xref: /openbmc/linux/drivers/misc/lkdtm/bugs.c (revision 4eb5928d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
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
6  * test source files.
7  */
8 #include "lkdtm.h"
9 #include <linux/list.h>
10 #include <linux/sched.h>
11 #include <linux/sched/signal.h>
12 #include <linux/sched/task_stack.h>
13 #include <linux/uaccess.h>
14 #include <linux/slab.h>
15 
16 #if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
17 #include <asm/desc.h>
18 #endif
19 
20 struct lkdtm_list {
21 	struct list_head node;
22 };
23 
24 /*
25  * Make sure our attempts to over run the kernel stack doesn't trigger
26  * a compiler warning when CONFIG_FRAME_WARN is set. Then make sure we
27  * recurse past the end of THREAD_SIZE by default.
28  */
29 #if defined(CONFIG_FRAME_WARN) && (CONFIG_FRAME_WARN > 0)
30 #define REC_STACK_SIZE (_AC(CONFIG_FRAME_WARN, UL) / 2)
31 #else
32 #define REC_STACK_SIZE (THREAD_SIZE / 8)
33 #endif
34 #define REC_NUM_DEFAULT ((THREAD_SIZE / REC_STACK_SIZE) * 2)
35 
36 static int recur_count = REC_NUM_DEFAULT;
37 
38 static DEFINE_SPINLOCK(lock_me_up);
39 
40 /*
41  * Make sure compiler does not optimize this function or stack frame away:
42  * - function marked noinline
43  * - stack variables are marked volatile
44  * - stack variables are written (memset()) and read (pr_info())
45  * - function has external effects (pr_info())
46  * */
47 static int noinline recursive_loop(int remaining)
48 {
49 	volatile char buf[REC_STACK_SIZE];
50 
51 	memset((void *)buf, remaining & 0xFF, sizeof(buf));
52 	pr_info("loop %d/%d ...\n", (int)buf[remaining % sizeof(buf)],
53 		recur_count);
54 	if (!remaining)
55 		return 0;
56 	else
57 		return recursive_loop(remaining - 1);
58 }
59 
60 /* If the depth is negative, use the default, otherwise keep parameter. */
61 void __init lkdtm_bugs_init(int *recur_param)
62 {
63 	if (*recur_param < 0)
64 		*recur_param = recur_count;
65 	else
66 		recur_count = *recur_param;
67 }
68 
69 void lkdtm_PANIC(void)
70 {
71 	panic("dumptest");
72 }
73 
74 void lkdtm_BUG(void)
75 {
76 	BUG();
77 }
78 
79 static int warn_counter;
80 
81 void lkdtm_WARNING(void)
82 {
83 	WARN_ON(++warn_counter);
84 }
85 
86 void lkdtm_WARNING_MESSAGE(void)
87 {
88 	WARN(1, "Warning message trigger count: %d\n", ++warn_counter);
89 }
90 
91 void lkdtm_EXCEPTION(void)
92 {
93 	*((volatile int *) 0) = 0;
94 }
95 
96 void lkdtm_LOOP(void)
97 {
98 	for (;;)
99 		;
100 }
101 
102 void lkdtm_EXHAUST_STACK(void)
103 {
104 	pr_info("Calling function with %lu frame size to depth %d ...\n",
105 		REC_STACK_SIZE, recur_count);
106 	recursive_loop(recur_count);
107 	pr_info("FAIL: survived without exhausting stack?!\n");
108 }
109 
110 static noinline void __lkdtm_CORRUPT_STACK(void *stack)
111 {
112 	memset(stack, '\xff', 64);
113 }
114 
115 /* This should trip the stack canary, not corrupt the return address. */
116 noinline void lkdtm_CORRUPT_STACK(void)
117 {
118 	/* Use default char array length that triggers stack protection. */
119 	char data[8] __aligned(sizeof(void *));
120 
121 	pr_info("Corrupting stack containing char array ...\n");
122 	__lkdtm_CORRUPT_STACK((void *)&data);
123 }
124 
125 /* Same as above but will only get a canary with -fstack-protector-strong */
126 noinline void lkdtm_CORRUPT_STACK_STRONG(void)
127 {
128 	union {
129 		unsigned short shorts[4];
130 		unsigned long *ptr;
131 	} data __aligned(sizeof(void *));
132 
133 	pr_info("Corrupting stack containing union ...\n");
134 	__lkdtm_CORRUPT_STACK((void *)&data);
135 }
136 
137 void lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)
138 {
139 	static u8 data[5] __attribute__((aligned(4))) = {1, 2, 3, 4, 5};
140 	u32 *p;
141 	u32 val = 0x12345678;
142 
143 	p = (u32 *)(data + 1);
144 	if (*p == 0)
145 		val = 0x87654321;
146 	*p = val;
147 }
148 
149 void lkdtm_SOFTLOCKUP(void)
150 {
151 	preempt_disable();
152 	for (;;)
153 		cpu_relax();
154 }
155 
156 void lkdtm_HARDLOCKUP(void)
157 {
158 	local_irq_disable();
159 	for (;;)
160 		cpu_relax();
161 }
162 
163 void lkdtm_SPINLOCKUP(void)
164 {
165 	/* Must be called twice to trigger. */
166 	spin_lock(&lock_me_up);
167 	/* Let sparse know we intended to exit holding the lock. */
168 	__release(&lock_me_up);
169 }
170 
171 void lkdtm_HUNG_TASK(void)
172 {
173 	set_current_state(TASK_UNINTERRUPTIBLE);
174 	schedule();
175 }
176 
177 volatile unsigned int huge = INT_MAX - 2;
178 volatile unsigned int ignored;
179 
180 void lkdtm_OVERFLOW_SIGNED(void)
181 {
182 	int value;
183 
184 	value = huge;
185 	pr_info("Normal signed addition ...\n");
186 	value += 1;
187 	ignored = value;
188 
189 	pr_info("Overflowing signed addition ...\n");
190 	value += 4;
191 	ignored = value;
192 }
193 
194 
195 void lkdtm_OVERFLOW_UNSIGNED(void)
196 {
197 	unsigned int value;
198 
199 	value = huge;
200 	pr_info("Normal unsigned addition ...\n");
201 	value += 1;
202 	ignored = value;
203 
204 	pr_info("Overflowing unsigned addition ...\n");
205 	value += 4;
206 	ignored = value;
207 }
208 
209 /* Intentionally using old-style flex array definition of 1 byte. */
210 struct array_bounds_flex_array {
211 	int one;
212 	int two;
213 	char data[1];
214 };
215 
216 struct array_bounds {
217 	int one;
218 	int two;
219 	char data[8];
220 	int three;
221 };
222 
223 void lkdtm_ARRAY_BOUNDS(void)
224 {
225 	struct array_bounds_flex_array *not_checked;
226 	struct array_bounds *checked;
227 	volatile int i;
228 
229 	not_checked = kmalloc(sizeof(*not_checked) * 2, GFP_KERNEL);
230 	checked = kmalloc(sizeof(*checked) * 2, GFP_KERNEL);
231 
232 	pr_info("Array access within bounds ...\n");
233 	/* For both, touch all bytes in the actual member size. */
234 	for (i = 0; i < sizeof(checked->data); i++)
235 		checked->data[i] = 'A';
236 	/*
237 	 * For the uninstrumented flex array member, also touch 1 byte
238 	 * beyond to verify it is correctly uninstrumented.
239 	 */
240 	for (i = 0; i < sizeof(not_checked->data) + 1; i++)
241 		not_checked->data[i] = 'A';
242 
243 	pr_info("Array access beyond bounds ...\n");
244 	for (i = 0; i < sizeof(checked->data) + 1; i++)
245 		checked->data[i] = 'B';
246 
247 	kfree(not_checked);
248 	kfree(checked);
249 	pr_err("FAIL: survived array bounds overflow!\n");
250 }
251 
252 void lkdtm_CORRUPT_LIST_ADD(void)
253 {
254 	/*
255 	 * Initially, an empty list via LIST_HEAD:
256 	 *	test_head.next = &test_head
257 	 *	test_head.prev = &test_head
258 	 */
259 	LIST_HEAD(test_head);
260 	struct lkdtm_list good, bad;
261 	void *target[2] = { };
262 	void *redirection = &target;
263 
264 	pr_info("attempting good list addition\n");
265 
266 	/*
267 	 * Adding to the list performs these actions:
268 	 *	test_head.next->prev = &good.node
269 	 *	good.node.next = test_head.next
270 	 *	good.node.prev = test_head
271 	 *	test_head.next = good.node
272 	 */
273 	list_add(&good.node, &test_head);
274 
275 	pr_info("attempting corrupted list addition\n");
276 	/*
277 	 * In simulating this "write what where" primitive, the "what" is
278 	 * the address of &bad.node, and the "where" is the address held
279 	 * by "redirection".
280 	 */
281 	test_head.next = redirection;
282 	list_add(&bad.node, &test_head);
283 
284 	if (target[0] == NULL && target[1] == NULL)
285 		pr_err("Overwrite did not happen, but no BUG?!\n");
286 	else
287 		pr_err("list_add() corruption not detected!\n");
288 }
289 
290 void lkdtm_CORRUPT_LIST_DEL(void)
291 {
292 	LIST_HEAD(test_head);
293 	struct lkdtm_list item;
294 	void *target[2] = { };
295 	void *redirection = &target;
296 
297 	list_add(&item.node, &test_head);
298 
299 	pr_info("attempting good list removal\n");
300 	list_del(&item.node);
301 
302 	pr_info("attempting corrupted list removal\n");
303 	list_add(&item.node, &test_head);
304 
305 	/* As with the list_add() test above, this corrupts "next". */
306 	item.node.next = redirection;
307 	list_del(&item.node);
308 
309 	if (target[0] == NULL && target[1] == NULL)
310 		pr_err("Overwrite did not happen, but no BUG?!\n");
311 	else
312 		pr_err("list_del() corruption not detected!\n");
313 }
314 
315 /* Test if unbalanced set_fs(KERNEL_DS)/set_fs(USER_DS) check exists. */
316 void lkdtm_CORRUPT_USER_DS(void)
317 {
318 	pr_info("setting bad task size limit\n");
319 	set_fs(KERNEL_DS);
320 
321 	/* Make sure we do not keep running with a KERNEL_DS! */
322 	force_sig(SIGKILL);
323 }
324 
325 /* Test that VMAP_STACK is actually allocating with a leading guard page */
326 void lkdtm_STACK_GUARD_PAGE_LEADING(void)
327 {
328 	const unsigned char *stack = task_stack_page(current);
329 	const unsigned char *ptr = stack - 1;
330 	volatile unsigned char byte;
331 
332 	pr_info("attempting bad read from page below current stack\n");
333 
334 	byte = *ptr;
335 
336 	pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
337 }
338 
339 /* Test that VMAP_STACK is actually allocating with a trailing guard page */
340 void lkdtm_STACK_GUARD_PAGE_TRAILING(void)
341 {
342 	const unsigned char *stack = task_stack_page(current);
343 	const unsigned char *ptr = stack + THREAD_SIZE;
344 	volatile unsigned char byte;
345 
346 	pr_info("attempting bad read from page above current stack\n");
347 
348 	byte = *ptr;
349 
350 	pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
351 }
352 
353 void lkdtm_UNSET_SMEP(void)
354 {
355 #if IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_UML)
356 #define MOV_CR4_DEPTH	64
357 	void (*direct_write_cr4)(unsigned long val);
358 	unsigned char *insn;
359 	unsigned long cr4;
360 	int i;
361 
362 	cr4 = native_read_cr4();
363 
364 	if ((cr4 & X86_CR4_SMEP) != X86_CR4_SMEP) {
365 		pr_err("FAIL: SMEP not in use\n");
366 		return;
367 	}
368 	cr4 &= ~(X86_CR4_SMEP);
369 
370 	pr_info("trying to clear SMEP normally\n");
371 	native_write_cr4(cr4);
372 	if (cr4 == native_read_cr4()) {
373 		pr_err("FAIL: pinning SMEP failed!\n");
374 		cr4 |= X86_CR4_SMEP;
375 		pr_info("restoring SMEP\n");
376 		native_write_cr4(cr4);
377 		return;
378 	}
379 	pr_info("ok: SMEP did not get cleared\n");
380 
381 	/*
382 	 * To test the post-write pinning verification we need to call
383 	 * directly into the middle of native_write_cr4() where the
384 	 * cr4 write happens, skipping any pinning. This searches for
385 	 * the cr4 writing instruction.
386 	 */
387 	insn = (unsigned char *)native_write_cr4;
388 	for (i = 0; i < MOV_CR4_DEPTH; i++) {
389 		/* mov %rdi, %cr4 */
390 		if (insn[i] == 0x0f && insn[i+1] == 0x22 && insn[i+2] == 0xe7)
391 			break;
392 		/* mov %rdi,%rax; mov %rax, %cr4 */
393 		if (insn[i]   == 0x48 && insn[i+1] == 0x89 &&
394 		    insn[i+2] == 0xf8 && insn[i+3] == 0x0f &&
395 		    insn[i+4] == 0x22 && insn[i+5] == 0xe0)
396 			break;
397 	}
398 	if (i >= MOV_CR4_DEPTH) {
399 		pr_info("ok: cannot locate cr4 writing call gadget\n");
400 		return;
401 	}
402 	direct_write_cr4 = (void *)(insn + i);
403 
404 	pr_info("trying to clear SMEP with call gadget\n");
405 	direct_write_cr4(cr4);
406 	if (native_read_cr4() & X86_CR4_SMEP) {
407 		pr_info("ok: SMEP removal was reverted\n");
408 	} else {
409 		pr_err("FAIL: cleared SMEP not detected!\n");
410 		cr4 |= X86_CR4_SMEP;
411 		pr_info("restoring SMEP\n");
412 		native_write_cr4(cr4);
413 	}
414 #else
415 	pr_err("XFAIL: this test is x86_64-only\n");
416 #endif
417 }
418 
419 void lkdtm_DOUBLE_FAULT(void)
420 {
421 #if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
422 	/*
423 	 * Trigger #DF by setting the stack limit to zero.  This clobbers
424 	 * a GDT TLS slot, which is okay because the current task will die
425 	 * anyway due to the double fault.
426 	 */
427 	struct desc_struct d = {
428 		.type = 3,	/* expand-up, writable, accessed data */
429 		.p = 1,		/* present */
430 		.d = 1,		/* 32-bit */
431 		.g = 0,		/* limit in bytes */
432 		.s = 1,		/* not system */
433 	};
434 
435 	local_irq_disable();
436 	write_gdt_entry(get_cpu_gdt_rw(smp_processor_id()),
437 			GDT_ENTRY_TLS_MIN, &d, DESCTYPE_S);
438 
439 	/*
440 	 * Put our zero-limit segment in SS and then trigger a fault.  The
441 	 * 4-byte access to (%esp) will fault with #SS, and the attempt to
442 	 * deliver the fault will recursively cause #SS and result in #DF.
443 	 * This whole process happens while NMIs and MCEs are blocked by the
444 	 * MOV SS window.  This is nice because an NMI with an invalid SS
445 	 * would also double-fault, resulting in the NMI or MCE being lost.
446 	 */
447 	asm volatile ("movw %0, %%ss; addl $0, (%%esp)" ::
448 		      "r" ((unsigned short)(GDT_ENTRY_TLS_MIN << 3)));
449 
450 	pr_err("FAIL: tried to double fault but didn't die\n");
451 #else
452 	pr_err("XFAIL: this test is ia32-only\n");
453 #endif
454 }
455 
456 #ifdef CONFIG_ARM64
457 static noinline void change_pac_parameters(void)
458 {
459 	if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH)) {
460 		/* Reset the keys of current task */
461 		ptrauth_thread_init_kernel(current);
462 		ptrauth_thread_switch_kernel(current);
463 	}
464 }
465 #endif
466 
467 noinline void lkdtm_CORRUPT_PAC(void)
468 {
469 #ifdef CONFIG_ARM64
470 #define CORRUPT_PAC_ITERATE	10
471 	int i;
472 
473 	if (!IS_ENABLED(CONFIG_ARM64_PTR_AUTH))
474 		pr_err("FAIL: kernel not built with CONFIG_ARM64_PTR_AUTH\n");
475 
476 	if (!system_supports_address_auth()) {
477 		pr_err("FAIL: CPU lacks pointer authentication feature\n");
478 		return;
479 	}
480 
481 	pr_info("changing PAC parameters to force function return failure...\n");
482 	/*
483 	 * PAC is a hash value computed from input keys, return address and
484 	 * stack pointer. As pac has fewer bits so there is a chance of
485 	 * collision, so iterate few times to reduce the collision probability.
486 	 */
487 	for (i = 0; i < CORRUPT_PAC_ITERATE; i++)
488 		change_pac_parameters();
489 
490 	pr_err("FAIL: survived PAC changes! Kernel may be unstable from here\n");
491 #else
492 	pr_err("XFAIL: this test is arm64-only\n");
493 #endif
494 }
495