mirror of
https://git.suyu.dev/suyu/suyu
synced 2024-11-01 12:57:52 +00:00
Merge pull request #4430 from bunnei/new-gpu-vmm
hle: nvdrv: Rewrite of GPU memory management.
This commit is contained in:
bunnei
GitHub
commit
f11628b9b7
6 changed files with 451 additions and 613 deletions
|
|
@ -16,11 +16,12 @@
|
|||
#include "video_core/renderer_base.h"
|
||||
|
||||
namespace Service::Nvidia::Devices {
|
||||
|
||||
namespace NvErrCodes {
|
||||
enum {
|
||||
InvalidNmapHandle = -22,
|
||||
};
|
||||
}
|
||||
constexpr u32 Success{};
|
||||
constexpr u32 OutOfMemory{static_cast<u32>(-12)};
|
||||
constexpr u32 InvalidInput{static_cast<u32>(-22)};
|
||||
} // namespace NvErrCodes
|
||||
|
||||
nvhost_as_gpu::nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev)
|
||||
: nvdevice(system), nvmap_dev(std::move(nvmap_dev)) {}
|
||||
|
|
@ -49,8 +50,9 @@ u32 nvhost_as_gpu::ioctl(Ioctl command, const std::vector<u8>& input, const std:
|
|||
break;
|
||||
}
|
||||
|
||||
if (static_cast<IoctlCommand>(command.cmd.Value()) == IoctlCommand::IocRemapCommand)
|
||||
if (static_cast<IoctlCommand>(command.cmd.Value()) == IoctlCommand::IocRemapCommand) {
|
||||
return Remap(input, output);
|
||||
}
|
||||
|
||||
UNIMPLEMENTED_MSG("Unimplemented ioctl command");
|
||||
return 0;
|
||||
|
|
@ -59,6 +61,7 @@ u32 nvhost_as_gpu::ioctl(Ioctl command, const std::vector<u8>& input, const std:
|
|||
u32 nvhost_as_gpu::InitalizeEx(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
IoctlInitalizeEx params{};
|
||||
std::memcpy(¶ms, input.data(), input.size());
|
||||
|
||||
LOG_WARNING(Service_NVDRV, "(STUBBED) called, big_page_size=0x{:X}", params.big_page_size);
|
||||
|
||||
return 0;
|
||||
|
|
@ -67,53 +70,61 @@ u32 nvhost_as_gpu::InitalizeEx(const std::vector<u8>& input, std::vector<u8>& ou
|
|||
u32 nvhost_as_gpu::AllocateSpace(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
IoctlAllocSpace params{};
|
||||
std::memcpy(¶ms, input.data(), input.size());
|
||||
|
||||
LOG_DEBUG(Service_NVDRV, "called, pages={:X}, page_size={:X}, flags={:X}", params.pages,
|
||||
params.page_size, params.flags);
|
||||
|
||||
auto& gpu = system.GPU();
|
||||
const u64 size{static_cast<u64>(params.pages) * static_cast<u64>(params.page_size)};
|
||||
if (params.flags & 1) {
|
||||
params.offset = gpu.MemoryManager().AllocateSpace(params.offset, size, 1);
|
||||
const auto size{static_cast<u64>(params.pages) * static_cast<u64>(params.page_size)};
|
||||
if ((params.flags & AddressSpaceFlags::FixedOffset) != AddressSpaceFlags::None) {
|
||||
params.offset = *system.GPU().MemoryManager().AllocateFixed(params.offset, size);
|
||||
} else {
|
||||
params.offset = gpu.MemoryManager().AllocateSpace(size, params.align);
|
||||
params.offset = system.GPU().MemoryManager().Allocate(size, params.align);
|
||||
}
|
||||
|
||||
auto result{NvErrCodes::Success};
|
||||
if (!params.offset) {
|
||||
LOG_CRITICAL(Service_NVDRV, "allocation failed for size {}", size);
|
||||
result = NvErrCodes::OutOfMemory;
|
||||
}
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return 0;
|
||||
return result;
|
||||
}
|
||||
|
||||
u32 nvhost_as_gpu::Remap(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
std::size_t num_entries = input.size() / sizeof(IoctlRemapEntry);
|
||||
const auto num_entries = input.size() / sizeof(IoctlRemapEntry);
|
||||
|
||||
LOG_WARNING(Service_NVDRV, "(STUBBED) called, num_entries=0x{:X}", num_entries);
|
||||
LOG_DEBUG(Service_NVDRV, "called, num_entries=0x{:X}", num_entries);
|
||||
|
||||
auto result{NvErrCodes::Success};
|
||||
std::vector<IoctlRemapEntry> entries(num_entries);
|
||||
std::memcpy(entries.data(), input.data(), input.size());
|
||||
|
||||
auto& gpu = system.GPU();
|
||||
for (const auto& entry : entries) {
|
||||
LOG_WARNING(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}",
|
||||
entry.offset, entry.nvmap_handle, entry.pages);
|
||||
GPUVAddr offset = static_cast<GPUVAddr>(entry.offset) << 0x10;
|
||||
auto object = nvmap_dev->GetObject(entry.nvmap_handle);
|
||||
LOG_DEBUG(Service_NVDRV, "remap entry, offset=0x{:X} handle=0x{:X} pages=0x{:X}",
|
||||
entry.offset, entry.nvmap_handle, entry.pages);
|
||||
|
||||
const auto object{nvmap_dev->GetObject(entry.nvmap_handle)};
|
||||
if (!object) {
|
||||
LOG_CRITICAL(Service_NVDRV, "nvmap {} is an invalid handle!", entry.nvmap_handle);
|
||||
std::memcpy(output.data(), entries.data(), output.size());
|
||||
return static_cast<u32>(NvErrCodes::InvalidNmapHandle);
|
||||
LOG_CRITICAL(Service_NVDRV, "invalid nvmap_handle={:X}", entry.nvmap_handle);
|
||||
result = NvErrCodes::InvalidInput;
|
||||
break;
|
||||
}
|
||||
|
||||
ASSERT(object->status == nvmap::Object::Status::Allocated);
|
||||
const auto offset{static_cast<GPUVAddr>(entry.offset) << 0x10};
|
||||
const auto size{static_cast<u64>(entry.pages) << 0x10};
|
||||
const auto map_offset{static_cast<u64>(entry.map_offset) << 0x10};
|
||||
const auto addr{system.GPU().MemoryManager().Map(object->addr + map_offset, offset, size)};
|
||||
|
||||
const u64 size = static_cast<u64>(entry.pages) << 0x10;
|
||||
ASSERT(size <= object->size);
|
||||
const u64 map_offset = static_cast<u64>(entry.map_offset) << 0x10;
|
||||
|
||||
const GPUVAddr returned =
|
||||
gpu.MemoryManager().MapBufferEx(object->addr + map_offset, offset, size);
|
||||
ASSERT(returned == offset);
|
||||
if (!addr) {
|
||||
LOG_CRITICAL(Service_NVDRV, "map returned an invalid address!");
|
||||
result = NvErrCodes::InvalidInput;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
std::memcpy(output.data(), entries.data(), output.size());
|
||||
return 0;
|
||||
return result;
|
||||
}
|
||||
|
||||
u32 nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
|
|
@ -126,44 +137,76 @@ u32 nvhost_as_gpu::MapBufferEx(const std::vector<u8>& input, std::vector<u8>& ou
|
|||
params.flags, params.nvmap_handle, params.buffer_offset, params.mapping_size,
|
||||
params.offset);
|
||||
|
||||
if (!params.nvmap_handle) {
|
||||
return 0;
|
||||
const auto object{nvmap_dev->GetObject(params.nvmap_handle)};
|
||||
if (!object) {
|
||||
LOG_CRITICAL(Service_NVDRV, "invalid nvmap_handle={:X}", params.nvmap_handle);
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return NvErrCodes::InvalidInput;
|
||||
}
|
||||
|
||||
auto object = nvmap_dev->GetObject(params.nvmap_handle);
|
||||
ASSERT(object);
|
||||
|
||||
// We can only map objects that have already been assigned a CPU address.
|
||||
ASSERT(object->status == nvmap::Object::Status::Allocated);
|
||||
|
||||
ASSERT(params.buffer_offset == 0);
|
||||
|
||||
// The real nvservices doesn't make a distinction between handles and ids, and
|
||||
// object can only have one handle and it will be the same as its id. Assert that this is the
|
||||
// case to prevent unexpected behavior.
|
||||
ASSERT(object->id == params.nvmap_handle);
|
||||
|
||||
auto& gpu = system.GPU();
|
||||
|
||||
if (params.flags & 1) {
|
||||
params.offset = gpu.MemoryManager().MapBufferEx(object->addr, params.offset, object->size);
|
||||
} else {
|
||||
params.offset = gpu.MemoryManager().MapBufferEx(object->addr, object->size);
|
||||
u64 page_size{params.page_size};
|
||||
if (!page_size) {
|
||||
page_size = object->align;
|
||||
}
|
||||
|
||||
// Create a new mapping entry for this operation.
|
||||
ASSERT_MSG(buffer_mappings.find(params.offset) == buffer_mappings.end(),
|
||||
"Offset is already mapped");
|
||||
if ((params.flags & AddressSpaceFlags::Remap) != AddressSpaceFlags::None) {
|
||||
if (const auto buffer_map{FindBufferMap(params.offset)}; buffer_map) {
|
||||
const auto cpu_addr{static_cast<VAddr>(buffer_map->CpuAddr() + params.buffer_offset)};
|
||||
const auto gpu_addr{static_cast<GPUVAddr>(params.offset + params.buffer_offset)};
|
||||
|
||||
BufferMapping mapping{};
|
||||
mapping.nvmap_handle = params.nvmap_handle;
|
||||
mapping.offset = params.offset;
|
||||
mapping.size = object->size;
|
||||
if (!gpu.MemoryManager().Map(cpu_addr, gpu_addr, params.mapping_size)) {
|
||||
LOG_CRITICAL(Service_NVDRV,
|
||||
"remap failed, flags={:X}, nvmap_handle={:X}, buffer_offset={}, "
|
||||
"mapping_size = {}, offset={}",
|
||||
params.flags, params.nvmap_handle, params.buffer_offset,
|
||||
params.mapping_size, params.offset);
|
||||
|
||||
buffer_mappings[params.offset] = mapping;
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return NvErrCodes::InvalidInput;
|
||||
}
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return NvErrCodes::Success;
|
||||
} else {
|
||||
LOG_CRITICAL(Service_NVDRV, "address not mapped offset={}", params.offset);
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return NvErrCodes::InvalidInput;
|
||||
}
|
||||
}
|
||||
|
||||
// We can only map objects that have already been assigned a CPU address.
|
||||
ASSERT(object->status == nvmap::Object::Status::Allocated);
|
||||
|
||||
const auto physical_address{object->addr + params.buffer_offset};
|
||||
u64 size{params.mapping_size};
|
||||
if (!size) {
|
||||
size = object->size;
|
||||
}
|
||||
|
||||
const bool is_alloc{(params.flags & AddressSpaceFlags::FixedOffset) == AddressSpaceFlags::None};
|
||||
if (is_alloc) {
|
||||
params.offset = gpu.MemoryManager().MapAllocate(physical_address, size, page_size);
|
||||
} else {
|
||||
params.offset = gpu.MemoryManager().Map(physical_address, params.offset, size);
|
||||
}
|
||||
|
||||
auto result{NvErrCodes::Success};
|
||||
if (!params.offset) {
|
||||
LOG_CRITICAL(Service_NVDRV, "failed to map size={}", size);
|
||||
result = NvErrCodes::InvalidInput;
|
||||
} else {
|
||||
AddBufferMap(params.offset, size, physical_address, is_alloc);
|
||||
}
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return 0;
|
||||
return result;
|
||||
}
|
||||
|
||||
u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
|
|
@ -172,24 +215,20 @@ u32 nvhost_as_gpu::UnmapBuffer(const std::vector<u8>& input, std::vector<u8>& ou
|
|||
|
||||
LOG_DEBUG(Service_NVDRV, "called, offset=0x{:X}", params.offset);
|
||||
|
||||
const auto itr = buffer_mappings.find(params.offset);
|
||||
if (itr == buffer_mappings.end()) {
|
||||
LOG_WARNING(Service_NVDRV, "Tried to unmap an invalid offset 0x{:X}", params.offset);
|
||||
// Hardware tests shows that unmapping an already unmapped buffer always returns successful
|
||||
// and doesn't fail.
|
||||
return 0;
|
||||
if (const auto size{RemoveBufferMap(params.offset)}; size) {
|
||||
system.GPU().MemoryManager().Unmap(params.offset, *size);
|
||||
} else {
|
||||
LOG_ERROR(Service_NVDRV, "invalid offset=0x{:X}", params.offset);
|
||||
}
|
||||
|
||||
params.offset = system.GPU().MemoryManager().UnmapBuffer(params.offset, itr->second.size);
|
||||
buffer_mappings.erase(itr->second.offset);
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return 0;
|
||||
return NvErrCodes::Success;
|
||||
}
|
||||
|
||||
u32 nvhost_as_gpu::BindChannel(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
IoctlBindChannel params{};
|
||||
std::memcpy(¶ms, input.data(), input.size());
|
||||
|
||||
LOG_DEBUG(Service_NVDRV, "called, fd={:X}", params.fd);
|
||||
|
||||
channel = params.fd;
|
||||
|
|
@ -199,6 +238,7 @@ u32 nvhost_as_gpu::BindChannel(const std::vector<u8>& input, std::vector<u8>& ou
|
|||
u32 nvhost_as_gpu::GetVARegions(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
IoctlGetVaRegions params{};
|
||||
std::memcpy(¶ms, input.data(), input.size());
|
||||
|
||||
LOG_WARNING(Service_NVDRV, "(STUBBED) called, buf_addr={:X}, buf_size={:X}", params.buf_addr,
|
||||
params.buf_size);
|
||||
|
||||
|
|
@ -210,9 +250,43 @@ u32 nvhost_as_gpu::GetVARegions(const std::vector<u8>& input, std::vector<u8>& o
|
|||
params.regions[1].offset = 0x04000000;
|
||||
params.regions[1].page_size = 0x10000;
|
||||
params.regions[1].pages = 0x1bffff;
|
||||
|
||||
// TODO(ogniK): This probably can stay stubbed but should add support way way later
|
||||
|
||||
std::memcpy(output.data(), ¶ms, output.size());
|
||||
return 0;
|
||||
}
|
||||
|
||||
std::optional<nvhost_as_gpu::BufferMap> nvhost_as_gpu::FindBufferMap(GPUVAddr gpu_addr) const {
|
||||
const auto end{buffer_mappings.upper_bound(gpu_addr)};
|
||||
for (auto iter{buffer_mappings.begin()}; iter != end; ++iter) {
|
||||
if (gpu_addr >= iter->second.StartAddr() && gpu_addr < iter->second.EndAddr()) {
|
||||
return iter->second;
|
||||
}
|
||||
}
|
||||
|
||||
return {};
|
||||
}
|
||||
|
||||
void nvhost_as_gpu::AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr,
|
||||
bool is_allocated) {
|
||||
buffer_mappings[gpu_addr] = {gpu_addr, size, cpu_addr, is_allocated};
|
||||
}
|
||||
|
||||
std::optional<std::size_t> nvhost_as_gpu::RemoveBufferMap(GPUVAddr gpu_addr) {
|
||||
if (const auto iter{buffer_mappings.find(gpu_addr)}; iter != buffer_mappings.end()) {
|
||||
std::size_t size{};
|
||||
|
||||
if (iter->second.IsAllocated()) {
|
||||
size = iter->second.Size();
|
||||
}
|
||||
|
||||
buffer_mappings.erase(iter);
|
||||
|
||||
return size;
|
||||
}
|
||||
|
||||
return {};
|
||||
}
|
||||
|
||||
} // namespace Service::Nvidia::Devices
|
||||
|
|
|
|||
|
|
@ -4,9 +4,12 @@
|
|||
|
||||
#pragma once
|
||||
|
||||
#include <map>
|
||||
#include <memory>
|
||||
#include <unordered_map>
|
||||
#include <optional>
|
||||
#include <vector>
|
||||
|
||||
#include "common/common_funcs.h"
|
||||
#include "common/common_types.h"
|
||||
#include "common/swap.h"
|
||||
#include "core/hle/service/nvdrv/devices/nvdevice.h"
|
||||
|
|
@ -15,6 +18,13 @@ namespace Service::Nvidia::Devices {
|
|||
|
||||
class nvmap;
|
||||
|
||||
enum class AddressSpaceFlags : u32 {
|
||||
None = 0x0,
|
||||
FixedOffset = 0x1,
|
||||
Remap = 0x100,
|
||||
};
|
||||
DECLARE_ENUM_FLAG_OPERATORS(AddressSpaceFlags);
|
||||
|
||||
class nvhost_as_gpu final : public nvdevice {
|
||||
public:
|
||||
explicit nvhost_as_gpu(Core::System& system, std::shared_ptr<nvmap> nvmap_dev);
|
||||
|
|
@ -25,6 +35,45 @@ public:
|
|||
IoctlVersion version) override;
|
||||
|
||||
private:
|
||||
class BufferMap final {
|
||||
public:
|
||||
constexpr BufferMap() = default;
|
||||
|
||||
constexpr BufferMap(GPUVAddr start_addr, std::size_t size)
|
||||
: start_addr{start_addr}, end_addr{start_addr + size} {}
|
||||
|
||||
constexpr BufferMap(GPUVAddr start_addr, std::size_t size, VAddr cpu_addr,
|
||||
bool is_allocated)
|
||||
: start_addr{start_addr}, end_addr{start_addr + size}, cpu_addr{cpu_addr},
|
||||
is_allocated{is_allocated} {}
|
||||
|
||||
constexpr VAddr StartAddr() const {
|
||||
return start_addr;
|
||||
}
|
||||
|
||||
constexpr VAddr EndAddr() const {
|
||||
return end_addr;
|
||||
}
|
||||
|
||||
constexpr std::size_t Size() const {
|
||||
return end_addr - start_addr;
|
||||
}
|
||||
|
||||
constexpr VAddr CpuAddr() const {
|
||||
return cpu_addr;
|
||||
}
|
||||
|
||||
constexpr bool IsAllocated() const {
|
||||
return is_allocated;
|
||||
}
|
||||
|
||||
private:
|
||||
GPUVAddr start_addr{};
|
||||
GPUVAddr end_addr{};
|
||||
VAddr cpu_addr{};
|
||||
bool is_allocated{};
|
||||
};
|
||||
|
||||
enum class IoctlCommand : u32_le {
|
||||
IocInitalizeExCommand = 0x40284109,
|
||||
IocAllocateSpaceCommand = 0xC0184102,
|
||||
|
|
@ -49,7 +98,7 @@ private:
|
|||
struct IoctlAllocSpace {
|
||||
u32_le pages;
|
||||
u32_le page_size;
|
||||
u32_le flags;
|
||||
AddressSpaceFlags flags;
|
||||
INSERT_PADDING_WORDS(1);
|
||||
union {
|
||||
u64_le offset;
|
||||
|
|
@ -69,18 +118,18 @@ private:
|
|||
static_assert(sizeof(IoctlRemapEntry) == 20, "IoctlRemapEntry is incorrect size");
|
||||
|
||||
struct IoctlMapBufferEx {
|
||||
u32_le flags; // bit0: fixed_offset, bit2: cacheable
|
||||
u32_le kind; // -1 is default
|
||||
AddressSpaceFlags flags; // bit0: fixed_offset, bit2: cacheable
|
||||
u32_le kind; // -1 is default
|
||||
u32_le nvmap_handle;
|
||||
u32_le page_size; // 0 means don't care
|
||||
u64_le buffer_offset;
|
||||
s64_le buffer_offset;
|
||||
u64_le mapping_size;
|
||||
u64_le offset;
|
||||
s64_le offset;
|
||||
};
|
||||
static_assert(sizeof(IoctlMapBufferEx) == 40, "IoctlMapBufferEx is incorrect size");
|
||||
|
||||
struct IoctlUnmapBuffer {
|
||||
u64_le offset;
|
||||
s64_le offset;
|
||||
};
|
||||
static_assert(sizeof(IoctlUnmapBuffer) == 8, "IoctlUnmapBuffer is incorrect size");
|
||||
|
||||
|
|
@ -106,15 +155,6 @@ private:
|
|||
static_assert(sizeof(IoctlGetVaRegions) == 16 + sizeof(IoctlVaRegion) * 2,
|
||||
"IoctlGetVaRegions is incorrect size");
|
||||
|
||||
struct BufferMapping {
|
||||
u64 offset;
|
||||
u64 size;
|
||||
u32 nvmap_handle;
|
||||
};
|
||||
|
||||
/// Map containing the nvmap object mappings in GPU memory.
|
||||
std::unordered_map<u64, BufferMapping> buffer_mappings;
|
||||
|
||||
u32 channel{};
|
||||
|
||||
u32 InitalizeEx(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
|
|
@ -125,7 +165,14 @@ private:
|
|||
u32 BindChannel(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
u32 GetVARegions(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
|
||||
std::optional<BufferMap> FindBufferMap(GPUVAddr gpu_addr) const;
|
||||
void AddBufferMap(GPUVAddr gpu_addr, std::size_t size, VAddr cpu_addr, bool is_allocated);
|
||||
std::optional<std::size_t> RemoveBufferMap(GPUVAddr gpu_addr);
|
||||
|
||||
std::shared_ptr<nvmap> nvmap_dev;
|
||||
|
||||
// This is expected to be ordered, therefore we must use a map, not unordered_map
|
||||
std::map<GPUVAddr, BufferMap> buffer_mappings;
|
||||
};
|
||||
|
||||
} // namespace Service::Nvidia::Devices
|
||||
|
|
|
|||
|
|
@ -18,7 +18,12 @@ enum {
|
|||
};
|
||||
}
|
||||
|
||||
nvmap::nvmap(Core::System& system) : nvdevice(system) {}
|
||||
nvmap::nvmap(Core::System& system) : nvdevice(system) {
|
||||
// Handle 0 appears to be used when remapping, so we create a placeholder empty nvmap object to
|
||||
// represent this.
|
||||
CreateObject(0);
|
||||
}
|
||||
|
||||
nvmap::~nvmap() = default;
|
||||
|
||||
VAddr nvmap::GetObjectAddress(u32 handle) const {
|
||||
|
|
@ -50,6 +55,21 @@ u32 nvmap::ioctl(Ioctl command, const std::vector<u8>& input, const std::vector<
|
|||
return 0;
|
||||
}
|
||||
|
||||
u32 nvmap::CreateObject(u32 size) {
|
||||
// Create a new nvmap object and obtain a handle to it.
|
||||
auto object = std::make_shared<Object>();
|
||||
object->id = next_id++;
|
||||
object->size = size;
|
||||
object->status = Object::Status::Created;
|
||||
object->refcount = 1;
|
||||
|
||||
const u32 handle = next_handle++;
|
||||
|
||||
handles.insert_or_assign(handle, std::move(object));
|
||||
|
||||
return handle;
|
||||
}
|
||||
|
||||
u32 nvmap::IocCreate(const std::vector<u8>& input, std::vector<u8>& output) {
|
||||
IocCreateParams params;
|
||||
std::memcpy(¶ms, input.data(), sizeof(params));
|
||||
|
|
@ -59,17 +79,8 @@ u32 nvmap::IocCreate(const std::vector<u8>& input, std::vector<u8>& output) {
|
|||
LOG_ERROR(Service_NVDRV, "Size is 0");
|
||||
return static_cast<u32>(NvErrCodes::InvalidValue);
|
||||
}
|
||||
// Create a new nvmap object and obtain a handle to it.
|
||||
auto object = std::make_shared<Object>();
|
||||
object->id = next_id++;
|
||||
object->size = params.size;
|
||||
object->status = Object::Status::Created;
|
||||
object->refcount = 1;
|
||||
|
||||
u32 handle = next_handle++;
|
||||
handles[handle] = std::move(object);
|
||||
|
||||
params.handle = handle;
|
||||
params.handle = CreateObject(params.size);
|
||||
|
||||
std::memcpy(output.data(), ¶ms, sizeof(params));
|
||||
return 0;
|
||||
|
|
|
|||
|
|
@ -49,10 +49,10 @@ public:
|
|||
|
||||
private:
|
||||
/// Id to use for the next handle that is created.
|
||||
u32 next_handle = 1;
|
||||
u32 next_handle = 0;
|
||||
|
||||
/// Id to use for the next object that is created.
|
||||
u32 next_id = 1;
|
||||
u32 next_id = 0;
|
||||
|
||||
/// Mapping of currently allocated handles to the objects they represent.
|
||||
std::unordered_map<u32, std::shared_ptr<Object>> handles;
|
||||
|
|
@ -119,6 +119,8 @@ private:
|
|||
};
|
||||
static_assert(sizeof(IocGetIdParams) == 8, "IocGetIdParams has wrong size");
|
||||
|
||||
u32 CreateObject(u32 size);
|
||||
|
||||
u32 IocCreate(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
u32 IocAlloc(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
u32 IocGetId(const std::vector<u8>& input, std::vector<u8>& output);
|
||||
|
|
|
|||
|
|
@ -4,7 +4,6 @@
|
|||
|
||||
#include "common/alignment.h"
|
||||
#include "common/assert.h"
|
||||
#include "common/logging/log.h"
|
||||
#include "core/core.h"
|
||||
#include "core/hle/kernel/memory/page_table.h"
|
||||
#include "core/hle/kernel/process.h"
|
||||
|
|
@ -16,121 +15,137 @@
|
|||
namespace Tegra {
|
||||
|
||||
MemoryManager::MemoryManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer)
|
||||
: rasterizer{rasterizer}, system{system} {
|
||||
page_table.Resize(address_space_width, page_bits, false);
|
||||
|
||||
// Initialize the map with a single free region covering the entire managed space.
|
||||
VirtualMemoryArea initial_vma;
|
||||
initial_vma.size = address_space_end;
|
||||
vma_map.emplace(initial_vma.base, initial_vma);
|
||||
|
||||
UpdatePageTableForVMA(initial_vma);
|
||||
}
|
||||
: system{system}, rasterizer{rasterizer}, page_table(page_table_size) {}
|
||||
|
||||
MemoryManager::~MemoryManager() = default;
|
||||
|
||||
GPUVAddr MemoryManager::AllocateSpace(u64 size, u64 align) {
|
||||
const u64 aligned_size{Common::AlignUp(size, page_size)};
|
||||
const GPUVAddr gpu_addr{FindFreeRegion(address_space_base, aligned_size)};
|
||||
|
||||
AllocateMemory(gpu_addr, 0, aligned_size);
|
||||
|
||||
GPUVAddr MemoryManager::UpdateRange(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size) {
|
||||
u64 remaining_size{size};
|
||||
for (u64 offset{}; offset < size; offset += page_size) {
|
||||
if (remaining_size < page_size) {
|
||||
SetPageEntry(gpu_addr + offset, page_entry + offset, remaining_size);
|
||||
} else {
|
||||
SetPageEntry(gpu_addr + offset, page_entry + offset);
|
||||
}
|
||||
remaining_size -= page_size;
|
||||
}
|
||||
return gpu_addr;
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::AllocateSpace(GPUVAddr gpu_addr, u64 size, u64 align) {
|
||||
const u64 aligned_size{Common::AlignUp(size, page_size)};
|
||||
|
||||
AllocateMemory(gpu_addr, 0, aligned_size);
|
||||
|
||||
return gpu_addr;
|
||||
GPUVAddr MemoryManager::Map(VAddr cpu_addr, GPUVAddr gpu_addr, std::size_t size) {
|
||||
return UpdateRange(gpu_addr, cpu_addr, size);
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, u64 size) {
|
||||
const u64 aligned_size{Common::AlignUp(size, page_size)};
|
||||
const GPUVAddr gpu_addr{FindFreeRegion(address_space_base, aligned_size)};
|
||||
|
||||
MapBackingMemory(gpu_addr, system.Memory().GetPointer(cpu_addr), aligned_size, cpu_addr);
|
||||
ASSERT(
|
||||
system.CurrentProcess()->PageTable().LockForDeviceAddressSpace(cpu_addr, size).IsSuccess());
|
||||
|
||||
return gpu_addr;
|
||||
GPUVAddr MemoryManager::MapAllocate(VAddr cpu_addr, std::size_t size, std::size_t align) {
|
||||
return Map(cpu_addr, *FindFreeRange(size, align), size);
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::MapBufferEx(VAddr cpu_addr, GPUVAddr gpu_addr, u64 size) {
|
||||
ASSERT((gpu_addr & page_mask) == 0);
|
||||
|
||||
const u64 aligned_size{Common::AlignUp(size, page_size)};
|
||||
|
||||
MapBackingMemory(gpu_addr, system.Memory().GetPointer(cpu_addr), aligned_size, cpu_addr);
|
||||
ASSERT(
|
||||
system.CurrentProcess()->PageTable().LockForDeviceAddressSpace(cpu_addr, size).IsSuccess());
|
||||
return gpu_addr;
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::UnmapBuffer(GPUVAddr gpu_addr, u64 size) {
|
||||
ASSERT((gpu_addr & page_mask) == 0);
|
||||
|
||||
const u64 aligned_size{Common::AlignUp(size, page_size)};
|
||||
const auto cpu_addr = GpuToCpuAddress(gpu_addr);
|
||||
ASSERT(cpu_addr);
|
||||
void MemoryManager::Unmap(GPUVAddr gpu_addr, std::size_t size) {
|
||||
if (!size) {
|
||||
return;
|
||||
}
|
||||
|
||||
// Flush and invalidate through the GPU interface, to be asynchronous if possible.
|
||||
system.GPU().FlushAndInvalidateRegion(*cpu_addr, aligned_size);
|
||||
system.GPU().FlushAndInvalidateRegion(*GpuToCpuAddress(gpu_addr), size);
|
||||
|
||||
UpdateRange(gpu_addr, PageEntry::State::Unmapped, size);
|
||||
}
|
||||
|
||||
std::optional<GPUVAddr> MemoryManager::AllocateFixed(GPUVAddr gpu_addr, std::size_t size) {
|
||||
for (u64 offset{}; offset < size; offset += page_size) {
|
||||
if (!GetPageEntry(gpu_addr + offset).IsUnmapped()) {
|
||||
return {};
|
||||
}
|
||||
}
|
||||
|
||||
return UpdateRange(gpu_addr, PageEntry::State::Allocated, size);
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::Allocate(std::size_t size, std::size_t align) {
|
||||
return *AllocateFixed(*FindFreeRange(size, align), size);
|
||||
}
|
||||
|
||||
void MemoryManager::TryLockPage(PageEntry page_entry, std::size_t size) {
|
||||
if (!page_entry.IsValid()) {
|
||||
return;
|
||||
}
|
||||
|
||||
UnmapRange(gpu_addr, aligned_size);
|
||||
ASSERT(system.CurrentProcess()
|
||||
->PageTable()
|
||||
.UnlockForDeviceAddressSpace(cpu_addr.value(), size)
|
||||
.LockForDeviceAddressSpace(page_entry.ToAddress(), size)
|
||||
.IsSuccess());
|
||||
|
||||
return gpu_addr;
|
||||
}
|
||||
|
||||
GPUVAddr MemoryManager::FindFreeRegion(GPUVAddr region_start, u64 size) const {
|
||||
// Find the first Free VMA.
|
||||
const VMAHandle vma_handle{
|
||||
std::find_if(vma_map.begin(), vma_map.end(), [region_start, size](const auto& vma) {
|
||||
if (vma.second.type != VirtualMemoryArea::Type::Unmapped) {
|
||||
return false;
|
||||
void MemoryManager::TryUnlockPage(PageEntry page_entry, std::size_t size) {
|
||||
if (!page_entry.IsValid()) {
|
||||
return;
|
||||
}
|
||||
|
||||
ASSERT(system.CurrentProcess()
|
||||
->PageTable()
|
||||
.UnlockForDeviceAddressSpace(page_entry.ToAddress(), size)
|
||||
.IsSuccess());
|
||||
}
|
||||
|
||||
PageEntry MemoryManager::GetPageEntry(GPUVAddr gpu_addr) const {
|
||||
return page_table[PageEntryIndex(gpu_addr)];
|
||||
}
|
||||
|
||||
void MemoryManager::SetPageEntry(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size) {
|
||||
// TODO(bunnei): We should lock/unlock device regions. This currently causes issues due to
|
||||
// improper tracking, but should be fixed in the future.
|
||||
|
||||
//// Unlock the old page
|
||||
// TryUnlockPage(page_table[PageEntryIndex(gpu_addr)], size);
|
||||
|
||||
//// Lock the new page
|
||||
// TryLockPage(page_entry, size);
|
||||
|
||||
page_table[PageEntryIndex(gpu_addr)] = page_entry;
|
||||
}
|
||||
|
||||
std::optional<GPUVAddr> MemoryManager::FindFreeRange(std::size_t size, std::size_t align) const {
|
||||
if (!align) {
|
||||
align = page_size;
|
||||
} else {
|
||||
align = Common::AlignUp(align, page_size);
|
||||
}
|
||||
|
||||
u64 available_size{};
|
||||
GPUVAddr gpu_addr{address_space_start};
|
||||
while (gpu_addr + available_size < address_space_size) {
|
||||
if (GetPageEntry(gpu_addr + available_size).IsUnmapped()) {
|
||||
available_size += page_size;
|
||||
|
||||
if (available_size >= size) {
|
||||
return gpu_addr;
|
||||
}
|
||||
} else {
|
||||
gpu_addr += available_size + page_size;
|
||||
available_size = 0;
|
||||
|
||||
const VAddr vma_end{vma.second.base + vma.second.size};
|
||||
return vma_end > region_start && vma_end >= region_start + size;
|
||||
})};
|
||||
|
||||
if (vma_handle == vma_map.end()) {
|
||||
return {};
|
||||
}
|
||||
|
||||
return std::max(region_start, vma_handle->second.base);
|
||||
}
|
||||
|
||||
bool MemoryManager::IsAddressValid(GPUVAddr addr) const {
|
||||
return (addr >> page_bits) < page_table.pointers.size();
|
||||
}
|
||||
|
||||
std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr addr) const {
|
||||
if (!IsAddressValid(addr)) {
|
||||
return {};
|
||||
}
|
||||
|
||||
const VAddr cpu_addr{page_table.backing_addr[addr >> page_bits]};
|
||||
if (cpu_addr) {
|
||||
return cpu_addr + (addr & page_mask);
|
||||
const auto remainder{gpu_addr % align};
|
||||
if (remainder) {
|
||||
gpu_addr = (gpu_addr - remainder) + align;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
return {};
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
T MemoryManager::Read(GPUVAddr addr) const {
|
||||
if (!IsAddressValid(addr)) {
|
||||
std::optional<VAddr> MemoryManager::GpuToCpuAddress(GPUVAddr gpu_addr) const {
|
||||
const auto page_entry{GetPageEntry(gpu_addr)};
|
||||
if (!page_entry.IsValid()) {
|
||||
return {};
|
||||
}
|
||||
|
||||
const u8* page_pointer{GetPointer(addr)};
|
||||
if (page_pointer) {
|
||||
return page_entry.ToAddress() + (gpu_addr & page_mask);
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
T MemoryManager::Read(GPUVAddr addr) const {
|
||||
if (auto page_pointer{GetPointer(addr)}; page_pointer) {
|
||||
// NOTE: Avoid adding any extra logic to this fast-path block
|
||||
T value;
|
||||
std::memcpy(&value, page_pointer, sizeof(T));
|
||||
|
|
@ -144,12 +159,7 @@ T MemoryManager::Read(GPUVAddr addr) const {
|
|||
|
||||
template <typename T>
|
||||
void MemoryManager::Write(GPUVAddr addr, T data) {
|
||||
if (!IsAddressValid(addr)) {
|
||||
return;
|
||||
}
|
||||
|
||||
u8* page_pointer{GetPointer(addr)};
|
||||
if (page_pointer) {
|
||||
if (auto page_pointer{GetPointer(addr)}; page_pointer) {
|
||||
// NOTE: Avoid adding any extra logic to this fast-path block
|
||||
std::memcpy(page_pointer, &data, sizeof(T));
|
||||
return;
|
||||
|
|
@ -167,66 +177,49 @@ template void MemoryManager::Write<u16>(GPUVAddr addr, u16 data);
|
|||
template void MemoryManager::Write<u32>(GPUVAddr addr, u32 data);
|
||||
template void MemoryManager::Write<u64>(GPUVAddr addr, u64 data);
|
||||
|
||||
u8* MemoryManager::GetPointer(GPUVAddr addr) {
|
||||
if (!IsAddressValid(addr)) {
|
||||
u8* MemoryManager::GetPointer(GPUVAddr gpu_addr) {
|
||||
if (!GetPageEntry(gpu_addr).IsValid()) {
|
||||
return {};
|
||||
}
|
||||
|
||||
auto& memory = system.Memory();
|
||||
|
||||
const VAddr page_addr{page_table.backing_addr[addr >> page_bits]};
|
||||
|
||||
if (page_addr != 0) {
|
||||
return memory.GetPointer(page_addr + (addr & page_mask));
|
||||
}
|
||||
|
||||
LOG_ERROR(HW_GPU, "Unknown GetPointer @ 0x{:016X}", addr);
|
||||
return {};
|
||||
}
|
||||
|
||||
const u8* MemoryManager::GetPointer(GPUVAddr addr) const {
|
||||
if (!IsAddressValid(addr)) {
|
||||
const auto address{GpuToCpuAddress(gpu_addr)};
|
||||
if (!address) {
|
||||
return {};
|
||||
}
|
||||
|
||||
const auto& memory = system.Memory();
|
||||
return system.Memory().GetPointer(*address);
|
||||
}
|
||||
|
||||
const VAddr page_addr{page_table.backing_addr[addr >> page_bits]};
|
||||
|
||||
if (page_addr != 0) {
|
||||
return memory.GetPointer(page_addr + (addr & page_mask));
|
||||
const u8* MemoryManager::GetPointer(GPUVAddr gpu_addr) const {
|
||||
if (!GetPageEntry(gpu_addr).IsValid()) {
|
||||
return {};
|
||||
}
|
||||
|
||||
LOG_ERROR(HW_GPU, "Unknown GetPointer @ 0x{:016X}", addr);
|
||||
return {};
|
||||
const auto address{GpuToCpuAddress(gpu_addr)};
|
||||
if (!address) {
|
||||
return {};
|
||||
}
|
||||
|
||||
return system.Memory().GetPointer(*address);
|
||||
}
|
||||
|
||||
bool MemoryManager::IsBlockContinuous(const GPUVAddr start, const std::size_t size) const {
|
||||
const std::size_t inner_size = size - 1;
|
||||
const GPUVAddr end = start + inner_size;
|
||||
const auto host_ptr_start = reinterpret_cast<std::uintptr_t>(GetPointer(start));
|
||||
const auto host_ptr_end = reinterpret_cast<std::uintptr_t>(GetPointer(end));
|
||||
const auto range = static_cast<std::size_t>(host_ptr_end - host_ptr_start);
|
||||
return range == inner_size;
|
||||
}
|
||||
|
||||
void MemoryManager::ReadBlock(GPUVAddr gpu_src_addr, void* dest_buffer,
|
||||
const std::size_t size) const {
|
||||
void MemoryManager::ReadBlock(GPUVAddr gpu_src_addr, void* dest_buffer, std::size_t size) const {
|
||||
std::size_t remaining_size{size};
|
||||
std::size_t page_index{gpu_src_addr >> page_bits};
|
||||
std::size_t page_offset{gpu_src_addr & page_mask};
|
||||
|
||||
auto& memory = system.Memory();
|
||||
|
||||
while (remaining_size > 0) {
|
||||
const std::size_t copy_amount{
|
||||
std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
|
||||
|
||||
const VAddr src_addr{page_table.backing_addr[page_index] + page_offset};
|
||||
// Flush must happen on the rasterizer interface, such that memory is always synchronous
|
||||
// when it is read (even when in asynchronous GPU mode). Fixes Dead Cells title menu.
|
||||
rasterizer.FlushRegion(src_addr, copy_amount);
|
||||
memory.ReadBlockUnsafe(src_addr, dest_buffer, copy_amount);
|
||||
if (const auto page_addr{GpuToCpuAddress(page_index << page_bits)}; page_addr) {
|
||||
const auto src_addr{*page_addr + page_offset};
|
||||
|
||||
// Flush must happen on the rasterizer interface, such that memory is always synchronous
|
||||
// when it is read (even when in asynchronous GPU mode). Fixes Dead Cells title menu.
|
||||
rasterizer.FlushRegion(src_addr, copy_amount);
|
||||
system.Memory().ReadBlockUnsafe(src_addr, dest_buffer, copy_amount);
|
||||
}
|
||||
|
||||
page_index++;
|
||||
page_offset = 0;
|
||||
|
|
@ -241,18 +234,17 @@ void MemoryManager::ReadBlockUnsafe(GPUVAddr gpu_src_addr, void* dest_buffer,
|
|||
std::size_t page_index{gpu_src_addr >> page_bits};
|
||||
std::size_t page_offset{gpu_src_addr & page_mask};
|
||||
|
||||
auto& memory = system.Memory();
|
||||
|
||||
while (remaining_size > 0) {
|
||||
const std::size_t copy_amount{
|
||||
std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
|
||||
const u8* page_pointer = page_table.pointers[page_index];
|
||||
if (page_pointer) {
|
||||
const VAddr src_addr{page_table.backing_addr[page_index] + page_offset};
|
||||
memory.ReadBlockUnsafe(src_addr, dest_buffer, copy_amount);
|
||||
|
||||
if (const auto page_addr{GpuToCpuAddress(page_index << page_bits)}; page_addr) {
|
||||
const auto src_addr{*page_addr + page_offset};
|
||||
system.Memory().ReadBlockUnsafe(src_addr, dest_buffer, copy_amount);
|
||||
} else {
|
||||
std::memset(dest_buffer, 0, copy_amount);
|
||||
}
|
||||
|
||||
page_index++;
|
||||
page_offset = 0;
|
||||
dest_buffer = static_cast<u8*>(dest_buffer) + copy_amount;
|
||||
|
|
@ -260,23 +252,23 @@ void MemoryManager::ReadBlockUnsafe(GPUVAddr gpu_src_addr, void* dest_buffer,
|
|||
}
|
||||
}
|
||||
|
||||
void MemoryManager::WriteBlock(GPUVAddr gpu_dest_addr, const void* src_buffer,
|
||||
const std::size_t size) {
|
||||
void MemoryManager::WriteBlock(GPUVAddr gpu_dest_addr, const void* src_buffer, std::size_t size) {
|
||||
std::size_t remaining_size{size};
|
||||
std::size_t page_index{gpu_dest_addr >> page_bits};
|
||||
std::size_t page_offset{gpu_dest_addr & page_mask};
|
||||
|
||||
auto& memory = system.Memory();
|
||||
|
||||
while (remaining_size > 0) {
|
||||
const std::size_t copy_amount{
|
||||
std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
|
||||
|
||||
const VAddr dest_addr{page_table.backing_addr[page_index] + page_offset};
|
||||
// Invalidate must happen on the rasterizer interface, such that memory is always
|
||||
// synchronous when it is written (even when in asynchronous GPU mode).
|
||||
rasterizer.InvalidateRegion(dest_addr, copy_amount);
|
||||
memory.WriteBlockUnsafe(dest_addr, src_buffer, copy_amount);
|
||||
if (const auto page_addr{GpuToCpuAddress(page_index << page_bits)}; page_addr) {
|
||||
const auto dest_addr{*page_addr + page_offset};
|
||||
|
||||
// Invalidate must happen on the rasterizer interface, such that memory is always
|
||||
// synchronous when it is written (even when in asynchronous GPU mode).
|
||||
rasterizer.InvalidateRegion(dest_addr, copy_amount);
|
||||
system.Memory().WriteBlockUnsafe(dest_addr, src_buffer, copy_amount);
|
||||
}
|
||||
|
||||
page_index++;
|
||||
page_offset = 0;
|
||||
|
|
@ -286,21 +278,20 @@ void MemoryManager::WriteBlock(GPUVAddr gpu_dest_addr, const void* src_buffer,
|
|||
}
|
||||
|
||||
void MemoryManager::WriteBlockUnsafe(GPUVAddr gpu_dest_addr, const void* src_buffer,
|
||||
const std::size_t size) {
|
||||
std::size_t size) {
|
||||
std::size_t remaining_size{size};
|
||||
std::size_t page_index{gpu_dest_addr >> page_bits};
|
||||
std::size_t page_offset{gpu_dest_addr & page_mask};
|
||||
|
||||
auto& memory = system.Memory();
|
||||
|
||||
while (remaining_size > 0) {
|
||||
const std::size_t copy_amount{
|
||||
std::min(static_cast<std::size_t>(page_size) - page_offset, remaining_size)};
|
||||
u8* page_pointer = page_table.pointers[page_index];
|
||||
if (page_pointer) {
|
||||
const VAddr dest_addr{page_table.backing_addr[page_index] + page_offset};
|
||||
memory.WriteBlockUnsafe(dest_addr, src_buffer, copy_amount);
|
||||
|
||||
if (const auto page_addr{GpuToCpuAddress(page_index << page_bits)}; page_addr) {
|
||||
const auto dest_addr{*page_addr + page_offset};
|
||||
system.Memory().WriteBlockUnsafe(dest_addr, src_buffer, copy_amount);
|
||||
}
|
||||
|
||||
page_index++;
|
||||
page_offset = 0;
|
||||
src_buffer = static_cast<const u8*>(src_buffer) + copy_amount;
|
||||
|
|
@ -308,273 +299,26 @@ void MemoryManager::WriteBlockUnsafe(GPUVAddr gpu_dest_addr, const void* src_buf
|
|||
}
|
||||
}
|
||||
|
||||
void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr,
|
||||
const std::size_t size) {
|
||||
void MemoryManager::CopyBlock(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size) {
|
||||
std::vector<u8> tmp_buffer(size);
|
||||
ReadBlock(gpu_src_addr, tmp_buffer.data(), size);
|
||||
WriteBlock(gpu_dest_addr, tmp_buffer.data(), size);
|
||||
}
|
||||
|
||||
void MemoryManager::CopyBlockUnsafe(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr,
|
||||
const std::size_t size) {
|
||||
std::size_t size) {
|
||||
std::vector<u8> tmp_buffer(size);
|
||||
ReadBlockUnsafe(gpu_src_addr, tmp_buffer.data(), size);
|
||||
WriteBlockUnsafe(gpu_dest_addr, tmp_buffer.data(), size);
|
||||
}
|
||||
|
||||
bool MemoryManager::IsGranularRange(GPUVAddr gpu_addr, std::size_t size) {
|
||||
const VAddr addr = page_table.backing_addr[gpu_addr >> page_bits];
|
||||
const std::size_t page = (addr & Core::Memory::PAGE_MASK) + size;
|
||||
const auto cpu_addr{GpuToCpuAddress(gpu_addr)};
|
||||
if (!cpu_addr) {
|
||||
return {};
|
||||
}
|
||||
const std::size_t page{(*cpu_addr & Core::Memory::PAGE_MASK) + size};
|
||||
return page <= Core::Memory::PAGE_SIZE;
|
||||
}
|
||||
|
||||
void MemoryManager::MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
|
||||
VAddr backing_addr) {
|
||||
LOG_DEBUG(HW_GPU, "Mapping {} onto {:016X}-{:016X}", fmt::ptr(memory), base * page_size,
|
||||
(base + size) * page_size);
|
||||
|
||||
const VAddr end{base + size};
|
||||
ASSERT_MSG(end <= page_table.pointers.size(), "out of range mapping at {:016X}",
|
||||
base + page_table.pointers.size());
|
||||
|
||||
if (memory == nullptr) {
|
||||
while (base != end) {
|
||||
page_table.pointers[base] = nullptr;
|
||||
page_table.backing_addr[base] = 0;
|
||||
|
||||
base += 1;
|
||||
}
|
||||
} else {
|
||||
while (base != end) {
|
||||
page_table.pointers[base] = memory;
|
||||
page_table.backing_addr[base] = backing_addr;
|
||||
|
||||
base += 1;
|
||||
memory += page_size;
|
||||
backing_addr += page_size;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void MemoryManager::MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr) {
|
||||
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
|
||||
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
|
||||
MapPages(base / page_size, size / page_size, target, Common::PageType::Memory, backing_addr);
|
||||
}
|
||||
|
||||
void MemoryManager::UnmapRegion(GPUVAddr base, u64 size) {
|
||||
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: {:016X}", size);
|
||||
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: {:016X}", base);
|
||||
MapPages(base / page_size, size / page_size, nullptr, Common::PageType::Unmapped);
|
||||
}
|
||||
|
||||
bool VirtualMemoryArea::CanBeMergedWith(const VirtualMemoryArea& next) const {
|
||||
ASSERT(base + size == next.base);
|
||||
if (type != next.type) {
|
||||
return {};
|
||||
}
|
||||
if (type == VirtualMemoryArea::Type::Allocated && (offset + size != next.offset)) {
|
||||
return {};
|
||||
}
|
||||
if (type == VirtualMemoryArea::Type::Mapped && backing_memory + size != next.backing_memory) {
|
||||
return {};
|
||||
}
|
||||
return true;
|
||||
}
|
||||
|
||||
MemoryManager::VMAHandle MemoryManager::FindVMA(GPUVAddr target) const {
|
||||
if (target >= address_space_end) {
|
||||
return vma_map.end();
|
||||
} else {
|
||||
return std::prev(vma_map.upper_bound(target));
|
||||
}
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::Allocate(VMAIter vma_handle) {
|
||||
VirtualMemoryArea& vma{vma_handle->second};
|
||||
|
||||
vma.type = VirtualMemoryArea::Type::Allocated;
|
||||
vma.backing_addr = 0;
|
||||
vma.backing_memory = {};
|
||||
UpdatePageTableForVMA(vma);
|
||||
|
||||
return MergeAdjacent(vma_handle);
|
||||
}
|
||||
|
||||
MemoryManager::VMAHandle MemoryManager::AllocateMemory(GPUVAddr target, std::size_t offset,
|
||||
u64 size) {
|
||||
|
||||
// This is the appropriately sized VMA that will turn into our allocation.
|
||||
VMAIter vma_handle{CarveVMA(target, size)};
|
||||
VirtualMemoryArea& vma{vma_handle->second};
|
||||
|
||||
ASSERT(vma.size == size);
|
||||
|
||||
vma.offset = offset;
|
||||
|
||||
return Allocate(vma_handle);
|
||||
}
|
||||
|
||||
MemoryManager::VMAHandle MemoryManager::MapBackingMemory(GPUVAddr target, u8* memory, u64 size,
|
||||
VAddr backing_addr) {
|
||||
// This is the appropriately sized VMA that will turn into our allocation.
|
||||
VMAIter vma_handle{CarveVMA(target, size)};
|
||||
VirtualMemoryArea& vma{vma_handle->second};
|
||||
|
||||
ASSERT(vma.size == size);
|
||||
|
||||
vma.type = VirtualMemoryArea::Type::Mapped;
|
||||
vma.backing_memory = memory;
|
||||
vma.backing_addr = backing_addr;
|
||||
UpdatePageTableForVMA(vma);
|
||||
|
||||
return MergeAdjacent(vma_handle);
|
||||
}
|
||||
|
||||
void MemoryManager::UnmapRange(GPUVAddr target, u64 size) {
|
||||
VMAIter vma{CarveVMARange(target, size)};
|
||||
const VAddr target_end{target + size};
|
||||
const VMAIter end{vma_map.end()};
|
||||
|
||||
// The comparison against the end of the range must be done using addresses since VMAs can be
|
||||
// merged during this process, causing invalidation of the iterators.
|
||||
while (vma != end && vma->second.base < target_end) {
|
||||
// Unmapped ranges return to allocated state and can be reused
|
||||
// This behavior is used by Super Mario Odyssey, Sonic Forces, and likely other games
|
||||
vma = std::next(Allocate(vma));
|
||||
}
|
||||
|
||||
ASSERT(FindVMA(target)->second.size >= size);
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::StripIterConstness(const VMAHandle& iter) {
|
||||
// This uses a neat C++ trick to convert a const_iterator to a regular iterator, given
|
||||
// non-const access to its container.
|
||||
return vma_map.erase(iter, iter); // Erases an empty range of elements
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::CarveVMA(GPUVAddr base, u64 size) {
|
||||
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: 0x{:016X}", size);
|
||||
ASSERT_MSG((base & page_mask) == 0, "non-page aligned base: 0x{:016X}", base);
|
||||
|
||||
VMAIter vma_handle{StripIterConstness(FindVMA(base))};
|
||||
if (vma_handle == vma_map.end()) {
|
||||
// Target address is outside the managed range
|
||||
return {};
|
||||
}
|
||||
|
||||
const VirtualMemoryArea& vma{vma_handle->second};
|
||||
if (vma.type == VirtualMemoryArea::Type::Mapped) {
|
||||
// Region is already allocated
|
||||
return vma_handle;
|
||||
}
|
||||
|
||||
const VAddr start_in_vma{base - vma.base};
|
||||
const VAddr end_in_vma{start_in_vma + size};
|
||||
|
||||
ASSERT_MSG(end_in_vma <= vma.size, "region size 0x{:016X} is less than required size 0x{:016X}",
|
||||
vma.size, end_in_vma);
|
||||
|
||||
if (end_in_vma < vma.size) {
|
||||
// Split VMA at the end of the allocated region
|
||||
SplitVMA(vma_handle, end_in_vma);
|
||||
}
|
||||
if (start_in_vma != 0) {
|
||||
// Split VMA at the start of the allocated region
|
||||
vma_handle = SplitVMA(vma_handle, start_in_vma);
|
||||
}
|
||||
|
||||
return vma_handle;
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::CarveVMARange(GPUVAddr target, u64 size) {
|
||||
ASSERT_MSG((size & page_mask) == 0, "non-page aligned size: 0x{:016X}", size);
|
||||
ASSERT_MSG((target & page_mask) == 0, "non-page aligned base: 0x{:016X}", target);
|
||||
|
||||
const VAddr target_end{target + size};
|
||||
ASSERT(target_end >= target);
|
||||
ASSERT(size > 0);
|
||||
|
||||
VMAIter begin_vma{StripIterConstness(FindVMA(target))};
|
||||
const VMAIter i_end{vma_map.lower_bound(target_end)};
|
||||
if (std::any_of(begin_vma, i_end, [](const auto& entry) {
|
||||
return entry.second.type == VirtualMemoryArea::Type::Unmapped;
|
||||
})) {
|
||||
return {};
|
||||
}
|
||||
|
||||
if (target != begin_vma->second.base) {
|
||||
begin_vma = SplitVMA(begin_vma, target - begin_vma->second.base);
|
||||
}
|
||||
|
||||
VMAIter end_vma{StripIterConstness(FindVMA(target_end))};
|
||||
if (end_vma != vma_map.end() && target_end != end_vma->second.base) {
|
||||
end_vma = SplitVMA(end_vma, target_end - end_vma->second.base);
|
||||
}
|
||||
|
||||
return begin_vma;
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::SplitVMA(VMAIter vma_handle, u64 offset_in_vma) {
|
||||
VirtualMemoryArea& old_vma{vma_handle->second};
|
||||
VirtualMemoryArea new_vma{old_vma}; // Make a copy of the VMA
|
||||
|
||||
// For now, don't allow no-op VMA splits (trying to split at a boundary) because it's probably
|
||||
// a bug. This restriction might be removed later.
|
||||
ASSERT(offset_in_vma < old_vma.size);
|
||||
ASSERT(offset_in_vma > 0);
|
||||
|
||||
old_vma.size = offset_in_vma;
|
||||
new_vma.base += offset_in_vma;
|
||||
new_vma.size -= offset_in_vma;
|
||||
|
||||
switch (new_vma.type) {
|
||||
case VirtualMemoryArea::Type::Unmapped:
|
||||
break;
|
||||
case VirtualMemoryArea::Type::Allocated:
|
||||
new_vma.offset += offset_in_vma;
|
||||
break;
|
||||
case VirtualMemoryArea::Type::Mapped:
|
||||
new_vma.backing_memory += offset_in_vma;
|
||||
break;
|
||||
}
|
||||
|
||||
ASSERT(old_vma.CanBeMergedWith(new_vma));
|
||||
|
||||
return vma_map.emplace_hint(std::next(vma_handle), new_vma.base, new_vma);
|
||||
}
|
||||
|
||||
MemoryManager::VMAIter MemoryManager::MergeAdjacent(VMAIter iter) {
|
||||
const VMAIter next_vma{std::next(iter)};
|
||||
if (next_vma != vma_map.end() && iter->second.CanBeMergedWith(next_vma->second)) {
|
||||
iter->second.size += next_vma->second.size;
|
||||
vma_map.erase(next_vma);
|
||||
}
|
||||
|
||||
if (iter != vma_map.begin()) {
|
||||
VMAIter prev_vma{std::prev(iter)};
|
||||
if (prev_vma->second.CanBeMergedWith(iter->second)) {
|
||||
prev_vma->second.size += iter->second.size;
|
||||
vma_map.erase(iter);
|
||||
iter = prev_vma;
|
||||
}
|
||||
}
|
||||
|
||||
return iter;
|
||||
}
|
||||
|
||||
void MemoryManager::UpdatePageTableForVMA(const VirtualMemoryArea& vma) {
|
||||
switch (vma.type) {
|
||||
case VirtualMemoryArea::Type::Unmapped:
|
||||
UnmapRegion(vma.base, vma.size);
|
||||
break;
|
||||
case VirtualMemoryArea::Type::Allocated:
|
||||
MapMemoryRegion(vma.base, vma.size, nullptr, vma.backing_addr);
|
||||
break;
|
||||
case VirtualMemoryArea::Type::Mapped:
|
||||
MapMemoryRegion(vma.base, vma.size, vma.backing_memory, vma.backing_addr);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace Tegra
|
||||
|
|
|
|||
|
|
@ -6,9 +6,9 @@
|
|||
|
||||
#include <map>
|
||||
#include <optional>
|
||||
#include <vector>
|
||||
|
||||
#include "common/common_types.h"
|
||||
#include "common/page_table.h"
|
||||
|
||||
namespace VideoCore {
|
||||
class RasterizerInterface;
|
||||
|
|
@ -20,45 +20,57 @@ class System;
|
|||
|
||||
namespace Tegra {
|
||||
|
||||
/**
|
||||
* Represents a VMA in an address space. A VMA is a contiguous region of virtual addressing space
|
||||
* with homogeneous attributes across its extents. In this particular implementation each VMA is
|
||||
* also backed by a single host memory allocation.
|
||||
*/
|
||||
struct VirtualMemoryArea {
|
||||
enum class Type : u8 {
|
||||
Unmapped,
|
||||
Allocated,
|
||||
Mapped,
|
||||
class PageEntry final {
|
||||
public:
|
||||
enum class State : u32 {
|
||||
Unmapped = static_cast<u32>(-1),
|
||||
Allocated = static_cast<u32>(-2),
|
||||
};
|
||||
|
||||
/// Virtual base address of the region.
|
||||
GPUVAddr base{};
|
||||
/// Size of the region.
|
||||
u64 size{};
|
||||
/// Memory area mapping type.
|
||||
Type type{Type::Unmapped};
|
||||
/// CPU memory mapped address corresponding to this memory area.
|
||||
VAddr backing_addr{};
|
||||
/// Offset into the backing_memory the mapping starts from.
|
||||
std::size_t offset{};
|
||||
/// Pointer backing this VMA.
|
||||
u8* backing_memory{};
|
||||
constexpr PageEntry() = default;
|
||||
constexpr PageEntry(State state) : state{state} {}
|
||||
constexpr PageEntry(VAddr addr) : state{static_cast<State>(addr >> ShiftBits)} {}
|
||||
|
||||
/// Tests if this area can be merged to the right with `next`.
|
||||
bool CanBeMergedWith(const VirtualMemoryArea& next) const;
|
||||
constexpr bool IsUnmapped() const {
|
||||
return state == State::Unmapped;
|
||||
}
|
||||
|
||||
constexpr bool IsAllocated() const {
|
||||
return state == State::Allocated;
|
||||
}
|
||||
|
||||
constexpr bool IsValid() const {
|
||||
return !IsUnmapped() && !IsAllocated();
|
||||
}
|
||||
|
||||
constexpr VAddr ToAddress() const {
|
||||
if (!IsValid()) {
|
||||
return {};
|
||||
}
|
||||
|
||||
return static_cast<VAddr>(state) << ShiftBits;
|
||||
}
|
||||
|
||||
constexpr PageEntry operator+(u64 offset) {
|
||||
// If this is a reserved value, offsets do not apply
|
||||
if (!IsValid()) {
|
||||
return *this;
|
||||
}
|
||||
return PageEntry{(static_cast<VAddr>(state) << ShiftBits) + offset};
|
||||
}
|
||||
|
||||
private:
|
||||
static constexpr std::size_t ShiftBits{12};
|
||||
|
||||
State state{State::Unmapped};
|
||||
};
|
||||
static_assert(sizeof(PageEntry) == 4, "PageEntry is too large");
|
||||
|
||||
class MemoryManager final {
|
||||
public:
|
||||
explicit MemoryManager(Core::System& system, VideoCore::RasterizerInterface& rasterizer);
|
||||
~MemoryManager();
|
||||
|
||||
GPUVAddr AllocateSpace(u64 size, u64 align);
|
||||
GPUVAddr AllocateSpace(GPUVAddr addr, u64 size, u64 align);
|
||||
GPUVAddr MapBufferEx(VAddr cpu_addr, u64 size);
|
||||
GPUVAddr MapBufferEx(VAddr cpu_addr, GPUVAddr addr, u64 size);
|
||||
GPUVAddr UnmapBuffer(GPUVAddr addr, u64 size);
|
||||
std::optional<VAddr> GpuToCpuAddress(GPUVAddr addr) const;
|
||||
|
||||
template <typename T>
|
||||
|
|
@ -70,9 +82,6 @@ public:
|
|||
u8* GetPointer(GPUVAddr addr);
|
||||
const u8* GetPointer(GPUVAddr addr) const;
|
||||
|
||||
/// Returns true if the block is continuous in host memory, false otherwise
|
||||
bool IsBlockContinuous(GPUVAddr start, std::size_t size) const;
|
||||
|
||||
/**
|
||||
* ReadBlock and WriteBlock are full read and write operations over virtual
|
||||
* GPU Memory. It's important to use these when GPU memory may not be continuous
|
||||
|
|
@ -98,92 +107,43 @@ public:
|
|||
void CopyBlockUnsafe(GPUVAddr gpu_dest_addr, GPUVAddr gpu_src_addr, std::size_t size);
|
||||
|
||||
/**
|
||||
* IsGranularRange checks if a gpu region can be simply read with a pointer
|
||||
* IsGranularRange checks if a gpu region can be simply read with a pointer.
|
||||
*/
|
||||
bool IsGranularRange(GPUVAddr gpu_addr, std::size_t size);
|
||||
|
||||
private:
|
||||
using VMAMap = std::map<GPUVAddr, VirtualMemoryArea>;
|
||||
using VMAHandle = VMAMap::const_iterator;
|
||||
using VMAIter = VMAMap::iterator;
|
||||
|
||||
bool IsAddressValid(GPUVAddr addr) const;
|
||||
void MapPages(GPUVAddr base, u64 size, u8* memory, Common::PageType type,
|
||||
VAddr backing_addr = 0);
|
||||
void MapMemoryRegion(GPUVAddr base, u64 size, u8* target, VAddr backing_addr);
|
||||
void UnmapRegion(GPUVAddr base, u64 size);
|
||||
|
||||
/// Finds the VMA in which the given address is included in, or `vma_map.end()`.
|
||||
VMAHandle FindVMA(GPUVAddr target) const;
|
||||
|
||||
VMAHandle AllocateMemory(GPUVAddr target, std::size_t offset, u64 size);
|
||||
|
||||
/**
|
||||
* Maps an unmanaged host memory pointer at a given address.
|
||||
*
|
||||
* @param target The guest address to start the mapping at.
|
||||
* @param memory The memory to be mapped.
|
||||
* @param size Size of the mapping in bytes.
|
||||
* @param backing_addr The base address of the range to back this mapping.
|
||||
*/
|
||||
VMAHandle MapBackingMemory(GPUVAddr target, u8* memory, u64 size, VAddr backing_addr);
|
||||
|
||||
/// Unmaps a range of addresses, splitting VMAs as necessary.
|
||||
void UnmapRange(GPUVAddr target, u64 size);
|
||||
|
||||
/// Converts a VMAHandle to a mutable VMAIter.
|
||||
VMAIter StripIterConstness(const VMAHandle& iter);
|
||||
|
||||
/// Marks as the specified VMA as allocated.
|
||||
VMAIter Allocate(VMAIter vma);
|
||||
|
||||
/**
|
||||
* Carves a VMA of a specific size at the specified address by splitting Free VMAs while doing
|
||||
* the appropriate error checking.
|
||||
*/
|
||||
VMAIter CarveVMA(GPUVAddr base, u64 size);
|
||||
|
||||
/**
|
||||
* Splits the edges of the given range of non-Free VMAs so that there is a VMA split at each
|
||||
* end of the range.
|
||||
*/
|
||||
VMAIter CarveVMARange(GPUVAddr base, u64 size);
|
||||
|
||||
/**
|
||||
* Splits a VMA in two, at the specified offset.
|
||||
* @returns the right side of the split, with the original iterator becoming the left side.
|
||||
*/
|
||||
VMAIter SplitVMA(VMAIter vma, u64 offset_in_vma);
|
||||
|
||||
/**
|
||||
* Checks for and merges the specified VMA with adjacent ones if possible.
|
||||
* @returns the merged VMA or the original if no merging was possible.
|
||||
*/
|
||||
VMAIter MergeAdjacent(VMAIter vma);
|
||||
|
||||
/// Updates the pages corresponding to this VMA so they match the VMA's attributes.
|
||||
void UpdatePageTableForVMA(const VirtualMemoryArea& vma);
|
||||
|
||||
/// Finds a free (unmapped region) of the specified size starting at the specified address.
|
||||
GPUVAddr FindFreeRegion(GPUVAddr region_start, u64 size) const;
|
||||
GPUVAddr Map(VAddr cpu_addr, GPUVAddr gpu_addr, std::size_t size);
|
||||
GPUVAddr MapAllocate(VAddr cpu_addr, std::size_t size, std::size_t align);
|
||||
std::optional<GPUVAddr> AllocateFixed(GPUVAddr gpu_addr, std::size_t size);
|
||||
GPUVAddr Allocate(std::size_t size, std::size_t align);
|
||||
void Unmap(GPUVAddr gpu_addr, std::size_t size);
|
||||
|
||||
private:
|
||||
PageEntry GetPageEntry(GPUVAddr gpu_addr) const;
|
||||
void SetPageEntry(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size = page_size);
|
||||
GPUVAddr UpdateRange(GPUVAddr gpu_addr, PageEntry page_entry, std::size_t size);
|
||||
std::optional<GPUVAddr> FindFreeRange(std::size_t size, std::size_t align) const;
|
||||
|
||||
void TryLockPage(PageEntry page_entry, std::size_t size);
|
||||
void TryUnlockPage(PageEntry page_entry, std::size_t size);
|
||||
|
||||
static constexpr std::size_t PageEntryIndex(GPUVAddr gpu_addr) {
|
||||
return (gpu_addr >> page_bits) & page_table_mask;
|
||||
}
|
||||
|
||||
static constexpr u64 address_space_size = 1ULL << 40;
|
||||
static constexpr u64 address_space_start = 1ULL << 32;
|
||||
static constexpr u64 page_bits{16};
|
||||
static constexpr u64 page_size{1 << page_bits};
|
||||
static constexpr u64 page_mask{page_size - 1};
|
||||
|
||||
/// Address space in bits, according to Tegra X1 TRM
|
||||
static constexpr u32 address_space_width{40};
|
||||
/// Start address for mapping, this is fairly arbitrary but must be non-zero.
|
||||
static constexpr GPUVAddr address_space_base{0x100000};
|
||||
/// End of address space, based on address space in bits.
|
||||
static constexpr GPUVAddr address_space_end{1ULL << address_space_width};
|
||||
|
||||
Common::PageTable page_table;
|
||||
VMAMap vma_map;
|
||||
VideoCore::RasterizerInterface& rasterizer;
|
||||
static constexpr u64 page_table_bits{24};
|
||||
static constexpr u64 page_table_size{1 << page_table_bits};
|
||||
static constexpr u64 page_table_mask{page_table_size - 1};
|
||||
|
||||
Core::System& system;
|
||||
|
||||
VideoCore::RasterizerInterface& rasterizer;
|
||||
|
||||
std::vector<PageEntry> page_table;
|
||||
};
|
||||
|
||||
} // namespace Tegra
|
||||
|
|
|
|||
Loading…
Reference in a new issue