3
0
Fork 0
forked from suyu/suyu

Merge pull request #5249 from ReinUsesLisp/lock-free-pages

core/memory: Read and write page table atomically
This commit is contained in:
bunnei 2021-01-01 02:54:01 -08:00 committed by GitHub
commit eb318ffffc
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
7 changed files with 132 additions and 147 deletions

View file

@ -10,16 +10,10 @@ PageTable::PageTable() = default;
PageTable::~PageTable() noexcept = default;
void PageTable::Resize(std::size_t address_space_width_in_bits, std::size_t page_size_in_bits,
bool has_attribute) {
const std::size_t num_page_table_entries{1ULL
<< (address_space_width_in_bits - page_size_in_bits)};
void PageTable::Resize(size_t address_space_width_in_bits, size_t page_size_in_bits) {
const size_t num_page_table_entries{1ULL << (address_space_width_in_bits - page_size_in_bits)};
pointers.resize(num_page_table_entries);
backing_addr.resize(num_page_table_entries);
if (has_attribute) {
attributes.resize(num_page_table_entries);
}
}
} // namespace Common

View file

@ -4,6 +4,7 @@
#pragma once
#include <atomic>
#include <tuple>
#include "common/common_types.h"
@ -20,10 +21,6 @@ enum class PageType : u8 {
/// Page is mapped to regular memory, but also needs to check for rasterizer cache flushing and
/// invalidation
RasterizerCachedMemory,
/// Page is mapped to a I/O region. Writing and reading to this page is handled by functions.
Special,
/// Page is allocated for use.
Allocated,
};
struct SpecialRegion {
@ -48,6 +45,59 @@ struct SpecialRegion {
* mimics the way a real CPU page table works.
*/
struct PageTable {
/// Number of bits reserved for attribute tagging.
/// This can be at most the guaranteed alignment of the pointers in the page table.
static constexpr int ATTRIBUTE_BITS = 2;
/**
* Pair of host pointer and page type attribute.
* This uses the lower bits of a given pointer to store the attribute tag.
* Writing and reading the pointer attribute pair is guaranteed to be atomic for the same method
* call. In other words, they are guaranteed to be synchronized at all times.
*/
class PageInfo {
public:
/// Returns the page pointer
[[nodiscard]] u8* Pointer() const noexcept {
return ExtractPointer(raw.load(std::memory_order_relaxed));
}
/// Returns the page type attribute
[[nodiscard]] PageType Type() const noexcept {
return ExtractType(raw.load(std::memory_order_relaxed));
}
/// Returns the page pointer and attribute pair, extracted from the same atomic read
[[nodiscard]] std::pair<u8*, PageType> PointerType() const noexcept {
const uintptr_t non_atomic_raw = raw.load(std::memory_order_relaxed);
return {ExtractPointer(non_atomic_raw), ExtractType(non_atomic_raw)};
}
/// Returns the raw representation of the page information.
/// Use ExtractPointer and ExtractType to unpack the value.
[[nodiscard]] uintptr_t Raw() const noexcept {
return raw.load(std::memory_order_relaxed);
}
/// Write a page pointer and type pair atomically
void Store(u8* pointer, PageType type) noexcept {
raw.store(reinterpret_cast<uintptr_t>(pointer) | static_cast<uintptr_t>(type));
}
/// Unpack a pointer from a page info raw representation
[[nodiscard]] static u8* ExtractPointer(uintptr_t raw) noexcept {
return reinterpret_cast<u8*>(raw & (~uintptr_t{0} << ATTRIBUTE_BITS));
}
/// Unpack a page type from a page info raw representation
[[nodiscard]] static PageType ExtractType(uintptr_t raw) noexcept {
return static_cast<PageType>(raw & ((uintptr_t{1} << ATTRIBUTE_BITS) - 1));
}
private:
std::atomic<uintptr_t> raw;
};
PageTable();
~PageTable() noexcept;
@ -63,20 +113,16 @@ struct PageTable {
*
* @param address_space_width_in_bits The address size width in bits.
* @param page_size_in_bits The page size in bits.
* @param has_attribute Whether or not this page has any backing attributes.
*/
void Resize(std::size_t address_space_width_in_bits, std::size_t page_size_in_bits,
bool has_attribute);
void Resize(size_t address_space_width_in_bits, size_t page_size_in_bits);
/**
* Vector of memory pointers backing each page. An entry can only be non-null if the
* corresponding entry in the `attributes` vector is of type `Memory`.
* corresponding attribute element is of type `Memory`.
*/
VirtualBuffer<u8*> pointers;
VirtualBuffer<PageInfo> pointers;
VirtualBuffer<u64> backing_addr;
VirtualBuffer<PageType> attributes;
};
} // namespace Common

View file

@ -15,10 +15,12 @@ void FreeMemoryPages(void* base, std::size_t size) noexcept;
template <typename T>
class VirtualBuffer final {
public:
static_assert(
std::is_trivially_constructible_v<T>,
"T must be trivially constructible, as non-trivial constructors will not be executed "
"with the current allocator");
// TODO: Uncomment this and change Common::PageTable::PageInfo to be trivially constructible
// using std::atomic_ref once libc++ has support for it
// static_assert(
// std::is_trivially_constructible_v<T>,
// "T must be trivially constructible, as non-trivial constructors will not be executed "
// "with the current allocator");
constexpr VirtualBuffer() = default;
explicit VirtualBuffer(std::size_t count) : alloc_size{count * sizeof(T)} {

View file

@ -133,6 +133,7 @@ std::shared_ptr<Dynarmic::A32::Jit> ARM_Dynarmic_32::MakeJit(Common::PageTable&
config.page_table = reinterpret_cast<std::array<std::uint8_t*, NUM_PAGE_TABLE_ENTRIES>*>(
page_table.pointers.data());
config.absolute_offset_page_table = true;
config.page_table_pointer_mask_bits = Common::PageTable::ATTRIBUTE_BITS;
config.detect_misaligned_access_via_page_table = 16 | 32 | 64 | 128;
config.only_detect_misalignment_via_page_table_on_page_boundary = true;

View file

@ -152,6 +152,7 @@ std::shared_ptr<Dynarmic::A64::Jit> ARM_Dynarmic_64::MakeJit(Common::PageTable&
// Memory
config.page_table = reinterpret_cast<void**>(page_table.pointers.data());
config.page_table_address_space_bits = address_space_bits;
config.page_table_pointer_mask_bits = Common::PageTable::ATTRIBUTE_BITS;
config.silently_mirror_page_table = false;
config.absolute_offset_page_table = true;
config.detect_misaligned_access_via_page_table = 16 | 32 | 64 | 128;

View file

@ -265,7 +265,7 @@ ResultCode PageTable::InitializeForProcess(FileSys::ProgramAddressSpaceType as_t
physical_memory_usage = 0;
memory_pool = pool;
page_table_impl.Resize(address_space_width, PageBits, true);
page_table_impl.Resize(address_space_width, PageBits);
return InitializeMemoryLayout(start, end);
}

View file

@ -4,7 +4,6 @@
#include <algorithm>
#include <cstring>
#include <mutex>
#include <optional>
#include <utility>
@ -68,21 +67,8 @@ struct Memory::Impl {
bool IsValidVirtualAddress(const Kernel::Process& process, const VAddr vaddr) const {
const auto& page_table = process.PageTable().PageTableImpl();
const u8* const page_pointer = page_table.pointers[vaddr >> PAGE_BITS];
if (page_pointer != nullptr) {
return true;
}
if (page_table.attributes[vaddr >> PAGE_BITS] == Common::PageType::RasterizerCachedMemory) {
return true;
}
if (page_table.attributes[vaddr >> PAGE_BITS] != Common::PageType::Special) {
return false;
}
return false;
const auto [pointer, type] = page_table.pointers[vaddr >> PAGE_BITS].PointerType();
return pointer != nullptr || type == Common::PageType::RasterizerCachedMemory;
}
bool IsValidVirtualAddress(VAddr vaddr) const {
@ -100,17 +86,15 @@ struct Memory::Impl {
}
u8* GetPointer(const VAddr vaddr) const {
u8* const page_pointer{current_page_table->pointers[vaddr >> PAGE_BITS]};
if (page_pointer) {
return page_pointer + vaddr;
const uintptr_t raw_pointer = current_page_table->pointers[vaddr >> PAGE_BITS].Raw();
if (u8* const pointer = Common::PageTable::PageInfo::ExtractPointer(raw_pointer)) {
return pointer + vaddr;
}
if (current_page_table->attributes[vaddr >> PAGE_BITS] ==
Common::PageType::RasterizerCachedMemory) {
const auto type = Common::PageTable::PageInfo::ExtractType(raw_pointer);
if (type == Common::PageType::RasterizerCachedMemory) {
return GetPointerFromRasterizerCachedMemory(vaddr);
}
return {};
return nullptr;
}
u8 Read8(const VAddr addr) {
@ -222,7 +206,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -231,10 +216,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
const u8* const src_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
const u8* const src_ptr = pointer + page_offset + (page_index << PAGE_BITS);
std::memcpy(dest_buffer, src_ptr, copy_amount);
break;
}
@ -268,7 +251,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped ReadBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -277,10 +261,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
const u8* const src_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
const u8* const src_ptr = pointer + page_offset + (page_index << PAGE_BITS);
std::memcpy(dest_buffer, src_ptr, copy_amount);
break;
}
@ -320,7 +302,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -328,10 +311,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
u8* const dest_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
u8* const dest_ptr = pointer + page_offset + (page_index << PAGE_BITS);
std::memcpy(dest_ptr, src_buffer, copy_amount);
break;
}
@ -364,7 +345,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped WriteBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -372,10 +354,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
u8* const dest_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
u8* const dest_ptr = pointer + page_offset + (page_index << PAGE_BITS);
std::memcpy(dest_ptr, src_buffer, copy_amount);
break;
}
@ -414,7 +394,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped ZeroBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -422,10 +403,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
u8* dest_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
u8* const dest_ptr = pointer + page_offset + (page_index << PAGE_BITS);
std::memset(dest_ptr, 0, copy_amount);
break;
}
@ -461,7 +440,8 @@ struct Memory::Impl {
std::min(static_cast<std::size_t>(PAGE_SIZE) - page_offset, remaining_size);
const auto current_vaddr = static_cast<VAddr>((page_index << PAGE_BITS) + page_offset);
switch (page_table.attributes[page_index]) {
const auto [pointer, type] = page_table.pointers[page_index].PointerType();
switch (type) {
case Common::PageType::Unmapped: {
LOG_ERROR(HW_Memory,
"Unmapped CopyBlock @ 0x{:016X} (start address = 0x{:016X}, size = {})",
@ -470,9 +450,8 @@ struct Memory::Impl {
break;
}
case Common::PageType::Memory: {
DEBUG_ASSERT(page_table.pointers[page_index]);
const u8* src_ptr =
page_table.pointers[page_index] + page_offset + (page_index << PAGE_BITS);
DEBUG_ASSERT(pointer);
const u8* src_ptr = pointer + page_offset + (page_index << PAGE_BITS);
WriteBlock(process, dest_addr, src_ptr, copy_amount);
break;
}
@ -498,34 +477,19 @@ struct Memory::Impl {
return CopyBlock(*system.CurrentProcess(), dest_addr, src_addr, size);
}
struct PageEntry {
u8* const pointer;
const Common::PageType attribute;
};
PageEntry SafePageEntry(std::size_t base) const {
std::lock_guard lock{rasterizer_cache_guard};
return {
.pointer = current_page_table->pointers[base],
.attribute = current_page_table->attributes[base],
};
}
void RasterizerMarkRegionCached(VAddr vaddr, u64 size, bool cached) {
std::lock_guard lock{rasterizer_cache_guard};
if (vaddr == 0) {
return;
}
// Iterate over a contiguous CPU address space, which corresponds to the specified GPU
// address space, marking the region as un/cached. The region is marked un/cached at a
// granularity of CPU pages, hence why we iterate on a CPU page basis (note: GPU page size
// is different). This assumes the specified GPU address region is contiguous as well.
u64 num_pages = ((vaddr + size - 1) >> PAGE_BITS) - (vaddr >> PAGE_BITS) + 1;
for (unsigned i = 0; i < num_pages; ++i, vaddr += PAGE_SIZE) {
Common::PageType& page_type{current_page_table->attributes[vaddr >> PAGE_BITS]};
const u64 num_pages = ((vaddr + size - 1) >> PAGE_BITS) - (vaddr >> PAGE_BITS) + 1;
for (u64 i = 0; i < num_pages; ++i, vaddr += PAGE_SIZE) {
const Common::PageType page_type{
current_page_table->pointers[vaddr >> PAGE_BITS].Type()};
if (cached) {
// Switch page type to cached if now cached
switch (page_type) {
@ -534,8 +498,8 @@ struct Memory::Impl {
// space, for example, a system module need not have a VRAM mapping.
break;
case Common::PageType::Memory:
page_type = Common::PageType::RasterizerCachedMemory;
current_page_table->pointers[vaddr >> PAGE_BITS] = nullptr;
current_page_table->pointers[vaddr >> PAGE_BITS].Store(
nullptr, Common::PageType::RasterizerCachedMemory);
break;
case Common::PageType::RasterizerCachedMemory:
// There can be more than one GPU region mapped per CPU region, so it's common
@ -556,16 +520,16 @@ struct Memory::Impl {
// that this area is already unmarked as cached.
break;
case Common::PageType::RasterizerCachedMemory: {
u8* pointer{GetPointerFromRasterizerCachedMemory(vaddr & ~PAGE_MASK)};
u8* const pointer{GetPointerFromRasterizerCachedMemory(vaddr & ~PAGE_MASK)};
if (pointer == nullptr) {
// It's possible that this function has been called while updating the
// pagetable after unmapping a VMA. In that case the underlying VMA will no
// longer exist, and we should just leave the pagetable entry blank.
page_type = Common::PageType::Unmapped;
current_page_table->pointers[vaddr >> PAGE_BITS].Store(
nullptr, Common::PageType::Unmapped);
} else {
current_page_table->pointers[vaddr >> PAGE_BITS] =
pointer - (vaddr & ~PAGE_MASK);
page_type = Common::PageType::Memory;
current_page_table->pointers[vaddr >> PAGE_BITS].Store(
pointer - (vaddr & ~PAGE_MASK), Common::PageType::Memory);
}
break;
}
@ -595,7 +559,7 @@ struct Memory::Impl {
auto& gpu = system.GPU();
for (u64 i = 0; i < size; i++) {
const auto page = base + i;
if (page_table.attributes[page] == Common::PageType::RasterizerCachedMemory) {
if (page_table.pointers[page].Type() == Common::PageType::RasterizerCachedMemory) {
gpu.FlushAndInvalidateRegion(page << PAGE_BITS, PAGE_SIZE);
}
}
@ -610,20 +574,18 @@ struct Memory::Impl {
"Mapping memory page without a pointer @ {:016x}", base * PAGE_SIZE);
while (base != end) {
page_table.attributes[base] = type;
page_table.pointers[base] = nullptr;
page_table.pointers[base].Store(nullptr, type);
page_table.backing_addr[base] = 0;
base += 1;
}
} else {
while (base != end) {
page_table.pointers[base] =
system.DeviceMemory().GetPointer(target) - (base << PAGE_BITS);
page_table.attributes[base] = type;
page_table.pointers[base].Store(
system.DeviceMemory().GetPointer(target) - (base << PAGE_BITS), type);
page_table.backing_addr[base] = target - (base << PAGE_BITS);
ASSERT_MSG(page_table.pointers[base],
ASSERT_MSG(page_table.pointers[base].Pointer(),
"memory mapping base yield a nullptr within the table");
base += 1;
@ -646,21 +608,13 @@ struct Memory::Impl {
template <typename T>
T Read(const VAddr vaddr) {
// Avoid adding any extra logic to this fast-path block
if (const u8* const pointer = current_page_table->pointers[vaddr >> PAGE_BITS]) {
const uintptr_t raw_pointer = current_page_table->pointers[vaddr >> PAGE_BITS].Raw();
if (const u8* const pointer = Common::PageTable::PageInfo::ExtractPointer(raw_pointer)) {
T value;
std::memcpy(&value, &pointer[vaddr], sizeof(T));
return value;
}
// Otherwise, we need to grab the page with a lock, in case it is currently being modified
const auto entry = SafePageEntry(vaddr >> PAGE_BITS);
if (entry.pointer) {
T value;
std::memcpy(&value, &entry.pointer[vaddr], sizeof(T));
return value;
}
switch (entry.attribute) {
switch (Common::PageTable::PageInfo::ExtractType(raw_pointer)) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_Memory, "Unmapped Read{} @ 0x{:08X}", sizeof(T) * 8, vaddr);
return 0;
@ -692,20 +646,12 @@ struct Memory::Impl {
template <typename T>
void Write(const VAddr vaddr, const T data) {
// Avoid adding any extra logic to this fast-path block
if (u8* const pointer = current_page_table->pointers[vaddr >> PAGE_BITS]) {
const uintptr_t raw_pointer = current_page_table->pointers[vaddr >> PAGE_BITS].Raw();
if (u8* const pointer = Common::PageTable::PageInfo::ExtractPointer(raw_pointer)) {
std::memcpy(&pointer[vaddr], &data, sizeof(T));
return;
}
// Otherwise, we need to grab the page with a lock, in case it is currently being modified
const auto entry = SafePageEntry(vaddr >> PAGE_BITS);
if (entry.pointer) {
// Memory was mapped, we are done
std::memcpy(&entry.pointer[vaddr], &data, sizeof(T));
return;
}
switch (entry.attribute) {
switch (Common::PageTable::PageInfo::ExtractType(raw_pointer)) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_Memory, "Unmapped Write{} 0x{:08X} @ 0x{:016X}", sizeof(data) * 8,
static_cast<u32>(data), vaddr);
@ -726,15 +672,13 @@ struct Memory::Impl {
template <typename T>
bool WriteExclusive(const VAddr vaddr, const T data, const T expected) {
u8* page_pointer = current_page_table->pointers[vaddr >> PAGE_BITS];
if (page_pointer != nullptr) {
const uintptr_t raw_pointer = current_page_table->pointers[vaddr >> PAGE_BITS].Raw();
if (u8* const pointer = Common::PageTable::PageInfo::ExtractPointer(raw_pointer)) {
// NOTE: Avoid adding any extra logic to this fast-path block
auto* pointer = reinterpret_cast<volatile T*>(&page_pointer[vaddr]);
return Common::AtomicCompareAndSwap(pointer, data, expected);
const auto volatile_pointer = reinterpret_cast<volatile T*>(&pointer[vaddr]);
return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
}
const Common::PageType type = current_page_table->attributes[vaddr >> PAGE_BITS];
switch (type) {
switch (Common::PageTable::PageInfo::ExtractType(raw_pointer)) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_Memory, "Unmapped Write{} 0x{:08X} @ 0x{:016X}", sizeof(data) * 8,
static_cast<u32>(data), vaddr);
@ -755,15 +699,13 @@ struct Memory::Impl {
}
bool WriteExclusive128(const VAddr vaddr, const u128 data, const u128 expected) {
u8* const page_pointer = current_page_table->pointers[vaddr >> PAGE_BITS];
if (page_pointer != nullptr) {
const uintptr_t raw_pointer = current_page_table->pointers[vaddr >> PAGE_BITS].Raw();
if (u8* const pointer = Common::PageTable::PageInfo::ExtractPointer(raw_pointer)) {
// NOTE: Avoid adding any extra logic to this fast-path block
auto* pointer = reinterpret_cast<volatile u64*>(&page_pointer[vaddr]);
return Common::AtomicCompareAndSwap(pointer, data, expected);
const auto volatile_pointer = reinterpret_cast<volatile u64*>(&pointer[vaddr]);
return Common::AtomicCompareAndSwap(volatile_pointer, data, expected);
}
const Common::PageType type = current_page_table->attributes[vaddr >> PAGE_BITS];
switch (type) {
switch (Common::PageTable::PageInfo::ExtractType(raw_pointer)) {
case Common::PageType::Unmapped:
LOG_ERROR(HW_Memory, "Unmapped Write{} 0x{:08X} @ 0x{:016X}{:016X}", sizeof(data) * 8,
static_cast<u64>(data[1]), static_cast<u64>(data[0]), vaddr);
@ -783,7 +725,6 @@ struct Memory::Impl {
return true;
}
mutable std::mutex rasterizer_cache_guard;
Common::PageTable* current_page_table = nullptr;
Core::System& system;
};