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
20struct 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
36static int recur_count = REC_NUM_DEFAULT;
37
38static 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 * */
47static 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. */
61void __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
69void lkdtm_PANIC(void)
70{
71	panic("dumptest");
72}
73
74void lkdtm_BUG(void)
75{
76	BUG();
77}
78
79static int warn_counter;
80
81void lkdtm_WARNING(void)
82{
83	WARN_ON(++warn_counter);
84}
85
86void lkdtm_WARNING_MESSAGE(void)
87{
88	WARN(1, "Warning message trigger count: %d\n", ++warn_counter);
89}
90
91void lkdtm_EXCEPTION(void)
92{
93	*((volatile int *) 0) = 0;
94}
95
96void lkdtm_LOOP(void)
97{
98	for (;;)
99		;
100}
101
102void 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
110static 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. */
116noinline 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 */
126noinline 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
137void 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	if (IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
149		pr_err("XFAIL: arch has CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS\n");
150}
151
152void lkdtm_SOFTLOCKUP(void)
153{
154	preempt_disable();
155	for (;;)
156		cpu_relax();
157}
158
159void lkdtm_HARDLOCKUP(void)
160{
161	local_irq_disable();
162	for (;;)
163		cpu_relax();
164}
165
166void lkdtm_SPINLOCKUP(void)
167{
168	/* Must be called twice to trigger. */
169	spin_lock(&lock_me_up);
170	/* Let sparse know we intended to exit holding the lock. */
171	__release(&lock_me_up);
172}
173
174void lkdtm_HUNG_TASK(void)
175{
176	set_current_state(TASK_UNINTERRUPTIBLE);
177	schedule();
178}
179
180volatile unsigned int huge = INT_MAX - 2;
181volatile unsigned int ignored;
182
183void lkdtm_OVERFLOW_SIGNED(void)
184{
185	int value;
186
187	value = huge;
188	pr_info("Normal signed addition ...\n");
189	value += 1;
190	ignored = value;
191
192	pr_info("Overflowing signed addition ...\n");
193	value += 4;
194	ignored = value;
195}
196
197
198void lkdtm_OVERFLOW_UNSIGNED(void)
199{
200	unsigned int value;
201
202	value = huge;
203	pr_info("Normal unsigned addition ...\n");
204	value += 1;
205	ignored = value;
206
207	pr_info("Overflowing unsigned addition ...\n");
208	value += 4;
209	ignored = value;
210}
211
212/* Intentionally using old-style flex array definition of 1 byte. */
213struct array_bounds_flex_array {
214	int one;
215	int two;
216	char data[1];
217};
218
219struct array_bounds {
220	int one;
221	int two;
222	char data[8];
223	int three;
224};
225
226void lkdtm_ARRAY_BOUNDS(void)
227{
228	struct array_bounds_flex_array *not_checked;
229	struct array_bounds *checked;
230	volatile int i;
231
232	not_checked = kmalloc(sizeof(*not_checked) * 2, GFP_KERNEL);
233	checked = kmalloc(sizeof(*checked) * 2, GFP_KERNEL);
234	if (!not_checked || !checked) {
235		kfree(not_checked);
236		kfree(checked);
237		return;
238	}
239
240	pr_info("Array access within bounds ...\n");
241	/* For both, touch all bytes in the actual member size. */
242	for (i = 0; i < sizeof(checked->data); i++)
243		checked->data[i] = 'A';
244	/*
245	 * For the uninstrumented flex array member, also touch 1 byte
246	 * beyond to verify it is correctly uninstrumented.
247	 */
248	for (i = 0; i < sizeof(not_checked->data) + 1; i++)
249		not_checked->data[i] = 'A';
250
251	pr_info("Array access beyond bounds ...\n");
252	for (i = 0; i < sizeof(checked->data) + 1; i++)
253		checked->data[i] = 'B';
254
255	kfree(not_checked);
256	kfree(checked);
257	pr_err("FAIL: survived array bounds overflow!\n");
258}
259
260void lkdtm_CORRUPT_LIST_ADD(void)
261{
262	/*
263	 * Initially, an empty list via LIST_HEAD:
264	 *	test_head.next = &test_head
265	 *	test_head.prev = &test_head
266	 */
267	LIST_HEAD(test_head);
268	struct lkdtm_list good, bad;
269	void *target[2] = { };
270	void *redirection = &target;
271
272	pr_info("attempting good list addition\n");
273
274	/*
275	 * Adding to the list performs these actions:
276	 *	test_head.next->prev = &good.node
277	 *	good.node.next = test_head.next
278	 *	good.node.prev = test_head
279	 *	test_head.next = good.node
280	 */
281	list_add(&good.node, &test_head);
282
283	pr_info("attempting corrupted list addition\n");
284	/*
285	 * In simulating this "write what where" primitive, the "what" is
286	 * the address of &bad.node, and the "where" is the address held
287	 * by "redirection".
288	 */
289	test_head.next = redirection;
290	list_add(&bad.node, &test_head);
291
292	if (target[0] == NULL && target[1] == NULL)
293		pr_err("Overwrite did not happen, but no BUG?!\n");
294	else
295		pr_err("list_add() corruption not detected!\n");
296}
297
298void lkdtm_CORRUPT_LIST_DEL(void)
299{
300	LIST_HEAD(test_head);
301	struct lkdtm_list item;
302	void *target[2] = { };
303	void *redirection = &target;
304
305	list_add(&item.node, &test_head);
306
307	pr_info("attempting good list removal\n");
308	list_del(&item.node);
309
310	pr_info("attempting corrupted list removal\n");
311	list_add(&item.node, &test_head);
312
313	/* As with the list_add() test above, this corrupts "next". */
314	item.node.next = redirection;
315	list_del(&item.node);
316
317	if (target[0] == NULL && target[1] == NULL)
318		pr_err("Overwrite did not happen, but no BUG?!\n");
319	else
320		pr_err("list_del() corruption not detected!\n");
321}
322
323/* Test that VMAP_STACK is actually allocating with a leading guard page */
324void lkdtm_STACK_GUARD_PAGE_LEADING(void)
325{
326	const unsigned char *stack = task_stack_page(current);
327	const unsigned char *ptr = stack - 1;
328	volatile unsigned char byte;
329
330	pr_info("attempting bad read from page below current stack\n");
331
332	byte = *ptr;
333
334	pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
335}
336
337/* Test that VMAP_STACK is actually allocating with a trailing guard page */
338void lkdtm_STACK_GUARD_PAGE_TRAILING(void)
339{
340	const unsigned char *stack = task_stack_page(current);
341	const unsigned char *ptr = stack + THREAD_SIZE;
342	volatile unsigned char byte;
343
344	pr_info("attempting bad read from page above current stack\n");
345
346	byte = *ptr;
347
348	pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
349}
350
351void lkdtm_UNSET_SMEP(void)
352{
353#if IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_UML)
354#define MOV_CR4_DEPTH	64
355	void (*direct_write_cr4)(unsigned long val);
356	unsigned char *insn;
357	unsigned long cr4;
358	int i;
359
360	cr4 = native_read_cr4();
361
362	if ((cr4 & X86_CR4_SMEP) != X86_CR4_SMEP) {
363		pr_err("FAIL: SMEP not in use\n");
364		return;
365	}
366	cr4 &= ~(X86_CR4_SMEP);
367
368	pr_info("trying to clear SMEP normally\n");
369	native_write_cr4(cr4);
370	if (cr4 == native_read_cr4()) {
371		pr_err("FAIL: pinning SMEP failed!\n");
372		cr4 |= X86_CR4_SMEP;
373		pr_info("restoring SMEP\n");
374		native_write_cr4(cr4);
375		return;
376	}
377	pr_info("ok: SMEP did not get cleared\n");
378
379	/*
380	 * To test the post-write pinning verification we need to call
381	 * directly into the middle of native_write_cr4() where the
382	 * cr4 write happens, skipping any pinning. This searches for
383	 * the cr4 writing instruction.
384	 */
385	insn = (unsigned char *)native_write_cr4;
386	for (i = 0; i < MOV_CR4_DEPTH; i++) {
387		/* mov %rdi, %cr4 */
388		if (insn[i] == 0x0f && insn[i+1] == 0x22 && insn[i+2] == 0xe7)
389			break;
390		/* mov %rdi,%rax; mov %rax, %cr4 */
391		if (insn[i]   == 0x48 && insn[i+1] == 0x89 &&
392		    insn[i+2] == 0xf8 && insn[i+3] == 0x0f &&
393		    insn[i+4] == 0x22 && insn[i+5] == 0xe0)
394			break;
395	}
396	if (i >= MOV_CR4_DEPTH) {
397		pr_info("ok: cannot locate cr4 writing call gadget\n");
398		return;
399	}
400	direct_write_cr4 = (void *)(insn + i);
401
402	pr_info("trying to clear SMEP with call gadget\n");
403	direct_write_cr4(cr4);
404	if (native_read_cr4() & X86_CR4_SMEP) {
405		pr_info("ok: SMEP removal was reverted\n");
406	} else {
407		pr_err("FAIL: cleared SMEP not detected!\n");
408		cr4 |= X86_CR4_SMEP;
409		pr_info("restoring SMEP\n");
410		native_write_cr4(cr4);
411	}
412#else
413	pr_err("XFAIL: this test is x86_64-only\n");
414#endif
415}
416
417void lkdtm_DOUBLE_FAULT(void)
418{
419#if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
420	/*
421	 * Trigger #DF by setting the stack limit to zero.  This clobbers
422	 * a GDT TLS slot, which is okay because the current task will die
423	 * anyway due to the double fault.
424	 */
425	struct desc_struct d = {
426		.type = 3,	/* expand-up, writable, accessed data */
427		.p = 1,		/* present */
428		.d = 1,		/* 32-bit */
429		.g = 0,		/* limit in bytes */
430		.s = 1,		/* not system */
431	};
432
433	local_irq_disable();
434	write_gdt_entry(get_cpu_gdt_rw(smp_processor_id()),
435			GDT_ENTRY_TLS_MIN, &d, DESCTYPE_S);
436
437	/*
438	 * Put our zero-limit segment in SS and then trigger a fault.  The
439	 * 4-byte access to (%esp) will fault with #SS, and the attempt to
440	 * deliver the fault will recursively cause #SS and result in #DF.
441	 * This whole process happens while NMIs and MCEs are blocked by the
442	 * MOV SS window.  This is nice because an NMI with an invalid SS
443	 * would also double-fault, resulting in the NMI or MCE being lost.
444	 */
445	asm volatile ("movw %0, %%ss; addl $0, (%%esp)" ::
446		      "r" ((unsigned short)(GDT_ENTRY_TLS_MIN << 3)));
447
448	pr_err("FAIL: tried to double fault but didn't die\n");
449#else
450	pr_err("XFAIL: this test is ia32-only\n");
451#endif
452}
453
454#ifdef CONFIG_ARM64
455static noinline void change_pac_parameters(void)
456{
457	if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH)) {
458		/* Reset the keys of current task */
459		ptrauth_thread_init_kernel(current);
460		ptrauth_thread_switch_kernel(current);
461	}
462}
463#endif
464
465noinline void lkdtm_CORRUPT_PAC(void)
466{
467#ifdef CONFIG_ARM64
468#define CORRUPT_PAC_ITERATE	10
469	int i;
470
471	if (!IS_ENABLED(CONFIG_ARM64_PTR_AUTH))
472		pr_err("FAIL: kernel not built with CONFIG_ARM64_PTR_AUTH\n");
473
474	if (!system_supports_address_auth()) {
475		pr_err("FAIL: CPU lacks pointer authentication feature\n");
476		return;
477	}
478
479	pr_info("changing PAC parameters to force function return failure...\n");
480	/*
481	 * PAC is a hash value computed from input keys, return address and
482	 * stack pointer. As pac has fewer bits so there is a chance of
483	 * collision, so iterate few times to reduce the collision probability.
484	 */
485	for (i = 0; i < CORRUPT_PAC_ITERATE; i++)
486		change_pac_parameters();
487
488	pr_err("FAIL: survived PAC changes! Kernel may be unstable from here\n");
489#else
490	pr_err("XFAIL: this test is arm64-only\n");
491#endif
492}
493