From 5145133a604f626c05f832465ac22019b003c32a Mon Sep 17 00:00:00 2001
From: Markus Wick
Date: Wed, 7 Apr 2021 08:42:54 +0200
Subject: video_core/gpu_thread: Implement a ShutDown method.

This was implicitly done by `is_powered_on = false`, however the explicit method allows us to block until the GPU is actually gone.

This should fix a race condition while removing the other subsystems while the GPU is still active.
---
 src/video_core/gpu_thread.h | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

(limited to 'src/video_core/gpu_thread.h')

diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index 18269e51c..d384164de 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -132,8 +132,8 @@ public:
     /// Notify rasterizer that any caches of the specified region should be flushed and invalidated
     void FlushAndInvalidateRegion(VAddr addr, u64 size);
 
-    // Wait until the gpu thread is idle.
-    void WaitIdle() const;
+    // Stops the GPU execution and waits for the GPU to finish working
+    void ShutDown();
 
     void OnCommandListEnd();
 
@@ -141,6 +141,9 @@ private:
     /// Pushes a command to be executed by the GPU thread
     u64 PushCommand(CommandData&& command_data);
 
+    // Wait until the gpu thread is idle.
+    void WaitIdle() const;
+
     Core::System& system;
     const bool is_async;
     VideoCore::RasterizerInterface* rasterizer = nullptr;
-- 
cgit v1.2.3


From e6fb49fa4bb2864702abcefc14f6bb62eaba7a7e Mon Sep 17 00:00:00 2001
From: Markus Wick
Date: Wed, 7 Apr 2021 13:57:49 +0200
Subject: video_core/gpu_thread: Keep the write lock for allocating the fence.

Else the fence might get submited out-of-order into the queue, which makes testing them pointless.
Overhead should be tiny as the mutex is just moved from the queue to the writing code.
---
 src/video_core/gpu_thread.h | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

(limited to 'src/video_core/gpu_thread.h')

diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index d384164de..cb901c22a 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -101,7 +101,8 @@ struct CommandDataContainer {
 struct SynchState final {
     std::atomic_bool is_running{true};
 
-    using CommandQueue = Common::MPSCQueue<CommandDataContainer>;
+    using CommandQueue = Common::SPSCQueue<CommandDataContainer>;
+    std::mutex write_lock;
     CommandQueue queue;
     u64 last_fence{};
     std::atomic<u64> signaled_fence{};
-- 
cgit v1.2.3


From e8bd9aed8bf0f60455d0ae6a8f6f3abf92dd8305 Mon Sep 17 00:00:00 2001
From: Markus Wick
Date: Wed, 7 Apr 2021 11:41:31 +0200
Subject: video_core: Use a CV for blocking commands.

There is no need for a busy loop here. Let's just use a condition variable to save some power.
---
 src/video_core/gpu_thread.h | 11 +++++------
 1 file changed, 5 insertions(+), 6 deletions(-)

(limited to 'src/video_core/gpu_thread.h')

diff --git a/src/video_core/gpu_thread.h b/src/video_core/gpu_thread.h
index cb901c22a..11a648f38 100644
--- a/src/video_core/gpu_thread.h
+++ b/src/video_core/gpu_thread.h
@@ -90,11 +90,12 @@ using CommandData =
 struct CommandDataContainer {
     CommandDataContainer() = default;
 
-    explicit CommandDataContainer(CommandData&& data_, u64 next_fence_)
-        : data{std::move(data_)}, fence{next_fence_} {}
+    explicit CommandDataContainer(CommandData&& data_, u64 next_fence_, bool block_)
+        : data{std::move(data_)}, fence{next_fence_}, block(block_) {}
 
     CommandData data;
     u64 fence{};
+    bool block{};
 };
 
 /// Struct used to synchronize the GPU thread
@@ -106,6 +107,7 @@ struct SynchState final {
     CommandQueue queue;
     u64 last_fence{};
     std::atomic<u64> signaled_fence{};
+    std::condition_variable cv;
 };
 
 /// Class used to manage the GPU thread
@@ -140,10 +142,7 @@ public:
 
 private:
     /// Pushes a command to be executed by the GPU thread
-    u64 PushCommand(CommandData&& command_data);
-
-    // Wait until the gpu thread is idle.
-    void WaitIdle() const;
+    u64 PushCommand(CommandData&& command_data, bool block = false);
 
     Core::System& system;
     const bool is_async;
-- 
cgit v1.2.3