rsx: Narrow the race condition window further

- Needs aliased paging to be implemented to fix properly or a re-entrant global IO lock

Author: kd-11

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -939,25 +939,30 @@ namespace rsx
 				}
 
 				//TODO: This bit can cause race conditions if other threads are accessing this memory
-				//1. Unprotect all memory in case of overlapping pages
+				//1. Force readback if surface is not synchronized yet to make unlocked part finish quickly
 				for (auto &tex : data.sections_to_flush)
 				{
 					if (tex->is_locked())
 					{
-						tex->unprotect();
+						if (!tex->is_synchronized())
+							tex->copy_texture(true, std::forward<Args>(extras)...);
+
 						m_cache[get_block_address(tex->get_section_base())].remove_one();
 					}
 				}
 
-				//2. Write all the memory
+				//TODO: Acquire global io lock here
+
+				//2. Unprotect all the memory
 				for (auto &tex : data.sections_to_flush)
 				{
-					if (!tex->flush(std::forward<Args>(extras)...))
-					{
-						//Missed address, note this
-						//TODO: Lower severity when successful to keep the cache from overworking
-						record_cache_miss(*tex);
-					}
+					tex->unprotect();
+				}
+
+				//3. Write all the memory
+				for (auto &tex : data.sections_to_flush)
+				{
+					tex->flush(std::forward<Args>(extras)...);
 				}
 
 				//Restore protection on the sections to reprotect

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -794,7 +794,7 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 	vk::texture_cache::thrashed_set result;
 	{
 		std::lock_guard<std::mutex> lock(m_secondary_cb_guard);
-		result = std::move(m_texture_cache.invalidate_address(address, is_writing, false, *m_device, m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue()));
+		result = std::move(m_texture_cache.invalidate_address(address, is_writing, false, m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue()));
 	}
 
 	if (!result.violation_handled)
@@ -882,7 +882,7 @@ bool VKGSRender::on_access_violation(u32 address, bool is_writing)
 			}
 		}
 
-		m_texture_cache.flush_all(result, *m_device, m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue());
+		m_texture_cache.flush_all(result, m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue());
 
 		if (has_queue_ref)
 		{
@@ -897,7 +897,7 @@ void VKGSRender::on_notify_memory_unmapped(u32 address_base, u32 size)
 {
 	std::lock_guard<std::mutex> lock(m_secondary_cb_guard);
 	if (m_texture_cache.invalidate_range(address_base, size, true, true, false,
-		*m_device, m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue()).violation_handled)
+		m_secondary_command_buffer, m_memory_type_mapping, m_swap_chain->get_present_queue()).violation_handled)
 	{
 		m_texture_cache.purge_dirty();
 		{

rpcs3/Emu/RSX/VK/VKTextureCache.h

@@ -244,12 +244,14 @@ namespace vk
 			}
 		}
 
-		bool flush(vk::render_device& dev, vk::command_buffer& cmd, vk::memory_type_mapping& memory_types, VkQueue submit_queue)
+		bool flush(vk::command_buffer& cmd, vk::memory_type_mapping& memory_types, VkQueue submit_queue)
 		{
 			if (flushed) return true;
 
 			if (m_device == nullptr)
-				m_device = &dev;
+			{
+				m_device = &cmd.get_command_pool().get_owner();
+			}
 
 			// Return false if a flush occured 'late', i.e we had a miss
 			bool result = true;
@@ -883,7 +885,7 @@ namespace vk
 		template<typename RsxTextureType>
 		sampled_image_descriptor _upload_texture(vk::command_buffer& cmd, RsxTextureType& tex, rsx::vk_render_targets& m_rtts)
 		{
-			return upload_texture(cmd, tex, m_rtts, *m_device, cmd, m_memory_types, const_cast<const VkQueue>(m_submit_queue));
+			return upload_texture(cmd, tex, m_rtts, cmd, m_memory_types, const_cast<const VkQueue>(m_submit_queue));
 		}
 
 		bool blit(rsx::blit_src_info& src, rsx::blit_dst_info& dst, bool interpolate, rsx::vk_render_targets& m_rtts, vk::command_buffer& cmd)
@@ -924,7 +926,7 @@ namespace vk
 			}
 			helper(&cmd);
 
-			return upload_scaled_image(src, dst, interpolate, cmd, m_rtts, helper, *m_device, cmd, m_memory_types, const_cast<const VkQueue>(m_submit_queue));
+			return upload_scaled_image(src, dst, interpolate, cmd, m_rtts, helper, cmd, m_memory_types, const_cast<const VkQueue>(m_submit_queue));
 		}
 
 		const u32 get_unreleased_textures_count() const override