x86/hypervisor/pvclock.cpp

// Copyright 2018 The Fuchsia Authors
//
// Use of this source code is governed by a MIT-style
// license that can be found in the LICENSE file or at
// https://opensource.org/licenses/MIT

#include "pvclock_priv.h"
#include <arch/hypervisor.h>
#include <arch/x86/pvclock.h>
#include <bits.h>
#include <hypervisor/guest_physical_address_space.h>
#include <platform.h>
#include <vm/physmap.h>
#include <zircon/types.h>

namespace {

void calculate_scale_factor(uint64_t tsc_freq, uint32_t* mul, int8_t* shift) {
    // Guests converts TSC ticks to nanoseconds using this formula:
    //   ns = #TSCticks * mul * 2^(shift - 32).
    // mul * 2^(shift - 32) is a fractional number used as a scale factor in conversion.
    // It's very similar to how floating point numbers are usually represented in memory.
    static const uint64_t target_freq = 1000000000ul;

    DEBUG_ASSERT(tsc_freq != 0);

    // We maintain the folowing invariant: 2^(exponent - 32) * x/y ~ target_freq / tsc_freq,
    int8_t exponent = 32;
    uint64_t x = target_freq;
    uint64_t y = tsc_freq;

    // First make y small enough so that (y << 31) doesn't overflow in the next step. Adjust
    // exponent along the way to maintain invariant.
    while (y >= (1ull << 31)) {
        y >>= 1;
        exponent--;
    }

    // We scale x/y multiplying x by 2 until it gets big enough or we run out of bits.
    while (x < (y << 31) && BIT(x, 63) == 0) {
        x <<= 1;
        exponent--;
    }

    // Though it's very unlikely lets also consider a situation when x/y is still too small.
    while (x < y) {
        y >>= 1;
        exponent++;
    }

    // Finally make sure that x/y fits within 32 bits.
    while (x >= (y << 32)) {
        x >>= 1;
        exponent++;
    }

    *shift = static_cast<int8_t>(exponent);
    *mul = static_cast<uint32_t>(x / y);
}

} // namespace

extern fbl::atomic<int64_t> utc_offset;

zx_status_t pvclock_update_boot_time(hypervisor::GuestPhysicalAddressSpace* gpas,
                                     zx_vaddr_t guest_paddr) {
    // KVM doesn't provide any protection against concurrent wall time requests from different
    // VCPUs, but documentation doesn't mention that it cannot happen and moreover it properly
    // protects per VCPU system time. Therefore to be on the safer side we use one global mutex
    // for protection.
    static fbl::Mutex mutex;
    static uint32_t version __TA_GUARDED(mutex);

    hypervisor::GuestPtr guest_ptr;
    zx_status_t status = gpas->CreateGuestPtr(guest_paddr, sizeof(pvclock_boot_time),
                                              "pvclock-boot-time-guest-mapping", &guest_ptr);
    if (status != ZX_OK) {
        return status;
    }
    auto boot_time = guest_ptr.as<pvclock_boot_time>();
    ZX_DEBUG_ASSERT(boot_time != nullptr);
    memset(boot_time, 0, sizeof(*boot_time));

    fbl::AutoLock lock(&mutex);
    zx_time_t time = utc_offset.load();
    // See the comment for pvclock_boot_time structure in arch/x86/pvclock.h
    atomic_store_relaxed_u32(&boot_time->version, version + 1);
    atomic_fence();
    boot_time->seconds = static_cast<uint32_t>(time / ZX_SEC(1));
    boot_time->nseconds = static_cast<uint32_t>(time % ZX_SEC(1));
    atomic_fence();
    atomic_store_relaxed_u32(&boot_time->version, version + 2);
    version += 2;
    return ZX_OK;
}

zx_status_t pvclock_reset_clock(PvClockState* pvclock, hypervisor::GuestPhysicalAddressSpace* gpas,
                                zx_vaddr_t guest_paddr) {
    zx_status_t status =
        gpas->CreateGuestPtr(guest_paddr, sizeof(pvclock_system_time),
                             "pvclock-system-time-guest-mapping", &pvclock->guest_ptr);
    if (status != ZX_OK) {
        return status;
    }
    pvclock->system_time = pvclock->guest_ptr.as<pvclock_system_time>();
    ZX_DEBUG_ASSERT(pvclock->system_time != nullptr);
    memset(pvclock->system_time, 0, sizeof(*pvclock->system_time));
    return ZX_OK;
}

void pvclock_update_system_time(PvClockState* pvclock,
                                hypervisor::GuestPhysicalAddressSpace* gpas) {
    if (!pvclock->system_time) {
        return;
    }

    uint32_t tsc_mul;
    int8_t tsc_shift;
    calculate_scale_factor(ticks_per_second(), &tsc_mul, &tsc_shift);

    // See the comment for pvclock_boot_time structure in arch/x86/pvclock.h
    pvclock_system_time* system_time = pvclock->system_time;
    atomic_store_relaxed_u32(&system_time->version, pvclock->version + 1);
    atomic_fence();
    system_time->tsc_mul = tsc_mul;
    system_time->tsc_shift = tsc_shift;
    system_time->system_time = current_time();
    system_time->tsc_timestamp = rdtsc();
    system_time->flags = pvclock->is_stable ? kKvmSystemTimeStable : 0;
    atomic_fence();
    atomic_store_relaxed_u32(&system_time->version, pvclock->version + 2);
    pvclock->version += 2;
}

void pvclock_stop_clock(PvClockState* pvclock) {
    pvclock->system_time = nullptr;
    pvclock->guest_ptr.reset();
}

zx_status_t pvclock_populate_offset(hypervisor::GuestPhysicalAddressSpace* gpas,
                                    zx_vaddr_t guest_paddr) {
    hypervisor::GuestPtr guest_ptr;
    zx_status_t status =
        gpas->CreateGuestPtr(guest_paddr, sizeof(PvClockOffset),
                             "pvclock-offset-guest-mapping", &guest_ptr);
    if (status != ZX_OK) {
        return status;
    }
    auto offset = guest_ptr.as<PvClockOffset>();
    ZX_DEBUG_ASSERT(offset != nullptr);
    memset(offset, 0, sizeof(*offset));
    zx_time_t time = utc_offset.load() + current_time();
    uint64_t tsc = rdtsc();
    offset->sec = time / ZX_SEC(1);
    offset->nsec = time % ZX_SEC(1);
    offset->tsc = tsc;
    return ZX_OK;
}