1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * sigreturn.c - tests for x86 sigreturn(2) and exit-to-userspace
4  * Copyright (c) 2014-2015 Andrew Lutomirski
5  *
6  * This is a series of tests that exercises the sigreturn(2) syscall and
7  * the IRET / SYSRET paths in the kernel.
8  *
9  * For now, this focuses on the effects of unusual CS and SS values,
10  * and it has a bunch of tests to make sure that ESP/RSP is restored
11  * properly.
12  *
13  * The basic idea behind these tests is to raise(SIGUSR1) to create a
14  * sigcontext frame, plug in the values to be tested, and then return,
15  * which implicitly invokes sigreturn(2) and programs the user context
16  * as desired.
17  *
18  * For tests for which we expect sigreturn and the subsequent return to
19  * user mode to succeed, we return to a short trampoline that generates
20  * SIGTRAP so that the meat of the tests can be ordinary C code in a
21  * SIGTRAP handler.
22  *
23  * The inner workings of each test is documented below.
24  *
25  * Do not run on outdated, unpatched kernels at risk of nasty crashes.
26  */
27 
28 #define _GNU_SOURCE
29 
30 #include <sys/time.h>
31 #include <time.h>
32 #include <stdlib.h>
33 #include <sys/syscall.h>
34 #include <unistd.h>
35 #include <stdio.h>
36 #include <string.h>
37 #include <inttypes.h>
38 #include <sys/mman.h>
39 #include <sys/signal.h>
40 #include <sys/ucontext.h>
41 #include <asm/ldt.h>
42 #include <err.h>
43 #include <setjmp.h>
44 #include <stddef.h>
45 #include <stdbool.h>
46 #include <sys/ptrace.h>
47 #include <sys/user.h>
48 
49 /* Pull in AR_xyz defines. */
50 typedef unsigned int u32;
51 typedef unsigned short u16;
52 #include "../../../../arch/x86/include/asm/desc_defs.h"
53 
54 /*
55  * Copied from asm/ucontext.h, as asm/ucontext.h conflicts badly with the glibc
56  * headers.
57  */
58 #ifdef __x86_64__
59 /*
60  * UC_SIGCONTEXT_SS will be set when delivering 64-bit or x32 signals on
61  * kernels that save SS in the sigcontext.  All kernels that set
62  * UC_SIGCONTEXT_SS will correctly restore at least the low 32 bits of esp
63  * regardless of SS (i.e. they implement espfix).
64  *
65  * Kernels that set UC_SIGCONTEXT_SS will also set UC_STRICT_RESTORE_SS
66  * when delivering a signal that came from 64-bit code.
67  *
68  * Sigreturn restores SS as follows:
69  *
70  * if (saved SS is valid || UC_STRICT_RESTORE_SS is set ||
71  *     saved CS is not 64-bit)
72  *         new SS = saved SS  (will fail IRET and signal if invalid)
73  * else
74  *         new SS = a flat 32-bit data segment
75  */
76 #define UC_SIGCONTEXT_SS       0x2
77 #define UC_STRICT_RESTORE_SS   0x4
78 #endif
79 
80 /*
81  * In principle, this test can run on Linux emulation layers (e.g.
82  * Illumos "LX branded zones").  Solaris-based kernels reserve LDT
83  * entries 0-5 for their own internal purposes, so start our LDT
84  * allocations above that reservation.  (The tests don't pass on LX
85  * branded zones, but at least this lets them run.)
86  */
87 #define LDT_OFFSET 6
88 
89 /* An aligned stack accessible through some of our segments. */
90 static unsigned char stack16[65536] __attribute__((aligned(4096)));
91 
92 /*
93  * An aligned int3 instruction used as a trampoline.  Some of the tests
94  * want to fish out their ss values, so this trampoline copies ss to eax
95  * before the int3.
96  */
97 asm (".pushsection .text\n\t"
98      ".type int3, @function\n\t"
99      ".align 4096\n\t"
100      "int3:\n\t"
101      "mov %ss,%ecx\n\t"
102      "int3\n\t"
103      ".size int3, . - int3\n\t"
104      ".align 4096, 0xcc\n\t"
105      ".popsection");
106 extern char int3[4096];
107 
108 /*
109  * At startup, we prepapre:
110  *
111  * - ldt_nonexistent_sel: An LDT entry that doesn't exist (all-zero
112  *   descriptor or out of bounds).
113  * - code16_sel: A 16-bit LDT code segment pointing to int3.
114  * - data16_sel: A 16-bit LDT data segment pointing to stack16.
115  * - npcode32_sel: A 32-bit not-present LDT code segment pointing to int3.
116  * - npdata32_sel: A 32-bit not-present LDT data segment pointing to stack16.
117  * - gdt_data16_idx: A 16-bit GDT data segment pointing to stack16.
118  * - gdt_npdata32_idx: A 32-bit not-present GDT data segment pointing to
119  *   stack16.
120  *
121  * For no particularly good reason, xyz_sel is a selector value with the
122  * RPL and LDT bits filled in, whereas xyz_idx is just an index into the
123  * descriptor table.  These variables will be zero if their respective
124  * segments could not be allocated.
125  */
126 static unsigned short ldt_nonexistent_sel;
127 static unsigned short code16_sel, data16_sel, npcode32_sel, npdata32_sel;
128 
129 static unsigned short gdt_data16_idx, gdt_npdata32_idx;
130 
GDT3(int idx)131 static unsigned short GDT3(int idx)
132 {
133 	return (idx << 3) | 3;
134 }
135 
LDT3(int idx)136 static unsigned short LDT3(int idx)
137 {
138 	return (idx << 3) | 7;
139 }
140 
sethandler(int sig,void (* handler)(int,siginfo_t *,void *),int flags)141 static void sethandler(int sig, void (*handler)(int, siginfo_t *, void *),
142 		       int flags)
143 {
144 	struct sigaction sa;
145 	memset(&sa, 0, sizeof(sa));
146 	sa.sa_sigaction = handler;
147 	sa.sa_flags = SA_SIGINFO | flags;
148 	sigemptyset(&sa.sa_mask);
149 	if (sigaction(sig, &sa, 0))
150 		err(1, "sigaction");
151 }
152 
clearhandler(int sig)153 static void clearhandler(int sig)
154 {
155 	struct sigaction sa;
156 	memset(&sa, 0, sizeof(sa));
157 	sa.sa_handler = SIG_DFL;
158 	sigemptyset(&sa.sa_mask);
159 	if (sigaction(sig, &sa, 0))
160 		err(1, "sigaction");
161 }
162 
add_ldt(const struct user_desc * desc,unsigned short * var,const char * name)163 static void add_ldt(const struct user_desc *desc, unsigned short *var,
164 		    const char *name)
165 {
166 	if (syscall(SYS_modify_ldt, 1, desc, sizeof(*desc)) == 0) {
167 		*var = LDT3(desc->entry_number);
168 	} else {
169 		printf("[NOTE]\tFailed to create %s segment\n", name);
170 		*var = 0;
171 	}
172 }
173 
setup_ldt(void)174 static void setup_ldt(void)
175 {
176 	if ((unsigned long)stack16 > (1ULL << 32) - sizeof(stack16))
177 		errx(1, "stack16 is too high\n");
178 	if ((unsigned long)int3 > (1ULL << 32) - sizeof(int3))
179 		errx(1, "int3 is too high\n");
180 
181 	ldt_nonexistent_sel = LDT3(LDT_OFFSET + 2);
182 
183 	const struct user_desc code16_desc = {
184 		.entry_number    = LDT_OFFSET + 0,
185 		.base_addr       = (unsigned long)int3,
186 		.limit           = 4095,
187 		.seg_32bit       = 0,
188 		.contents        = 2, /* Code, not conforming */
189 		.read_exec_only  = 0,
190 		.limit_in_pages  = 0,
191 		.seg_not_present = 0,
192 		.useable         = 0
193 	};
194 	add_ldt(&code16_desc, &code16_sel, "code16");
195 
196 	const struct user_desc data16_desc = {
197 		.entry_number    = LDT_OFFSET + 1,
198 		.base_addr       = (unsigned long)stack16,
199 		.limit           = 0xffff,
200 		.seg_32bit       = 0,
201 		.contents        = 0, /* Data, grow-up */
202 		.read_exec_only  = 0,
203 		.limit_in_pages  = 0,
204 		.seg_not_present = 0,
205 		.useable         = 0
206 	};
207 	add_ldt(&data16_desc, &data16_sel, "data16");
208 
209 	const struct user_desc npcode32_desc = {
210 		.entry_number    = LDT_OFFSET + 3,
211 		.base_addr       = (unsigned long)int3,
212 		.limit           = 4095,
213 		.seg_32bit       = 1,
214 		.contents        = 2, /* Code, not conforming */
215 		.read_exec_only  = 0,
216 		.limit_in_pages  = 0,
217 		.seg_not_present = 1,
218 		.useable         = 0
219 	};
220 	add_ldt(&npcode32_desc, &npcode32_sel, "npcode32");
221 
222 	const struct user_desc npdata32_desc = {
223 		.entry_number    = LDT_OFFSET + 4,
224 		.base_addr       = (unsigned long)stack16,
225 		.limit           = 0xffff,
226 		.seg_32bit       = 1,
227 		.contents        = 0, /* Data, grow-up */
228 		.read_exec_only  = 0,
229 		.limit_in_pages  = 0,
230 		.seg_not_present = 1,
231 		.useable         = 0
232 	};
233 	add_ldt(&npdata32_desc, &npdata32_sel, "npdata32");
234 
235 	struct user_desc gdt_data16_desc = {
236 		.entry_number    = -1,
237 		.base_addr       = (unsigned long)stack16,
238 		.limit           = 0xffff,
239 		.seg_32bit       = 0,
240 		.contents        = 0, /* Data, grow-up */
241 		.read_exec_only  = 0,
242 		.limit_in_pages  = 0,
243 		.seg_not_present = 0,
244 		.useable         = 0
245 	};
246 
247 	if (syscall(SYS_set_thread_area, &gdt_data16_desc) == 0) {
248 		/*
249 		 * This probably indicates vulnerability to CVE-2014-8133.
250 		 * Merely getting here isn't definitive, though, and we'll
251 		 * diagnose the problem for real later on.
252 		 */
253 		printf("[WARN]\tset_thread_area allocated data16 at index %d\n",
254 		       gdt_data16_desc.entry_number);
255 		gdt_data16_idx = gdt_data16_desc.entry_number;
256 	} else {
257 		printf("[OK]\tset_thread_area refused 16-bit data\n");
258 	}
259 
260 	struct user_desc gdt_npdata32_desc = {
261 		.entry_number    = -1,
262 		.base_addr       = (unsigned long)stack16,
263 		.limit           = 0xffff,
264 		.seg_32bit       = 1,
265 		.contents        = 0, /* Data, grow-up */
266 		.read_exec_only  = 0,
267 		.limit_in_pages  = 0,
268 		.seg_not_present = 1,
269 		.useable         = 0
270 	};
271 
272 	if (syscall(SYS_set_thread_area, &gdt_npdata32_desc) == 0) {
273 		/*
274 		 * As a hardening measure, newer kernels don't allow this.
275 		 */
276 		printf("[WARN]\tset_thread_area allocated npdata32 at index %d\n",
277 		       gdt_npdata32_desc.entry_number);
278 		gdt_npdata32_idx = gdt_npdata32_desc.entry_number;
279 	} else {
280 		printf("[OK]\tset_thread_area refused 16-bit data\n");
281 	}
282 }
283 
284 /* State used by our signal handlers. */
285 static gregset_t initial_regs, requested_regs, resulting_regs;
286 
287 /* Instructions for the SIGUSR1 handler. */
288 static volatile unsigned short sig_cs, sig_ss;
289 static volatile sig_atomic_t sig_trapped, sig_err, sig_trapno;
290 #ifdef __x86_64__
291 static volatile sig_atomic_t sig_corrupt_final_ss;
292 #endif
293 
294 /* Abstractions for some 32-bit vs 64-bit differences. */
295 #ifdef __x86_64__
296 # define REG_IP REG_RIP
297 # define REG_SP REG_RSP
298 # define REG_CX REG_RCX
299 
300 struct selectors {
301 	unsigned short cs, gs, fs, ss;
302 };
303 
ssptr(ucontext_t * ctx)304 static unsigned short *ssptr(ucontext_t *ctx)
305 {
306 	struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
307 	return &sels->ss;
308 }
309 
csptr(ucontext_t * ctx)310 static unsigned short *csptr(ucontext_t *ctx)
311 {
312 	struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
313 	return &sels->cs;
314 }
315 #else
316 # define REG_IP REG_EIP
317 # define REG_SP REG_ESP
318 # define REG_CX REG_ECX
319 
ssptr(ucontext_t * ctx)320 static greg_t *ssptr(ucontext_t *ctx)
321 {
322 	return &ctx->uc_mcontext.gregs[REG_SS];
323 }
324 
csptr(ucontext_t * ctx)325 static greg_t *csptr(ucontext_t *ctx)
326 {
327 	return &ctx->uc_mcontext.gregs[REG_CS];
328 }
329 #endif
330 
331 /*
332  * Checks a given selector for its code bitness or returns -1 if it's not
333  * a usable code segment selector.
334  */
cs_bitness(unsigned short cs)335 int cs_bitness(unsigned short cs)
336 {
337 	uint32_t valid = 0, ar;
338 	asm ("lar %[cs], %[ar]\n\t"
339 	     "jnz 1f\n\t"
340 	     "mov $1, %[valid]\n\t"
341 	     "1:"
342 	     : [ar] "=r" (ar), [valid] "+rm" (valid)
343 	     : [cs] "r" (cs));
344 
345 	if (!valid)
346 		return -1;
347 
348 	bool db = (ar & (1 << 22));
349 	bool l = (ar & (1 << 21));
350 
351 	if (!(ar & (1<<11)))
352 	    return -1;	/* Not code. */
353 
354 	if (l && !db)
355 		return 64;
356 	else if (!l && db)
357 		return 32;
358 	else if (!l && !db)
359 		return 16;
360 	else
361 		return -1;	/* Unknown bitness. */
362 }
363 
364 /*
365  * Checks a given selector for its code bitness or returns -1 if it's not
366  * a usable code segment selector.
367  */
is_valid_ss(unsigned short cs)368 bool is_valid_ss(unsigned short cs)
369 {
370 	uint32_t valid = 0, ar;
371 	asm ("lar %[cs], %[ar]\n\t"
372 	     "jnz 1f\n\t"
373 	     "mov $1, %[valid]\n\t"
374 	     "1:"
375 	     : [ar] "=r" (ar), [valid] "+rm" (valid)
376 	     : [cs] "r" (cs));
377 
378 	if (!valid)
379 		return false;
380 
381 	if ((ar & AR_TYPE_MASK) != AR_TYPE_RWDATA &&
382 	    (ar & AR_TYPE_MASK) != AR_TYPE_RWDATA_EXPDOWN)
383 		return false;
384 
385 	return (ar & AR_P);
386 }
387 
388 /* Number of errors in the current test case. */
389 static volatile sig_atomic_t nerrs;
390 
validate_signal_ss(int sig,ucontext_t * ctx)391 static void validate_signal_ss(int sig, ucontext_t *ctx)
392 {
393 #ifdef __x86_64__
394 	bool was_64bit = (cs_bitness(*csptr(ctx)) == 64);
395 
396 	if (!(ctx->uc_flags & UC_SIGCONTEXT_SS)) {
397 		printf("[FAIL]\tUC_SIGCONTEXT_SS was not set\n");
398 		nerrs++;
399 
400 		/*
401 		 * This happens on Linux 4.1.  The rest will fail, too, so
402 		 * return now to reduce the noise.
403 		 */
404 		return;
405 	}
406 
407 	/* UC_STRICT_RESTORE_SS is set iff we came from 64-bit mode. */
408 	if (!!(ctx->uc_flags & UC_STRICT_RESTORE_SS) != was_64bit) {
409 		printf("[FAIL]\tUC_STRICT_RESTORE_SS was wrong in signal %d\n",
410 		       sig);
411 		nerrs++;
412 	}
413 
414 	if (is_valid_ss(*ssptr(ctx))) {
415 		/*
416 		 * DOSEMU was written before 64-bit sigcontext had SS, and
417 		 * it tries to figure out the signal source SS by looking at
418 		 * the physical register.  Make sure that keeps working.
419 		 */
420 		unsigned short hw_ss;
421 		asm ("mov %%ss, %0" : "=rm" (hw_ss));
422 		if (hw_ss != *ssptr(ctx)) {
423 			printf("[FAIL]\tHW SS didn't match saved SS\n");
424 			nerrs++;
425 		}
426 	}
427 #endif
428 }
429 
430 /*
431  * SIGUSR1 handler.  Sets CS and SS as requested and points IP to the
432  * int3 trampoline.  Sets SP to a large known value so that we can see
433  * whether the value round-trips back to user mode correctly.
434  */
sigusr1(int sig,siginfo_t * info,void * ctx_void)435 static void sigusr1(int sig, siginfo_t *info, void *ctx_void)
436 {
437 	ucontext_t *ctx = (ucontext_t*)ctx_void;
438 
439 	validate_signal_ss(sig, ctx);
440 
441 	memcpy(&initial_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
442 
443 	*csptr(ctx) = sig_cs;
444 	*ssptr(ctx) = sig_ss;
445 
446 	ctx->uc_mcontext.gregs[REG_IP] =
447 		sig_cs == code16_sel ? 0 : (unsigned long)&int3;
448 	ctx->uc_mcontext.gregs[REG_SP] = (unsigned long)0x8badf00d5aadc0deULL;
449 	ctx->uc_mcontext.gregs[REG_CX] = 0;
450 
451 #ifdef __i386__
452 	/*
453 	 * Make sure the kernel doesn't inadvertently use DS or ES-relative
454 	 * accesses in a region where user DS or ES is loaded.
455 	 *
456 	 * Skip this for 64-bit builds because long mode doesn't care about
457 	 * DS and ES and skipping it increases test coverage a little bit,
458 	 * since 64-bit kernels can still run the 32-bit build.
459 	 */
460 	ctx->uc_mcontext.gregs[REG_DS] = 0;
461 	ctx->uc_mcontext.gregs[REG_ES] = 0;
462 #endif
463 
464 	memcpy(&requested_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
465 	requested_regs[REG_CX] = *ssptr(ctx);	/* The asm code does this. */
466 
467 	return;
468 }
469 
470 /*
471  * Called after a successful sigreturn (via int3) or from a failed
472  * sigreturn (directly by kernel).  Restores our state so that the
473  * original raise(SIGUSR1) returns.
474  */
sigtrap(int sig,siginfo_t * info,void * ctx_void)475 static void sigtrap(int sig, siginfo_t *info, void *ctx_void)
476 {
477 	ucontext_t *ctx = (ucontext_t*)ctx_void;
478 
479 	validate_signal_ss(sig, ctx);
480 
481 	sig_err = ctx->uc_mcontext.gregs[REG_ERR];
482 	sig_trapno = ctx->uc_mcontext.gregs[REG_TRAPNO];
483 
484 	unsigned short ss;
485 	asm ("mov %%ss,%0" : "=r" (ss));
486 
487 	greg_t asm_ss = ctx->uc_mcontext.gregs[REG_CX];
488 	if (asm_ss != sig_ss && sig == SIGTRAP) {
489 		/* Sanity check failure. */
490 		printf("[FAIL]\tSIGTRAP: ss = %hx, frame ss = %hx, ax = %llx\n",
491 		       ss, *ssptr(ctx), (unsigned long long)asm_ss);
492 		nerrs++;
493 	}
494 
495 	memcpy(&resulting_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
496 	memcpy(&ctx->uc_mcontext.gregs, &initial_regs, sizeof(gregset_t));
497 
498 #ifdef __x86_64__
499 	if (sig_corrupt_final_ss) {
500 		if (ctx->uc_flags & UC_STRICT_RESTORE_SS) {
501 			printf("[FAIL]\tUC_STRICT_RESTORE_SS was set inappropriately\n");
502 			nerrs++;
503 		} else {
504 			/*
505 			 * DOSEMU transitions from 32-bit to 64-bit mode by
506 			 * adjusting sigcontext, and it requires that this work
507 			 * even if the saved SS is bogus.
508 			 */
509 			printf("\tCorrupting SS on return to 64-bit mode\n");
510 			*ssptr(ctx) = 0;
511 		}
512 	}
513 #endif
514 
515 	sig_trapped = sig;
516 }
517 
518 #ifdef __x86_64__
519 /* Tests recovery if !UC_STRICT_RESTORE_SS */
sigusr2(int sig,siginfo_t * info,void * ctx_void)520 static void sigusr2(int sig, siginfo_t *info, void *ctx_void)
521 {
522 	ucontext_t *ctx = (ucontext_t*)ctx_void;
523 
524 	if (!(ctx->uc_flags & UC_STRICT_RESTORE_SS)) {
525 		printf("[FAIL]\traise(2) didn't set UC_STRICT_RESTORE_SS\n");
526 		nerrs++;
527 		return;  /* We can't do the rest. */
528 	}
529 
530 	ctx->uc_flags &= ~UC_STRICT_RESTORE_SS;
531 	*ssptr(ctx) = 0;
532 
533 	/* Return.  The kernel should recover without sending another signal. */
534 }
535 
test_nonstrict_ss(void)536 static int test_nonstrict_ss(void)
537 {
538 	clearhandler(SIGUSR1);
539 	clearhandler(SIGTRAP);
540 	clearhandler(SIGSEGV);
541 	clearhandler(SIGILL);
542 	sethandler(SIGUSR2, sigusr2, 0);
543 
544 	nerrs = 0;
545 
546 	printf("[RUN]\tClear UC_STRICT_RESTORE_SS and corrupt SS\n");
547 	raise(SIGUSR2);
548 	if (!nerrs)
549 		printf("[OK]\tIt worked\n");
550 
551 	return nerrs;
552 }
553 #endif
554 
555 /* Finds a usable code segment of the requested bitness. */
find_cs(int bitness)556 int find_cs(int bitness)
557 {
558 	unsigned short my_cs;
559 
560 	asm ("mov %%cs,%0" :  "=r" (my_cs));
561 
562 	if (cs_bitness(my_cs) == bitness)
563 		return my_cs;
564 	if (cs_bitness(my_cs + (2 << 3)) == bitness)
565 		return my_cs + (2 << 3);
566 	if (my_cs > (2<<3) && cs_bitness(my_cs - (2 << 3)) == bitness)
567 	    return my_cs - (2 << 3);
568 	if (cs_bitness(code16_sel) == bitness)
569 		return code16_sel;
570 
571 	printf("[WARN]\tCould not find %d-bit CS\n", bitness);
572 	return -1;
573 }
574 
test_valid_sigreturn(int cs_bits,bool use_16bit_ss,int force_ss)575 static int test_valid_sigreturn(int cs_bits, bool use_16bit_ss, int force_ss)
576 {
577 	int cs = find_cs(cs_bits);
578 	if (cs == -1) {
579 		printf("[SKIP]\tCode segment unavailable for %d-bit CS, %d-bit SS\n",
580 		       cs_bits, use_16bit_ss ? 16 : 32);
581 		return 0;
582 	}
583 
584 	if (force_ss != -1) {
585 		sig_ss = force_ss;
586 	} else {
587 		if (use_16bit_ss) {
588 			if (!data16_sel) {
589 				printf("[SKIP]\tData segment unavailable for %d-bit CS, 16-bit SS\n",
590 				       cs_bits);
591 				return 0;
592 			}
593 			sig_ss = data16_sel;
594 		} else {
595 			asm volatile ("mov %%ss,%0" : "=r" (sig_ss));
596 		}
597 	}
598 
599 	sig_cs = cs;
600 
601 	printf("[RUN]\tValid sigreturn: %d-bit CS (%hx), %d-bit SS (%hx%s)\n",
602 	       cs_bits, sig_cs, use_16bit_ss ? 16 : 32, sig_ss,
603 	       (sig_ss & 4) ? "" : ", GDT");
604 
605 	raise(SIGUSR1);
606 
607 	nerrs = 0;
608 
609 	/*
610 	 * Check that each register had an acceptable value when the
611 	 * int3 trampoline was invoked.
612 	 */
613 	for (int i = 0; i < NGREG; i++) {
614 		greg_t req = requested_regs[i], res = resulting_regs[i];
615 
616 		if (i == REG_TRAPNO || i == REG_IP)
617 			continue;	/* don't care */
618 
619 		if (i == REG_SP) {
620 			/*
621 			 * If we were using a 16-bit stack segment, then
622 			 * the kernel is a bit stuck: IRET only restores
623 			 * the low 16 bits of ESP/RSP if SS is 16-bit.
624 			 * The kernel uses a hack to restore bits 31:16,
625 			 * but that hack doesn't help with bits 63:32.
626 			 * On Intel CPUs, bits 63:32 end up zeroed, and, on
627 			 * AMD CPUs, they leak the high bits of the kernel
628 			 * espfix64 stack pointer.  There's very little that
629 			 * the kernel can do about it.
630 			 *
631 			 * Similarly, if we are returning to a 32-bit context,
632 			 * the CPU will often lose the high 32 bits of RSP.
633 			 */
634 
635 			if (res == req)
636 				continue;
637 
638 			if (cs_bits != 64 && ((res ^ req) & 0xFFFFFFFF) == 0) {
639 				printf("[NOTE]\tSP: %llx -> %llx\n",
640 				       (unsigned long long)req,
641 				       (unsigned long long)res);
642 				continue;
643 			}
644 
645 			printf("[FAIL]\tSP mismatch: requested 0x%llx; got 0x%llx\n",
646 			       (unsigned long long)requested_regs[i],
647 			       (unsigned long long)resulting_regs[i]);
648 			nerrs++;
649 			continue;
650 		}
651 
652 		bool ignore_reg = false;
653 #if __i386__
654 		if (i == REG_UESP)
655 			ignore_reg = true;
656 #else
657 		if (i == REG_CSGSFS) {
658 			struct selectors *req_sels =
659 				(void *)&requested_regs[REG_CSGSFS];
660 			struct selectors *res_sels =
661 				(void *)&resulting_regs[REG_CSGSFS];
662 			if (req_sels->cs != res_sels->cs) {
663 				printf("[FAIL]\tCS mismatch: requested 0x%hx; got 0x%hx\n",
664 				       req_sels->cs, res_sels->cs);
665 				nerrs++;
666 			}
667 
668 			if (req_sels->ss != res_sels->ss) {
669 				printf("[FAIL]\tSS mismatch: requested 0x%hx; got 0x%hx\n",
670 				       req_sels->ss, res_sels->ss);
671 				nerrs++;
672 			}
673 
674 			continue;
675 		}
676 #endif
677 
678 		/* Sanity check on the kernel */
679 		if (i == REG_CX && req != res) {
680 			printf("[FAIL]\tCX (saved SP) mismatch: requested 0x%llx; got 0x%llx\n",
681 			       (unsigned long long)req,
682 			       (unsigned long long)res);
683 			nerrs++;
684 			continue;
685 		}
686 
687 		if (req != res && !ignore_reg) {
688 			printf("[FAIL]\tReg %d mismatch: requested 0x%llx; got 0x%llx\n",
689 			       i, (unsigned long long)req,
690 			       (unsigned long long)res);
691 			nerrs++;
692 		}
693 	}
694 
695 	if (nerrs == 0)
696 		printf("[OK]\tall registers okay\n");
697 
698 	return nerrs;
699 }
700 
test_bad_iret(int cs_bits,unsigned short ss,int force_cs)701 static int test_bad_iret(int cs_bits, unsigned short ss, int force_cs)
702 {
703 	int cs = force_cs == -1 ? find_cs(cs_bits) : force_cs;
704 	if (cs == -1)
705 		return 0;
706 
707 	sig_cs = cs;
708 	sig_ss = ss;
709 
710 	printf("[RUN]\t%d-bit CS (%hx), bogus SS (%hx)\n",
711 	       cs_bits, sig_cs, sig_ss);
712 
713 	sig_trapped = 0;
714 	raise(SIGUSR1);
715 	if (sig_trapped) {
716 		char errdesc[32] = "";
717 		if (sig_err) {
718 			const char *src = (sig_err & 1) ? " EXT" : "";
719 			const char *table;
720 			if ((sig_err & 0x6) == 0x0)
721 				table = "GDT";
722 			else if ((sig_err & 0x6) == 0x4)
723 				table = "LDT";
724 			else if ((sig_err & 0x6) == 0x2)
725 				table = "IDT";
726 			else
727 				table = "???";
728 
729 			sprintf(errdesc, "%s%s index %d, ",
730 				table, src, sig_err >> 3);
731 		}
732 
733 		char trapname[32];
734 		if (sig_trapno == 13)
735 			strcpy(trapname, "GP");
736 		else if (sig_trapno == 11)
737 			strcpy(trapname, "NP");
738 		else if (sig_trapno == 12)
739 			strcpy(trapname, "SS");
740 		else if (sig_trapno == 32)
741 			strcpy(trapname, "IRET");  /* X86_TRAP_IRET */
742 		else
743 			sprintf(trapname, "%d", sig_trapno);
744 
745 		printf("[OK]\tGot #%s(0x%lx) (i.e. %s%s)\n",
746 		       trapname, (unsigned long)sig_err,
747 		       errdesc, strsignal(sig_trapped));
748 		return 0;
749 	} else {
750 		/*
751 		 * This also implicitly tests UC_STRICT_RESTORE_SS:
752 		 * We check that these signals set UC_STRICT_RESTORE_SS and,
753 		 * if UC_STRICT_RESTORE_SS doesn't cause strict behavior,
754 		 * then we won't get SIGSEGV.
755 		 */
756 		printf("[FAIL]\tDid not get SIGSEGV\n");
757 		return 1;
758 	}
759 }
760 
main()761 int main()
762 {
763 	int total_nerrs = 0;
764 	unsigned short my_cs, my_ss;
765 
766 	asm volatile ("mov %%cs,%0" : "=r" (my_cs));
767 	asm volatile ("mov %%ss,%0" : "=r" (my_ss));
768 	setup_ldt();
769 
770 	stack_t stack = {
771 		/* Our sigaltstack scratch space. */
772 		.ss_sp = malloc(sizeof(char) * SIGSTKSZ),
773 		.ss_size = SIGSTKSZ,
774 	};
775 	if (sigaltstack(&stack, NULL) != 0)
776 		err(1, "sigaltstack");
777 
778 	sethandler(SIGUSR1, sigusr1, 0);
779 	sethandler(SIGTRAP, sigtrap, SA_ONSTACK);
780 
781 	/* Easy cases: return to a 32-bit SS in each possible CS bitness. */
782 	total_nerrs += test_valid_sigreturn(64, false, -1);
783 	total_nerrs += test_valid_sigreturn(32, false, -1);
784 	total_nerrs += test_valid_sigreturn(16, false, -1);
785 
786 	/*
787 	 * Test easy espfix cases: return to a 16-bit LDT SS in each possible
788 	 * CS bitness.  NB: with a long mode CS, the SS bitness is irrelevant.
789 	 *
790 	 * This catches the original missing-espfix-on-64-bit-kernels issue
791 	 * as well as CVE-2014-8134.
792 	 */
793 	total_nerrs += test_valid_sigreturn(64, true, -1);
794 	total_nerrs += test_valid_sigreturn(32, true, -1);
795 	total_nerrs += test_valid_sigreturn(16, true, -1);
796 
797 	if (gdt_data16_idx) {
798 		/*
799 		 * For performance reasons, Linux skips espfix if SS points
800 		 * to the GDT.  If we were able to allocate a 16-bit SS in
801 		 * the GDT, see if it leaks parts of the kernel stack pointer.
802 		 *
803 		 * This tests for CVE-2014-8133.
804 		 */
805 		total_nerrs += test_valid_sigreturn(64, true,
806 						    GDT3(gdt_data16_idx));
807 		total_nerrs += test_valid_sigreturn(32, true,
808 						    GDT3(gdt_data16_idx));
809 		total_nerrs += test_valid_sigreturn(16, true,
810 						    GDT3(gdt_data16_idx));
811 	}
812 
813 #ifdef __x86_64__
814 	/* Nasty ABI case: check SS corruption handling. */
815 	sig_corrupt_final_ss = 1;
816 	total_nerrs += test_valid_sigreturn(32, false, -1);
817 	total_nerrs += test_valid_sigreturn(32, true, -1);
818 	sig_corrupt_final_ss = 0;
819 #endif
820 
821 	/*
822 	 * We're done testing valid sigreturn cases.  Now we test states
823 	 * for which sigreturn itself will succeed but the subsequent
824 	 * entry to user mode will fail.
825 	 *
826 	 * Depending on the failure mode and the kernel bitness, these
827 	 * entry failures can generate SIGSEGV, SIGBUS, or SIGILL.
828 	 */
829 	clearhandler(SIGTRAP);
830 	sethandler(SIGSEGV, sigtrap, SA_ONSTACK);
831 	sethandler(SIGBUS, sigtrap, SA_ONSTACK);
832 	sethandler(SIGILL, sigtrap, SA_ONSTACK);  /* 32-bit kernels do this */
833 
834 	/* Easy failures: invalid SS, resulting in #GP(0) */
835 	test_bad_iret(64, ldt_nonexistent_sel, -1);
836 	test_bad_iret(32, ldt_nonexistent_sel, -1);
837 	test_bad_iret(16, ldt_nonexistent_sel, -1);
838 
839 	/* These fail because SS isn't a data segment, resulting in #GP(SS) */
840 	test_bad_iret(64, my_cs, -1);
841 	test_bad_iret(32, my_cs, -1);
842 	test_bad_iret(16, my_cs, -1);
843 
844 	/* Try to return to a not-present code segment, triggering #NP(SS). */
845 	test_bad_iret(32, my_ss, npcode32_sel);
846 
847 	/*
848 	 * Try to return to a not-present but otherwise valid data segment.
849 	 * This will cause IRET to fail with #SS on the espfix stack.  This
850 	 * exercises CVE-2014-9322.
851 	 *
852 	 * Note that, if espfix is enabled, 64-bit Linux will lose track
853 	 * of the actual cause of failure and report #GP(0) instead.
854 	 * This would be very difficult for Linux to avoid, because
855 	 * espfix64 causes IRET failures to be promoted to #DF, so the
856 	 * original exception frame is never pushed onto the stack.
857 	 */
858 	test_bad_iret(32, npdata32_sel, -1);
859 
860 	/*
861 	 * Try to return to a not-present but otherwise valid data
862 	 * segment without invoking espfix.  Newer kernels don't allow
863 	 * this to happen in the first place.  On older kernels, though,
864 	 * this can trigger CVE-2014-9322.
865 	 */
866 	if (gdt_npdata32_idx)
867 		test_bad_iret(32, GDT3(gdt_npdata32_idx), -1);
868 
869 #ifdef __x86_64__
870 	total_nerrs += test_nonstrict_ss();
871 #endif
872 
873 	free(stack.ss_sp);
874 	return total_nerrs ? 1 : 0;
875 }
876