sys/dev/hyperv/vmbus/hyperv.c

   1 /*-
   2  * Copyright (c) 2009-2012,2016 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 #include <dev/hyperv/include/hyperv_busdma.h>
  47 #include <dev/hyperv/vmbus/hv_vmbus_priv.h>
  48 #include <dev/hyperv/vmbus/hyperv_reg.h>
  49 #include <dev/hyperv/vmbus/hyperv_var.h>
  50 #include <dev/hyperv/vmbus/vmbus_var.h>
  51
  52 #define HYPERV_FREEBSD_BUILD            0ULL
  53 #define HYPERV_FREEBSD_VERSION          ((uint64_t)__FreeBSD_version)
  54 #define HYPERV_FREEBSD_OSID             0ULL
  55
  56 #define MSR_HV_GUESTID_BUILD_FREEBSD    \
  57         (HYPERV_FREEBSD_BUILD & MSR_HV_GUESTID_BUILD_MASK)
  58 #define MSR_HV_GUESTID_VERSION_FREEBSD  \
  59         ((HYPERV_FREEBSD_VERSION << MSR_HV_GUESTID_VERSION_SHIFT) & \
  60          MSR_HV_GUESTID_VERSION_MASK)
  61 #define MSR_HV_GUESTID_OSID_FREEBSD     \
  62         ((HYPERV_FREEBSD_OSID << MSR_HV_GUESTID_OSID_SHIFT) & \
  63          MSR_HV_GUESTID_OSID_MASK)
  64
  65 #define MSR_HV_GUESTID_FREEBSD          \
  66         (MSR_HV_GUESTID_BUILD_FREEBSD | \
  67          MSR_HV_GUESTID_VERSION_FREEBSD | \
  68          MSR_HV_GUESTID_OSID_FREEBSD |  \
  69          MSR_HV_GUESTID_OSTYPE_FREEBSD)
  70
  71 struct hypercall_ctx {
  72         void                    *hc_addr;
  73         struct hyperv_dma       hc_dma;
  74 };
  75
  76 static u_int    hyperv_get_timecount(struct timecounter *tc);
  77
  78 u_int           hyperv_features;
  79 u_int           hyperv_recommends;
  80
  81 static u_int    hyperv_pm_features;
  82 static u_int    hyperv_features3;
  83
  84 static struct timecounter       hyperv_timecounter = {
  85         .tc_get_timecount       = hyperv_get_timecount,
  86         .tc_poll_pps            = NULL,
  87         .tc_counter_mask        = 0xffffffff,
  88         .tc_frequency           = HYPERV_TIMER_FREQ,
  89         .tc_name                = "Hyper-V",
  90         .tc_quality             = 2000,
  91         .tc_flags               = 0,
  92         .tc_priv                = NULL
  93 };
  94
  95 static struct hypercall_ctx     hypercall_context;
  96
  97 static u_int
  98 hyperv_get_timecount(struct timecounter *tc __unused)
  99 {
 100         return rdmsr(MSR_HV_TIME_REF_COUNT);
 101 }
 102
 103 /**
 104  * @brief Invoke the specified hypercall
 105  */
 106 static uint64_t
 107 hv_vmbus_do_hypercall(uint64_t control, void* input, void* output)
 108 {
 109 #ifdef __x86_64__
 110         uint64_t hv_status = 0;
 111         uint64_t input_address = (input) ? hv_get_phys_addr(input) : 0;
 112         uint64_t output_address = (output) ? hv_get_phys_addr(output) : 0;
 113         volatile void *hypercall_page = hypercall_context.hc_addr;
 114
 115         __asm__ __volatile__ ("mov %0, %%r8" : : "r" (output_address): "r8");
 116         __asm__ __volatile__ ("call *%3" : "=a"(hv_status):
 117                                 "c" (control), "d" (input_address),
 118                                 "m" (hypercall_page));
 119         return (hv_status);
 120 #else
 121         uint32_t control_high = control >> 32;
 122         uint32_t control_low = control & 0xFFFFFFFF;
 123         uint32_t hv_status_high = 1;
 124         uint32_t hv_status_low = 1;
 125         uint64_t input_address = (input) ? hv_get_phys_addr(input) : 0;
 126         uint32_t input_address_high = input_address >> 32;
 127         uint32_t input_address_low = input_address & 0xFFFFFFFF;
 128         uint64_t output_address = (output) ? hv_get_phys_addr(output) : 0;
 129         uint32_t output_address_high = output_address >> 32;
 130         uint32_t output_address_low = output_address & 0xFFFFFFFF;
 131         volatile void *hypercall_page = hypercall_context.hc_addr;
 132
 133         __asm__ __volatile__ ("call *%8" : "=d"(hv_status_high),
 134                                 "=a"(hv_status_low) : "d" (control_high),
 135                                 "a" (control_low), "b" (input_address_high),
 136                                 "c" (input_address_low),
 137                                 "D"(output_address_high),
 138                                 "S"(output_address_low), "m" (hypercall_page));
 139         return (hv_status_low | ((uint64_t)hv_status_high << 32));
 140 #endif /* __x86_64__ */
 141 }
 142
 143 /**
 144  * @brief Post a message using the hypervisor message IPC.
 145  * (This involves a hypercall.)
 146  */
 147 hv_vmbus_status
 148 hv_vmbus_post_msg_via_msg_ipc(
 149         hv_vmbus_connection_id  connection_id,
 150         hv_vmbus_msg_type       message_type,
 151         void*                   payload,
 152         size_t                  payload_size)
 153 {
 154         struct alignedinput {
 155             uint64_t alignment8;
 156             hv_vmbus_input_post_message msg;
 157         };
 158
 159         hv_vmbus_input_post_message*    aligned_msg;
 160         hv_vmbus_status                 status;
 161         size_t                          addr;
 162
 163         if (payload_size > HV_MESSAGE_PAYLOAD_BYTE_COUNT)
 164             return (EMSGSIZE);
 165
 166         addr = (size_t) malloc(sizeof(struct alignedinput), M_DEVBUF,
 167                             M_ZERO | M_NOWAIT);
 168         KASSERT(addr != 0,
 169             ("Error VMBUS: malloc failed to allocate message buffer!"));
 170         if (addr == 0)
 171             return (ENOMEM);
 172
 173         aligned_msg = (hv_vmbus_input_post_message*)
 174             (HV_ALIGN_UP(addr, HV_HYPERCALL_PARAM_ALIGN));
 175
 176         aligned_msg->connection_id = connection_id;
 177         aligned_msg->message_type = message_type;
 178         aligned_msg->payload_size = payload_size;
 179         memcpy((void*) aligned_msg->payload, payload, payload_size);
 180
 181         status = hv_vmbus_do_hypercall(
 182                     HV_CALL_POST_MESSAGE, aligned_msg, 0) & 0xFFFF;
 183
 184         free((void *) addr, M_DEVBUF);
 185         return (status);
 186 }
 187
 188 /**
 189  * @brief Signal an event on the specified connection using the hypervisor
 190  * event IPC. (This involves a hypercall.)
 191  */
 192 hv_vmbus_status
 193 hv_vmbus_signal_event(void *con_id)
 194 {
 195         hv_vmbus_status status;
 196
 197         status = hv_vmbus_do_hypercall(
 198                     HV_CALL_SIGNAL_EVENT,
 199                     con_id,
 200                     0) & 0xFFFF;
 201
 202         return (status);
 203 }
 204
 205 int
 206 hyperv_guid2str(const struct hv_guid *guid, char *buf, size_t sz)
 207 {
 208         const uint8_t *d = guid->data;
 209
 210         return snprintf(buf, sz, "%02x%02x%02x%02x-"
 211             "%02x%02x-%02x%02x-%02x%02x-"
 212             "%02x%02x%02x%02x%02x%02x",
 213             d[3], d[2], d[1], d[0],
 214             d[5], d[4], d[7], d[6], d[8], d[9],
 215             d[10], d[11], d[12], d[13], d[14], d[15]);
 216 }
 217
 218 static bool
 219 hyperv_identify(void)
 220 {
 221         u_int regs[4];
 222         unsigned int maxleaf;
 223
 224         if (vm_guest != VM_GUEST_HV)
 225                 return (false);
 226
 227         do_cpuid(CPUID_LEAF_HV_MAXLEAF, regs);
 228         maxleaf = regs[0];
 229         if (maxleaf < CPUID_LEAF_HV_LIMITS)
 230                 return (false);
 231
 232         do_cpuid(CPUID_LEAF_HV_INTERFACE, regs);
 233         if (regs[0] != CPUID_HV_IFACE_HYPERV)
 234                 return (false);
 235
 236         do_cpuid(CPUID_LEAF_HV_FEATURES, regs);
 237         if ((regs[0] & CPUID_HV_MSR_HYPERCALL) == 0) {
 238                 /*
 239                  * Hyper-V w/o Hypercall is impossible; someone
 240                  * is faking Hyper-V.
 241                  */
 242                 return (false);
 243         }
 244         hyperv_features = regs[0];
 245         hyperv_pm_features = regs[2];
 246         hyperv_features3 = regs[3];
 247
 248         do_cpuid(CPUID_LEAF_HV_IDENTITY, regs);
 249         printf("Hyper-V Version: %d.%d.%d [SP%d]\n",
 250             regs[1] >> 16, regs[1] & 0xffff, regs[0], regs[2]);
 251
 252         printf("  Features=0x%b\n", hyperv_features,
 253             "\020"
 254             "\001VPRUNTIME"     /* MSR_HV_VP_RUNTIME */
 255             "\002TMREFCNT"      /* MSR_HV_TIME_REF_COUNT */
 256             "\003SYNIC"         /* MSRs for SynIC */
 257             "\004SYNTM"         /* MSRs for SynTimer */
 258             "\005APIC"          /* MSR_HV_{EOI,ICR,TPR} */
 259             "\006HYPERCALL"     /* MSR_HV_{GUEST_OS_ID,HYPERCALL} */
 260             "\007VPINDEX"       /* MSR_HV_VP_INDEX */
 261             "\010RESET"         /* MSR_HV_RESET */
 262             "\011STATS"         /* MSR_HV_STATS_ */
 263             "\012REFTSC"        /* MSR_HV_REFERENCE_TSC */
 264             "\013IDLE"          /* MSR_HV_GUEST_IDLE */
 265             "\014TMFREQ"        /* MSR_HV_{TSC,APIC}_FREQUENCY */
 266             "\015DEBUG");       /* MSR_HV_SYNTH_DEBUG_ */
 267         printf("  PM Features=0x%b [C%u]\n",
 268             (hyperv_pm_features & ~CPUPM_HV_CSTATE_MASK),
 269             "\020"
 270             "\005C3HPET",       /* HPET is required for C3 state */
 271             CPUPM_HV_CSTATE(hyperv_pm_features));
 272         printf("  Features3=0x%b\n", hyperv_features3,
 273             "\020"
 274             "\001MWAIT"         /* MWAIT */
 275             "\002DEBUG"         /* guest debug support */
 276             "\003PERFMON"       /* performance monitor */
 277             "\004PCPUDPE"       /* physical CPU dynamic partition event */
 278             "\005XMMHC"         /* hypercall input through XMM regs */
 279             "\006IDLE"          /* guest idle support */
 280             "\007SLEEP"         /* hypervisor sleep support */
 281             "\010NUMA"          /* NUMA distance query support */
 282             "\011TMFREQ"        /* timer frequency query (TSC, LAPIC) */
 283             "\012SYNCMC"        /* inject synthetic machine checks */
 284             "\013CRASH"         /* MSRs for guest crash */
 285             "\014DEBUGMSR"      /* MSRs for guest debug */
 286             "\015NPIEP"         /* NPIEP */
 287             "\016HVDIS");       /* disabling hypervisor */
 288
 289         do_cpuid(CPUID_LEAF_HV_RECOMMENDS, regs);
 290         hyperv_recommends = regs[0];
 291         if (bootverbose)
 292                 printf("  Recommends: %08x %08x\n", regs[0], regs[1]);
 293
 294         do_cpuid(CPUID_LEAF_HV_LIMITS, regs);
 295         if (bootverbose) {
 296                 printf("  Limits: Vcpu:%d Lcpu:%d Int:%d\n",
 297                     regs[0], regs[1], regs[2]);
 298         }
 299
 300         if (maxleaf >= CPUID_LEAF_HV_HWFEATURES) {
 301                 do_cpuid(CPUID_LEAF_HV_HWFEATURES, regs);
 302                 if (bootverbose) {
 303                         printf("  HW Features: %08x, AMD: %08x\n",
 304                             regs[0], regs[3]);
 305                 }
 306         }
 307
 308         return (true);
 309 }
 310
 311 static void
 312 hyperv_init(void *dummy __unused)
 313 {
 314         if (!hyperv_identify()) {
 315                 /* Not Hyper-V; reset guest id to the generic one. */
 316                 if (vm_guest == VM_GUEST_HV)
 317                         vm_guest = VM_GUEST_VM;
 318                 return;
 319         }
 320
 321         /* Set guest id */
 322         wrmsr(MSR_HV_GUEST_OS_ID, MSR_HV_GUESTID_FREEBSD);
 323
 324         if (hyperv_features & CPUID_HV_MSR_TIME_REFCNT) {
 325                 /* Register Hyper-V timecounter */
 326                 tc_init(&hyperv_timecounter);
 327         }
 328 }
 329 SYSINIT(hyperv_initialize, SI_SUB_HYPERVISOR, SI_ORDER_FIRST, hyperv_init,
 330     NULL);
 331
 332 static void
 333 hypercall_memfree(void)
 334 {
 335         hyperv_dmamem_free(&hypercall_context.hc_dma,
 336             hypercall_context.hc_addr);
 337         hypercall_context.hc_addr = NULL;
 338 }
 339
 340 static void
 341 hypercall_create(void *arg __unused)
 342 {
 343         uint64_t hc, hc_orig;
 344
 345         if (vm_guest != VM_GUEST_HV)
 346                 return;
 347
 348         hypercall_context.hc_addr = hyperv_dmamem_alloc(NULL, PAGE_SIZE, 0,
 349             PAGE_SIZE, &hypercall_context.hc_dma, BUS_DMA_WAITOK);
 350         if (hypercall_context.hc_addr == NULL) {
 351                 printf("hyperv: Hypercall page allocation failed\n");
 352                 /* Can't perform any Hyper-V specific actions */
 353                 vm_guest = VM_GUEST_VM;
 354                 return;
 355         }
 356
 357         /* Get the 'reserved' bits, which requires preservation. */
 358         hc_orig = rdmsr(MSR_HV_HYPERCALL);
 359
 360         /*
 361          * Setup the Hypercall page.
 362          *
 363          * NOTE: 'reserved' bits MUST be preserved.
 364          */
 365         hc = ((hypercall_context.hc_dma.hv_paddr >> PAGE_SHIFT) <<
 366             MSR_HV_HYPERCALL_PGSHIFT) |
 367             (hc_orig & MSR_HV_HYPERCALL_RSVD_MASK) |
 368             MSR_HV_HYPERCALL_ENABLE;
 369         wrmsr(MSR_HV_HYPERCALL, hc);
 370
 371         /*
 372          * Confirm that Hypercall page did get setup.
 373          */
 374         hc = rdmsr(MSR_HV_HYPERCALL);
 375         if ((hc & MSR_HV_HYPERCALL_ENABLE) == 0) {
 376                 printf("hyperv: Hypercall setup failed\n");
 377                 hypercall_memfree();
 378                 /* Can't perform any Hyper-V specific actions */
 379                 vm_guest = VM_GUEST_VM;
 380                 return;
 381         }
 382         if (bootverbose)
 383                 printf("hyperv: Hypercall created\n");
 384 }
 385 SYSINIT(hypercall_ctor, SI_SUB_DRIVERS, SI_ORDER_FIRST, hypercall_create, NULL);
 386
 387 static void
 388 hypercall_destroy(void *arg __unused)
 389 {
 390         uint64_t hc;
 391
 392         if (hypercall_context.hc_addr == NULL)
 393                 return;
 394
 395         /* Disable Hypercall */
 396         hc = rdmsr(MSR_HV_HYPERCALL);
 397         wrmsr(MSR_HV_HYPERCALL, (hc & MSR_HV_HYPERCALL_RSVD_MASK));
 398         hypercall_memfree();
 399
 400         if (bootverbose)
 401                 printf("hyperv: Hypercall destroyed\n");
 402 }
 403 SYSUNINIT(hypercall_dtor, SI_SUB_DRIVERS, SI_ORDER_FIRST, hypercall_destroy,
 404     NULL);