Merge branch 'main' into http-uri-sweeppath

src/core/libraries/fiber/fiber.cpp

@@ -6,8 +6,8 @@
 #include "common/elf_info.h"
 #include "common/logging/log.h"
 #include "core/libraries/fiber/fiber_error.h"
-#include "core/libraries/kernel/threads/pthread.h"
 #include "core/libraries/libs.h"
+#include "core/tls.h"
 
 namespace Libraries::Fiber {
 
@@ -20,7 +20,7 @@ static constexpr u64 kFiberStackSizeCheck = 0xdeadbeefdeadbeef;
 static std::atomic<u32> context_size_check = false;
 
 OrbisFiberContext* GetFiberContext() {
-    return Libraries::Kernel::g_curthread->tcb->tcb_fiber;
+    return Core::GetTcbBase()->tcb_fiber;
 }
 
 extern "C" s32 PS4_SYSV_ABI _sceFiberSetJmp(OrbisFiberContext* ctx) asm("_sceFiberSetJmp");
@@ -269,7 +269,7 @@ s32 PS4_SYSV_ABI sceFiberRunImpl(OrbisFiber* fiber, void* addr_context, u64 size
         return ORBIS_FIBER_ERROR_INVALID;
     }
 
-    Core::Tcb* tcb = Libraries::Kernel::g_curthread->tcb;
+    Core::Tcb* tcb = Core::GetTcbBase();
     if (tcb->tcb_fiber) {
         return ORBIS_FIBER_ERROR_PERMISSION;
     }

src/core/libraries/kernel/threads/pthread.cpp

@@ -663,6 +663,10 @@ void RegisterThread(Core::Loader::SymbolsResolver* sym) {
     LIB_FUNCTION("Z4QosVuAsA0", "libkernel", 1, "libkernel", posix_pthread_once);
     LIB_FUNCTION("EotR8a3ASf4", "libkernel", 1, "libkernel", posix_pthread_self);
     LIB_FUNCTION("OxhIB8LB-PQ", "libkernel", 1, "libkernel", posix_pthread_create);
+    LIB_FUNCTION("lZzFeSxPl08", "libkernel", 1, "libkernel", posix_pthread_setcancelstate);
+    LIB_FUNCTION("CBNtXOoef-E", "libkernel", 1, "libkernel", posix_sched_get_priority_max);
+    LIB_FUNCTION("m0iS6jNsXds", "libkernel", 1, "libkernel", posix_sched_get_priority_min);
+    LIB_FUNCTION("Xs9hdiD7sAA", "libkernel", 1, "libkernel", posix_pthread_setschedparam);
     LIB_FUNCTION("+U1R4WtXvoc", "libkernel", 1, "libkernel", posix_pthread_detach);
     LIB_FUNCTION("7Xl257M4VNI", "libkernel", 1, "libkernel", posix_pthread_equal);
     LIB_FUNCTION("h9CcP3J0oVM", "libkernel", 1, "libkernel", posix_pthread_join);

src/core/linker.cpp

@@ -368,7 +368,7 @@ bool Linker::Resolve(const std::string& name, Loader::SymbolType sym_type, Modul
 void* Linker::TlsGetAddr(u64 module_index, u64 offset) {
     std::scoped_lock lk{mutex};
 
-    DtvEntry* dtv_table = Libraries::Kernel::g_curthread->tcb->tcb_dtv;
+    DtvEntry* dtv_table = GetTcbBase()->tcb_dtv;
     if (dtv_table[0].counter != dtv_generation_counter) {
         // Generation counter changed, a dynamic module was either loaded or unloaded.
         const u32 old_num_dtvs = dtv_table[1].counter;
@@ -381,7 +381,7 @@ void* Linker::TlsGetAddr(u64 module_index, u64 offset) {
         delete[] dtv_table;
 
         // Update TCB pointer.
-        Libraries::Kernel::g_curthread->tcb->tcb_dtv = new_dtv_table;
+        GetTcbBase()->tcb_dtv = new_dtv_table;
         dtv_table = new_dtv_table;
     }
 

src/core/tls.cpp

@@ -46,6 +46,10 @@ void SetTcbBase(void* image_address) {
     ASSERT(result != 0);
 }
 
+Tcb* GetTcbBase() {
+    return reinterpret_cast<Tcb*>(TlsGetValue(GetTcbKey()));
+}
+
 #elif defined(__APPLE__) && defined(ARCH_X86_64)
 
 // Apple x86_64
@@ -145,6 +149,12 @@ void SetTcbBase(void* image_address) {
                "Failed to store thread LDT page pointer: {}", errno);
 }
 
+Tcb* GetTcbBase() {
+    Tcb* tcb;
+    asm volatile("mov %%fs:0x0, %0" : "=r"(tcb));
+    return tcb;
+}
+
 #elif defined(ARCH_X86_64)
 
 // Other POSIX x86_64
@@ -154,6 +164,10 @@ void SetTcbBase(void* image_address) {
     ASSERT_MSG(ret == 0, "Failed to set GS base: errno {}", errno);
 }
 
+Tcb* GetTcbBase() {
+    return Libraries::Kernel::g_curthread->tcb;
+}
+
 #else
 
 // POSIX non-x86_64
@@ -176,6 +190,10 @@ void SetTcbBase(void* image_address) {
     ASSERT(pthread_setspecific(GetTcbKey(), image_address) == 0);
 }
 
+Tcb* GetTcbBase() {
+    return static_cast<Tcb*>(pthread_getspecific(GetTcbKey()));
+}
+
 #endif
 
 thread_local std::once_flag init_tls_flag;

src/core/tls.h

@@ -39,6 +39,9 @@ u32 GetTcbKey();
 /// Sets the data pointer to the TCB block.
 void SetTcbBase(void* image_address);
 
+/// Retrieves Tcb structure for the calling thread.
+Tcb* GetTcbBase();
+
 /// Makes sure TLS is initialized for the thread before entering guest.
 void EnsureThreadInitialized();
 

src/video_core/renderer_vulkan/vk_scheduler.cpp

@@ -3,6 +3,7 @@
 
 #include "common/assert.h"
 #include "common/debug.h"
+#include "common/thread.h"
 #include "imgui/renderer/texture_manager.h"
 #include "video_core/renderer_vulkan/vk_instance.h"
 #include "video_core/renderer_vulkan/vk_scheduler.h"
@@ -17,6 +18,8 @@ Scheduler::Scheduler(const Instance& instance)
     profiler_scope = reinterpret_cast<tracy::VkCtxScope*>(std::malloc(sizeof(tracy::VkCtxScope)));
 #endif
     AllocateWorkerCommandBuffers();
+    priority_pending_ops_thread =
+        std::jthread(std::bind_front(&Scheduler::PriorityPendingOpsThread, this));
 }
 
 Scheduler::~Scheduler() {
@@ -167,6 +170,32 @@ void Scheduler::SubmitExecution(SubmitInfo& info) {
     PopPendingOperations();
 }
 
+void Scheduler::PriorityPendingOpsThread(std::stop_token stoken) {
+    Common::SetCurrentThreadName("shadPS4:GpuSchedPriorityPendingOpsRunner");
+
+    while (!stoken.stop_requested()) {
+        PendingOp op;
+        {
+            std::unique_lock lk(priority_pending_ops_mutex);
+            priority_pending_ops_cv.wait(lk, stoken,
+                                         [this] { return !priority_pending_ops.empty(); });
+            if (stoken.stop_requested()) {
+                break;
+            }
+
+            op = std::move(priority_pending_ops.front());
+            priority_pending_ops.pop();
+        }
+
+        master_semaphore.Wait(op.gpu_tick);
+        if (stoken.stop_requested()) {
+            break;
+        }
+
+        op.callback();
+    }
+}
+
 void DynamicState::Commit(const Instance& instance, const vk::CommandBuffer& cmdbuf) {
     if (dirty_state.viewports) {
         dirty_state.viewports = false;

src/video_core/renderer_vulkan/vk_scheduler.h

@@ -5,6 +5,7 @@
 
 #include <condition_variable>
 #include <mutex>
+#include <thread>
 #include <queue>
 
 #include "common/unique_function.h"
@@ -401,10 +402,21 @@ public:
     }
 
     /// Defers an operation until the gpu has reached the current cpu tick.
+    /// Will be run when submitting or calling PopPendingOperations.
     void DeferOperation(Common::UniqueFunction<void>&& func) {
         pending_ops.emplace(std::move(func), CurrentTick());
     }
 
+    /// Defers an operation until the gpu has reached the current cpu tick.
+    /// Runs as soon as possible in another thread.
+    void DeferPriorityOperation(Common::UniqueFunction<void>&& func) {
+        {
+            std::unique_lock lk(priority_pending_ops_mutex);
+            priority_pending_ops.emplace(std::move(func), CurrentTick());
+        }
+        priority_pending_ops_cv.notify_one();
+    }
+
     static std::mutex submit_mutex;
 
 private:
@@ -412,6 +424,8 @@ private:
 
     void SubmitExecution(SubmitInfo& info);
 
+    void PriorityPendingOpsThread(std::stop_token stoken);
+
 private:
     const Instance& instance;
     MasterSemaphore master_semaphore;
@@ -424,6 +438,10 @@ private:
         u64 gpu_tick;
     };
     std::queue<PendingOp> pending_ops;
+    std::queue<PendingOp> priority_pending_ops;
+    std::mutex priority_pending_ops_mutex;
+    std::condition_variable_any priority_pending_ops_cv;
+    std::jthread priority_pending_ops_thread;
     RenderState render_state;
     bool is_rendering = false;
     tracy::VkCtxScope* profiler_scope{};

src/video_core/texture_cache/texture_cache.cpp

@@ -52,9 +52,6 @@ TextureCache::TextureCache(const Vulkan::Instance& instance_, Vulkan::Scheduler&
         std::max<u64>(std::min(device_local_memory - min_vacancy_critical, min_spacing_critical),
                       DEFAULT_CRITICAL_GC_MEMORY));
     trigger_gc_memory = static_cast<u64>((device_local_memory - mem_threshold) / 2);
-
-    downloaded_images_thread =
-        std::jthread([&](const std::stop_token& token) { DownloadedImagesThread(token); });
 }
 
 TextureCache::~TextureCache() = default;
@@ -125,33 +122,11 @@ void TextureCache::DownloadImageMemory(ImageId image_id) {
     cmdbuf.copyImageToBuffer(image.GetImage(), vk::ImageLayout::eTransferSrcOptimal,
                              download_buffer.Handle(), image_download);
 
-    {
-        std::unique_lock lock(downloaded_images_mutex);
-        downloaded_images_queue.emplace(scheduler.CurrentTick(), image.info.guest_address, download,
-                                        download_size);
-        downloaded_images_cv.notify_one();
-    }
-}
-
-void TextureCache::DownloadedImagesThread(const std::stop_token& token) {
-    auto* memory = Core::Memory::Instance();
-    while (!token.stop_requested()) {
-        DownloadedImage image;
-        {
-            std::unique_lock lock{downloaded_images_mutex};
-            downloaded_images_cv.wait(lock, token,
-                                      [this] { return !downloaded_images_queue.empty(); });
-            if (token.stop_requested()) {
-                break;
-            }
-            image = downloaded_images_queue.front();
-            downloaded_images_queue.pop();
-        }
-
-        scheduler.GetMasterSemaphore()->Wait(image.tick);
-        memory->TryWriteBacking(std::bit_cast<u8*>(image.device_addr), image.download,
-                                image.download_size);
-    }
+    scheduler.DeferPriorityOperation(
+        [this, device_addr = image.info.guest_address, download, download_size] {
+            Core::Memory::Instance()->TryWriteBacking(std::bit_cast<u8*>(device_addr), download,
+                                                      download_size);
+        });
 }
 
 void TextureCache::MarkAsMaybeDirty(ImageId image_id, Image& image) {

src/video_core/texture_cache/texture_cache.h

@@ -314,16 +314,6 @@ private:
     Common::LeastRecentlyUsedCache<ImageId, u64> lru_cache;
     PageTable page_table;
     std::mutex mutex;
-    struct DownloadedImage {
-        u64 tick;
-        VAddr device_addr;
-        void* download;
-        size_t download_size;
-    };
-    std::queue<DownloadedImage> downloaded_images_queue;
-    std::mutex downloaded_images_mutex;
-    std::condition_variable_any downloaded_images_cv;
-    std::jthread downloaded_images_thread;
     struct MetaDataInfo {
         enum class Type {
             CMask,