HV_X64_MSR_SCONTROL, sctrl.as_uint64_t);
 382
 383         hv_vmbus_g_context.syn_ic_initialized = TRUE;
 384
 385         /*
 386          * Set up the cpuid mapping from Hyper-V to FreeBSD.
 387          * The array is indexed using FreeBSD cpuid.
 388          */
 389         hv_vcpu_index = rdmsr(HV_X64_MSR_VP_INDEX);
 390         hv_vmbus_g_context.hv_vcpu_index[cpu] = (uint32_t)hv_vcpu_index;
 391
 392         return;
 393 }
 394
 395 /**
 396  * @brief Cleanup routine for hv_vmbus_synic_init()
 397  */
 398 void hv_vmbus_synic_cleanup(void *arg)
 399 {
 400         hv_vmbus_synic_sint     shared_sint;
 401         hv_vmbus_synic_simp     simp;
 402         hv_vmbus_synic_siefp    siefp;
 403
 404         if (!hv_vmbus_g_context.syn_ic_initialized)
 405             return;
 406
 407         shared_sint.as_uint64_t = rdmsr(
 408             HV_X64_MSR_SINT0 + HV_VMBUS_MESSAGE_SINT);
 409
 410         shared_sint.u.masked = 1;
 411
 412         /*
 413          * Disable the interrupt
 414          */
 415         wrmsr(
 416             HV_X64_MSR_SINT0 + HV_VMBUS_MESSAGE_SINT,
 417             shared_sint.as_uint64_t);
 418
 419         simp.as_uint64_t = rdmsr(HV_X64_MSR_SIMP);
 420         simp.u.simp_enabled = 0;
 421         simp.u.base_simp_gpa = 0;
 422
 423         wrmsr(HV_X64_MSR_SIMP, simp.as_uint64_t);
 424
 425         siefp.as_uint64_t = rdmsr(HV_X64_MSR_SIEFP);
 426         siefp.u.siefp_enabled = 0;
 427         siefp.u.base_siefp_gpa = 0;
 428
 429         wrmsr(HV_X64_MSR_SIEFP, siefp.as_uint64_t);
 430 }
 431
 432 static bool
 433 hyperv_identify(void)
 434 {
 435         u_int regs[4];
 436         unsigned int maxLeaf;
 437         unsigned int op;
 438
 439         if (vm_guest != VM_GUEST_HV)
 440                 return (false);
 441
 442         op = HV_CPU_ID_FUNCTION_HV_VENDOR_AND_MAX_FUNCTION;
 443         do_cpuid(op, regs);
 444         maxLeaf = regs[0];
 445         if (maxLeaf < HV_CPU_ID_FUNCTION_MS_HV_IMPLEMENTATION_LIMITS)
 446                 return (false);
 447
 448         op = HV_CPU_ID_FUNCTION_HV_INTERFACE;
 449         do_cpuid(op, regs);
 450         if (regs[0] != 0x31237648 /* HV#1 */)
 451                 return (false);
 452
 453         op = HV_CPU_ID_FUNCTION_MS_HV_FEATURES;
 454         do_cpuid(op, regs);
 455         if ((regs[0] & HV_FEATURE_MSR_HYPERCALL) == 0) {
 456                 /*
 457                  * Hyper-V w/o Hypercall is impossible; someone
 458                  * is faking Hyper-V.
 459                  */
 460                 return (false);
 461         }
 462         hyperv_features = regs[0];
 463
 464         op = HV_CPU_ID_FUNCTION_MS_HV_VERSION;
 465         do_cpuid(op, regs);
 466         printf("Hyper-V Version: %d.%d.%d [SP%d]\n",
 467             regs[1] >> 16, regs[1] & 0xffff, regs[0], regs[2]);
 468
 469         printf("  Features: 0x%b\n", hyperv_features,
 470             "\020"
 471             "\001VPRUNTIME"
 472             "\002TMREFCNT"
 473             "\003SYNCIC"
 474             "\004SYNCTM"
 475             "\005APIC"
 476             "\006HYERCALL"
 477             "\007VPINDEX"
 478             "\010RESET"
 479             "\011STATS"
 480             "\012REFTSC"
 481             "\013IDLE"
 482             "\014TMFREQ"
 483             "\015DEBUG");
 484
 485         op = HV_CPU_ID_FUNCTION_MS_HV_ENLIGHTENMENT_INFORMATION;
 486         do_cpuid(op, regs);
 487         hyperv_recommends = regs[0];
 488         if (bootverbose)
 489                 printf("  Recommends: %08x %08x\n", regs[0], regs[1]);
 490
 491         op = HV_CPU_ID_FUNCTION_MS_HV_IMPLEMENTATION_LIMITS;
 492         do_cpuid(op, regs);
 493         if (bootverbose) {
 494                 printf("  Limits: Vcpu:%d Lcpu:%d Int:%d\n",
 495                     regs[0], regs[1], regs[2]);
 496         }
 497
 498         if (maxLeaf >= HV_CPU_ID_FUNCTION_MS_HV_HARDWARE_FEATURE) {
 499                 op = HV_CPU_ID_FUNCTION_MS_HV_HARDWARE_FEATURE;
 500                 do_cpuid(op, regs);
 501                 if (bootverbose) {
 502                         printf("  HW Features: %08x AMD: %08x\n",
 503                             regs[0], regs[3]);
 504                 }
 505         }
 506
 507         return (true);
 508 }
 509
 510 static void
 511 hyperv_init(void *dummy __unused)
 512 {
 513         if (!hyperv_identify())
 514                 return;
 515
 516         if (hyperv_features & HV_FEATURE_MSR_TIME_REFCNT) {
 517                 /* Register virtual timecount */
 518                 tc_init(&hv_timecounter);
 519         }
 520 }
 521 SYSINIT(hyperv_initialize, SI_SUB_HYPERVISOR, SI_ORDER_FIRST, hyperv_init, NULL);