RSX/Savestates: Replace GCM hack with a proper fix

Utilities/Thread.cpp

@@ -1536,14 +1536,14 @@ bool handle_access_violation(u32 addr, bool is_writing, ucontext_t* context) noe
 			// If we fail due to being busy, wait a bit and try again.
 			while (static_cast<u32>(sending_error) == CELL_EBUSY)
 			{
+				thread_ctrl::wait_for(1000);
+				sending_error = sys_event_port_send(pf_port_id, data1, data2, data3);
+
 				if (cpu->is_stopped())
 				{
 					sending_error = {};
 					break;
 				}
-
-				thread_ctrl::wait_for(1000);
-				sending_error = sys_event_port_send(pf_port_id, data1, data2, data3);
 			}
 
 			if (sending_error)

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1,4 +1,4 @@
-#include "stdafx.h"
+#include "stdafx.h"
 #include "Utilities/JIT.h"
 #include "Utilities/date_time.h"
 #include "Emu/Memory/vm.h"
@@ -1704,7 +1704,7 @@ void spu_thread::serialize_common(utils::serial& ar)
 		, ch_out_mbox.data, ch_out_intr_mbox.data, snr_config, ch_snr1.data, ch_snr2.data, ch_events.raw().all, interrupts_enabled
 		, run_ctrl, exit_status.data, status_npc.raw().status);
 
-	if (GET_SERIALIZATION_VERSION(spu) != 1u)
+	if (GET_SERIALIZATION_VERSION(spu) != 1)
 	{
 		ar(ch_dec_start_timestamp, ch_dec_value, is_dec_frozen);
 	}

rpcs3/Emu/Cell/lv2/sys_event.cpp

@@ -239,7 +239,7 @@ error_code sys_event_queue_destroy(ppu_thread& ppu, u32 equeue_id, s32 mode)
 
 	const auto queue = idm::withdraw<lv2_obj, lv2_event_queue>(equeue_id, [&](lv2_event_queue& queue) -> CellError
 	{
-		qlock.lock(queue.mutex);
+		qlock = std::unique_lock{queue.mutex};
 
 		if (!mode && !queue.sq.empty())
 		{
@@ -456,8 +456,14 @@ error_code sys_event_queue_receive(ppu_thread& ppu, u32 equeue_id, vm::ptr<sys_e
 
 		if (is_stopped(state))
 		{
-			extern void signal_gcm_intr_thread_offline(lv2_event_queue&);
-			signal_gcm_intr_thread_offline(*queue);
+			std::lock_guard lock_rsx(queue->mutex);
+
+			if (std::find(queue->sq.begin(), queue->sq.end(), &ppu) == queue->sq.end())
+			{
+				break;
+			}
+
+			ppu.state += cpu_flag::again;
 			return {};
 		}
 
@@ -686,7 +692,8 @@ error_code sys_event_port_send(u32 eport_id, u64 data1, u64 data2, u64 data3)
 	{
 		if (port.ret == CELL_EAGAIN)
 		{
-			return CELL_OK;
+			// Not really an error code exposed to games (thread has raised cpu_flag::again)
+			return not_an_error(CELL_EAGAIN);
 		}
 
 		if (port.ret == CELL_EBUSY)

rpcs3/Emu/Cell/lv2/sys_rsx.cpp

@@ -37,7 +37,7 @@ u64 rsxTimeStamp()
 	return get_timebased_time();
 }
 
-void rsx::thread::send_event(u64 data1, u64 event_flags, u64 data3) const
+bool rsx::thread::send_event(u64 data1, u64 event_flags, u64 data3)
 {
 	// Filter event bits, send them only if they are masked by gcm
 	// Except the upper 32-bits, they are reserved for unmapped io events and execute unconditionally
@@ -46,7 +46,7 @@ void rsx::thread::send_event(u64 data1, u64 event_flags, u64 data3) const
 	if (!event_flags)
 	{
 		// Nothing to do
-		return;
+		return true;
 	}
 
 	auto error = sys_event_port_send(rsx_event_port, data1, event_flags, data3);
@@ -76,35 +76,19 @@ void rsx::thread::send_event(u64 data1, u64 event_flags, u64 data3) const
 		error = sys_event_port_send(rsx_event_port, data1, event_flags, data3);
 	}
 
-	if (!Emu.IsPaused() && error && error + 0u != CELL_ENOTCONN)
+	if (error + 0u == CELL_EAGAIN)
+	{
+		// Thread has aborted when sending event (VBLANK duplicates are allowed)
+		ensure((unsent_gcm_events.fetch_or(event_flags) & event_flags & ~(SYS_RSX_EVENT_VBLANK | SYS_RSX_EVENT_SECOND_VBLANK_BASE | SYS_RSX_EVENT_SECOND_VBLANK_BASE * 2)) == 0);
+		return false;
+	}
+
+	if (error && error + 0u != CELL_ENOTCONN)
 	{
 		fmt::throw_exception("rsx::thread::send_event() Failed to send event! (error=%x)", +error);
 	}
-}
 
-// Returns true on success of receiving the event
-void signal_gcm_intr_thread_offline(lv2_event_queue& q)
-{
-	const auto render = rsx::get_current_renderer();
-	const auto cpu = cpu_thread::get_current();
-
-	static shared_mutex s_dummy;
-
-	std::scoped_lock lock_rsx(render ? render->sys_rsx_mtx : s_dummy, q.mutex);
-
-	if (std::find(q.sq.begin(), q.sq.end(), cpu) == q.sq.end())
-	{
-		return;
-	}
-
-	cpu->state += cpu_flag::again;
-
-	if (!vm::check_addr(render->driver_info) || vm::_ref<RsxDriverInfo>(render->driver_info).handler_queue != q.id)
-	{
-		return;
-	}
-
-	render->gcm_intr_thread_offline = true;
+	return true;
 }
 
 error_code sys_rsx_device_open(cpu_thread& cpu)
@@ -534,7 +518,10 @@ error_code sys_rsx_context_attribute(u32 context_id, u32 package_id, u64 a3, u64
 		driverInfo.head[a3].flipFlags |= 0x40000000 | (1 << a4);
 
 		render->on_frame_end(static_cast<u32>(a4));
-		render->send_event(0, SYS_RSX_EVENT_QUEUE_BASE << a3, 0);
+		if (!render->send_event(0, SYS_RSX_EVENT_QUEUE_BASE << a3, 0))
+		{
+			break;
+		}
 
 		if (g_cfg.video.frame_limit == frame_limit_type::infinite)
 		{
@@ -780,19 +767,6 @@ error_code sys_rsx_context_attribute(u32 context_id, u32 package_id, u64 a3, u64
 
 	case 0xFEC: // hack: flip event notification
 	{
-		reader_lock lock(render->sys_rsx_mtx);
-
-		if (render->gcm_intr_thread_offline)
-		{
-			if (auto cpu = get_current_cpu_thread())
-			{
-				cpu->state += cpu_flag::exit;
-				cpu->state += cpu_flag::again;
-			}
-
-			break;
-		}
-
 		// we only ever use head 1 for now
 		driverInfo.head[1].flipFlags |= 0x80000000;
 		driverInfo.head[1].lastFlipTime = rsxTimeStamp(); // should rsxthread set this?
@@ -817,13 +791,6 @@ error_code sys_rsx_context_attribute(u32 context_id, u32 package_id, u64 a3, u64
 		// NOTE: There currently seem to only be 2 active heads on PS3
 		ensure(a3 < 2);
 
-		reader_lock lock(render->sys_rsx_mtx);
-
-		if (render->gcm_intr_thread_offline)
-		{
-			break;
-		}
-
 		// todo: this is wrong and should be 'second' vblank handler and freq, but since currently everything is reported as being 59.94, this should be fine
 		driverInfo.head[a3].lastSecondVTime.atomic_op([&](be_t<u64>& time)
 		{
@@ -852,19 +819,6 @@ error_code sys_rsx_context_attribute(u32 context_id, u32 package_id, u64 a3, u64
 
 	case 0xFEF: // hack: user command
 	{
-		reader_lock lock(render->sys_rsx_mtx);
-
-		if (render->gcm_intr_thread_offline)
-		{
-			if (auto cpu = get_current_cpu_thread())
-			{
-				cpu->state += cpu_flag::exit;
-				cpu->state += cpu_flag::again;
-			}
-
-			break;
-		}
-
 		// 'custom' invalid package id for now
 		// as i think we need custom lv1 interrupts to handle this accurately
 		// this also should probly be set by rsxthread

rpcs3/Emu/RSX/RSXThread.cpp

@@ -453,6 +453,11 @@ namespace rsx
 		ar(device_addr, label_addr, main_mem_size, local_mem_size, rsx_event_port, driver_info);
 		ar(in_begin_end, zcull_stats_enabled, zcull_rendering_enabled, zcull_pixel_cnt_enabled);
 		ar(display_buffers, display_buffers_count, current_display_buffer);
+
+		if (ar.is_writing() || GET_SERIALIZATION_VERSION(rsx) != 1u)
+		{
+			ar(unsent_gcm_events);
+		}
 	}
 
 	thread::thread(utils::serial* _ar)
@@ -756,6 +761,14 @@ namespace rsx
 
 		vblank_count = 0;
 
+		if (u64 event_flags = unsent_gcm_events.exchange(0))
+		{
+			if (!send_event(0, event_flags, 0))
+			{
+				return;
+			}
+		}
+
 		g_fxo->init<named_thread>("VBlank Thread", [this]()
 		{
 			// See sys_timer_usleep for details
@@ -772,7 +785,7 @@ namespace rsx
 			u64 local_vblank_count = 0;
 
 			// TODO: exit condition
-			while (!is_stopped())
+			while (!is_stopped() && !unsent_gcm_events)
 			{
 				// Get current time
 				const u64 current = get_system_time();
@@ -3407,6 +3420,13 @@ namespace rsx
 			if (!isHLE)
 			{
 				sys_rsx_context_attribute(0x55555555, 0xFEC, buffer, 0, 0, 0);
+
+				if (unsent_gcm_events)
+				{
+					// TODO: A proper fix
+					return;
+				}
+
 				continue;
 			}
 

rpcs3/Emu/RSX/RSXThread.h

@@ -584,9 +584,10 @@ namespace rsx
 		u32 local_mem_size{0};
 		u32 rsx_event_port{0};
 		u32 driver_info{0};
-		bool gcm_intr_thread_offline = false; // Hack for savestates
 
-		void send_event(u64, u64, u64) const;
+		atomic_t<u64> unsent_gcm_events = 0; // Unsent event bits when aborting RSX/VBLANK thread (will be sent on savestate load)
+
+		bool send_event(u64, u64, u64);
 
 		bool m_rtts_dirty = true;
 		std::array<bool, 16> m_textures_dirty;

rpcs3/Emu/System.cpp

@@ -67,8 +67,8 @@ u64 g_rtm_tx_limit2 = 0;
 struct serial_ver_t
 {
 	bool used = false;
-	u32 current_version = 0;
-	std::set<u32> compatible_versions;
+	s32 current_version = 0;
+	std::set<s32> compatible_versions;
 };
 
 static std::array<serial_ver_t, 23> s_serial_versions;
@@ -83,7 +83,7 @@ static std::array<serial_ver_t, 23> s_serial_versions;
 		::s_serial_versions[identifier].used = true;\
 	}\
 \
-	extern u32 get_##name##_serialization_version()\
+	extern s32 get_##name##_serialization_version()\
 	{\
 		return ::s_serial_versions[identifier].current_version;\
 	}
@@ -101,7 +101,7 @@ SERIALIZATION_VER(lv2_config, 9,                                1)
 
 namespace rsx
 {
-	SERIALIZATION_VER(rsx, 10,                                  1)
+	SERIALIZATION_VER(rsx, 10,                                  1, 2)
 }
 
 namespace np
@@ -111,7 +111,7 @@ namespace np
 
 #ifdef _MSC_VER
 // Compiler bug, lambda function body does seem to inherit used namespace atleast for function decleration 
-SERIALIZATION_VER(rsx, 10,                                      1)
+SERIALIZATION_VER(rsx, 10,                                      1, 2)
 SERIALIZATION_VER(sceNp, 11,                                    1)
 #endif
 

rpcs3/util/types.hpp

@@ -1118,7 +1118,7 @@ extern bool serialize(utils::serial& ar, T& obj);
 
 #define GET_SERIALIZATION_VERSION(name) []()\
 {\
-	extern u32 get_##name##_serialization_version();\
+	extern s32 get_##name##_serialization_version();\
 	return get_##name##_serialization_version();\
 }()