WindowServer: Optimize msync calls to synchronize less bytes

This *probably* makes framebuffer much faster :D Window server does not
have to do page aligned and page sized syncs which used way too many
pixels.

kernel/include/kernel/Processor.h

@@ -60,6 +60,7 @@ namespace Kernel
 		static void wait_until_processors_ready();
 
 		static void toggle_should_print_cpu_load() { s_should_print_cpu_load = !s_should_print_cpu_load; }
+		static bool get_should_print_cpu_load() { return s_should_print_cpu_load; }
 
 		static ProcessorID bsb_id() { return s_bsb_id; }
 		static bool current_is_bsb() { return current_id() == bsb_id(); }

kernel/kernel/Device/FramebufferDevice.cpp

@@ -3,11 +3,14 @@
 #include <kernel/Device/FramebufferDevice.h>
 #include <kernel/FS/DevFS/FileSystem.h>
 #include <kernel/Memory/Heap.h>
+#include <kernel/Terminal/TerminalDriver.h>
 
 #include <sys/framebuffer.h>
 #include <sys/mman.h>
 #include <sys/sysmacros.h>
 
+extern Kernel::TerminalDriver* g_terminal_driver;
+
 namespace Kernel
 {
 
@@ -250,16 +253,21 @@ namespace Kernel
 		{
 			if (flags != MS_SYNC)
 				return BAN::Error::from_errno(ENOTSUP);
+			if (vaddr % (BANAN_FB_BPP / 8))
+				return BAN::Error::from_errno(EINVAL);
 
-			if (vaddr < m_vaddr)
-				vaddr = m_vaddr;
-			if (vaddr + size > m_vaddr + m_size)
-				size = (vaddr - m_vaddr) + m_size;
+			if (auto rem = size % (BANAN_FB_BPP / 8))
+				size += (BANAN_FB_BPP / 8) - rem;
 
-			m_framebuffer->sync_pixels_linear(
-				(vaddr - m_vaddr) / (BANAN_FB_BPP / 8),
-				BAN::Math::div_round_up<uint32_t>((vaddr % (BANAN_FB_BPP / 8)) + size, (BANAN_FB_BPP / 8))
-			);
+			const vaddr_t start = BAN::Math::max(vaddr, m_vaddr);
+			const size_t  end   = BAN::Math::min(vaddr + size, m_vaddr + m_size);
+			if (start < end)
+			{
+				do_msync(
+					(start - m_vaddr) / (BANAN_FB_BPP / 8),
+					(end   - start)   / (BANAN_FB_BPP / 8)
+				);
+			}
 
 			return {};
 		}
@@ -300,6 +308,53 @@ namespace Kernel
 			, m_framebuffer(framebuffer)
 		{ }
 
+		void do_msync(uint32_t first_pixel, uint32_t pixel_count)
+		{
+			if (!Processor::get_should_print_cpu_load())
+				return m_framebuffer->sync_pixels_linear(first_pixel, pixel_count);
+
+			const uint32_t fb_width = m_framebuffer->width();
+
+			const auto& font = g_terminal_driver->font();
+
+			const uint32_t x = first_pixel % fb_width;
+			const uint32_t y = first_pixel / fb_width;
+
+			const uint32_t load_w = 16                 * font.width();
+			const uint32_t load_h = Processor::count() * font.height();
+
+			if (y >= load_h || x + pixel_count <= fb_width - load_w)
+				return m_framebuffer->sync_pixels_linear(first_pixel, pixel_count);
+
+			if (x >= fb_width - load_w && x + pixel_count <= fb_width)
+				return;
+
+			if (x < fb_width - load_w)
+				m_framebuffer->sync_pixels_linear(first_pixel, fb_width - load_w - x);
+
+			if (x + pixel_count > fb_width)
+			{
+				const uint32_t past_last_pixel = first_pixel + pixel_count;
+
+				first_pixel = (y + 1) * fb_width;
+				pixel_count = past_last_pixel - first_pixel;
+
+				const uint32_t cpu_load_end = load_h * fb_width;
+
+				while (pixel_count && first_pixel < cpu_load_end)
+				{
+					m_framebuffer->sync_pixels_linear(first_pixel, BAN::Math::min(pixel_count, fb_width - load_w));
+
+					const uint32_t advance = BAN::Math::min(pixel_count, fb_width);
+					pixel_count -= advance;
+					first_pixel += advance;
+				}
+
+				if (pixel_count)
+					m_framebuffer->sync_pixels_linear(first_pixel, pixel_count);
+			}
+		}
+
 	private:
 		BAN::RefPtr<FramebufferDevice> m_framebuffer;
 	};

kernel/kernel/IDT.cpp

@@ -179,6 +179,8 @@ namespace Kernel
 		const pid_t tid = Thread::current_tid();
 		const pid_t pid = (tid && Thread::current().has_process()) ? Process::current().pid() : 0;
 
+		const char* process_name = "";
+
 		if (tid)
 		{
 			auto& thread = Thread::current();
@@ -245,15 +247,22 @@ namespace Kernel
 			);
 		}
 
+		if (Thread::current().has_process() && Process::current().is_userspace())
+		{
+			const char* const* argv = Process::current().userspace_info().argv;
+			if (argv && *argv)
+				process_name = *argv;
+		}
+
 #if ARCH(x86_64)
 		dwarnln(
-			"CPU {}: {} (error code: 0x{8H}), pid {}, tid {}\r\n"
+			"CPU {}: {} (error code: 0x{8H}), pid {}, tid {}: {}\r\n"
 			"Register dump\r\n"
 			"rax=0x{16H}, rbx=0x{16H}, rcx=0x{16H}, rdx=0x{16H}\r\n"
 			"rsp=0x{16H}, rbp=0x{16H}, rdi=0x{16H}, rsi=0x{16H}\r\n"
 			"rip=0x{16H}, rflags=0x{16H}\r\n"
 			"cr0=0x{16H}, cr2=0x{16H}, cr3=0x{16H}, cr4=0x{16H}",
-			Processor::current_id(), isr_exceptions[isr], error, pid, tid,
+			Processor::current_id(), isr_exceptions[isr], error, pid, tid, process_name,
 			regs->rax, regs->rbx, regs->rcx, regs->rdx,
 			interrupt_stack->sp, regs->rbp, regs->rdi, regs->rsi,
 			interrupt_stack->ip, interrupt_stack->flags,
@@ -261,13 +270,13 @@ namespace Kernel
 		);
 #elif ARCH(i686)
 		dwarnln(
-			"CPU {}: {} (error code: 0x{8H}), pid {}, tid {}\r\n"
+			"CPU {}: {} (error code: 0x{8H}), pid {}, tid {}: {}\r\n"
 			"Register dump\r\n"
 			"eax=0x{8H}, ebx=0x{8H}, ecx=0x{8H}, edx=0x{8H}\r\n"
 			"esp=0x{8H}, ebp=0x{8H}, edi=0x{8H}, esi=0x{8H}\r\n"
 			"eip=0x{8H}, eflags=0x{8H}\r\n"
 			"cr0=0x{8H}, cr2=0x{8H}, cr3=0x{8H}, cr4=0x{8H}",
-			Processor::current_id(), isr_exceptions[isr], error, pid, tid,
+			Processor::current_id(), isr_exceptions[isr], error, pid, tid, process_name,
 			regs->eax, regs->ebx, regs->ecx, regs->edx,
 			interrupt_stack->sp, regs->ebp, regs->edi, regs->esi,
 			interrupt_stack->ip, interrupt_stack->flags,

kernel/kernel/Process.cpp

@@ -1682,12 +1682,9 @@ namespace Kernel
 		if (flags != MS_SYNC && flags != MS_ASYNC && flags != MS_INVALIDATE)
 			return BAN::Error::from_errno(EINVAL);
 
-		vaddr_t vaddr = (vaddr_t)addr;
-		if (vaddr % PAGE_SIZE != 0)
-			return BAN::Error::from_errno(EINVAL);
-
 		LockGuard _(m_process_lock);
 
+		const vaddr_t vaddr = reinterpret_cast<vaddr_t>(addr);
 		for (auto& mapped_region : m_mapped_regions)
 			if (mapped_region->overlaps(vaddr, len))
 				TRY(mapped_region->msync(vaddr, len, flags));

userspace/programs/WindowServer/Utils.h

@@ -80,3 +80,32 @@ struct Circle
 	}
 
 };
+
+struct Range
+{
+	uint32_t start { 0 };
+	uint32_t count { 0 };
+
+	bool is_continuous_with(const Range& range) const
+	{
+		return start <= range.start + range.count && range.start <= start + count;
+	}
+
+	uint32_t distance_between(const Range& range) const
+	{
+		if (is_continuous_with(range))
+			return 0;
+		if (start < range.start)
+			return range.start - (start + count);
+		return start - (range.start + range.count);
+	}
+
+	void merge_with(const Range& range)
+	{
+		const uint32_t new_start = BAN::Math::min(start, range.start);
+		const uint32_t new_end   = BAN::Math::max(start + count, range.start + range.count);
+
+		start = new_start;
+		count = new_end - new_start;
+	}
+};

userspace/programs/WindowServer/WindowServer.cpp

@@ -20,7 +20,12 @@ WindowServer::WindowServer(Framebuffer& framebuffer, int32_t corner_radius)
 	, m_cursor({ framebuffer.width / 2, framebuffer.height / 2 })
 	, m_font(MUST(LibFont::Font::load("/usr/share/fonts/lat0-16.psfu"_sv)))
 {
-	MUST(m_pages_to_sync_bitmap.resize(BAN::Math::div_round_up<size_t>(m_framebuffer.width * m_framebuffer.height * sizeof(uint32_t), 4096 * 8), 0));
+	BAN::Vector<LibImage::Image::Color> bitmap;
+	MUST(bitmap.resize(m_framebuffer.width * m_framebuffer.height, { 0x10, 0x10, 0x10, 0xFF }));
+	m_background_image = MUST(BAN::UniqPtr<LibImage::Image>::create(m_framebuffer.width, m_framebuffer.height, BAN::move(bitmap)));
+
+	MUST(m_pending_syncs.resize(m_framebuffer.height));
+
 	invalidate(m_framebuffer.area());
 }
 
@@ -535,6 +540,7 @@ void WindowServer::on_mouse_scroll(LibInput::MouseScrollEvent event)
 
 void WindowServer::set_focused_window(BAN::RefPtr<Window> window)
 {
+	ASSERT(window->get_attributes().focusable);
 	if (m_focused_window == window)
 		return;
 
@@ -574,6 +580,9 @@ static uint32_t alpha_blend(uint32_t color_a, uint32_t color_b)
 
 void WindowServer::invalidate(Rectangle area)
 {
+	ASSERT(m_background_image->width() == (uint64_t)m_framebuffer.width);
+	ASSERT(m_background_image->height() == (uint64_t)m_framebuffer.height);
+
 	const auto get_cursor_pixel =
 		[](int32_t rel_x, int32_t rel_y) -> BAN::Optional<uint32_t>
 		{
@@ -605,11 +614,28 @@ void WindowServer::invalidate(Rectangle area)
 			return;
 		area = focused_overlap.release_value();
 
+		const bool should_alpha_blend = m_focused_window->get_attributes().alpha_channel;
+
 		if (client_area == m_framebuffer.area())
 		{
-			for (int32_t y = area.y; y < area.y + area.height; y++)
-				for (int32_t x = area.x; x < area.x + area.width; x++)
-					m_framebuffer.mmap[y * m_framebuffer.width + x] = m_focused_window->framebuffer()[y * m_focused_window->client_width() + x];
+			if (!should_alpha_blend)
+			{
+				for (int32_t y = area.y; y < area.y + area.height; y++)
+					for (int32_t x = area.x; x < area.x + area.width; x++)
+						m_framebuffer.mmap[y * m_framebuffer.width + x] = m_focused_window->framebuffer()[y * m_focused_window->client_width() + x];
+			}
+			else
+			{
+				for (int32_t y = area.y; y < area.y + area.height; y++)
+				{
+					for (int32_t x = area.x; x < area.x + area.width; x++)
+					{
+						const uint32_t src_pixel = m_focused_window->framebuffer()[y * m_focused_window->client_width() + x];
+						const uint32_t bg_pixel = m_background_image->get_color(x, y).as_argb();
+						m_framebuffer.mmap[y * m_framebuffer.width + x] = alpha_blend(src_pixel, bg_pixel);
+					}
+				}
+			}
 			mark_pending_sync(area);
 		}
 		else
@@ -630,7 +656,12 @@ void WindowServer::invalidate(Rectangle area)
 				{
 					const int32_t src_x = BAN::Math::clamp<int32_t>(dst_x * m_focused_window->client_width()  / m_framebuffer.width,  0, m_focused_window->client_width());
 					const int32_t src_y = BAN::Math::clamp<int32_t>(dst_y * m_focused_window->client_height() / m_framebuffer.height, 0, m_focused_window->client_height());
-					m_framebuffer.mmap[dst_y * m_framebuffer.width + dst_x] = m_focused_window->framebuffer()[src_y * m_focused_window->client_width() + src_x];
+
+					const uint32_t src_pixel = m_focused_window->framebuffer()[src_y * m_focused_window->client_width() + src_x];
+					const uint32_t bg_pixel = m_background_image->get_color(dst_x, dst_y).as_argb();
+
+					uint32_t& dst_pixel = m_framebuffer.mmap[dst_y * m_framebuffer.width + dst_x];
+					dst_pixel = should_alpha_blend ? alpha_blend(src_pixel, bg_pixel) : src_pixel;
 				}
 			}
 
@@ -682,20 +713,9 @@ void WindowServer::invalidate(Rectangle area)
 		return;
 	area = fb_overlap.release_value();
 
-	if (m_background_image)
-	{
-		ASSERT(m_background_image->width() == (uint64_t)m_framebuffer.width);
-		ASSERT(m_background_image->height() == (uint64_t)m_framebuffer.height);
-		for (int32_t y = area.y; y < area.y + area.height; y++)
-			for (int32_t x = area.x; x < area.x + area.width; x++)
-				m_framebuffer.mmap[y * m_framebuffer.width + x] = m_background_image->get_color(x, y).as_argb();
-	}
-	else
-	{
-		for (int32_t y = area.y; y < area.y + area.height; y++)
-			for (int32_t x = area.x; x < area.x + area.width; x++)
-				m_framebuffer.mmap[y * m_framebuffer.width + x] = 0xFF101010;
-	}
+	for (int32_t y = area.y; y < area.y + area.height; y++)
+		for (int32_t x = area.x; x < area.x + area.width; x++)
+			m_framebuffer.mmap[y * m_framebuffer.width + x] = m_background_image->get_color(x, y).as_argb();
 
 	// FIXME: this loop should be inverse order and terminate
 	//        after window without alpha channel is found
@@ -892,24 +912,71 @@ void WindowServer::invalidate(Rectangle area)
 	mark_pending_sync(area);
 }
 
-void WindowServer::mark_pending_sync(Rectangle area)
+void WindowServer::RangeList::add_range(const Range& range)
 {
-	// FIXME: this marks too many pages
-
-	const uintptr_t mmap_start = reinterpret_cast<uintptr_t>(m_framebuffer.mmap) + area.y * m_framebuffer.width * 4;
-	const uintptr_t mmap_end = mmap_start + (area.height + 1) * m_framebuffer.width * 4;
-
-	uintptr_t mmap_addr = mmap_start & ~(uintptr_t)0xFFF;
-	while (mmap_addr < mmap_end)
+	if (range_count == 0)
 	{
-		size_t index = (mmap_addr - reinterpret_cast<uintptr_t>(m_framebuffer.mmap)) / 4096;
-		size_t byte = index / 8;
-		size_t bit  = index % 8;
-		//dprintln("{}/{}", byte, m_pages_to_sync_bitmap.size());
-		if (byte < m_pages_to_sync_bitmap.size())
-			m_pages_to_sync_bitmap[byte] |= 1 << bit;
-		mmap_addr += 4096;
+		ranges[0] = range;
+		range_count++;
+		return;
 	}
+
+	size_t min_distance_value = SIZE_MAX;
+	size_t min_distance_index = 0;
+	for (size_t i = 0; i < range_count; i++)
+	{
+		if (ranges[i].is_continuous_with(range))
+		{
+			ranges[i].merge_with(range);
+
+			size_t last_continuous = i;
+			for (size_t j = i + 1; j < range_count; j++)
+			{
+				if (!ranges[i].is_continuous_with(ranges[j]))
+					break;
+				last_continuous = j;
+			}
+
+			if (last_continuous != i)
+			{
+				ranges[i].merge_with(ranges[last_continuous]);
+				for (size_t j = 1; last_continuous + j < range_count; j++)
+					ranges[i + j] = ranges[last_continuous + j];
+				range_count -= last_continuous - i;
+			}
+
+			return;
+		}
+
+		const auto distance = ranges[i].distance_between(range);
+		if (distance < min_distance_value)
+		{
+			min_distance_value = distance;
+			min_distance_index = i;
+		}
+	}
+
+	if (range_count >= ranges.size())
+	{
+		ranges[min_distance_index].merge_with(range);
+		return;
+	}
+
+	size_t insert_idx = 0;
+	for (; insert_idx < range_count; insert_idx++)
+		if (range.start < ranges[insert_idx].start)
+			break;
+	for (size_t i = range_count; i > insert_idx; i--)
+		ranges[i] = ranges[i - 1];
+	ranges[insert_idx] = range;
+	range_count++;
+}
+
+void WindowServer::mark_pending_sync(Rectangle to_sync)
+{
+	ASSERT(to_sync == to_sync.get_overlap(m_framebuffer.area()).value());
+	for (int32_t y_off = 0; y_off < to_sync.height; y_off++)
+		m_pending_syncs[to_sync.y + y_off].add_range({ static_cast<uint32_t>(to_sync.x), static_cast<uint32_t>(to_sync.width) });
 }
 
 void WindowServer::sync()
@@ -933,33 +1000,44 @@ void WindowServer::sync()
 			dir_y = -dir_y;
 	}
 
-	for (size_t i = 0; i < m_pages_to_sync_bitmap.size() * 8; i++)
+	size_t range_start = 0;
+	size_t range_count = 0;
+	for (int32_t y = 0; y < m_framebuffer.height; y++)
 	{
-		size_t byte = i / 8;
-		size_t bit  = i % 8;
-		if (!(m_pages_to_sync_bitmap[byte] & (1 << bit)))
-			continue;
+		auto& range_list = m_pending_syncs[y];
 
-		size_t len = 1;
-		while (i + len < m_pages_to_sync_bitmap.size() * 8)
+		for (size_t i = 0; i < range_list.range_count; i++)
 		{
-			size_t byte = (i + len) / 8;
-			size_t bit  = (i + len) % 8;
-			if (!(m_pages_to_sync_bitmap[byte] & (1 << bit)))
-				break;
-			len++;
+			const size_t cur_start = y * m_framebuffer.width + range_list.ranges[i].start;
+			const size_t cur_count =                           range_list.ranges[i].count;
+
+			if (range_count == 0)
+			{
+				range_start = cur_start;
+				range_count = cur_count;
+			}
+			else
+			{
+				const size_t distance = cur_start - (range_start + range_count);
+
+				// combine nearby ranges to reduce msync calls
+				// NOTE: value of 128 is an arbitary constant that *just* felt nice
+				if (distance <= 128)
+					range_count = (cur_start + cur_count) - range_start;
+				else
+				{
+					msync(m_framebuffer.mmap + range_start, range_count * 4, MS_SYNC);
+					range_start = cur_start;
+					range_count = cur_count;
+				}
+			}
 		}
 
-		msync(
-			reinterpret_cast<uint8_t*>(m_framebuffer.mmap) + i * 4096,
-			len * 4096,
-			MS_SYNC
-		);
-
-		i += len;
+		range_list.range_count = 0;
 	}
 
-	memset(m_pages_to_sync_bitmap.data(), 0, m_pages_to_sync_bitmap.size());
+	if (range_count)
+		msync(m_framebuffer.mmap + range_start, range_count * 4, MS_SYNC);
 }
 
 Rectangle WindowServer::cursor_area() const
@@ -1002,8 +1080,14 @@ void WindowServer::remove_client_fd(int fd)
 			if (window == m_focused_window)
 			{
 				m_focused_window = nullptr;
-				if (!m_client_windows.empty())
-					set_focused_window(m_client_windows.back());
+				for (size_t j = m_client_windows.size(); j > 0; j--)
+				{
+					auto& client_window = m_client_windows[j - 1];
+					if (!client_window->get_attributes().focusable)
+						continue;
+					set_focused_window(client_window);
+					break;
+				}
 			}
 
 			break;

userspace/programs/WindowServer/WindowServer.h

@@ -3,6 +3,7 @@
 #include "Framebuffer.h"
 #include "Window.h"
 
+#include <BAN/Array.h>
 #include <BAN/Function.h>
 #include <BAN/Iteration.h>
 #include <BAN/Vector.h>
@@ -59,6 +60,14 @@ public:
 private:
 	void mark_pending_sync(Rectangle area);
 
+private:
+	struct RangeList
+	{
+		size_t range_count { 0 };
+		BAN::Array<Range, 32> ranges;
+		void add_range(const Range& range);
+	};
+
 private:
 	Framebuffer& m_framebuffer;
 	BAN::Vector<BAN::RefPtr<Window>> m_client_windows;
@@ -67,7 +76,7 @@ private:
 
 	const int32_t m_corner_radius;
 
-	BAN::Vector<uint8_t> m_pages_to_sync_bitmap;
+	BAN::Vector<RangeList> m_pending_syncs;
 
 	BAN::UniqPtr<LibImage::Image> m_background_image;
 

userspace/programs/WindowServer/main.cpp

@@ -198,19 +198,23 @@ int main()
 		if (auto ret = window_server.set_background_image(BAN::move(config.background_image)); ret.is_error())
 			dwarnln("Could not set background image: {}", ret.error());
 
+	const auto get_current_us =
+		[]() -> uint64_t
+		{
+			timespec current_ts;
+			clock_gettime(CLOCK_MONOTONIC, &current_ts);
+			return (current_ts.tv_sec * 1'000'000) + (current_ts.tv_nsec / 1'000);
+		};
+
 	constexpr uint64_t sync_interval_us = 1'000'000 / 60;
-	timespec last_sync { .tv_sec = 0, .tv_nsec = 0 };
+	uint64_t last_sync_us = 0;
 	while (!window_server.is_stopped())
 	{
-		timespec current_ts;
-		clock_gettime(CLOCK_MONOTONIC, &current_ts);
-
-		uint64_t us_since_last_sync = (current_ts.tv_sec - last_sync.tv_sec) * 1'000'000 + (current_ts.tv_nsec - last_sync.tv_nsec) / 1000;
-		if (us_since_last_sync > sync_interval_us)
+		const auto current_us = get_current_us();
+		if (current_us - last_sync_us > sync_interval_us)
 		{
 			window_server.sync();
-			us_since_last_sync = 0;
-			last_sync = current_ts;
+			last_sync_us += sync_interval_us;
 		}
 
 		int max_fd = server_fd;
@@ -230,10 +234,10 @@ int main()
 		}
 		max_fd = BAN::Math::max(max_fd, window_server.get_client_fds(fds));
 
-		timeval select_timeout {
-			.tv_sec = 0,
-			.tv_usec = static_cast<long>(sync_interval_us - us_since_last_sync)
-		};
+		timeval select_timeout {};
+		if (auto current_us = get_current_us(); current_us - last_sync_us < sync_interval_us)
+			select_timeout.tv_usec = sync_interval_us - (current_us - last_sync_us);
+
 		int nselect = select(max_fd + 1, &fds, nullptr, nullptr, &select_timeout);
 		if (nselect == 0)
 			continue;