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, NONE } vsyscall_mode =
  46 #ifdef 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("none", str))
  58                         vsyscall_mode = NONE;
  59                 else
  60                         return -EINVAL;
  61
  62                 return 0;
  63         }
  64
  65         return -EINVAL;
  66 }
  67 early_param("vsyscall", vsyscall_setup);
  68
  69 static void warn_bad_vsyscall(const char *level, struct pt_regs *regs,
  70                               const char *message)
  71 {
  72         if (!show_unhandled_signals)
  73                 return;
  74
  75         printk_ratelimited("%s%s[%d] %s ip:%lx cs:%lx sp:%lx ax:%lx si:%lx di:%lx\n",
  76                            level, current->comm, task_pid_nr(current),
  77                            message, regs->ip, regs->cs,
  78                            regs->sp, regs->ax, regs->si, regs->di);
  79 }
  80
  81 static int addr_to_vsyscall_nr(unsigned long addr)
  82 {
  83         int nr;
  84
  85         if ((addr & ~0xC00UL) != VSYSCALL_ADDR)
  86                 return -EINVAL;
  87
  88         nr = (addr & 0xC00UL) >> 10;
  89         if (nr >= 3)
  90                 return -EINVAL;
  91
  92         return nr;
  93 }
  94
  95 static bool write_ok_or_segv(unsigned long ptr, size_t size)
  96 {
  97         /*
  98          * XXX: if access_ok, get_user, and put_user handled
  99          * sig_on_uaccess_err, this could go away.
 100          */
 101
 102         if (!access_ok(VERIFY_WRITE, (void __user *)ptr, size)) {
 103                 struct thread_struct *thread = &current->thread;
 104
 105                 thread->error_code      = 6;  /* user fault, no page, write */
 106                 thread->cr2             = ptr;
 107                 thread->trap_nr         = X86_TRAP_PF;
 108
 109                 force_sig_fault(SIGSEGV, SEGV_MAPERR, (void __user *)ptr, current);
 110                 return false;
 111         } else {
 112                 return true;
 113         }
 114 }
 115
 116 bool emulate_vsyscall(struct pt_regs *regs, unsigned long address)
 117 {
 118         struct task_struct *tsk;
 119         unsigned long caller;
 120         int vsyscall_nr, syscall_nr, tmp;
 121         int prev_sig_on_uaccess_err;
 122         long ret;
 123         unsigned long orig_dx;
 124
 125         /*
 126          * No point in checking CS -- the only way to get here is a user mode
 127          * trap to a high address, which means that we're in 64-bit user code.
 128          */
 129
 130         WARN_ON_ONCE(address != regs->ip);
 131
 132         if (vsyscall_mode == NONE) {
 133                 warn_bad_vsyscall(KERN_INFO, regs,
 134                                   "vsyscall attempted with vsyscall=none");
 135                 return false;
 136         }
 137
 138         vsyscall_nr = addr_to_vsyscall_nr(address);
 139
 140         trace_emulate_vsyscall(vsyscall_nr);
 141
 142         if (vsyscall_nr < 0) {
 143                 warn_bad_vsyscall(KERN_WARNING, regs,
 144                                   "misaligned vsyscall (exploit attempt or buggy program) -- look up the vsyscall kernel parameter if you need a workaround");
 145                 goto sigsegv;
 146         }
 147
 148         if (get_user(caller, (unsigned long __user *)regs->sp) != 0) {
 149                 warn_bad_vsyscall(KERN_WARNING, regs,
 150                                   "vsyscall with bad stack (exploit attempt?)");
 151                 goto sigsegv;
 152         }
 153
 154         tsk = current;
 155
 156         /*
 157          * Check for access_ok violations and find the syscall nr.
 158          *
 159          * NULL is a valid user pointer (in the access_ok sense) on 32-bit and
 160          * 64-bit, so we don't need to special-case it here.  For all the
 161          * vsyscalls, NULL means "don't write anything" not "write it at
 162          * address 0".
 163          */
 164         switch (vsyscall_nr) {
 165         case 0:
 166                 if (!write_ok_or_segv(regs->di, sizeof(struct timeval)) ||
 167                     !write_ok_or_segv(regs->si, sizeof(struct timezone))) {
 168                         ret = -EFAULT;
 169                         goto check_fault;
 170                 }
 171
 172                 syscall_nr = __NR_gettimeofday;
 173                 break;
 174
 175         case 1:
 176                 if (!write_ok_or_segv(regs->di, sizeof(time_t))) {
 177                         ret = -EFAULT;
 178                         goto check_fault;
 179                 }
 180
 181                 syscall_nr = __NR_time;
 182                 break;
 183
 184         case 2:
 185                 if (!write_ok_or_segv(regs->di, sizeof(unsigned)) ||
 186                     !write_ok_or_segv(regs->si, sizeof(unsigned))) {
 187                         ret = -EFAULT;
 188                         goto check_fault;
 189                 }
 190
 191                 syscall_nr = __NR_getcpu;
 192                 break;
 193         }
 194
 195         /*
 196          * Handle seccomp.  regs->ip must be the original value.
 197          * See seccomp_send_sigsys and Documentation/userspace-api/seccomp_filter.rst.
 198          *
 199          * We could optimize the seccomp disabled case, but performance
 200          * here doesn't matter.
 201          */
 202         regs->orig_ax = syscall_nr;
 203         regs->ax = -ENOSYS;
 204         tmp = secure_computing(NULL);
 205         if ((!tmp && regs->orig_ax != syscall_nr) || regs->ip != address) {
 206                 warn_bad_vsyscall(KERN_DEBUG, regs,
 207                                   "seccomp tried to change syscall nr or ip");
 208                 do_exit(SIGSYS);
 209         }
 210         regs->orig_ax = -1;
 211         if (tmp)
 212                 goto do_ret;  /* skip requested */
 213
 214         /*
 215          * With a real vsyscall, page faults cause SIGSEGV.  We want to
 216          * preserve that behavior to make writing exploits harder.
 217          */
 218         prev_sig_on_uaccess_err = current->thread.sig_on_uaccess_err;
 219         current->thread.sig_on_uaccess_err = 1;
 220
 221         ret = -EFAULT;
 222         switch (vsyscall_nr) {
 223         case 0:
 224                 /* this decodes regs->di and regs->si on its own */
 225                 ret = __x64_sys_gettimeofday(regs);
 226                 break;
 227
 228         case 1:
 229                 /* this decodes regs->di on its own */
 230                 ret = __x64_sys_time(regs);
 231                 break;
 232
 233         case 2:
 234                 /* while we could clobber regs->dx, we didn't in the past... */
 235                 orig_dx = regs->dx;
 236                 regs->dx = 0;
 237                 /* this decodes regs->di, regs->si and regs->dx on its own */
 238                 ret = __x64_sys_getcpu(regs);
 239                 regs->dx = orig_dx;
 240                 break;
 241         }
 242
 243         current->thread.sig_on_uaccess_err = prev_sig_on_uaccess_err;
 244
 245 check_fault:
 246         if (ret == -EFAULT) {
 247                 /* Bad news -- userspace fed a bad pointer to a vsyscall. */
 248                 warn_bad_vsyscall(KERN_INFO, regs,
 249                                   "vsyscall fault (exploit attempt?)");
 250
 251                 /*
 252                  * If we failed to generate a signal for any reason,
 253                  * generate one here.  (This should be impossible.)
 254                  */
 255                 if (WARN_ON_ONCE(!sigismember(&tsk->pending.signal, SIGBUS) &&
 256                                  !sigismember(&tsk->pending.signal, SIGSEGV)))
 257                         goto sigsegv;
 258
 259                 return true;  /* Don't emulate the ret. */
 260         }
 261
 262         regs->ax = ret;
 263
 264 do_ret:
 265         /* Emulate a ret instruction. */
 266         regs->ip = caller;
 267         regs->sp += 8;
 268         return true;
 269
 270 sigsegv:
 271         force_sig(SIGSEGV, current);
 272         return true;
 273 }
 274
 275 /*
 276  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 277  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 278  * not need special handling anymore:
 279  */
 280 static const char *gate_vma_name(struct vm_area_struct *vma)
 281 {
 282         return "[vsyscall]";
 283 }
 284 static const struct vm_operations_struct gate_vma_ops = {
 285         .name = gate_vma_name,
 286 };
 287 static struct vm_area_struct gate_vma = {
 288         .vm_start       = VSYSCALL_ADDR,
 289         .vm_end         = VSYSCALL_ADDR + PAGE_SIZE,
 290         .vm_page_prot   = PAGE_READONLY_EXEC,
 291         .vm_flags       = VM_READ | VM_EXEC,
 292         .vm_ops         = &gate_vma_ops,
 293 };
 294
 295 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 296 {
 297 #ifdef CONFIG_COMPAT
 298         if (!mm || mm->context.ia32_compat)
 299                 return NULL;
 300 #endif
 301         if (vsyscall_mode == NONE)
 302                 return NULL;
 303         return &gate_vma;
 304 }
 305
 306 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 307 {
 308         struct vm_area_struct *vma = get_gate_vma(mm);
 309
 310         if (!vma)
 311                 return 0;
 312
 313         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 314 }
 315
 316 /*
 317  * Use this when you have no reliable mm, typically from interrupt
 318  * context. It is less reliable than using a task's mm and may give
 319  * false positives.
 320  */
 321 int in_gate_area_no_mm(unsigned long addr)
 322 {
 323         return vsyscall_mode != NONE && (addr & PAGE_MASK) == VSYSCALL_ADDR;
 324 }
 325
 326 /*
 327  * The VSYSCALL page is the only user-accessible page in the kernel address
 328  * range.  Normally, the kernel page tables can have _PAGE_USER clear, but
 329  * the tables covering VSYSCALL_ADDR need _PAGE_USER set if vsyscalls
 330  * are enabled.
 331  *
 332  * Some day we may create a "minimal" vsyscall mode in which we emulate
 333  * vsyscalls but leave the page not present.  If so, we skip calling
 334  * this.
 335  */
 336 void __init set_vsyscall_pgtable_user_bits(pgd_t *root)
 337 {
 338         pgd_t *pgd;
 339         p4d_t *p4d;
 340         pud_t *pud;
 341         pmd_t *pmd;
 342
 343         pgd = pgd_offset_pgd(root, VSYSCALL_ADDR);
 344         set_pgd(pgd, __pgd(pgd_val(*pgd) | _PAGE_USER));
 345         p4d = p4d_offset(pgd, VSYSCALL_ADDR);
 346 #if CONFIG_PGTABLE_LEVELS >= 5
 347         set_p4d(p4d, __p4d(p4d_val(*p4d) | _PAGE_USER));
 348 #endif
 349         pud = pud_offset(p4d, VSYSCALL_ADDR);
 350         set_pud(pud, __pud(pud_val(*pud) | _PAGE_USER));
 351         pmd = pmd_offset(pud, VSYSCALL_ADDR);
 352         set_pmd(pmd, __pmd(pmd_val(*pmd) | _PAGE_USER));
 353 }
 354
 355 void __init map_vsyscall(void)
 356 {
 357         extern char __vsyscall_page;
 358         unsigned long physaddr_vsyscall = __pa_symbol(&__vsyscall_page);
 359
 360         if (vsyscall_mode != NONE) {
 361                 __set_fixmap(VSYSCALL_PAGE, physaddr_vsyscall,
 362                              PAGE_KERNEL_VVAR);
 363                 set_vsyscall_pgtable_user_bits(swapper_pg_dir);
 364         }
 365
 366         BUILD_BUG_ON((unsigned long)__fix_to_virt(VSYSCALL_PAGE) !=
 367                      (unsigned long)VSYSCALL_ADDR);
 368 }