arch/x86/entry/vsyscall/vsyscall_64.c

   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, NATIVE, NONE } vsyscall_mode =
  46 #if defined(CONFIG_LEGACY_VSYSCALL_NATIVE)
  47         NATIVE;
  48 #elif defined(CONFIG_LEGACY_VSYSCALL_NONE)
  49         NONE;
  50 #else
  51         EMULATE;
  52 #endif
  53
  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("native", str))
  60                         vsyscall_mode = NATIVE;
  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
  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
  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
  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(VERIFY_WRITE, (void __user *)ptr, size)) {
 107                 siginfo_t info;
 108                 struct thread_struct *thread = &current->thread;
 109
 110                 thread->error_code      = 6;  /* user fault, no page, write */
 111                 thread->cr2             = ptr;
 112                 thread->trap_nr         = X86_TRAP_PF;
 113
 114                 memset(&info, 0, sizeof(info));
 115                 info.si_signo           = SIGSEGV;
 116                 info.si_errno           = 0;
 117                 info.si_code            = SEGV_MAPERR;
 118                 info.si_addr            = (void __user *)ptr;
 119
 120                 force_sig_info(SIGSEGV, &info, current);
 121                 return false;
 122         } else {
 123                 return true;
 124         }
 125 }
 126
 127 bool emulate_vsyscall(struct pt_regs *regs, unsigned long address)
 128 {
 129         struct task_struct *tsk;
 130         unsigned long caller;
 131         int vsyscall_nr, syscall_nr, tmp;
 132         int prev_sig_on_uaccess_err;
 133         long ret;
 134
 135         /*
 136          * No point in checking CS -- the only way to get here is a user mode
 137          * trap to a high address, which means that we're in 64-bit user code.
 138          */
 139
 140         WARN_ON_ONCE(address != regs->ip);
 141
 142         /* This should be unreachable in NATIVE mode. */
 143         if (WARN_ON(vsyscall_mode == NATIVE))
 144                 return false;
 145
 146         if (vsyscall_mode == NONE) {
 147                 warn_bad_vsyscall(KERN_INFO, regs,
 148                                   "vsyscall attempted with vsyscall=none");
 149                 return false;
 150         }
 151
 152         vsyscall_nr = addr_to_vsyscall_nr(address);
 153
 154         trace_emulate_vsyscall(vsyscall_nr);
 155
 156         if (vsyscall_nr < 0) {
 157                 warn_bad_vsyscall(KERN_WARNING, regs,
 158                                   "misaligned vsyscall (exploit attempt or buggy program) -- look up the vsyscall kernel parameter if you need a workaround");
 159                 goto sigsegv;
 160         }
 161
 162         if (get_user(caller, (unsigned long __user *)regs->sp) != 0) {
 163                 warn_bad_vsyscall(KERN_WARNING, regs,
 164                                   "vsyscall with bad stack (exploit attempt?)");
 165                 goto sigsegv;
 166         }
 167
 168         tsk = current;
 169
 170         /*
 171          * Check for access_ok violations and find the syscall nr.
 172          *
 173          * NULL is a valid user pointer (in the access_ok sense) on 32-bit and
 174          * 64-bit, so we don't need to special-case it here.  For all the
 175          * vsyscalls, NULL means "don't write anything" not "write it at
 176          * address 0".
 177          */
 178         switch (vsyscall_nr) {
 179         case 0:
 180                 if (!write_ok_or_segv(regs->di, sizeof(struct timeval)) ||
 181                     !write_ok_or_segv(regs->si, sizeof(struct timezone))) {
 182                         ret = -EFAULT;
 183                         goto check_fault;
 184                 }
 185
 186                 syscall_nr = __NR_gettimeofday;
 187                 break;
 188
 189         case 1:
 190                 if (!write_ok_or_segv(regs->di, sizeof(time_t))) {
 191                         ret = -EFAULT;
 192                         goto check_fault;
 193                 }
 194
 195                 syscall_nr = __NR_time;
 196                 break;
 197
 198         case 2:
 199                 if (!write_ok_or_segv(regs->di, sizeof(unsigned)) ||
 200                     !write_ok_or_segv(regs->si, sizeof(unsigned))) {
 201                         ret = -EFAULT;
 202                         goto check_fault;
 203                 }
 204
 205                 syscall_nr = __NR_getcpu;
 206                 break;
 207         }
 208
 209         /*
 210          * Handle seccomp.  regs->ip must be the original value.
 211          * See seccomp_send_sigsys and Documentation/prctl/seccomp_filter.txt.
 212          *
 213          * We could optimize the seccomp disabled case, but performance
 214          * here doesn't matter.
 215          */
 216         regs->orig_ax = syscall_nr;
 217         regs->ax = -ENOSYS;
 218         tmp = secure_computing(NULL);
 219         if ((!tmp && regs->orig_ax != syscall_nr) || regs->ip != address) {
 220                 warn_bad_vsyscall(KERN_DEBUG, regs,
 221                                   "seccomp tried to change syscall nr or ip");
 222                 do_exit(SIGSYS);
 223         }
 224         regs->orig_ax = -1;
 225         if (tmp)
 226                 goto do_ret;  /* skip requested */
 227
 228         /*
 229          * With a real vsyscall, page faults cause SIGSEGV.  We want to
 230          * preserve that behavior to make writing exploits harder.
 231          */
 232         prev_sig_on_uaccess_err = current->thread.sig_on_uaccess_err;
 233         current->thread.sig_on_uaccess_err = 1;
 234
 235         ret = -EFAULT;
 236         switch (vsyscall_nr) {
 237         case 0:
 238                 ret = sys_gettimeofday(
 239                         (struct timeval __user *)regs->di,
 240                         (struct timezone __user *)regs->si);
 241                 break;
 242
 243         case 1:
 244                 ret = sys_time((time_t __user *)regs->di);
 245                 break;
 246
 247         case 2:
 248                 ret = sys_getcpu((unsigned __user *)regs->di,
 249                                  (unsigned __user *)regs->si,
 250                                  NULL);
 251                 break;
 252         }
 253
 254         current->thread.sig_on_uaccess_err = prev_sig_on_uaccess_err;
 255
 256 check_fault:
 257         if (ret == -EFAULT) {
 258                 /* Bad news -- userspace fed a bad pointer to a vsyscall. */
 259                 warn_bad_vsyscall(KERN_INFO, regs,
 260                                   "vsyscall fault (exploit attempt?)");
 261
 262                 /*
 263                  * If we failed to generate a signal for any reason,
 264                  * generate one here.  (This should be impossible.)
 265                  */
 266                 if (WARN_ON_ONCE(!sigismember(&tsk->pending.signal, SIGBUS) &&
 267                                  !sigismember(&tsk->pending.signal, SIGSEGV)))
 268                         goto sigsegv;
 269
 270                 return true;  /* Don't emulate the ret. */
 271         }
 272
 273         regs->ax = ret;
 274
 275 do_ret:
 276         /* Emulate a ret instruction. */
 277         regs->ip = caller;
 278         regs->sp += 8;
 279         return true;
 280
 281 sigsegv:
 282         force_sig(SIGSEGV, current);
 283         return true;
 284 }
 285
 286 /*
 287  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 288  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 289  * not need special handling anymore:
 290  */
 291 static const char *gate_vma_name(struct vm_area_struct *vma)
 292 {
 293         return "[vsyscall]";
 294 }
 295 static const struct vm_operations_struct gate_vma_ops = {
 296         .name = gate_vma_name,
 297 };
 298 static struct vm_area_struct gate_vma = {
 299         .vm_start       = VSYSCALL_ADDR,
 300         .vm_end         = VSYSCALL_ADDR + PAGE_SIZE,
 301         .vm_page_prot   = PAGE_READONLY_EXEC,
 302         .vm_flags       = VM_READ | VM_EXEC,
 303         .vm_ops         = &gate_vma_ops,
 304 };
 305
 306 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 307 {
 308 #ifdef CONFIG_COMPAT
 309         if (!mm || mm->context.ia32_compat)
 310                 return NULL;
 311 #endif
 312         if (vsyscall_mode == NONE)
 313                 return NULL;
 314         return &gate_vma;
 315 }
 316
 317 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 318 {
 319         struct vm_area_struct *vma = get_gate_vma(mm);
 320
 321         if (!vma)
 322                 return 0;
 323
 324         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 325 }
 326
 327 /*
 328  * Use this when you have no reliable mm, typically from interrupt
 329  * context. It is less reliable than using a task's mm and may give
 330  * false positives.
 331  */
 332 int in_gate_area_no_mm(unsigned long addr)
 333 {
 334         return vsyscall_mode != NONE && (addr & PAGE_MASK) == VSYSCALL_ADDR;
 335 }
 336
 337 /*
 338  * The VSYSCALL page is the only user-accessible page in the kernel address
 339  * range.  Normally, the kernel page tables can have _PAGE_USER clear, but
 340  * the tables covering VSYSCALL_ADDR need _PAGE_USER set if vsyscalls
 341  * are enabled.
 342  *
 343  * Some day we may create a "minimal" vsyscall mode in which we emulate
 344  * vsyscalls but leave the page not present.  If so, we skip calling
 345  * this.
 346  */
 347 void __init set_vsyscall_pgtable_user_bits(pgd_t *root)
 348 {
 349         pgd_t *pgd;
 350         p4d_t *p4d;
 351         pud_t *pud;
 352         pmd_t *pmd;
 353
 354         pgd = pgd_offset_pgd(root, VSYSCALL_ADDR);
 355         set_pgd(pgd, __pgd(pgd_val(*pgd) | _PAGE_USER));
 356         p4d = p4d_offset(pgd, VSYSCALL_ADDR);
 357 #if CONFIG_PGTABLE_LEVELS >= 5
 358         p4d->p4d |= _PAGE_USER;
 359 #endif
 360         pud = pud_offset(p4d, VSYSCALL_ADDR);
 361         set_pud(pud, __pud(pud_val(*pud) | _PAGE_USER));
 362         pmd = pmd_offset(pud, VSYSCALL_ADDR);
 363         set_pmd(pmd, __pmd(pmd_val(*pmd) | _PAGE_USER));
 364 }
 365
 366 void __init map_vsyscall(void)
 367 {
 368         extern char __vsyscall_page;
 369         unsigned long physaddr_vsyscall = __pa_symbol(&__vsyscall_page);
 370
 371         if (vsyscall_mode != NONE) {
 372                 __set_fixmap(VSYSCALL_PAGE, physaddr_vsyscall,
 373                              vsyscall_mode == NATIVE
 374                              ? PAGE_KERNEL_VSYSCALL
 375                              : PAGE_KERNEL_VVAR);
 376                 set_vsyscall_pgtable_user_bits(swapper_pg_dir);
 377         }
 378
 379         BUILD_BUG_ON((unsigned long)__fix_to_virt(VSYSCALL_PAGE) !=
 380                      (unsigned long)VSYSCALL_ADDR);
 381 }