forked from ~ljy/RK356X_SDK_RELEASE

blame | history | patch | commit | commitdiff
hc
2024-02-20 102a0743326a03cd1a1202ceda21e175b7d3575c 
 kernel/drivers/misc/lkdtm/bugs.c
....@@ -11,6 +11,11 @@
1111 #include <linux/sched/signal.h>
1212 #include <linux/sched/task_stack.h>
1313 #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
1419 
1520 struct lkdtm_list {
1621 	struct list_head node;
....@@ -22,7 +27,7 @@
2227  * recurse past the end of THREAD_SIZE by default.
2328  */
2429 #if defined(CONFIG_FRAME_WARN) && (CONFIG_FRAME_WARN > 0)
25
-#define REC_STACK_SIZE (CONFIG_FRAME_WARN / 2)
30
+#define REC_STACK_SIZE (_AC(CONFIG_FRAME_WARN, UL) / 2)
2631 #else
2732 #define REC_STACK_SIZE (THREAD_SIZE / 8)
2833 #endif
....@@ -32,12 +37,20 @@
3237 
3338 static DEFINE_SPINLOCK(lock_me_up);
3439 
35
-static int recursive_loop(int remaining)
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)
3648 {
37
-	char buf[REC_STACK_SIZE];
49
+	volatile char buf[REC_STACK_SIZE];
3850 
39
-	/* Make sure compiler does not optimize this away. */
40
-	memset(buf, (remaining & 0xff) | 0x1, REC_STACK_SIZE);
51
+	memset((void *)buf, remaining & 0xFF, sizeof(buf));
52
+	pr_info("loop %d/%d ...\n", (int)buf[remaining % sizeof(buf)],
53
+		recur_count);
4154 	if (!remaining)
4255 		return 0;
4356 	else
....@@ -67,7 +80,12 @@
6780 
6881 void lkdtm_WARNING(void)
6982 {
70
-	WARN(1, "Warning message trigger count: %d\n", warn_counter++);
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);
7189 }
7290 
7391 void lkdtm_EXCEPTION(void)
....@@ -81,9 +99,12 @@
8199 		;
82100 }
83101 
84
-void lkdtm_OVERFLOW(void)
102
+void lkdtm_EXHAUST_STACK(void)
85103 {
86
-	(void) recursive_loop(recur_count);
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");
87108 }
88109 
89110 static noinline void __lkdtm_CORRUPT_STACK(void *stack)
....@@ -97,9 +118,8 @@
97118 	/* Use default char array length that triggers stack protection. */
98119 	char data[8] __aligned(sizeof(void *));
99120 
100
-	__lkdtm_CORRUPT_STACK(&data);
101
-
102
-	pr_info("Corrupted stack containing char array ...\n");
121
+	pr_info("Corrupting stack containing char array ...\n");
122
+	__lkdtm_CORRUPT_STACK((void *)&data);
103123 }
104124 
105125 /* Same as above but will only get a canary with -fstack-protector-strong */
....@@ -110,9 +130,8 @@
110130 		unsigned long *ptr;
111131 	} data __aligned(sizeof(void *));
112132 
113
-	__lkdtm_CORRUPT_STACK(&data);
114
-
115
-	pr_info("Corrupted stack containing union ...\n");
133
+	pr_info("Corrupting stack containing union ...\n");
134
+	__lkdtm_CORRUPT_STACK((void *)&data);
116135 }
117136 
118137 void lkdtm_UNALIGNED_LOAD_STORE_WRITE(void)
....@@ -125,6 +144,9 @@
125144 	if (*p == 0)
126145 		val = 0x87654321;
127146 	*p = val;
147
+
148
+	if (IS_ENABLED(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS))
149
+		pr_err("XFAIL: arch has CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS\n");
128150 }
129151 
130152 void lkdtm_SOFTLOCKUP(void)
....@@ -153,6 +175,86 @@
153175 {
154176 	set_current_state(TASK_UNINTERRUPTIBLE);
155177 	schedule();
178
+}
179
+
180
+volatile unsigned int huge = INT_MAX - 2;
181
+volatile unsigned int ignored;
182
+
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
+
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
+
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
+	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");
156258 }
157259 
158260 void lkdtm_CORRUPT_LIST_ADD(void)
....@@ -218,16 +320,6 @@
218320 		pr_err("list_del() corruption not detected!\n");
219321 }
220322 
221
-/* Test if unbalanced set_fs(KERNEL_DS)/set_fs(USER_DS) check exists. */
222
-void lkdtm_CORRUPT_USER_DS(void)
223
-{
224
-	pr_info("setting bad task size limit\n");
225
-	set_fs(KERNEL_DS);
226
-
227
-	/* Make sure we do not keep running with a KERNEL_DS! */
228
-	force_sig(SIGKILL, current);
229
-}
230
-
231323 /* Test that VMAP_STACK is actually allocating with a leading guard page */
232324 void lkdtm_STACK_GUARD_PAGE_LEADING(void)
233325 {
....@@ -239,7 +331,7 @@
239331 
240332 	byte = *ptr;
241333 
242
-	pr_err("FAIL: accessed page before stack!\n");
334
+	pr_err("FAIL: accessed page before stack! (byte: %x)\n", byte);
243335 }
244336 
245337 /* Test that VMAP_STACK is actually allocating with a trailing guard page */
....@@ -253,5 +345,148 @@
253345 
254346 	byte = *ptr;
255347 
256
-	pr_err("FAIL: accessed page after stack!\n");
348
+	pr_err("FAIL: accessed page after stack! (byte: %x)\n", byte);
349
+}
350
+
351
+void 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
+
417
+void 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
455
+static 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
+
465
+noinline 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
257492 }