perf(libkernel): commit Windows direct backing when mapped

core/libs/prx/libc/include/WindowsMappings.hpp

@@ -82,6 +82,9 @@ public:
             split(cursor + done, pageBytes);
             void* page = reinterpret_cast<void*>(cursor + done);
             if (!map(section, GetCurrentProcess(), page, offset + done, pageBytes, MEM_REPLACE_PLACEHOLDER, PAGE_EXECUTE_READWRITE, nullptr, 0)) fail("map shared guest page");
+            MEMORY_BASIC_INFORMATION memory{};
+            if (VirtualQuery(page, &memory, sizeof(memory)) != sizeof(memory)) fail("query shared guest page");
+            if (memory.State == MEM_RESERVE && !VirtualAlloc(page, pageBytes, MEM_COMMIT, PAGE_EXECUTE_READWRITE)) fail("commit shared guest page");
             DWORD previous;
             if (!VirtualProtect(page, pageBytes, protection, &previous)) fail("protect shared guest page");
             const auto key = std::make_pair(reinterpret_cast<std::uintptr_t>(section), offset + done);

core/libs/prx/libkernel/DirectMemory/DirectMemory.cpp

@@ -270,7 +270,7 @@ public:
 #ifdef _WIN32
         static_cast<void>(start);
         const auto size = static_cast<std::uint64_t>(bytes);
-        section = CreateFileMappingW(INVALID_HANDLE_VALUE, nullptr, PAGE_EXECUTE_READWRITE, static_cast<DWORD>(size >> 32), static_cast<DWORD>(size), nullptr);
+        section = CreateFileMappingW(INVALID_HANDLE_VALUE, nullptr, PAGE_EXECUTE_READWRITE | SEC_RESERVE, static_cast<DWORD>(size >> 32), static_cast<DWORD>(size), nullptr);
         if (!section) {
             const auto error = static_cast<int>(GetLastError());
             char message[96];

core/libs/tests/GuestMemory.cpp

@@ -289,6 +289,34 @@ static void CheckDirectMemoryFollowsPhysicalPages() {
     Require(sceKernelReleaseDirectMemory(again, page * 2) == 0);
 }
 
+static void CheckSparseDirectMemoryBacking() {
+    constexpr std::size_t page = 0x4000;
+    constexpr std::size_t size = 256u * 1024u * 1024u;
+    std::int64_t physical = 0;
+    Require(sceKernelAllocateDirectMemory(0, 0x7fffffffffll, size, 0, 0, &physical) == 0);
+    void* first = nullptr;
+    void* last = nullptr;
+    Require(sceKernelMapDirectMemory(&first, page, 3, 0, physical, 0) == 0);
+    Require(sceKernelMapDirectMemory(&last, page, 3, 0, physical + size - page, 0) == 0);
+    auto* firstBytes = static_cast<unsigned char*>(first);
+    auto* lastBytes = static_cast<unsigned char*>(last);
+    Require(firstBytes[0] == 0 && firstBytes[page - 1] == 0 && lastBytes[0] == 0 && lastBytes[page - 1] == 0);
+    firstBytes[0] = 17;
+    lastBytes[page - 1] = 29;
+    void* alias = nullptr;
+    Require(sceKernelMapDirectMemory(&alias, page, 3, 0, physical + size - page, 0) == 0);
+    auto* aliasBytes = static_cast<unsigned char*>(alias);
+    Require(aliasBytes[page - 1] == 29 && aliasBytes[0] == 0);
+    aliasBytes[page - 1] = 41;
+    Require(lastBytes[page - 1] == 41 && firstBytes[0] == 17);
+    Require(sceKernelMunmap(last, page) == 0 && sceKernelMunmap(alias, page) == 0);
+    last = nullptr;
+    Require(sceKernelMapDirectMemory(&last, page, 3, 0, physical + size - page, 0) == 0);
+    Require(static_cast<unsigned char*>(last)[page - 1] == 41);
+    Require(sceKernelMunmap(first, page) == 0 && sceKernelMunmap(last, page) == 0);
+    Require(sceKernelReleaseDirectMemory(physical, size) == 0);
+}
+
 static void CheckReleaseDirectMemoryClearsMappings() {
     constexpr std::size_t page = 0x4000;
     std::int64_t phys = 0;
@@ -1715,6 +1743,7 @@ int main() {
     CheckCheckedReleaseDirectMemory();
     CheckAudioCoprocessorProtection();
     CheckDirectMemoryFollowsPhysicalPages();
+    CheckSparseDirectMemoryBacking();
     CheckReleaseDirectMemoryClearsMappings();
     CheckReleaseDirectMemoryRejectsInvalidRanges();
     CheckFixedMappingReplacesPartialOverlap();

docs/dev/TechnicalDebt.md

@@ -9,6 +9,10 @@
   - libwinpthread-1.dll
 - winpthreads initializes a statically initialized rwlock on its first lock and fails a concurrent first lock with `EINVAL`, which `std::shared_mutex` and `std::shared_timed_mutex` ignore, so that lock does not exclude. Guest rwlocks ([Pthread.hpp](../../core/libs/prx/libkernel/Pthread/include/Pthread.hpp)) are locked once when created. The host `std::shared_mutex` objects in [Recompiler.cpp](../../core/shader/recompiler/Recompiler.cpp) and [GuestWriteWatch.cpp](../../core/libs/prx/libc/src/GuestWriteWatch.cpp) are not, and can still race on their first contended use.
 
+### Direct memory
+
+- Windows direct-memory backing reserves a paging-file section and commits each mapped guest page before setting its protection. Committed physical contents remain in the section across unmap/remap and are released when the backing and its remaining views are destroyed; commitment is not recovered merely by unmapping a page. Linux retains its existing backing. Peak gameplay memory and Windows versions other than the tested host remain unverified.
+
 ### Silent stubs
 
 Throughout the project, every function at every stage either **does exactly what it's supposed to or throws an exception**. Everywhere... except: