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25  * Do not run on outdated, unpatched kernels at risk of nasty crashes.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