1/*
2 * single_step_syscall.c - single-steps various x86 syscalls
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 very simple series of tests that makes system calls with
15 * the TF flag set.  This exercises some nasty kernel code in the
16 * SYSENTER case: SYSENTER does not clear TF, so SYSENTER with TF set
17 * immediately issues #DB from CPL 0.  This requires special handling in
18 * the kernel.
19 */
20
21#define _GNU_SOURCE
22
23#include <sys/time.h>
24#include <time.h>
25#include <stdlib.h>
26#include <sys/syscall.h>
27#include <unistd.h>
28#include <stdio.h>
29#include <string.h>
30#include <inttypes.h>
31#include <sys/mman.h>
32#include <sys/signal.h>
33#include <sys/ucontext.h>
34#include <asm/ldt.h>
35#include <err.h>
36#include <setjmp.h>
37#include <stddef.h>
38#include <stdbool.h>
39#include <sys/ptrace.h>
40#include <sys/user.h>
41
42static void sethandler(int sig, void (*handler)(int, siginfo_t *, void *),
43		       int flags)
44{
45	struct sigaction sa;
46	memset(&sa, 0, sizeof(sa));
47	sa.sa_sigaction = handler;
48	sa.sa_flags = SA_SIGINFO | flags;
49	sigemptyset(&sa.sa_mask);
50	if (sigaction(sig, &sa, 0))
51		err(1, "sigaction");
52}
53
54static volatile sig_atomic_t sig_traps;
55
56#ifdef __x86_64__
57# define REG_IP REG_RIP
58# define WIDTH "q"
59#else
60# define REG_IP REG_EIP
61# define WIDTH "l"
62#endif
63
64static unsigned long get_eflags(void)
65{
66	unsigned long eflags;
67	asm volatile ("pushf" WIDTH "\n\tpop" WIDTH " %0" : "=rm" (eflags));
68	return eflags;
69}
70
71static void set_eflags(unsigned long eflags)
72{
73	asm volatile ("push" WIDTH " %0\n\tpopf" WIDTH
74		      : : "rm" (eflags) : "flags");
75}
76
77#define X86_EFLAGS_TF (1UL << 8)
78
79static void sigtrap(int sig, siginfo_t *info, void *ctx_void)
80{
81	ucontext_t *ctx = (ucontext_t*)ctx_void;
82
83	if (get_eflags() & X86_EFLAGS_TF) {
84		set_eflags(get_eflags() & ~X86_EFLAGS_TF);
85		printf("[WARN]\tSIGTRAP handler had TF set\n");
86		_exit(1);
87	}
88
89	sig_traps++;
90
91	if (sig_traps == 10000 || sig_traps == 10001) {
92		printf("[WARN]\tHit %d SIGTRAPs with si_addr 0x%lx, ip 0x%lx\n",
93		       (int)sig_traps,
94		       (unsigned long)info->si_addr,
95		       (unsigned long)ctx->uc_mcontext.gregs[REG_IP]);
96	}
97}
98
99static void check_result(void)
100{
101	unsigned long new_eflags = get_eflags();
102	set_eflags(new_eflags & ~X86_EFLAGS_TF);
103
104	if (!sig_traps) {
105		printf("[FAIL]\tNo SIGTRAP\n");
106		exit(1);
107	}
108
109	if (!(new_eflags & X86_EFLAGS_TF)) {
110		printf("[FAIL]\tTF was cleared\n");
111		exit(1);
112	}
113
114	printf("[OK]\tSurvived with TF set and %d traps\n", (int)sig_traps);
115	sig_traps = 0;
116}
117
118int main()
119{
120	int tmp;
121
122	sethandler(SIGTRAP, sigtrap, 0);
123
124	printf("[RUN]\tSet TF and check nop\n");
125	set_eflags(get_eflags() | X86_EFLAGS_TF);
126	asm volatile ("nop");
127	check_result();
128
129#ifdef __x86_64__
130	printf("[RUN]\tSet TF and check syscall-less opportunistic sysret\n");
131	set_eflags(get_eflags() | X86_EFLAGS_TF);
132	extern unsigned char post_nop[];
133	asm volatile ("pushf" WIDTH "\n\t"
134		      "pop" WIDTH " %%r11\n\t"
135		      "nop\n\t"
136		      "post_nop:"
137		      : : "c" (post_nop) : "r11");
138	check_result();
139#endif
140
141	printf("[RUN]\tSet TF and check int80\n");
142	set_eflags(get_eflags() | X86_EFLAGS_TF);
143	asm volatile ("int $0x80" : "=a" (tmp) : "a" (SYS_getpid));
144	check_result();
145
146	/*
147	 * This test is particularly interesting if fast syscalls use
148	 * SYSENTER: it triggers a nasty design flaw in SYSENTER.
149	 * Specifically, SYSENTER does not clear TF, so either SYSENTER
150	 * or the next instruction traps at CPL0.  (Of course, Intel
151	 * mostly forgot to document exactly what happens here.)  So we
152	 * get a CPL0 fault with usergs (on 64-bit kernels) and possibly
153	 * no stack.  The only sane way the kernel can possibly handle
154	 * it is to clear TF on return from the #DB handler, but this
155	 * happens way too early to set TF in the saved pt_regs, so the
156	 * kernel has to do something clever to avoid losing track of
157	 * the TF bit.
158	 *
159	 * Needless to say, we've had bugs in this area.
160	 */
161	syscall(SYS_getpid);  /* Force symbol binding without TF set. */
162	printf("[RUN]\tSet TF and check a fast syscall\n");
163	set_eflags(get_eflags() | X86_EFLAGS_TF);
164	syscall(SYS_getpid);
165	check_result();
166
167	/* Now make sure that another fast syscall doesn't set TF again. */
168	printf("[RUN]\tFast syscall with TF cleared\n");
169	fflush(stdout);  /* Force a syscall */
170	if (get_eflags() & X86_EFLAGS_TF) {
171		printf("[FAIL]\tTF is now set\n");
172		exit(1);
173	}
174	if (sig_traps) {
175		printf("[FAIL]\tGot SIGTRAP\n");
176		exit(1);
177	}
178	printf("[OK]\tNothing unexpected happened\n");
179
180	return 0;
181}
182