rsx: Do not clobber CELL memory when a surface is partially inherited

rpcs3/Emu/Cell/SPUThread.cpp

@@ -4326,7 +4326,7 @@ s64 spu_thread::get_ch_value(u32 ch)
 					atomic_wait_engine::set_one_time_use_wait_callback(mask1 != SPU_EVENT_LR ? nullptr : +[](u64 attempts) -> bool
 					{
 						const auto _this = static_cast<spu_thread*>(cpu_thread::get_current());
-						AUDIT(_this->id_type() == 1);
+						AUDIT(_this->id_type() == 2);
 
 						const auto old = +_this->state;
 

rpcs3/Emu/Memory/vm.cpp

@@ -1544,7 +1544,7 @@ namespace vm
 
 	static void load_memory_bytes(utils::serial& ar, u8* ptr, usz size)
 	{
-		AUDIT(ar.is_writing() && !(size % 128));
+		AUDIT(!ar.is_writing() && !(size % 128));
 
 		for (; size; ptr += 128 * 8, size -= 128 * 8)
 		{

rpcs3/Emu/RSX/Common/simple_array.hpp

@@ -1,6 +1,7 @@
 #pragma once
 
 #include <util/types.hpp>
+#include <functional>
 
 namespace rsx
 {
@@ -282,5 +283,35 @@ namespace rsx
 		{
 			return _data ? _data + _size : nullptr;
 		}
+
+		bool any(std::predicate<const Ty&> auto predicate) const
+		{
+			for (auto it = begin(); it != end(); ++it)
+			{
+				if (std::invoke(predicate, *it))
+				{
+					return true;
+				}
+			}
+			return false;
+		}
+
+		void filter(std::predicate<const Ty&> auto predicate)
+		{
+			if (!_size)
+			{
+				return;
+			}
+
+			for (auto ptr = _data, last = _data + _size - 1; ptr < last; ptr++)
+			{
+				if (!predicate(*ptr))
+				{
+					// Move item to the end of the list and shrink by 1
+					std::memcpy(ptr, last, sizeof(Ty));
+					last = _data + (--_size);
+				}
+			}
+		}
 	};
 }

rpcs3/Emu/RSX/Common/surface_store.h

@@ -68,7 +68,7 @@ namespace rsx
 		std::pair<u32, surface_type> m_bound_depth_stencil = {};
 
 		// List of sections derived from a section that has been split and invalidated
-		std::vector<surface_type> orphaned_surfaces;
+		std::vector<std::pair<u32, surface_type>> orphaned_surfaces;
 
 		// List of sections that have been wholly inherited and invalidated
 		std::vector<surface_type> superseded_surfaces;
@@ -156,7 +156,7 @@ namespace rsx
 				}
 
 				ensure(region.target == Traits::get(sink));
-				orphaned_surfaces.push_back(region.target);
+				orphaned_surfaces.push_back({ address, region.target });
 				data.emplace(region.target->get_memory_range(), std::move(sink));
 			};
 

rpcs3/Emu/RSX/Common/texture_cache.h

@@ -3237,6 +3237,24 @@ namespace rsx
 			return m_predictor;
 		}
 
+		bool is_protected(u32 section_base_address, const address_range& test_range, rsx::texture_upload_context context)
+		{
+			reader_lock lock(m_cache_mutex);
+
+			const auto& block = m_storage.block_for(section_base_address);
+			for (const auto& tex : block)
+			{
+				if (tex.get_section_base() == section_base_address)
+				{
+					return tex.get_context() == context &&
+						tex.is_locked() &&
+						test_range.inside(tex.get_section_range());
+				}
+			}
+
+			return false;
+		}
+
 
 		/**
 		 * The read only texture invalidate flag is set if a read only texture is trampled by framebuffer memory

rpcs3/Emu/RSX/GL/GLRenderTargets.cpp

@@ -310,54 +310,26 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool /*
 		m_rtts.superseded_surfaces.clear();
 	}
 
-	const auto color_format = rsx::internals::surface_color_format_to_gl(m_framebuffer_layout.color_format);
-	for (u8 i = 0; i < rsx::limits::color_buffers_count; ++i)
-	{
-		if (!m_surface_info[i].address || !m_surface_info[i].pitch) continue;
-
-		const auto surface_range = m_surface_info[i].get_memory_range();
-		if (g_cfg.video.write_color_buffers)
-		{
-			// Mark buffer regions as NO_ACCESS on Cell-visible side
-			m_gl_texture_cache.lock_memory_region(
-				cmd, m_rtts.m_bound_render_targets[i].second, surface_range, true,
-				m_surface_info[i].width, m_surface_info[i].height, m_surface_info[i].pitch,
-				color_format.format, color_format.type, color_format.swap_bytes);
-		}
-		else
-		{
-			m_gl_texture_cache.commit_framebuffer_memory_region(cmd, surface_range);
-		}
-	}
-
-	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
-	{
-		const auto surface_range = m_depth_surface_info.get_memory_range();
-		if (g_cfg.video.write_depth_buffer)
-		{
-			const auto depth_format_gl = rsx::internals::surface_depth_format_to_gl(m_framebuffer_layout.depth_format);
-			m_gl_texture_cache.lock_memory_region(
-				cmd, m_rtts.m_bound_depth_stencil.second, surface_range, true,
-				m_depth_surface_info.width, m_depth_surface_info.height, m_depth_surface_info.pitch,
-				depth_format_gl.format, depth_format_gl.type, depth_format_gl.type != gl::texture::type::uint_24_8);
-		}
-		else
-		{
-			m_gl_texture_cache.commit_framebuffer_memory_region(cmd, surface_range);
-		}
-	}
-
 	if (!m_rtts.orphaned_surfaces.empty())
 	{
 		gl::texture::format format;
 		gl::texture::type type;
 		bool swap_bytes;
 
-		for (auto& surface : m_rtts.orphaned_surfaces)
+		for (auto& [base_addr, surface] : m_rtts.orphaned_surfaces)
 		{
-			const bool lock = surface->is_depth_surface() ? !!g_cfg.video.write_depth_buffer :
+			bool lock = surface->is_depth_surface() ? !!g_cfg.video.write_depth_buffer :
 				!!g_cfg.video.write_color_buffers;
 
+			if (lock &&
+				!m_gl_texture_cache.is_protected(
+					base_addr,
+					surface->get_memory_range(),
+					rsx::texture_upload_context::framebuffer_storage))
+			{
+				lock = false;
+			}
+
 			if (!lock) [[likely]]
 			{
 				m_gl_texture_cache.commit_framebuffer_memory_region(cmd, surface->get_memory_range());
@@ -388,6 +360,43 @@ void GLGSRender::init_buffers(rsx::framebuffer_creation_context context, bool /*
 		m_rtts.orphaned_surfaces.clear();
 	}
 
+	const auto color_format = rsx::internals::surface_color_format_to_gl(m_framebuffer_layout.color_format);
+	for (u8 i = 0; i < rsx::limits::color_buffers_count; ++i)
+	{
+		if (!m_surface_info[i].address || !m_surface_info[i].pitch) continue;
+
+		const auto surface_range = m_surface_info[i].get_memory_range();
+		if (g_cfg.video.write_color_buffers)
+		{
+			// Mark buffer regions as NO_ACCESS on Cell-visible side
+			m_gl_texture_cache.lock_memory_region(
+				cmd, m_rtts.m_bound_render_targets[i].second, surface_range, true,
+				m_surface_info[i].width, m_surface_info[i].height, m_surface_info[i].pitch,
+				color_format.format, color_format.type, color_format.swap_bytes);
+		}
+		else
+		{
+			m_gl_texture_cache.commit_framebuffer_memory_region(cmd, surface_range);
+		}
+	}
+
+	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
+	{
+		const auto surface_range = m_depth_surface_info.get_memory_range();
+		if (g_cfg.video.write_depth_buffer)
+		{
+			const auto depth_format_gl = rsx::internals::surface_depth_format_to_gl(m_framebuffer_layout.depth_format);
+			m_gl_texture_cache.lock_memory_region(
+				cmd, m_rtts.m_bound_depth_stencil.second, surface_range, true,
+				m_depth_surface_info.width, m_depth_surface_info.height, m_depth_surface_info.pitch,
+				depth_format_gl.format, depth_format_gl.type, depth_format_gl.type != gl::texture::type::uint_24_8);
+		}
+		else
+		{
+			m_gl_texture_cache.commit_framebuffer_memory_region(cmd, surface_range);
+		}
+	}
+
 	if (m_gl_texture_cache.get_ro_tex_invalidate_intr())
 	{
 		// Invalidate cached sampler state

rpcs3/Emu/RSX/VK/VKGSRender.cpp

@@ -2416,51 +2416,25 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 		m_rtts.superseded_surfaces.clear();
 	}
 
-	const auto color_fmt_info = get_compatible_gcm_format(m_framebuffer_layout.color_format);
-	for (u8 index : m_draw_buffers)
-	{
-		if (!m_surface_info[index].address || !m_surface_info[index].pitch) continue;
-
-		const utils::address_range surface_range = m_surface_info[index].get_memory_range();
-		if (g_cfg.video.write_color_buffers)
-		{
-			m_texture_cache.lock_memory_region(
-				*m_current_command_buffer, m_rtts.m_bound_render_targets[index].second, surface_range, true,
-				m_surface_info[index].width, m_surface_info[index].height, m_framebuffer_layout.actual_color_pitch[index],
-				color_fmt_info.first, color_fmt_info.second);
-		}
-		else
-		{
-			m_texture_cache.commit_framebuffer_memory_region(*m_current_command_buffer, surface_range);
-		}
-	}
-
-	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
-	{
-		const utils::address_range surface_range = m_depth_surface_info.get_memory_range();
-		if (g_cfg.video.write_depth_buffer)
-		{
-			const u32 gcm_format = (m_depth_surface_info.depth_format == rsx::surface_depth_format::z16) ? CELL_GCM_TEXTURE_DEPTH16 : CELL_GCM_TEXTURE_DEPTH24_D8;
-			m_texture_cache.lock_memory_region(
-				*m_current_command_buffer, m_rtts.m_bound_depth_stencil.second, surface_range, true,
-				m_depth_surface_info.width, m_depth_surface_info.height, m_framebuffer_layout.actual_zeta_pitch, gcm_format, true);
-		}
-		else
-		{
-			m_texture_cache.commit_framebuffer_memory_region(*m_current_command_buffer, surface_range);
-		}
-	}
-
 	if (!m_rtts.orphaned_surfaces.empty())
 	{
 		u32 gcm_format;
 		bool swap_bytes;
 
-		for (auto& surface : m_rtts.orphaned_surfaces)
+		for (auto& [base_addr, surface] : m_rtts.orphaned_surfaces)
 		{
-			const bool lock = surface->is_depth_surface() ? !!g_cfg.video.write_depth_buffer :
+			bool lock = surface->is_depth_surface() ? !!g_cfg.video.write_depth_buffer :
 				!!g_cfg.video.write_color_buffers;
 
+			if (lock &&
+				!m_texture_cache.is_protected(
+					base_addr,
+					surface->get_memory_range(),
+					rsx::texture_upload_context::framebuffer_storage))
+			{
+				lock = false;
+			}
+
 			if (!lock) [[likely]]
 			{
 				m_texture_cache.commit_framebuffer_memory_region(*m_current_command_buffer, surface->get_memory_range());
@@ -2488,6 +2462,41 @@ void VKGSRender::prepare_rtts(rsx::framebuffer_creation_context context)
 		m_rtts.orphaned_surfaces.clear();
 	}
 
+	const auto color_fmt_info = get_compatible_gcm_format(m_framebuffer_layout.color_format);
+	for (u8 index : m_draw_buffers)
+	{
+		if (!m_surface_info[index].address || !m_surface_info[index].pitch) continue;
+
+		const utils::address_range surface_range = m_surface_info[index].get_memory_range();
+		if (g_cfg.video.write_color_buffers)
+		{
+			m_texture_cache.lock_memory_region(
+				*m_current_command_buffer, m_rtts.m_bound_render_targets[index].second, surface_range, true,
+				m_surface_info[index].width, m_surface_info[index].height, m_framebuffer_layout.actual_color_pitch[index],
+				color_fmt_info.first, color_fmt_info.second);
+		}
+		else
+		{
+			m_texture_cache.commit_framebuffer_memory_region(*m_current_command_buffer, surface_range);
+		}
+	}
+
+	if (m_depth_surface_info.address && m_depth_surface_info.pitch)
+	{
+		const utils::address_range surface_range = m_depth_surface_info.get_memory_range();
+		if (g_cfg.video.write_depth_buffer)
+		{
+			const u32 gcm_format = (m_depth_surface_info.depth_format == rsx::surface_depth_format::z16) ? CELL_GCM_TEXTURE_DEPTH16 : CELL_GCM_TEXTURE_DEPTH24_D8;
+			m_texture_cache.lock_memory_region(
+				*m_current_command_buffer, m_rtts.m_bound_depth_stencil.second, surface_range, true,
+				m_depth_surface_info.width, m_depth_surface_info.height, m_framebuffer_layout.actual_zeta_pitch, gcm_format, true);
+		}
+		else
+		{
+			m_texture_cache.commit_framebuffer_memory_region(*m_current_command_buffer, surface_range);
+		}
+	}
+
 	m_current_renderpass_key = vk::get_renderpass_key(m_fbo_images);
 	m_cached_renderpass = vk::get_renderpass(*m_device, m_current_renderpass_key);