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