1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (c) 2012-2014 Andy Lutomirski <luto@amacapital.net>
4 *
5 * Based on the original implementation which is:
6 * Copyright (C) 2001 Andrea Arcangeli <andrea@suse.de> SuSE
7 * Copyright 2003 Andi Kleen, SuSE Labs.
8 *
9 * Parts of the original code have been moved to arch/x86/vdso/vma.c
10 *
11 * This file implements vsyscall emulation. vsyscalls are a legacy ABI:
12 * Userspace can request certain kernel services by calling fixed
13 * addresses. This concept is problematic:
14 *
15 * - It interferes with ASLR.
16 * - It's awkward to write code that lives in kernel addresses but is
17 * callable by userspace at fixed addresses.
18 * - The whole concept is impossible for 32-bit compat userspace.
19 * - UML cannot easily virtualize a vsyscall.
20 *
21 * As of mid-2014, I believe that there is no new userspace code that
22 * will use a vsyscall if the vDSO is present. I hope that there will
23 * soon be no new userspace code that will ever use a vsyscall.
24 *
25 * The code in this file emulates vsyscalls when notified of a page
26 * fault to a vsyscall address.
27 */
28
29 #include <linux/kernel.h>
30 #include <linux/timer.h>
31 #include <linux/sched/signal.h>
32 #include <linux/mm_types.h>
33 #include <linux/syscalls.h>
34 #include <linux/ratelimit.h>
35
36 #include <asm/vsyscall.h>
37 #include <asm/unistd.h>
38 #include <asm/fixmap.h>
39 #include <asm/traps.h>
40 #include <asm/paravirt.h>
41
42 #define CREATE_TRACE_POINTS
43 #include "vsyscall_trace.h"
44
45 static enum { EMULATE, XONLY, NONE } vsyscall_mode __ro_after_init =
46 #ifdef CONFIG_LEGACY_VSYSCALL_NONE
47 NONE;
48 #elif defined(CONFIG_LEGACY_VSYSCALL_XONLY)
49 XONLY;
50 #else
51 EMULATE;
52 #endif
53
vsyscall_setup(char * str)54 static int __init vsyscall_setup(char *str)
55 {
56 if (str) {
57 if (!strcmp("emulate", str))
58 vsyscall_mode = EMULATE;
59 else if (!strcmp("xonly", str))
60 vsyscall_mode = XONLY;
61 else if (!strcmp("none", str))
62 vsyscall_mode = NONE;
63 else
64 return -EINVAL;
65
66 return 0;
67 }
68
69 return -EINVAL;
70 }
71 early_param("vsyscall", vsyscall_setup);
72
warn_bad_vsyscall(const char * level,struct pt_regs * regs,const char * message)73 static void warn_bad_vsyscall(const char *level, struct pt_regs *regs,
74 const char *message)
75 {
76 if (!show_unhandled_signals)
77 return;
78
79 printk_ratelimited("%s%s[%d] %s ip:%lx cs:%lx sp:%lx ax:%lx si:%lx di:%lx\n",
80 level, current->comm, task_pid_nr(current),
81 message, regs->ip, regs->cs,
82 regs->sp, regs->ax, regs->si, regs->di);
83 }
84
addr_to_vsyscall_nr(unsigned long addr)85 static int addr_to_vsyscall_nr(unsigned long addr)
86 {
87 int nr;
88
89 if ((addr & ~0xC00UL) != VSYSCALL_ADDR)
90 return -EINVAL;
91
92 nr = (addr & 0xC00UL) >> 10;
93 if (nr >= 3)
94 return -EINVAL;
95
96 return nr;
97 }
98
write_ok_or_segv(unsigned long ptr,size_t size)99 static bool write_ok_or_segv(unsigned long ptr, size_t size)
100 {
101 /*
102 * XXX: if access_ok, get_user, and put_user handled
103 * sig_on_uaccess_err, this could go away.
104 */
105
106 if (!access_ok((void __user *)ptr, size)) {
107 struct thread_struct *thread = ¤t->thread;
108
109 thread->error_code = X86_PF_USER | X86_PF_WRITE;
110 thread->cr2 = ptr;
111 thread->trap_nr = X86_TRAP_PF;
112
113 force_sig_fault(SIGSEGV, SEGV_MAPERR, (void __user *)ptr);
114 return false;
115 } else {
116 return true;
117 }
118 }
119
emulate_vsyscall(unsigned long error_code,struct pt_regs * regs,unsigned long address)120 bool emulate_vsyscall(unsigned long error_code,
121 struct pt_regs *regs, unsigned long address)
122 {
123 struct task_struct *tsk;
124 unsigned long caller;
125 int vsyscall_nr, syscall_nr, tmp;
126 int prev_sig_on_uaccess_err;
127 long ret;
128 unsigned long orig_dx;
129
130 /* Write faults or kernel-privilege faults never get fixed up. */
131 if ((error_code & (X86_PF_WRITE | X86_PF_USER)) != X86_PF_USER)
132 return false;
133
134 if (!(error_code & X86_PF_INSTR)) {
135 /* Failed vsyscall read */
136 if (vsyscall_mode == EMULATE)
137 return false;
138
139 /*
140 * User code tried and failed to read the vsyscall page.
141 */
142 warn_bad_vsyscall(KERN_INFO, regs, "vsyscall read attempt denied -- look up the vsyscall kernel parameter if you need a workaround");
143 return false;
144 }
145
146 /*
147 * No point in checking CS -- the only way to get here is a user mode
148 * trap to a high address, which means that we're in 64-bit user code.
149 */
150
151 WARN_ON_ONCE(address != regs->ip);
152
153 if (vsyscall_mode == NONE) {
154 warn_bad_vsyscall(KERN_INFO, regs,
155 "vsyscall attempted with vsyscall=none");
156 return false;
157 }
158
159 vsyscall_nr = addr_to_vsyscall_nr(address);
160
161 trace_emulate_vsyscall(vsyscall_nr);
162
163 if (vsyscall_nr < 0) {
164 warn_bad_vsyscall(KERN_WARNING, regs,
165 "misaligned vsyscall (exploit attempt or buggy program) -- look up the vsyscall kernel parameter if you need a workaround");
166 goto sigsegv;
167 }
168
169 if (get_user(caller, (unsigned long __user *)regs->sp) != 0) {
170 warn_bad_vsyscall(KERN_WARNING, regs,
171 "vsyscall with bad stack (exploit attempt?)");
172 goto sigsegv;
173 }
174
175 tsk = current;
176
177 /*
178 * Check for access_ok violations and find the syscall nr.
179 *
180 * NULL is a valid user pointer (in the access_ok sense) on 32-bit and
181 * 64-bit, so we don't need to special-case it here. For all the
182 * vsyscalls, NULL means "don't write anything" not "write it at
183 * address 0".
184 */
185 switch (vsyscall_nr) {
186 case 0:
187 if (!write_ok_or_segv(regs->di, sizeof(struct __kernel_old_timeval)) ||
188 !write_ok_or_segv(regs->si, sizeof(struct timezone))) {
189 ret = -EFAULT;
190 goto check_fault;
191 }
192
193 syscall_nr = __NR_gettimeofday;
194 break;
195
196 case 1:
197 if (!write_ok_or_segv(regs->di, sizeof(__kernel_old_time_t))) {
198 ret = -EFAULT;
199 goto check_fault;
200 }
201
202 syscall_nr = __NR_time;
203 break;
204
205 case 2:
206 if (!write_ok_or_segv(regs->di, sizeof(unsigned)) ||
207 !write_ok_or_segv(regs->si, sizeof(unsigned))) {
208 ret = -EFAULT;
209 goto check_fault;
210 }
211
212 syscall_nr = __NR_getcpu;
213 break;
214 }
215
216 /*
217 * Handle seccomp. regs->ip must be the original value.
218 * See seccomp_send_sigsys and Documentation/userspace-api/seccomp_filter.rst.
219 *
220 * We could optimize the seccomp disabled case, but performance
221 * here doesn't matter.
222 */
223 regs->orig_ax = syscall_nr;
224 regs->ax = -ENOSYS;
225 tmp = secure_computing();
226 if ((!tmp && regs->orig_ax != syscall_nr) || regs->ip != address) {
227 warn_bad_vsyscall(KERN_DEBUG, regs,
228 "seccomp tried to change syscall nr or ip");
229 do_exit(SIGSYS);
230 }
231 regs->orig_ax = -1;
232 if (tmp)
233 goto do_ret; /* skip requested */
234
235 /*
236 * With a real vsyscall, page faults cause SIGSEGV. We want to
237 * preserve that behavior to make writing exploits harder.
238 */
239 prev_sig_on_uaccess_err = current->thread.sig_on_uaccess_err;
240 current->thread.sig_on_uaccess_err = 1;
241
242 ret = -EFAULT;
243 switch (vsyscall_nr) {
244 case 0:
245 /* this decodes regs->di and regs->si on its own */
246 ret = __x64_sys_gettimeofday(regs);
247 break;
248
249 case 1:
250 /* this decodes regs->di on its own */
251 ret = __x64_sys_time(regs);
252 break;
253
254 case 2:
255 /* while we could clobber regs->dx, we didn't in the past... */
256 orig_dx = regs->dx;
257 regs->dx = 0;
258 /* this decodes regs->di, regs->si and regs->dx on its own */
259 ret = __x64_sys_getcpu(regs);
260 regs->dx = orig_dx;
261 break;
262 }
263
264 current->thread.sig_on_uaccess_err = prev_sig_on_uaccess_err;
265
266 check_fault:
267 if (ret == -EFAULT) {
268 /* Bad news -- userspace fed a bad pointer to a vsyscall. */
269 warn_bad_vsyscall(KERN_INFO, regs,
270 "vsyscall fault (exploit attempt?)");
271
272 /*
273 * If we failed to generate a signal for any reason,
274 * generate one here. (This should be impossible.)
275 */
276 if (WARN_ON_ONCE(!sigismember(&tsk->pending.signal, SIGBUS) &&
277 !sigismember(&tsk->pending.signal, SIGSEGV)))
278 goto sigsegv;
279
280 return true; /* Don't emulate the ret. */
281 }
282
283 regs->ax = ret;
284
285 do_ret:
286 /* Emulate a ret instruction. */
287 regs->ip = caller;
288 regs->sp += 8;
289 return true;
290
291 sigsegv:
292 force_sig(SIGSEGV);
293 return true;
294 }
295
296 /*
297 * A pseudo VMA to allow ptrace access for the vsyscall page. This only
298 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
299 * not need special handling anymore:
300 */
gate_vma_name(struct vm_area_struct * vma)301 static const char *gate_vma_name(struct vm_area_struct *vma)
302 {
303 return "[vsyscall]";
304 }
305 static const struct vm_operations_struct gate_vma_ops = {
306 .name = gate_vma_name,
307 };
308 static struct vm_area_struct gate_vma __ro_after_init = {
309 .vm_start = VSYSCALL_ADDR,
310 .vm_end = VSYSCALL_ADDR + PAGE_SIZE,
311 .vm_page_prot = PAGE_READONLY_EXEC,
312 .vm_flags = VM_READ | VM_EXEC,
313 .vm_ops = &gate_vma_ops,
314 };
315
get_gate_vma(struct mm_struct * mm)316 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
317 {
318 #ifdef CONFIG_COMPAT
319 if (!mm || mm->context.ia32_compat)
320 return NULL;
321 #endif
322 if (vsyscall_mode == NONE)
323 return NULL;
324 return &gate_vma;
325 }
326
in_gate_area(struct mm_struct * mm,unsigned long addr)327 int in_gate_area(struct mm_struct *mm, unsigned long addr)
328 {
329 struct vm_area_struct *vma = get_gate_vma(mm);
330
331 if (!vma)
332 return 0;
333
334 return (addr >= vma->vm_start) && (addr < vma->vm_end);
335 }
336
337 /*
338 * Use this when you have no reliable mm, typically from interrupt
339 * context. It is less reliable than using a task's mm and may give
340 * false positives.
341 */
in_gate_area_no_mm(unsigned long addr)342 int in_gate_area_no_mm(unsigned long addr)
343 {
344 return vsyscall_mode != NONE && (addr & PAGE_MASK) == VSYSCALL_ADDR;
345 }
346
347 /*
348 * The VSYSCALL page is the only user-accessible page in the kernel address
349 * range. Normally, the kernel page tables can have _PAGE_USER clear, but
350 * the tables covering VSYSCALL_ADDR need _PAGE_USER set if vsyscalls
351 * are enabled.
352 *
353 * Some day we may create a "minimal" vsyscall mode in which we emulate
354 * vsyscalls but leave the page not present. If so, we skip calling
355 * this.
356 */
set_vsyscall_pgtable_user_bits(pgd_t * root)357 void __init set_vsyscall_pgtable_user_bits(pgd_t *root)
358 {
359 pgd_t *pgd;
360 p4d_t *p4d;
361 pud_t *pud;
362 pmd_t *pmd;
363
364 pgd = pgd_offset_pgd(root, VSYSCALL_ADDR);
365 set_pgd(pgd, __pgd(pgd_val(*pgd) | _PAGE_USER));
366 p4d = p4d_offset(pgd, VSYSCALL_ADDR);
367 #if CONFIG_PGTABLE_LEVELS >= 5
368 set_p4d(p4d, __p4d(p4d_val(*p4d) | _PAGE_USER));
369 #endif
370 pud = pud_offset(p4d, VSYSCALL_ADDR);
371 set_pud(pud, __pud(pud_val(*pud) | _PAGE_USER));
372 pmd = pmd_offset(pud, VSYSCALL_ADDR);
373 set_pmd(pmd, __pmd(pmd_val(*pmd) | _PAGE_USER));
374 }
375
map_vsyscall(void)376 void __init map_vsyscall(void)
377 {
378 extern char __vsyscall_page;
379 unsigned long physaddr_vsyscall = __pa_symbol(&__vsyscall_page);
380
381 /*
382 * For full emulation, the page needs to exist for real. In
383 * execute-only mode, there is no PTE at all backing the vsyscall
384 * page.
385 */
386 if (vsyscall_mode == EMULATE) {
387 __set_fixmap(VSYSCALL_PAGE, physaddr_vsyscall,
388 PAGE_KERNEL_VVAR);
389 set_vsyscall_pgtable_user_bits(swapper_pg_dir);
390 }
391
392 if (vsyscall_mode == XONLY)
393 gate_vma.vm_flags = VM_EXEC;
394
395 BUILD_BUG_ON((unsigned long)__fix_to_virt(VSYSCALL_PAGE) !=
396 (unsigned long)VSYSCALL_ADDR);
397 }
398