1 // SPDX-License-Identifier: GPL-2.0
3 * This is for all the tests related to logic bugs (e.g. bad dereferences,
4 * bad alignment, bad loops, bad locking, bad scheduling, deep stacks, and
5 * lockups) along with other things that don't fit well into existing LKDTM
10 #include <linux/list.h>
11 #include <linux/sched.h>
12 #include <linux/sched/signal.h>
13 #include <linux/sched/task_stack.h>
14 #include <linux/slab.h>
15 #include <linux/stop_machine.h>
16 #include <linux/uaccess.h>
18 #if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
23 struct list_head node;
27 * Make sure our attempts to over run the kernel stack doesn't trigger
28 * a compiler warning when CONFIG_FRAME_WARN is set. Then make sure we
29 * recurse past the end of THREAD_SIZE by default.
31 #if defined(CONFIG_FRAME_WARN) && (CONFIG_FRAME_WARN > 0)
32 #define REC_STACK_SIZE (_AC(CONFIG_FRAME_WARN, UL) / 2)
34 #define REC_STACK_SIZE (THREAD_SIZE / 8UL)
36 #define REC_NUM_DEFAULT ((THREAD_SIZE / REC_STACK_SIZE) * 2)
38 static int recur_count = REC_NUM_DEFAULT;
40 static DEFINE_SPINLOCK(lock_me_up);
43 * Make sure compiler does not optimize this function or stack frame away:
44 * - function marked noinline
45 * - stack variables are marked volatile
46 * - stack variables are written (memset()) and read (buf[..] passed as arg)
47 * - function may have external effects (memzero_explicit())
48 * - no tail recursion possible
50 static int noinline recursive_loop(int remaining)
52 volatile char buf[REC_STACK_SIZE];
55 memset((void *)buf, remaining & 0xFF, sizeof(buf));
59 ret = recursive_loop((int)buf[remaining % sizeof(buf)] - 1);
60 memzero_explicit((void *)buf, sizeof(buf));
64 /* If the depth is negative, use the default, otherwise keep parameter. */
65 void __init lkdtm_bugs_init(int *recur_param)
68 *recur_param = recur_count;
70 recur_count = *recur_param;
73 static void lkdtm_PANIC(void)
78 static int panic_stop_irqoff_fn(void *arg)
83 * As stop_machine() disables interrupts, all CPUs within this function
84 * have interrupts disabled and cannot take a regular IPI.
86 * The last CPU which enters here will trigger a panic, and as all CPUs
87 * cannot take a regular IPI, we'll only be able to stop secondaries if
88 * smp_send_stop() or crash_smp_send_stop() uses an NMI.
90 if (atomic_inc_return(v) == num_online_cpus())
91 panic("panic stop irqoff test");
97 static void lkdtm_PANIC_STOP_IRQOFF(void)
99 atomic_t v = ATOMIC_INIT(0);
100 stop_machine(panic_stop_irqoff_fn, &v, cpu_online_mask);
103 static void lkdtm_BUG(void)
108 static int warn_counter;
110 static void lkdtm_WARNING(void)
112 WARN_ON(++warn_counter);
115 static void lkdtm_WARNING_MESSAGE(void)
117 WARN(1, "Warning message trigger count: %d\n", ++warn_counter);
120 static void lkdtm_EXCEPTION(void)
122 *((volatile int *) 0) = 0;
125 static void lkdtm_LOOP(void)
131 static void lkdtm_EXHAUST_STACK(void)
133 pr_info("Calling function with %lu frame size to depth %d ...\n",
134 REC_STACK_SIZE, recur_count);
135 recursive_loop(recur_count);
136 pr_info("FAIL: survived without exhausting stack?!\n");
139 static noinline void __lkdtm_CORRUPT_STACK(void *stack)
141 memset(stack, '\xff', 64);
144 /* This should trip the stack canary, not corrupt the return address. */
145 static noinline void lkdtm_CORRUPT_STACK(void)
147 /* Use default char array length that triggers stack protection. */
148 char data[8] __aligned(sizeof(void *));
150 pr_info("Corrupting stack containing char array ...\n");
151 __lkdtm_CORRUPT_STACK((void *)&data);
154 /* Same as above but will only get a canary with -fstack-protector-strong */
155 static noinline void lkdtm_CORRUPT_STACK_STRONG(void)
158 unsigned short shorts[4];
160 } data __aligned(sizeof(void *));
162 pr_info("Corrupting stack containing union ...\n");
163 __lkdtm_CORRUPT_STACK((void *)&data);
166 static pid_t stack_pid;
167 static unsigned long stack_addr;
169 static void lkdtm_REPORT_STACK(void)
171 volatile uintptr_t magic;
172 pid_t pid = task_pid_nr(current);
174 if (pid != stack_pid) {
175 pr_info("Starting stack offset tracking for pid %d\n", pid);
177 stack_addr = (uintptr_t)&magic;
180 pr_info("Stack offset: %d\n", (int)(stack_addr - (uintptr_t)&magic));
183 static pid_t stack_canary_pid;
184 static unsigned long stack_canary;
185 static unsigned long stack_canary_offset;
187 static noinline void __lkdtm_REPORT_STACK_CANARY(void *stack)
190 pid_t pid = task_pid_nr(current);
191 unsigned long *canary = (unsigned long *)stack;
192 unsigned long current_offset = 0, init_offset = 0;
194 /* Do our best to find the canary in a 16 word window ... */
195 for (i = 1; i < 16; i++) {
196 canary = (unsigned long *)stack + i;
197 #ifdef CONFIG_STACKPROTECTOR
198 if (*canary == current->stack_canary)
200 if (*canary == init_task.stack_canary)
205 if (current_offset == 0) {
207 * If the canary doesn't match what's in the task_struct,
208 * we're either using a global canary or the stack frame
211 if (init_offset != 0) {
212 pr_err("FAIL: global stack canary found at offset %ld (canary for pid %d matches init_task's)!\n",
215 pr_warn("FAIL: did not correctly locate stack canary :(\n");
216 pr_expected_config(CONFIG_STACKPROTECTOR);
220 } else if (init_offset != 0) {
221 pr_warn("WARNING: found both current and init_task canaries nearby?!\n");
224 canary = (unsigned long *)stack + current_offset;
225 if (stack_canary_pid == 0) {
226 stack_canary = *canary;
227 stack_canary_pid = pid;
228 stack_canary_offset = current_offset;
229 pr_info("Recorded stack canary for pid %d at offset %ld\n",
230 stack_canary_pid, stack_canary_offset);
231 } else if (pid == stack_canary_pid) {
232 pr_warn("ERROR: saw pid %d again -- please use a new pid\n", pid);
234 if (current_offset != stack_canary_offset) {
235 pr_warn("ERROR: canary offset changed from %ld to %ld!?\n",
236 stack_canary_offset, current_offset);
240 if (*canary == stack_canary) {
241 pr_warn("FAIL: canary identical for pid %d and pid %d at offset %ld!\n",
242 stack_canary_pid, pid, current_offset);
244 pr_info("ok: stack canaries differ between pid %d and pid %d at offset %ld.\n",
245 stack_canary_pid, pid, current_offset);
246 /* Reset the test. */
247 stack_canary_pid = 0;
252 static void lkdtm_REPORT_STACK_CANARY(void)
254 /* Use default char array length that triggers stack protection. */
255 char data[8] __aligned(sizeof(void *)) = { };
257 __lkdtm_REPORT_STACK_CANARY((void *)&data);
260 static void lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)
262 static u8 data[5] __attribute__((aligned(4))) = {1, 2, 3, 4, 5};
264 u32 val = 0x12345678;
266 p = (u32 *)(data + 1);
271 if (IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
272 pr_err("XFAIL: arch has CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS\n");
275 static void lkdtm_SOFTLOCKUP(void)
282 static void lkdtm_HARDLOCKUP(void)
289 static void __lkdtm_SMP_CALL_LOCKUP(void *unused)
295 static void lkdtm_SMP_CALL_LOCKUP(void)
297 unsigned int cpu, target;
302 target = cpumask_any_but(cpu_online_mask, cpu);
304 if (target >= nr_cpu_ids) {
305 pr_err("FAIL: no other online CPUs\n");
309 smp_call_function_single(target, __lkdtm_SMP_CALL_LOCKUP, NULL, 1);
311 pr_err("FAIL: did not hang\n");
318 static void lkdtm_SPINLOCKUP(void)
320 /* Must be called twice to trigger. */
321 spin_lock(&lock_me_up);
322 /* Let sparse know we intended to exit holding the lock. */
323 __release(&lock_me_up);
326 static void __noreturn lkdtm_HUNG_TASK(void)
328 set_current_state(TASK_UNINTERRUPTIBLE);
333 static volatile unsigned int huge = INT_MAX - 2;
334 static volatile unsigned int ignored;
336 static void lkdtm_OVERFLOW_SIGNED(void)
341 pr_info("Normal signed addition ...\n");
345 pr_info("Overflowing signed addition ...\n");
351 static void lkdtm_OVERFLOW_UNSIGNED(void)
356 pr_info("Normal unsigned addition ...\n");
360 pr_info("Overflowing unsigned addition ...\n");
365 /* Intentionally using unannotated flex array definition. */
366 struct array_bounds_flex_array {
372 struct array_bounds {
379 static void lkdtm_ARRAY_BOUNDS(void)
381 struct array_bounds_flex_array *not_checked;
382 struct array_bounds *checked;
385 not_checked = kmalloc(sizeof(*not_checked) * 2, GFP_KERNEL);
386 checked = kmalloc(sizeof(*checked) * 2, GFP_KERNEL);
387 if (!not_checked || !checked) {
393 pr_info("Array access within bounds ...\n");
394 /* For both, touch all bytes in the actual member size. */
395 for (i = 0; i < sizeof(checked->data); i++)
396 checked->data[i] = 'A';
398 * For the uninstrumented flex array member, also touch 1 byte
399 * beyond to verify it is correctly uninstrumented.
401 for (i = 0; i < 2; i++)
402 not_checked->data[i] = 'A';
404 pr_info("Array access beyond bounds ...\n");
405 for (i = 0; i < sizeof(checked->data) + 1; i++)
406 checked->data[i] = 'B';
410 pr_err("FAIL: survived array bounds overflow!\n");
411 if (IS_ENABLED(CONFIG_UBSAN_BOUNDS))
412 pr_expected_config(CONFIG_UBSAN_TRAP);
414 pr_expected_config(CONFIG_UBSAN_BOUNDS);
417 struct lkdtm_annotated {
420 int array[] __counted_by(count);
423 static volatile int fam_count = 4;
425 static void lkdtm_FAM_BOUNDS(void)
427 struct lkdtm_annotated *inst;
429 inst = kzalloc(struct_size(inst, array, fam_count + 1), GFP_KERNEL);
431 pr_err("FAIL: could not allocate test struct!\n");
435 inst->count = fam_count;
436 pr_info("Array access within bounds ...\n");
437 inst->array[1] = fam_count;
438 ignored = inst->array[1];
440 pr_info("Array access beyond bounds ...\n");
441 inst->array[fam_count] = fam_count;
442 ignored = inst->array[fam_count];
446 pr_err("FAIL: survived access of invalid flexible array member index!\n");
448 if (!IS_ENABLED(CONFIG_CC_HAS_COUNTED_BY))
449 pr_warn("This is expected since this %s was built with a compiler that does not support __counted_by\n",
451 else if (IS_ENABLED(CONFIG_UBSAN_BOUNDS))
452 pr_expected_config(CONFIG_UBSAN_TRAP);
454 pr_expected_config(CONFIG_UBSAN_BOUNDS);
457 static void lkdtm_CORRUPT_LIST_ADD(void)
460 * Initially, an empty list via LIST_HEAD:
461 * test_head.next = &test_head
462 * test_head.prev = &test_head
464 LIST_HEAD(test_head);
465 struct lkdtm_list good, bad;
466 void *target[2] = { };
467 void *redirection = ⌖
469 pr_info("attempting good list addition\n");
472 * Adding to the list performs these actions:
473 * test_head.next->prev = &good.node
474 * good.node.next = test_head.next
475 * good.node.prev = test_head
476 * test_head.next = good.node
478 list_add(&good.node, &test_head);
480 pr_info("attempting corrupted list addition\n");
482 * In simulating this "write what where" primitive, the "what" is
483 * the address of &bad.node, and the "where" is the address held
486 test_head.next = redirection;
487 list_add(&bad.node, &test_head);
489 if (target[0] == NULL && target[1] == NULL)
490 pr_err("Overwrite did not happen, but no BUG?!\n");
492 pr_err("list_add() corruption not detected!\n");
493 pr_expected_config(CONFIG_LIST_HARDENED);
497 static void lkdtm_CORRUPT_LIST_DEL(void)
499 LIST_HEAD(test_head);
500 struct lkdtm_list item;
501 void *target[2] = { };
502 void *redirection = ⌖
504 list_add(&item.node, &test_head);
506 pr_info("attempting good list removal\n");
507 list_del(&item.node);
509 pr_info("attempting corrupted list removal\n");
510 list_add(&item.node, &test_head);
512 /* As with the list_add() test above, this corrupts "next". */
513 item.node.next = redirection;
514 list_del(&item.node);
516 if (target[0] == NULL && target[1] == NULL)
517 pr_err("Overwrite did not happen, but no BUG?!\n");
519 pr_err("list_del() corruption not detected!\n");
520 pr_expected_config(CONFIG_LIST_HARDENED);
524 /* Test that VMAP_STACK is actually allocating with a leading guard page */
525 static void lkdtm_STACK_GUARD_PAGE_LEADING(void)
527 const unsigned char *stack = task_stack_page(current);
528 const unsigned char *ptr = stack - 1;
529 volatile unsigned char byte;
531 pr_info("attempting bad read from page below current stack\n");
535 pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
538 /* Test that VMAP_STACK is actually allocating with a trailing guard page */
539 static void lkdtm_STACK_GUARD_PAGE_TRAILING(void)
541 const unsigned char *stack = task_stack_page(current);
542 const unsigned char *ptr = stack + THREAD_SIZE;
543 volatile unsigned char byte;
545 pr_info("attempting bad read from page above current stack\n");
549 pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
552 static void lkdtm_UNSET_SMEP(void)
554 #if IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_UML)
555 #define MOV_CR4_DEPTH 64
556 void (*direct_write_cr4)(unsigned long val);
561 cr4 = native_read_cr4();
563 if ((cr4 & X86_CR4_SMEP) != X86_CR4_SMEP) {
564 pr_err("FAIL: SMEP not in use\n");
567 cr4 &= ~(X86_CR4_SMEP);
569 pr_info("trying to clear SMEP normally\n");
570 native_write_cr4(cr4);
571 if (cr4 == native_read_cr4()) {
572 pr_err("FAIL: pinning SMEP failed!\n");
574 pr_info("restoring SMEP\n");
575 native_write_cr4(cr4);
578 pr_info("ok: SMEP did not get cleared\n");
581 * To test the post-write pinning verification we need to call
582 * directly into the middle of native_write_cr4() where the
583 * cr4 write happens, skipping any pinning. This searches for
584 * the cr4 writing instruction.
586 insn = (unsigned char *)native_write_cr4;
587 OPTIMIZER_HIDE_VAR(insn);
588 for (i = 0; i < MOV_CR4_DEPTH; i++) {
590 if (insn[i] == 0x0f && insn[i+1] == 0x22 && insn[i+2] == 0xe7)
592 /* mov %rdi,%rax; mov %rax, %cr4 */
593 if (insn[i] == 0x48 && insn[i+1] == 0x89 &&
594 insn[i+2] == 0xf8 && insn[i+3] == 0x0f &&
595 insn[i+4] == 0x22 && insn[i+5] == 0xe0)
598 if (i >= MOV_CR4_DEPTH) {
599 pr_info("ok: cannot locate cr4 writing call gadget\n");
602 direct_write_cr4 = (void *)(insn + i);
604 pr_info("trying to clear SMEP with call gadget\n");
605 direct_write_cr4(cr4);
606 if (native_read_cr4() & X86_CR4_SMEP) {
607 pr_info("ok: SMEP removal was reverted\n");
609 pr_err("FAIL: cleared SMEP not detected!\n");
611 pr_info("restoring SMEP\n");
612 native_write_cr4(cr4);
615 pr_err("XFAIL: this test is x86_64-only\n");
619 static void lkdtm_DOUBLE_FAULT(void)
621 #if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
623 * Trigger #DF by setting the stack limit to zero. This clobbers
624 * a GDT TLS slot, which is okay because the current task will die
625 * anyway due to the double fault.
627 struct desc_struct d = {
628 .type = 3, /* expand-up, writable, accessed data */
629 .p = 1, /* present */
631 .g = 0, /* limit in bytes */
632 .s = 1, /* not system */
636 write_gdt_entry(get_cpu_gdt_rw(smp_processor_id()),
637 GDT_ENTRY_TLS_MIN, &d, DESCTYPE_S);
640 * Put our zero-limit segment in SS and then trigger a fault. The
641 * 4-byte access to (%esp) will fault with #SS, and the attempt to
642 * deliver the fault will recursively cause #SS and result in #DF.
643 * This whole process happens while NMIs and MCEs are blocked by the
644 * MOV SS window. This is nice because an NMI with an invalid SS
645 * would also double-fault, resulting in the NMI or MCE being lost.
647 asm volatile ("movw %0, %%ss; addl $0, (%%esp)" ::
648 "r" ((unsigned short)(GDT_ENTRY_TLS_MIN << 3)));
650 pr_err("FAIL: tried to double fault but didn't die\n");
652 pr_err("XFAIL: this test is ia32-only\n");
657 static noinline void change_pac_parameters(void)
659 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL)) {
660 /* Reset the keys of current task */
661 ptrauth_thread_init_kernel(current);
662 ptrauth_thread_switch_kernel(current);
667 static noinline void lkdtm_CORRUPT_PAC(void)
670 #define CORRUPT_PAC_ITERATE 10
673 if (!IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
674 pr_err("FAIL: kernel not built with CONFIG_ARM64_PTR_AUTH_KERNEL\n");
676 if (!system_supports_address_auth()) {
677 pr_err("FAIL: CPU lacks pointer authentication feature\n");
681 pr_info("changing PAC parameters to force function return failure...\n");
683 * PAC is a hash value computed from input keys, return address and
684 * stack pointer. As pac has fewer bits so there is a chance of
685 * collision, so iterate few times to reduce the collision probability.
687 for (i = 0; i < CORRUPT_PAC_ITERATE; i++)
688 change_pac_parameters();
690 pr_err("FAIL: survived PAC changes! Kernel may be unstable from here\n");
692 pr_err("XFAIL: this test is arm64-only\n");
696 static struct crashtype crashtypes[] = {
698 CRASHTYPE(PANIC_STOP_IRQOFF),
701 CRASHTYPE(WARNING_MESSAGE),
702 CRASHTYPE(EXCEPTION),
704 CRASHTYPE(EXHAUST_STACK),
705 CRASHTYPE(CORRUPT_STACK),
706 CRASHTYPE(CORRUPT_STACK_STRONG),
707 CRASHTYPE(REPORT_STACK),
708 CRASHTYPE(REPORT_STACK_CANARY),
709 CRASHTYPE(UNALIGNED_LOAD_STORE_WRITE),
710 CRASHTYPE(SOFTLOCKUP),
711 CRASHTYPE(HARDLOCKUP),
712 CRASHTYPE(SMP_CALL_LOCKUP),
713 CRASHTYPE(SPINLOCKUP),
714 CRASHTYPE(HUNG_TASK),
715 CRASHTYPE(OVERFLOW_SIGNED),
716 CRASHTYPE(OVERFLOW_UNSIGNED),
717 CRASHTYPE(ARRAY_BOUNDS),
718 CRASHTYPE(FAM_BOUNDS),
719 CRASHTYPE(CORRUPT_LIST_ADD),
720 CRASHTYPE(CORRUPT_LIST_DEL),
721 CRASHTYPE(STACK_GUARD_PAGE_LEADING),
722 CRASHTYPE(STACK_GUARD_PAGE_TRAILING),
723 CRASHTYPE(UNSET_SMEP),
724 CRASHTYPE(DOUBLE_FAULT),
725 CRASHTYPE(CORRUPT_PAC),
728 struct crashtype_category bugs_crashtypes = {
729 .crashtypes = crashtypes,
730 .len = ARRAY_SIZE(crashtypes),