1 /*
2  * sigreturn.c - tests for x86 sigreturn(2) and exit-to-userspace
3  * Copyright (c) 2014-2015 Andrew Lutomirski
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but
10  * WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * General Public License for more details.
13  *
14  * This is a series of tests that exercises the sigreturn(2) syscall and
15  * the IRET / SYSRET paths in the kernel.
16  *
17  * For now, this focuses on the effects of unusual CS and SS values,
18  * and it has a bunch of tests to make sure that ESP/RSP is restored
19  * properly.
20  *
21  * The basic idea behind these tests is to raise(SIGUSR1) to create a
22  * sigcontext frame, plug in the values to be tested, and then return,
23  * which implicitly invokes sigreturn(2) and programs the user context
24  * as desired.
25  *
26  * For tests for which we expect sigreturn and the subsequent return to
27  * user mode to succeed, we return to a short trampoline that generates
28  * SIGTRAP so that the meat of the tests can be ordinary C code in a
29  * SIGTRAP handler.
30  *
31  * The inner workings of each test is documented below.
32  *
33  * Do not run on outdated, unpatched kernels at risk of nasty crashes.
34  */
35 
36 #define _GNU_SOURCE
37 
38 #include <sys/time.h>
39 #include <time.h>
40 #include <stdlib.h>
41 #include <sys/syscall.h>
42 #include <unistd.h>
43 #include <stdio.h>
44 #include <string.h>
45 #include <inttypes.h>
46 #include <sys/mman.h>
47 #include <sys/signal.h>
48 #include <sys/ucontext.h>
49 #include <asm/ldt.h>
50 #include <err.h>
51 #include <setjmp.h>
52 #include <stddef.h>
53 #include <stdbool.h>
54 #include <sys/ptrace.h>
55 #include <sys/user.h>
56 
57 /*
58  * In principle, this test can run on Linux emulation layers (e.g.
59  * Illumos "LX branded zones").  Solaris-based kernels reserve LDT
60  * entries 0-5 for their own internal purposes, so start our LDT
61  * allocations above that reservation.  (The tests don't pass on LX
62  * branded zones, but at least this lets them run.)
63  */
64 #define LDT_OFFSET 6
65 
66 /* An aligned stack accessible through some of our segments. */
67 static unsigned char stack16[65536] __attribute__((aligned(4096)));
68 
69 /*
70  * An aligned int3 instruction used as a trampoline.  Some of the tests
71  * want to fish out their ss values, so this trampoline copies ss to eax
72  * before the int3.
73  */
74 asm (".pushsection .text\n\t"
75      ".type int3, @function\n\t"
76      ".align 4096\n\t"
77      "int3:\n\t"
78      "mov %ss,%eax\n\t"
79      "int3\n\t"
80      ".size int3, . - int3\n\t"
81      ".align 4096, 0xcc\n\t"
82      ".popsection");
83 extern char int3[4096];
84 
85 /*
86  * At startup, we prepapre:
87  *
88  * - ldt_nonexistent_sel: An LDT entry that doesn't exist (all-zero
89  *   descriptor or out of bounds).
90  * - code16_sel: A 16-bit LDT code segment pointing to int3.
91  * - data16_sel: A 16-bit LDT data segment pointing to stack16.
92  * - npcode32_sel: A 32-bit not-present LDT code segment pointing to int3.
93  * - npdata32_sel: A 32-bit not-present LDT data segment pointing to stack16.
94  * - gdt_data16_idx: A 16-bit GDT data segment pointing to stack16.
95  * - gdt_npdata32_idx: A 32-bit not-present GDT data segment pointing to
96  *   stack16.
97  *
98  * For no particularly good reason, xyz_sel is a selector value with the
99  * RPL and LDT bits filled in, whereas xyz_idx is just an index into the
100  * descriptor table.  These variables will be zero if their respective
101  * segments could not be allocated.
102  */
103 static unsigned short ldt_nonexistent_sel;
104 static unsigned short code16_sel, data16_sel, npcode32_sel, npdata32_sel;
105 
106 static unsigned short gdt_data16_idx, gdt_npdata32_idx;
107 
108 static unsigned short GDT3(int idx)
109 {
110 	return (idx << 3) | 3;
111 }
112 
113 static unsigned short LDT3(int idx)
114 {
115 	return (idx << 3) | 7;
116 }
117 
118 /* Our sigaltstack scratch space. */
119 static char altstack_data[SIGSTKSZ];
120 
121 static void sethandler(int sig, void (*handler)(int, siginfo_t *, void *),
122 		       int flags)
123 {
124 	struct sigaction sa;
125 	memset(&sa, 0, sizeof(sa));
126 	sa.sa_sigaction = handler;
127 	sa.sa_flags = SA_SIGINFO | flags;
128 	sigemptyset(&sa.sa_mask);
129 	if (sigaction(sig, &sa, 0))
130 		err(1, "sigaction");
131 }
132 
133 static void clearhandler(int sig)
134 {
135 	struct sigaction sa;
136 	memset(&sa, 0, sizeof(sa));
137 	sa.sa_handler = SIG_DFL;
138 	sigemptyset(&sa.sa_mask);
139 	if (sigaction(sig, &sa, 0))
140 		err(1, "sigaction");
141 }
142 
143 static void add_ldt(const struct user_desc *desc, unsigned short *var,
144 		    const char *name)
145 {
146 	if (syscall(SYS_modify_ldt, 1, desc, sizeof(*desc)) == 0) {
147 		*var = LDT3(desc->entry_number);
148 	} else {
149 		printf("[NOTE]\tFailed to create %s segment\n", name);
150 		*var = 0;
151 	}
152 }
153 
154 static void setup_ldt(void)
155 {
156 	if ((unsigned long)stack16 > (1ULL << 32) - sizeof(stack16))
157 		errx(1, "stack16 is too high\n");
158 	if ((unsigned long)int3 > (1ULL << 32) - sizeof(int3))
159 		errx(1, "int3 is too high\n");
160 
161 	ldt_nonexistent_sel = LDT3(LDT_OFFSET + 2);
162 
163 	const struct user_desc code16_desc = {
164 		.entry_number    = LDT_OFFSET + 0,
165 		.base_addr       = (unsigned long)int3,
166 		.limit           = 4095,
167 		.seg_32bit       = 0,
168 		.contents        = 2, /* Code, not conforming */
169 		.read_exec_only  = 0,
170 		.limit_in_pages  = 0,
171 		.seg_not_present = 0,
172 		.useable         = 0
173 	};
174 	add_ldt(&code16_desc, &code16_sel, "code16");
175 
176 	const struct user_desc data16_desc = {
177 		.entry_number    = LDT_OFFSET + 1,
178 		.base_addr       = (unsigned long)stack16,
179 		.limit           = 0xffff,
180 		.seg_32bit       = 0,
181 		.contents        = 0, /* Data, grow-up */
182 		.read_exec_only  = 0,
183 		.limit_in_pages  = 0,
184 		.seg_not_present = 0,
185 		.useable         = 0
186 	};
187 	add_ldt(&data16_desc, &data16_sel, "data16");
188 
189 	const struct user_desc npcode32_desc = {
190 		.entry_number    = LDT_OFFSET + 3,
191 		.base_addr       = (unsigned long)int3,
192 		.limit           = 4095,
193 		.seg_32bit       = 1,
194 		.contents        = 2, /* Code, not conforming */
195 		.read_exec_only  = 0,
196 		.limit_in_pages  = 0,
197 		.seg_not_present = 1,
198 		.useable         = 0
199 	};
200 	add_ldt(&npcode32_desc, &npcode32_sel, "npcode32");
201 
202 	const struct user_desc npdata32_desc = {
203 		.entry_number    = LDT_OFFSET + 4,
204 		.base_addr       = (unsigned long)stack16,
205 		.limit           = 0xffff,
206 		.seg_32bit       = 1,
207 		.contents        = 0, /* Data, grow-up */
208 		.read_exec_only  = 0,
209 		.limit_in_pages  = 0,
210 		.seg_not_present = 1,
211 		.useable         = 0
212 	};
213 	add_ldt(&npdata32_desc, &npdata32_sel, "npdata32");
214 
215 	struct user_desc gdt_data16_desc = {
216 		.entry_number    = -1,
217 		.base_addr       = (unsigned long)stack16,
218 		.limit           = 0xffff,
219 		.seg_32bit       = 0,
220 		.contents        = 0, /* Data, grow-up */
221 		.read_exec_only  = 0,
222 		.limit_in_pages  = 0,
223 		.seg_not_present = 0,
224 		.useable         = 0
225 	};
226 
227 	if (syscall(SYS_set_thread_area, &gdt_data16_desc) == 0) {
228 		/*
229 		 * This probably indicates vulnerability to CVE-2014-8133.
230 		 * Merely getting here isn't definitive, though, and we'll
231 		 * diagnose the problem for real later on.
232 		 */
233 		printf("[WARN]\tset_thread_area allocated data16 at index %d\n",
234 		       gdt_data16_desc.entry_number);
235 		gdt_data16_idx = gdt_data16_desc.entry_number;
236 	} else {
237 		printf("[OK]\tset_thread_area refused 16-bit data\n");
238 	}
239 
240 	struct user_desc gdt_npdata32_desc = {
241 		.entry_number    = -1,
242 		.base_addr       = (unsigned long)stack16,
243 		.limit           = 0xffff,
244 		.seg_32bit       = 1,
245 		.contents        = 0, /* Data, grow-up */
246 		.read_exec_only  = 0,
247 		.limit_in_pages  = 0,
248 		.seg_not_present = 1,
249 		.useable         = 0
250 	};
251 
252 	if (syscall(SYS_set_thread_area, &gdt_npdata32_desc) == 0) {
253 		/*
254 		 * As a hardening measure, newer kernels don't allow this.
255 		 */
256 		printf("[WARN]\tset_thread_area allocated npdata32 at index %d\n",
257 		       gdt_npdata32_desc.entry_number);
258 		gdt_npdata32_idx = gdt_npdata32_desc.entry_number;
259 	} else {
260 		printf("[OK]\tset_thread_area refused 16-bit data\n");
261 	}
262 }
263 
264 /* State used by our signal handlers. */
265 static gregset_t initial_regs, requested_regs, resulting_regs;
266 
267 /* Instructions for the SIGUSR1 handler. */
268 static volatile unsigned short sig_cs, sig_ss;
269 static volatile sig_atomic_t sig_trapped, sig_err, sig_trapno;
270 
271 /* Abstractions for some 32-bit vs 64-bit differences. */
272 #ifdef __x86_64__
273 # define REG_IP REG_RIP
274 # define REG_SP REG_RSP
275 # define REG_AX REG_RAX
276 
277 struct selectors {
278 	unsigned short cs, gs, fs, ss;
279 };
280 
281 static unsigned short *ssptr(ucontext_t *ctx)
282 {
283 	struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
284 	return &sels->ss;
285 }
286 
287 static unsigned short *csptr(ucontext_t *ctx)
288 {
289 	struct selectors *sels = (void *)&ctx->uc_mcontext.gregs[REG_CSGSFS];
290 	return &sels->cs;
291 }
292 #else
293 # define REG_IP REG_EIP
294 # define REG_SP REG_ESP
295 # define REG_AX REG_EAX
296 
297 static greg_t *ssptr(ucontext_t *ctx)
298 {
299 	return &ctx->uc_mcontext.gregs[REG_SS];
300 }
301 
302 static greg_t *csptr(ucontext_t *ctx)
303 {
304 	return &ctx->uc_mcontext.gregs[REG_CS];
305 }
306 #endif
307 
308 /* Number of errors in the current test case. */
309 static volatile sig_atomic_t nerrs;
310 
311 /*
312  * SIGUSR1 handler.  Sets CS and SS as requested and points IP to the
313  * int3 trampoline.  Sets SP to a large known value so that we can see
314  * whether the value round-trips back to user mode correctly.
315  */
316 static void sigusr1(int sig, siginfo_t *info, void *ctx_void)
317 {
318 	ucontext_t *ctx = (ucontext_t*)ctx_void;
319 
320 	memcpy(&initial_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
321 
322 	*csptr(ctx) = sig_cs;
323 	*ssptr(ctx) = sig_ss;
324 
325 	ctx->uc_mcontext.gregs[REG_IP] =
326 		sig_cs == code16_sel ? 0 : (unsigned long)&int3;
327 	ctx->uc_mcontext.gregs[REG_SP] = (unsigned long)0x8badf00d5aadc0deULL;
328 	ctx->uc_mcontext.gregs[REG_AX] = 0;
329 
330 	memcpy(&requested_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
331 	requested_regs[REG_AX] = *ssptr(ctx);	/* The asm code does this. */
332 
333 	return;
334 }
335 
336 /*
337  * Called after a successful sigreturn.  Restores our state so that
338  * the original raise(SIGUSR1) returns.
339  */
340 static void sigtrap(int sig, siginfo_t *info, void *ctx_void)
341 {
342 	ucontext_t *ctx = (ucontext_t*)ctx_void;
343 
344 	sig_err = ctx->uc_mcontext.gregs[REG_ERR];
345 	sig_trapno = ctx->uc_mcontext.gregs[REG_TRAPNO];
346 
347 	unsigned short ss;
348 	asm ("mov %%ss,%0" : "=r" (ss));
349 
350 	greg_t asm_ss = ctx->uc_mcontext.gregs[REG_AX];
351 	if (asm_ss != sig_ss && sig == SIGTRAP) {
352 		/* Sanity check failure. */
353 		printf("[FAIL]\tSIGTRAP: ss = %hx, frame ss = %hx, ax = %llx\n",
354 		       ss, *ssptr(ctx), (unsigned long long)asm_ss);
355 		nerrs++;
356 	}
357 
358 	memcpy(&resulting_regs, &ctx->uc_mcontext.gregs, sizeof(gregset_t));
359 	memcpy(&ctx->uc_mcontext.gregs, &initial_regs, sizeof(gregset_t));
360 
361 	sig_trapped = sig;
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  */
368 int cs_bitness(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 -1;
380 
381 	bool db = (ar & (1 << 22));
382 	bool l = (ar & (1 << 21));
383 
384 	if (!(ar & (1<<11)))
385 	    return -1;	/* Not code. */
386 
387 	if (l && !db)
388 		return 64;
389 	else if (!l && db)
390 		return 32;
391 	else if (!l && !db)
392 		return 16;
393 	else
394 		return -1;	/* Unknown bitness. */
395 }
396 
397 /* Finds a usable code segment of the requested bitness. */
398 int find_cs(int bitness)
399 {
400 	unsigned short my_cs;
401 
402 	asm ("mov %%cs,%0" :  "=r" (my_cs));
403 
404 	if (cs_bitness(my_cs) == bitness)
405 		return my_cs;
406 	if (cs_bitness(my_cs + (2 << 3)) == bitness)
407 		return my_cs + (2 << 3);
408 	if (my_cs > (2<<3) && cs_bitness(my_cs - (2 << 3)) == bitness)
409 	    return my_cs - (2 << 3);
410 	if (cs_bitness(code16_sel) == bitness)
411 		return code16_sel;
412 
413 	printf("[WARN]\tCould not find %d-bit CS\n", bitness);
414 	return -1;
415 }
416 
417 static int test_valid_sigreturn(int cs_bits, bool use_16bit_ss, int force_ss)
418 {
419 	int cs = find_cs(cs_bits);
420 	if (cs == -1) {
421 		printf("[SKIP]\tCode segment unavailable for %d-bit CS, %d-bit SS\n",
422 		       cs_bits, use_16bit_ss ? 16 : 32);
423 		return 0;
424 	}
425 
426 	if (force_ss != -1) {
427 		sig_ss = force_ss;
428 	} else {
429 		if (use_16bit_ss) {
430 			if (!data16_sel) {
431 				printf("[SKIP]\tData segment unavailable for %d-bit CS, 16-bit SS\n",
432 				       cs_bits);
433 				return 0;
434 			}
435 			sig_ss = data16_sel;
436 		} else {
437 			asm volatile ("mov %%ss,%0" : "=r" (sig_ss));
438 		}
439 	}
440 
441 	sig_cs = cs;
442 
443 	printf("[RUN]\tValid sigreturn: %d-bit CS (%hx), %d-bit SS (%hx%s)\n",
444 	       cs_bits, sig_cs, use_16bit_ss ? 16 : 32, sig_ss,
445 	       (sig_ss & 4) ? "" : ", GDT");
446 
447 	raise(SIGUSR1);
448 
449 	nerrs = 0;
450 
451 	/*
452 	 * Check that each register had an acceptable value when the
453 	 * int3 trampoline was invoked.
454 	 */
455 	for (int i = 0; i < NGREG; i++) {
456 		greg_t req = requested_regs[i], res = resulting_regs[i];
457 		if (i == REG_TRAPNO || i == REG_IP)
458 			continue;	/* don't care */
459 		if (i == REG_SP) {
460 			printf("\tSP: %llx -> %llx\n", (unsigned long long)req,
461 			       (unsigned long long)res);
462 
463 			/*
464 			 * In many circumstances, the high 32 bits of rsp
465 			 * are zeroed.  For example, we could be a real
466 			 * 32-bit program, or we could hit any of a number
467 			 * of poorly-documented IRET or segmented ESP
468 			 * oddities.  If this happens, it's okay.
469 			 */
470 			if (res == (req & 0xFFFFFFFF))
471 				continue;  /* OK; not expected to work */
472 		}
473 
474 		bool ignore_reg = false;
475 #if __i386__
476 		if (i == REG_UESP)
477 			ignore_reg = true;
478 #else
479 		if (i == REG_CSGSFS) {
480 			struct selectors *req_sels =
481 				(void *)&requested_regs[REG_CSGSFS];
482 			struct selectors *res_sels =
483 				(void *)&resulting_regs[REG_CSGSFS];
484 			if (req_sels->cs != res_sels->cs) {
485 				printf("[FAIL]\tCS mismatch: requested 0x%hx; got 0x%hx\n",
486 				       req_sels->cs, res_sels->cs);
487 				nerrs++;
488 			}
489 
490 			if (req_sels->ss != res_sels->ss) {
491 				printf("[FAIL]\tSS mismatch: requested 0x%hx; got 0x%hx\n",
492 				       req_sels->ss, res_sels->ss);
493 				nerrs++;
494 			}
495 
496 			continue;
497 		}
498 #endif
499 
500 		/* Sanity check on the kernel */
501 		if (i == REG_AX && requested_regs[i] != resulting_regs[i]) {
502 			printf("[FAIL]\tAX (saved SP) mismatch: requested 0x%llx; got 0x%llx\n",
503 			       (unsigned long long)requested_regs[i],
504 			       (unsigned long long)resulting_regs[i]);
505 			nerrs++;
506 			continue;
507 		}
508 
509 		if (requested_regs[i] != resulting_regs[i] && !ignore_reg) {
510 			/*
511 			 * SP is particularly interesting here.  The
512 			 * usual cause of failures is that we hit the
513 			 * nasty IRET case of returning to a 16-bit SS,
514 			 * in which case bits 16:31 of the *kernel*
515 			 * stack pointer persist in ESP.
516 			 */
517 			printf("[FAIL]\tReg %d mismatch: requested 0x%llx; got 0x%llx\n",
518 			       i, (unsigned long long)requested_regs[i],
519 			       (unsigned long long)resulting_regs[i]);
520 			nerrs++;
521 		}
522 	}
523 
524 	if (nerrs == 0)
525 		printf("[OK]\tall registers okay\n");
526 
527 	return nerrs;
528 }
529 
530 static int test_bad_iret(int cs_bits, unsigned short ss, int force_cs)
531 {
532 	int cs = force_cs == -1 ? find_cs(cs_bits) : force_cs;
533 	if (cs == -1)
534 		return 0;
535 
536 	sig_cs = cs;
537 	sig_ss = ss;
538 
539 	printf("[RUN]\t%d-bit CS (%hx), bogus SS (%hx)\n",
540 	       cs_bits, sig_cs, sig_ss);
541 
542 	sig_trapped = 0;
543 	raise(SIGUSR1);
544 	if (sig_trapped) {
545 		char errdesc[32] = "";
546 		if (sig_err) {
547 			const char *src = (sig_err & 1) ? " EXT" : "";
548 			const char *table;
549 			if ((sig_err & 0x6) == 0x0)
550 				table = "GDT";
551 			else if ((sig_err & 0x6) == 0x4)
552 				table = "LDT";
553 			else if ((sig_err & 0x6) == 0x2)
554 				table = "IDT";
555 			else
556 				table = "???";
557 
558 			sprintf(errdesc, "%s%s index %d, ",
559 				table, src, sig_err >> 3);
560 		}
561 
562 		char trapname[32];
563 		if (sig_trapno == 13)
564 			strcpy(trapname, "GP");
565 		else if (sig_trapno == 11)
566 			strcpy(trapname, "NP");
567 		else if (sig_trapno == 12)
568 			strcpy(trapname, "SS");
569 		else if (sig_trapno == 32)
570 			strcpy(trapname, "IRET");  /* X86_TRAP_IRET */
571 		else
572 			sprintf(trapname, "%d", sig_trapno);
573 
574 		printf("[OK]\tGot #%s(0x%lx) (i.e. %s%s)\n",
575 		       trapname, (unsigned long)sig_err,
576 		       errdesc, strsignal(sig_trapped));
577 		return 0;
578 	} else {
579 		printf("[FAIL]\tDid not get SIGSEGV\n");
580 		return 1;
581 	}
582 }
583 
584 int main()
585 {
586 	int total_nerrs = 0;
587 	unsigned short my_cs, my_ss;
588 
589 	asm volatile ("mov %%cs,%0" : "=r" (my_cs));
590 	asm volatile ("mov %%ss,%0" : "=r" (my_ss));
591 	setup_ldt();
592 
593 	stack_t stack = {
594 		.ss_sp = altstack_data,
595 		.ss_size = SIGSTKSZ,
596 	};
597 	if (sigaltstack(&stack, NULL) != 0)
598 		err(1, "sigaltstack");
599 
600 	sethandler(SIGUSR1, sigusr1, 0);
601 	sethandler(SIGTRAP, sigtrap, SA_ONSTACK);
602 
603 	/* Easy cases: return to a 32-bit SS in each possible CS bitness. */
604 	total_nerrs += test_valid_sigreturn(64, false, -1);
605 	total_nerrs += test_valid_sigreturn(32, false, -1);
606 	total_nerrs += test_valid_sigreturn(16, false, -1);
607 
608 	/*
609 	 * Test easy espfix cases: return to a 16-bit LDT SS in each possible
610 	 * CS bitness.  NB: with a long mode CS, the SS bitness is irrelevant.
611 	 *
612 	 * This catches the original missing-espfix-on-64-bit-kernels issue
613 	 * as well as CVE-2014-8134.
614 	 */
615 	total_nerrs += test_valid_sigreturn(64, true, -1);
616 	total_nerrs += test_valid_sigreturn(32, true, -1);
617 	total_nerrs += test_valid_sigreturn(16, true, -1);
618 
619 	if (gdt_data16_idx) {
620 		/*
621 		 * For performance reasons, Linux skips espfix if SS points
622 		 * to the GDT.  If we were able to allocate a 16-bit SS in
623 		 * the GDT, see if it leaks parts of the kernel stack pointer.
624 		 *
625 		 * This tests for CVE-2014-8133.
626 		 */
627 		total_nerrs += test_valid_sigreturn(64, true,
628 						    GDT3(gdt_data16_idx));
629 		total_nerrs += test_valid_sigreturn(32, true,
630 						    GDT3(gdt_data16_idx));
631 		total_nerrs += test_valid_sigreturn(16, true,
632 						    GDT3(gdt_data16_idx));
633 	}
634 
635 	/*
636 	 * We're done testing valid sigreturn cases.  Now we test states
637 	 * for which sigreturn itself will succeed but the subsequent
638 	 * entry to user mode will fail.
639 	 *
640 	 * Depending on the failure mode and the kernel bitness, these
641 	 * entry failures can generate SIGSEGV, SIGBUS, or SIGILL.
642 	 */
643 	clearhandler(SIGTRAP);
644 	sethandler(SIGSEGV, sigtrap, SA_ONSTACK);
645 	sethandler(SIGBUS, sigtrap, SA_ONSTACK);
646 	sethandler(SIGILL, sigtrap, SA_ONSTACK);  /* 32-bit kernels do this */
647 
648 	/* Easy failures: invalid SS, resulting in #GP(0) */
649 	test_bad_iret(64, ldt_nonexistent_sel, -1);
650 	test_bad_iret(32, ldt_nonexistent_sel, -1);
651 	test_bad_iret(16, ldt_nonexistent_sel, -1);
652 
653 	/* These fail because SS isn't a data segment, resulting in #GP(SS) */
654 	test_bad_iret(64, my_cs, -1);
655 	test_bad_iret(32, my_cs, -1);
656 	test_bad_iret(16, my_cs, -1);
657 
658 	/* Try to return to a not-present code segment, triggering #NP(SS). */
659 	test_bad_iret(32, my_ss, npcode32_sel);
660 
661 	/*
662 	 * Try to return to a not-present but otherwise valid data segment.
663 	 * This will cause IRET to fail with #SS on the espfix stack.  This
664 	 * exercises CVE-2014-9322.
665 	 *
666 	 * Note that, if espfix is enabled, 64-bit Linux will lose track
667 	 * of the actual cause of failure and report #GP(0) instead.
668 	 * This would be very difficult for Linux to avoid, because
669 	 * espfix64 causes IRET failures to be promoted to #DF, so the
670 	 * original exception frame is never pushed onto the stack.
671 	 */
672 	test_bad_iret(32, npdata32_sel, -1);
673 
674 	/*
675 	 * Try to return to a not-present but otherwise valid data
676 	 * segment without invoking espfix.  Newer kernels don't allow
677 	 * this to happen in the first place.  On older kernels, though,
678 	 * this can trigger CVE-2014-9322.
679 	 */
680 	if (gdt_npdata32_idx)
681 		test_bad_iret(32, GDT3(gdt_npdata32_idx), -1);
682 
683 	return total_nerrs ? 1 : 0;
684 }
685