Kernel: Rework scheduler/processor stacks.

kernel/arch/i686/Thread.S

@@ -4,40 +4,31 @@ read_ip:
 	popl %eax
 	jmp *%eax
 
-# void start_thread(uint32_t sp, uint32_t ip)
-.global start_thread
-start_thread:
-	movl 8(%esp), %edi # ip
-	movl 4(%esp), %esp # sp
+# void start_kernel_thread()
+.global start_kernel_thread
+start_kernel_thread:
+	call get_start_kernel_thread_sp
+	movl %eax, %esp
 
 	# STACK LAYOUT
-	#   NULL
-	#   thread ptr
-	#   &Thread::on_exit
-	#   data
+	#   on_exit arg
+	#   on_exit func
+	#   entry arg
+	#   entry func
 
-	xorl %ebp, %ebp
+	movl 4(%esp), %edi
+	movl 0(%esp), %esi
 
+	subl $12, %esp
+	pushl %edi
 	sti
-	call *%edi
+	call *%esi
+	addl $16, %esp
 
-	movl 4(%esp), %edi # &Thread::on_exit
+	movl 12(%esp), %edi
+	movl  8(%esp), %esi
 
-	movl 8(%esp), %eax # thread ptr
-	movl %eax, (%esp)
-
-	call *%edi
-
-
-# void continue_thread(uint32_t sp, uint32_t ip)
-.global continue_thread
-continue_thread:
-	movl 8(%esp), %edi # ip
-	movl 4(%esp), %esp # sp
-	xorl %eax, %eax
-	jmp *%edi
-
-# void thread_userspace_trampoline(uint32_t sp, uint32_t ip, int argc, char** argv, char** envp)
-.global thread_userspace_trampoline
-thread_userspace_trampoline:
-	ud2
+	subl $12, %esp
+	pushl %edi
+	call *%esi
+	addl $16, %esp

kernel/arch/i686/interrupts.S

@@ -26,17 +26,31 @@ isr_stub:
 irq_stub:
 	pusha
 
-	leal 40(%esp), %eax // interrupt stack ptr
-	movl 32(%esp), %ebx // irq number
+	movl 32(%esp), %eax # interrupt number
+
+	subl $8, %esp
+	pushl %eax
+	call cpp_irq_handler
+	addl $12, %esp
+
+	popa
+	addl $8, %esp
+	iret
+
+.global asm_reschedule_handler
+asm_reschedule_handler:
+	pusha
+
+	movl %esp,     %eax # interrupt registers ptr
+	leal 32(%esp), %ebx # interrupt stack ptr
 
 	subl $12, %esp
 	pushl %eax
 	pushl %ebx
-	call cpp_irq_handler
+	call cpp_reschedule_handler
 	addl $20, %esp
 
 	popa
-	addl $8, %esp
 	iret
 
 // arguments in EAX, EBX, ECX, EDX, ESI, EDI
@@ -158,4 +172,3 @@ irq 28
 irq 29
 irq 30
 irq 31
-irq 32

kernel/arch/x86_64/Thread.S

@@ -4,9 +4,12 @@ read_ip:
 	popq %rax
 	jmp *%rax
 
-# void start_thread()
+# void start_kernel_thread()
 .global start_kernel_thread
 start_kernel_thread:
+	call get_start_kernel_thread_sp
+	movq %rax, %rsp
+
 	# STACK LAYOUT
 	#   on_exit arg
 	#   on_exit func
@@ -15,6 +18,7 @@ start_kernel_thread:
 
 	movq 8(%rsp), %rdi
 	movq 0(%rsp), %rsi
+	sti
 	call *%rsi
 
 	movq 24(%rsp), %rdi

kernel/include/kernel/InterruptStack.h

@@ -14,6 +14,7 @@ namespace Kernel
 		uintptr_t ss;
 	};
 
+#if ARCH(x86_64)
 	struct InterruptRegisters
 	{
 		uintptr_t r15;
@@ -33,5 +34,18 @@ namespace Kernel
 		uintptr_t rcx;
 		uintptr_t rax;
 	};
+#elif ARCH(i686)
+	struct InterruptRegisters
+	{
+		uintptr_t edi;
+		uintptr_t esi;
+		uintptr_t ebp;
+		uintptr_t unused;
+		uintptr_t ebx;
+		uintptr_t edx;
+		uintptr_t ecx;
+		uintptr_t eax;
+	};
+#endif
 
 }

kernel/include/kernel/Thread.h

@@ -77,7 +77,6 @@ namespace Kernel
 		size_t virtual_page_count() const { return (m_kernel_stack->size() / PAGE_SIZE) + (m_userspace_stack->size() / PAGE_SIZE); }
 		size_t physical_page_count() const { return virtual_page_count(); }
 
-		uintptr_t& interrupt_sp() { return m_interrupt_sp; }
 		InterruptStack& interrupt_stack() { return m_interrupt_stack; }
 		InterruptRegisters& interrupt_registers() { return m_interrupt_registers; }
 
@@ -89,6 +88,8 @@ namespace Kernel
 
 	private:
 		Thread(pid_t tid, Process*);
+
+		static void on_exit_trampoline(Thread*);
 		void on_exit();
 
 	private:
@@ -104,7 +105,6 @@ namespace Kernel
 
 		InterruptStack				m_interrupt_stack		{ };
 		InterruptRegisters			m_interrupt_registers	{ };
-		uintptr_t					m_interrupt_sp			{ };
 
 		uint64_t					m_signal_pending_mask	{ 0 };
 		uint64_t					m_signal_block_mask		{ 0 };

kernel/kernel/Scheduler.cpp

@@ -14,17 +14,6 @@ namespace Kernel
 	static Scheduler* s_instance = nullptr;
 	static BAN::Atomic<bool> s_started { false };
 
-	ALWAYS_INLINE static void load_temp_stack()
-	{
-#if ARCH(x86_64)
-		asm volatile("movq %0, %%rsp" :: "rm"(Processor::current_stack_top()));
-#elif ARCH(i686)
-		asm volatile("movl %0, %%esp" :: "rm"(Processor::current_stack_top()));
-#else
-		#error
-#endif
-	}
-
 	BAN::ErrorOr<void> Scheduler::initialize()
 	{
 		ASSERT(s_instance == nullptr);
@@ -40,6 +29,11 @@ namespace Kernel
 		return *s_instance;
 	}
 
+	extern "C" uintptr_t get_start_kernel_thread_sp()
+	{
+		return Scheduler::get().current_thread().kernel_stack_top() - 4 * sizeof(uintptr_t);
+	}
+
 	void Scheduler::start()
 	{
 		ASSERT(Processor::get_interrupt_state() == InterruptState::Disabled);
@@ -86,7 +80,6 @@ namespace Kernel
 				if (thread->state() != Thread::State::NotStarted)
 				{
 					thread->interrupt_stack() = Processor::get_interrupt_stack();
-					thread->interrupt_stack().sp = thread->interrupt_sp();
 					thread->interrupt_registers() = Processor::get_interrupt_registers();
 				}
 
@@ -155,15 +148,29 @@ namespace Kernel
 		auto state = Processor::get_interrupt_state();
 		Processor::set_interrupt_state(InterruptState::Disabled);
 
+#if ARCH(x86_64)
 		asm volatile(
-			"movq %%rsp, %[save_sp];"
+			"movq %%rsp, %%rcx;"
 			"movq %[load_sp], %%rsp;"
 			"int %[ipi];"
-			: [save_sp]"=m"(Thread::current().interrupt_sp())
-			: [load_sp]"r"(Processor::current_stack_top()),
+			"movq %%rcx, %%rsp;"
+			:: [load_sp]"r"(Processor::current_stack_top()),
 			   [ipi]"i"(IRQ_VECTOR_BASE + IRQ_IPI)
-			: "memory"
+			:  "memory", "rcx"
 		);
+#elif ARCH(i686)
+		asm volatile(
+			"movl %%esp, %%ecx;"
+			"movl %[load_sp], %%esp;"
+			"int %[ipi];"
+			"movl %%ecx, %%esp;"
+			:: [load_sp]"r"(Processor::current_stack_top()),
+			   [ipi]"i"(IRQ_VECTOR_BASE + IRQ_IPI)
+			:  "memory", "ecx"
+		);
+#else
+		#error
+#endif
 
 		Processor::set_interrupt_state(state);
 	}

kernel/kernel/Thread.cpp

@@ -12,19 +12,15 @@
 namespace Kernel
 {
 
-	extern "C" [[noreturn]] void start_userspace_thread();
 	extern "C" [[noreturn]] void start_kernel_thread();
 
 	extern "C" void signal_trampoline();
 
 	template<typename T>
-	static void write_to_stack(uintptr_t& rsp, const T& value)
+	static void write_to_stack(uintptr_t& rsp, const T& value) requires(sizeof(T) <= sizeof(uintptr_t))
 	{
 		rsp -= sizeof(uintptr_t);
-		if constexpr(sizeof(T) < sizeof(uintptr_t))
 		*(uintptr_t*)rsp = (uintptr_t)value;
-		else
-			memcpy((void*)rsp, (void*)&value, sizeof(uintptr_t));
 	}
 
 	static pid_t s_next_tid = 1;
@@ -50,13 +46,13 @@ namespace Kernel
 		// Initialize stack for returning
 		uintptr_t sp = thread->kernel_stack_top();
 		write_to_stack(sp, thread);
-		write_to_stack(sp, &Thread::on_exit);
+		write_to_stack(sp, &Thread::on_exit_trampoline);
 		write_to_stack(sp, data);
 		write_to_stack(sp, entry);
 
 		thread->m_interrupt_stack.ip = reinterpret_cast<vaddr_t>(start_kernel_thread);
 		thread->m_interrupt_stack.cs = 0x08;
-		thread->m_interrupt_stack.flags = 0x202;
+		thread->m_interrupt_stack.flags = 0x002;
 		thread->m_interrupt_stack.sp = sp;
 		thread->m_interrupt_stack.ss = 0x10;
 
@@ -245,7 +241,7 @@ namespace Kernel
 		PageTable::with_fast_page(process().page_table().physical_address_of(kernel_stack_top() - PAGE_SIZE), [&] {
 			uintptr_t sp = PageTable::fast_page() + PAGE_SIZE;
 			write_to_stack(sp, this);
-			write_to_stack(sp, &Thread::on_exit);
+			write_to_stack(sp, &Thread::on_exit_trampoline);
 			write_to_stack(sp, m_process);
 			write_to_stack(sp, entry);
 		});
@@ -422,6 +418,11 @@ namespace Kernel
 		return {};
 	}
 
+	void Thread::on_exit_trampoline(Thread* thread)
+	{
+		thread->on_exit();
+	}
+
 	void Thread::on_exit()
 	{
 		ASSERT(this == &Thread::current());