Linux kernel system call table hooking

afw_main.c

#include <asm/uaccess.h>

#include <linux/init.h>
#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/version.h>

#include "locate_sct.h"
#include "ttgl.h"

MODULE_LICENSE("GPL");

MODULE_AUTHOR("ilammy <a.lozovsky@gmail.com>");

MODULE_DESCRIPTION("Locates and patches the system call table to hook execve() "
                   "and install additional customizable restrictions on the "
                   "processes that can be launched from userland.");

#if 0

#if !(defined(ARCH_X86) || defined(ARCH_X64))
#  error "Only x86(_64) kernel architectures are supported"
#endif

#if LINUX_KERNEL_VERSION != KERNEL_VERSION(3,2,0)
#  error "Only Linux kernel 3.2.0 is supported"
#endif

#endif

static
void hex_dump(unsigned char *bytes, size_t count)
{
	size_t i;

	printk(KERN_INFO "Dumping %zu bytes at %p:", count, bytes);
	for (i = 0; i < count; i++)
	{
		if (i % 16 == 0)
		{
			printk("\n    ");
		}
		printk("%02X ", bytes[i]);
	}
	printk("\n");
}

long hijacked_sys_execve(const char __user *filename,
                         const char __user *const __user *argv,
                         const char __user *const __user *envp)
{
	char buffer[256];
	strncpy_from_user(buffer, filename, 256);

	printk(KERN_INFO "hijacked_sys_execve: i see what you did there: %s\n",
		buffer);

	return sys_execve(filename, argv, envp);
}

static
int print_sct_1(struct gl_region regions[], size_t region_count, void* arg)
{
	size_t i;
	unsigned long* sys_call_table = regions[0].writeable;

	for (i = 0; i < 256; i++)
		if ((void*) sys_call_table[i] == (void*) sys_execve)
			sys_call_table[i] = (unsigned long*) hijacked_sys_execve;

	return 0;
}

static
int print_sct_2(struct gl_region regions[], size_t region_count, void* arg)
{
	size_t i;
	unsigned long* sys_call_table = regions[0].writeable;

	for (i = 0; i < 256; i++)
		if ((void*) sys_call_table[i] == (void*) hijacked_sys_execve)
			sys_call_table[i] = (unsigned long*) sys_execve;

	return 0;
}

static
int __init afw_init(void)
{
	struct gl_region sys_call_table;

	printk(KERN_INFO "init_module()\n");

	printk(KERN_INFO "located sys_call_table: %p\n"
	                 "located ia32_sys_call_table: %p\n",
	       afw_locate_sys_call_table(),
	       afw_locate_ia32_sys_call_table());

	sys_call_table = (struct gl_region) {
		.source = afw_locate_sys_call_table(),
		.length = 256 * sizeof(unsigned long)
	};

	afw_do_with_write_permissions(print_sct_1, &sys_call_table, 1, NULL);

	return 0;
}

static
void __exit afw_exit(void)
{
	struct gl_region sys_call_table;

	printk(KERN_INFO "exit_module()\n");

	sys_call_table = (struct gl_region) {
		.source = afw_locate_sys_call_table(),
		.length = 256 * sizeof(unsigned long)
	};

	afw_do_with_write_permissions(print_sct_2, &sys_call_table, 1, NULL);
}

module_init(afw_init);
module_exit(afw_exit);

locate_sct.c

#include "locate_sct.h"

#include <asm/desc.h>
#include <asm/desc_defs.h>
#include <asm/irq_vectors.h>
#include <asm/msr.h>
#include <asm/msr-index.h>

/*
**  We need to find the address of the sys_call_table. Since 2.6 kernels this
**  symbol is no longer exported, so there is no easy way to find it. One way
**  is to use "cat /boot/System.map-`uname -r` | grep sys_call_table". However,
**  it may be absent or invalid, and just reading files is not fun. So we're
**  going the brute-force way. Introspection! The system call handler should
**  bo somewhere in memory and it obviosly uses sys_call_table, so if we
**  analyze its code, we can get the address we want.
**
**  Actually, on 64-bit x86 systems there are two tables: sys_call_table with
**  64-bit handlers and ia32_sys_call_table with 32-bit handlers for 32-bit
**  emulation mode. We need to locate them both.
**
**  Also, Linux has two mechanisms for handling system calls on x86 systems:
**  the legacy one that uses "int $0x80" to get into ring 0, and more modern
**  one which uses special "sysenter"/"syscall" instructions for this. Also
**  there is vsyscall mechanism for fast user-level system calls, but it is
**  not covered here.
**
**  Legacy mechanism uses interrupt traps to obtain control. It is initialized
**  by the function trap_init() which can be found in arch/x86/kernel/traps.c.
**  The interrupt handler for 0x80 always contains a 32-bit handler.
**
**  Modern handler uses model-specific registers (MSRs) to register itself as
**  a handler for "sysenter"/"syscall" instructions. It is initialized by the
**  function syscall_init() from arch/x86/kernel/cpu/common.c. Native handler
**  system_call is loaded to the LSTAR MSR, and IA-32 emulated ones are loaded
**  to IA32_SYSENTER_EIP and CSTAR MSRs by the function syscall32_cpu_init()
**  from arch/x86/vdso/vdso32-setup.c.
**
**  So, to sum it up, to get the 32-bit syscall handler we consult the 0x80
**  interrupt handler in the interrupt descriptor table. To get the 64-bit
**  syscall handler we consult the LSTAR MSR.
*/

static inline
u8* get_32bit_system_call_handler(void)
{
	struct desc_ptr interrupt_descriptor_table;
	gate_desc* interrupt_gates;

	store_idt(&interrupt_descriptor_table);
	interrupt_gates = (gate_desc*) interrupt_descriptor_table.address;

	return (u8*) gate_offset(interrupt_gates[IA32_SYSCALL_VECTOR]);
}

static inline
u8* get_64bit_system_call_handler(void)
{
	u64 system_call_entry;
	rdmsrl(MSR_LSTAR, system_call_entry);
	return (u8*) system_call_entry;
}

/*
**  Previous functions return pointers to system call handlers. Native system
**  call handler is named system_call, its 32-bit implementation is located in
**  arch/x86/kernel/entry_32.S and 64-bit one is in arch/x86/kernel/entry_64.S.
**  IA-32 emulation handlers are implemented in arch/x86/ia32/ia32entry.S, they
**  are named ia32_syscall, ia32_sysenter_target, and ia32_cstar_target.
**
**  We are interested in this:
**
**      call *sys_call_table(,%rax,8)
**      movq %rax,RAX-ARGOFFSET(%rsp)
**
**  or this:
**
**      call *sys_call_table(,%eax,4)
**      movl %eax,PT_EAX(%esp)
**
**  These snippets do the actual system call and store its return value.
**  sys_call_table is 32-bit offset, should be expanded to 64-bit if necessary.
**  RAX, ARGOFFSET, and PT_EAX are all macros that govern calling conventions,
**  they expand into small numbers that fit into one byte.
**
**  According to the source, these instructions should be found within the
**  first 256 bytes of the handlers. We should look for a call instruction
**  followed by a mov instruction. Opcodes are the following:
**
**      call disp32(,%eax,4)   ==>  FF 14 85 -- -- -- --
**      movl %eax,disp8(%esp)  ==>  89 44 24 --
**
**      call disp32(,%rax,8)   ==>  FF 14 C5 -- -- -- --
**      movq %rax,disp8(%rsp)  ==>  48 89 44 24 --
*/

static
unsigned long* find_sys_call_table_ref(u8* code)
{
	size_t i;

	for (i = 0; i < 256; i++)
	{
#ifdef CONFIG_X86_64
		if (code[i + 0] == 0xFF && code[i + 1] == 0x14 &&
		    code[i + 2] == 0xC5 && code[i + 7] == 0x48 &&
		    code[i + 8] == 0x89 && code[i + 9] == 0x44 &&
		    code[i +10] == 0x24)
		{
			u32 offset = *((u32*) &code[i + 3]);
			return (unsigned long*) (0xFFFFFFFF00000000 | offset);
		}
#else
		if (code[i + 0] == 0xFF && code[i + 1] == 0x14 &&
		    code[i + 2] == 0x85 && code[i + 7] == 0x89 &&
		    code[i + 8] == 0x44 && code[i + 9] == 0x24)
		{
			u32 offset = *((u32*) &code[i + 3]);
			return (unsigned long*) offset;
		}
#endif
	}

	return NULL;
}

/*
**  And now, when everything's in place...
*/

unsigned long* afw_locate_sys_call_table(void)
{
#ifdef CONFIG_X86_64
	return find_sys_call_table_ref(get_64bit_system_call_handler());
#else
	return find_sys_call_table_ref(get_32bit_system_call_handler());
#endif
}

#ifdef CONFIG_IA32_EMULATION
unsigned long* afw_locate_ia32_sys_call_table(void)
{
	return find_sys_call_table_ref(get_32bit_system_call_handler());
}
#endif

locate_sct.h

#ifndef AFW_LOCATE_SCT
#define AFW_LOCATE_SCT

/**
 *  Locate the native system call table (sys_call_table).
 *
 *  @return
 *  	Returns a pointer to the native system call table,
 *  	or `NULL` in case of failure.
 */
unsigned long* afw_locate_sys_call_table(void);

#ifdef CONFIG_IA32_EMULATION
/**
 *  Locate the system call table used for IA-32 emulation (ia32_sys_call_table).
 *
 *  @return
 *  	Returns a pointer to the IA-32 emulation system call table,
 *  	or `NULL` in case of failure.
 */
unsigned long* afw_locate_ia32_sys_call_table(void);
#endif

#endif // AFW_LOCATE_SCT

Makefile

obj-m += afw.o
afw-objs := afw_main.o locate_sct.o ttgl.o

ccflags-y := -std=gnu99 -O2

all:
	make -C /lib/modules/$(shell uname -r)/build M=$(PWD) modules

clean:
	make -C /lib/modules/$(shell uname -r)/build M=$(PWD) clean

ttgl.c

#include "ttgl.h"

#include <asm/page.h>

#include <linux/mm.h>
#include <linux/module.h>
#include <linux/slab.h>
#include <linux/stop_machine.h>

/*
**  Having troubles with read-only memory? Fuck this shit, we're in kernel mode!
**  To obtain write permissions for a region of kernel memory we can just remap
**  it somewhere else with necessary permissions. God bless MMU and paging.
*/

#define base_of_page(x) ((void*)((unsigned long)(x) & PAGE_MASK))

static
int enumerate_pages(void* region, struct page *pages[], size_t page_num)
{
	size_t i;
	void* page_addr = base_of_page(region);

	for (i = 0; i < page_num; i++) {
		// explain
		if (__module_address((unsigned long) page_addr)) {
			pages[i] = vmalloc_to_page(page_addr);
		}
		else {
			pages[i] = virt_to_page(page_addr);
			WARN_ON(!PageReserved(pages[i]));
		}

		if (!pages[i])
			return -EFAULT;

		page_addr += PAGE_SIZE;
	}

	return 0;
}

static
void* remap_with_write_permissions(void* region, size_t len)
{
	void* writeable_region;
	size_t page_num = DIV_ROUND_UP(offset_in_page(region) + len, PAGE_SIZE);
	struct page **pages = kmalloc(page_num * sizeof(*pages), GFP_KERNEL);

	if (!pages)
		goto err;

	if (enumerate_pages(region, pages, page_num))
		goto err;

	writeable_region = vmap(pages, page_num, VM_MAP, PAGE_KERNEL);
	if (!writeable_region)
		goto err;

	kfree(pages);
	return writeable_region + offset_in_page(region);

err:
	kfree(pages);
	return NULL;
}

/*
**  One needs to write the unwritable only for monkey-patching the kernel code,
**  so it is wise to forbid anybody else to mess with global memory when we're
**  doing our evil stuff. That's why stop_machine() is used.
**
**  Its interface is not identical to our callback, so we need a thunk to pass
**  all the arguments we want.
*/

struct stop_machine_work {
	int (*fn)(struct gl_region[], size_t, void*);
	struct gl_region *regions;
	size_t region_count;
	void* args;
};

static
int stop_machine_thunk(void* arg)
{
	struct stop_machine_work *work = arg;
	return work->fn(work->regions, work->region_count, work->args);
}

int afw_do_with_write_permissions(int (*fn)(struct gl_region[], size_t, void*),
                                  struct gl_region regions[],
                                  size_t region_count,
                                  void* args)
{
	size_t i;
	int result = 0;
	struct stop_machine_work work;

	if (!fn)
		return -EINVAL;

	if (!regions || region_count == 0)
		return fn(NULL, 0, args);

	for (i = 0; i < region_count; i++) {
		regions[i].writeable =
			remap_with_write_permissions(regions[i].source,
			                             regions[i].length);
		if (!regions[i].writeable) {
			size_t j;
			for (j = 0; j < i; j++)
				vunmap(base_of_page(regions[j].writeable));

			return -ENOMEM;
		}
	}

	work = (struct stop_machine_work) {
		.fn = fn, .regions = regions,
		.region_count = region_count,
		.args = args
	};

	/* Stop the machines, prepare to die! */
	result = stop_machine(stop_machine_thunk, &work, 0); // 0?

	for (i = 0; i < region_count; i++)
		vunmap(base_of_page(regions[i].writeable));

	return result;
}

ttgl.h

#ifndef AFW_TTGL_H
#define AFW_TTGL_H

#include <stddef.h>

/** 
 *  Region descriptor for `afw_do_with_write_permissions()`.
 *
 *  `source` and `length` should be filled by the caller. `writable` is filled
 *  by `afw_do_with_write_permissions()`, it will be valid inside the `fn`
 *  callback (and only there).
 */
struct gl_region {
	void* source;		//!< read-only region of memory
	void* writeable;	//!< writeable mapping of the `source` region
	size_t length;		//!< length of the region in bytes
};

/**
 *  Execute a function with write permissions for specified memory regions.
 *
 *  If `regions == NULL` or `region_count == 0` then the call is equivalent to
 *  `fn(NULL, 0, args)`.
 *
 *  `fn` is executed in atomic context, so please no locks inside it.
 *
 *  @param fn
 *  	The function to execute. Receives `regions` with updated `writeable`
 *  	fields, `region_count`, and `args`. Must be non-NULL.
 *
 *  @param regions
 *  	Array of `gl_region` structures which describe the read-only regions
 *  	that should be made writable for `fn`.
 *
 *  @param region_count
 *  	Size of `regions` (in elements).
 *
 *  @param args
 *  	Additional arguments to `fn`.
 *
 *  @return
 *  	On success returns the value `fn` returned.
 *  	`-EINVAL` if `fn == NULL`.
 *  	`-ENOMEM` if failed to obtain write permissions.
 */
int afw_do_with_write_permissions(int (*fn)(struct gl_region[], size_t, void*),
                                  struct gl_region regions[],
                                  size_t region_count,
                                  void* args);

#endif // AFW_TTGL_H