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 / 8UL)
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
137 static pid_t stack_pid;
138 static unsigned long stack_addr;
139
lkdtm_REPORT_STACK(void)140 void lkdtm_REPORT_STACK(void)
141 {
142 volatile uintptr_t magic;
143 pid_t pid = task_pid_nr(current);
144
145 if (pid != stack_pid) {
146 pr_info("Starting stack offset tracking for pid %d\n", pid);
147 stack_pid = pid;
148 stack_addr = (uintptr_t)&magic;
149 }
150
151 pr_info("Stack offset: %d\n", (int)(stack_addr - (uintptr_t)&magic));
152 }
153
lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)154 void lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)
155 {
156 static u8 data[5] __attribute__((aligned(4))) = {1, 2, 3, 4, 5};
157 u32 *p;
158 u32 val = 0x12345678;
159
160 p = (u32 *)(data + 1);
161 if (*p == 0)
162 val = 0x87654321;
163 *p = val;
164
165 if (IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
166 pr_err("XFAIL: arch has CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS\n");
167 }
168
lkdtm_SOFTLOCKUP(void)169 void lkdtm_SOFTLOCKUP(void)
170 {
171 preempt_disable();
172 for (;;)
173 cpu_relax();
174 }
175
lkdtm_HARDLOCKUP(void)176 void lkdtm_HARDLOCKUP(void)
177 {
178 local_irq_disable();
179 for (;;)
180 cpu_relax();
181 }
182
lkdtm_SPINLOCKUP(void)183 void lkdtm_SPINLOCKUP(void)
184 {
185 /* Must be called twice to trigger. */
186 spin_lock(&lock_me_up);
187 /* Let sparse know we intended to exit holding the lock. */
188 __release(&lock_me_up);
189 }
190
lkdtm_HUNG_TASK(void)191 void lkdtm_HUNG_TASK(void)
192 {
193 set_current_state(TASK_UNINTERRUPTIBLE);
194 schedule();
195 }
196
197 volatile unsigned int huge = INT_MAX - 2;
198 volatile unsigned int ignored;
199
lkdtm_OVERFLOW_SIGNED(void)200 void lkdtm_OVERFLOW_SIGNED(void)
201 {
202 int value;
203
204 value = huge;
205 pr_info("Normal signed addition ...\n");
206 value += 1;
207 ignored = value;
208
209 pr_info("Overflowing signed addition ...\n");
210 value += 4;
211 ignored = value;
212 }
213
214
lkdtm_OVERFLOW_UNSIGNED(void)215 void lkdtm_OVERFLOW_UNSIGNED(void)
216 {
217 unsigned int value;
218
219 value = huge;
220 pr_info("Normal unsigned addition ...\n");
221 value += 1;
222 ignored = value;
223
224 pr_info("Overflowing unsigned addition ...\n");
225 value += 4;
226 ignored = value;
227 }
228
229 /* Intentionally using old-style flex array definition of 1 byte. */
230 struct array_bounds_flex_array {
231 int one;
232 int two;
233 char data[1];
234 };
235
236 struct array_bounds {
237 int one;
238 int two;
239 char data[8];
240 int three;
241 };
242
lkdtm_ARRAY_BOUNDS(void)243 void lkdtm_ARRAY_BOUNDS(void)
244 {
245 struct array_bounds_flex_array *not_checked;
246 struct array_bounds *checked;
247 volatile int i;
248
249 not_checked = kmalloc(sizeof(*not_checked) * 2, GFP_KERNEL);
250 checked = kmalloc(sizeof(*checked) * 2, GFP_KERNEL);
251 if (!not_checked || !checked) {
252 kfree(not_checked);
253 kfree(checked);
254 return;
255 }
256
257 pr_info("Array access within bounds ...\n");
258 /* For both, touch all bytes in the actual member size. */
259 for (i = 0; i < sizeof(checked->data); i++)
260 checked->data[i] = 'A';
261 /*
262 * For the uninstrumented flex array member, also touch 1 byte
263 * beyond to verify it is correctly uninstrumented.
264 */
265 for (i = 0; i < sizeof(not_checked->data) + 1; i++)
266 not_checked->data[i] = 'A';
267
268 pr_info("Array access beyond bounds ...\n");
269 for (i = 0; i < sizeof(checked->data) + 1; i++)
270 checked->data[i] = 'B';
271
272 kfree(not_checked);
273 kfree(checked);
274 pr_err("FAIL: survived array bounds overflow!\n");
275 if (IS_ENABLED(CONFIG_UBSAN_BOUNDS))
276 pr_expected_config(CONFIG_UBSAN_TRAP);
277 else
278 pr_expected_config(CONFIG_UBSAN_BOUNDS);
279 }
280
lkdtm_CORRUPT_LIST_ADD(void)281 void lkdtm_CORRUPT_LIST_ADD(void)
282 {
283 /*
284 * Initially, an empty list via LIST_HEAD:
285 * test_head.next = &test_head
286 * test_head.prev = &test_head
287 */
288 LIST_HEAD(test_head);
289 struct lkdtm_list good, bad;
290 void *target[2] = { };
291 void *redirection = ⌖
292
293 pr_info("attempting good list addition\n");
294
295 /*
296 * Adding to the list performs these actions:
297 * test_head.next->prev = &good.node
298 * good.node.next = test_head.next
299 * good.node.prev = test_head
300 * test_head.next = good.node
301 */
302 list_add(&good.node, &test_head);
303
304 pr_info("attempting corrupted list addition\n");
305 /*
306 * In simulating this "write what where" primitive, the "what" is
307 * the address of &bad.node, and the "where" is the address held
308 * by "redirection".
309 */
310 test_head.next = redirection;
311 list_add(&bad.node, &test_head);
312
313 if (target[0] == NULL && target[1] == NULL)
314 pr_err("Overwrite did not happen, but no BUG?!\n");
315 else {
316 pr_err("list_add() corruption not detected!\n");
317 pr_expected_config(CONFIG_DEBUG_LIST);
318 }
319 }
320
lkdtm_CORRUPT_LIST_DEL(void)321 void lkdtm_CORRUPT_LIST_DEL(void)
322 {
323 LIST_HEAD(test_head);
324 struct lkdtm_list item;
325 void *target[2] = { };
326 void *redirection = ⌖
327
328 list_add(&item.node, &test_head);
329
330 pr_info("attempting good list removal\n");
331 list_del(&item.node);
332
333 pr_info("attempting corrupted list removal\n");
334 list_add(&item.node, &test_head);
335
336 /* As with the list_add() test above, this corrupts "next". */
337 item.node.next = redirection;
338 list_del(&item.node);
339
340 if (target[0] == NULL && target[1] == NULL)
341 pr_err("Overwrite did not happen, but no BUG?!\n");
342 else {
343 pr_err("list_del() corruption not detected!\n");
344 pr_expected_config(CONFIG_DEBUG_LIST);
345 }
346 }
347
348 /* Test that VMAP_STACK is actually allocating with a leading guard page */
lkdtm_STACK_GUARD_PAGE_LEADING(void)349 void lkdtm_STACK_GUARD_PAGE_LEADING(void)
350 {
351 const unsigned char *stack = task_stack_page(current);
352 const unsigned char *ptr = stack - 1;
353 volatile unsigned char byte;
354
355 pr_info("attempting bad read from page below current stack\n");
356
357 byte = *ptr;
358
359 pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
360 }
361
362 /* Test that VMAP_STACK is actually allocating with a trailing guard page */
lkdtm_STACK_GUARD_PAGE_TRAILING(void)363 void lkdtm_STACK_GUARD_PAGE_TRAILING(void)
364 {
365 const unsigned char *stack = task_stack_page(current);
366 const unsigned char *ptr = stack + THREAD_SIZE;
367 volatile unsigned char byte;
368
369 pr_info("attempting bad read from page above current stack\n");
370
371 byte = *ptr;
372
373 pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
374 }
375
lkdtm_UNSET_SMEP(void)376 void lkdtm_UNSET_SMEP(void)
377 {
378 #if IS_ENABLED(CONFIG_X86_64) && !IS_ENABLED(CONFIG_UML)
379 #define MOV_CR4_DEPTH 64
380 void (*direct_write_cr4)(unsigned long val);
381 unsigned char *insn;
382 unsigned long cr4;
383 int i;
384
385 cr4 = native_read_cr4();
386
387 if ((cr4 & X86_CR4_SMEP) != X86_CR4_SMEP) {
388 pr_err("FAIL: SMEP not in use\n");
389 return;
390 }
391 cr4 &= ~(X86_CR4_SMEP);
392
393 pr_info("trying to clear SMEP normally\n");
394 native_write_cr4(cr4);
395 if (cr4 == native_read_cr4()) {
396 pr_err("FAIL: pinning SMEP failed!\n");
397 cr4 |= X86_CR4_SMEP;
398 pr_info("restoring SMEP\n");
399 native_write_cr4(cr4);
400 return;
401 }
402 pr_info("ok: SMEP did not get cleared\n");
403
404 /*
405 * To test the post-write pinning verification we need to call
406 * directly into the middle of native_write_cr4() where the
407 * cr4 write happens, skipping any pinning. This searches for
408 * the cr4 writing instruction.
409 */
410 insn = (unsigned char *)native_write_cr4;
411 for (i = 0; i < MOV_CR4_DEPTH; i++) {
412 /* mov %rdi, %cr4 */
413 if (insn[i] == 0x0f && insn[i+1] == 0x22 && insn[i+2] == 0xe7)
414 break;
415 /* mov %rdi,%rax; mov %rax, %cr4 */
416 if (insn[i] == 0x48 && insn[i+1] == 0x89 &&
417 insn[i+2] == 0xf8 && insn[i+3] == 0x0f &&
418 insn[i+4] == 0x22 && insn[i+5] == 0xe0)
419 break;
420 }
421 if (i >= MOV_CR4_DEPTH) {
422 pr_info("ok: cannot locate cr4 writing call gadget\n");
423 return;
424 }
425 direct_write_cr4 = (void *)(insn + i);
426
427 pr_info("trying to clear SMEP with call gadget\n");
428 direct_write_cr4(cr4);
429 if (native_read_cr4() & X86_CR4_SMEP) {
430 pr_info("ok: SMEP removal was reverted\n");
431 } else {
432 pr_err("FAIL: cleared SMEP not detected!\n");
433 cr4 |= X86_CR4_SMEP;
434 pr_info("restoring SMEP\n");
435 native_write_cr4(cr4);
436 }
437 #else
438 pr_err("XFAIL: this test is x86_64-only\n");
439 #endif
440 }
441
lkdtm_DOUBLE_FAULT(void)442 void lkdtm_DOUBLE_FAULT(void)
443 {
444 #if IS_ENABLED(CONFIG_X86_32) && !IS_ENABLED(CONFIG_UML)
445 /*
446 * Trigger #DF by setting the stack limit to zero. This clobbers
447 * a GDT TLS slot, which is okay because the current task will die
448 * anyway due to the double fault.
449 */
450 struct desc_struct d = {
451 .type = 3, /* expand-up, writable, accessed data */
452 .p = 1, /* present */
453 .d = 1, /* 32-bit */
454 .g = 0, /* limit in bytes */
455 .s = 1, /* not system */
456 };
457
458 local_irq_disable();
459 write_gdt_entry(get_cpu_gdt_rw(smp_processor_id()),
460 GDT_ENTRY_TLS_MIN, &d, DESCTYPE_S);
461
462 /*
463 * Put our zero-limit segment in SS and then trigger a fault. The
464 * 4-byte access to (%esp) will fault with #SS, and the attempt to
465 * deliver the fault will recursively cause #SS and result in #DF.
466 * This whole process happens while NMIs and MCEs are blocked by the
467 * MOV SS window. This is nice because an NMI with an invalid SS
468 * would also double-fault, resulting in the NMI or MCE being lost.
469 */
470 asm volatile ("movw %0, %%ss; addl $0, (%%esp)" ::
471 "r" ((unsigned short)(GDT_ENTRY_TLS_MIN << 3)));
472
473 pr_err("FAIL: tried to double fault but didn't die\n");
474 #else
475 pr_err("XFAIL: this test is ia32-only\n");
476 #endif
477 }
478
479 #ifdef CONFIG_ARM64
change_pac_parameters(void)480 static noinline void change_pac_parameters(void)
481 {
482 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL)) {
483 /* Reset the keys of current task */
484 ptrauth_thread_init_kernel(current);
485 ptrauth_thread_switch_kernel(current);
486 }
487 }
488 #endif
489
lkdtm_CORRUPT_PAC(void)490 noinline void lkdtm_CORRUPT_PAC(void)
491 {
492 #ifdef CONFIG_ARM64
493 #define CORRUPT_PAC_ITERATE 10
494 int i;
495
496 if (!IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL))
497 pr_err("FAIL: kernel not built with CONFIG_ARM64_PTR_AUTH_KERNEL\n");
498
499 if (!system_supports_address_auth()) {
500 pr_err("FAIL: CPU lacks pointer authentication feature\n");
501 return;
502 }
503
504 pr_info("changing PAC parameters to force function return failure...\n");
505 /*
506 * PAC is a hash value computed from input keys, return address and
507 * stack pointer. As pac has fewer bits so there is a chance of
508 * collision, so iterate few times to reduce the collision probability.
509 */
510 for (i = 0; i < CORRUPT_PAC_ITERATE; i++)
511 change_pac_parameters();
512
513 pr_err("FAIL: survived PAC changes! Kernel may be unstable from here\n");
514 #else
515 pr_err("XFAIL: this test is arm64-only\n");
516 #endif
517 }
518