gpu: address extended guest memory without widening dense cache tables

This commit is contained in:
nmzik
2026-09-28 16:32:32 +04:00
parent c91fe80d64
commit 2016f8e2fe
12 changed files with 81 additions and 34 deletions
+2 -2
View File
@@ -9,9 +9,9 @@
namespace {
using Owners = std::vector<uint32_t>;
using Table = Libs::Graphics::MultiLevelPageTable<Owners>;
using Table = Libs::Graphics::MultiLevelPageTable<Owners, 20, 44, 14>;
using PageOwners = Libs::Graphics::InlinePageOwnerList<uint32_t, 16>;
using OwnerTable = Libs::Graphics::MultiLevelPageTable<PageOwners, 20, 40, 10>;
using OwnerTable = Libs::Graphics::MultiLevelPageTable<PageOwners, 20, 44, 14>;
void Check(bool value, const char* text) {
if (!value) {
+9 -2
View File
@@ -219,17 +219,24 @@ void TestRangeSet() {
void TestGuestRange() {
constexpr GuestRange empty{};
constexpr GuestRange first_byte{1, 1};
constexpr GuestRange last_byte{TRACKER_ADDRESS_SIZE - 1, 1};
constexpr uint64_t extended_end = Libs::LibKernel::Memory::kExtendedMemoryBase +
Libs::LibKernel::Memory::kExtendedMemorySize;
constexpr GuestRange last_byte{extended_end - 1, 1};
static_assert(empty.Empty() && !empty.Valid() && empty.ValidOrEmpty());
static_assert(!first_byte.Empty() && first_byte.Valid() &&
first_byte.ValidOrEmpty() && first_byte.End() == 2);
static_assert(last_byte.Valid() && last_byte.End() == TRACKER_ADDRESS_SIZE);
static_assert(last_byte.Valid() && last_byte.End() == extended_end);
Check(!GuestRange{0, 1}.Empty() && !GuestRange{0, 1}.ValidOrEmpty(),
"zero-address nonempty guest range is rejected");
Check(!GuestRange{1, 0}.Empty() && !GuestRange{1, 0}.ValidOrEmpty(),
"nonzero-address empty guest range is rejected");
Check(GuestRange{Libs::LibKernel::Memory::kExtendedMemoryBase, 1}.Valid() &&
!GuestRange{Libs::Graphics::LOWER_ADDRESS_SIZE, 1}.Valid() &&
!GuestRange{Libs::LibKernel::Memory::kExtendedMemoryBase - 1, 2}.Valid() &&
!GuestRange{extended_end - 1, 2}.Valid(),
"extended range and gap boundaries are enforced");
Check(!GuestRange{TRACKER_ADDRESS_SIZE, 1}.Valid(),
"first address beyond the guest range is rejected");
Check(!GuestRange{TRACKER_ADDRESS_SIZE - 1, 2}.Valid(),
+5 -4
View File
@@ -136,8 +136,8 @@ bool ProtectAddressSpace(uint64_t vaddr, uint64_t size,
&old_protection) != 0;
}
uint8_t *Allocate(uint64_t size, uint32_t protection = PAGE_READWRITE) {
constexpr uintptr_t test_address = 0x0000000200010000ull;
uint8_t *Allocate(uint64_t size, uint32_t protection = PAGE_READWRITE,
uintptr_t test_address = 0x0000000200010000ull) {
#if KYTY_PLATFORM == KYTY_PLATFORM_WINDOWS
auto *memory = static_cast<uint8_t *>(
VirtualAlloc(reinterpret_cast<void *>(test_address), size,
@@ -155,11 +155,11 @@ uint8_t *Allocate(uint64_t size, uint32_t protection = PAGE_READWRITE) {
return memory;
}
void TestWatchAndUnwatch() {
void TestWatchAndUnwatch(uint64_t base = 0x0000000200010000ull) {
g_protection_calls = 0;
PageManager manager;
const auto page_size = manager.GetPageSize();
auto *memory = Allocate(page_size * 2);
auto *memory = Allocate(page_size * 2, PAGE_READWRITE, base);
const auto address = reinterpret_cast<uint64_t>(memory);
manager.UpdatePageWatchers<true>(address, page_size);
@@ -580,6 +580,7 @@ int main(int argc, char **argv) {
RunDeathCase(argv[2]);
}
TestWatchAndUnwatch();
TestWatchAndUnwatch(Libs::LibKernel::Memory::kExtendedMemoryBase);
TestSharedWatcherCounts();
TestCrossRegionRange();
TestBatchedWatcherRanges();
+15 -5
View File
@@ -208,8 +208,8 @@ struct TextureCacheTestAccess {
static_assert(std::same_as<decltype(TextureCache::m_slot_images),
Common::SlotVector<Image>>);
static_assert(TextureCache::ImagePageTable::kPageBits == 20);
static_assert(TextureCache::ImagePageTable::kAddressSpaceBits == 40);
static_assert(TextureCache::ImagePageTable::kFirstLevelBits == 10);
static_assert(TextureCache::ImagePageTable::kAddressSpaceBits == 44);
static_assert(TextureCache::ImagePageTable::kFirstLevelBits == 14);
static std::unique_lock<TrackingSpinLock> Lock(TextureCache &cache) {
return std::unique_lock(cache.m_lock);
@@ -13694,7 +13694,7 @@ public:
(bytes + BufferCache::CACHING_PAGESIZE - 1) /
BufferCache::CACHING_PAGESIZE;
auto address = buffer.device_address + backing - page_offset;
auto page = guest_base >> BufferCache::CACHING_PAGEBITS;
auto page = BufferCache::PageIndex(guest_base);
for (uint64_t mapped = 0; mapped < pages; mapped++) {
cmd.updateBuffer(m_bda_pagetable_buffer.buffer,
(page + mapped) * sizeof(vk::DeviceAddress),
@@ -25957,6 +25957,11 @@ TestCase FlatVirtualAddressRebasesGuestAllocation() {
using O = ShaderOpcode;
constexpr uint64_t GuestBase = 0x0000000110000000ull;
constexpr uint64_t ExtendedBase = Libs::LibKernel::Memory::kExtendedMemoryBase + GuestBase;
static_assert(BufferCache::PageIndex(ExtendedBase) ==
((Libs::Graphics::LOWER_ADDRESS_SIZE + GuestBase) >> BufferCache::CACHING_PAGEBITS));
static_assert(BufferCache::GuestAddress(BufferCache::PageIndex(ExtendedBase) <<
BufferCache::CACHING_PAGEBITS) == ExtendedBase);
std::vector<u32> code;
AppendVMovLiteral(&code, 20, static_cast<u32>(GuestBase + 4u));
AppendVMovLiteral(&code, 21, static_cast<u32>(GuestBase >> 32u));
@@ -25967,17 +25972,22 @@ TestCase FlatVirtualAddressRebasesGuestAllocation() {
AppendVMovLiteral(&code, 21, static_cast<u32>((GuestBase >> 32u) + 1u));
code.push_back(EncodeFlat0(0x0c, 0, 0));
code.push_back(EncodeFlat1(2, 0x7d, 0, 20));
AppendVMovLiteral(&code, 20, static_cast<u32>(ExtendedBase + 4u));
AppendVMovLiteral(&code, 21, static_cast<u32>(ExtendedBase >> 32u));
code.push_back(EncodeFlat0(0x0c, 0, 0));
code.push_back(EncodeFlat1(3, 0x7d, 0, 20));
AppendStoreVgpr(&code, 0, 0);
AppendStoreVgpr(&code, 1, 1);
AppendStoreVgpr(&code, 2, 2);
AppendStoreVgpr(&code, 3, 3);
AppendEnd(&code);
TestCase test;
test.name = "FlatVirtualAddressRebasesGuestAllocation";
test.code = std::move(code);
test.initial = {0xfeedfaceu, 0xcafebabeu, 0, 0x12345678u};
test.expected = {0x12345678u, 0x12345678u, 0};
test.bda_mappings = {{GuestBase, 8}};
test.expected = {0x12345678u, 0x12345678u, 0, 0xcafebabeu};
test.bda_mappings = {{GuestBase, 8}, {ExtendedBase, 0}};
test.opcodes = {O::V_MOV_B32, O::FLAT_LOAD_DWORD, O::BUFFER_STORE_DWORD,
O::S_ENDPGM};
return test;