[boot][kernel] Split programs into sections

To enable setting sections as NX or read-only, the boot program loader
now loads programs as lists of sections, and the kernel args are updated
accordingly. The kernel's loader now just takes a program pointer to
iterate the sections. Also enable NX in IA32_EFER in the bootloader.

src/boot/hardware.cpp

@@ -37,8 +37,25 @@ find_acpi_table(uefi::system_table *st)
 	return reinterpret_cast<void*>(acpi1_table);
 }
 
+static uint64_t
+rdmsr(uint32_t addr)
+{
+	uint32_t low, high;
+	__asm__ __volatile__ ("rdmsr" : "=a"(low), "=d"(high) : "c"(addr));
+	return (static_cast<uint64_t>(high) << 32) | low;
+}
+
+static void
+wrmsr(uint32_t addr, uint64_t value)
+{
+	uint32_t low = value & 0xffffffff;
+	uint32_t high = value >> 32;
+	__asm__ __volatile__ ("wrmsr" :: "c"(addr), "a"(low), "d"(high));
+}
+
+
 void
-setup_cr4()
+setup_control_regs()
 {
 	uint64_t cr4 = 0;
 	asm volatile ( "mov %%cr4, %0" : "=r" (cr4) );
@@ -49,6 +66,15 @@ setup_cr4()
 		0x020000 | // Enable PCIDs
 		0;
 	asm volatile ( "mov %0, %%cr4" :: "r" (cr4) );
+
+	// Set up IA32_EFER
+	constexpr uint32_t IA32_EFER = 0xC0000080;
+	uint64_t efer = rdmsr(IA32_EFER);
+	efer |=
+		0x0001 | // Enable SYSCALL
+		0x0800 | // Enable NX bit
+		0;
+	wrmsr(IA32_EFER, efer);
 }
 
 } // namespace hw

src/boot/hardware.h

@@ -13,8 +13,8 @@ namespace hw {
 /// significant bit set to 1.
 void * find_acpi_table(uefi::system_table *st);
 
-/// Enable CPU options in CR4 for the kernel starting state.
+/// Enable CPU options in CR4 etc for the kernel starting state.
-void setup_cr4();
+void setup_control_regs();
 
 } // namespace hw
 } // namespace boot

src/boot/loader.cpp

@@ -87,6 +87,9 @@ load_program(
 
 	bs->set_mem(pages, total_size, 0);
 
+	program.base = prog_base;
+	program.total_size = total_size;
+	program.num_sections = 0;
 	for (int i = 0; i < header->ph_num; ++i) {
 		ptrdiff_t offset = header->ph_offset + i * header->ph_entsize;
 		const elf::program_header *pheader =
@@ -95,14 +98,18 @@ load_program(
 		if (pheader->type != elf::PT_LOAD)
 			continue;
 
+		args::program_section &section = program.sections[program.num_sections++];
+
 		void *src_start = offset_ptr<void>(data.data, pheader->offset);
 		void *dest_start = offset_ptr<void>(pages, pheader->vaddr - prog_base);
+
 		bs->copy_mem(dest_start, src_start, pheader->file_size);
+		section.phys_addr = reinterpret_cast<uintptr_t>(dest_start);
+		section.virt_addr = pheader->vaddr;
+		section.size = pheader->mem_size;
+		section.type = static_cast<args::section_flags>(pheader->flags);
 	}
 
-	program.phys_addr = reinterpret_cast<uintptr_t>(pages);
-	program.size = total_size;
-	program.virt_addr = prog_base;
 	program.entrypoint = header->entrypoint;
 }
 

src/boot/main.cpp

@@ -192,16 +192,20 @@ efi_main(uefi::handle image, uefi::system_table *st)
 	status_bar status {con.fb()}; // Switch to fb status display
 
 	args::program &kernel = args->programs[0];
-	paging::map_pages(args, kernel.phys_addr, kernel.virt_addr, kernel.size);
+	for (auto &section : kernel.sections)
+		if (section.size)
+			paging::map_section(args, section);
+
 	kernel::entrypoint kentry =
 		reinterpret_cast<kernel::entrypoint>(kernel.entrypoint);
 	status.next();
 
+
+	hw::setup_control_regs();
 	memory::virtualize(args->pml4, map, st->runtime_services);
 	status.next();
 
 	change_pointer(args->pml4);
-	hw::setup_cr4();
 	status.next();
 
 	kentry(args);

src/boot/paging.cpp

@@ -15,7 +15,7 @@ using memory::page_size;
 using ::memory::pml4e_kernel;
 using ::memory::table_entries;
 
-// Flags: 0 0 0 1  0 0 0 0  0 0 1 1 = 0x0103
+// Flags: 0 0 0 1  0 0 0 0  0 0 0 1 = 0x0101
 //         IGN  |  | | | |  | | | +- Present
 //              |  | | | |  | | +--- Writeable
 //              |  | | | |  | +----- Usermode access (supervisor only)
@@ -26,7 +26,7 @@ using ::memory::table_entries;
 //              |  +---------------- PAT (determining memory type for page)
 //              +------------------- Global
 /// Page table entry flags for entries pointing at a page
-constexpr uint16_t page_flags = 0x103;
+constexpr uint64_t page_flags = 0x101;
 
 // Flags: 0  0 0 0 1  1 0 0 0  1 0 1 1 = 0x018b
 //        |   IGN  |  | | | |  | | | +- Present
@@ -40,7 +40,7 @@ constexpr uint16_t page_flags = 0x103;
 //        |        +------------------- Global
 //        +---------------------------- PAT (determining memory type for page)
 /// Page table entry flags for entries pointing at a huge page
-constexpr uint16_t huge_page_flags = 0x183;
+constexpr uint64_t huge_page_flags = 0x18b;
 
 // Flags: 0 0 0 0  0 0 0 0  0 0 1 1 = 0x0003
 //        IGNORED  | | | |  | | | +- Present
@@ -52,7 +52,7 @@ constexpr uint16_t huge_page_flags = 0x183;
 //                 | +-------------- Ignored
 //                 +---------------- Reserved 0 (Table pointer, not page)
 /// Page table entry flags for entries pointing at another table
-constexpr uint16_t table_flags = 0x003;
+constexpr uint64_t table_flags = 0x003;
 
 /// Iterator over page table entries.
 template <unsigned D = 4>
@@ -191,7 +191,7 @@ allocate_tables(kernel::args::header *args, uefi::boot_services *bs)
 	status_line status(L"Allocating initial page tables");
 
 	static constexpr size_t pd_tables = 256;   // number of pages for kernelspace PDs
-	static constexpr size_t extra_tables = 49; // number of extra pages
+	static constexpr size_t extra_tables = 64; // number of extra pages
 
 	// number of pages for kernelspace PDs + PML4
 	static constexpr size_t kernel_tables = pd_tables + 1;
@@ -223,23 +223,39 @@ allocate_tables(kernel::args::header *args, uefi::boot_services *bs)
 	console::print(L"    Set up initial mappings, %d spare tables.\r\n", args->table_count);
 }
 
+template <typename E>
+constexpr bool has_flag(E set, E flag) {
+	return
+		(static_cast<uint64_t>(set) & static_cast<uint64_t>(flag)) ==
+		 static_cast<uint64_t>(flag);
+}
+
 void
-map_pages(
+map_section(
 	kernel::args::header *args,
-	uintptr_t phys, uintptr_t virt,
+	const kernel::args::program_section &section)
-	size_t size)
 {
 	paging::page_table *pml4 =
 		reinterpret_cast<paging::page_table*>(args->pml4);
 
-	size_t pages = memory::bytes_to_pages(size);
+	size_t pages = memory::bytes_to_pages(section.size);
 	page_entry_iterator<4> iterator{
-		virt, pml4,
+		section.virt_addr, pml4,
 		args->page_tables,
 		args->table_count};
 
+	using kernel::args::section_flags;
+
+	uint64_t flags = page_flags;
+	if (!has_flag(section.type, section_flags::execute))
+		flags |= (1ull << 63); // set NX bit
+
+	if (has_flag(section.type, section_flags::write))
+		flags |= 2;
+
+	uintptr_t phys = section.phys_addr;
 	while (true) {
-		*iterator = phys | page_flags;
+		*iterator = phys | flags;
 		if (--pages == 0)
 			break;
 

src/boot/paging.h

@@ -38,15 +38,14 @@ void allocate_tables(
 /// tables in the current PML4.
 void add_current_mappings(page_table *new_pml4);
 
-/// Map a physical address to a virtual address in the given page tables.
+/// Map a program section in physical memory to its virtual address in the
+/// given page tables.
 /// \arg args    The kernel args header, used for the page table cache and pml4
-/// \arg phys  The phyiscal address to map in
+/// \arg section The program section to load
-/// \arg virt  The virtual address to map in
+void map_section(
-/// \arg size  The size in bytes of the mapping 
-void map_pages(
 	kernel::args::header *args,
-	uintptr_t phys, uintptr_t virt,
+	const kernel::args::program_section &section);
-	size_t bytes);
+
 
 } // namespace paging
 } // namespace boot

src/include/kernel_args.h

@@ -35,11 +35,26 @@ struct framebuffer {
 	fb_type type;
 };
 
-struct program {
+enum class section_flags : uint32_t {
+	none    = 0,
+	execute = 1,
+	write   = 2,
+	read    = 4,
+};
+
+struct program_section {
 	uintptr_t phys_addr;
 	uintptr_t virt_addr;
+	uint32_t size;
+	section_flags type;
+};
+
+struct program {
 	uintptr_t entrypoint;
-	size_t size;
+	uintptr_t base;
+	size_t total_size;
+	size_t num_sections;
+	program_section sections[3];
 };
 
 enum class mem_type : uint32_t {

src/include/kernel_memory.h

@@ -58,6 +58,11 @@ namespace memory {
 		return reinterpret_cast<T*>(a|page_offset);
 	}
 
+	/// Convert a physical address to a virtual one (in the offset-mapped area)
+	template <typename T> T * to_virtual(T *p) {
+		return to_virtual<T>(reinterpret_cast<uintptr_t>(p));
+	}
+
 	/// Get the number of pages needed for a given number of bytes.
 	/// \arg bytes  The number of bytes desired
 	/// \returns    The number of pages needed to contain the desired bytes

src/kernel/loader.s

@@ -8,11 +8,7 @@ preloaded_process_init:
 	; create_process already pushed the arguments for load_process_image and
 	; the following iretq onto the stack for us
 
-	pop rdi ; the physical address of the program image
+	pop rdi ; a pointer to the program descriptor
-	pop rsi ; the virtual address of the program image
-	pop rdx ; the size in bytes of the program image
-	pop rcx ; the address of this thread's TCB
-
 	call load_process_image
 
 	; user rsp is now in rax, put it in the right place for iret

src/kernel/main.cpp

@@ -192,8 +192,8 @@ kernel_main(args::header *header)
 
 	// Skip program 0, which is the kernel itself
 	for (size_t i = 1; i < header->num_programs; ++i) {
-		args::program &prog = header->programs[i];
+		args::program *prog = memory::to_virtual(&header->programs[i]);
-		thread *th = sched->load_process(prog.phys_addr, prog.virt_addr, prog.size, prog.entrypoint); 
+		thread *th = sched->load_process(*prog);
 		if (i == 2) {
 			//th->set_state(thread::state::constant);
 		}

src/kernel/scheduler.cpp

@@ -34,7 +34,7 @@ const uint64_t rflags_int = 0x202;
 
 extern "C" {
 	void preloaded_process_init();
-	uintptr_t load_process_image(uintptr_t phys, uintptr_t virt, size_t bytes, TCB *tcb);
+	uintptr_t load_process_image(const kernel::args::program*);
 };
 
 extern uint64_t idle_stack_end;
@@ -76,20 +76,28 @@ inline T * push(uintptr_t &rsp, size_t size = sizeof(T)) {
 }
 
 uintptr_t
-load_process_image(uintptr_t phys, uintptr_t virt, size_t bytes, TCB *tcb)
+load_process_image(const kernel::args::program *program)
 {
 	using memory::page_align_down;
 	using memory::page_align_up;
+	using kernel::args::section_flags;
 
 	// We're now in the process space for this process, allocate memory for the
 	// process code and load it
 	process &proc = process::current();
 	thread &th = thread::current();
+	TCB *tcb = th.tcb();
 	vm_space &space = proc.space();
 
-	vm_area *vma = new vm_area_open(bytes, space, vm_flags::zero|vm_flags::write);
+	for (const auto &sect : program->sections) {
-	space.add(virt, vma);
+		vm_flags flags =
-	vma->commit(phys, 0, memory::page_count(bytes));
+			(bitfield_has(sect.type, section_flags::execute) ? vm_flags::exec : vm_flags::none) |
+			(bitfield_has(sect.type, section_flags::write) ? vm_flags::write : vm_flags::none);
+
+		vm_area *vma = new vm_area_open(sect.size, space, flags);
+		space.add(sect.virt_addr, vma);
+		vma->commit(sect.phys_addr, 0, memory::page_count(sect.size));
+	}
 
 	// double zero stack sentinel
 	*push<uint64_t>(tcb->rsp3) = 0;
@@ -141,7 +149,8 @@ load_process_image(uintptr_t phys, uintptr_t virt, size_t bytes, TCB *tcb)
 
 	// Crazypants framebuffer part
 	if (fb) {
-		vma = new vm_area_open(fb->size, space, vm_flags::write|vm_flags::mmio|vm_flags::write_combine);
+		vm_area *vma = new vm_area_open(fb->size, space,
+				vm_flags::write|vm_flags::mmio|vm_flags::write_combine);
 		space.add(0x100000000, vma);
 		vma->commit(fb->phys_addr, 0, memory::page_count(fb->size));
 	}
@@ -167,7 +176,7 @@ scheduler::create_process(bool user)
 }
 
 thread *
-scheduler::load_process(uintptr_t phys, uintptr_t virt, size_t size, uintptr_t entry)
+scheduler::load_process(kernel::args::program &program)
 {
 
 	uint16_t kcs = (1 << 3) | 0; // Kernel CS is GDT entry 1, ring 0
@@ -180,23 +189,20 @@ scheduler::load_process(uintptr_t phys, uintptr_t virt, size_t size, uintptr_t e
 	auto *tcb = th->tcb();
 
 	// Create an initial kernel stack space
-	uintptr_t *stack = reinterpret_cast<uintptr_t *>(tcb->rsp0) - 9;
+	uintptr_t *stack = reinterpret_cast<uintptr_t *>(tcb->rsp0) - 6;
 
 	// Pass args to preloaded_process_init on the stack
-	stack[0] = reinterpret_cast<uintptr_t>(phys);
+	stack[0] = reinterpret_cast<uintptr_t>(&program);
-	stack[1] = reinterpret_cast<uintptr_t>(virt);
-	stack[2] = reinterpret_cast<uintptr_t>(size);
-	stack[3] = reinterpret_cast<uintptr_t>(tcb);
 
 	tcb->rsp = reinterpret_cast<uintptr_t>(stack);
 	th->add_thunk_kernel(reinterpret_cast<uintptr_t>(preloaded_process_init));
 
 	// Arguments for iret - rip will be pushed on before these
-	stack[4] = reinterpret_cast<uintptr_t>(entry);
+	stack[1] = reinterpret_cast<uintptr_t>(program.entrypoint);
-	stack[5] = cs;
+	stack[2] = cs;
-	stack[6] = rflags_int | (3 << 12);
+	stack[3] = rflags_int | (3 << 12);
-	stack[7] = process::stacks_top;
+	stack[4] = process::stacks_top;
-	stack[8] = ss;
+	stack[5] = ss;
 
 	tcb->rsp3 = process::stacks_top;
 

src/kernel/scheduler.h

@@ -5,6 +5,11 @@
 #include <stdint.h>
 #include "objects/thread.h"
 
+namespace kernel {
+namespace args {
+	struct program;
+}}
+
 class lapic;
 class process;
 struct page_table;
@@ -41,12 +46,9 @@ public:
 	scheduler(lapic *apic);
 
 	/// Create a new process from a program image in memory.
-	/// \arg phys  Physical address of the loaded program image
+	/// \arg program  The descriptor of the pogram in memory
-	/// \arg virt  Virtual address of the loaded program image
-	/// \arg size  Size of the program image, in bytes
-	/// \arg entry Virtual address of the program entrypoint
 	/// \returns      The main thread of the loaded process
-	thread * load_process(uintptr_t phys, uintptr_t virt, size_t size, uintptr_t entry);
+	thread * load_process(kernel::args::program &program);
 
 	/// Create a new kernel task
 	/// \arg proc     Function to run as a kernel task

src/libraries/kutil/include/kutil/enum_bitfields.h

@@ -131,10 +131,13 @@ operator ! (E rhs)
 }
 
 template <typename E>
-constexpr typename std::enable_if<is_enum_bitfield<E>::value,bool>::type
+constexpr bool
 bitfield_has(E set, E flag)
 {
-	return (set & flag) == flag;
+	return
+		(static_cast<typename std::underlying_type<E>::type>(set) &
+		 static_cast<typename std::underlying_type<E>::type>(flag)) ==
+		 static_cast<typename std::underlying_type<E>::type>(flag);
 }
 
 // Overload the logical-and operator to be 'bitwise-and, bool-cast'