Merge pull request #6097 from XD3V3LOP3RX/fix/host-import-probe-use-stamp

fix(graphics): keep discarded import probes from pinning windows

core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp

@@ -1400,7 +1400,10 @@ const HostImport* importAllocation(const Context& context, HostImports& state, s
     if (fresh && end != 0) {
         if (const auto* current = findImport(state, address, end)) return current;
     } else if (!fresh) {
-        if (const auto* current = findImport(state, address, address + 1); current != nullptr && current->joined && sameRange(*current, lease)) return current;
+        if (const auto* current = peekImport(state, address, address + 1); current != nullptr && current->joined && sameRange(*current, lease)) {
+            current->lastUse = ++state.useCounter;
+            return current;
+        }
     }
     auto run = importRun(lease, address);
     // Without an end the caller views the whole run through the import (AcquireRegistered): a window that holds the
@@ -1431,7 +1434,10 @@ const HostImport* importRunAt(const Context& context, HostImports& state, Import
         failures.insert_or_assign(base, ImportFailure{base, bytes, GuestAllocations::Mapped(run.ranges.begin(), run.ranges.end())});
         return nullptr;
     }
-    if (const auto found = state.imports.find(base); found != state.imports.end() && sameRun(found->second, run)) return &found->second;
+    if (const auto found = state.imports.find(base); found != state.imports.end() && sameRun(found->second, run)) {
+        found->second.lastUse = ++state.useCounter;
+        return &found->second;
+    }
     if (alignment == 0 || base % alignment != 0 || bytes % alignment != 0) return nullptr;
     if (const auto failed = failures.find(base); failed != failures.end()) {
         if (sameRun(failed->second, run)) return nullptr;

core/libs/prx/libSceAgcDriver/tests/BdaResources.cpp

@@ -22,6 +22,7 @@
 #include <iostream>
 #include <limits>
 #include <memory>
+#include <mutex>
 #include <new>
 #include <string>
 #include <vector>
@@ -739,6 +740,151 @@ void importWholeRangeOverWindowTests(const Context& context) {
     leased.WriteBack();
 }
 
+struct ObservedHostImport {
+    const void* pointer;
+    VkDeviceSize bytes;
+};
+
+class ImportAllocationProbe {
+public:
+    explicit ImportAllocationProbe(Context& context) : context(context), original(context.deviceProc) {
+        Require(active == nullptr && original != nullptr, "invalid import allocation probe scope");
+        active = this;
+        context.deviceProc = Resolve;
+    }
+    ~ImportAllocationProbe() {
+        context.deviceProc = original;
+        active = nullptr;
+    }
+    ImportAllocationProbe(const ImportAllocationProbe&) = delete;
+    ImportAllocationProbe& operator=(const ImportAllocationProbe&) = delete;
+
+    std::vector<ObservedHostImport> calls;
+
+private:
+    static PFN_vkVoidFunction VKAPI_CALL Resolve(VkDevice device, const char* name) {
+        Require(active != nullptr, "import allocation probe escaped its scope");
+        if (std::strcmp(name, "vkAllocateMemory") == 0) return reinterpret_cast<PFN_vkVoidFunction>(Allocate);
+        return active->original(device, name);
+    }
+    static VkResult VKAPI_CALL Allocate(VkDevice device, const VkMemoryAllocateInfo* info, const VkAllocationCallbacks* allocator, VkDeviceMemory* memory) {
+        Require(active != nullptr, "import allocation callback escaped its scope");
+        for (const auto* next = static_cast<const VkBaseInStructure*>(info->pNext); next != nullptr; next = next->pNext) {
+            if (next->sType != VK_STRUCTURE_TYPE_IMPORT_MEMORY_HOST_POINTER_INFO_EXT) continue;
+            const auto* imported = reinterpret_cast<const VkImportMemoryHostPointerInfoEXT*>(next);
+            active->calls.push_back({imported->pHostPointer, info->allocationSize});
+        }
+        const auto allocate = reinterpret_cast<PFN_vkAllocateMemory>(active->original(device, "vkAllocateMemory"));
+        Require(allocate != nullptr, "import allocation probe has no original allocator");
+        return allocate(device, info, allocator, memory);
+    }
+
+    Context& context;
+    PFN_vkGetDeviceProcAddr original;
+    inline static thread_local ImportAllocationProbe* active = nullptr;
+};
+
+class UnrelatedImportProbeRange {
+public:
+    UnrelatedImportProbeRange() {
+        GuestAllocations::Mutation().Add(storage.data(), storage.size(), true, false, false);
+    }
+    ~UnrelatedImportProbeRange() {
+        GuestAllocations::Mutation().Remove(storage.data());
+    }
+    UnrelatedImportProbeRange(const UnrelatedImportProbeRange&) = delete;
+    UnrelatedImportProbeRange& operator=(const UnrelatedImportProbeRange&) = delete;
+
+private:
+    std::array<std::byte, 64> storage{};
+};
+
+void importStaleProbeTests(const Context& context, const BdaTestAccess& access) {
+    for (const auto scenario : {0u, 1u, 2u}) {
+        const bool held = scenario == 1u;
+        const bool exactWhole = scenario == 2u;
+        WindowedRange range(context, 2 * WindowedRange::piece, 2);
+        ImportAllocationProbe probe(range.importing);
+        const auto first = range.First();
+        auto* guest = reinterpret_cast<std::byte*>(first);
+        for (std::size_t offset = 17u; offset < range.bytes; offset += 4096u) guest[offset] = static_cast<std::byte>((offset / 4096u) & 0xffu);
+        const std::vector<std::byte> originalGuest(guest, guest + range.bytes);
+        const auto* window = HostImportFor(range.importing, first, exactWhole ? range.bytes : 32u);
+        Require(window != nullptr && !window->joined && window->base == first && window->bytes == (exactWhole ? range.bytes : range.window), "the stale-probe fixture did not create its ordinary import extent");
+        const auto originalBuffer = window->buffer;
+        const auto originalUse = window->lastUse;
+        const auto originalSerial = HostImportSerial(range.importing, first, 32u, false);
+        auto mockContents = access.bytes(window->buffer);
+        Require(mockContents.size() == window->bytes, "the mock import buffer has an incorrect extent");
+        std::copy_n(originalGuest.begin(), window->bytes, mockContents.begin());
+        std::uint32_t scratch = 0x12345678u;
+        {
+            GuestBufferMemory copied(range.importing);
+            copied.AddReadable(reinterpret_cast<std::uintptr_t>(&scratch), sizeof(scratch));
+            copied.Upload(false);
+            copied.WriteBack();
+        }
+        Require(window->lastUse == originalUse && HostImportSerial(range.importing, first, 32u, false) == originalSerial, "an unrelated copied build touched or replaced the prefix window");
+        GuestBufferMemory reader(range.importing);
+        VkDescriptorBufferInfo heldView{};
+        std::uint32_t heldAdjustment = 0u;
+        if (held) {
+            reader.AddReadable(first + 16u, 32u);
+            reader.Upload(false);
+            heldView = reader.Descriptor(first + 16u, 32u, heldAdjustment);
+            Require(heldView.buffer == originalBuffer, "the held control did not bind the prefix window");
+        }
+        GuestAllocations::Mapped mapped;
+        {
+            const auto lease = GuestAllocations::GuestAllocationsAcquire_nid_postfix();
+            for (const auto& owner : lease) if (owner->address == first && owner->bytes == range.bytes) mapped.push_back(owner);
+        }
+        Require(mapped.size() == 1u, "the stale-probe fixture lost its registered range");
+        const auto originalOwner = mapped.front().lock();
+        const auto generation = GuestAllocations::GuestAllocationsGeneration_nid_postfix();
+        UnrelatedImportProbeRange unrelated;
+        Require(GuestAllocations::GuestAllocationsGeneration_nid_postfix() != generation && HostImportsIdentity(range.importing).device == VK_NULL_HANDLE,
+            "the unrelated registry change did not make imports stale");
+        const auto attempts = access.allocationAttempts();
+        probe.calls.clear();
+        {
+            std::lock_guard gpu(AgcDriver::GuestMemory::GpuMutex());
+            Require(ImportMappedRanges(range.importing, mapped, GuestAllocations::GuestAllocationsGeneration_nid_postfix(), false), "the current device did not admit the map-time import pass");
+        }
+        Require(mapped.front().lock() == originalOwner, "an unrelated registry change replaced the import's memory owner");
+        if (held) {
+            Require(access.allocationAttempts() == attempts && probe.calls.empty(), "a stale probe retired an import held by an actual descriptor");
+            const auto* kept = HostImportFor(range.importing, first + 16u, 32u);
+            Require(kept != nullptr && kept->buffer == originalBuffer && kept->bytes == range.window && HostImportSerial(range.importing, first, 32u, false) == originalSerial,
+                "the actual held import lost its buffer, extent or serial");
+            std::uint32_t adjustment = 0u;
+            const auto view = reader.Descriptor(first + 16u, 32u, adjustment);
+            Require(view.buffer == heldView.buffer && view.offset == heldView.offset && view.range == heldView.range && adjustment == heldAdjustment,
+                "a map-time stale probe invalidated a descriptor that still uses its import");
+            const auto contents = access.bytes(view.buffer);
+            Require(view.offset + adjustment + 32u <= contents.size() && std::memcmp(contents.data() + view.offset + adjustment, reinterpret_cast<const void*>(first + 16u), 32u) == 0,
+                "the protected descriptor no longer reads the registered bytes");
+            reader.WriteBack();
+        } else if (exactWhole) {
+            Require(access.allocationAttempts() == attempts && probe.calls.empty() && window->lastUse > originalUse,
+                "accepting the stale exact whole import reallocated it or did not protect its returned pointer");
+            const auto* kept = HostImportFor(range.importing, first, range.bytes);
+            Require(kept != nullptr && kept->buffer == originalBuffer && kept->ranges.size() == 1u && kept->ranges.front() == originalOwner && HostImportSerial(range.importing, first, range.bytes, false) == originalSerial,
+                "accepting the exact whole import changed its native or registered identity");
+        } else {
+            Require(access.allocationAttempts() == attempts + 1u, "a discarded stale probe made an unheld prefix window block the whole import");
+            const auto* whole = HostImportFor(range.importing, first, range.bytes);
+            Require(whole != nullptr && whole->base == first && whole->bytes == range.bytes && whole->ranges.size() == 1u && whole->ranges.front() == originalOwner,
+                "the stale-probe repair did not retain the exact registered owner over the whole import");
+            Require(HostImportSerial(range.importing, first, range.bytes, false) != originalSerial, "replacing the unheld window reused its old import identity");
+            Require(probe.calls.size() == 1u && probe.calls.front().pointer == reinterpret_cast<const void*>(first) && probe.calls.front().bytes == range.bytes,
+                "the replacement native import did not reference the exact whole registered memory");
+            Require(access.bytes(whole->buffer).size() == range.bytes, "the replacement import buffer does not cover its registered extent");
+        }
+        Require(scratch == 0x12345678u && std::equal(originalGuest.begin(), originalGuest.end(), guest), "the import probe changed guest bytes or unrelated copied data");
+    }
+}
+
 // An address-based build whose whole-range import does not fit the budget leaves the range's window alone: before,
 // every such build retired the window and the next access imported it again.
 void importRefusedWholeRangeTests(const Context& context) {
@@ -2068,6 +2214,7 @@ void RunBdaResourceTests(const Context& context, const BdaTestAccess& access) {
     importCoverageHintBudgetTests(context, access);
     failedImportJoinProbeBudgetTests(context, access);
     importWholeRangeOverWindowTests(context);
+    importStaleProbeTests(context, access);
     importRefusedWholeRangeTests(context);
     importNeighbouringEdgesTests(context);
     importAtMapOverBudgetTests(context);