1 /*
2  * Copyright (c) 2012-2014 Andy Lutomirski <luto@amacapital.net>
3  *
4  * Based on the original implementation which is:
5  *  Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE
6  *  Copyright 2003 Andi Kleen, SuSE Labs.
7  *
8  *  Parts of the original code have been moved to arch/x86/vdso/vma.c
9  *
10  * This file implements vsyscall emulation.  vsyscalls are a legacy ABI:
11  * Userspace can request certain kernel services by calling fixed
12  * addresses.  This concept is problematic:
13  *
14  * - It interferes with ASLR.
15  * - It's awkward to write code that lives in kernel addresses but is
16  *   callable by userspace at fixed addresses.
17  * - The whole concept is impossible for 32-bit compat userspace.
18  * - UML cannot easily virtualize a vsyscall.
19  *
20  * As of mid-2014, I believe that there is no new userspace code that
21  * will use a vsyscall if the vDSO is present.  I hope that there will
22  * soon be no new userspace code that will ever use a vsyscall.
23  *
24  * The code in this file emulates vsyscalls when notified of a page
25  * fault to a vsyscall address.
26  */
27 
28 #include <linux/kernel.h>
29 #include <linux/timer.h>
30 #include <linux/sched/signal.h>
31 #include <linux/mm_types.h>
32 #include <linux/syscalls.h>
33 #include <linux/ratelimit.h>
34 
35 #include <asm/vsyscall.h>
36 #include <asm/unistd.h>
37 #include <asm/fixmap.h>
38 #include <asm/traps.h>
39 
40 #define CREATE_TRACE_POINTS
41 #include "vsyscall_trace.h"
42 
43 static enum { EMULATE, NATIVE, NONE } vsyscall_mode =
44 #if defined(CONFIG_LEGACY_VSYSCALL_NATIVE)
45 	NATIVE;
46 #elif defined(CONFIG_LEGACY_VSYSCALL_NONE)
47 	NONE;
48 #else
49 	EMULATE;
50 #endif
51 
52 static int __init vsyscall_setup(char *str)
53 {
54 	if (str) {
55 		if (!strcmp("emulate", str))
56 			vsyscall_mode = EMULATE;
57 		else if (!strcmp("native", str))
58 			vsyscall_mode = NATIVE;
59 		else if (!strcmp("none", str))
60 			vsyscall_mode = NONE;
61 		else
62 			return -EINVAL;
63 
64 		return 0;
65 	}
66 
67 	return -EINVAL;
68 }
69 early_param("vsyscall", vsyscall_setup);
70 
71 static void warn_bad_vsyscall(const char *level, struct pt_regs *regs,
72 			      const char *message)
73 {
74 	if (!show_unhandled_signals)
75 		return;
76 
77 	printk_ratelimited("%s%s[%d] %s ip:%lx cs:%lx sp:%lx ax:%lx si:%lx di:%lx\n",
78 			   level, current->comm, task_pid_nr(current),
79 			   message, regs->ip, regs->cs,
80 			   regs->sp, regs->ax, regs->si, regs->di);
81 }
82 
83 static int addr_to_vsyscall_nr(unsigned long addr)
84 {
85 	int nr;
86 
87 	if ((addr & ~0xC00UL) != VSYSCALL_ADDR)
88 		return -EINVAL;
89 
90 	nr = (addr & 0xC00UL) >> 10;
91 	if (nr >= 3)
92 		return -EINVAL;
93 
94 	return nr;
95 }
96 
97 static bool write_ok_or_segv(unsigned long ptr, size_t size)
98 {
99 	/*
100 	 * XXX: if access_ok, get_user, and put_user handled
101 	 * sig_on_uaccess_err, this could go away.
102 	 */
103 
104 	if (!access_ok(VERIFY_WRITE, (void __user *)ptr, size)) {
105 		siginfo_t info;
106 		struct thread_struct *thread = &current->thread;
107 
108 		thread->error_code	= 6;  /* user fault, no page, write */
109 		thread->cr2		= ptr;
110 		thread->trap_nr		= X86_TRAP_PF;
111 
112 		memset(&info, 0, sizeof(info));
113 		info.si_signo		= SIGSEGV;
114 		info.si_errno		= 0;
115 		info.si_code		= SEGV_MAPERR;
116 		info.si_addr		= (void __user *)ptr;
117 
118 		force_sig_info(SIGSEGV, &info, current);
119 		return false;
120 	} else {
121 		return true;
122 	}
123 }
124 
125 bool emulate_vsyscall(struct pt_regs *regs, unsigned long address)
126 {
127 	struct task_struct *tsk;
128 	unsigned long caller;
129 	int vsyscall_nr, syscall_nr, tmp;
130 	int prev_sig_on_uaccess_err;
131 	long ret;
132 
133 	/*
134 	 * No point in checking CS -- the only way to get here is a user mode
135 	 * trap to a high address, which means that we're in 64-bit user code.
136 	 */
137 
138 	WARN_ON_ONCE(address != regs->ip);
139 
140 	if (vsyscall_mode == NONE) {
141 		warn_bad_vsyscall(KERN_INFO, regs,
142 				  "vsyscall attempted with vsyscall=none");
143 		return false;
144 	}
145 
146 	vsyscall_nr = addr_to_vsyscall_nr(address);
147 
148 	trace_emulate_vsyscall(vsyscall_nr);
149 
150 	if (vsyscall_nr < 0) {
151 		warn_bad_vsyscall(KERN_WARNING, regs,
152 				  "misaligned vsyscall (exploit attempt or buggy program) -- look up the vsyscall kernel parameter if you need a workaround");
153 		goto sigsegv;
154 	}
155 
156 	if (get_user(caller, (unsigned long __user *)regs->sp) != 0) {
157 		warn_bad_vsyscall(KERN_WARNING, regs,
158 				  "vsyscall with bad stack (exploit attempt?)");
159 		goto sigsegv;
160 	}
161 
162 	tsk = current;
163 
164 	/*
165 	 * Check for access_ok violations and find the syscall nr.
166 	 *
167 	 * NULL is a valid user pointer (in the access_ok sense) on 32-bit and
168 	 * 64-bit, so we don't need to special-case it here.  For all the
169 	 * vsyscalls, NULL means "don't write anything" not "write it at
170 	 * address 0".
171 	 */
172 	switch (vsyscall_nr) {
173 	case 0:
174 		if (!write_ok_or_segv(regs->di, sizeof(struct timeval)) ||
175 		    !write_ok_or_segv(regs->si, sizeof(struct timezone))) {
176 			ret = -EFAULT;
177 			goto check_fault;
178 		}
179 
180 		syscall_nr = __NR_gettimeofday;
181 		break;
182 
183 	case 1:
184 		if (!write_ok_or_segv(regs->di, sizeof(time_t))) {
185 			ret = -EFAULT;
186 			goto check_fault;
187 		}
188 
189 		syscall_nr = __NR_time;
190 		break;
191 
192 	case 2:
193 		if (!write_ok_or_segv(regs->di, sizeof(unsigned)) ||
194 		    !write_ok_or_segv(regs->si, sizeof(unsigned))) {
195 			ret = -EFAULT;
196 			goto check_fault;
197 		}
198 
199 		syscall_nr = __NR_getcpu;
200 		break;
201 	}
202 
203 	/*
204 	 * Handle seccomp.  regs->ip must be the original value.
205 	 * See seccomp_send_sigsys and Documentation/prctl/seccomp_filter.txt.
206 	 *
207 	 * We could optimize the seccomp disabled case, but performance
208 	 * here doesn't matter.
209 	 */
210 	regs->orig_ax = syscall_nr;
211 	regs->ax = -ENOSYS;
212 	tmp = secure_computing(NULL);
213 	if ((!tmp && regs->orig_ax != syscall_nr) || regs->ip != address) {
214 		warn_bad_vsyscall(KERN_DEBUG, regs,
215 				  "seccomp tried to change syscall nr or ip");
216 		do_exit(SIGSYS);
217 	}
218 	regs->orig_ax = -1;
219 	if (tmp)
220 		goto do_ret;  /* skip requested */
221 
222 	/*
223 	 * With a real vsyscall, page faults cause SIGSEGV.  We want to
224 	 * preserve that behavior to make writing exploits harder.
225 	 */
226 	prev_sig_on_uaccess_err = current->thread.sig_on_uaccess_err;
227 	current->thread.sig_on_uaccess_err = 1;
228 
229 	ret = -EFAULT;
230 	switch (vsyscall_nr) {
231 	case 0:
232 		ret = sys_gettimeofday(
233 			(struct timeval __user *)regs->di,
234 			(struct timezone __user *)regs->si);
235 		break;
236 
237 	case 1:
238 		ret = sys_time((time_t __user *)regs->di);
239 		break;
240 
241 	case 2:
242 		ret = sys_getcpu((unsigned __user *)regs->di,
243 				 (unsigned __user *)regs->si,
244 				 NULL);
245 		break;
246 	}
247 
248 	current->thread.sig_on_uaccess_err = prev_sig_on_uaccess_err;
249 
250 check_fault:
251 	if (ret == -EFAULT) {
252 		/* Bad news -- userspace fed a bad pointer to a vsyscall. */
253 		warn_bad_vsyscall(KERN_INFO, regs,
254 				  "vsyscall fault (exploit attempt?)");
255 
256 		/*
257 		 * If we failed to generate a signal for any reason,
258 		 * generate one here.  (This should be impossible.)
259 		 */
260 		if (WARN_ON_ONCE(!sigismember(&tsk->pending.signal, SIGBUS) &&
261 				 !sigismember(&tsk->pending.signal, SIGSEGV)))
262 			goto sigsegv;
263 
264 		return true;  /* Don't emulate the ret. */
265 	}
266 
267 	regs->ax = ret;
268 
269 do_ret:
270 	/* Emulate a ret instruction. */
271 	regs->ip = caller;
272 	regs->sp += 8;
273 	return true;
274 
275 sigsegv:
276 	force_sig(SIGSEGV, current);
277 	return true;
278 }
279 
280 /*
281  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
282  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
283  * not need special handling anymore:
284  */
285 static const char *gate_vma_name(struct vm_area_struct *vma)
286 {
287 	return "[vsyscall]";
288 }
289 static const struct vm_operations_struct gate_vma_ops = {
290 	.name = gate_vma_name,
291 };
292 static struct vm_area_struct gate_vma = {
293 	.vm_start	= VSYSCALL_ADDR,
294 	.vm_end		= VSYSCALL_ADDR + PAGE_SIZE,
295 	.vm_page_prot	= PAGE_READONLY_EXEC,
296 	.vm_flags	= VM_READ | VM_EXEC,
297 	.vm_ops		= &gate_vma_ops,
298 };
299 
300 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
301 {
302 #ifdef CONFIG_COMPAT
303 	if (!mm || mm->context.ia32_compat)
304 		return NULL;
305 #endif
306 	if (vsyscall_mode == NONE)
307 		return NULL;
308 	return &gate_vma;
309 }
310 
311 int in_gate_area(struct mm_struct *mm, unsigned long addr)
312 {
313 	struct vm_area_struct *vma = get_gate_vma(mm);
314 
315 	if (!vma)
316 		return 0;
317 
318 	return (addr >= vma->vm_start) && (addr < vma->vm_end);
319 }
320 
321 /*
322  * Use this when you have no reliable mm, typically from interrupt
323  * context. It is less reliable than using a task's mm and may give
324  * false positives.
325  */
326 int in_gate_area_no_mm(unsigned long addr)
327 {
328 	return vsyscall_mode != NONE && (addr & PAGE_MASK) == VSYSCALL_ADDR;
329 }
330 
331 void __init map_vsyscall(void)
332 {
333 	extern char __vsyscall_page;
334 	unsigned long physaddr_vsyscall = __pa_symbol(&__vsyscall_page);
335 
336 	if (vsyscall_mode != NONE)
337 		__set_fixmap(VSYSCALL_PAGE, physaddr_vsyscall,
338 			     vsyscall_mode == NATIVE
339 			     ? PAGE_KERNEL_VSYSCALL
340 			     : PAGE_KERNEL_VVAR);
341 
342 	BUILD_BUG_ON((unsigned long)__fix_to_virt(VSYSCALL_PAGE) !=
343 		     (unsigned long)VSYSCALL_ADDR);
344 }
345