// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <atomic>
#include <limits>
#include <memory>
#include <type_traits>
#include "common/address_space.h"
#include "common/address_space.inc"
#include "common/alignment.h"
#include "common/assert.h"
#include "common/div_ceil.h"
#include "common/scope_exit.h"
#include "core/device_memory.h"
#include "core/device_memory_manager.h"
#include "core/memory.h"
namespace Core {
namespace {
class MultiAddressContainer {
public:
MultiAddressContainer() = default;
~MultiAddressContainer() = default;
void GatherValues(u32 start_entry, Common::ScratchBuffer<u32>& buffer) {
buffer.resize(8);
buffer.resize(0);
size_t index = 0;
const auto add_value = [&](u32 value) {
buffer[index] = value;
index++;
buffer.resize(index);
};
u32 iter_entry = start_entry;
Entry* current = &storage[iter_entry - 1];
add_value(current->value);
while (current->next_entry != 0) {
iter_entry = current->next_entry;
current = &storage[iter_entry - 1];
add_value(current->value);
}
}
u32 Register(u32 value) {
return RegisterImplementation(value);
}
void Register(u32 value, u32 start_entry) {
auto entry_id = RegisterImplementation(value);
u32 iter_entry = start_entry;
Entry* current = &storage[iter_entry - 1];
while (current->next_entry != 0) {
iter_entry = current->next_entry;
current = &storage[iter_entry - 1];
}
current->next_entry = entry_id;
}
std::pair<bool, u32> Unregister(u32 value, u32 start_entry) {
u32 iter_entry = start_entry;
Entry* previous{};
Entry* current = &storage[iter_entry - 1];
Entry* next{};
bool more_than_one_remaining = false;
u32 result_start{start_entry};
size_t count = 0;
while (current->value != value) {
count++;
previous = current;
iter_entry = current->next_entry;
current = &storage[iter_entry - 1];
}
// Find next
u32 next_entry = current->next_entry;
if (next_entry != 0) {
next = &storage[next_entry - 1];
more_than_one_remaining = next->next_entry != 0 || previous != nullptr;
}
if (previous) {
previous->next_entry = next_entry;
} else {
result_start = next_entry;
}
free_entries.emplace_back(iter_entry);
return std::make_pair(more_than_one_remaining || count > 1, result_start);
}
u32 ReleaseEntry(u32 start_entry) {
Entry* current = &storage[start_entry - 1];
free_entries.emplace_back(start_entry);
return current->value;
}
private:
u32 RegisterImplementation(u32 value) {
auto entry_id = GetNewEntry();
auto& entry = storage[entry_id - 1];
entry.next_entry = 0;
entry.value = value;
return entry_id;
}
u32 GetNewEntry() {
if (!free_entries.empty()) {
u32 result = free_entries.front();
free_entries.pop_front();
return result;
}
storage.emplace_back();
u32 new_entry = static_cast<u32>(storage.size());
return new_entry;
}
struct Entry {
u32 next_entry{};
u32 value{};
};
std::deque<Entry> storage;
std::deque<u32> free_entries;
};
struct EmptyAllocator {
EmptyAllocator([[maybe_unused]] DAddr address) {}
};
} // namespace
template <typename DTraits>
struct DeviceMemoryManagerAllocator {
static constexpr bool supports_pinning = DTraits::supports_pinning;
static constexpr size_t device_virtual_bits = DTraits::device_virtual_bits;
static constexpr size_t pin_bits = 32;
static constexpr DAddr first_address = 1ULL << Memory::YUZU_PAGEBITS;
static constexpr DAddr max_pin_area = supports_pinning ? 1ULL << pin_bits : first_address;
static constexpr DAddr max_device_area = 1ULL << device_virtual_bits;
DeviceMemoryManagerAllocator()
: pin_allocator(first_address),
main_allocator(supports_pinning ? 1ULL << pin_bits : first_address) {}
std::conditional_t<supports_pinning, Common::FlatAllocator<DAddr, 0, pin_bits>, EmptyAllocator>
pin_allocator;
Common::FlatAllocator<DAddr, 0, device_virtual_bits> main_allocator;
MultiAddressContainer multi_dev_address;
/// Returns true when vaddr -> vaddr+size is fully contained in the buffer
template <bool pin_area>
[[nodiscard]] bool IsInBounds(VAddr addr, u64 size) const noexcept {
if constexpr (pin_area) {
return addr >= 0 && addr + size <= max_pin_area;
} else {
return addr >= max_pin_area && addr + size <= max_device_area;
}
}
DAddr Allocate(size_t size) {
return main_allocator.Allocate(size);
}
DAddr AllocatePinned(size_t size) {
if constexpr (supports_pinning) {
return pin_allocator.Allocate(size);
} else {
return DAddr{};
}
}
void DoInRange(DAddr address, size_t size, auto pin_func, auto main_func) {
if (IsInBounds<true>(address, size)) {
pin_func(address, size);
return;
}
if (IsInBounds<false>(address, size)) {
main_func(address, size);
return;
}
DAddr end_size = address + size - max_pin_area;
DAddr end_size2 = max_pin_area - address;
pin_func(address, end_size2);
main_func(max_pin_area, end_size);
}
void AllocateFixed(DAddr b_address, size_t b_size) {
if constexpr (supports_pinning) {
DoInRange(
b_address, b_size,
[this](DAddr address, size_t size) { pin_allocator.AllocateFixed(address, size); },
[this](DAddr address, size_t size) {
main_allocator.AllocateFixed(address, size);
});
} else {
main_allocator.AllocateFixed(b_address, b_size);
}
}
void Free(DAddr b_address, size_t b_size) {
if constexpr (supports_pinning) {
DoInRange(
b_address, b_size,
[this](DAddr address, size_t size) { pin_allocator.Free(address, size); },
[this](DAddr address, size_t size) { main_allocator.Free(address, size); });
} else {
main_allocator.Free(b_address, b_size);
}
}
};
template <typename Traits>
DeviceMemoryManager<Traits>::DeviceMemoryManager(const DeviceMemory& device_memory_)
: physical_base{reinterpret_cast<const uintptr_t>(device_memory_.buffer.BackingBasePointer())},
interface{nullptr}, compressed_physical_ptr(device_as_size >> Memory::YUZU_PAGEBITS),
compressed_device_addr(1ULL << (physical_max_bits - Memory::YUZU_PAGEBITS)),
cpu_backing_address(device_as_size >> Memory::YUZU_PAGEBITS) {
impl = std::make_unique<DeviceMemoryManagerAllocator<Traits>>();
cached_pages = std::make_unique<CachedPages>();
}
template <typename Traits>
DeviceMemoryManager<Traits>::~DeviceMemoryManager() = default;
template <typename Traits>
void DeviceMemoryManager<Traits>::BindInterface(DeviceInterface* interface_) {
interface = interface_;
}
template <typename Traits>
DAddr DeviceMemoryManager<Traits>::Allocate(size_t size) {
return impl->Allocate(size);
}
template <typename Traits>
void DeviceMemoryManager<Traits>::AllocateFixed(DAddr start, size_t size) {
return impl->AllocateFixed(start, size);
}
template <typename Traits>
DAddr DeviceMemoryManager<Traits>::AllocatePinned(size_t size) {
return impl->AllocatePinned(size);
}
template <typename Traits>
void DeviceMemoryManager<Traits>::Free(DAddr start, size_t size) {
impl->Free(start, size);
}
template <typename Traits>
void DeviceMemoryManager<Traits>::Map(DAddr address, VAddr virtual_address, size_t size,
size_t process_id) {
Core::Memory::Memory* process_memory = registered_processes[process_id];
size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
std::scoped_lock lk(mapping_guard);
for (size_t i = 0; i < num_pages; i++) {
const VAddr new_vaddress = virtual_address + i * Memory::YUZU_PAGESIZE;
auto* ptr = process_memory->GetPointerSilent(Common::ProcessAddress(new_vaddress));
if (ptr == nullptr) [[unlikely]] {
compressed_physical_ptr[start_page_d + i] = 0;
continue;
}
auto phys_addr = static_cast<u32>(GetRawPhysicalAddr(ptr) >> Memory::YUZU_PAGEBITS) + 1U;
compressed_physical_ptr[start_page_d + i] = phys_addr;
InsertCPUBacking(start_page_d + i, new_vaddress, process_id);
const u32 base_dev = compressed_device_addr[phys_addr - 1U];
const u32 new_dev = static_cast<u32>(start_page_d + i);
if (base_dev == 0) [[likely]] {
compressed_device_addr[phys_addr - 1U] = new_dev;
continue;
}
u32 start_id = base_dev & MULTI_MASK;
if ((base_dev >> MULTI_FLAG_BITS) == 0) {
start_id = impl->multi_dev_address.Register(base_dev);
compressed_device_addr[phys_addr - 1U] = MULTI_FLAG | start_id;
}
impl->multi_dev_address.Register(new_dev, start_id);
}
}
template <typename Traits>
void DeviceMemoryManager<Traits>::Unmap(DAddr address, size_t size) {
size_t start_page_d = address >> Memory::YUZU_PAGEBITS;
size_t num_pages = Common::AlignUp(size, Memory::YUZU_PAGESIZE) >> Memory::YUZU_PAGEBITS;
interface->InvalidateRegion(address, size);
std::scoped_lock lk(mapping_guard);
for (size_t i = 0; i < num_pages; i++) {
auto phys_addr = compressed_physical_ptr[start_page_d + i];
compressed_physical_ptr[start_page_d + i] = 0;
cpu_backing_address[start_page_d + i] = 0;
if (phys_addr != 0) [[likely]] {
const u32 base_dev = compressed_device_addr[phys_addr - 1U];
if ((base_dev >> MULTI_FLAG_BITS) == 0) [[likely]] {
compressed_device_addr[phys_addr - 1] = 0;
continue;
}
const auto [more_entries, new_start] = impl->multi_dev_address.Unregister(
static_cast<u32>(start_page_d + i), base_dev & MULTI_MASK);
if (!more_entries) {
compressed_device_addr[phys_addr - 1] =
impl->multi_dev_address.ReleaseEntry(new_start);
continue;
}
compressed_device_addr[phys_addr - 1] = new_start | MULTI_FLAG;
}
}
}
template <typename Traits>
void DeviceMemoryManager<Traits>::InnerGatherDeviceAddresses(Common::ScratchBuffer<u32>& buffer,
PAddr address) {
size_t phys_addr = address >> page_bits;
std::scoped_lock lk(mapping_guard);
u32 backing = compressed_device_addr[phys_addr];
if ((backing >> MULTI_FLAG_BITS) != 0) {
impl->multi_dev_address.GatherValues(backing & MULTI_MASK, buffer);
return;
}
buffer.resize(1);
buffer[0] = backing;
}
template <typename Traits>
template <typename T>
T* DeviceMemoryManager<Traits>::GetPointer(DAddr address) {
const size_t index = address >> Memory::YUZU_PAGEBITS;
const size_t offset = address & Memory::YUZU_PAGEMASK;
auto phys_addr = compressed_physical_ptr[index];
if (phys_addr == 0) [[unlikely]] {
return nullptr;
}
return GetPointerFromRaw<T>(
static_cast<PAddr>(((phys_addr - 1) << Memory::YUZU_PAGEBITS) + offset));
}
template <typename Traits>
template <typename T>
const T* DeviceMemoryManager<Traits>::GetPointer(DAddr address) const {
const size_t index = address >> Memory::YUZU_PAGEBITS;
const size_t offset = address & Memory::YUZU_PAGEMASK;
auto phys_addr = compressed_physical_ptr[index];
if (phys_addr == 0) [[unlikely]] {
return nullptr;
}
return GetPointerFromRaw<T>(
static_cast<PAddr>(((phys_addr - 1) << Memory::YUZU_PAGEBITS) + offset));
}
template <typename Traits>
template <typename T>
void DeviceMemoryManager<Traits>::Write(DAddr address, T value) {
T* ptr = GetPointer<T>(address);
if (!ptr) [[unlikely]] {
return;
}
std::memcpy(ptr, &value, sizeof(T));
}
template <typename Traits>
template <typename T>
T DeviceMemoryManager<Traits>::Read(DAddr address) const {
const T* ptr = GetPointer<T>(address);
T result{};
if (!ptr) [[unlikely]] {
return result;
}
std::memcpy(&result, ptr, sizeof(T));
return result;
}
template <typename Traits>
void DeviceMemoryManager<Traits>::WalkBlock(DAddr addr, std::size_t size, auto on_unmapped,
auto on_memory, auto increment) {
std::size_t remaining_size = size;
std::size_t page_index = addr >> Memory::YUZU_PAGEBITS;
std::size_t page_offset = addr & Memory::YUZU_PAGEMASK;
while (remaining_size) {
const std::size_t copy_amount =
std::min(static_cast<std::size_t>(Memory::YUZU_PAGESIZE) - page_offset, remaining_size);
const auto current_vaddr =
static_cast<u64>((page_index << Memory::YUZU_PAGEBITS) + page_offset);
SCOPE_EXIT({
page_index++;
page_offset = 0;
increment(copy_amount);
remaining_size -= copy_amount;
});
auto phys_addr = compressed_physical_ptr[page_index];
if (phys_addr == 0) {
on_unmapped(copy_amount, current_vaddr);
continue;
}
auto* mem_ptr = GetPointerFromRaw<u8>(
static_cast<PAddr>(((phys_addr - 1) << Memory::YUZU_PAGEBITS) + page_offset));
on_memory(copy_amount, mem_ptr);
}
}
template <typename Traits>
void DeviceMemoryManager<Traits>::ReadBlock(DAddr address, void* dest_pointer, size_t size) {
interface->FlushRegion(address, size);
WalkBlock(
address, size,
[&](size_t copy_amount, DAddr current_vaddr) {
LOG_ERROR(
HW_Memory,
"Unmapped Device ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
current_vaddr, address, size);
std::memset(dest_pointer, 0, copy_amount);
},
[&](size_t copy_amount, const u8* const src_ptr) {
std::memcpy(dest_pointer, src_ptr, copy_amount);
},
[&](const std::size_t copy_amount) {
dest_pointer = static_cast<u8*>(dest_pointer) + copy_amount;
});
}
template <typename Traits>
void DeviceMemoryManager<Traits>::WriteBlock(DAddr address, const void* src_pointer, size_t size) {
WalkBlock(
address, size,
[&](size_t copy_amount, DAddr current_vaddr) {
LOG_ERROR(
HW_Memory,
"Unmapped Device WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
current_vaddr, address, size);
},
[&](size_t copy_amount, u8* const dst_ptr) {
std::memcpy(dst_ptr, src_pointer, copy_amount);
},
[&](const std::size_t copy_amount) {
src_pointer = static_cast<const u8*>(src_pointer) + copy_amount;
});
interface->InvalidateRegion(address, size);
}
template <typename Traits>
void DeviceMemoryManager<Traits>::ReadBlockUnsafe(DAddr address, void* dest_pointer, size_t size) {
WalkBlock(
address, size,
[&](size_t copy_amount, DAddr current_vaddr) {
LOG_ERROR(
HW_Memory,
"Unmapped Device ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
current_vaddr, address, size);
std::memset(dest_pointer, 0, copy_amount);
},
[&](size_t copy_amount, const u8* const src_ptr) {
std::memcpy(dest_pointer, src_ptr, copy_amount);
},
[&](const std::size_t copy_amount) {
dest_pointer = static_cast<u8*>(dest_pointer) + copy_amount;
});
}
template <typename Traits>
void DeviceMemoryManager<Traits>::WriteBlockUnsafe(DAddr address, const void* src_pointer,
size_t size) {
WalkBlock(
address, size,
[&](size_t copy_amount, DAddr current_vaddr) {
LOG_ERROR(
HW_Memory,
"Unmapped Device WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
current_vaddr, address, size);
},
[&](size_t copy_amount, u8* const dst_ptr) {
std::memcpy(dst_ptr, src_pointer, copy_amount);
},
[&](const std::size_t copy_amount) {
src_pointer = static_cast<const u8*>(src_pointer) + copy_amount;
});
}
template <typename Traits>
size_t DeviceMemoryManager<Traits>::RegisterProcess(Memory::Memory* memory_interface) {
size_t new_id;
if (!id_pool.empty()) {
new_id = id_pool.front();
id_pool.pop_front();
registered_processes[new_id] = memory_interface;
} else {
registered_processes.emplace_back(memory_interface);
new_id = registered_processes.size() - 1U;
}
return new_id;
}
template <typename Traits>
void DeviceMemoryManager<Traits>::UnregisterProcess(size_t id) {
registered_processes[id] = nullptr;
id_pool.push_front(id);
}
template <typename Traits>
void DeviceMemoryManager<Traits>::UpdatePagesCachedCount(DAddr addr, size_t size, s32 delta) {
bool locked = false;
auto lock = [&] {
if (!locked) {
counter_guard.lock();
locked = true;
}
};
SCOPE_EXIT({
if (locked) {
counter_guard.unlock();
}
});
u64 uncache_begin = 0;
u64 cache_begin = 0;
u64 uncache_bytes = 0;
u64 cache_bytes = 0;
const auto MarkRegionCaching = &DeviceMemoryManager<Traits>::DeviceMethods::MarkRegionCaching;
std::atomic_thread_fence(std::memory_order_acquire);
const size_t page_end = Common::DivCeil(addr + size, Memory::YUZU_PAGESIZE);
size_t page = addr >> Memory::YUZU_PAGEBITS;
auto [process_id, base_vaddress] = ExtractCPUBacking(page);
size_t vpage = base_vaddress >> Memory::YUZU_PAGEBITS;
auto* memory_interface = registered_processes[process_id];
for (; page != page_end; ++page) {
std::atomic_uint16_t& count = cached_pages->at(page >> 2).Count(page);
if (delta > 0) {
ASSERT_MSG(count.load(std::memory_order::relaxed) < std::numeric_limits<u16>::max(),
"Count may overflow!");
} else if (delta < 0) {
ASSERT_MSG(count.load(std::memory_order::relaxed) > 0, "Count may underflow!");
} else {
ASSERT_MSG(false, "Delta must be non-zero!");
}
// Adds or subtracts 1, as count is a unsigned 8-bit value
count.fetch_add(static_cast<u16>(delta), std::memory_order_release);
// Assume delta is either -1 or 1
if (count.load(std::memory_order::relaxed) == 0) {
if (uncache_bytes == 0) {
uncache_begin = vpage;
}
uncache_bytes += Memory::YUZU_PAGESIZE;
} else if (uncache_bytes > 0) {
lock();
MarkRegionCaching(memory_interface, uncache_begin << Memory::YUZU_PAGEBITS,
uncache_bytes, false);
uncache_bytes = 0;
}
if (count.load(std::memory_order::relaxed) == 1 && delta > 0) {
if (cache_bytes == 0) {
cache_begin = vpage;
}
cache_bytes += Memory::YUZU_PAGESIZE;
} else if (cache_bytes > 0) {
lock();
MarkRegionCaching(memory_interface, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
true);
cache_bytes = 0;
}
vpage++;
}
if (uncache_bytes > 0) {
lock();
MarkRegionCaching(memory_interface, uncache_begin << Memory::YUZU_PAGEBITS, uncache_bytes,
false);
}
if (cache_bytes > 0) {
lock();
MarkRegionCaching(memory_interface, cache_begin << Memory::YUZU_PAGEBITS, cache_bytes,
true);
}
}
} // namespace Core